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
)
355 TCP_SKB_CB(skb
)->flags
= flags
;
356 TCP_SKB_CB(skb
)->sacked
= 0;
358 skb_shinfo(skb
)->gso_segs
= 1;
359 skb_shinfo(skb
)->gso_size
= 0;
360 skb_shinfo(skb
)->gso_type
= 0;
362 TCP_SKB_CB(skb
)->seq
= seq
;
363 if (flags
& (TCPCB_FLAG_SYN
| TCPCB_FLAG_FIN
))
365 TCP_SKB_CB(skb
)->end_seq
= seq
;
368 static inline int tcp_urg_mode(const struct tcp_sock
*tp
)
370 return tp
->snd_una
!= tp
->snd_up
;
373 #define OPTION_SACK_ADVERTISE (1 << 0)
374 #define OPTION_TS (1 << 1)
375 #define OPTION_MD5 (1 << 2)
376 #define OPTION_WSCALE (1 << 3)
377 #define OPTION_COOKIE_EXTENSION (1 << 4)
379 struct tcp_out_options
{
380 u8 options
; /* bit field of OPTION_* */
381 u8 ws
; /* window scale, 0 to disable */
382 u8 num_sack_blocks
; /* number of SACK blocks to include */
383 u8 hash_size
; /* bytes in hash_location */
384 u16 mss
; /* 0 to disable */
385 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
386 __u8
*hash_location
; /* temporary pointer, overloaded */
389 /* The sysctl int routines are generic, so check consistency here.
391 static u8
tcp_cookie_size_check(u8 desired
)
394 /* previously specified */
397 if (sysctl_tcp_cookie_size
<= 0) {
398 /* no default specified */
401 if (sysctl_tcp_cookie_size
<= TCP_COOKIE_MIN
) {
402 /* value too small, specify minimum */
403 return TCP_COOKIE_MIN
;
405 if (sysctl_tcp_cookie_size
>= TCP_COOKIE_MAX
) {
406 /* value too large, specify maximum */
407 return TCP_COOKIE_MAX
;
409 if (0x1 & sysctl_tcp_cookie_size
) {
410 /* 8-bit multiple, illegal, fix it */
411 return (u8
)(sysctl_tcp_cookie_size
+ 0x1);
413 return (u8
)sysctl_tcp_cookie_size
;
416 /* Write previously computed TCP options to the packet.
418 * Beware: Something in the Internet is very sensitive to the ordering of
419 * TCP options, we learned this through the hard way, so be careful here.
420 * Luckily we can at least blame others for their non-compliance but from
421 * inter-operatibility perspective it seems that we're somewhat stuck with
422 * the ordering which we have been using if we want to keep working with
423 * those broken things (not that it currently hurts anybody as there isn't
424 * particular reason why the ordering would need to be changed).
426 * At least SACK_PERM as the first option is known to lead to a disaster
427 * (but it may well be that other scenarios fail similarly).
429 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
430 struct tcp_out_options
*opts
)
432 u8 options
= opts
->options
; /* mungable copy */
434 /* Having both authentication and cookies for security is redundant,
435 * and there's certainly not enough room. Instead, the cookie-less
436 * extension variant is proposed.
438 * Consider the pessimal case with authentication. The options
440 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
442 if (unlikely(OPTION_MD5
& options
)) {
443 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
444 *ptr
++ = htonl((TCPOPT_COOKIE
<< 24) |
445 (TCPOLEN_COOKIE_BASE
<< 16) |
446 (TCPOPT_MD5SIG
<< 8) |
449 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
451 (TCPOPT_MD5SIG
<< 8) |
454 options
&= ~OPTION_COOKIE_EXTENSION
;
455 /* overload cookie hash location */
456 opts
->hash_location
= (__u8
*)ptr
;
460 if (unlikely(opts
->mss
)) {
461 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
462 (TCPOLEN_MSS
<< 16) |
466 if (likely(OPTION_TS
& options
)) {
467 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
468 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
469 (TCPOLEN_SACK_PERM
<< 16) |
470 (TCPOPT_TIMESTAMP
<< 8) |
472 options
&= ~OPTION_SACK_ADVERTISE
;
474 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
476 (TCPOPT_TIMESTAMP
<< 8) |
479 *ptr
++ = htonl(opts
->tsval
);
480 *ptr
++ = htonl(opts
->tsecr
);
483 /* Specification requires after timestamp, so do it now.
485 * Consider the pessimal case without authentication. The options
487 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
489 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
490 __u8
*cookie_copy
= opts
->hash_location
;
491 u8 cookie_size
= opts
->hash_size
;
493 /* 8-bit multiple handled in tcp_cookie_size_check() above,
496 if (0x2 & cookie_size
) {
497 __u8
*p
= (__u8
*)ptr
;
499 /* 16-bit multiple */
500 *p
++ = TCPOPT_COOKIE
;
501 *p
++ = TCPOLEN_COOKIE_BASE
+ cookie_size
;
502 *p
++ = *cookie_copy
++;
503 *p
++ = *cookie_copy
++;
507 /* 32-bit multiple */
508 *ptr
++ = htonl(((TCPOPT_NOP
<< 24) |
510 (TCPOPT_COOKIE
<< 8) |
511 TCPOLEN_COOKIE_BASE
) +
515 if (cookie_size
> 0) {
516 memcpy(ptr
, cookie_copy
, cookie_size
);
517 ptr
+= (cookie_size
/ 4);
521 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
522 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
524 (TCPOPT_SACK_PERM
<< 8) |
528 if (unlikely(OPTION_WSCALE
& options
)) {
529 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
530 (TCPOPT_WINDOW
<< 16) |
531 (TCPOLEN_WINDOW
<< 8) |
535 if (unlikely(opts
->num_sack_blocks
)) {
536 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
537 tp
->duplicate_sack
: tp
->selective_acks
;
540 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
543 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
544 TCPOLEN_SACK_PERBLOCK
)));
546 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
548 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
549 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
552 tp
->rx_opt
.dsack
= 0;
556 /* Compute TCP options for SYN packets. This is not the final
557 * network wire format yet.
559 static unsigned tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
560 struct tcp_out_options
*opts
,
561 struct tcp_md5sig_key
**md5
) {
562 struct tcp_sock
*tp
= tcp_sk(sk
);
563 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
564 unsigned remaining
= MAX_TCP_OPTION_SPACE
;
565 u8 cookie_size
= (!tp
->rx_opt
.cookie_out_never
&& cvp
!= NULL
) ?
566 tcp_cookie_size_check(cvp
->cookie_desired
) :
569 #ifdef CONFIG_TCP_MD5SIG
570 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
572 opts
->options
|= OPTION_MD5
;
573 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
579 /* We always get an MSS option. The option bytes which will be seen in
580 * normal data packets should timestamps be used, must be in the MSS
581 * advertised. But we subtract them from tp->mss_cache so that
582 * calculations in tcp_sendmsg are simpler etc. So account for this
583 * fact here if necessary. If we don't do this correctly, as a
584 * receiver we won't recognize data packets as being full sized when we
585 * should, and thus we won't abide by the delayed ACK rules correctly.
586 * SACKs don't matter, we never delay an ACK when we have any of those
588 opts
->mss
= tcp_advertise_mss(sk
);
589 remaining
-= TCPOLEN_MSS_ALIGNED
;
591 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
592 opts
->options
|= OPTION_TS
;
593 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
594 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
595 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
597 if (likely(sysctl_tcp_window_scaling
)) {
598 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
599 opts
->options
|= OPTION_WSCALE
;
600 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
602 if (likely(sysctl_tcp_sack
)) {
603 opts
->options
|= OPTION_SACK_ADVERTISE
;
604 if (unlikely(!(OPTION_TS
& opts
->options
)))
605 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
608 /* Note that timestamps are required by the specification.
610 * Odd numbers of bytes are prohibited by the specification, ensuring
611 * that the cookie is 16-bit aligned, and the resulting cookie pair is
615 (OPTION_TS
& opts
->options
) &&
617 int need
= TCPOLEN_COOKIE_BASE
+ cookie_size
;
620 /* 32-bit multiple */
621 need
+= 2; /* NOPs */
623 if (need
> remaining
) {
624 /* try shrinking cookie to fit */
629 while (need
> remaining
&& TCP_COOKIE_MIN
<= cookie_size
) {
633 if (TCP_COOKIE_MIN
<= cookie_size
) {
634 opts
->options
|= OPTION_COOKIE_EXTENSION
;
635 opts
->hash_location
= (__u8
*)&cvp
->cookie_pair
[0];
636 opts
->hash_size
= cookie_size
;
638 /* Remember for future incarnations. */
639 cvp
->cookie_desired
= cookie_size
;
641 if (cvp
->cookie_desired
!= cvp
->cookie_pair_size
) {
642 /* Currently use random bytes as a nonce,
643 * assuming these are completely unpredictable
644 * by hostile users of the same system.
646 get_random_bytes(&cvp
->cookie_pair
[0],
648 cvp
->cookie_pair_size
= cookie_size
;
654 return MAX_TCP_OPTION_SPACE
- remaining
;
657 /* Set up TCP options for SYN-ACKs. */
658 static unsigned tcp_synack_options(struct sock
*sk
,
659 struct request_sock
*req
,
660 unsigned mss
, struct sk_buff
*skb
,
661 struct tcp_out_options
*opts
,
662 struct tcp_md5sig_key
**md5
,
663 struct tcp_extend_values
*xvp
)
665 struct inet_request_sock
*ireq
= inet_rsk(req
);
666 unsigned remaining
= MAX_TCP_OPTION_SPACE
;
667 u8 cookie_plus
= (xvp
!= NULL
&& !xvp
->cookie_out_never
) ?
670 bool doing_ts
= ireq
->tstamp_ok
;
672 #ifdef CONFIG_TCP_MD5SIG
673 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
675 opts
->options
|= OPTION_MD5
;
676 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
678 /* We can't fit any SACK blocks in a packet with MD5 + TS
679 * options. There was discussion about disabling SACK
680 * rather than TS in order to fit in better with old,
681 * buggy kernels, but that was deemed to be unnecessary.
683 doing_ts
&= !ireq
->sack_ok
;
689 /* We always send an MSS option. */
691 remaining
-= TCPOLEN_MSS_ALIGNED
;
693 if (likely(ireq
->wscale_ok
)) {
694 opts
->ws
= ireq
->rcv_wscale
;
695 opts
->options
|= OPTION_WSCALE
;
696 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
698 if (likely(doing_ts
)) {
699 opts
->options
|= OPTION_TS
;
700 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
701 opts
->tsecr
= req
->ts_recent
;
702 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
704 if (likely(ireq
->sack_ok
)) {
705 opts
->options
|= OPTION_SACK_ADVERTISE
;
706 if (unlikely(!doing_ts
))
707 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
710 /* Similar rationale to tcp_syn_options() applies here, too.
711 * If the <SYN> options fit, the same options should fit now!
715 cookie_plus
> TCPOLEN_COOKIE_BASE
) {
716 int need
= cookie_plus
; /* has TCPOLEN_COOKIE_BASE */
719 /* 32-bit multiple */
720 need
+= 2; /* NOPs */
722 if (need
<= remaining
) {
723 opts
->options
|= OPTION_COOKIE_EXTENSION
;
724 opts
->hash_size
= cookie_plus
- TCPOLEN_COOKIE_BASE
;
727 /* There's no error return, so flag it. */
728 xvp
->cookie_out_never
= 1; /* true */
732 return MAX_TCP_OPTION_SPACE
- remaining
;
735 /* Compute TCP options for ESTABLISHED sockets. This is not the
736 * final wire format yet.
738 static unsigned tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
739 struct tcp_out_options
*opts
,
740 struct tcp_md5sig_key
**md5
) {
741 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
742 struct tcp_sock
*tp
= tcp_sk(sk
);
744 unsigned int eff_sacks
;
746 #ifdef CONFIG_TCP_MD5SIG
747 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
748 if (unlikely(*md5
)) {
749 opts
->options
|= OPTION_MD5
;
750 size
+= TCPOLEN_MD5SIG_ALIGNED
;
756 if (likely(tp
->rx_opt
.tstamp_ok
)) {
757 opts
->options
|= OPTION_TS
;
758 opts
->tsval
= tcb
? tcb
->when
: 0;
759 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
760 size
+= TCPOLEN_TSTAMP_ALIGNED
;
763 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
764 if (unlikely(eff_sacks
)) {
765 const unsigned remaining
= MAX_TCP_OPTION_SPACE
- size
;
766 opts
->num_sack_blocks
=
767 min_t(unsigned, eff_sacks
,
768 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
769 TCPOLEN_SACK_PERBLOCK
);
770 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
771 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
777 /* This routine actually transmits TCP packets queued in by
778 * tcp_do_sendmsg(). This is used by both the initial
779 * transmission and possible later retransmissions.
780 * All SKB's seen here are completely headerless. It is our
781 * job to build the TCP header, and pass the packet down to
782 * IP so it can do the same plus pass the packet off to the
785 * We are working here with either a clone of the original
786 * SKB, or a fresh unique copy made by the retransmit engine.
788 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
791 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
792 struct inet_sock
*inet
;
794 struct tcp_skb_cb
*tcb
;
795 struct tcp_out_options opts
;
796 unsigned tcp_options_size
, tcp_header_size
;
797 struct tcp_md5sig_key
*md5
;
801 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
803 /* If congestion control is doing timestamping, we must
804 * take such a timestamp before we potentially clone/copy.
806 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
807 __net_timestamp(skb
);
809 if (likely(clone_it
)) {
810 if (unlikely(skb_cloned(skb
)))
811 skb
= pskb_copy(skb
, gfp_mask
);
813 skb
= skb_clone(skb
, gfp_mask
);
820 tcb
= TCP_SKB_CB(skb
);
821 memset(&opts
, 0, sizeof(opts
));
823 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
))
824 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
826 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
828 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
830 if (tcp_packets_in_flight(tp
) == 0)
831 tcp_ca_event(sk
, CA_EVENT_TX_START
);
833 skb_push(skb
, tcp_header_size
);
834 skb_reset_transport_header(skb
);
835 skb_set_owner_w(skb
, sk
);
837 /* Build TCP header and checksum it. */
839 th
->source
= inet
->inet_sport
;
840 th
->dest
= inet
->inet_dport
;
841 th
->seq
= htonl(tcb
->seq
);
842 th
->ack_seq
= htonl(tp
->rcv_nxt
);
843 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
846 if (unlikely(tcb
->flags
& TCPCB_FLAG_SYN
)) {
847 /* RFC1323: The window in SYN & SYN/ACK segments
850 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
852 th
->window
= htons(tcp_select_window(sk
));
857 /* The urg_mode check is necessary during a below snd_una win probe */
858 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
859 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
860 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
862 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
863 th
->urg_ptr
= 0xFFFF;
868 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
869 if (likely((tcb
->flags
& TCPCB_FLAG_SYN
) == 0))
870 TCP_ECN_send(sk
, skb
, tcp_header_size
);
872 #ifdef CONFIG_TCP_MD5SIG
873 /* Calculate the MD5 hash, as we have all we need now */
875 sk
->sk_route_caps
&= ~NETIF_F_GSO_MASK
;
876 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
881 icsk
->icsk_af_ops
->send_check(sk
, skb
->len
, skb
);
883 if (likely(tcb
->flags
& TCPCB_FLAG_ACK
))
884 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
886 if (skb
->len
!= tcp_header_size
)
887 tcp_event_data_sent(tp
, skb
, sk
);
889 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
890 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
892 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, 0);
893 if (likely(err
<= 0))
896 tcp_enter_cwr(sk
, 1);
898 return net_xmit_eval(err
);
901 /* This routine just queues the buffer for sending.
903 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
904 * otherwise socket can stall.
906 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
908 struct tcp_sock
*tp
= tcp_sk(sk
);
910 /* Advance write_seq and place onto the write_queue. */
911 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
912 skb_header_release(skb
);
913 tcp_add_write_queue_tail(sk
, skb
);
914 sk
->sk_wmem_queued
+= skb
->truesize
;
915 sk_mem_charge(sk
, skb
->truesize
);
918 /* Initialize TSO segments for a packet. */
919 static void tcp_set_skb_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
920 unsigned int mss_now
)
922 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
) ||
923 skb
->ip_summed
== CHECKSUM_NONE
) {
924 /* Avoid the costly divide in the normal
927 skb_shinfo(skb
)->gso_segs
= 1;
928 skb_shinfo(skb
)->gso_size
= 0;
929 skb_shinfo(skb
)->gso_type
= 0;
931 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
932 skb_shinfo(skb
)->gso_size
= mss_now
;
933 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
937 /* When a modification to fackets out becomes necessary, we need to check
938 * skb is counted to fackets_out or not.
940 static void tcp_adjust_fackets_out(struct sock
*sk
, struct sk_buff
*skb
,
943 struct tcp_sock
*tp
= tcp_sk(sk
);
945 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
948 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
949 tp
->fackets_out
-= decr
;
952 /* Pcount in the middle of the write queue got changed, we need to do various
953 * tweaks to fix counters
955 static void tcp_adjust_pcount(struct sock
*sk
, struct sk_buff
*skb
, int decr
)
957 struct tcp_sock
*tp
= tcp_sk(sk
);
959 tp
->packets_out
-= decr
;
961 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
962 tp
->sacked_out
-= decr
;
963 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
964 tp
->retrans_out
-= decr
;
965 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
966 tp
->lost_out
-= decr
;
968 /* Reno case is special. Sigh... */
969 if (tcp_is_reno(tp
) && decr
> 0)
970 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
972 tcp_adjust_fackets_out(sk
, skb
, decr
);
974 if (tp
->lost_skb_hint
&&
975 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
976 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
977 tp
->lost_cnt_hint
-= decr
;
979 tcp_verify_left_out(tp
);
982 /* Function to create two new TCP segments. Shrinks the given segment
983 * to the specified size and appends a new segment with the rest of the
984 * packet to the list. This won't be called frequently, I hope.
985 * Remember, these are still headerless SKBs at this point.
987 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
988 unsigned int mss_now
)
990 struct tcp_sock
*tp
= tcp_sk(sk
);
991 struct sk_buff
*buff
;
992 int nsize
, old_factor
;
996 BUG_ON(len
> skb
->len
);
998 nsize
= skb_headlen(skb
) - len
;
1002 if (skb_cloned(skb
) &&
1003 skb_is_nonlinear(skb
) &&
1004 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1007 /* Get a new skb... force flag on. */
1008 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
1010 return -ENOMEM
; /* We'll just try again later. */
1012 sk
->sk_wmem_queued
+= buff
->truesize
;
1013 sk_mem_charge(sk
, buff
->truesize
);
1014 nlen
= skb
->len
- len
- nsize
;
1015 buff
->truesize
+= nlen
;
1016 skb
->truesize
-= nlen
;
1018 /* Correct the sequence numbers. */
1019 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1020 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1021 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1023 /* PSH and FIN should only be set in the second packet. */
1024 flags
= TCP_SKB_CB(skb
)->flags
;
1025 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
1026 TCP_SKB_CB(buff
)->flags
= flags
;
1027 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1029 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1030 /* Copy and checksum data tail into the new buffer. */
1031 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1032 skb_put(buff
, nsize
),
1037 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1039 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1040 skb_split(skb
, buff
, len
);
1043 buff
->ip_summed
= skb
->ip_summed
;
1045 /* Looks stupid, but our code really uses when of
1046 * skbs, which it never sent before. --ANK
1048 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
1049 buff
->tstamp
= skb
->tstamp
;
1051 old_factor
= tcp_skb_pcount(skb
);
1053 /* Fix up tso_factor for both original and new SKB. */
1054 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1055 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1057 /* If this packet has been sent out already, we must
1058 * adjust the various packet counters.
1060 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1061 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1062 tcp_skb_pcount(buff
);
1065 tcp_adjust_pcount(sk
, skb
, diff
);
1068 /* Link BUFF into the send queue. */
1069 skb_header_release(buff
);
1070 tcp_insert_write_queue_after(skb
, buff
, sk
);
1075 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1076 * eventually). The difference is that pulled data not copied, but
1077 * immediately discarded.
1079 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1085 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
1086 if (skb_shinfo(skb
)->frags
[i
].size
<= eat
) {
1087 put_page(skb_shinfo(skb
)->frags
[i
].page
);
1088 eat
-= skb_shinfo(skb
)->frags
[i
].size
;
1090 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
1092 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
1093 skb_shinfo(skb
)->frags
[k
].size
-= eat
;
1099 skb_shinfo(skb
)->nr_frags
= k
;
1101 skb_reset_tail_pointer(skb
);
1102 skb
->data_len
-= len
;
1103 skb
->len
= skb
->data_len
;
1106 /* Remove acked data from a packet in the transmit queue. */
1107 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1109 if (skb_cloned(skb
) && pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1112 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
1113 if (unlikely(len
< skb_headlen(skb
)))
1114 __skb_pull(skb
, len
);
1116 __pskb_trim_head(skb
, len
- skb_headlen(skb
));
1118 TCP_SKB_CB(skb
)->seq
+= len
;
1119 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1121 skb
->truesize
-= len
;
1122 sk
->sk_wmem_queued
-= len
;
1123 sk_mem_uncharge(sk
, len
);
1124 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1126 /* Any change of skb->len requires recalculation of tso
1129 if (tcp_skb_pcount(skb
) > 1)
1130 tcp_set_skb_tso_segs(sk
, skb
, tcp_current_mss(sk
));
1135 /* Calculate MSS. Not accounting for SACKs here. */
1136 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1138 struct tcp_sock
*tp
= tcp_sk(sk
);
1139 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1142 /* Calculate base mss without TCP options:
1143 It is MMS_S - sizeof(tcphdr) of rfc1122
1145 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1147 /* Clamp it (mss_clamp does not include tcp options) */
1148 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1149 mss_now
= tp
->rx_opt
.mss_clamp
;
1151 /* Now subtract optional transport overhead */
1152 mss_now
-= icsk
->icsk_ext_hdr_len
;
1154 /* Then reserve room for full set of TCP options and 8 bytes of data */
1158 /* Now subtract TCP options size, not including SACKs */
1159 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
1164 /* Inverse of above */
1165 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1167 struct tcp_sock
*tp
= tcp_sk(sk
);
1168 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1172 tp
->tcp_header_len
+
1173 icsk
->icsk_ext_hdr_len
+
1174 icsk
->icsk_af_ops
->net_header_len
;
1179 /* MTU probing init per socket */
1180 void tcp_mtup_init(struct sock
*sk
)
1182 struct tcp_sock
*tp
= tcp_sk(sk
);
1183 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1185 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
1186 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1187 icsk
->icsk_af_ops
->net_header_len
;
1188 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
1189 icsk
->icsk_mtup
.probe_size
= 0;
1192 /* This function synchronize snd mss to current pmtu/exthdr set.
1194 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1195 for TCP options, but includes only bare TCP header.
1197 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1198 It is minimum of user_mss and mss received with SYN.
1199 It also does not include TCP options.
1201 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1203 tp->mss_cache is current effective sending mss, including
1204 all tcp options except for SACKs. It is evaluated,
1205 taking into account current pmtu, but never exceeds
1206 tp->rx_opt.mss_clamp.
1208 NOTE1. rfc1122 clearly states that advertised MSS
1209 DOES NOT include either tcp or ip options.
1211 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1212 are READ ONLY outside this function. --ANK (980731)
1214 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1216 struct tcp_sock
*tp
= tcp_sk(sk
);
1217 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1220 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1221 icsk
->icsk_mtup
.search_high
= pmtu
;
1223 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1224 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1226 /* And store cached results */
1227 icsk
->icsk_pmtu_cookie
= pmtu
;
1228 if (icsk
->icsk_mtup
.enabled
)
1229 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1230 tp
->mss_cache
= mss_now
;
1235 /* Compute the current effective MSS, taking SACKs and IP options,
1236 * and even PMTU discovery events into account.
1238 unsigned int tcp_current_mss(struct sock
*sk
)
1240 struct tcp_sock
*tp
= tcp_sk(sk
);
1241 struct dst_entry
*dst
= __sk_dst_get(sk
);
1243 unsigned header_len
;
1244 struct tcp_out_options opts
;
1245 struct tcp_md5sig_key
*md5
;
1247 mss_now
= tp
->mss_cache
;
1250 u32 mtu
= dst_mtu(dst
);
1251 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1252 mss_now
= tcp_sync_mss(sk
, mtu
);
1255 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1256 sizeof(struct tcphdr
);
1257 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1258 * some common options. If this is an odd packet (because we have SACK
1259 * blocks etc) then our calculated header_len will be different, and
1260 * we have to adjust mss_now correspondingly */
1261 if (header_len
!= tp
->tcp_header_len
) {
1262 int delta
= (int) header_len
- tp
->tcp_header_len
;
1269 /* Congestion window validation. (RFC2861) */
1270 static void tcp_cwnd_validate(struct sock
*sk
)
1272 struct tcp_sock
*tp
= tcp_sk(sk
);
1274 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1275 /* Network is feed fully. */
1276 tp
->snd_cwnd_used
= 0;
1277 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1279 /* Network starves. */
1280 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1281 tp
->snd_cwnd_used
= tp
->packets_out
;
1283 if (sysctl_tcp_slow_start_after_idle
&&
1284 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1285 tcp_cwnd_application_limited(sk
);
1289 /* Returns the portion of skb which can be sent right away without
1290 * introducing MSS oddities to segment boundaries. In rare cases where
1291 * mss_now != mss_cache, we will request caller to create a small skb
1292 * per input skb which could be mostly avoided here (if desired).
1294 * We explicitly want to create a request for splitting write queue tail
1295 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1296 * thus all the complexity (cwnd_len is always MSS multiple which we
1297 * return whenever allowed by the other factors). Basically we need the
1298 * modulo only when the receiver window alone is the limiting factor or
1299 * when we would be allowed to send the split-due-to-Nagle skb fully.
1301 static unsigned int tcp_mss_split_point(struct sock
*sk
, struct sk_buff
*skb
,
1302 unsigned int mss_now
, unsigned int cwnd
)
1304 struct tcp_sock
*tp
= tcp_sk(sk
);
1305 u32 needed
, window
, cwnd_len
;
1307 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1308 cwnd_len
= mss_now
* cwnd
;
1310 if (likely(cwnd_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1313 needed
= min(skb
->len
, window
);
1315 if (cwnd_len
<= needed
)
1318 return needed
- needed
% mss_now
;
1321 /* Can at least one segment of SKB be sent right now, according to the
1322 * congestion window rules? If so, return how many segments are allowed.
1324 static inline unsigned int tcp_cwnd_test(struct tcp_sock
*tp
,
1325 struct sk_buff
*skb
)
1327 u32 in_flight
, cwnd
;
1329 /* Don't be strict about the congestion window for the final FIN. */
1330 if ((TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
1331 tcp_skb_pcount(skb
) == 1)
1334 in_flight
= tcp_packets_in_flight(tp
);
1335 cwnd
= tp
->snd_cwnd
;
1336 if (in_flight
< cwnd
)
1337 return (cwnd
- in_flight
);
1342 /* Intialize TSO state of a skb.
1343 * This must be invoked the first time we consider transmitting
1344 * SKB onto the wire.
1346 static int tcp_init_tso_segs(struct sock
*sk
, struct sk_buff
*skb
,
1347 unsigned int mss_now
)
1349 int tso_segs
= tcp_skb_pcount(skb
);
1351 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1352 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1353 tso_segs
= tcp_skb_pcount(skb
);
1358 /* Minshall's variant of the Nagle send check. */
1359 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1361 return after(tp
->snd_sml
, tp
->snd_una
) &&
1362 !after(tp
->snd_sml
, tp
->snd_nxt
);
1365 /* Return 0, if packet can be sent now without violation Nagle's rules:
1366 * 1. It is full sized.
1367 * 2. Or it contains FIN. (already checked by caller)
1368 * 3. Or TCP_NODELAY was set.
1369 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1370 * With Minshall's modification: all sent small packets are ACKed.
1372 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1373 const struct sk_buff
*skb
,
1374 unsigned mss_now
, int nonagle
)
1376 return (skb
->len
< mss_now
&&
1377 ((nonagle
& TCP_NAGLE_CORK
) ||
1378 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
))));
1381 /* Return non-zero if the Nagle test allows this packet to be
1384 static inline int tcp_nagle_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1385 unsigned int cur_mss
, int nonagle
)
1387 /* Nagle rule does not apply to frames, which sit in the middle of the
1388 * write_queue (they have no chances to get new data).
1390 * This is implemented in the callers, where they modify the 'nonagle'
1391 * argument based upon the location of SKB in the send queue.
1393 if (nonagle
& TCP_NAGLE_PUSH
)
1396 /* Don't use the nagle rule for urgent data (or for the final FIN).
1397 * Nagle can be ignored during F-RTO too (see RFC4138).
1399 if (tcp_urg_mode(tp
) || (tp
->frto_counter
== 2) ||
1400 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
))
1403 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1409 /* Does at least the first segment of SKB fit into the send window? */
1410 static inline int tcp_snd_wnd_test(struct tcp_sock
*tp
, struct sk_buff
*skb
,
1411 unsigned int cur_mss
)
1413 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1415 if (skb
->len
> cur_mss
)
1416 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1418 return !after(end_seq
, tcp_wnd_end(tp
));
1421 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1422 * should be put on the wire right now. If so, it returns the number of
1423 * packets allowed by the congestion window.
1425 static unsigned int tcp_snd_test(struct sock
*sk
, struct sk_buff
*skb
,
1426 unsigned int cur_mss
, int nonagle
)
1428 struct tcp_sock
*tp
= tcp_sk(sk
);
1429 unsigned int cwnd_quota
;
1431 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1433 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1436 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1437 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1443 /* Test if sending is allowed right now. */
1444 int tcp_may_send_now(struct sock
*sk
)
1446 struct tcp_sock
*tp
= tcp_sk(sk
);
1447 struct sk_buff
*skb
= tcp_send_head(sk
);
1450 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1451 (tcp_skb_is_last(sk
, skb
) ?
1452 tp
->nonagle
: TCP_NAGLE_PUSH
)));
1455 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1456 * which is put after SKB on the list. It is very much like
1457 * tcp_fragment() except that it may make several kinds of assumptions
1458 * in order to speed up the splitting operation. In particular, we
1459 * know that all the data is in scatter-gather pages, and that the
1460 * packet has never been sent out before (and thus is not cloned).
1462 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1463 unsigned int mss_now
)
1465 struct sk_buff
*buff
;
1466 int nlen
= skb
->len
- len
;
1469 /* All of a TSO frame must be composed of paged data. */
1470 if (skb
->len
!= skb
->data_len
)
1471 return tcp_fragment(sk
, skb
, len
, mss_now
);
1473 buff
= sk_stream_alloc_skb(sk
, 0, GFP_ATOMIC
);
1474 if (unlikely(buff
== NULL
))
1477 sk
->sk_wmem_queued
+= buff
->truesize
;
1478 sk_mem_charge(sk
, buff
->truesize
);
1479 buff
->truesize
+= nlen
;
1480 skb
->truesize
-= nlen
;
1482 /* Correct the sequence numbers. */
1483 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1484 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1485 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1487 /* PSH and FIN should only be set in the second packet. */
1488 flags
= TCP_SKB_CB(skb
)->flags
;
1489 TCP_SKB_CB(skb
)->flags
= flags
& ~(TCPCB_FLAG_FIN
| TCPCB_FLAG_PSH
);
1490 TCP_SKB_CB(buff
)->flags
= flags
;
1492 /* This packet was never sent out yet, so no SACK bits. */
1493 TCP_SKB_CB(buff
)->sacked
= 0;
1495 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1496 skb_split(skb
, buff
, len
);
1498 /* Fix up tso_factor for both original and new SKB. */
1499 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1500 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1502 /* Link BUFF into the send queue. */
1503 skb_header_release(buff
);
1504 tcp_insert_write_queue_after(skb
, buff
, sk
);
1509 /* Try to defer sending, if possible, in order to minimize the amount
1510 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1512 * This algorithm is from John Heffner.
1514 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1516 struct tcp_sock
*tp
= tcp_sk(sk
);
1517 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1518 u32 send_win
, cong_win
, limit
, in_flight
;
1520 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
)
1523 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1526 /* Defer for less than two clock ticks. */
1527 if (tp
->tso_deferred
&&
1528 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1531 in_flight
= tcp_packets_in_flight(tp
);
1533 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1535 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1537 /* From in_flight test above, we know that cwnd > in_flight. */
1538 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1540 limit
= min(send_win
, cong_win
);
1542 /* If a full-sized TSO skb can be sent, do it. */
1543 if (limit
>= sk
->sk_gso_max_size
)
1546 /* Middle in queue won't get any more data, full sendable already? */
1547 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1550 if (sysctl_tcp_tso_win_divisor
) {
1551 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1553 /* If at least some fraction of a window is available,
1556 chunk
/= sysctl_tcp_tso_win_divisor
;
1560 /* Different approach, try not to defer past a single
1561 * ACK. Receiver should ACK every other full sized
1562 * frame, so if we have space for more than 3 frames
1565 if (limit
> tcp_max_burst(tp
) * tp
->mss_cache
)
1569 /* Ok, it looks like it is advisable to defer. */
1570 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1575 tp
->tso_deferred
= 0;
1579 /* Create a new MTU probe if we are ready.
1580 * MTU probe is regularly attempting to increase the path MTU by
1581 * deliberately sending larger packets. This discovers routing
1582 * changes resulting in larger path MTUs.
1584 * Returns 0 if we should wait to probe (no cwnd available),
1585 * 1 if a probe was sent,
1588 static int tcp_mtu_probe(struct sock
*sk
)
1590 struct tcp_sock
*tp
= tcp_sk(sk
);
1591 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1592 struct sk_buff
*skb
, *nskb
, *next
;
1599 /* Not currently probing/verifying,
1601 * have enough cwnd, and
1602 * not SACKing (the variable headers throw things off) */
1603 if (!icsk
->icsk_mtup
.enabled
||
1604 icsk
->icsk_mtup
.probe_size
||
1605 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1606 tp
->snd_cwnd
< 11 ||
1607 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1610 /* Very simple search strategy: just double the MSS. */
1611 mss_now
= tcp_current_mss(sk
);
1612 probe_size
= 2 * tp
->mss_cache
;
1613 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1614 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1615 /* TODO: set timer for probe_converge_event */
1619 /* Have enough data in the send queue to probe? */
1620 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1623 if (tp
->snd_wnd
< size_needed
)
1625 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1628 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1629 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1630 if (!tcp_packets_in_flight(tp
))
1636 /* We're allowed to probe. Build it now. */
1637 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1639 sk
->sk_wmem_queued
+= nskb
->truesize
;
1640 sk_mem_charge(sk
, nskb
->truesize
);
1642 skb
= tcp_send_head(sk
);
1644 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1645 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1646 TCP_SKB_CB(nskb
)->flags
= TCPCB_FLAG_ACK
;
1647 TCP_SKB_CB(nskb
)->sacked
= 0;
1649 nskb
->ip_summed
= skb
->ip_summed
;
1651 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1654 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1655 copy
= min_t(int, skb
->len
, probe_size
- len
);
1656 if (nskb
->ip_summed
)
1657 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1659 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1660 skb_put(nskb
, copy
),
1663 if (skb
->len
<= copy
) {
1664 /* We've eaten all the data from this skb.
1666 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
;
1667 tcp_unlink_write_queue(skb
, sk
);
1668 sk_wmem_free_skb(sk
, skb
);
1670 TCP_SKB_CB(nskb
)->flags
|= TCP_SKB_CB(skb
)->flags
&
1671 ~(TCPCB_FLAG_FIN
|TCPCB_FLAG_PSH
);
1672 if (!skb_shinfo(skb
)->nr_frags
) {
1673 skb_pull(skb
, copy
);
1674 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1675 skb
->csum
= csum_partial(skb
->data
,
1678 __pskb_trim_head(skb
, copy
);
1679 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1681 TCP_SKB_CB(skb
)->seq
+= copy
;
1686 if (len
>= probe_size
)
1689 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1691 /* We're ready to send. If this fails, the probe will
1692 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1693 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1694 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1695 /* Decrement cwnd here because we are sending
1696 * effectively two packets. */
1698 tcp_event_new_data_sent(sk
, nskb
);
1700 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1701 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1702 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1710 /* This routine writes packets to the network. It advances the
1711 * send_head. This happens as incoming acks open up the remote
1714 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1715 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1716 * account rare use of URG, this is not a big flaw.
1718 * Returns 1, if no segments are in flight and we have queued segments, but
1719 * cannot send anything now because of SWS or another problem.
1721 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1722 int push_one
, gfp_t gfp
)
1724 struct tcp_sock
*tp
= tcp_sk(sk
);
1725 struct sk_buff
*skb
;
1726 unsigned int tso_segs
, sent_pkts
;
1733 /* Do MTU probing. */
1734 result
= tcp_mtu_probe(sk
);
1737 } else if (result
> 0) {
1742 while ((skb
= tcp_send_head(sk
))) {
1745 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1748 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1752 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1755 if (tso_segs
== 1) {
1756 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1757 (tcp_skb_is_last(sk
, skb
) ?
1758 nonagle
: TCP_NAGLE_PUSH
))))
1761 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
1766 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
1767 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1770 if (skb
->len
> limit
&&
1771 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
)))
1774 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1776 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
1779 /* Advance the send_head. This one is sent out.
1780 * This call will increment packets_out.
1782 tcp_event_new_data_sent(sk
, skb
);
1784 tcp_minshall_update(tp
, mss_now
, skb
);
1791 if (likely(sent_pkts
)) {
1792 tcp_cwnd_validate(sk
);
1795 return !tp
->packets_out
&& tcp_send_head(sk
);
1798 /* Push out any pending frames which were held back due to
1799 * TCP_CORK or attempt at coalescing tiny packets.
1800 * The socket must be locked by the caller.
1802 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1805 /* If we are closed, the bytes will have to remain here.
1806 * In time closedown will finish, we empty the write queue and
1807 * all will be happy.
1809 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1812 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0, GFP_ATOMIC
))
1813 tcp_check_probe_timer(sk
);
1816 /* Send _single_ skb sitting at the send head. This function requires
1817 * true push pending frames to setup probe timer etc.
1819 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1821 struct sk_buff
*skb
= tcp_send_head(sk
);
1823 BUG_ON(!skb
|| skb
->len
< mss_now
);
1825 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
1828 /* This function returns the amount that we can raise the
1829 * usable window based on the following constraints
1831 * 1. The window can never be shrunk once it is offered (RFC 793)
1832 * 2. We limit memory per socket
1835 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1836 * RECV.NEXT + RCV.WIN fixed until:
1837 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1839 * i.e. don't raise the right edge of the window until you can raise
1840 * it at least MSS bytes.
1842 * Unfortunately, the recommended algorithm breaks header prediction,
1843 * since header prediction assumes th->window stays fixed.
1845 * Strictly speaking, keeping th->window fixed violates the receiver
1846 * side SWS prevention criteria. The problem is that under this rule
1847 * a stream of single byte packets will cause the right side of the
1848 * window to always advance by a single byte.
1850 * Of course, if the sender implements sender side SWS prevention
1851 * then this will not be a problem.
1853 * BSD seems to make the following compromise:
1855 * If the free space is less than the 1/4 of the maximum
1856 * space available and the free space is less than 1/2 mss,
1857 * then set the window to 0.
1858 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1859 * Otherwise, just prevent the window from shrinking
1860 * and from being larger than the largest representable value.
1862 * This prevents incremental opening of the window in the regime
1863 * where TCP is limited by the speed of the reader side taking
1864 * data out of the TCP receive queue. It does nothing about
1865 * those cases where the window is constrained on the sender side
1866 * because the pipeline is full.
1868 * BSD also seems to "accidentally" limit itself to windows that are a
1869 * multiple of MSS, at least until the free space gets quite small.
1870 * This would appear to be a side effect of the mbuf implementation.
1871 * Combining these two algorithms results in the observed behavior
1872 * of having a fixed window size at almost all times.
1874 * Below we obtain similar behavior by forcing the offered window to
1875 * a multiple of the mss when it is feasible to do so.
1877 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1878 * Regular options like TIMESTAMP are taken into account.
1880 u32
__tcp_select_window(struct sock
*sk
)
1882 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1883 struct tcp_sock
*tp
= tcp_sk(sk
);
1884 /* MSS for the peer's data. Previous versions used mss_clamp
1885 * here. I don't know if the value based on our guesses
1886 * of peer's MSS is better for the performance. It's more correct
1887 * but may be worse for the performance because of rcv_mss
1888 * fluctuations. --SAW 1998/11/1
1890 int mss
= icsk
->icsk_ack
.rcv_mss
;
1891 int free_space
= tcp_space(sk
);
1892 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1895 if (mss
> full_space
)
1898 if (free_space
< (full_space
>> 1)) {
1899 icsk
->icsk_ack
.quick
= 0;
1901 if (tcp_memory_pressure
)
1902 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
1905 if (free_space
< mss
)
1909 if (free_space
> tp
->rcv_ssthresh
)
1910 free_space
= tp
->rcv_ssthresh
;
1912 /* Don't do rounding if we are using window scaling, since the
1913 * scaled window will not line up with the MSS boundary anyway.
1915 window
= tp
->rcv_wnd
;
1916 if (tp
->rx_opt
.rcv_wscale
) {
1917 window
= free_space
;
1919 /* Advertise enough space so that it won't get scaled away.
1920 * Import case: prevent zero window announcement if
1921 * 1<<rcv_wscale > mss.
1923 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1924 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1925 << tp
->rx_opt
.rcv_wscale
);
1927 /* Get the largest window that is a nice multiple of mss.
1928 * Window clamp already applied above.
1929 * If our current window offering is within 1 mss of the
1930 * free space we just keep it. This prevents the divide
1931 * and multiply from happening most of the time.
1932 * We also don't do any window rounding when the free space
1935 if (window
<= free_space
- mss
|| window
> free_space
)
1936 window
= (free_space
/ mss
) * mss
;
1937 else if (mss
== full_space
&&
1938 free_space
> window
+ (full_space
>> 1))
1939 window
= free_space
;
1945 /* Collapses two adjacent SKB's during retransmission. */
1946 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
1948 struct tcp_sock
*tp
= tcp_sk(sk
);
1949 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1950 int skb_size
, next_skb_size
;
1952 skb_size
= skb
->len
;
1953 next_skb_size
= next_skb
->len
;
1955 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
1957 tcp_highest_sack_combine(sk
, next_skb
, skb
);
1959 tcp_unlink_write_queue(next_skb
, sk
);
1961 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
1964 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
1965 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1967 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1968 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
1970 /* Update sequence range on original skb. */
1971 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
1973 /* Merge over control information. This moves PSH/FIN etc. over */
1974 TCP_SKB_CB(skb
)->flags
|= TCP_SKB_CB(next_skb
)->flags
;
1976 /* All done, get rid of second SKB and account for it so
1977 * packet counting does not break.
1979 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
1981 /* changed transmit queue under us so clear hints */
1982 tcp_clear_retrans_hints_partial(tp
);
1983 if (next_skb
== tp
->retransmit_skb_hint
)
1984 tp
->retransmit_skb_hint
= skb
;
1986 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
1988 sk_wmem_free_skb(sk
, next_skb
);
1991 /* Check if coalescing SKBs is legal. */
1992 static int tcp_can_collapse(struct sock
*sk
, struct sk_buff
*skb
)
1994 if (tcp_skb_pcount(skb
) > 1)
1996 /* TODO: SACK collapsing could be used to remove this condition */
1997 if (skb_shinfo(skb
)->nr_frags
!= 0)
1999 if (skb_cloned(skb
))
2001 if (skb
== tcp_send_head(sk
))
2003 /* Some heurestics for collapsing over SACK'd could be invented */
2004 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2010 /* Collapse packets in the retransmit queue to make to create
2011 * less packets on the wire. This is only done on retransmission.
2013 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2016 struct tcp_sock
*tp
= tcp_sk(sk
);
2017 struct sk_buff
*skb
= to
, *tmp
;
2020 if (!sysctl_tcp_retrans_collapse
)
2022 if (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
)
2025 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2026 if (!tcp_can_collapse(sk
, skb
))
2038 /* Punt if not enough space exists in the first SKB for
2039 * the data in the second
2041 if (skb
->len
> skb_tailroom(to
))
2044 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2047 tcp_collapse_retrans(sk
, to
);
2051 /* This retransmits one SKB. Policy decisions and retransmit queue
2052 * state updates are done by the caller. Returns non-zero if an
2053 * error occurred which prevented the send.
2055 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2057 struct tcp_sock
*tp
= tcp_sk(sk
);
2058 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2059 unsigned int cur_mss
;
2062 /* Inconslusive MTU probe */
2063 if (icsk
->icsk_mtup
.probe_size
) {
2064 icsk
->icsk_mtup
.probe_size
= 0;
2067 /* Do not sent more than we queued. 1/4 is reserved for possible
2068 * copying overhead: fragmentation, tunneling, mangling etc.
2070 if (atomic_read(&sk
->sk_wmem_alloc
) >
2071 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2074 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2075 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2077 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2081 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2082 return -EHOSTUNREACH
; /* Routing failure or similar. */
2084 cur_mss
= tcp_current_mss(sk
);
2086 /* If receiver has shrunk his window, and skb is out of
2087 * new window, do not retransmit it. The exception is the
2088 * case, when window is shrunk to zero. In this case
2089 * our retransmit serves as a zero window probe.
2091 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2092 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2095 if (skb
->len
> cur_mss
) {
2096 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
2097 return -ENOMEM
; /* We'll try again later. */
2099 int oldpcount
= tcp_skb_pcount(skb
);
2101 if (unlikely(oldpcount
> 1)) {
2102 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2103 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2107 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2109 /* Some Solaris stacks overoptimize and ignore the FIN on a
2110 * retransmit when old data is attached. So strip it off
2111 * since it is cheap to do so and saves bytes on the network.
2114 (TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_FIN
) &&
2115 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
2116 if (!pskb_trim(skb
, 0)) {
2117 /* Reuse, even though it does some unnecessary work */
2118 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
2119 TCP_SKB_CB(skb
)->flags
);
2120 skb
->ip_summed
= CHECKSUM_NONE
;
2124 /* Make a copy, if the first transmission SKB clone we made
2125 * is still in somebody's hands, else make a clone.
2127 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2129 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2132 /* Update global TCP statistics. */
2133 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2135 tp
->total_retrans
++;
2137 #if FASTRETRANS_DEBUG > 0
2138 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2139 if (net_ratelimit())
2140 printk(KERN_DEBUG
"retrans_out leaked.\n");
2143 if (!tp
->retrans_out
)
2144 tp
->lost_retrans_low
= tp
->snd_nxt
;
2145 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2146 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2148 /* Save stamp of the first retransmit. */
2149 if (!tp
->retrans_stamp
)
2150 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2154 /* snd_nxt is stored to detect loss of retransmitted segment,
2155 * see tcp_input.c tcp_sacktag_write_queue().
2157 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2162 /* Check if we forward retransmits are possible in the current
2163 * window/congestion state.
2165 static int tcp_can_forward_retransmit(struct sock
*sk
)
2167 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2168 struct tcp_sock
*tp
= tcp_sk(sk
);
2170 /* Forward retransmissions are possible only during Recovery. */
2171 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2174 /* No forward retransmissions in Reno are possible. */
2175 if (tcp_is_reno(tp
))
2178 /* Yeah, we have to make difficult choice between forward transmission
2179 * and retransmission... Both ways have their merits...
2181 * For now we do not retransmit anything, while we have some new
2182 * segments to send. In the other cases, follow rule 3 for
2183 * NextSeg() specified in RFC3517.
2186 if (tcp_may_send_now(sk
))
2192 /* This gets called after a retransmit timeout, and the initially
2193 * retransmitted data is acknowledged. It tries to continue
2194 * resending the rest of the retransmit queue, until either
2195 * we've sent it all or the congestion window limit is reached.
2196 * If doing SACK, the first ACK which comes back for a timeout
2197 * based retransmit packet might feed us FACK information again.
2198 * If so, we use it to avoid unnecessarily retransmissions.
2200 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2202 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2203 struct tcp_sock
*tp
= tcp_sk(sk
);
2204 struct sk_buff
*skb
;
2205 struct sk_buff
*hole
= NULL
;
2208 int fwd_rexmitting
= 0;
2211 tp
->retransmit_high
= tp
->snd_una
;
2213 if (tp
->retransmit_skb_hint
) {
2214 skb
= tp
->retransmit_skb_hint
;
2215 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2216 if (after(last_lost
, tp
->retransmit_high
))
2217 last_lost
= tp
->retransmit_high
;
2219 skb
= tcp_write_queue_head(sk
);
2220 last_lost
= tp
->snd_una
;
2223 tcp_for_write_queue_from(skb
, sk
) {
2224 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2226 if (skb
== tcp_send_head(sk
))
2228 /* we could do better than to assign each time */
2230 tp
->retransmit_skb_hint
= skb
;
2232 /* Assume this retransmit will generate
2233 * only one packet for congestion window
2234 * calculation purposes. This works because
2235 * tcp_retransmit_skb() will chop up the
2236 * packet to be MSS sized and all the
2237 * packet counting works out.
2239 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2242 if (fwd_rexmitting
) {
2244 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2246 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2248 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2249 tp
->retransmit_high
= last_lost
;
2250 if (!tcp_can_forward_retransmit(sk
))
2252 /* Backtrack if necessary to non-L'ed skb */
2260 } else if (!(sacked
& TCPCB_LOST
)) {
2261 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2266 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2267 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2268 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2270 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2273 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2276 if (tcp_retransmit_skb(sk
, skb
))
2278 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2280 if (skb
== tcp_write_queue_head(sk
))
2281 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2282 inet_csk(sk
)->icsk_rto
,
2287 /* Send a fin. The caller locks the socket for us. This cannot be
2288 * allowed to fail queueing a FIN frame under any circumstances.
2290 void tcp_send_fin(struct sock
*sk
)
2292 struct tcp_sock
*tp
= tcp_sk(sk
);
2293 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2296 /* Optimization, tack on the FIN if we have a queue of
2297 * unsent frames. But be careful about outgoing SACKS
2300 mss_now
= tcp_current_mss(sk
);
2302 if (tcp_send_head(sk
) != NULL
) {
2303 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_FIN
;
2304 TCP_SKB_CB(skb
)->end_seq
++;
2307 /* Socket is locked, keep trying until memory is available. */
2309 skb
= alloc_skb_fclone(MAX_TCP_HEADER
,
2316 /* Reserve space for headers and prepare control bits. */
2317 skb_reserve(skb
, MAX_TCP_HEADER
);
2318 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2319 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2320 TCPCB_FLAG_ACK
| TCPCB_FLAG_FIN
);
2321 tcp_queue_skb(sk
, skb
);
2323 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2326 /* We get here when a process closes a file descriptor (either due to
2327 * an explicit close() or as a byproduct of exit()'ing) and there
2328 * was unread data in the receive queue. This behavior is recommended
2329 * by RFC 2525, section 2.17. -DaveM
2331 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2333 struct sk_buff
*skb
;
2335 /* NOTE: No TCP options attached and we never retransmit this. */
2336 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2338 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2342 /* Reserve space for headers and prepare control bits. */
2343 skb_reserve(skb
, MAX_TCP_HEADER
);
2344 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2345 TCPCB_FLAG_ACK
| TCPCB_FLAG_RST
);
2347 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2348 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2349 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2351 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2354 /* Send a crossed SYN-ACK during socket establishment.
2355 * WARNING: This routine must only be called when we have already sent
2356 * a SYN packet that crossed the incoming SYN that caused this routine
2357 * to get called. If this assumption fails then the initial rcv_wnd
2358 * and rcv_wscale values will not be correct.
2360 int tcp_send_synack(struct sock
*sk
)
2362 struct sk_buff
*skb
;
2364 skb
= tcp_write_queue_head(sk
);
2365 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_SYN
)) {
2366 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2369 if (!(TCP_SKB_CB(skb
)->flags
& TCPCB_FLAG_ACK
)) {
2370 if (skb_cloned(skb
)) {
2371 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2374 tcp_unlink_write_queue(skb
, sk
);
2375 skb_header_release(nskb
);
2376 __tcp_add_write_queue_head(sk
, nskb
);
2377 sk_wmem_free_skb(sk
, skb
);
2378 sk
->sk_wmem_queued
+= nskb
->truesize
;
2379 sk_mem_charge(sk
, nskb
->truesize
);
2383 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_ACK
;
2384 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2386 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2387 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2390 /* Prepare a SYN-ACK. */
2391 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2392 struct request_sock
*req
,
2393 struct request_values
*rvp
)
2395 struct tcp_out_options opts
;
2396 struct tcp_extend_values
*xvp
= tcp_xv(rvp
);
2397 struct inet_request_sock
*ireq
= inet_rsk(req
);
2398 struct tcp_sock
*tp
= tcp_sk(sk
);
2399 const struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
2401 struct sk_buff
*skb
;
2402 struct tcp_md5sig_key
*md5
;
2403 int tcp_header_size
;
2405 int s_data_desired
= 0;
2407 if (cvp
!= NULL
&& cvp
->s_data_constant
&& cvp
->s_data_desired
)
2408 s_data_desired
= cvp
->s_data_desired
;
2409 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15 + s_data_desired
, 1, GFP_ATOMIC
);
2413 /* Reserve space for headers. */
2414 skb_reserve(skb
, MAX_TCP_HEADER
);
2416 skb_dst_set(skb
, dst_clone(dst
));
2418 mss
= dst_metric(dst
, RTAX_ADVMSS
);
2419 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2420 mss
= tp
->rx_opt
.user_mss
;
2422 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2424 /* Set this up on the first call only */
2425 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2426 /* tcp_full_space because it is guaranteed to be the first packet */
2427 tcp_select_initial_window(tcp_full_space(sk
),
2428 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2433 dst_metric(dst
, RTAX_INITRWND
));
2434 ireq
->rcv_wscale
= rcv_wscale
;
2437 memset(&opts
, 0, sizeof(opts
));
2438 #ifdef CONFIG_SYN_COOKIES
2439 if (unlikely(req
->cookie_ts
))
2440 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2443 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2444 tcp_header_size
= tcp_synack_options(sk
, req
, mss
,
2445 skb
, &opts
, &md5
, xvp
)
2448 skb_push(skb
, tcp_header_size
);
2449 skb_reset_transport_header(skb
);
2452 memset(th
, 0, sizeof(struct tcphdr
));
2455 TCP_ECN_make_synack(req
, th
);
2456 th
->source
= ireq
->loc_port
;
2457 th
->dest
= ireq
->rmt_port
;
2458 /* Setting of flags are superfluous here for callers (and ECE is
2459 * not even correctly set)
2461 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2462 TCPCB_FLAG_SYN
| TCPCB_FLAG_ACK
);
2464 if (OPTION_COOKIE_EXTENSION
& opts
.options
) {
2465 if (s_data_desired
) {
2466 u8
*buf
= skb_put(skb
, s_data_desired
);
2468 /* copy data directly from the listening socket. */
2469 memcpy(buf
, cvp
->s_data_payload
, s_data_desired
);
2470 TCP_SKB_CB(skb
)->end_seq
+= s_data_desired
;
2473 if (opts
.hash_size
> 0) {
2474 __u32 workspace
[SHA_WORKSPACE_WORDS
];
2475 u32
*mess
= &xvp
->cookie_bakery
[COOKIE_DIGEST_WORDS
];
2476 u32
*tail
= &mess
[COOKIE_MESSAGE_WORDS
-1];
2478 /* Secret recipe depends on the Timestamp, (future)
2479 * Sequence and Acknowledgment Numbers, Initiator
2480 * Cookie, and others handled by IP variant caller.
2482 *tail
-- ^= opts
.tsval
;
2483 *tail
-- ^= tcp_rsk(req
)->rcv_isn
+ 1;
2484 *tail
-- ^= TCP_SKB_CB(skb
)->seq
+ 1;
2487 *tail
-- ^= ((th
->dest
<< 16) | th
->source
);
2488 *tail
-- ^= (u32
)(unsigned long)cvp
; /* per sockopt */
2490 sha_transform((__u32
*)&xvp
->cookie_bakery
[0],
2493 opts
.hash_location
=
2494 (__u8
*)&xvp
->cookie_bakery
[0];
2498 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2499 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2501 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2502 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2503 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2504 th
->doff
= (tcp_header_size
>> 2);
2505 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
2507 #ifdef CONFIG_TCP_MD5SIG
2508 /* Okay, we have all we need - do the md5 hash if needed */
2510 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2511 md5
, NULL
, req
, skb
);
2518 /* Do all connect socket setups that can be done AF independent. */
2519 static void tcp_connect_init(struct sock
*sk
)
2521 struct dst_entry
*dst
= __sk_dst_get(sk
);
2522 struct tcp_sock
*tp
= tcp_sk(sk
);
2525 /* We'll fix this up when we get a response from the other end.
2526 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2528 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2529 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2531 #ifdef CONFIG_TCP_MD5SIG
2532 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2533 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2536 /* If user gave his TCP_MAXSEG, record it to clamp */
2537 if (tp
->rx_opt
.user_mss
)
2538 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2541 tcp_sync_mss(sk
, dst_mtu(dst
));
2543 if (!tp
->window_clamp
)
2544 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2545 tp
->advmss
= dst_metric(dst
, RTAX_ADVMSS
);
2546 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2547 tp
->advmss
= tp
->rx_opt
.user_mss
;
2549 tcp_initialize_rcv_mss(sk
);
2551 tcp_select_initial_window(tcp_full_space(sk
),
2552 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2555 sysctl_tcp_window_scaling
,
2557 dst_metric(dst
, RTAX_INITRWND
));
2559 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2560 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2563 sock_reset_flag(sk
, SOCK_DONE
);
2566 tp
->snd_una
= tp
->write_seq
;
2567 tp
->snd_sml
= tp
->write_seq
;
2568 tp
->snd_up
= tp
->write_seq
;
2573 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2574 inet_csk(sk
)->icsk_retransmits
= 0;
2575 tcp_clear_retrans(tp
);
2578 /* Build a SYN and send it off. */
2579 int tcp_connect(struct sock
*sk
)
2581 struct tcp_sock
*tp
= tcp_sk(sk
);
2582 struct sk_buff
*buff
;
2584 tcp_connect_init(sk
);
2586 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2587 if (unlikely(buff
== NULL
))
2590 /* Reserve space for headers. */
2591 skb_reserve(buff
, MAX_TCP_HEADER
);
2593 tp
->snd_nxt
= tp
->write_seq
;
2594 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPCB_FLAG_SYN
);
2595 TCP_ECN_send_syn(sk
, buff
);
2598 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2599 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2600 skb_header_release(buff
);
2601 __tcp_add_write_queue_tail(sk
, buff
);
2602 sk
->sk_wmem_queued
+= buff
->truesize
;
2603 sk_mem_charge(sk
, buff
->truesize
);
2604 tp
->packets_out
+= tcp_skb_pcount(buff
);
2605 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
2607 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2608 * in order to make this packet get counted in tcpOutSegs.
2610 tp
->snd_nxt
= tp
->write_seq
;
2611 tp
->pushed_seq
= tp
->write_seq
;
2612 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2614 /* Timer for repeating the SYN until an answer. */
2615 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2616 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2620 /* Send out a delayed ack, the caller does the policy checking
2621 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2624 void tcp_send_delayed_ack(struct sock
*sk
)
2626 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2627 int ato
= icsk
->icsk_ack
.ato
;
2628 unsigned long timeout
;
2630 if (ato
> TCP_DELACK_MIN
) {
2631 const struct tcp_sock
*tp
= tcp_sk(sk
);
2632 int max_ato
= HZ
/ 2;
2634 if (icsk
->icsk_ack
.pingpong
||
2635 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2636 max_ato
= TCP_DELACK_MAX
;
2638 /* Slow path, intersegment interval is "high". */
2640 /* If some rtt estimate is known, use it to bound delayed ack.
2641 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2645 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
2651 ato
= min(ato
, max_ato
);
2654 /* Stay within the limit we were given */
2655 timeout
= jiffies
+ ato
;
2657 /* Use new timeout only if there wasn't a older one earlier. */
2658 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2659 /* If delack timer was blocked or is about to expire,
2662 if (icsk
->icsk_ack
.blocked
||
2663 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2668 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2669 timeout
= icsk
->icsk_ack
.timeout
;
2671 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2672 icsk
->icsk_ack
.timeout
= timeout
;
2673 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2676 /* This routine sends an ack and also updates the window. */
2677 void tcp_send_ack(struct sock
*sk
)
2679 struct sk_buff
*buff
;
2681 /* If we have been reset, we may not send again. */
2682 if (sk
->sk_state
== TCP_CLOSE
)
2685 /* We are not putting this on the write queue, so
2686 * tcp_transmit_skb() will set the ownership to this
2689 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2691 inet_csk_schedule_ack(sk
);
2692 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2693 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2694 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2698 /* Reserve space for headers and prepare control bits. */
2699 skb_reserve(buff
, MAX_TCP_HEADER
);
2700 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPCB_FLAG_ACK
);
2702 /* Send it off, this clears delayed acks for us. */
2703 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2704 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2707 /* This routine sends a packet with an out of date sequence
2708 * number. It assumes the other end will try to ack it.
2710 * Question: what should we make while urgent mode?
2711 * 4.4BSD forces sending single byte of data. We cannot send
2712 * out of window data, because we have SND.NXT==SND.MAX...
2714 * Current solution: to send TWO zero-length segments in urgent mode:
2715 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2716 * out-of-date with SND.UNA-1 to probe window.
2718 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2720 struct tcp_sock
*tp
= tcp_sk(sk
);
2721 struct sk_buff
*skb
;
2723 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2724 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2728 /* Reserve space for headers and set control bits. */
2729 skb_reserve(skb
, MAX_TCP_HEADER
);
2730 /* Use a previous sequence. This should cause the other
2731 * end to send an ack. Don't queue or clone SKB, just
2734 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPCB_FLAG_ACK
);
2735 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2736 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2739 /* Initiate keepalive or window probe from timer. */
2740 int tcp_write_wakeup(struct sock
*sk
)
2742 struct tcp_sock
*tp
= tcp_sk(sk
);
2743 struct sk_buff
*skb
;
2745 if (sk
->sk_state
== TCP_CLOSE
)
2748 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2749 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
2751 unsigned int mss
= tcp_current_mss(sk
);
2752 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2754 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2755 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2757 /* We are probing the opening of a window
2758 * but the window size is != 0
2759 * must have been a result SWS avoidance ( sender )
2761 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2763 seg_size
= min(seg_size
, mss
);
2764 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2765 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2767 } else if (!tcp_skb_pcount(skb
))
2768 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2770 TCP_SKB_CB(skb
)->flags
|= TCPCB_FLAG_PSH
;
2771 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2772 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2774 tcp_event_new_data_sent(sk
, skb
);
2777 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
2778 tcp_xmit_probe_skb(sk
, 1);
2779 return tcp_xmit_probe_skb(sk
, 0);
2783 /* A window probe timeout has occurred. If window is not closed send
2784 * a partial packet else a zero probe.
2786 void tcp_send_probe0(struct sock
*sk
)
2788 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2789 struct tcp_sock
*tp
= tcp_sk(sk
);
2792 err
= tcp_write_wakeup(sk
);
2794 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2795 /* Cancel probe timer, if it is not required. */
2796 icsk
->icsk_probes_out
= 0;
2797 icsk
->icsk_backoff
= 0;
2802 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2803 icsk
->icsk_backoff
++;
2804 icsk
->icsk_probes_out
++;
2805 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2806 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2809 /* If packet was not sent due to local congestion,
2810 * do not backoff and do not remember icsk_probes_out.
2811 * Let local senders to fight for local resources.
2813 * Use accumulated backoff yet.
2815 if (!icsk
->icsk_probes_out
)
2816 icsk
->icsk_probes_out
= 1;
2817 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2818 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2819 TCP_RESOURCE_PROBE_INTERVAL
),
2824 EXPORT_SYMBOL(tcp_select_initial_window
);
2825 EXPORT_SYMBOL(tcp_connect
);
2826 EXPORT_SYMBOL(tcp_make_synack
);
2827 EXPORT_SYMBOL(tcp_simple_retransmit
);
2828 EXPORT_SYMBOL(tcp_sync_mss
);
2829 EXPORT_SYMBOL(tcp_mtup_init
);