2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
39 #include <linux/compiler.h>
40 #include <linux/gfp.h>
41 #include <linux/module.h>
43 /* People can turn this off for buggy TCP's found in printers etc. */
44 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
46 /* People can turn this on to work with those rare, broken TCPs that
47 * interpret the window field as a signed quantity.
49 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
51 /* This limits the percentage of the congestion window which we
52 * will allow a single TSO frame to consume. Building TSO frames
53 * which are too large can cause TCP streams to be bursty.
55 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
57 int sysctl_tcp_mtu_probing __read_mostly
= 0;
58 int sysctl_tcp_base_mss __read_mostly
= TCP_BASE_MSS
;
60 /* By default, RFC2861 behavior. */
61 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
63 int sysctl_tcp_cookie_size __read_mostly
= 0; /* TCP_COOKIE_MAX */
64 EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size
);
67 /* Account for new data that has been sent to the network. */
68 static void tcp_event_new_data_sent(struct sock
*sk
, const 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 indefinitely 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(const struct sock
*sk
)
94 const 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 const struct dst_entry
*dst
= __sk_dst_get(sk
);
120 int mss
= tp
->advmss
;
123 unsigned int metric
= dst_metric_advmss(dst
);
134 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
135 * This is the first part of cwnd validation mechanism. */
136 static void tcp_cwnd_restart(struct sock
*sk
, const struct dst_entry
*dst
)
138 struct tcp_sock
*tp
= tcp_sk(sk
);
139 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
140 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
141 u32 cwnd
= tp
->snd_cwnd
;
143 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
145 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
146 restart_cwnd
= min(restart_cwnd
, cwnd
);
148 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
150 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
151 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
152 tp
->snd_cwnd_used
= 0;
155 /* Congestion state accounting after a packet has been sent. */
156 static void tcp_event_data_sent(struct tcp_sock
*tp
,
159 struct inet_connection_sock
*icsk
= inet_csk(sk
);
160 const u32 now
= tcp_time_stamp
;
162 if (sysctl_tcp_slow_start_after_idle
&&
163 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
164 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
168 /* If it is a reply for ato after last received
169 * packet, enter pingpong mode.
171 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
172 icsk
->icsk_ack
.pingpong
= 1;
175 /* Account for an ACK we sent. */
176 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
178 tcp_dec_quickack_mode(sk
, pkts
);
179 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
182 /* Determine a window scaling and initial window to offer.
183 * Based on the assumption that the given amount of space
184 * will be offered. Store the results in the tp structure.
185 * NOTE: for smooth operation initial space offering should
186 * be a multiple of mss if possible. We assume here that mss >= 1.
187 * This MUST be enforced by all callers.
189 void tcp_select_initial_window(int __space
, __u32 mss
,
190 __u32
*rcv_wnd
, __u32
*window_clamp
,
191 int wscale_ok
, __u8
*rcv_wscale
,
194 unsigned int space
= (__space
< 0 ? 0 : __space
);
196 /* If no clamp set the clamp to the max possible scaled window */
197 if (*window_clamp
== 0)
198 (*window_clamp
) = (65535 << 14);
199 space
= min(*window_clamp
, space
);
201 /* Quantize space offering to a multiple of mss if possible. */
203 space
= (space
/ mss
) * mss
;
205 /* NOTE: offering an initial window larger than 32767
206 * will break some buggy TCP stacks. If the admin tells us
207 * it is likely we could be speaking with such a buggy stack
208 * we will truncate our initial window offering to 32K-1
209 * unless the remote has sent us a window scaling option,
210 * which we interpret as a sign the remote TCP is not
211 * misinterpreting the window field as a signed quantity.
213 if (sysctl_tcp_workaround_signed_windows
)
214 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
220 /* Set window scaling on max possible window
221 * See RFC1323 for an explanation of the limit to 14
223 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
224 space
= min_t(u32
, space
, *window_clamp
);
225 while (space
> 65535 && (*rcv_wscale
) < 14) {
231 /* Set initial window to a value enough for senders starting with
232 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
233 * a limit on the initial window when mss is larger than 1460.
235 if (mss
> (1 << *rcv_wscale
)) {
236 int init_cwnd
= TCP_DEFAULT_INIT_RCVWND
;
239 max_t(u32
, (1460 * TCP_DEFAULT_INIT_RCVWND
) / mss
, 2);
240 /* when initializing use the value from init_rcv_wnd
241 * rather than the default from above
244 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
246 *rcv_wnd
= min(*rcv_wnd
, init_cwnd
* mss
);
249 /* Set the clamp no higher than max representable value */
250 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
252 EXPORT_SYMBOL(tcp_select_initial_window
);
254 /* Chose a new window to advertise, update state in tcp_sock for the
255 * socket, and return result with RFC1323 scaling applied. The return
256 * value can be stuffed directly into th->window for an outgoing
259 static u16
tcp_select_window(struct sock
*sk
)
261 struct tcp_sock
*tp
= tcp_sk(sk
);
262 u32 cur_win
= tcp_receive_window(tp
);
263 u32 new_win
= __tcp_select_window(sk
);
265 /* Never shrink the offered window */
266 if (new_win
< cur_win
) {
267 /* Danger Will Robinson!
268 * Don't update rcv_wup/rcv_wnd here or else
269 * we will not be able to advertise a zero
270 * window in time. --DaveM
272 * Relax Will Robinson.
274 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
276 tp
->rcv_wnd
= new_win
;
277 tp
->rcv_wup
= tp
->rcv_nxt
;
279 /* Make sure we do not exceed the maximum possible
282 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
283 new_win
= min(new_win
, MAX_TCP_WINDOW
);
285 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
287 /* RFC1323 scaling applied */
288 new_win
>>= tp
->rx_opt
.rcv_wscale
;
290 /* If we advertise zero window, disable fast path. */
297 /* Packet ECN state for a SYN-ACK */
298 static inline void TCP_ECN_send_synack(const struct tcp_sock
*tp
, struct sk_buff
*skb
)
300 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
301 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
302 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
305 /* Packet ECN state for a SYN. */
306 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
308 struct tcp_sock
*tp
= tcp_sk(sk
);
311 if (sysctl_tcp_ecn
== 1) {
312 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
313 tp
->ecn_flags
= TCP_ECN_OK
;
317 static __inline__
void
318 TCP_ECN_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
320 if (inet_rsk(req
)->ecn_ok
)
324 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
327 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
330 struct tcp_sock
*tp
= tcp_sk(sk
);
332 if (tp
->ecn_flags
& TCP_ECN_OK
) {
333 /* Not-retransmitted data segment: set ECT and inject CWR. */
334 if (skb
->len
!= tcp_header_len
&&
335 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
337 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
338 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
339 tcp_hdr(skb
)->cwr
= 1;
340 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
343 /* ACK or retransmitted segment: clear ECT|CE */
344 INET_ECN_dontxmit(sk
);
346 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
347 tcp_hdr(skb
)->ece
= 1;
351 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
352 * auto increment end seqno.
354 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
356 skb
->ip_summed
= CHECKSUM_PARTIAL
;
359 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
360 TCP_SKB_CB(skb
)->sacked
= 0;
362 skb_shinfo(skb
)->gso_segs
= 1;
363 skb_shinfo(skb
)->gso_size
= 0;
364 skb_shinfo(skb
)->gso_type
= 0;
366 TCP_SKB_CB(skb
)->seq
= seq
;
367 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
369 TCP_SKB_CB(skb
)->end_seq
= seq
;
372 static inline int tcp_urg_mode(const struct tcp_sock
*tp
)
374 return tp
->snd_una
!= tp
->snd_up
;
377 #define OPTION_SACK_ADVERTISE (1 << 0)
378 #define OPTION_TS (1 << 1)
379 #define OPTION_MD5 (1 << 2)
380 #define OPTION_WSCALE (1 << 3)
381 #define OPTION_COOKIE_EXTENSION (1 << 4)
383 struct tcp_out_options
{
384 u8 options
; /* bit field of OPTION_* */
385 u8 ws
; /* window scale, 0 to disable */
386 u8 num_sack_blocks
; /* number of SACK blocks to include */
387 u8 hash_size
; /* bytes in hash_location */
388 u16 mss
; /* 0 to disable */
389 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
390 __u8
*hash_location
; /* temporary pointer, overloaded */
393 /* The sysctl int routines are generic, so check consistency here.
395 static u8
tcp_cookie_size_check(u8 desired
)
400 /* previously specified */
403 cookie_size
= ACCESS_ONCE(sysctl_tcp_cookie_size
);
404 if (cookie_size
<= 0)
405 /* no default specified */
408 if (cookie_size
<= TCP_COOKIE_MIN
)
409 /* value too small, specify minimum */
410 return TCP_COOKIE_MIN
;
412 if (cookie_size
>= TCP_COOKIE_MAX
)
413 /* value too large, specify maximum */
414 return TCP_COOKIE_MAX
;
417 /* 8-bit multiple, illegal, fix it */
420 return (u8
)cookie_size
;
423 /* Write previously computed TCP options to the packet.
425 * Beware: Something in the Internet is very sensitive to the ordering of
426 * TCP options, we learned this through the hard way, so be careful here.
427 * Luckily we can at least blame others for their non-compliance but from
428 * inter-operatibility perspective it seems that we're somewhat stuck with
429 * the ordering which we have been using if we want to keep working with
430 * those broken things (not that it currently hurts anybody as there isn't
431 * particular reason why the ordering would need to be changed).
433 * At least SACK_PERM as the first option is known to lead to a disaster
434 * (but it may well be that other scenarios fail similarly).
436 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
437 struct tcp_out_options
*opts
)
439 u8 options
= opts
->options
; /* mungable copy */
441 /* Having both authentication and cookies for security is redundant,
442 * and there's certainly not enough room. Instead, the cookie-less
443 * extension variant is proposed.
445 * Consider the pessimal case with authentication. The options
447 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
449 if (unlikely(OPTION_MD5
& options
)) {
450 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
451 *ptr
++ = htonl((TCPOPT_COOKIE
<< 24) |
452 (TCPOLEN_COOKIE_BASE
<< 16) |
453 (TCPOPT_MD5SIG
<< 8) |
456 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
458 (TCPOPT_MD5SIG
<< 8) |
461 options
&= ~OPTION_COOKIE_EXTENSION
;
462 /* overload cookie hash location */
463 opts
->hash_location
= (__u8
*)ptr
;
467 if (unlikely(opts
->mss
)) {
468 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
469 (TCPOLEN_MSS
<< 16) |
473 if (likely(OPTION_TS
& options
)) {
474 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
475 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
476 (TCPOLEN_SACK_PERM
<< 16) |
477 (TCPOPT_TIMESTAMP
<< 8) |
479 options
&= ~OPTION_SACK_ADVERTISE
;
481 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
483 (TCPOPT_TIMESTAMP
<< 8) |
486 *ptr
++ = htonl(opts
->tsval
);
487 *ptr
++ = htonl(opts
->tsecr
);
490 /* Specification requires after timestamp, so do it now.
492 * Consider the pessimal case without authentication. The options
494 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
496 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
497 __u8
*cookie_copy
= opts
->hash_location
;
498 u8 cookie_size
= opts
->hash_size
;
500 /* 8-bit multiple handled in tcp_cookie_size_check() above,
503 if (0x2 & cookie_size
) {
504 __u8
*p
= (__u8
*)ptr
;
506 /* 16-bit multiple */
507 *p
++ = TCPOPT_COOKIE
;
508 *p
++ = TCPOLEN_COOKIE_BASE
+ cookie_size
;
509 *p
++ = *cookie_copy
++;
510 *p
++ = *cookie_copy
++;
514 /* 32-bit multiple */
515 *ptr
++ = htonl(((TCPOPT_NOP
<< 24) |
517 (TCPOPT_COOKIE
<< 8) |
518 TCPOLEN_COOKIE_BASE
) +
522 if (cookie_size
> 0) {
523 memcpy(ptr
, cookie_copy
, cookie_size
);
524 ptr
+= (cookie_size
/ 4);
528 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
529 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
531 (TCPOPT_SACK_PERM
<< 8) |
535 if (unlikely(OPTION_WSCALE
& options
)) {
536 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
537 (TCPOPT_WINDOW
<< 16) |
538 (TCPOLEN_WINDOW
<< 8) |
542 if (unlikely(opts
->num_sack_blocks
)) {
543 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
544 tp
->duplicate_sack
: tp
->selective_acks
;
547 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
550 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
551 TCPOLEN_SACK_PERBLOCK
)));
553 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
555 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
556 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
559 tp
->rx_opt
.dsack
= 0;
563 /* Compute TCP options for SYN packets. This is not the final
564 * network wire format yet.
566 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
567 struct tcp_out_options
*opts
,
568 struct tcp_md5sig_key
**md5
)
570 struct tcp_sock
*tp
= tcp_sk(sk
);
571 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
572 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
573 u8 cookie_size
= (!tp
->rx_opt
.cookie_out_never
&& cvp
!= NULL
) ?
574 tcp_cookie_size_check(cvp
->cookie_desired
) :
577 #ifdef CONFIG_TCP_MD5SIG
578 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
580 opts
->options
|= OPTION_MD5
;
581 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
587 /* We always get an MSS option. The option bytes which will be seen in
588 * normal data packets should timestamps be used, must be in the MSS
589 * advertised. But we subtract them from tp->mss_cache so that
590 * calculations in tcp_sendmsg are simpler etc. So account for this
591 * fact here if necessary. If we don't do this correctly, as a
592 * receiver we won't recognize data packets as being full sized when we
593 * should, and thus we won't abide by the delayed ACK rules correctly.
594 * SACKs don't matter, we never delay an ACK when we have any of those
596 opts
->mss
= tcp_advertise_mss(sk
);
597 remaining
-= TCPOLEN_MSS_ALIGNED
;
599 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
600 opts
->options
|= OPTION_TS
;
601 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
602 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
603 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
605 if (likely(sysctl_tcp_window_scaling
)) {
606 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
607 opts
->options
|= OPTION_WSCALE
;
608 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
610 if (likely(sysctl_tcp_sack
)) {
611 opts
->options
|= OPTION_SACK_ADVERTISE
;
612 if (unlikely(!(OPTION_TS
& opts
->options
)))
613 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
616 /* Note that timestamps are required by the specification.
618 * Odd numbers of bytes are prohibited by the specification, ensuring
619 * that the cookie is 16-bit aligned, and the resulting cookie pair is
623 (OPTION_TS
& opts
->options
) &&
625 int need
= TCPOLEN_COOKIE_BASE
+ cookie_size
;
628 /* 32-bit multiple */
629 need
+= 2; /* NOPs */
631 if (need
> remaining
) {
632 /* try shrinking cookie to fit */
637 while (need
> remaining
&& TCP_COOKIE_MIN
<= cookie_size
) {
641 if (TCP_COOKIE_MIN
<= cookie_size
) {
642 opts
->options
|= OPTION_COOKIE_EXTENSION
;
643 opts
->hash_location
= (__u8
*)&cvp
->cookie_pair
[0];
644 opts
->hash_size
= cookie_size
;
646 /* Remember for future incarnations. */
647 cvp
->cookie_desired
= cookie_size
;
649 if (cvp
->cookie_desired
!= cvp
->cookie_pair_size
) {
650 /* Currently use random bytes as a nonce,
651 * assuming these are completely unpredictable
652 * by hostile users of the same system.
654 get_random_bytes(&cvp
->cookie_pair
[0],
656 cvp
->cookie_pair_size
= cookie_size
;
662 return MAX_TCP_OPTION_SPACE
- remaining
;
665 /* Set up TCP options for SYN-ACKs. */
666 static unsigned int tcp_synack_options(struct sock
*sk
,
667 struct request_sock
*req
,
668 unsigned int mss
, struct sk_buff
*skb
,
669 struct tcp_out_options
*opts
,
670 struct tcp_md5sig_key
**md5
,
671 struct tcp_extend_values
*xvp
)
673 struct inet_request_sock
*ireq
= inet_rsk(req
);
674 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
675 u8 cookie_plus
= (xvp
!= NULL
&& !xvp
->cookie_out_never
) ?
679 #ifdef CONFIG_TCP_MD5SIG
680 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
682 opts
->options
|= OPTION_MD5
;
683 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
685 /* We can't fit any SACK blocks in a packet with MD5 + TS
686 * options. There was discussion about disabling SACK
687 * rather than TS in order to fit in better with old,
688 * buggy kernels, but that was deemed to be unnecessary.
690 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
696 /* We always send an MSS option. */
698 remaining
-= TCPOLEN_MSS_ALIGNED
;
700 if (likely(ireq
->wscale_ok
)) {
701 opts
->ws
= ireq
->rcv_wscale
;
702 opts
->options
|= OPTION_WSCALE
;
703 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
705 if (likely(ireq
->tstamp_ok
)) {
706 opts
->options
|= OPTION_TS
;
707 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
708 opts
->tsecr
= req
->ts_recent
;
709 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
711 if (likely(ireq
->sack_ok
)) {
712 opts
->options
|= OPTION_SACK_ADVERTISE
;
713 if (unlikely(!ireq
->tstamp_ok
))
714 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
717 /* Similar rationale to tcp_syn_options() applies here, too.
718 * If the <SYN> options fit, the same options should fit now!
722 cookie_plus
> TCPOLEN_COOKIE_BASE
) {
723 int need
= cookie_plus
; /* has TCPOLEN_COOKIE_BASE */
726 /* 32-bit multiple */
727 need
+= 2; /* NOPs */
729 if (need
<= remaining
) {
730 opts
->options
|= OPTION_COOKIE_EXTENSION
;
731 opts
->hash_size
= cookie_plus
- TCPOLEN_COOKIE_BASE
;
734 /* There's no error return, so flag it. */
735 xvp
->cookie_out_never
= 1; /* true */
739 return MAX_TCP_OPTION_SPACE
- remaining
;
742 /* Compute TCP options for ESTABLISHED sockets. This is not the
743 * final wire format yet.
745 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
746 struct tcp_out_options
*opts
,
747 struct tcp_md5sig_key
**md5
)
749 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
750 struct tcp_sock
*tp
= tcp_sk(sk
);
751 unsigned int size
= 0;
752 unsigned int eff_sacks
;
754 #ifdef CONFIG_TCP_MD5SIG
755 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
756 if (unlikely(*md5
)) {
757 opts
->options
|= OPTION_MD5
;
758 size
+= TCPOLEN_MD5SIG_ALIGNED
;
764 if (likely(tp
->rx_opt
.tstamp_ok
)) {
765 opts
->options
|= OPTION_TS
;
766 opts
->tsval
= tcb
? tcb
->when
: 0;
767 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
768 size
+= TCPOLEN_TSTAMP_ALIGNED
;
771 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
772 if (unlikely(eff_sacks
)) {
773 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
774 opts
->num_sack_blocks
=
775 min_t(unsigned int, eff_sacks
,
776 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
777 TCPOLEN_SACK_PERBLOCK
);
778 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
779 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
785 /* This routine actually transmits TCP packets queued in by
786 * tcp_do_sendmsg(). This is used by both the initial
787 * transmission and possible later retransmissions.
788 * All SKB's seen here are completely headerless. It is our
789 * job to build the TCP header, and pass the packet down to
790 * IP so it can do the same plus pass the packet off to the
793 * We are working here with either a clone of the original
794 * SKB, or a fresh unique copy made by the retransmit engine.
796 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
799 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
800 struct inet_sock
*inet
;
802 struct tcp_skb_cb
*tcb
;
803 struct tcp_out_options opts
;
804 unsigned int tcp_options_size
, tcp_header_size
;
805 struct tcp_md5sig_key
*md5
;
809 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
811 /* If congestion control is doing timestamping, we must
812 * take such a timestamp before we potentially clone/copy.
814 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
815 __net_timestamp(skb
);
817 if (likely(clone_it
)) {
818 if (unlikely(skb_cloned(skb
)))
819 skb
= pskb_copy(skb
, gfp_mask
);
821 skb
= skb_clone(skb
, gfp_mask
);
828 tcb
= TCP_SKB_CB(skb
);
829 memset(&opts
, 0, sizeof(opts
));
831 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
832 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
834 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
836 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
838 if (tcp_packets_in_flight(tp
) == 0) {
839 tcp_ca_event(sk
, CA_EVENT_TX_START
);
844 skb_push(skb
, tcp_header_size
);
845 skb_reset_transport_header(skb
);
846 skb_set_owner_w(skb
, sk
);
848 /* Build TCP header and checksum it. */
850 th
->source
= inet
->inet_sport
;
851 th
->dest
= inet
->inet_dport
;
852 th
->seq
= htonl(tcb
->seq
);
853 th
->ack_seq
= htonl(tp
->rcv_nxt
);
854 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
857 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
858 /* RFC1323: The window in SYN & SYN/ACK segments
861 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
863 th
->window
= htons(tcp_select_window(sk
));
868 /* The urg_mode check is necessary during a below snd_una win probe */
869 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
870 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
871 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
873 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
874 th
->urg_ptr
= htons(0xFFFF);
879 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
880 if (likely((tcb
->tcp_flags
& TCPHDR_SYN
) == 0))
881 TCP_ECN_send(sk
, skb
, tcp_header_size
);
883 #ifdef CONFIG_TCP_MD5SIG
884 /* Calculate the MD5 hash, as we have all we need now */
886 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
887 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
892 icsk
->icsk_af_ops
->send_check(sk
, skb
);
894 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
895 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
897 if (skb
->len
!= tcp_header_size
)
898 tcp_event_data_sent(tp
, sk
);
900 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
901 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
902 tcp_skb_pcount(skb
));
904 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, &inet
->cork
.fl
);
905 if (likely(err
<= 0))
908 tcp_enter_cwr(sk
, 1);
910 return net_xmit_eval(err
);
913 /* This routine just queues the buffer for sending.
915 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
916 * otherwise socket can stall.
918 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
920 struct tcp_sock
*tp
= tcp_sk(sk
);
922 /* Advance write_seq and place onto the write_queue. */
923 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
924 skb_header_release(skb
);
925 tcp_add_write_queue_tail(sk
, skb
);
926 sk
->sk_wmem_queued
+= skb
->truesize
;
927 sk_mem_charge(sk
, skb
->truesize
);
930 /* Initialize TSO segments for a packet. */
931 static void tcp_set_skb_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
932 unsigned int mss_now
)
934 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
) ||
935 skb
->ip_summed
== CHECKSUM_NONE
) {
936 /* Avoid the costly divide in the normal
939 skb_shinfo(skb
)->gso_segs
= 1;
940 skb_shinfo(skb
)->gso_size
= 0;
941 skb_shinfo(skb
)->gso_type
= 0;
943 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
944 skb_shinfo(skb
)->gso_size
= mss_now
;
945 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
949 /* When a modification to fackets out becomes necessary, we need to check
950 * skb is counted to fackets_out or not.
952 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
955 struct tcp_sock
*tp
= tcp_sk(sk
);
957 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
960 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
961 tp
->fackets_out
-= decr
;
964 /* Pcount in the middle of the write queue got changed, we need to do various
965 * tweaks to fix counters
967 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
969 struct tcp_sock
*tp
= tcp_sk(sk
);
971 tp
->packets_out
-= decr
;
973 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
974 tp
->sacked_out
-= decr
;
975 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
976 tp
->retrans_out
-= decr
;
977 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
978 tp
->lost_out
-= decr
;
980 /* Reno case is special. Sigh... */
981 if (tcp_is_reno(tp
) && decr
> 0)
982 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
984 tcp_adjust_fackets_out(sk
, skb
, decr
);
986 if (tp
->lost_skb_hint
&&
987 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
988 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
989 tp
->lost_cnt_hint
-= decr
;
991 tcp_verify_left_out(tp
);
994 /* Function to create two new TCP segments. Shrinks the given segment
995 * to the specified size and appends a new segment with the rest of the
996 * packet to the list. This won't be called frequently, I hope.
997 * Remember, these are still headerless SKBs at this point.
999 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1000 unsigned int mss_now
)
1002 struct tcp_sock
*tp
= tcp_sk(sk
);
1003 struct sk_buff
*buff
;
1004 int nsize
, old_factor
;
1008 if (WARN_ON(len
> skb
->len
))
1011 nsize
= skb_headlen(skb
) - len
;
1015 if (skb_cloned(skb
) &&
1016 skb_is_nonlinear(skb
) &&
1017 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1020 /* Get a new skb... force flag on. */
1021 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
1023 return -ENOMEM
; /* We'll just try again later. */
1025 sk
->sk_wmem_queued
+= buff
->truesize
;
1026 sk_mem_charge(sk
, buff
->truesize
);
1027 nlen
= skb
->len
- len
- nsize
;
1028 buff
->truesize
+= nlen
;
1029 skb
->truesize
-= nlen
;
1031 /* Correct the sequence numbers. */
1032 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1033 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1034 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1036 /* PSH and FIN should only be set in the second packet. */
1037 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1038 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1039 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1040 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1042 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1043 /* Copy and checksum data tail into the new buffer. */
1044 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1045 skb_put(buff
, nsize
),
1050 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1052 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1053 skb_split(skb
, buff
, len
);
1056 buff
->ip_summed
= skb
->ip_summed
;
1058 /* Looks stupid, but our code really uses when of
1059 * skbs, which it never sent before. --ANK
1061 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
1062 buff
->tstamp
= skb
->tstamp
;
1064 old_factor
= tcp_skb_pcount(skb
);
1066 /* Fix up tso_factor for both original and new SKB. */
1067 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1068 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1070 /* If this packet has been sent out already, we must
1071 * adjust the various packet counters.
1073 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1074 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1075 tcp_skb_pcount(buff
);
1078 tcp_adjust_pcount(sk
, skb
, diff
);
1081 /* Link BUFF into the send queue. */
1082 skb_header_release(buff
);
1083 tcp_insert_write_queue_after(skb
, buff
, sk
);
1088 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1089 * eventually). The difference is that pulled data not copied, but
1090 * immediately discarded.
1092 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1096 eat
= min_t(int, len
, skb_headlen(skb
));
1098 __skb_pull(skb
, eat
);
1099 skb
->avail_size
-= eat
;
1106 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
1107 int size
= skb_frag_size(&skb_shinfo(skb
)->frags
[i
]);
1110 skb_frag_unref(skb
, i
);
1113 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
1115 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
1116 skb_frag_size_sub(&skb_shinfo(skb
)->frags
[k
], eat
);
1122 skb_shinfo(skb
)->nr_frags
= k
;
1124 skb_reset_tail_pointer(skb
);
1125 skb
->data_len
-= len
;
1126 skb
->len
= skb
->data_len
;
1129 /* Remove acked data from a packet in the transmit queue. */
1130 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1132 if (skb_cloned(skb
) && pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1135 __pskb_trim_head(skb
, len
);
1137 TCP_SKB_CB(skb
)->seq
+= len
;
1138 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1140 skb
->truesize
-= len
;
1141 sk
->sk_wmem_queued
-= len
;
1142 sk_mem_uncharge(sk
, len
);
1143 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1145 /* Any change of skb->len requires recalculation of tso factor. */
1146 if (tcp_skb_pcount(skb
) > 1)
1147 tcp_set_skb_tso_segs(sk
, skb
, tcp_skb_mss(skb
));
1152 /* Calculate MSS. Not accounting for SACKs here. */
1153 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1155 const struct tcp_sock
*tp
= tcp_sk(sk
);
1156 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1159 /* Calculate base mss without TCP options:
1160 It is MMS_S - sizeof(tcphdr) of rfc1122
1162 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1164 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1165 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1166 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1168 if (dst
&& dst_allfrag(dst
))
1169 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1172 /* Clamp it (mss_clamp does not include tcp options) */
1173 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1174 mss_now
= tp
->rx_opt
.mss_clamp
;
1176 /* Now subtract optional transport overhead */
1177 mss_now
-= icsk
->icsk_ext_hdr_len
;
1179 /* Then reserve room for full set of TCP options and 8 bytes of data */
1183 /* Now subtract TCP options size, not including SACKs */
1184 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
1189 /* Inverse of above */
1190 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1192 const struct tcp_sock
*tp
= tcp_sk(sk
);
1193 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1197 tp
->tcp_header_len
+
1198 icsk
->icsk_ext_hdr_len
+
1199 icsk
->icsk_af_ops
->net_header_len
;
1201 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1202 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1203 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1205 if (dst
&& dst_allfrag(dst
))
1206 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1211 /* MTU probing init per socket */
1212 void tcp_mtup_init(struct sock
*sk
)
1214 struct tcp_sock
*tp
= tcp_sk(sk
);
1215 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1217 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
1218 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1219 icsk
->icsk_af_ops
->net_header_len
;
1220 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
1221 icsk
->icsk_mtup
.probe_size
= 0;
1223 EXPORT_SYMBOL(tcp_mtup_init
);
1225 /* This function synchronize snd mss to current pmtu/exthdr set.
1227 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1228 for TCP options, but includes only bare TCP header.
1230 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1231 It is minimum of user_mss and mss received with SYN.
1232 It also does not include TCP options.
1234 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1236 tp->mss_cache is current effective sending mss, including
1237 all tcp options except for SACKs. It is evaluated,
1238 taking into account current pmtu, but never exceeds
1239 tp->rx_opt.mss_clamp.
1241 NOTE1. rfc1122 clearly states that advertised MSS
1242 DOES NOT include either tcp or ip options.
1244 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1245 are READ ONLY outside this function. --ANK (980731)
1247 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1249 struct tcp_sock
*tp
= tcp_sk(sk
);
1250 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1253 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1254 icsk
->icsk_mtup
.search_high
= pmtu
;
1256 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1257 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1259 /* And store cached results */
1260 icsk
->icsk_pmtu_cookie
= pmtu
;
1261 if (icsk
->icsk_mtup
.enabled
)
1262 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1263 tp
->mss_cache
= mss_now
;
1267 EXPORT_SYMBOL(tcp_sync_mss
);
1269 /* Compute the current effective MSS, taking SACKs and IP options,
1270 * and even PMTU discovery events into account.
1272 unsigned int tcp_current_mss(struct sock
*sk
)
1274 const struct tcp_sock
*tp
= tcp_sk(sk
);
1275 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1277 unsigned int header_len
;
1278 struct tcp_out_options opts
;
1279 struct tcp_md5sig_key
*md5
;
1281 mss_now
= tp
->mss_cache
;
1284 u32 mtu
= dst_mtu(dst
);
1285 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1286 mss_now
= tcp_sync_mss(sk
, mtu
);
1289 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1290 sizeof(struct tcphdr
);
1291 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1292 * some common options. If this is an odd packet (because we have SACK
1293 * blocks etc) then our calculated header_len will be different, and
1294 * we have to adjust mss_now correspondingly */
1295 if (header_len
!= tp
->tcp_header_len
) {
1296 int delta
= (int) header_len
- tp
->tcp_header_len
;
1303 /* Congestion window validation. (RFC2861) */
1304 static void tcp_cwnd_validate(struct sock
*sk
)
1306 struct tcp_sock
*tp
= tcp_sk(sk
);
1308 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1309 /* Network is feed fully. */
1310 tp
->snd_cwnd_used
= 0;
1311 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1313 /* Network starves. */
1314 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1315 tp
->snd_cwnd_used
= tp
->packets_out
;
1317 if (sysctl_tcp_slow_start_after_idle
&&
1318 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1319 tcp_cwnd_application_limited(sk
);
1323 /* Returns the portion of skb which can be sent right away without
1324 * introducing MSS oddities to segment boundaries. In rare cases where
1325 * mss_now != mss_cache, we will request caller to create a small skb
1326 * per input skb which could be mostly avoided here (if desired).
1328 * We explicitly want to create a request for splitting write queue tail
1329 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1330 * thus all the complexity (cwnd_len is always MSS multiple which we
1331 * return whenever allowed by the other factors). Basically we need the
1332 * modulo only when the receiver window alone is the limiting factor or
1333 * when we would be allowed to send the split-due-to-Nagle skb fully.
1335 static unsigned int tcp_mss_split_point(const struct sock
*sk
, const struct sk_buff
*skb
,
1336 unsigned int mss_now
, unsigned int cwnd
)
1338 const struct tcp_sock
*tp
= tcp_sk(sk
);
1339 u32 needed
, window
, cwnd_len
;
1341 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1342 cwnd_len
= mss_now
* cwnd
;
1344 if (likely(cwnd_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1347 needed
= min(skb
->len
, window
);
1349 if (cwnd_len
<= needed
)
1352 return needed
- needed
% mss_now
;
1355 /* Can at least one segment of SKB be sent right now, according to the
1356 * congestion window rules? If so, return how many segments are allowed.
1358 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1359 const struct sk_buff
*skb
)
1361 u32 in_flight
, cwnd
;
1363 /* Don't be strict about the congestion window for the final FIN. */
1364 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1365 tcp_skb_pcount(skb
) == 1)
1368 in_flight
= tcp_packets_in_flight(tp
);
1369 cwnd
= tp
->snd_cwnd
;
1370 if (in_flight
< cwnd
)
1371 return (cwnd
- in_flight
);
1376 /* Initialize TSO state of a skb.
1377 * This must be invoked the first time we consider transmitting
1378 * SKB onto the wire.
1380 static int tcp_init_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
1381 unsigned int mss_now
)
1383 int tso_segs
= tcp_skb_pcount(skb
);
1385 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1386 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1387 tso_segs
= tcp_skb_pcount(skb
);
1392 /* Minshall's variant of the Nagle send check. */
1393 static inline int tcp_minshall_check(const struct tcp_sock
*tp
)
1395 return after(tp
->snd_sml
, tp
->snd_una
) &&
1396 !after(tp
->snd_sml
, tp
->snd_nxt
);
1399 /* Return 0, if packet can be sent now without violation Nagle's rules:
1400 * 1. It is full sized.
1401 * 2. Or it contains FIN. (already checked by caller)
1402 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1403 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1404 * With Minshall's modification: all sent small packets are ACKed.
1406 static inline int tcp_nagle_check(const struct tcp_sock
*tp
,
1407 const struct sk_buff
*skb
,
1408 unsigned int mss_now
, int nonagle
)
1410 return skb
->len
< mss_now
&&
1411 ((nonagle
& TCP_NAGLE_CORK
) ||
1412 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1415 /* Return non-zero if the Nagle test allows this packet to be
1418 static inline int tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1419 unsigned int cur_mss
, int nonagle
)
1421 /* Nagle rule does not apply to frames, which sit in the middle of the
1422 * write_queue (they have no chances to get new data).
1424 * This is implemented in the callers, where they modify the 'nonagle'
1425 * argument based upon the location of SKB in the send queue.
1427 if (nonagle
& TCP_NAGLE_PUSH
)
1430 /* Don't use the nagle rule for urgent data (or for the final FIN).
1431 * Nagle can be ignored during F-RTO too (see RFC4138).
1433 if (tcp_urg_mode(tp
) || (tp
->frto_counter
== 2) ||
1434 (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1437 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1443 /* Does at least the first segment of SKB fit into the send window? */
1444 static inline int tcp_snd_wnd_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1445 unsigned int cur_mss
)
1447 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1449 if (skb
->len
> cur_mss
)
1450 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1452 return !after(end_seq
, tcp_wnd_end(tp
));
1455 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1456 * should be put on the wire right now. If so, it returns the number of
1457 * packets allowed by the congestion window.
1459 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1460 unsigned int cur_mss
, int nonagle
)
1462 const struct tcp_sock
*tp
= tcp_sk(sk
);
1463 unsigned int cwnd_quota
;
1465 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1467 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1470 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1471 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1477 /* Test if sending is allowed right now. */
1478 int tcp_may_send_now(struct sock
*sk
)
1480 const struct tcp_sock
*tp
= tcp_sk(sk
);
1481 struct sk_buff
*skb
= tcp_send_head(sk
);
1484 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1485 (tcp_skb_is_last(sk
, skb
) ?
1486 tp
->nonagle
: TCP_NAGLE_PUSH
));
1489 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1490 * which is put after SKB on the list. It is very much like
1491 * tcp_fragment() except that it may make several kinds of assumptions
1492 * in order to speed up the splitting operation. In particular, we
1493 * know that all the data is in scatter-gather pages, and that the
1494 * packet has never been sent out before (and thus is not cloned).
1496 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1497 unsigned int mss_now
, gfp_t gfp
)
1499 struct sk_buff
*buff
;
1500 int nlen
= skb
->len
- len
;
1503 /* All of a TSO frame must be composed of paged data. */
1504 if (skb
->len
!= skb
->data_len
)
1505 return tcp_fragment(sk
, skb
, len
, mss_now
);
1507 buff
= sk_stream_alloc_skb(sk
, 0, gfp
);
1508 if (unlikely(buff
== NULL
))
1511 sk
->sk_wmem_queued
+= buff
->truesize
;
1512 sk_mem_charge(sk
, buff
->truesize
);
1513 buff
->truesize
+= nlen
;
1514 skb
->truesize
-= nlen
;
1516 /* Correct the sequence numbers. */
1517 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1518 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1519 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1521 /* PSH and FIN should only be set in the second packet. */
1522 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1523 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1524 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1526 /* This packet was never sent out yet, so no SACK bits. */
1527 TCP_SKB_CB(buff
)->sacked
= 0;
1529 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1530 skb_split(skb
, buff
, len
);
1532 /* Fix up tso_factor for both original and new SKB. */
1533 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1534 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1536 /* Link BUFF into the send queue. */
1537 skb_header_release(buff
);
1538 tcp_insert_write_queue_after(skb
, buff
, sk
);
1543 /* Try to defer sending, if possible, in order to minimize the amount
1544 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1546 * This algorithm is from John Heffner.
1548 static int tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1550 struct tcp_sock
*tp
= tcp_sk(sk
);
1551 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1552 u32 send_win
, cong_win
, limit
, in_flight
;
1555 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1558 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1561 /* Defer for less than two clock ticks. */
1562 if (tp
->tso_deferred
&&
1563 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1566 in_flight
= tcp_packets_in_flight(tp
);
1568 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1570 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1572 /* From in_flight test above, we know that cwnd > in_flight. */
1573 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1575 limit
= min(send_win
, cong_win
);
1577 /* If a full-sized TSO skb can be sent, do it. */
1578 if (limit
>= sk
->sk_gso_max_size
)
1581 /* Middle in queue won't get any more data, full sendable already? */
1582 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1585 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1587 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1589 /* If at least some fraction of a window is available,
1592 chunk
/= win_divisor
;
1596 /* Different approach, try not to defer past a single
1597 * ACK. Receiver should ACK every other full sized
1598 * frame, so if we have space for more than 3 frames
1601 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1605 /* Ok, it looks like it is advisable to defer. */
1606 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1611 tp
->tso_deferred
= 0;
1615 /* Create a new MTU probe if we are ready.
1616 * MTU probe is regularly attempting to increase the path MTU by
1617 * deliberately sending larger packets. This discovers routing
1618 * changes resulting in larger path MTUs.
1620 * Returns 0 if we should wait to probe (no cwnd available),
1621 * 1 if a probe was sent,
1624 static int tcp_mtu_probe(struct sock
*sk
)
1626 struct tcp_sock
*tp
= tcp_sk(sk
);
1627 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1628 struct sk_buff
*skb
, *nskb
, *next
;
1635 /* Not currently probing/verifying,
1637 * have enough cwnd, and
1638 * not SACKing (the variable headers throw things off) */
1639 if (!icsk
->icsk_mtup
.enabled
||
1640 icsk
->icsk_mtup
.probe_size
||
1641 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1642 tp
->snd_cwnd
< 11 ||
1643 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1646 /* Very simple search strategy: just double the MSS. */
1647 mss_now
= tcp_current_mss(sk
);
1648 probe_size
= 2 * tp
->mss_cache
;
1649 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1650 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1651 /* TODO: set timer for probe_converge_event */
1655 /* Have enough data in the send queue to probe? */
1656 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1659 if (tp
->snd_wnd
< size_needed
)
1661 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1664 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1665 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1666 if (!tcp_packets_in_flight(tp
))
1672 /* We're allowed to probe. Build it now. */
1673 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1675 sk
->sk_wmem_queued
+= nskb
->truesize
;
1676 sk_mem_charge(sk
, nskb
->truesize
);
1678 skb
= tcp_send_head(sk
);
1680 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1681 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1682 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
1683 TCP_SKB_CB(nskb
)->sacked
= 0;
1685 nskb
->ip_summed
= skb
->ip_summed
;
1687 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1690 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1691 copy
= min_t(int, skb
->len
, probe_size
- len
);
1692 if (nskb
->ip_summed
)
1693 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1695 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1696 skb_put(nskb
, copy
),
1699 if (skb
->len
<= copy
) {
1700 /* We've eaten all the data from this skb.
1702 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1703 tcp_unlink_write_queue(skb
, sk
);
1704 sk_wmem_free_skb(sk
, skb
);
1706 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
1707 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1708 if (!skb_shinfo(skb
)->nr_frags
) {
1709 skb_pull(skb
, copy
);
1710 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1711 skb
->csum
= csum_partial(skb
->data
,
1714 __pskb_trim_head(skb
, copy
);
1715 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1717 TCP_SKB_CB(skb
)->seq
+= copy
;
1722 if (len
>= probe_size
)
1725 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1727 /* We're ready to send. If this fails, the probe will
1728 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1729 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1730 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1731 /* Decrement cwnd here because we are sending
1732 * effectively two packets. */
1734 tcp_event_new_data_sent(sk
, nskb
);
1736 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1737 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1738 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1746 /* This routine writes packets to the network. It advances the
1747 * send_head. This happens as incoming acks open up the remote
1750 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1751 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1752 * account rare use of URG, this is not a big flaw.
1754 * Returns 1, if no segments are in flight and we have queued segments, but
1755 * cannot send anything now because of SWS or another problem.
1757 static int tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1758 int push_one
, gfp_t gfp
)
1760 struct tcp_sock
*tp
= tcp_sk(sk
);
1761 struct sk_buff
*skb
;
1762 unsigned int tso_segs
, sent_pkts
;
1769 /* Do MTU probing. */
1770 result
= tcp_mtu_probe(sk
);
1773 } else if (result
> 0) {
1778 while ((skb
= tcp_send_head(sk
))) {
1781 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1784 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1788 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1791 if (tso_segs
== 1) {
1792 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1793 (tcp_skb_is_last(sk
, skb
) ?
1794 nonagle
: TCP_NAGLE_PUSH
))))
1797 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
1802 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
1803 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1806 if (skb
->len
> limit
&&
1807 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
1810 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1812 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
1815 /* Advance the send_head. This one is sent out.
1816 * This call will increment packets_out.
1818 tcp_event_new_data_sent(sk
, skb
);
1820 tcp_minshall_update(tp
, mss_now
, skb
);
1821 sent_pkts
+= tcp_skb_pcount(skb
);
1826 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Recovery
)
1827 tp
->prr_out
+= sent_pkts
;
1829 if (likely(sent_pkts
)) {
1830 tcp_cwnd_validate(sk
);
1833 return !tp
->packets_out
&& tcp_send_head(sk
);
1836 /* Push out any pending frames which were held back due to
1837 * TCP_CORK or attempt at coalescing tiny packets.
1838 * The socket must be locked by the caller.
1840 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1843 /* If we are closed, the bytes will have to remain here.
1844 * In time closedown will finish, we empty the write queue and
1845 * all will be happy.
1847 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1850 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0, GFP_ATOMIC
))
1851 tcp_check_probe_timer(sk
);
1854 /* Send _single_ skb sitting at the send head. This function requires
1855 * true push pending frames to setup probe timer etc.
1857 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1859 struct sk_buff
*skb
= tcp_send_head(sk
);
1861 BUG_ON(!skb
|| skb
->len
< mss_now
);
1863 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
1866 /* This function returns the amount that we can raise the
1867 * usable window based on the following constraints
1869 * 1. The window can never be shrunk once it is offered (RFC 793)
1870 * 2. We limit memory per socket
1873 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1874 * RECV.NEXT + RCV.WIN fixed until:
1875 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1877 * i.e. don't raise the right edge of the window until you can raise
1878 * it at least MSS bytes.
1880 * Unfortunately, the recommended algorithm breaks header prediction,
1881 * since header prediction assumes th->window stays fixed.
1883 * Strictly speaking, keeping th->window fixed violates the receiver
1884 * side SWS prevention criteria. The problem is that under this rule
1885 * a stream of single byte packets will cause the right side of the
1886 * window to always advance by a single byte.
1888 * Of course, if the sender implements sender side SWS prevention
1889 * then this will not be a problem.
1891 * BSD seems to make the following compromise:
1893 * If the free space is less than the 1/4 of the maximum
1894 * space available and the free space is less than 1/2 mss,
1895 * then set the window to 0.
1896 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1897 * Otherwise, just prevent the window from shrinking
1898 * and from being larger than the largest representable value.
1900 * This prevents incremental opening of the window in the regime
1901 * where TCP is limited by the speed of the reader side taking
1902 * data out of the TCP receive queue. It does nothing about
1903 * those cases where the window is constrained on the sender side
1904 * because the pipeline is full.
1906 * BSD also seems to "accidentally" limit itself to windows that are a
1907 * multiple of MSS, at least until the free space gets quite small.
1908 * This would appear to be a side effect of the mbuf implementation.
1909 * Combining these two algorithms results in the observed behavior
1910 * of having a fixed window size at almost all times.
1912 * Below we obtain similar behavior by forcing the offered window to
1913 * a multiple of the mss when it is feasible to do so.
1915 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1916 * Regular options like TIMESTAMP are taken into account.
1918 u32
__tcp_select_window(struct sock
*sk
)
1920 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1921 struct tcp_sock
*tp
= tcp_sk(sk
);
1922 /* MSS for the peer's data. Previous versions used mss_clamp
1923 * here. I don't know if the value based on our guesses
1924 * of peer's MSS is better for the performance. It's more correct
1925 * but may be worse for the performance because of rcv_mss
1926 * fluctuations. --SAW 1998/11/1
1928 int mss
= icsk
->icsk_ack
.rcv_mss
;
1929 int free_space
= tcp_space(sk
);
1930 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1933 if (mss
> full_space
)
1936 if (free_space
< (full_space
>> 1)) {
1937 icsk
->icsk_ack
.quick
= 0;
1939 if (sk_under_memory_pressure(sk
))
1940 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
1943 if (free_space
< mss
)
1947 if (free_space
> tp
->rcv_ssthresh
)
1948 free_space
= tp
->rcv_ssthresh
;
1950 /* Don't do rounding if we are using window scaling, since the
1951 * scaled window will not line up with the MSS boundary anyway.
1953 window
= tp
->rcv_wnd
;
1954 if (tp
->rx_opt
.rcv_wscale
) {
1955 window
= free_space
;
1957 /* Advertise enough space so that it won't get scaled away.
1958 * Import case: prevent zero window announcement if
1959 * 1<<rcv_wscale > mss.
1961 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1962 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1963 << tp
->rx_opt
.rcv_wscale
);
1965 /* Get the largest window that is a nice multiple of mss.
1966 * Window clamp already applied above.
1967 * If our current window offering is within 1 mss of the
1968 * free space we just keep it. This prevents the divide
1969 * and multiply from happening most of the time.
1970 * We also don't do any window rounding when the free space
1973 if (window
<= free_space
- mss
|| window
> free_space
)
1974 window
= (free_space
/ mss
) * mss
;
1975 else if (mss
== full_space
&&
1976 free_space
> window
+ (full_space
>> 1))
1977 window
= free_space
;
1983 /* Collapses two adjacent SKB's during retransmission. */
1984 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
1986 struct tcp_sock
*tp
= tcp_sk(sk
);
1987 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1988 int skb_size
, next_skb_size
;
1990 skb_size
= skb
->len
;
1991 next_skb_size
= next_skb
->len
;
1993 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
1995 tcp_highest_sack_combine(sk
, next_skb
, skb
);
1997 tcp_unlink_write_queue(next_skb
, sk
);
1999 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
2002 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2003 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2005 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2006 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2008 /* Update sequence range on original skb. */
2009 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2011 /* Merge over control information. This moves PSH/FIN etc. over */
2012 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2014 /* All done, get rid of second SKB and account for it so
2015 * packet counting does not break.
2017 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2019 /* changed transmit queue under us so clear hints */
2020 tcp_clear_retrans_hints_partial(tp
);
2021 if (next_skb
== tp
->retransmit_skb_hint
)
2022 tp
->retransmit_skb_hint
= skb
;
2024 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2026 sk_wmem_free_skb(sk
, next_skb
);
2029 /* Check if coalescing SKBs is legal. */
2030 static int tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2032 if (tcp_skb_pcount(skb
) > 1)
2034 /* TODO: SACK collapsing could be used to remove this condition */
2035 if (skb_shinfo(skb
)->nr_frags
!= 0)
2037 if (skb_cloned(skb
))
2039 if (skb
== tcp_send_head(sk
))
2041 /* Some heurestics for collapsing over SACK'd could be invented */
2042 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2048 /* Collapse packets in the retransmit queue to make to create
2049 * less packets on the wire. This is only done on retransmission.
2051 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2054 struct tcp_sock
*tp
= tcp_sk(sk
);
2055 struct sk_buff
*skb
= to
, *tmp
;
2058 if (!sysctl_tcp_retrans_collapse
)
2060 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2063 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2064 if (!tcp_can_collapse(sk
, skb
))
2076 /* Punt if not enough space exists in the first SKB for
2077 * the data in the second
2079 if (skb
->len
> skb_availroom(to
))
2082 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2085 tcp_collapse_retrans(sk
, to
);
2089 /* This retransmits one SKB. Policy decisions and retransmit queue
2090 * state updates are done by the caller. Returns non-zero if an
2091 * error occurred which prevented the send.
2093 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2095 struct tcp_sock
*tp
= tcp_sk(sk
);
2096 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2097 unsigned int cur_mss
;
2100 /* Inconslusive MTU probe */
2101 if (icsk
->icsk_mtup
.probe_size
) {
2102 icsk
->icsk_mtup
.probe_size
= 0;
2105 /* Do not sent more than we queued. 1/4 is reserved for possible
2106 * copying overhead: fragmentation, tunneling, mangling etc.
2108 if (atomic_read(&sk
->sk_wmem_alloc
) >
2109 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2112 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2113 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2115 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2119 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2120 return -EHOSTUNREACH
; /* Routing failure or similar. */
2122 cur_mss
= tcp_current_mss(sk
);
2124 /* If receiver has shrunk his window, and skb is out of
2125 * new window, do not retransmit it. The exception is the
2126 * case, when window is shrunk to zero. In this case
2127 * our retransmit serves as a zero window probe.
2129 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2130 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2133 if (skb
->len
> cur_mss
) {
2134 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
2135 return -ENOMEM
; /* We'll try again later. */
2137 int oldpcount
= tcp_skb_pcount(skb
);
2139 if (unlikely(oldpcount
> 1)) {
2140 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2141 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2145 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2147 /* Some Solaris stacks overoptimize and ignore the FIN on a
2148 * retransmit when old data is attached. So strip it off
2149 * since it is cheap to do so and saves bytes on the network.
2152 (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
2153 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
2154 if (!pskb_trim(skb
, 0)) {
2155 /* Reuse, even though it does some unnecessary work */
2156 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
2157 TCP_SKB_CB(skb
)->tcp_flags
);
2158 skb
->ip_summed
= CHECKSUM_NONE
;
2162 /* Make a copy, if the first transmission SKB clone we made
2163 * is still in somebody's hands, else make a clone.
2165 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2167 /* make sure skb->data is aligned on arches that require it */
2168 if (unlikely(NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3))) {
2169 struct sk_buff
*nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
,
2171 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2174 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2178 /* Update global TCP statistics. */
2179 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2181 tp
->total_retrans
++;
2183 #if FASTRETRANS_DEBUG > 0
2184 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2185 if (net_ratelimit())
2186 printk(KERN_DEBUG
"retrans_out leaked.\n");
2189 if (!tp
->retrans_out
)
2190 tp
->lost_retrans_low
= tp
->snd_nxt
;
2191 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2192 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2194 /* Save stamp of the first retransmit. */
2195 if (!tp
->retrans_stamp
)
2196 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2198 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2200 /* snd_nxt is stored to detect loss of retransmitted segment,
2201 * see tcp_input.c tcp_sacktag_write_queue().
2203 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2208 /* Check if we forward retransmits are possible in the current
2209 * window/congestion state.
2211 static int tcp_can_forward_retransmit(struct sock
*sk
)
2213 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2214 const struct tcp_sock
*tp
= tcp_sk(sk
);
2216 /* Forward retransmissions are possible only during Recovery. */
2217 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2220 /* No forward retransmissions in Reno are possible. */
2221 if (tcp_is_reno(tp
))
2224 /* Yeah, we have to make difficult choice between forward transmission
2225 * and retransmission... Both ways have their merits...
2227 * For now we do not retransmit anything, while we have some new
2228 * segments to send. In the other cases, follow rule 3 for
2229 * NextSeg() specified in RFC3517.
2232 if (tcp_may_send_now(sk
))
2238 /* This gets called after a retransmit timeout, and the initially
2239 * retransmitted data is acknowledged. It tries to continue
2240 * resending the rest of the retransmit queue, until either
2241 * we've sent it all or the congestion window limit is reached.
2242 * If doing SACK, the first ACK which comes back for a timeout
2243 * based retransmit packet might feed us FACK information again.
2244 * If so, we use it to avoid unnecessarily retransmissions.
2246 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2248 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2249 struct tcp_sock
*tp
= tcp_sk(sk
);
2250 struct sk_buff
*skb
;
2251 struct sk_buff
*hole
= NULL
;
2254 int fwd_rexmitting
= 0;
2256 if (!tp
->packets_out
)
2260 tp
->retransmit_high
= tp
->snd_una
;
2262 if (tp
->retransmit_skb_hint
) {
2263 skb
= tp
->retransmit_skb_hint
;
2264 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2265 if (after(last_lost
, tp
->retransmit_high
))
2266 last_lost
= tp
->retransmit_high
;
2268 skb
= tcp_write_queue_head(sk
);
2269 last_lost
= tp
->snd_una
;
2272 tcp_for_write_queue_from(skb
, sk
) {
2273 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2275 if (skb
== tcp_send_head(sk
))
2277 /* we could do better than to assign each time */
2279 tp
->retransmit_skb_hint
= skb
;
2281 /* Assume this retransmit will generate
2282 * only one packet for congestion window
2283 * calculation purposes. This works because
2284 * tcp_retransmit_skb() will chop up the
2285 * packet to be MSS sized and all the
2286 * packet counting works out.
2288 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2291 if (fwd_rexmitting
) {
2293 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2295 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2297 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2298 tp
->retransmit_high
= last_lost
;
2299 if (!tcp_can_forward_retransmit(sk
))
2301 /* Backtrack if necessary to non-L'ed skb */
2309 } else if (!(sacked
& TCPCB_LOST
)) {
2310 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2315 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2316 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2317 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2319 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2322 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2325 if (tcp_retransmit_skb(sk
, skb
)) {
2326 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2329 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2331 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Recovery
)
2332 tp
->prr_out
+= tcp_skb_pcount(skb
);
2334 if (skb
== tcp_write_queue_head(sk
))
2335 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2336 inet_csk(sk
)->icsk_rto
,
2341 /* Send a fin. The caller locks the socket for us. This cannot be
2342 * allowed to fail queueing a FIN frame under any circumstances.
2344 void tcp_send_fin(struct sock
*sk
)
2346 struct tcp_sock
*tp
= tcp_sk(sk
);
2347 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2350 /* Optimization, tack on the FIN if we have a queue of
2351 * unsent frames. But be careful about outgoing SACKS
2354 mss_now
= tcp_current_mss(sk
);
2356 if (tcp_send_head(sk
) != NULL
) {
2357 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_FIN
;
2358 TCP_SKB_CB(skb
)->end_seq
++;
2361 /* Socket is locked, keep trying until memory is available. */
2363 skb
= alloc_skb_fclone(MAX_TCP_HEADER
,
2370 /* Reserve space for headers and prepare control bits. */
2371 skb_reserve(skb
, MAX_TCP_HEADER
);
2372 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2373 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2374 TCPHDR_ACK
| TCPHDR_FIN
);
2375 tcp_queue_skb(sk
, skb
);
2377 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2380 /* We get here when a process closes a file descriptor (either due to
2381 * an explicit close() or as a byproduct of exit()'ing) and there
2382 * was unread data in the receive queue. This behavior is recommended
2383 * by RFC 2525, section 2.17. -DaveM
2385 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2387 struct sk_buff
*skb
;
2389 /* NOTE: No TCP options attached and we never retransmit this. */
2390 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2392 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2396 /* Reserve space for headers and prepare control bits. */
2397 skb_reserve(skb
, MAX_TCP_HEADER
);
2398 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2399 TCPHDR_ACK
| TCPHDR_RST
);
2401 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2402 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2403 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2405 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2408 /* Send a crossed SYN-ACK during socket establishment.
2409 * WARNING: This routine must only be called when we have already sent
2410 * a SYN packet that crossed the incoming SYN that caused this routine
2411 * to get called. If this assumption fails then the initial rcv_wnd
2412 * and rcv_wscale values will not be correct.
2414 int tcp_send_synack(struct sock
*sk
)
2416 struct sk_buff
*skb
;
2418 skb
= tcp_write_queue_head(sk
);
2419 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2420 printk(KERN_DEBUG
"tcp_send_synack: wrong queue state\n");
2423 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
2424 if (skb_cloned(skb
)) {
2425 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2428 tcp_unlink_write_queue(skb
, sk
);
2429 skb_header_release(nskb
);
2430 __tcp_add_write_queue_head(sk
, nskb
);
2431 sk_wmem_free_skb(sk
, skb
);
2432 sk
->sk_wmem_queued
+= nskb
->truesize
;
2433 sk_mem_charge(sk
, nskb
->truesize
);
2437 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
2438 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2440 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2441 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2444 /* Prepare a SYN-ACK. */
2445 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2446 struct request_sock
*req
,
2447 struct request_values
*rvp
)
2449 struct tcp_out_options opts
;
2450 struct tcp_extend_values
*xvp
= tcp_xv(rvp
);
2451 struct inet_request_sock
*ireq
= inet_rsk(req
);
2452 struct tcp_sock
*tp
= tcp_sk(sk
);
2453 const struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
2455 struct sk_buff
*skb
;
2456 struct tcp_md5sig_key
*md5
;
2457 int tcp_header_size
;
2459 int s_data_desired
= 0;
2461 if (cvp
!= NULL
&& cvp
->s_data_constant
&& cvp
->s_data_desired
)
2462 s_data_desired
= cvp
->s_data_desired
;
2463 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15 + s_data_desired
, 1, GFP_ATOMIC
);
2467 /* Reserve space for headers. */
2468 skb_reserve(skb
, MAX_TCP_HEADER
);
2470 skb_dst_set(skb
, dst_clone(dst
));
2472 mss
= dst_metric_advmss(dst
);
2473 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2474 mss
= tp
->rx_opt
.user_mss
;
2476 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2478 /* Set this up on the first call only */
2479 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2481 /* limit the window selection if the user enforce a smaller rx buffer */
2482 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2483 (req
->window_clamp
> tcp_full_space(sk
) || req
->window_clamp
== 0))
2484 req
->window_clamp
= tcp_full_space(sk
);
2486 /* tcp_full_space because it is guaranteed to be the first packet */
2487 tcp_select_initial_window(tcp_full_space(sk
),
2488 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2493 dst_metric(dst
, RTAX_INITRWND
));
2494 ireq
->rcv_wscale
= rcv_wscale
;
2497 memset(&opts
, 0, sizeof(opts
));
2498 #ifdef CONFIG_SYN_COOKIES
2499 if (unlikely(req
->cookie_ts
))
2500 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2503 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2504 tcp_header_size
= tcp_synack_options(sk
, req
, mss
,
2505 skb
, &opts
, &md5
, xvp
)
2508 skb_push(skb
, tcp_header_size
);
2509 skb_reset_transport_header(skb
);
2512 memset(th
, 0, sizeof(struct tcphdr
));
2515 TCP_ECN_make_synack(req
, th
);
2516 th
->source
= ireq
->loc_port
;
2517 th
->dest
= ireq
->rmt_port
;
2518 /* Setting of flags are superfluous here for callers (and ECE is
2519 * not even correctly set)
2521 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2522 TCPHDR_SYN
| TCPHDR_ACK
);
2524 if (OPTION_COOKIE_EXTENSION
& opts
.options
) {
2525 if (s_data_desired
) {
2526 u8
*buf
= skb_put(skb
, s_data_desired
);
2528 /* copy data directly from the listening socket. */
2529 memcpy(buf
, cvp
->s_data_payload
, s_data_desired
);
2530 TCP_SKB_CB(skb
)->end_seq
+= s_data_desired
;
2533 if (opts
.hash_size
> 0) {
2534 __u32 workspace
[SHA_WORKSPACE_WORDS
];
2535 u32
*mess
= &xvp
->cookie_bakery
[COOKIE_DIGEST_WORDS
];
2536 u32
*tail
= &mess
[COOKIE_MESSAGE_WORDS
-1];
2538 /* Secret recipe depends on the Timestamp, (future)
2539 * Sequence and Acknowledgment Numbers, Initiator
2540 * Cookie, and others handled by IP variant caller.
2542 *tail
-- ^= opts
.tsval
;
2543 *tail
-- ^= tcp_rsk(req
)->rcv_isn
+ 1;
2544 *tail
-- ^= TCP_SKB_CB(skb
)->seq
+ 1;
2547 *tail
-- ^= (((__force u32
)th
->dest
<< 16) | (__force u32
)th
->source
);
2548 *tail
-- ^= (u32
)(unsigned long)cvp
; /* per sockopt */
2550 sha_transform((__u32
*)&xvp
->cookie_bakery
[0],
2553 opts
.hash_location
=
2554 (__u8
*)&xvp
->cookie_bakery
[0];
2558 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2559 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2561 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2562 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2563 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2564 th
->doff
= (tcp_header_size
>> 2);
2565 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
, tcp_skb_pcount(skb
));
2567 #ifdef CONFIG_TCP_MD5SIG
2568 /* Okay, we have all we need - do the md5 hash if needed */
2570 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2571 md5
, NULL
, req
, skb
);
2577 EXPORT_SYMBOL(tcp_make_synack
);
2579 /* Do all connect socket setups that can be done AF independent. */
2580 void tcp_connect_init(struct sock
*sk
)
2582 const struct dst_entry
*dst
= __sk_dst_get(sk
);
2583 struct tcp_sock
*tp
= tcp_sk(sk
);
2586 /* We'll fix this up when we get a response from the other end.
2587 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2589 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2590 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2592 #ifdef CONFIG_TCP_MD5SIG
2593 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2594 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2597 /* If user gave his TCP_MAXSEG, record it to clamp */
2598 if (tp
->rx_opt
.user_mss
)
2599 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2602 tcp_sync_mss(sk
, dst_mtu(dst
));
2604 if (!tp
->window_clamp
)
2605 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2606 tp
->advmss
= dst_metric_advmss(dst
);
2607 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2608 tp
->advmss
= tp
->rx_opt
.user_mss
;
2610 tcp_initialize_rcv_mss(sk
);
2612 /* limit the window selection if the user enforce a smaller rx buffer */
2613 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2614 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
2615 tp
->window_clamp
= tcp_full_space(sk
);
2617 tcp_select_initial_window(tcp_full_space(sk
),
2618 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2621 sysctl_tcp_window_scaling
,
2623 dst_metric(dst
, RTAX_INITRWND
));
2625 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2626 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2629 sock_reset_flag(sk
, SOCK_DONE
);
2632 tp
->snd_una
= tp
->write_seq
;
2633 tp
->snd_sml
= tp
->write_seq
;
2634 tp
->snd_up
= tp
->write_seq
;
2635 tp
->snd_nxt
= tp
->write_seq
;
2637 if (likely(!tp
->repair
))
2639 tp
->rcv_wup
= tp
->rcv_nxt
;
2640 tp
->copied_seq
= tp
->rcv_nxt
;
2642 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2643 inet_csk(sk
)->icsk_retransmits
= 0;
2644 tcp_clear_retrans(tp
);
2647 /* Build a SYN and send it off. */
2648 int tcp_connect(struct sock
*sk
)
2650 struct tcp_sock
*tp
= tcp_sk(sk
);
2651 struct sk_buff
*buff
;
2654 tcp_connect_init(sk
);
2656 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2657 if (unlikely(buff
== NULL
))
2660 /* Reserve space for headers. */
2661 skb_reserve(buff
, MAX_TCP_HEADER
);
2663 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
2664 TCP_ECN_send_syn(sk
, buff
);
2667 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2668 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2669 skb_header_release(buff
);
2670 __tcp_add_write_queue_tail(sk
, buff
);
2671 sk
->sk_wmem_queued
+= buff
->truesize
;
2672 sk_mem_charge(sk
, buff
->truesize
);
2673 tp
->packets_out
+= tcp_skb_pcount(buff
);
2674 err
= tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
2675 if (err
== -ECONNREFUSED
)
2678 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2679 * in order to make this packet get counted in tcpOutSegs.
2681 tp
->snd_nxt
= tp
->write_seq
;
2682 tp
->pushed_seq
= tp
->write_seq
;
2683 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2685 /* Timer for repeating the SYN until an answer. */
2686 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2687 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2690 EXPORT_SYMBOL(tcp_connect
);
2692 /* Send out a delayed ack, the caller does the policy checking
2693 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2696 void tcp_send_delayed_ack(struct sock
*sk
)
2698 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2699 int ato
= icsk
->icsk_ack
.ato
;
2700 unsigned long timeout
;
2702 if (ato
> TCP_DELACK_MIN
) {
2703 const struct tcp_sock
*tp
= tcp_sk(sk
);
2704 int max_ato
= HZ
/ 2;
2706 if (icsk
->icsk_ack
.pingpong
||
2707 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2708 max_ato
= TCP_DELACK_MAX
;
2710 /* Slow path, intersegment interval is "high". */
2712 /* If some rtt estimate is known, use it to bound delayed ack.
2713 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2717 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
2723 ato
= min(ato
, max_ato
);
2726 /* Stay within the limit we were given */
2727 timeout
= jiffies
+ ato
;
2729 /* Use new timeout only if there wasn't a older one earlier. */
2730 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2731 /* If delack timer was blocked or is about to expire,
2734 if (icsk
->icsk_ack
.blocked
||
2735 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2740 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2741 timeout
= icsk
->icsk_ack
.timeout
;
2743 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2744 icsk
->icsk_ack
.timeout
= timeout
;
2745 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2748 /* This routine sends an ack and also updates the window. */
2749 void tcp_send_ack(struct sock
*sk
)
2751 struct sk_buff
*buff
;
2753 /* If we have been reset, we may not send again. */
2754 if (sk
->sk_state
== TCP_CLOSE
)
2757 /* We are not putting this on the write queue, so
2758 * tcp_transmit_skb() will set the ownership to this
2761 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2763 inet_csk_schedule_ack(sk
);
2764 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2765 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2766 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2770 /* Reserve space for headers and prepare control bits. */
2771 skb_reserve(buff
, MAX_TCP_HEADER
);
2772 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
2774 /* Send it off, this clears delayed acks for us. */
2775 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2776 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2779 /* This routine sends a packet with an out of date sequence
2780 * number. It assumes the other end will try to ack it.
2782 * Question: what should we make while urgent mode?
2783 * 4.4BSD forces sending single byte of data. We cannot send
2784 * out of window data, because we have SND.NXT==SND.MAX...
2786 * Current solution: to send TWO zero-length segments in urgent mode:
2787 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2788 * out-of-date with SND.UNA-1 to probe window.
2790 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2792 struct tcp_sock
*tp
= tcp_sk(sk
);
2793 struct sk_buff
*skb
;
2795 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2796 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2800 /* Reserve space for headers and set control bits. */
2801 skb_reserve(skb
, MAX_TCP_HEADER
);
2802 /* Use a previous sequence. This should cause the other
2803 * end to send an ack. Don't queue or clone SKB, just
2806 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
2807 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2808 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2811 void tcp_send_window_probe(struct sock
*sk
)
2813 if (sk
->sk_state
== TCP_ESTABLISHED
) {
2814 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
2815 tcp_sk(sk
)->snd_nxt
= tcp_sk(sk
)->write_seq
;
2816 tcp_xmit_probe_skb(sk
, 0);
2820 /* Initiate keepalive or window probe from timer. */
2821 int tcp_write_wakeup(struct sock
*sk
)
2823 struct tcp_sock
*tp
= tcp_sk(sk
);
2824 struct sk_buff
*skb
;
2826 if (sk
->sk_state
== TCP_CLOSE
)
2829 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2830 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
2832 unsigned int mss
= tcp_current_mss(sk
);
2833 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2835 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2836 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2838 /* We are probing the opening of a window
2839 * but the window size is != 0
2840 * must have been a result SWS avoidance ( sender )
2842 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2844 seg_size
= min(seg_size
, mss
);
2845 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
2846 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2848 } else if (!tcp_skb_pcount(skb
))
2849 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2851 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
2852 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2853 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2855 tcp_event_new_data_sent(sk
, skb
);
2858 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
2859 tcp_xmit_probe_skb(sk
, 1);
2860 return tcp_xmit_probe_skb(sk
, 0);
2864 /* A window probe timeout has occurred. If window is not closed send
2865 * a partial packet else a zero probe.
2867 void tcp_send_probe0(struct sock
*sk
)
2869 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2870 struct tcp_sock
*tp
= tcp_sk(sk
);
2873 err
= tcp_write_wakeup(sk
);
2875 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2876 /* Cancel probe timer, if it is not required. */
2877 icsk
->icsk_probes_out
= 0;
2878 icsk
->icsk_backoff
= 0;
2883 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2884 icsk
->icsk_backoff
++;
2885 icsk
->icsk_probes_out
++;
2886 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2887 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2890 /* If packet was not sent due to local congestion,
2891 * do not backoff and do not remember icsk_probes_out.
2892 * Let local senders to fight for local resources.
2894 * Use accumulated backoff yet.
2896 if (!icsk
->icsk_probes_out
)
2897 icsk
->icsk_probes_out
= 1;
2898 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2899 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2900 TCP_RESOURCE_PROBE_INTERVAL
),