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
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
53 /* This limits the percentage of the congestion window which we
54 * will allow a single TSO frame to consume. Building TSO frames
55 * which are too large can cause TCP streams to be bursty.
57 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
59 int sysctl_tcp_mtu_probing __read_mostly
= 0;
60 int sysctl_tcp_base_mss __read_mostly
= TCP_BASE_MSS
;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
65 int sysctl_tcp_cookie_size __read_mostly
= 0; /* TCP_COOKIE_MAX */
66 EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size
);
69 /* Account for new data that has been sent to the network. */
70 static void tcp_event_new_data_sent(struct sock
*sk
, const struct sk_buff
*skb
)
72 struct tcp_sock
*tp
= tcp_sk(sk
);
73 unsigned int prior_packets
= tp
->packets_out
;
75 tcp_advance_send_head(sk
, skb
);
76 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
78 /* Don't override Nagle indefinitely with F-RTO */
79 if (tp
->frto_counter
== 2)
82 tp
->packets_out
+= tcp_skb_pcount(skb
);
83 if (!prior_packets
|| tp
->early_retrans_delayed
)
87 /* SND.NXT, if window was not shrunk.
88 * If window has been shrunk, what should we make? It is not clear at all.
89 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
90 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
91 * invalid. OK, let's make this for now:
93 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
95 const struct tcp_sock
*tp
= tcp_sk(sk
);
97 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
100 return tcp_wnd_end(tp
);
103 /* Calculate mss to advertise in SYN segment.
104 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
106 * 1. It is independent of path mtu.
107 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
108 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
109 * attached devices, because some buggy hosts are confused by
111 * 4. We do not make 3, we advertise MSS, calculated from first
112 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
113 * This may be overridden via information stored in routing table.
114 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
115 * probably even Jumbo".
117 static __u16
tcp_advertise_mss(struct sock
*sk
)
119 struct tcp_sock
*tp
= tcp_sk(sk
);
120 const struct dst_entry
*dst
= __sk_dst_get(sk
);
121 int mss
= tp
->advmss
;
124 unsigned int metric
= dst_metric_advmss(dst
);
135 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
136 * This is the first part of cwnd validation mechanism. */
137 static void tcp_cwnd_restart(struct sock
*sk
, const struct dst_entry
*dst
)
139 struct tcp_sock
*tp
= tcp_sk(sk
);
140 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
141 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
142 u32 cwnd
= tp
->snd_cwnd
;
144 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
146 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
147 restart_cwnd
= min(restart_cwnd
, cwnd
);
149 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
151 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
152 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
153 tp
->snd_cwnd_used
= 0;
156 /* Congestion state accounting after a packet has been sent. */
157 static void tcp_event_data_sent(struct tcp_sock
*tp
,
160 struct inet_connection_sock
*icsk
= inet_csk(sk
);
161 const u32 now
= tcp_time_stamp
;
163 if (sysctl_tcp_slow_start_after_idle
&&
164 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
165 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
169 /* If it is a reply for ato after last received
170 * packet, enter pingpong mode.
172 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
173 icsk
->icsk_ack
.pingpong
= 1;
176 /* Account for an ACK we sent. */
177 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
179 tcp_dec_quickack_mode(sk
, pkts
);
180 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
183 /* Determine a window scaling and initial window to offer.
184 * Based on the assumption that the given amount of space
185 * will be offered. Store the results in the tp structure.
186 * NOTE: for smooth operation initial space offering should
187 * be a multiple of mss if possible. We assume here that mss >= 1.
188 * This MUST be enforced by all callers.
190 void tcp_select_initial_window(int __space
, __u32 mss
,
191 __u32
*rcv_wnd
, __u32
*window_clamp
,
192 int wscale_ok
, __u8
*rcv_wscale
,
195 unsigned int space
= (__space
< 0 ? 0 : __space
);
197 /* If no clamp set the clamp to the max possible scaled window */
198 if (*window_clamp
== 0)
199 (*window_clamp
) = (65535 << 14);
200 space
= min(*window_clamp
, space
);
202 /* Quantize space offering to a multiple of mss if possible. */
204 space
= (space
/ mss
) * mss
;
206 /* NOTE: offering an initial window larger than 32767
207 * will break some buggy TCP stacks. If the admin tells us
208 * it is likely we could be speaking with such a buggy stack
209 * we will truncate our initial window offering to 32K-1
210 * unless the remote has sent us a window scaling option,
211 * which we interpret as a sign the remote TCP is not
212 * misinterpreting the window field as a signed quantity.
214 if (sysctl_tcp_workaround_signed_windows
)
215 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
221 /* Set window scaling on max possible window
222 * See RFC1323 for an explanation of the limit to 14
224 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
225 space
= min_t(u32
, space
, *window_clamp
);
226 while (space
> 65535 && (*rcv_wscale
) < 14) {
232 /* Set initial window to a value enough for senders starting with
233 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
234 * a limit on the initial window when mss is larger than 1460.
236 if (mss
> (1 << *rcv_wscale
)) {
237 int init_cwnd
= TCP_DEFAULT_INIT_RCVWND
;
240 max_t(u32
, (1460 * TCP_DEFAULT_INIT_RCVWND
) / mss
, 2);
241 /* when initializing use the value from init_rcv_wnd
242 * rather than the default from above
245 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
247 *rcv_wnd
= min(*rcv_wnd
, init_cwnd
* mss
);
250 /* Set the clamp no higher than max representable value */
251 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
253 EXPORT_SYMBOL(tcp_select_initial_window
);
255 /* Chose a new window to advertise, update state in tcp_sock for the
256 * socket, and return result with RFC1323 scaling applied. The return
257 * value can be stuffed directly into th->window for an outgoing
260 static u16
tcp_select_window(struct sock
*sk
)
262 struct tcp_sock
*tp
= tcp_sk(sk
);
263 u32 cur_win
= tcp_receive_window(tp
);
264 u32 new_win
= __tcp_select_window(sk
);
266 /* Never shrink the offered window */
267 if (new_win
< cur_win
) {
268 /* Danger Will Robinson!
269 * Don't update rcv_wup/rcv_wnd here or else
270 * we will not be able to advertise a zero
271 * window in time. --DaveM
273 * Relax Will Robinson.
275 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
277 tp
->rcv_wnd
= new_win
;
278 tp
->rcv_wup
= tp
->rcv_nxt
;
280 /* Make sure we do not exceed the maximum possible
283 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
284 new_win
= min(new_win
, MAX_TCP_WINDOW
);
286 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
288 /* RFC1323 scaling applied */
289 new_win
>>= tp
->rx_opt
.rcv_wscale
;
291 /* If we advertise zero window, disable fast path. */
298 /* Packet ECN state for a SYN-ACK */
299 static inline void TCP_ECN_send_synack(const struct tcp_sock
*tp
, struct sk_buff
*skb
)
301 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
302 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
303 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
306 /* Packet ECN state for a SYN. */
307 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
309 struct tcp_sock
*tp
= tcp_sk(sk
);
312 if (sysctl_tcp_ecn
== 1) {
313 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
314 tp
->ecn_flags
= TCP_ECN_OK
;
318 static __inline__
void
319 TCP_ECN_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
321 if (inet_rsk(req
)->ecn_ok
)
325 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
328 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
331 struct tcp_sock
*tp
= tcp_sk(sk
);
333 if (tp
->ecn_flags
& TCP_ECN_OK
) {
334 /* Not-retransmitted data segment: set ECT and inject CWR. */
335 if (skb
->len
!= tcp_header_len
&&
336 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
338 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
339 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
340 tcp_hdr(skb
)->cwr
= 1;
341 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
344 /* ACK or retransmitted segment: clear ECT|CE */
345 INET_ECN_dontxmit(sk
);
347 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
348 tcp_hdr(skb
)->ece
= 1;
352 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
353 * auto increment end seqno.
355 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
357 skb
->ip_summed
= CHECKSUM_PARTIAL
;
360 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
361 TCP_SKB_CB(skb
)->sacked
= 0;
363 skb_shinfo(skb
)->gso_segs
= 1;
364 skb_shinfo(skb
)->gso_size
= 0;
365 skb_shinfo(skb
)->gso_type
= 0;
367 TCP_SKB_CB(skb
)->seq
= seq
;
368 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
370 TCP_SKB_CB(skb
)->end_seq
= seq
;
373 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
375 return tp
->snd_una
!= tp
->snd_up
;
378 #define OPTION_SACK_ADVERTISE (1 << 0)
379 #define OPTION_TS (1 << 1)
380 #define OPTION_MD5 (1 << 2)
381 #define OPTION_WSCALE (1 << 3)
382 #define OPTION_COOKIE_EXTENSION (1 << 4)
384 struct tcp_out_options
{
385 u8 options
; /* bit field of OPTION_* */
386 u8 ws
; /* window scale, 0 to disable */
387 u8 num_sack_blocks
; /* number of SACK blocks to include */
388 u8 hash_size
; /* bytes in hash_location */
389 u16 mss
; /* 0 to disable */
390 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
391 __u8
*hash_location
; /* temporary pointer, overloaded */
394 /* The sysctl int routines are generic, so check consistency here.
396 static u8
tcp_cookie_size_check(u8 desired
)
401 /* previously specified */
404 cookie_size
= ACCESS_ONCE(sysctl_tcp_cookie_size
);
405 if (cookie_size
<= 0)
406 /* no default specified */
409 if (cookie_size
<= TCP_COOKIE_MIN
)
410 /* value too small, specify minimum */
411 return TCP_COOKIE_MIN
;
413 if (cookie_size
>= TCP_COOKIE_MAX
)
414 /* value too large, specify maximum */
415 return TCP_COOKIE_MAX
;
418 /* 8-bit multiple, illegal, fix it */
421 return (u8
)cookie_size
;
424 /* Write previously computed TCP options to the packet.
426 * Beware: Something in the Internet is very sensitive to the ordering of
427 * TCP options, we learned this through the hard way, so be careful here.
428 * Luckily we can at least blame others for their non-compliance but from
429 * inter-operatibility perspective it seems that we're somewhat stuck with
430 * the ordering which we have been using if we want to keep working with
431 * those broken things (not that it currently hurts anybody as there isn't
432 * particular reason why the ordering would need to be changed).
434 * At least SACK_PERM as the first option is known to lead to a disaster
435 * (but it may well be that other scenarios fail similarly).
437 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
438 struct tcp_out_options
*opts
)
440 u8 options
= opts
->options
; /* mungable copy */
442 /* Having both authentication and cookies for security is redundant,
443 * and there's certainly not enough room. Instead, the cookie-less
444 * extension variant is proposed.
446 * Consider the pessimal case with authentication. The options
448 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
450 if (unlikely(OPTION_MD5
& options
)) {
451 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
452 *ptr
++ = htonl((TCPOPT_COOKIE
<< 24) |
453 (TCPOLEN_COOKIE_BASE
<< 16) |
454 (TCPOPT_MD5SIG
<< 8) |
457 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
459 (TCPOPT_MD5SIG
<< 8) |
462 options
&= ~OPTION_COOKIE_EXTENSION
;
463 /* overload cookie hash location */
464 opts
->hash_location
= (__u8
*)ptr
;
468 if (unlikely(opts
->mss
)) {
469 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
470 (TCPOLEN_MSS
<< 16) |
474 if (likely(OPTION_TS
& options
)) {
475 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
476 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
477 (TCPOLEN_SACK_PERM
<< 16) |
478 (TCPOPT_TIMESTAMP
<< 8) |
480 options
&= ~OPTION_SACK_ADVERTISE
;
482 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
484 (TCPOPT_TIMESTAMP
<< 8) |
487 *ptr
++ = htonl(opts
->tsval
);
488 *ptr
++ = htonl(opts
->tsecr
);
491 /* Specification requires after timestamp, so do it now.
493 * Consider the pessimal case without authentication. The options
495 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
497 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
498 __u8
*cookie_copy
= opts
->hash_location
;
499 u8 cookie_size
= opts
->hash_size
;
501 /* 8-bit multiple handled in tcp_cookie_size_check() above,
504 if (0x2 & cookie_size
) {
505 __u8
*p
= (__u8
*)ptr
;
507 /* 16-bit multiple */
508 *p
++ = TCPOPT_COOKIE
;
509 *p
++ = TCPOLEN_COOKIE_BASE
+ cookie_size
;
510 *p
++ = *cookie_copy
++;
511 *p
++ = *cookie_copy
++;
515 /* 32-bit multiple */
516 *ptr
++ = htonl(((TCPOPT_NOP
<< 24) |
518 (TCPOPT_COOKIE
<< 8) |
519 TCPOLEN_COOKIE_BASE
) +
523 if (cookie_size
> 0) {
524 memcpy(ptr
, cookie_copy
, cookie_size
);
525 ptr
+= (cookie_size
/ 4);
529 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
530 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
532 (TCPOPT_SACK_PERM
<< 8) |
536 if (unlikely(OPTION_WSCALE
& options
)) {
537 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
538 (TCPOPT_WINDOW
<< 16) |
539 (TCPOLEN_WINDOW
<< 8) |
543 if (unlikely(opts
->num_sack_blocks
)) {
544 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
545 tp
->duplicate_sack
: tp
->selective_acks
;
548 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
551 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
552 TCPOLEN_SACK_PERBLOCK
)));
554 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
556 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
557 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
560 tp
->rx_opt
.dsack
= 0;
564 /* Compute TCP options for SYN packets. This is not the final
565 * network wire format yet.
567 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
568 struct tcp_out_options
*opts
,
569 struct tcp_md5sig_key
**md5
)
571 struct tcp_sock
*tp
= tcp_sk(sk
);
572 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
573 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
574 u8 cookie_size
= (!tp
->rx_opt
.cookie_out_never
&& cvp
!= NULL
) ?
575 tcp_cookie_size_check(cvp
->cookie_desired
) :
578 #ifdef CONFIG_TCP_MD5SIG
579 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
581 opts
->options
|= OPTION_MD5
;
582 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
588 /* We always get an MSS option. The option bytes which will be seen in
589 * normal data packets should timestamps be used, must be in the MSS
590 * advertised. But we subtract them from tp->mss_cache so that
591 * calculations in tcp_sendmsg are simpler etc. So account for this
592 * fact here if necessary. If we don't do this correctly, as a
593 * receiver we won't recognize data packets as being full sized when we
594 * should, and thus we won't abide by the delayed ACK rules correctly.
595 * SACKs don't matter, we never delay an ACK when we have any of those
597 opts
->mss
= tcp_advertise_mss(sk
);
598 remaining
-= TCPOLEN_MSS_ALIGNED
;
600 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
601 opts
->options
|= OPTION_TS
;
602 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
603 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
604 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
606 if (likely(sysctl_tcp_window_scaling
)) {
607 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
608 opts
->options
|= OPTION_WSCALE
;
609 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
611 if (likely(sysctl_tcp_sack
)) {
612 opts
->options
|= OPTION_SACK_ADVERTISE
;
613 if (unlikely(!(OPTION_TS
& opts
->options
)))
614 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
617 /* Note that timestamps are required by the specification.
619 * Odd numbers of bytes are prohibited by the specification, ensuring
620 * that the cookie is 16-bit aligned, and the resulting cookie pair is
624 (OPTION_TS
& opts
->options
) &&
626 int need
= TCPOLEN_COOKIE_BASE
+ cookie_size
;
629 /* 32-bit multiple */
630 need
+= 2; /* NOPs */
632 if (need
> remaining
) {
633 /* try shrinking cookie to fit */
638 while (need
> remaining
&& TCP_COOKIE_MIN
<= cookie_size
) {
642 if (TCP_COOKIE_MIN
<= cookie_size
) {
643 opts
->options
|= OPTION_COOKIE_EXTENSION
;
644 opts
->hash_location
= (__u8
*)&cvp
->cookie_pair
[0];
645 opts
->hash_size
= cookie_size
;
647 /* Remember for future incarnations. */
648 cvp
->cookie_desired
= cookie_size
;
650 if (cvp
->cookie_desired
!= cvp
->cookie_pair_size
) {
651 /* Currently use random bytes as a nonce,
652 * assuming these are completely unpredictable
653 * by hostile users of the same system.
655 get_random_bytes(&cvp
->cookie_pair
[0],
657 cvp
->cookie_pair_size
= cookie_size
;
663 return MAX_TCP_OPTION_SPACE
- remaining
;
666 /* Set up TCP options for SYN-ACKs. */
667 static unsigned int tcp_synack_options(struct sock
*sk
,
668 struct request_sock
*req
,
669 unsigned int mss
, struct sk_buff
*skb
,
670 struct tcp_out_options
*opts
,
671 struct tcp_md5sig_key
**md5
,
672 struct tcp_extend_values
*xvp
)
674 struct inet_request_sock
*ireq
= inet_rsk(req
);
675 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
676 u8 cookie_plus
= (xvp
!= NULL
&& !xvp
->cookie_out_never
) ?
680 #ifdef CONFIG_TCP_MD5SIG
681 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
683 opts
->options
|= OPTION_MD5
;
684 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
686 /* We can't fit any SACK blocks in a packet with MD5 + TS
687 * options. There was discussion about disabling SACK
688 * rather than TS in order to fit in better with old,
689 * buggy kernels, but that was deemed to be unnecessary.
691 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
697 /* We always send an MSS option. */
699 remaining
-= TCPOLEN_MSS_ALIGNED
;
701 if (likely(ireq
->wscale_ok
)) {
702 opts
->ws
= ireq
->rcv_wscale
;
703 opts
->options
|= OPTION_WSCALE
;
704 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
706 if (likely(ireq
->tstamp_ok
)) {
707 opts
->options
|= OPTION_TS
;
708 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
709 opts
->tsecr
= req
->ts_recent
;
710 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
712 if (likely(ireq
->sack_ok
)) {
713 opts
->options
|= OPTION_SACK_ADVERTISE
;
714 if (unlikely(!ireq
->tstamp_ok
))
715 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
718 /* Similar rationale to tcp_syn_options() applies here, too.
719 * If the <SYN> options fit, the same options should fit now!
723 cookie_plus
> TCPOLEN_COOKIE_BASE
) {
724 int need
= cookie_plus
; /* has TCPOLEN_COOKIE_BASE */
727 /* 32-bit multiple */
728 need
+= 2; /* NOPs */
730 if (need
<= remaining
) {
731 opts
->options
|= OPTION_COOKIE_EXTENSION
;
732 opts
->hash_size
= cookie_plus
- TCPOLEN_COOKIE_BASE
;
735 /* There's no error return, so flag it. */
736 xvp
->cookie_out_never
= 1; /* true */
740 return MAX_TCP_OPTION_SPACE
- remaining
;
743 /* Compute TCP options for ESTABLISHED sockets. This is not the
744 * final wire format yet.
746 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
747 struct tcp_out_options
*opts
,
748 struct tcp_md5sig_key
**md5
)
750 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
751 struct tcp_sock
*tp
= tcp_sk(sk
);
752 unsigned int size
= 0;
753 unsigned int eff_sacks
;
755 #ifdef CONFIG_TCP_MD5SIG
756 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
757 if (unlikely(*md5
)) {
758 opts
->options
|= OPTION_MD5
;
759 size
+= TCPOLEN_MD5SIG_ALIGNED
;
765 if (likely(tp
->rx_opt
.tstamp_ok
)) {
766 opts
->options
|= OPTION_TS
;
767 opts
->tsval
= tcb
? tcb
->when
: 0;
768 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
769 size
+= TCPOLEN_TSTAMP_ALIGNED
;
772 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
773 if (unlikely(eff_sacks
)) {
774 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
775 opts
->num_sack_blocks
=
776 min_t(unsigned int, eff_sacks
,
777 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
778 TCPOLEN_SACK_PERBLOCK
);
779 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
780 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
786 /* This routine actually transmits TCP packets queued in by
787 * tcp_do_sendmsg(). This is used by both the initial
788 * transmission and possible later retransmissions.
789 * All SKB's seen here are completely headerless. It is our
790 * job to build the TCP header, and pass the packet down to
791 * IP so it can do the same plus pass the packet off to the
794 * We are working here with either a clone of the original
795 * SKB, or a fresh unique copy made by the retransmit engine.
797 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
800 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
801 struct inet_sock
*inet
;
803 struct tcp_skb_cb
*tcb
;
804 struct tcp_out_options opts
;
805 unsigned int tcp_options_size
, tcp_header_size
;
806 struct tcp_md5sig_key
*md5
;
810 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
812 /* If congestion control is doing timestamping, we must
813 * take such a timestamp before we potentially clone/copy.
815 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
816 __net_timestamp(skb
);
818 if (likely(clone_it
)) {
819 if (unlikely(skb_cloned(skb
)))
820 skb
= pskb_copy(skb
, gfp_mask
);
822 skb
= skb_clone(skb
, gfp_mask
);
829 tcb
= TCP_SKB_CB(skb
);
830 memset(&opts
, 0, sizeof(opts
));
832 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
833 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
835 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
837 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
839 if (tcp_packets_in_flight(tp
) == 0) {
840 tcp_ca_event(sk
, CA_EVENT_TX_START
);
845 skb_push(skb
, tcp_header_size
);
846 skb_reset_transport_header(skb
);
847 skb_set_owner_w(skb
, sk
);
849 /* Build TCP header and checksum it. */
851 th
->source
= inet
->inet_sport
;
852 th
->dest
= inet
->inet_dport
;
853 th
->seq
= htonl(tcb
->seq
);
854 th
->ack_seq
= htonl(tp
->rcv_nxt
);
855 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
858 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
859 /* RFC1323: The window in SYN & SYN/ACK segments
862 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
864 th
->window
= htons(tcp_select_window(sk
));
869 /* The urg_mode check is necessary during a below snd_una win probe */
870 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
871 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
872 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
874 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
875 th
->urg_ptr
= htons(0xFFFF);
880 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
881 if (likely((tcb
->tcp_flags
& TCPHDR_SYN
) == 0))
882 TCP_ECN_send(sk
, skb
, tcp_header_size
);
884 #ifdef CONFIG_TCP_MD5SIG
885 /* Calculate the MD5 hash, as we have all we need now */
887 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
888 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
893 icsk
->icsk_af_ops
->send_check(sk
, skb
);
895 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
896 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
898 if (skb
->len
!= tcp_header_size
)
899 tcp_event_data_sent(tp
, sk
);
901 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
902 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
903 tcp_skb_pcount(skb
));
905 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, &inet
->cork
.fl
);
906 if (likely(err
<= 0))
909 tcp_enter_cwr(sk
, 1);
911 return net_xmit_eval(err
);
914 /* This routine just queues the buffer for sending.
916 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
917 * otherwise socket can stall.
919 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
921 struct tcp_sock
*tp
= tcp_sk(sk
);
923 /* Advance write_seq and place onto the write_queue. */
924 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
925 skb_header_release(skb
);
926 tcp_add_write_queue_tail(sk
, skb
);
927 sk
->sk_wmem_queued
+= skb
->truesize
;
928 sk_mem_charge(sk
, skb
->truesize
);
931 /* Initialize TSO segments for a packet. */
932 static void tcp_set_skb_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
933 unsigned int mss_now
)
935 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
) ||
936 skb
->ip_summed
== CHECKSUM_NONE
) {
937 /* Avoid the costly divide in the normal
940 skb_shinfo(skb
)->gso_segs
= 1;
941 skb_shinfo(skb
)->gso_size
= 0;
942 skb_shinfo(skb
)->gso_type
= 0;
944 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
945 skb_shinfo(skb
)->gso_size
= mss_now
;
946 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
950 /* When a modification to fackets out becomes necessary, we need to check
951 * skb is counted to fackets_out or not.
953 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
956 struct tcp_sock
*tp
= tcp_sk(sk
);
958 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
961 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
962 tp
->fackets_out
-= decr
;
965 /* Pcount in the middle of the write queue got changed, we need to do various
966 * tweaks to fix counters
968 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
970 struct tcp_sock
*tp
= tcp_sk(sk
);
972 tp
->packets_out
-= decr
;
974 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
975 tp
->sacked_out
-= decr
;
976 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
977 tp
->retrans_out
-= decr
;
978 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
979 tp
->lost_out
-= decr
;
981 /* Reno case is special. Sigh... */
982 if (tcp_is_reno(tp
) && decr
> 0)
983 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
985 tcp_adjust_fackets_out(sk
, skb
, decr
);
987 if (tp
->lost_skb_hint
&&
988 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
989 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
990 tp
->lost_cnt_hint
-= decr
;
992 tcp_verify_left_out(tp
);
995 /* Function to create two new TCP segments. Shrinks the given segment
996 * to the specified size and appends a new segment with the rest of the
997 * packet to the list. This won't be called frequently, I hope.
998 * Remember, these are still headerless SKBs at this point.
1000 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1001 unsigned int mss_now
)
1003 struct tcp_sock
*tp
= tcp_sk(sk
);
1004 struct sk_buff
*buff
;
1005 int nsize
, old_factor
;
1009 if (WARN_ON(len
> skb
->len
))
1012 nsize
= skb_headlen(skb
) - len
;
1016 if (skb_cloned(skb
) &&
1017 skb_is_nonlinear(skb
) &&
1018 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1021 /* Get a new skb... force flag on. */
1022 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
1024 return -ENOMEM
; /* We'll just try again later. */
1026 sk
->sk_wmem_queued
+= buff
->truesize
;
1027 sk_mem_charge(sk
, buff
->truesize
);
1028 nlen
= skb
->len
- len
- nsize
;
1029 buff
->truesize
+= nlen
;
1030 skb
->truesize
-= nlen
;
1032 /* Correct the sequence numbers. */
1033 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1034 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1035 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1037 /* PSH and FIN should only be set in the second packet. */
1038 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1039 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1040 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1041 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1043 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1044 /* Copy and checksum data tail into the new buffer. */
1045 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1046 skb_put(buff
, nsize
),
1051 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1053 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1054 skb_split(skb
, buff
, len
);
1057 buff
->ip_summed
= skb
->ip_summed
;
1059 /* Looks stupid, but our code really uses when of
1060 * skbs, which it never sent before. --ANK
1062 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
1063 buff
->tstamp
= skb
->tstamp
;
1065 old_factor
= tcp_skb_pcount(skb
);
1067 /* Fix up tso_factor for both original and new SKB. */
1068 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1069 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1071 /* If this packet has been sent out already, we must
1072 * adjust the various packet counters.
1074 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1075 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1076 tcp_skb_pcount(buff
);
1079 tcp_adjust_pcount(sk
, skb
, diff
);
1082 /* Link BUFF into the send queue. */
1083 skb_header_release(buff
);
1084 tcp_insert_write_queue_after(skb
, buff
, sk
);
1089 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1090 * eventually). The difference is that pulled data not copied, but
1091 * immediately discarded.
1093 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1097 eat
= min_t(int, len
, skb_headlen(skb
));
1099 __skb_pull(skb
, eat
);
1100 skb
->avail_size
-= eat
;
1107 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
1108 int size
= skb_frag_size(&skb_shinfo(skb
)->frags
[i
]);
1111 skb_frag_unref(skb
, i
);
1114 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
1116 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
1117 skb_frag_size_sub(&skb_shinfo(skb
)->frags
[k
], eat
);
1123 skb_shinfo(skb
)->nr_frags
= k
;
1125 skb_reset_tail_pointer(skb
);
1126 skb
->data_len
-= len
;
1127 skb
->len
= skb
->data_len
;
1130 /* Remove acked data from a packet in the transmit queue. */
1131 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1133 if (skb_cloned(skb
) && pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1136 __pskb_trim_head(skb
, len
);
1138 TCP_SKB_CB(skb
)->seq
+= len
;
1139 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1141 skb
->truesize
-= len
;
1142 sk
->sk_wmem_queued
-= len
;
1143 sk_mem_uncharge(sk
, len
);
1144 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1146 /* Any change of skb->len requires recalculation of tso factor. */
1147 if (tcp_skb_pcount(skb
) > 1)
1148 tcp_set_skb_tso_segs(sk
, skb
, tcp_skb_mss(skb
));
1153 /* Calculate MSS. Not accounting for SACKs here. */
1154 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1156 const struct tcp_sock
*tp
= tcp_sk(sk
);
1157 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1160 /* Calculate base mss without TCP options:
1161 It is MMS_S - sizeof(tcphdr) of rfc1122
1163 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1165 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1166 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1167 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1169 if (dst
&& dst_allfrag(dst
))
1170 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1173 /* Clamp it (mss_clamp does not include tcp options) */
1174 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1175 mss_now
= tp
->rx_opt
.mss_clamp
;
1177 /* Now subtract optional transport overhead */
1178 mss_now
-= icsk
->icsk_ext_hdr_len
;
1180 /* Then reserve room for full set of TCP options and 8 bytes of data */
1184 /* Now subtract TCP options size, not including SACKs */
1185 mss_now
-= tp
->tcp_header_len
- sizeof(struct tcphdr
);
1190 /* Inverse of above */
1191 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1193 const struct tcp_sock
*tp
= tcp_sk(sk
);
1194 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1198 tp
->tcp_header_len
+
1199 icsk
->icsk_ext_hdr_len
+
1200 icsk
->icsk_af_ops
->net_header_len
;
1202 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1203 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1204 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1206 if (dst
&& dst_allfrag(dst
))
1207 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1212 /* MTU probing init per socket */
1213 void tcp_mtup_init(struct sock
*sk
)
1215 struct tcp_sock
*tp
= tcp_sk(sk
);
1216 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1218 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
1219 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1220 icsk
->icsk_af_ops
->net_header_len
;
1221 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
1222 icsk
->icsk_mtup
.probe_size
= 0;
1224 EXPORT_SYMBOL(tcp_mtup_init
);
1226 /* This function synchronize snd mss to current pmtu/exthdr set.
1228 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1229 for TCP options, but includes only bare TCP header.
1231 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1232 It is minimum of user_mss and mss received with SYN.
1233 It also does not include TCP options.
1235 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1237 tp->mss_cache is current effective sending mss, including
1238 all tcp options except for SACKs. It is evaluated,
1239 taking into account current pmtu, but never exceeds
1240 tp->rx_opt.mss_clamp.
1242 NOTE1. rfc1122 clearly states that advertised MSS
1243 DOES NOT include either tcp or ip options.
1245 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1246 are READ ONLY outside this function. --ANK (980731)
1248 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1250 struct tcp_sock
*tp
= tcp_sk(sk
);
1251 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1254 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1255 icsk
->icsk_mtup
.search_high
= pmtu
;
1257 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1258 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1260 /* And store cached results */
1261 icsk
->icsk_pmtu_cookie
= pmtu
;
1262 if (icsk
->icsk_mtup
.enabled
)
1263 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1264 tp
->mss_cache
= mss_now
;
1268 EXPORT_SYMBOL(tcp_sync_mss
);
1270 /* Compute the current effective MSS, taking SACKs and IP options,
1271 * and even PMTU discovery events into account.
1273 unsigned int tcp_current_mss(struct sock
*sk
)
1275 const struct tcp_sock
*tp
= tcp_sk(sk
);
1276 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1278 unsigned int header_len
;
1279 struct tcp_out_options opts
;
1280 struct tcp_md5sig_key
*md5
;
1282 mss_now
= tp
->mss_cache
;
1285 u32 mtu
= dst_mtu(dst
);
1286 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1287 mss_now
= tcp_sync_mss(sk
, mtu
);
1290 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1291 sizeof(struct tcphdr
);
1292 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1293 * some common options. If this is an odd packet (because we have SACK
1294 * blocks etc) then our calculated header_len will be different, and
1295 * we have to adjust mss_now correspondingly */
1296 if (header_len
!= tp
->tcp_header_len
) {
1297 int delta
= (int) header_len
- tp
->tcp_header_len
;
1304 /* Congestion window validation. (RFC2861) */
1305 static void tcp_cwnd_validate(struct sock
*sk
)
1307 struct tcp_sock
*tp
= tcp_sk(sk
);
1309 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1310 /* Network is feed fully. */
1311 tp
->snd_cwnd_used
= 0;
1312 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1314 /* Network starves. */
1315 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1316 tp
->snd_cwnd_used
= tp
->packets_out
;
1318 if (sysctl_tcp_slow_start_after_idle
&&
1319 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1320 tcp_cwnd_application_limited(sk
);
1324 /* Returns the portion of skb which can be sent right away without
1325 * introducing MSS oddities to segment boundaries. In rare cases where
1326 * mss_now != mss_cache, we will request caller to create a small skb
1327 * per input skb which could be mostly avoided here (if desired).
1329 * We explicitly want to create a request for splitting write queue tail
1330 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1331 * thus all the complexity (cwnd_len is always MSS multiple which we
1332 * return whenever allowed by the other factors). Basically we need the
1333 * modulo only when the receiver window alone is the limiting factor or
1334 * when we would be allowed to send the split-due-to-Nagle skb fully.
1336 static unsigned int tcp_mss_split_point(const struct sock
*sk
, const struct sk_buff
*skb
,
1337 unsigned int mss_now
, unsigned int cwnd
)
1339 const struct tcp_sock
*tp
= tcp_sk(sk
);
1340 u32 needed
, window
, cwnd_len
;
1342 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1343 cwnd_len
= mss_now
* cwnd
;
1345 if (likely(cwnd_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1348 needed
= min(skb
->len
, window
);
1350 if (cwnd_len
<= needed
)
1353 return needed
- needed
% mss_now
;
1356 /* Can at least one segment of SKB be sent right now, according to the
1357 * congestion window rules? If so, return how many segments are allowed.
1359 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1360 const struct sk_buff
*skb
)
1362 u32 in_flight
, cwnd
;
1364 /* Don't be strict about the congestion window for the final FIN. */
1365 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1366 tcp_skb_pcount(skb
) == 1)
1369 in_flight
= tcp_packets_in_flight(tp
);
1370 cwnd
= tp
->snd_cwnd
;
1371 if (in_flight
< cwnd
)
1372 return (cwnd
- in_flight
);
1377 /* Initialize TSO state of a skb.
1378 * This must be invoked the first time we consider transmitting
1379 * SKB onto the wire.
1381 static int tcp_init_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
1382 unsigned int mss_now
)
1384 int tso_segs
= tcp_skb_pcount(skb
);
1386 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1387 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1388 tso_segs
= tcp_skb_pcount(skb
);
1393 /* Minshall's variant of the Nagle send check. */
1394 static inline bool tcp_minshall_check(const struct tcp_sock
*tp
)
1396 return after(tp
->snd_sml
, tp
->snd_una
) &&
1397 !after(tp
->snd_sml
, tp
->snd_nxt
);
1400 /* Return false, if packet can be sent now without violation Nagle's rules:
1401 * 1. It is full sized.
1402 * 2. Or it contains FIN. (already checked by caller)
1403 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1404 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1405 * With Minshall's modification: all sent small packets are ACKed.
1407 static inline bool tcp_nagle_check(const struct tcp_sock
*tp
,
1408 const struct sk_buff
*skb
,
1409 unsigned int mss_now
, int nonagle
)
1411 return skb
->len
< mss_now
&&
1412 ((nonagle
& TCP_NAGLE_CORK
) ||
1413 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1416 /* Return true if the Nagle test allows this packet to be
1419 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1420 unsigned int cur_mss
, int nonagle
)
1422 /* Nagle rule does not apply to frames, which sit in the middle of the
1423 * write_queue (they have no chances to get new data).
1425 * This is implemented in the callers, where they modify the 'nonagle'
1426 * argument based upon the location of SKB in the send queue.
1428 if (nonagle
& TCP_NAGLE_PUSH
)
1431 /* Don't use the nagle rule for urgent data (or for the final FIN).
1432 * Nagle can be ignored during F-RTO too (see RFC4138).
1434 if (tcp_urg_mode(tp
) || (tp
->frto_counter
== 2) ||
1435 (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1438 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1444 /* Does at least the first segment of SKB fit into the send window? */
1445 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1446 const struct sk_buff
*skb
,
1447 unsigned int cur_mss
)
1449 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1451 if (skb
->len
> cur_mss
)
1452 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1454 return !after(end_seq
, tcp_wnd_end(tp
));
1457 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1458 * should be put on the wire right now. If so, it returns the number of
1459 * packets allowed by the congestion window.
1461 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1462 unsigned int cur_mss
, int nonagle
)
1464 const struct tcp_sock
*tp
= tcp_sk(sk
);
1465 unsigned int cwnd_quota
;
1467 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1469 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1472 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1473 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1479 /* Test if sending is allowed right now. */
1480 bool tcp_may_send_now(struct sock
*sk
)
1482 const struct tcp_sock
*tp
= tcp_sk(sk
);
1483 struct sk_buff
*skb
= tcp_send_head(sk
);
1486 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1487 (tcp_skb_is_last(sk
, skb
) ?
1488 tp
->nonagle
: TCP_NAGLE_PUSH
));
1491 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1492 * which is put after SKB on the list. It is very much like
1493 * tcp_fragment() except that it may make several kinds of assumptions
1494 * in order to speed up the splitting operation. In particular, we
1495 * know that all the data is in scatter-gather pages, and that the
1496 * packet has never been sent out before (and thus is not cloned).
1498 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1499 unsigned int mss_now
, gfp_t gfp
)
1501 struct sk_buff
*buff
;
1502 int nlen
= skb
->len
- len
;
1505 /* All of a TSO frame must be composed of paged data. */
1506 if (skb
->len
!= skb
->data_len
)
1507 return tcp_fragment(sk
, skb
, len
, mss_now
);
1509 buff
= sk_stream_alloc_skb(sk
, 0, gfp
);
1510 if (unlikely(buff
== NULL
))
1513 sk
->sk_wmem_queued
+= buff
->truesize
;
1514 sk_mem_charge(sk
, buff
->truesize
);
1515 buff
->truesize
+= nlen
;
1516 skb
->truesize
-= nlen
;
1518 /* Correct the sequence numbers. */
1519 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1520 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1521 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1523 /* PSH and FIN should only be set in the second packet. */
1524 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1525 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1526 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1528 /* This packet was never sent out yet, so no SACK bits. */
1529 TCP_SKB_CB(buff
)->sacked
= 0;
1531 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1532 skb_split(skb
, buff
, len
);
1534 /* Fix up tso_factor for both original and new SKB. */
1535 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1536 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1538 /* Link BUFF into the send queue. */
1539 skb_header_release(buff
);
1540 tcp_insert_write_queue_after(skb
, buff
, sk
);
1545 /* Try to defer sending, if possible, in order to minimize the amount
1546 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1548 * This algorithm is from John Heffner.
1550 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1552 struct tcp_sock
*tp
= tcp_sk(sk
);
1553 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1554 u32 send_win
, cong_win
, limit
, in_flight
;
1557 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1560 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1563 /* Defer for less than two clock ticks. */
1564 if (tp
->tso_deferred
&&
1565 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1568 in_flight
= tcp_packets_in_flight(tp
);
1570 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1572 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1574 /* From in_flight test above, we know that cwnd > in_flight. */
1575 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1577 limit
= min(send_win
, cong_win
);
1579 /* If a full-sized TSO skb can be sent, do it. */
1580 if (limit
>= sk
->sk_gso_max_size
)
1583 /* Middle in queue won't get any more data, full sendable already? */
1584 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1587 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1589 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1591 /* If at least some fraction of a window is available,
1594 chunk
/= win_divisor
;
1598 /* Different approach, try not to defer past a single
1599 * ACK. Receiver should ACK every other full sized
1600 * frame, so if we have space for more than 3 frames
1603 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1607 /* Ok, it looks like it is advisable to defer. */
1608 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1613 tp
->tso_deferred
= 0;
1617 /* Create a new MTU probe if we are ready.
1618 * MTU probe is regularly attempting to increase the path MTU by
1619 * deliberately sending larger packets. This discovers routing
1620 * changes resulting in larger path MTUs.
1622 * Returns 0 if we should wait to probe (no cwnd available),
1623 * 1 if a probe was sent,
1626 static int tcp_mtu_probe(struct sock
*sk
)
1628 struct tcp_sock
*tp
= tcp_sk(sk
);
1629 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1630 struct sk_buff
*skb
, *nskb
, *next
;
1637 /* Not currently probing/verifying,
1639 * have enough cwnd, and
1640 * not SACKing (the variable headers throw things off) */
1641 if (!icsk
->icsk_mtup
.enabled
||
1642 icsk
->icsk_mtup
.probe_size
||
1643 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1644 tp
->snd_cwnd
< 11 ||
1645 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1648 /* Very simple search strategy: just double the MSS. */
1649 mss_now
= tcp_current_mss(sk
);
1650 probe_size
= 2 * tp
->mss_cache
;
1651 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1652 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1653 /* TODO: set timer for probe_converge_event */
1657 /* Have enough data in the send queue to probe? */
1658 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1661 if (tp
->snd_wnd
< size_needed
)
1663 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1666 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1667 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1668 if (!tcp_packets_in_flight(tp
))
1674 /* We're allowed to probe. Build it now. */
1675 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1677 sk
->sk_wmem_queued
+= nskb
->truesize
;
1678 sk_mem_charge(sk
, nskb
->truesize
);
1680 skb
= tcp_send_head(sk
);
1682 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1683 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1684 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
1685 TCP_SKB_CB(nskb
)->sacked
= 0;
1687 nskb
->ip_summed
= skb
->ip_summed
;
1689 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1692 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1693 copy
= min_t(int, skb
->len
, probe_size
- len
);
1694 if (nskb
->ip_summed
)
1695 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1697 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1698 skb_put(nskb
, copy
),
1701 if (skb
->len
<= copy
) {
1702 /* We've eaten all the data from this skb.
1704 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1705 tcp_unlink_write_queue(skb
, sk
);
1706 sk_wmem_free_skb(sk
, skb
);
1708 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
1709 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1710 if (!skb_shinfo(skb
)->nr_frags
) {
1711 skb_pull(skb
, copy
);
1712 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1713 skb
->csum
= csum_partial(skb
->data
,
1716 __pskb_trim_head(skb
, copy
);
1717 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1719 TCP_SKB_CB(skb
)->seq
+= copy
;
1724 if (len
>= probe_size
)
1727 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1729 /* We're ready to send. If this fails, the probe will
1730 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1731 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1732 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1733 /* Decrement cwnd here because we are sending
1734 * effectively two packets. */
1736 tcp_event_new_data_sent(sk
, nskb
);
1738 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1739 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1740 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1748 /* This routine writes packets to the network. It advances the
1749 * send_head. This happens as incoming acks open up the remote
1752 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1753 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1754 * account rare use of URG, this is not a big flaw.
1756 * Returns true, if no segments are in flight and we have queued segments,
1757 * but cannot send anything now because of SWS or another problem.
1759 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1760 int push_one
, gfp_t gfp
)
1762 struct tcp_sock
*tp
= tcp_sk(sk
);
1763 struct sk_buff
*skb
;
1764 unsigned int tso_segs
, sent_pkts
;
1771 /* Do MTU probing. */
1772 result
= tcp_mtu_probe(sk
);
1775 } else if (result
> 0) {
1780 while ((skb
= tcp_send_head(sk
))) {
1783 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1786 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1790 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1793 if (tso_segs
== 1) {
1794 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1795 (tcp_skb_is_last(sk
, skb
) ?
1796 nonagle
: TCP_NAGLE_PUSH
))))
1799 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
1804 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
1805 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1808 if (skb
->len
> limit
&&
1809 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
1812 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1814 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
1817 /* Advance the send_head. This one is sent out.
1818 * This call will increment packets_out.
1820 tcp_event_new_data_sent(sk
, skb
);
1822 tcp_minshall_update(tp
, mss_now
, skb
);
1823 sent_pkts
+= tcp_skb_pcount(skb
);
1828 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Recovery
)
1829 tp
->prr_out
+= sent_pkts
;
1831 if (likely(sent_pkts
)) {
1832 tcp_cwnd_validate(sk
);
1835 return !tp
->packets_out
&& tcp_send_head(sk
);
1838 /* Push out any pending frames which were held back due to
1839 * TCP_CORK or attempt at coalescing tiny packets.
1840 * The socket must be locked by the caller.
1842 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
1845 /* If we are closed, the bytes will have to remain here.
1846 * In time closedown will finish, we empty the write queue and
1847 * all will be happy.
1849 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
1852 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0, GFP_ATOMIC
))
1853 tcp_check_probe_timer(sk
);
1856 /* Send _single_ skb sitting at the send head. This function requires
1857 * true push pending frames to setup probe timer etc.
1859 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
1861 struct sk_buff
*skb
= tcp_send_head(sk
);
1863 BUG_ON(!skb
|| skb
->len
< mss_now
);
1865 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
1868 /* This function returns the amount that we can raise the
1869 * usable window based on the following constraints
1871 * 1. The window can never be shrunk once it is offered (RFC 793)
1872 * 2. We limit memory per socket
1875 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1876 * RECV.NEXT + RCV.WIN fixed until:
1877 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1879 * i.e. don't raise the right edge of the window until you can raise
1880 * it at least MSS bytes.
1882 * Unfortunately, the recommended algorithm breaks header prediction,
1883 * since header prediction assumes th->window stays fixed.
1885 * Strictly speaking, keeping th->window fixed violates the receiver
1886 * side SWS prevention criteria. The problem is that under this rule
1887 * a stream of single byte packets will cause the right side of the
1888 * window to always advance by a single byte.
1890 * Of course, if the sender implements sender side SWS prevention
1891 * then this will not be a problem.
1893 * BSD seems to make the following compromise:
1895 * If the free space is less than the 1/4 of the maximum
1896 * space available and the free space is less than 1/2 mss,
1897 * then set the window to 0.
1898 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1899 * Otherwise, just prevent the window from shrinking
1900 * and from being larger than the largest representable value.
1902 * This prevents incremental opening of the window in the regime
1903 * where TCP is limited by the speed of the reader side taking
1904 * data out of the TCP receive queue. It does nothing about
1905 * those cases where the window is constrained on the sender side
1906 * because the pipeline is full.
1908 * BSD also seems to "accidentally" limit itself to windows that are a
1909 * multiple of MSS, at least until the free space gets quite small.
1910 * This would appear to be a side effect of the mbuf implementation.
1911 * Combining these two algorithms results in the observed behavior
1912 * of having a fixed window size at almost all times.
1914 * Below we obtain similar behavior by forcing the offered window to
1915 * a multiple of the mss when it is feasible to do so.
1917 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1918 * Regular options like TIMESTAMP are taken into account.
1920 u32
__tcp_select_window(struct sock
*sk
)
1922 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1923 struct tcp_sock
*tp
= tcp_sk(sk
);
1924 /* MSS for the peer's data. Previous versions used mss_clamp
1925 * here. I don't know if the value based on our guesses
1926 * of peer's MSS is better for the performance. It's more correct
1927 * but may be worse for the performance because of rcv_mss
1928 * fluctuations. --SAW 1998/11/1
1930 int mss
= icsk
->icsk_ack
.rcv_mss
;
1931 int free_space
= tcp_space(sk
);
1932 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
1935 if (mss
> full_space
)
1938 if (free_space
< (full_space
>> 1)) {
1939 icsk
->icsk_ack
.quick
= 0;
1941 if (sk_under_memory_pressure(sk
))
1942 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
1945 if (free_space
< mss
)
1949 if (free_space
> tp
->rcv_ssthresh
)
1950 free_space
= tp
->rcv_ssthresh
;
1952 /* Don't do rounding if we are using window scaling, since the
1953 * scaled window will not line up with the MSS boundary anyway.
1955 window
= tp
->rcv_wnd
;
1956 if (tp
->rx_opt
.rcv_wscale
) {
1957 window
= free_space
;
1959 /* Advertise enough space so that it won't get scaled away.
1960 * Import case: prevent zero window announcement if
1961 * 1<<rcv_wscale > mss.
1963 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
1964 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
1965 << tp
->rx_opt
.rcv_wscale
);
1967 /* Get the largest window that is a nice multiple of mss.
1968 * Window clamp already applied above.
1969 * If our current window offering is within 1 mss of the
1970 * free space we just keep it. This prevents the divide
1971 * and multiply from happening most of the time.
1972 * We also don't do any window rounding when the free space
1975 if (window
<= free_space
- mss
|| window
> free_space
)
1976 window
= (free_space
/ mss
) * mss
;
1977 else if (mss
== full_space
&&
1978 free_space
> window
+ (full_space
>> 1))
1979 window
= free_space
;
1985 /* Collapses two adjacent SKB's during retransmission. */
1986 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
1988 struct tcp_sock
*tp
= tcp_sk(sk
);
1989 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
1990 int skb_size
, next_skb_size
;
1992 skb_size
= skb
->len
;
1993 next_skb_size
= next_skb
->len
;
1995 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
1997 tcp_highest_sack_combine(sk
, next_skb
, skb
);
1999 tcp_unlink_write_queue(next_skb
, sk
);
2001 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
2004 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2005 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2007 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2008 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2010 /* Update sequence range on original skb. */
2011 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2013 /* Merge over control information. This moves PSH/FIN etc. over */
2014 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2016 /* All done, get rid of second SKB and account for it so
2017 * packet counting does not break.
2019 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2021 /* changed transmit queue under us so clear hints */
2022 tcp_clear_retrans_hints_partial(tp
);
2023 if (next_skb
== tp
->retransmit_skb_hint
)
2024 tp
->retransmit_skb_hint
= skb
;
2026 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2028 sk_wmem_free_skb(sk
, next_skb
);
2031 /* Check if coalescing SKBs is legal. */
2032 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2034 if (tcp_skb_pcount(skb
) > 1)
2036 /* TODO: SACK collapsing could be used to remove this condition */
2037 if (skb_shinfo(skb
)->nr_frags
!= 0)
2039 if (skb_cloned(skb
))
2041 if (skb
== tcp_send_head(sk
))
2043 /* Some heurestics for collapsing over SACK'd could be invented */
2044 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2050 /* Collapse packets in the retransmit queue to make to create
2051 * less packets on the wire. This is only done on retransmission.
2053 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2056 struct tcp_sock
*tp
= tcp_sk(sk
);
2057 struct sk_buff
*skb
= to
, *tmp
;
2060 if (!sysctl_tcp_retrans_collapse
)
2062 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2065 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2066 if (!tcp_can_collapse(sk
, skb
))
2078 /* Punt if not enough space exists in the first SKB for
2079 * the data in the second
2081 if (skb
->len
> skb_availroom(to
))
2084 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2087 tcp_collapse_retrans(sk
, to
);
2091 /* This retransmits one SKB. Policy decisions and retransmit queue
2092 * state updates are done by the caller. Returns non-zero if an
2093 * error occurred which prevented the send.
2095 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2097 struct tcp_sock
*tp
= tcp_sk(sk
);
2098 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2099 unsigned int cur_mss
;
2102 /* Inconslusive MTU probe */
2103 if (icsk
->icsk_mtup
.probe_size
) {
2104 icsk
->icsk_mtup
.probe_size
= 0;
2107 /* Do not sent more than we queued. 1/4 is reserved for possible
2108 * copying overhead: fragmentation, tunneling, mangling etc.
2110 if (atomic_read(&sk
->sk_wmem_alloc
) >
2111 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2114 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2115 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2117 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2121 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2122 return -EHOSTUNREACH
; /* Routing failure or similar. */
2124 cur_mss
= tcp_current_mss(sk
);
2126 /* If receiver has shrunk his window, and skb is out of
2127 * new window, do not retransmit it. The exception is the
2128 * case, when window is shrunk to zero. In this case
2129 * our retransmit serves as a zero window probe.
2131 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2132 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2135 if (skb
->len
> cur_mss
) {
2136 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
2137 return -ENOMEM
; /* We'll try again later. */
2139 int oldpcount
= tcp_skb_pcount(skb
);
2141 if (unlikely(oldpcount
> 1)) {
2142 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2143 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2147 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2149 /* Some Solaris stacks overoptimize and ignore the FIN on a
2150 * retransmit when old data is attached. So strip it off
2151 * since it is cheap to do so and saves bytes on the network.
2154 (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
2155 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
2156 if (!pskb_trim(skb
, 0)) {
2157 /* Reuse, even though it does some unnecessary work */
2158 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
2159 TCP_SKB_CB(skb
)->tcp_flags
);
2160 skb
->ip_summed
= CHECKSUM_NONE
;
2164 /* Make a copy, if the first transmission SKB clone we made
2165 * is still in somebody's hands, else make a clone.
2167 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2169 /* make sure skb->data is aligned on arches that require it */
2170 if (unlikely(NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3))) {
2171 struct sk_buff
*nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
,
2173 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2176 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2180 /* Update global TCP statistics. */
2181 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2183 tp
->total_retrans
++;
2185 #if FASTRETRANS_DEBUG > 0
2186 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2187 net_dbg_ratelimited("retrans_out leaked\n");
2190 if (!tp
->retrans_out
)
2191 tp
->lost_retrans_low
= tp
->snd_nxt
;
2192 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2193 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2195 /* Save stamp of the first retransmit. */
2196 if (!tp
->retrans_stamp
)
2197 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2199 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2201 /* snd_nxt is stored to detect loss of retransmitted segment,
2202 * see tcp_input.c tcp_sacktag_write_queue().
2204 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2209 /* Check if we forward retransmits are possible in the current
2210 * window/congestion state.
2212 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2214 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2215 const struct tcp_sock
*tp
= tcp_sk(sk
);
2217 /* Forward retransmissions are possible only during Recovery. */
2218 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2221 /* No forward retransmissions in Reno are possible. */
2222 if (tcp_is_reno(tp
))
2225 /* Yeah, we have to make difficult choice between forward transmission
2226 * and retransmission... Both ways have their merits...
2228 * For now we do not retransmit anything, while we have some new
2229 * segments to send. In the other cases, follow rule 3 for
2230 * NextSeg() specified in RFC3517.
2233 if (tcp_may_send_now(sk
))
2239 /* This gets called after a retransmit timeout, and the initially
2240 * retransmitted data is acknowledged. It tries to continue
2241 * resending the rest of the retransmit queue, until either
2242 * we've sent it all or the congestion window limit is reached.
2243 * If doing SACK, the first ACK which comes back for a timeout
2244 * based retransmit packet might feed us FACK information again.
2245 * If so, we use it to avoid unnecessarily retransmissions.
2247 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2249 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2250 struct tcp_sock
*tp
= tcp_sk(sk
);
2251 struct sk_buff
*skb
;
2252 struct sk_buff
*hole
= NULL
;
2255 int fwd_rexmitting
= 0;
2257 if (!tp
->packets_out
)
2261 tp
->retransmit_high
= tp
->snd_una
;
2263 if (tp
->retransmit_skb_hint
) {
2264 skb
= tp
->retransmit_skb_hint
;
2265 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2266 if (after(last_lost
, tp
->retransmit_high
))
2267 last_lost
= tp
->retransmit_high
;
2269 skb
= tcp_write_queue_head(sk
);
2270 last_lost
= tp
->snd_una
;
2273 tcp_for_write_queue_from(skb
, sk
) {
2274 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2276 if (skb
== tcp_send_head(sk
))
2278 /* we could do better than to assign each time */
2280 tp
->retransmit_skb_hint
= skb
;
2282 /* Assume this retransmit will generate
2283 * only one packet for congestion window
2284 * calculation purposes. This works because
2285 * tcp_retransmit_skb() will chop up the
2286 * packet to be MSS sized and all the
2287 * packet counting works out.
2289 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2292 if (fwd_rexmitting
) {
2294 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2296 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2298 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2299 tp
->retransmit_high
= last_lost
;
2300 if (!tcp_can_forward_retransmit(sk
))
2302 /* Backtrack if necessary to non-L'ed skb */
2310 } else if (!(sacked
& TCPCB_LOST
)) {
2311 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2316 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2317 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2318 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2320 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2323 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2326 if (tcp_retransmit_skb(sk
, skb
)) {
2327 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2330 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2332 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Recovery
)
2333 tp
->prr_out
+= tcp_skb_pcount(skb
);
2335 if (skb
== tcp_write_queue_head(sk
))
2336 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2337 inet_csk(sk
)->icsk_rto
,
2342 /* Send a fin. The caller locks the socket for us. This cannot be
2343 * allowed to fail queueing a FIN frame under any circumstances.
2345 void tcp_send_fin(struct sock
*sk
)
2347 struct tcp_sock
*tp
= tcp_sk(sk
);
2348 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2351 /* Optimization, tack on the FIN if we have a queue of
2352 * unsent frames. But be careful about outgoing SACKS
2355 mss_now
= tcp_current_mss(sk
);
2357 if (tcp_send_head(sk
) != NULL
) {
2358 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_FIN
;
2359 TCP_SKB_CB(skb
)->end_seq
++;
2362 /* Socket is locked, keep trying until memory is available. */
2364 skb
= alloc_skb_fclone(MAX_TCP_HEADER
,
2371 /* Reserve space for headers and prepare control bits. */
2372 skb_reserve(skb
, MAX_TCP_HEADER
);
2373 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2374 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2375 TCPHDR_ACK
| TCPHDR_FIN
);
2376 tcp_queue_skb(sk
, skb
);
2378 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2381 /* We get here when a process closes a file descriptor (either due to
2382 * an explicit close() or as a byproduct of exit()'ing) and there
2383 * was unread data in the receive queue. This behavior is recommended
2384 * by RFC 2525, section 2.17. -DaveM
2386 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2388 struct sk_buff
*skb
;
2390 /* NOTE: No TCP options attached and we never retransmit this. */
2391 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2393 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2397 /* Reserve space for headers and prepare control bits. */
2398 skb_reserve(skb
, MAX_TCP_HEADER
);
2399 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2400 TCPHDR_ACK
| TCPHDR_RST
);
2402 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2403 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2404 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2406 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2409 /* Send a crossed SYN-ACK during socket establishment.
2410 * WARNING: This routine must only be called when we have already sent
2411 * a SYN packet that crossed the incoming SYN that caused this routine
2412 * to get called. If this assumption fails then the initial rcv_wnd
2413 * and rcv_wscale values will not be correct.
2415 int tcp_send_synack(struct sock
*sk
)
2417 struct sk_buff
*skb
;
2419 skb
= tcp_write_queue_head(sk
);
2420 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2421 pr_debug("%s: wrong queue state\n", __func__
);
2424 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
2425 if (skb_cloned(skb
)) {
2426 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2429 tcp_unlink_write_queue(skb
, sk
);
2430 skb_header_release(nskb
);
2431 __tcp_add_write_queue_head(sk
, nskb
);
2432 sk_wmem_free_skb(sk
, skb
);
2433 sk
->sk_wmem_queued
+= nskb
->truesize
;
2434 sk_mem_charge(sk
, nskb
->truesize
);
2438 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
2439 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2441 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2442 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2445 /* Prepare a SYN-ACK. */
2446 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2447 struct request_sock
*req
,
2448 struct request_values
*rvp
)
2450 struct tcp_out_options opts
;
2451 struct tcp_extend_values
*xvp
= tcp_xv(rvp
);
2452 struct inet_request_sock
*ireq
= inet_rsk(req
);
2453 struct tcp_sock
*tp
= tcp_sk(sk
);
2454 const struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
2456 struct sk_buff
*skb
;
2457 struct tcp_md5sig_key
*md5
;
2458 int tcp_header_size
;
2460 int s_data_desired
= 0;
2462 if (cvp
!= NULL
&& cvp
->s_data_constant
&& cvp
->s_data_desired
)
2463 s_data_desired
= cvp
->s_data_desired
;
2464 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15 + s_data_desired
, 1, GFP_ATOMIC
);
2468 /* Reserve space for headers. */
2469 skb_reserve(skb
, MAX_TCP_HEADER
);
2471 skb_dst_set(skb
, dst_clone(dst
));
2473 mss
= dst_metric_advmss(dst
);
2474 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2475 mss
= tp
->rx_opt
.user_mss
;
2477 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2479 /* Set this up on the first call only */
2480 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2482 /* limit the window selection if the user enforce a smaller rx buffer */
2483 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2484 (req
->window_clamp
> tcp_full_space(sk
) || req
->window_clamp
== 0))
2485 req
->window_clamp
= tcp_full_space(sk
);
2487 /* tcp_full_space because it is guaranteed to be the first packet */
2488 tcp_select_initial_window(tcp_full_space(sk
),
2489 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2494 dst_metric(dst
, RTAX_INITRWND
));
2495 ireq
->rcv_wscale
= rcv_wscale
;
2498 memset(&opts
, 0, sizeof(opts
));
2499 #ifdef CONFIG_SYN_COOKIES
2500 if (unlikely(req
->cookie_ts
))
2501 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2504 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2505 tcp_header_size
= tcp_synack_options(sk
, req
, mss
,
2506 skb
, &opts
, &md5
, xvp
)
2509 skb_push(skb
, tcp_header_size
);
2510 skb_reset_transport_header(skb
);
2513 memset(th
, 0, sizeof(struct tcphdr
));
2516 TCP_ECN_make_synack(req
, th
);
2517 th
->source
= ireq
->loc_port
;
2518 th
->dest
= ireq
->rmt_port
;
2519 /* Setting of flags are superfluous here for callers (and ECE is
2520 * not even correctly set)
2522 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2523 TCPHDR_SYN
| TCPHDR_ACK
);
2525 if (OPTION_COOKIE_EXTENSION
& opts
.options
) {
2526 if (s_data_desired
) {
2527 u8
*buf
= skb_put(skb
, s_data_desired
);
2529 /* copy data directly from the listening socket. */
2530 memcpy(buf
, cvp
->s_data_payload
, s_data_desired
);
2531 TCP_SKB_CB(skb
)->end_seq
+= s_data_desired
;
2534 if (opts
.hash_size
> 0) {
2535 __u32 workspace
[SHA_WORKSPACE_WORDS
];
2536 u32
*mess
= &xvp
->cookie_bakery
[COOKIE_DIGEST_WORDS
];
2537 u32
*tail
= &mess
[COOKIE_MESSAGE_WORDS
-1];
2539 /* Secret recipe depends on the Timestamp, (future)
2540 * Sequence and Acknowledgment Numbers, Initiator
2541 * Cookie, and others handled by IP variant caller.
2543 *tail
-- ^= opts
.tsval
;
2544 *tail
-- ^= tcp_rsk(req
)->rcv_isn
+ 1;
2545 *tail
-- ^= TCP_SKB_CB(skb
)->seq
+ 1;
2548 *tail
-- ^= (((__force u32
)th
->dest
<< 16) | (__force u32
)th
->source
);
2549 *tail
-- ^= (u32
)(unsigned long)cvp
; /* per sockopt */
2551 sha_transform((__u32
*)&xvp
->cookie_bakery
[0],
2554 opts
.hash_location
=
2555 (__u8
*)&xvp
->cookie_bakery
[0];
2559 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2560 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_isn
+ 1);
2562 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2563 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2564 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2565 th
->doff
= (tcp_header_size
>> 2);
2566 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
, tcp_skb_pcount(skb
));
2568 #ifdef CONFIG_TCP_MD5SIG
2569 /* Okay, we have all we need - do the md5 hash if needed */
2571 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2572 md5
, NULL
, req
, skb
);
2578 EXPORT_SYMBOL(tcp_make_synack
);
2580 /* Do all connect socket setups that can be done AF independent. */
2581 void tcp_connect_init(struct sock
*sk
)
2583 const struct dst_entry
*dst
= __sk_dst_get(sk
);
2584 struct tcp_sock
*tp
= tcp_sk(sk
);
2587 /* We'll fix this up when we get a response from the other end.
2588 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2590 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2591 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2593 #ifdef CONFIG_TCP_MD5SIG
2594 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2595 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2598 /* If user gave his TCP_MAXSEG, record it to clamp */
2599 if (tp
->rx_opt
.user_mss
)
2600 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2603 tcp_sync_mss(sk
, dst_mtu(dst
));
2605 if (!tp
->window_clamp
)
2606 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2607 tp
->advmss
= dst_metric_advmss(dst
);
2608 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2609 tp
->advmss
= tp
->rx_opt
.user_mss
;
2611 tcp_initialize_rcv_mss(sk
);
2613 /* limit the window selection if the user enforce a smaller rx buffer */
2614 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2615 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
2616 tp
->window_clamp
= tcp_full_space(sk
);
2618 tcp_select_initial_window(tcp_full_space(sk
),
2619 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2622 sysctl_tcp_window_scaling
,
2624 dst_metric(dst
, RTAX_INITRWND
));
2626 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2627 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2630 sock_reset_flag(sk
, SOCK_DONE
);
2633 tp
->snd_una
= tp
->write_seq
;
2634 tp
->snd_sml
= tp
->write_seq
;
2635 tp
->snd_up
= tp
->write_seq
;
2636 tp
->snd_nxt
= tp
->write_seq
;
2638 if (likely(!tp
->repair
))
2640 tp
->rcv_wup
= tp
->rcv_nxt
;
2641 tp
->copied_seq
= tp
->rcv_nxt
;
2643 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2644 inet_csk(sk
)->icsk_retransmits
= 0;
2645 tcp_clear_retrans(tp
);
2648 /* Build a SYN and send it off. */
2649 int tcp_connect(struct sock
*sk
)
2651 struct tcp_sock
*tp
= tcp_sk(sk
);
2652 struct sk_buff
*buff
;
2655 tcp_connect_init(sk
);
2657 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2658 if (unlikely(buff
== NULL
))
2661 /* Reserve space for headers. */
2662 skb_reserve(buff
, MAX_TCP_HEADER
);
2664 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
2665 TCP_ECN_send_syn(sk
, buff
);
2668 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2669 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
;
2670 skb_header_release(buff
);
2671 __tcp_add_write_queue_tail(sk
, buff
);
2672 sk
->sk_wmem_queued
+= buff
->truesize
;
2673 sk_mem_charge(sk
, buff
->truesize
);
2674 tp
->packets_out
+= tcp_skb_pcount(buff
);
2675 err
= tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
2676 if (err
== -ECONNREFUSED
)
2679 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2680 * in order to make this packet get counted in tcpOutSegs.
2682 tp
->snd_nxt
= tp
->write_seq
;
2683 tp
->pushed_seq
= tp
->write_seq
;
2684 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2686 /* Timer for repeating the SYN until an answer. */
2687 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2688 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2691 EXPORT_SYMBOL(tcp_connect
);
2693 /* Send out a delayed ack, the caller does the policy checking
2694 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2697 void tcp_send_delayed_ack(struct sock
*sk
)
2699 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2700 int ato
= icsk
->icsk_ack
.ato
;
2701 unsigned long timeout
;
2703 if (ato
> TCP_DELACK_MIN
) {
2704 const struct tcp_sock
*tp
= tcp_sk(sk
);
2705 int max_ato
= HZ
/ 2;
2707 if (icsk
->icsk_ack
.pingpong
||
2708 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2709 max_ato
= TCP_DELACK_MAX
;
2711 /* Slow path, intersegment interval is "high". */
2713 /* If some rtt estimate is known, use it to bound delayed ack.
2714 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2718 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
2724 ato
= min(ato
, max_ato
);
2727 /* Stay within the limit we were given */
2728 timeout
= jiffies
+ ato
;
2730 /* Use new timeout only if there wasn't a older one earlier. */
2731 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
2732 /* If delack timer was blocked or is about to expire,
2735 if (icsk
->icsk_ack
.blocked
||
2736 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
2741 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
2742 timeout
= icsk
->icsk_ack
.timeout
;
2744 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
2745 icsk
->icsk_ack
.timeout
= timeout
;
2746 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
2749 /* This routine sends an ack and also updates the window. */
2750 void tcp_send_ack(struct sock
*sk
)
2752 struct sk_buff
*buff
;
2754 /* If we have been reset, we may not send again. */
2755 if (sk
->sk_state
== TCP_CLOSE
)
2758 /* We are not putting this on the write queue, so
2759 * tcp_transmit_skb() will set the ownership to this
2762 buff
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2764 inet_csk_schedule_ack(sk
);
2765 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
2766 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
2767 TCP_DELACK_MAX
, TCP_RTO_MAX
);
2771 /* Reserve space for headers and prepare control bits. */
2772 skb_reserve(buff
, MAX_TCP_HEADER
);
2773 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
2775 /* Send it off, this clears delayed acks for us. */
2776 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2777 tcp_transmit_skb(sk
, buff
, 0, GFP_ATOMIC
);
2780 /* This routine sends a packet with an out of date sequence
2781 * number. It assumes the other end will try to ack it.
2783 * Question: what should we make while urgent mode?
2784 * 4.4BSD forces sending single byte of data. We cannot send
2785 * out of window data, because we have SND.NXT==SND.MAX...
2787 * Current solution: to send TWO zero-length segments in urgent mode:
2788 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2789 * out-of-date with SND.UNA-1 to probe window.
2791 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
2793 struct tcp_sock
*tp
= tcp_sk(sk
);
2794 struct sk_buff
*skb
;
2796 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2797 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2801 /* Reserve space for headers and set control bits. */
2802 skb_reserve(skb
, MAX_TCP_HEADER
);
2803 /* Use a previous sequence. This should cause the other
2804 * end to send an ack. Don't queue or clone SKB, just
2807 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
2808 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2809 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
2812 void tcp_send_window_probe(struct sock
*sk
)
2814 if (sk
->sk_state
== TCP_ESTABLISHED
) {
2815 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
2816 tcp_sk(sk
)->snd_nxt
= tcp_sk(sk
)->write_seq
;
2817 tcp_xmit_probe_skb(sk
, 0);
2821 /* Initiate keepalive or window probe from timer. */
2822 int tcp_write_wakeup(struct sock
*sk
)
2824 struct tcp_sock
*tp
= tcp_sk(sk
);
2825 struct sk_buff
*skb
;
2827 if (sk
->sk_state
== TCP_CLOSE
)
2830 if ((skb
= tcp_send_head(sk
)) != NULL
&&
2831 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
2833 unsigned int mss
= tcp_current_mss(sk
);
2834 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2836 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
2837 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
2839 /* We are probing the opening of a window
2840 * but the window size is != 0
2841 * must have been a result SWS avoidance ( sender )
2843 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
2845 seg_size
= min(seg_size
, mss
);
2846 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
2847 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
2849 } else if (!tcp_skb_pcount(skb
))
2850 tcp_set_skb_tso_segs(sk
, skb
, mss
);
2852 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
2853 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2854 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2856 tcp_event_new_data_sent(sk
, skb
);
2859 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
2860 tcp_xmit_probe_skb(sk
, 1);
2861 return tcp_xmit_probe_skb(sk
, 0);
2865 /* A window probe timeout has occurred. If window is not closed send
2866 * a partial packet else a zero probe.
2868 void tcp_send_probe0(struct sock
*sk
)
2870 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2871 struct tcp_sock
*tp
= tcp_sk(sk
);
2874 err
= tcp_write_wakeup(sk
);
2876 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
2877 /* Cancel probe timer, if it is not required. */
2878 icsk
->icsk_probes_out
= 0;
2879 icsk
->icsk_backoff
= 0;
2884 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
2885 icsk
->icsk_backoff
++;
2886 icsk
->icsk_probes_out
++;
2887 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2888 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
2891 /* If packet was not sent due to local congestion,
2892 * do not backoff and do not remember icsk_probes_out.
2893 * Let local senders to fight for local resources.
2895 * Use accumulated backoff yet.
2897 if (!icsk
->icsk_probes_out
)
2898 icsk
->icsk_probes_out
= 1;
2899 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
2900 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
2901 TCP_RESOURCE_PROBE_INTERVAL
),