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 /* Default TSQ limit of two TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly
= 131072;
56 /* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
60 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
62 int sysctl_tcp_mtu_probing __read_mostly
= 0;
63 int sysctl_tcp_base_mss __read_mostly
= TCP_BASE_MSS
;
65 /* By default, RFC2861 behavior. */
66 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
68 int sysctl_tcp_cookie_size __read_mostly
= 0; /* TCP_COOKIE_MAX */
69 EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size
);
71 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
72 int push_one
, gfp_t gfp
);
74 /* Account for new data that has been sent to the network. */
75 static void tcp_event_new_data_sent(struct sock
*sk
, const struct sk_buff
*skb
)
77 struct tcp_sock
*tp
= tcp_sk(sk
);
78 unsigned int prior_packets
= tp
->packets_out
;
80 tcp_advance_send_head(sk
, skb
);
81 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
83 /* Don't override Nagle indefinitely with F-RTO */
84 if (tp
->frto_counter
== 2)
87 tp
->packets_out
+= tcp_skb_pcount(skb
);
88 if (!prior_packets
|| tp
->early_retrans_delayed
)
92 /* SND.NXT, if window was not shrunk.
93 * If window has been shrunk, what should we make? It is not clear at all.
94 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
95 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
96 * invalid. OK, let's make this for now:
98 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
100 const struct tcp_sock
*tp
= tcp_sk(sk
);
102 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
105 return tcp_wnd_end(tp
);
108 /* Calculate mss to advertise in SYN segment.
109 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
111 * 1. It is independent of path mtu.
112 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
113 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
114 * attached devices, because some buggy hosts are confused by
116 * 4. We do not make 3, we advertise MSS, calculated from first
117 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
118 * This may be overridden via information stored in routing table.
119 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
120 * probably even Jumbo".
122 static __u16
tcp_advertise_mss(struct sock
*sk
)
124 struct tcp_sock
*tp
= tcp_sk(sk
);
125 const struct dst_entry
*dst
= __sk_dst_get(sk
);
126 int mss
= tp
->advmss
;
129 unsigned int metric
= dst_metric_advmss(dst
);
140 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
141 * This is the first part of cwnd validation mechanism. */
142 static void tcp_cwnd_restart(struct sock
*sk
, const struct dst_entry
*dst
)
144 struct tcp_sock
*tp
= tcp_sk(sk
);
145 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
146 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
147 u32 cwnd
= tp
->snd_cwnd
;
149 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
151 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
152 restart_cwnd
= min(restart_cwnd
, cwnd
);
154 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
156 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
157 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
158 tp
->snd_cwnd_used
= 0;
161 /* Congestion state accounting after a packet has been sent. */
162 static void tcp_event_data_sent(struct tcp_sock
*tp
,
165 struct inet_connection_sock
*icsk
= inet_csk(sk
);
166 const u32 now
= tcp_time_stamp
;
168 if (sysctl_tcp_slow_start_after_idle
&&
169 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
170 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
174 /* If it is a reply for ato after last received
175 * packet, enter pingpong mode.
177 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
178 icsk
->icsk_ack
.pingpong
= 1;
181 /* Account for an ACK we sent. */
182 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
184 tcp_dec_quickack_mode(sk
, pkts
);
185 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
188 /* Determine a window scaling and initial window to offer.
189 * Based on the assumption that the given amount of space
190 * will be offered. Store the results in the tp structure.
191 * NOTE: for smooth operation initial space offering should
192 * be a multiple of mss if possible. We assume here that mss >= 1.
193 * This MUST be enforced by all callers.
195 void tcp_select_initial_window(int __space
, __u32 mss
,
196 __u32
*rcv_wnd
, __u32
*window_clamp
,
197 int wscale_ok
, __u8
*rcv_wscale
,
200 unsigned int space
= (__space
< 0 ? 0 : __space
);
202 /* If no clamp set the clamp to the max possible scaled window */
203 if (*window_clamp
== 0)
204 (*window_clamp
) = (65535 << 14);
205 space
= min(*window_clamp
, space
);
207 /* Quantize space offering to a multiple of mss if possible. */
209 space
= (space
/ mss
) * mss
;
211 /* NOTE: offering an initial window larger than 32767
212 * will break some buggy TCP stacks. If the admin tells us
213 * it is likely we could be speaking with such a buggy stack
214 * we will truncate our initial window offering to 32K-1
215 * unless the remote has sent us a window scaling option,
216 * which we interpret as a sign the remote TCP is not
217 * misinterpreting the window field as a signed quantity.
219 if (sysctl_tcp_workaround_signed_windows
)
220 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
226 /* Set window scaling on max possible window
227 * See RFC1323 for an explanation of the limit to 14
229 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
230 space
= min_t(u32
, space
, *window_clamp
);
231 while (space
> 65535 && (*rcv_wscale
) < 14) {
237 /* Set initial window to a value enough for senders starting with
238 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
239 * a limit on the initial window when mss is larger than 1460.
241 if (mss
> (1 << *rcv_wscale
)) {
242 int init_cwnd
= TCP_DEFAULT_INIT_RCVWND
;
245 max_t(u32
, (1460 * TCP_DEFAULT_INIT_RCVWND
) / mss
, 2);
246 /* when initializing use the value from init_rcv_wnd
247 * rather than the default from above
250 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
252 *rcv_wnd
= min(*rcv_wnd
, init_cwnd
* mss
);
255 /* Set the clamp no higher than max representable value */
256 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
258 EXPORT_SYMBOL(tcp_select_initial_window
);
260 /* Chose a new window to advertise, update state in tcp_sock for the
261 * socket, and return result with RFC1323 scaling applied. The return
262 * value can be stuffed directly into th->window for an outgoing
265 static u16
tcp_select_window(struct sock
*sk
)
267 struct tcp_sock
*tp
= tcp_sk(sk
);
268 u32 cur_win
= tcp_receive_window(tp
);
269 u32 new_win
= __tcp_select_window(sk
);
271 /* Never shrink the offered window */
272 if (new_win
< cur_win
) {
273 /* Danger Will Robinson!
274 * Don't update rcv_wup/rcv_wnd here or else
275 * we will not be able to advertise a zero
276 * window in time. --DaveM
278 * Relax Will Robinson.
280 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
282 tp
->rcv_wnd
= new_win
;
283 tp
->rcv_wup
= tp
->rcv_nxt
;
285 /* Make sure we do not exceed the maximum possible
288 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
289 new_win
= min(new_win
, MAX_TCP_WINDOW
);
291 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
293 /* RFC1323 scaling applied */
294 new_win
>>= tp
->rx_opt
.rcv_wscale
;
296 /* If we advertise zero window, disable fast path. */
303 /* Packet ECN state for a SYN-ACK */
304 static inline void TCP_ECN_send_synack(const struct tcp_sock
*tp
, struct sk_buff
*skb
)
306 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
307 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
308 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
311 /* Packet ECN state for a SYN. */
312 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
314 struct tcp_sock
*tp
= tcp_sk(sk
);
317 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1) {
318 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
319 tp
->ecn_flags
= TCP_ECN_OK
;
323 static __inline__
void
324 TCP_ECN_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
326 if (inet_rsk(req
)->ecn_ok
)
330 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
333 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
336 struct tcp_sock
*tp
= tcp_sk(sk
);
338 if (tp
->ecn_flags
& TCP_ECN_OK
) {
339 /* Not-retransmitted data segment: set ECT and inject CWR. */
340 if (skb
->len
!= tcp_header_len
&&
341 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
343 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
344 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
345 tcp_hdr(skb
)->cwr
= 1;
346 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
349 /* ACK or retransmitted segment: clear ECT|CE */
350 INET_ECN_dontxmit(sk
);
352 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
353 tcp_hdr(skb
)->ece
= 1;
357 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
358 * auto increment end seqno.
360 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
362 skb
->ip_summed
= CHECKSUM_PARTIAL
;
365 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
366 TCP_SKB_CB(skb
)->sacked
= 0;
368 skb_shinfo(skb
)->gso_segs
= 1;
369 skb_shinfo(skb
)->gso_size
= 0;
370 skb_shinfo(skb
)->gso_type
= 0;
372 TCP_SKB_CB(skb
)->seq
= seq
;
373 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
375 TCP_SKB_CB(skb
)->end_seq
= seq
;
378 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
380 return tp
->snd_una
!= tp
->snd_up
;
383 #define OPTION_SACK_ADVERTISE (1 << 0)
384 #define OPTION_TS (1 << 1)
385 #define OPTION_MD5 (1 << 2)
386 #define OPTION_WSCALE (1 << 3)
387 #define OPTION_COOKIE_EXTENSION (1 << 4)
388 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
390 struct tcp_out_options
{
391 u16 options
; /* bit field of OPTION_* */
392 u16 mss
; /* 0 to disable */
393 u8 ws
; /* window scale, 0 to disable */
394 u8 num_sack_blocks
; /* number of SACK blocks to include */
395 u8 hash_size
; /* bytes in hash_location */
396 __u8
*hash_location
; /* temporary pointer, overloaded */
397 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
398 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
401 /* The sysctl int routines are generic, so check consistency here.
403 static u8
tcp_cookie_size_check(u8 desired
)
408 /* previously specified */
411 cookie_size
= ACCESS_ONCE(sysctl_tcp_cookie_size
);
412 if (cookie_size
<= 0)
413 /* no default specified */
416 if (cookie_size
<= TCP_COOKIE_MIN
)
417 /* value too small, specify minimum */
418 return TCP_COOKIE_MIN
;
420 if (cookie_size
>= TCP_COOKIE_MAX
)
421 /* value too large, specify maximum */
422 return TCP_COOKIE_MAX
;
425 /* 8-bit multiple, illegal, fix it */
428 return (u8
)cookie_size
;
431 /* Write previously computed TCP options to the packet.
433 * Beware: Something in the Internet is very sensitive to the ordering of
434 * TCP options, we learned this through the hard way, so be careful here.
435 * Luckily we can at least blame others for their non-compliance but from
436 * inter-operatibility perspective it seems that we're somewhat stuck with
437 * the ordering which we have been using if we want to keep working with
438 * those broken things (not that it currently hurts anybody as there isn't
439 * particular reason why the ordering would need to be changed).
441 * At least SACK_PERM as the first option is known to lead to a disaster
442 * (but it may well be that other scenarios fail similarly).
444 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
445 struct tcp_out_options
*opts
)
447 u16 options
= opts
->options
; /* mungable copy */
449 /* Having both authentication and cookies for security is redundant,
450 * and there's certainly not enough room. Instead, the cookie-less
451 * extension variant is proposed.
453 * Consider the pessimal case with authentication. The options
455 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
457 if (unlikely(OPTION_MD5
& options
)) {
458 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
459 *ptr
++ = htonl((TCPOPT_COOKIE
<< 24) |
460 (TCPOLEN_COOKIE_BASE
<< 16) |
461 (TCPOPT_MD5SIG
<< 8) |
464 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
466 (TCPOPT_MD5SIG
<< 8) |
469 options
&= ~OPTION_COOKIE_EXTENSION
;
470 /* overload cookie hash location */
471 opts
->hash_location
= (__u8
*)ptr
;
475 if (unlikely(opts
->mss
)) {
476 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
477 (TCPOLEN_MSS
<< 16) |
481 if (likely(OPTION_TS
& options
)) {
482 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
483 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
484 (TCPOLEN_SACK_PERM
<< 16) |
485 (TCPOPT_TIMESTAMP
<< 8) |
487 options
&= ~OPTION_SACK_ADVERTISE
;
489 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
491 (TCPOPT_TIMESTAMP
<< 8) |
494 *ptr
++ = htonl(opts
->tsval
);
495 *ptr
++ = htonl(opts
->tsecr
);
498 /* Specification requires after timestamp, so do it now.
500 * Consider the pessimal case without authentication. The options
502 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
504 if (unlikely(OPTION_COOKIE_EXTENSION
& options
)) {
505 __u8
*cookie_copy
= opts
->hash_location
;
506 u8 cookie_size
= opts
->hash_size
;
508 /* 8-bit multiple handled in tcp_cookie_size_check() above,
511 if (0x2 & cookie_size
) {
512 __u8
*p
= (__u8
*)ptr
;
514 /* 16-bit multiple */
515 *p
++ = TCPOPT_COOKIE
;
516 *p
++ = TCPOLEN_COOKIE_BASE
+ cookie_size
;
517 *p
++ = *cookie_copy
++;
518 *p
++ = *cookie_copy
++;
522 /* 32-bit multiple */
523 *ptr
++ = htonl(((TCPOPT_NOP
<< 24) |
525 (TCPOPT_COOKIE
<< 8) |
526 TCPOLEN_COOKIE_BASE
) +
530 if (cookie_size
> 0) {
531 memcpy(ptr
, cookie_copy
, cookie_size
);
532 ptr
+= (cookie_size
/ 4);
536 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
537 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
539 (TCPOPT_SACK_PERM
<< 8) |
543 if (unlikely(OPTION_WSCALE
& options
)) {
544 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
545 (TCPOPT_WINDOW
<< 16) |
546 (TCPOLEN_WINDOW
<< 8) |
550 if (unlikely(opts
->num_sack_blocks
)) {
551 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
552 tp
->duplicate_sack
: tp
->selective_acks
;
555 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
558 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
559 TCPOLEN_SACK_PERBLOCK
)));
561 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
563 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
564 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
567 tp
->rx_opt
.dsack
= 0;
570 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
571 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
573 *ptr
++ = htonl((TCPOPT_EXP
<< 24) |
574 ((TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
) << 16) |
575 TCPOPT_FASTOPEN_MAGIC
);
577 memcpy(ptr
, foc
->val
, foc
->len
);
578 if ((foc
->len
& 3) == 2) {
579 u8
*align
= ((u8
*)ptr
) + foc
->len
;
580 align
[0] = align
[1] = TCPOPT_NOP
;
582 ptr
+= (foc
->len
+ 3) >> 2;
586 /* Compute TCP options for SYN packets. This is not the final
587 * network wire format yet.
589 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
590 struct tcp_out_options
*opts
,
591 struct tcp_md5sig_key
**md5
)
593 struct tcp_sock
*tp
= tcp_sk(sk
);
594 struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
595 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
596 u8 cookie_size
= (!tp
->rx_opt
.cookie_out_never
&& cvp
!= NULL
) ?
597 tcp_cookie_size_check(cvp
->cookie_desired
) :
599 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
601 #ifdef CONFIG_TCP_MD5SIG
602 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
604 opts
->options
|= OPTION_MD5
;
605 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
611 /* We always get an MSS option. The option bytes which will be seen in
612 * normal data packets should timestamps be used, must be in the MSS
613 * advertised. But we subtract them from tp->mss_cache so that
614 * calculations in tcp_sendmsg are simpler etc. So account for this
615 * fact here if necessary. If we don't do this correctly, as a
616 * receiver we won't recognize data packets as being full sized when we
617 * should, and thus we won't abide by the delayed ACK rules correctly.
618 * SACKs don't matter, we never delay an ACK when we have any of those
620 opts
->mss
= tcp_advertise_mss(sk
);
621 remaining
-= TCPOLEN_MSS_ALIGNED
;
623 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
624 opts
->options
|= OPTION_TS
;
625 opts
->tsval
= TCP_SKB_CB(skb
)->when
+ tp
->tsoffset
;
626 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
627 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
629 if (likely(sysctl_tcp_window_scaling
)) {
630 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
631 opts
->options
|= OPTION_WSCALE
;
632 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
634 if (likely(sysctl_tcp_sack
)) {
635 opts
->options
|= OPTION_SACK_ADVERTISE
;
636 if (unlikely(!(OPTION_TS
& opts
->options
)))
637 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
640 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
641 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ fastopen
->cookie
.len
;
642 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
643 if (remaining
>= need
) {
644 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
645 opts
->fastopen_cookie
= &fastopen
->cookie
;
647 tp
->syn_fastopen
= 1;
650 /* Note that timestamps are required by the specification.
652 * Odd numbers of bytes are prohibited by the specification, ensuring
653 * that the cookie is 16-bit aligned, and the resulting cookie pair is
657 (OPTION_TS
& opts
->options
) &&
659 int need
= TCPOLEN_COOKIE_BASE
+ cookie_size
;
662 /* 32-bit multiple */
663 need
+= 2; /* NOPs */
665 if (need
> remaining
) {
666 /* try shrinking cookie to fit */
671 while (need
> remaining
&& TCP_COOKIE_MIN
<= cookie_size
) {
675 if (TCP_COOKIE_MIN
<= cookie_size
) {
676 opts
->options
|= OPTION_COOKIE_EXTENSION
;
677 opts
->hash_location
= (__u8
*)&cvp
->cookie_pair
[0];
678 opts
->hash_size
= cookie_size
;
680 /* Remember for future incarnations. */
681 cvp
->cookie_desired
= cookie_size
;
683 if (cvp
->cookie_desired
!= cvp
->cookie_pair_size
) {
684 /* Currently use random bytes as a nonce,
685 * assuming these are completely unpredictable
686 * by hostile users of the same system.
688 get_random_bytes(&cvp
->cookie_pair
[0],
690 cvp
->cookie_pair_size
= cookie_size
;
696 return MAX_TCP_OPTION_SPACE
- remaining
;
699 /* Set up TCP options for SYN-ACKs. */
700 static unsigned int tcp_synack_options(struct sock
*sk
,
701 struct request_sock
*req
,
702 unsigned int mss
, struct sk_buff
*skb
,
703 struct tcp_out_options
*opts
,
704 struct tcp_md5sig_key
**md5
,
705 struct tcp_extend_values
*xvp
,
706 struct tcp_fastopen_cookie
*foc
)
708 struct inet_request_sock
*ireq
= inet_rsk(req
);
709 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
710 u8 cookie_plus
= (xvp
!= NULL
&& !xvp
->cookie_out_never
) ?
714 #ifdef CONFIG_TCP_MD5SIG
715 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
717 opts
->options
|= OPTION_MD5
;
718 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
720 /* We can't fit any SACK blocks in a packet with MD5 + TS
721 * options. There was discussion about disabling SACK
722 * rather than TS in order to fit in better with old,
723 * buggy kernels, but that was deemed to be unnecessary.
725 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
731 /* We always send an MSS option. */
733 remaining
-= TCPOLEN_MSS_ALIGNED
;
735 if (likely(ireq
->wscale_ok
)) {
736 opts
->ws
= ireq
->rcv_wscale
;
737 opts
->options
|= OPTION_WSCALE
;
738 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
740 if (likely(ireq
->tstamp_ok
)) {
741 opts
->options
|= OPTION_TS
;
742 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
743 opts
->tsecr
= req
->ts_recent
;
744 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
746 if (likely(ireq
->sack_ok
)) {
747 opts
->options
|= OPTION_SACK_ADVERTISE
;
748 if (unlikely(!ireq
->tstamp_ok
))
749 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
752 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
753 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
754 if (remaining
>= need
) {
755 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
756 opts
->fastopen_cookie
= foc
;
760 /* Similar rationale to tcp_syn_options() applies here, too.
761 * If the <SYN> options fit, the same options should fit now!
765 cookie_plus
> TCPOLEN_COOKIE_BASE
) {
766 int need
= cookie_plus
; /* has TCPOLEN_COOKIE_BASE */
769 /* 32-bit multiple */
770 need
+= 2; /* NOPs */
772 if (need
<= remaining
) {
773 opts
->options
|= OPTION_COOKIE_EXTENSION
;
774 opts
->hash_size
= cookie_plus
- TCPOLEN_COOKIE_BASE
;
777 /* There's no error return, so flag it. */
778 xvp
->cookie_out_never
= 1; /* true */
782 return MAX_TCP_OPTION_SPACE
- remaining
;
785 /* Compute TCP options for ESTABLISHED sockets. This is not the
786 * final wire format yet.
788 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
789 struct tcp_out_options
*opts
,
790 struct tcp_md5sig_key
**md5
)
792 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
793 struct tcp_sock
*tp
= tcp_sk(sk
);
794 unsigned int size
= 0;
795 unsigned int eff_sacks
;
797 #ifdef CONFIG_TCP_MD5SIG
798 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
799 if (unlikely(*md5
)) {
800 opts
->options
|= OPTION_MD5
;
801 size
+= TCPOLEN_MD5SIG_ALIGNED
;
807 if (likely(tp
->rx_opt
.tstamp_ok
)) {
808 opts
->options
|= OPTION_TS
;
809 opts
->tsval
= tcb
? tcb
->when
+ tp
->tsoffset
: 0;
810 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
811 size
+= TCPOLEN_TSTAMP_ALIGNED
;
814 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
815 if (unlikely(eff_sacks
)) {
816 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
817 opts
->num_sack_blocks
=
818 min_t(unsigned int, eff_sacks
,
819 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
820 TCPOLEN_SACK_PERBLOCK
);
821 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
822 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
829 /* TCP SMALL QUEUES (TSQ)
831 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
832 * to reduce RTT and bufferbloat.
833 * We do this using a special skb destructor (tcp_wfree).
835 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
836 * needs to be reallocated in a driver.
837 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
839 * Since transmit from skb destructor is forbidden, we use a tasklet
840 * to process all sockets that eventually need to send more skbs.
841 * We use one tasklet per cpu, with its own queue of sockets.
844 struct tasklet_struct tasklet
;
845 struct list_head head
; /* queue of tcp sockets */
847 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
849 static void tcp_tsq_handler(struct sock
*sk
)
851 if ((1 << sk
->sk_state
) &
852 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
853 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
))
854 tcp_write_xmit(sk
, tcp_current_mss(sk
), 0, 0, GFP_ATOMIC
);
857 * One tasklest per cpu tries to send more skbs.
858 * We run in tasklet context but need to disable irqs when
859 * transfering tsq->head because tcp_wfree() might
860 * interrupt us (non NAPI drivers)
862 static void tcp_tasklet_func(unsigned long data
)
864 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
867 struct list_head
*q
, *n
;
871 local_irq_save(flags
);
872 list_splice_init(&tsq
->head
, &list
);
873 local_irq_restore(flags
);
875 list_for_each_safe(q
, n
, &list
) {
876 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
877 list_del(&tp
->tsq_node
);
879 sk
= (struct sock
*)tp
;
882 if (!sock_owned_by_user(sk
)) {
885 /* defer the work to tcp_release_cb() */
886 set_bit(TCP_TSQ_DEFERRED
, &tp
->tsq_flags
);
890 clear_bit(TSQ_QUEUED
, &tp
->tsq_flags
);
895 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
896 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
897 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
898 (1UL << TCP_MTU_REDUCED_DEFERRED))
900 * tcp_release_cb - tcp release_sock() callback
903 * called from release_sock() to perform protocol dependent
904 * actions before socket release.
906 void tcp_release_cb(struct sock
*sk
)
908 struct tcp_sock
*tp
= tcp_sk(sk
);
909 unsigned long flags
, nflags
;
911 /* perform an atomic operation only if at least one flag is set */
913 flags
= tp
->tsq_flags
;
914 if (!(flags
& TCP_DEFERRED_ALL
))
916 nflags
= flags
& ~TCP_DEFERRED_ALL
;
917 } while (cmpxchg(&tp
->tsq_flags
, flags
, nflags
) != flags
);
919 if (flags
& (1UL << TCP_TSQ_DEFERRED
))
922 if (flags
& (1UL << TCP_WRITE_TIMER_DEFERRED
)) {
923 tcp_write_timer_handler(sk
);
926 if (flags
& (1UL << TCP_DELACK_TIMER_DEFERRED
)) {
927 tcp_delack_timer_handler(sk
);
930 if (flags
& (1UL << TCP_MTU_REDUCED_DEFERRED
)) {
931 sk
->sk_prot
->mtu_reduced(sk
);
935 EXPORT_SYMBOL(tcp_release_cb
);
937 void __init
tcp_tasklet_init(void)
941 for_each_possible_cpu(i
) {
942 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
944 INIT_LIST_HEAD(&tsq
->head
);
945 tasklet_init(&tsq
->tasklet
,
952 * Write buffer destructor automatically called from kfree_skb.
953 * We cant xmit new skbs from this context, as we might already
956 static void tcp_wfree(struct sk_buff
*skb
)
958 struct sock
*sk
= skb
->sk
;
959 struct tcp_sock
*tp
= tcp_sk(sk
);
961 if (test_and_clear_bit(TSQ_THROTTLED
, &tp
->tsq_flags
) &&
962 !test_and_set_bit(TSQ_QUEUED
, &tp
->tsq_flags
)) {
964 struct tsq_tasklet
*tsq
;
966 /* Keep a ref on socket.
967 * This last ref will be released in tcp_tasklet_func()
969 atomic_sub(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
971 /* queue this socket to tasklet queue */
972 local_irq_save(flags
);
973 tsq
= &__get_cpu_var(tsq_tasklet
);
974 list_add(&tp
->tsq_node
, &tsq
->head
);
975 tasklet_schedule(&tsq
->tasklet
);
976 local_irq_restore(flags
);
982 /* This routine actually transmits TCP packets queued in by
983 * tcp_do_sendmsg(). This is used by both the initial
984 * transmission and possible later retransmissions.
985 * All SKB's seen here are completely headerless. It is our
986 * job to build the TCP header, and pass the packet down to
987 * IP so it can do the same plus pass the packet off to the
990 * We are working here with either a clone of the original
991 * SKB, or a fresh unique copy made by the retransmit engine.
993 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
996 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
997 struct inet_sock
*inet
;
999 struct tcp_skb_cb
*tcb
;
1000 struct tcp_out_options opts
;
1001 unsigned int tcp_options_size
, tcp_header_size
;
1002 struct tcp_md5sig_key
*md5
;
1006 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
1008 /* If congestion control is doing timestamping, we must
1009 * take such a timestamp before we potentially clone/copy.
1011 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
1012 __net_timestamp(skb
);
1014 if (likely(clone_it
)) {
1015 if (unlikely(skb_cloned(skb
)))
1016 skb
= pskb_copy(skb
, gfp_mask
);
1018 skb
= skb_clone(skb
, gfp_mask
);
1025 tcb
= TCP_SKB_CB(skb
);
1026 memset(&opts
, 0, sizeof(opts
));
1028 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
1029 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
1031 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
1033 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
1035 if (tcp_packets_in_flight(tp
) == 0) {
1036 tcp_ca_event(sk
, CA_EVENT_TX_START
);
1041 skb_push(skb
, tcp_header_size
);
1042 skb_reset_transport_header(skb
);
1046 skb
->destructor
= (sysctl_tcp_limit_output_bytes
> 0) ?
1047 tcp_wfree
: sock_wfree
;
1048 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
1050 /* Build TCP header and checksum it. */
1052 th
->source
= inet
->inet_sport
;
1053 th
->dest
= inet
->inet_dport
;
1054 th
->seq
= htonl(tcb
->seq
);
1055 th
->ack_seq
= htonl(tp
->rcv_nxt
);
1056 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
1059 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
1060 /* RFC1323: The window in SYN & SYN/ACK segments
1063 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
1065 th
->window
= htons(tcp_select_window(sk
));
1070 /* The urg_mode check is necessary during a below snd_una win probe */
1071 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
1072 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
1073 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
1075 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
1076 th
->urg_ptr
= htons(0xFFFF);
1081 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
1082 if (likely((tcb
->tcp_flags
& TCPHDR_SYN
) == 0))
1083 TCP_ECN_send(sk
, skb
, tcp_header_size
);
1085 #ifdef CONFIG_TCP_MD5SIG
1086 /* Calculate the MD5 hash, as we have all we need now */
1088 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
1089 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
1090 md5
, sk
, NULL
, skb
);
1094 icsk
->icsk_af_ops
->send_check(sk
, skb
);
1096 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
1097 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
1099 if (skb
->len
!= tcp_header_size
)
1100 tcp_event_data_sent(tp
, sk
);
1102 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
1103 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
1104 tcp_skb_pcount(skb
));
1106 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, &inet
->cork
.fl
);
1107 if (likely(err
<= 0))
1110 tcp_enter_cwr(sk
, 1);
1112 return net_xmit_eval(err
);
1115 /* This routine just queues the buffer for sending.
1117 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1118 * otherwise socket can stall.
1120 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1122 struct tcp_sock
*tp
= tcp_sk(sk
);
1124 /* Advance write_seq and place onto the write_queue. */
1125 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
1126 skb_header_release(skb
);
1127 tcp_add_write_queue_tail(sk
, skb
);
1128 sk
->sk_wmem_queued
+= skb
->truesize
;
1129 sk_mem_charge(sk
, skb
->truesize
);
1132 /* Initialize TSO segments for a packet. */
1133 static void tcp_set_skb_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
1134 unsigned int mss_now
)
1136 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
) ||
1137 skb
->ip_summed
== CHECKSUM_NONE
) {
1138 /* Avoid the costly divide in the normal
1141 skb_shinfo(skb
)->gso_segs
= 1;
1142 skb_shinfo(skb
)->gso_size
= 0;
1143 skb_shinfo(skb
)->gso_type
= 0;
1145 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
1146 skb_shinfo(skb
)->gso_size
= mss_now
;
1147 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
1151 /* When a modification to fackets out becomes necessary, we need to check
1152 * skb is counted to fackets_out or not.
1154 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1157 struct tcp_sock
*tp
= tcp_sk(sk
);
1159 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1162 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1163 tp
->fackets_out
-= decr
;
1166 /* Pcount in the middle of the write queue got changed, we need to do various
1167 * tweaks to fix counters
1169 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1171 struct tcp_sock
*tp
= tcp_sk(sk
);
1173 tp
->packets_out
-= decr
;
1175 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1176 tp
->sacked_out
-= decr
;
1177 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1178 tp
->retrans_out
-= decr
;
1179 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1180 tp
->lost_out
-= decr
;
1182 /* Reno case is special. Sigh... */
1183 if (tcp_is_reno(tp
) && decr
> 0)
1184 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1186 tcp_adjust_fackets_out(sk
, skb
, decr
);
1188 if (tp
->lost_skb_hint
&&
1189 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1190 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1191 tp
->lost_cnt_hint
-= decr
;
1193 tcp_verify_left_out(tp
);
1196 /* Function to create two new TCP segments. Shrinks the given segment
1197 * to the specified size and appends a new segment with the rest of the
1198 * packet to the list. This won't be called frequently, I hope.
1199 * Remember, these are still headerless SKBs at this point.
1201 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1202 unsigned int mss_now
)
1204 struct tcp_sock
*tp
= tcp_sk(sk
);
1205 struct sk_buff
*buff
;
1206 int nsize
, old_factor
;
1210 if (WARN_ON(len
> skb
->len
))
1213 nsize
= skb_headlen(skb
) - len
;
1217 if (skb_cloned(skb
) &&
1218 skb_is_nonlinear(skb
) &&
1219 pskb_expand_head(skb
, 0, 0, GFP_ATOMIC
))
1222 /* Get a new skb... force flag on. */
1223 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
1225 return -ENOMEM
; /* We'll just try again later. */
1227 sk
->sk_wmem_queued
+= buff
->truesize
;
1228 sk_mem_charge(sk
, buff
->truesize
);
1229 nlen
= skb
->len
- len
- nsize
;
1230 buff
->truesize
+= nlen
;
1231 skb
->truesize
-= nlen
;
1233 /* Correct the sequence numbers. */
1234 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1235 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1236 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1238 /* PSH and FIN should only be set in the second packet. */
1239 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1240 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1241 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1242 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1244 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1245 /* Copy and checksum data tail into the new buffer. */
1246 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1247 skb_put(buff
, nsize
),
1252 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1254 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1255 skb_split(skb
, buff
, len
);
1258 buff
->ip_summed
= skb
->ip_summed
;
1260 /* Looks stupid, but our code really uses when of
1261 * skbs, which it never sent before. --ANK
1263 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
1264 buff
->tstamp
= skb
->tstamp
;
1266 old_factor
= tcp_skb_pcount(skb
);
1268 /* Fix up tso_factor for both original and new SKB. */
1269 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1270 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1272 /* If this packet has been sent out already, we must
1273 * adjust the various packet counters.
1275 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1276 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1277 tcp_skb_pcount(buff
);
1280 tcp_adjust_pcount(sk
, skb
, diff
);
1283 /* Link BUFF into the send queue. */
1284 skb_header_release(buff
);
1285 tcp_insert_write_queue_after(skb
, buff
, sk
);
1290 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1291 * eventually). The difference is that pulled data not copied, but
1292 * immediately discarded.
1294 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1298 eat
= min_t(int, len
, skb_headlen(skb
));
1300 __skb_pull(skb
, eat
);
1301 skb
->avail_size
-= eat
;
1308 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
1309 int size
= skb_frag_size(&skb_shinfo(skb
)->frags
[i
]);
1312 skb_frag_unref(skb
, i
);
1315 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
1317 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
1318 skb_frag_size_sub(&skb_shinfo(skb
)->frags
[k
], eat
);
1324 skb_shinfo(skb
)->nr_frags
= k
;
1326 skb_reset_tail_pointer(skb
);
1327 skb
->data_len
-= len
;
1328 skb
->len
= skb
->data_len
;
1331 /* Remove acked data from a packet in the transmit queue. */
1332 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1334 if (skb_unclone(skb
, GFP_ATOMIC
))
1337 __pskb_trim_head(skb
, len
);
1339 TCP_SKB_CB(skb
)->seq
+= len
;
1340 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1342 skb
->truesize
-= len
;
1343 sk
->sk_wmem_queued
-= len
;
1344 sk_mem_uncharge(sk
, len
);
1345 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1347 /* Any change of skb->len requires recalculation of tso factor. */
1348 if (tcp_skb_pcount(skb
) > 1)
1349 tcp_set_skb_tso_segs(sk
, skb
, tcp_skb_mss(skb
));
1354 /* Calculate MSS not accounting any TCP options. */
1355 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1357 const struct tcp_sock
*tp
= tcp_sk(sk
);
1358 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1361 /* Calculate base mss without TCP options:
1362 It is MMS_S - sizeof(tcphdr) of rfc1122
1364 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1366 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1367 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1368 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1370 if (dst
&& dst_allfrag(dst
))
1371 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1374 /* Clamp it (mss_clamp does not include tcp options) */
1375 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1376 mss_now
= tp
->rx_opt
.mss_clamp
;
1378 /* Now subtract optional transport overhead */
1379 mss_now
-= icsk
->icsk_ext_hdr_len
;
1381 /* Then reserve room for full set of TCP options and 8 bytes of data */
1387 /* Calculate MSS. Not accounting for SACKs here. */
1388 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1390 /* Subtract TCP options size, not including SACKs */
1391 return __tcp_mtu_to_mss(sk
, pmtu
) -
1392 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1395 /* Inverse of above */
1396 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1398 const struct tcp_sock
*tp
= tcp_sk(sk
);
1399 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1403 tp
->tcp_header_len
+
1404 icsk
->icsk_ext_hdr_len
+
1405 icsk
->icsk_af_ops
->net_header_len
;
1407 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1408 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1409 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1411 if (dst
&& dst_allfrag(dst
))
1412 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1417 /* MTU probing init per socket */
1418 void tcp_mtup_init(struct sock
*sk
)
1420 struct tcp_sock
*tp
= tcp_sk(sk
);
1421 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1423 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
1424 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1425 icsk
->icsk_af_ops
->net_header_len
;
1426 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
1427 icsk
->icsk_mtup
.probe_size
= 0;
1429 EXPORT_SYMBOL(tcp_mtup_init
);
1431 /* This function synchronize snd mss to current pmtu/exthdr set.
1433 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1434 for TCP options, but includes only bare TCP header.
1436 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1437 It is minimum of user_mss and mss received with SYN.
1438 It also does not include TCP options.
1440 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1442 tp->mss_cache is current effective sending mss, including
1443 all tcp options except for SACKs. It is evaluated,
1444 taking into account current pmtu, but never exceeds
1445 tp->rx_opt.mss_clamp.
1447 NOTE1. rfc1122 clearly states that advertised MSS
1448 DOES NOT include either tcp or ip options.
1450 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1451 are READ ONLY outside this function. --ANK (980731)
1453 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1455 struct tcp_sock
*tp
= tcp_sk(sk
);
1456 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1459 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1460 icsk
->icsk_mtup
.search_high
= pmtu
;
1462 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1463 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1465 /* And store cached results */
1466 icsk
->icsk_pmtu_cookie
= pmtu
;
1467 if (icsk
->icsk_mtup
.enabled
)
1468 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1469 tp
->mss_cache
= mss_now
;
1473 EXPORT_SYMBOL(tcp_sync_mss
);
1475 /* Compute the current effective MSS, taking SACKs and IP options,
1476 * and even PMTU discovery events into account.
1478 unsigned int tcp_current_mss(struct sock
*sk
)
1480 const struct tcp_sock
*tp
= tcp_sk(sk
);
1481 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1483 unsigned int header_len
;
1484 struct tcp_out_options opts
;
1485 struct tcp_md5sig_key
*md5
;
1487 mss_now
= tp
->mss_cache
;
1490 u32 mtu
= dst_mtu(dst
);
1491 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1492 mss_now
= tcp_sync_mss(sk
, mtu
);
1495 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1496 sizeof(struct tcphdr
);
1497 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1498 * some common options. If this is an odd packet (because we have SACK
1499 * blocks etc) then our calculated header_len will be different, and
1500 * we have to adjust mss_now correspondingly */
1501 if (header_len
!= tp
->tcp_header_len
) {
1502 int delta
= (int) header_len
- tp
->tcp_header_len
;
1509 /* Congestion window validation. (RFC2861) */
1510 static void tcp_cwnd_validate(struct sock
*sk
)
1512 struct tcp_sock
*tp
= tcp_sk(sk
);
1514 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1515 /* Network is feed fully. */
1516 tp
->snd_cwnd_used
= 0;
1517 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1519 /* Network starves. */
1520 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1521 tp
->snd_cwnd_used
= tp
->packets_out
;
1523 if (sysctl_tcp_slow_start_after_idle
&&
1524 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1525 tcp_cwnd_application_limited(sk
);
1529 /* Returns the portion of skb which can be sent right away without
1530 * introducing MSS oddities to segment boundaries. In rare cases where
1531 * mss_now != mss_cache, we will request caller to create a small skb
1532 * per input skb which could be mostly avoided here (if desired).
1534 * We explicitly want to create a request for splitting write queue tail
1535 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1536 * thus all the complexity (cwnd_len is always MSS multiple which we
1537 * return whenever allowed by the other factors). Basically we need the
1538 * modulo only when the receiver window alone is the limiting factor or
1539 * when we would be allowed to send the split-due-to-Nagle skb fully.
1541 static unsigned int tcp_mss_split_point(const struct sock
*sk
, const struct sk_buff
*skb
,
1542 unsigned int mss_now
, unsigned int max_segs
)
1544 const struct tcp_sock
*tp
= tcp_sk(sk
);
1545 u32 needed
, window
, max_len
;
1547 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1548 max_len
= mss_now
* max_segs
;
1550 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1553 needed
= min(skb
->len
, window
);
1555 if (max_len
<= needed
)
1558 return needed
- needed
% mss_now
;
1561 /* Can at least one segment of SKB be sent right now, according to the
1562 * congestion window rules? If so, return how many segments are allowed.
1564 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1565 const struct sk_buff
*skb
)
1567 u32 in_flight
, cwnd
;
1569 /* Don't be strict about the congestion window for the final FIN. */
1570 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1571 tcp_skb_pcount(skb
) == 1)
1574 in_flight
= tcp_packets_in_flight(tp
);
1575 cwnd
= tp
->snd_cwnd
;
1576 if (in_flight
< cwnd
)
1577 return (cwnd
- in_flight
);
1582 /* Initialize TSO state of a skb.
1583 * This must be invoked the first time we consider transmitting
1584 * SKB onto the wire.
1586 static int tcp_init_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
1587 unsigned int mss_now
)
1589 int tso_segs
= tcp_skb_pcount(skb
);
1591 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1592 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1593 tso_segs
= tcp_skb_pcount(skb
);
1598 /* Minshall's variant of the Nagle send check. */
1599 static inline bool tcp_minshall_check(const struct tcp_sock
*tp
)
1601 return after(tp
->snd_sml
, tp
->snd_una
) &&
1602 !after(tp
->snd_sml
, tp
->snd_nxt
);
1605 /* Return false, if packet can be sent now without violation Nagle's rules:
1606 * 1. It is full sized.
1607 * 2. Or it contains FIN. (already checked by caller)
1608 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1609 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1610 * With Minshall's modification: all sent small packets are ACKed.
1612 static inline bool tcp_nagle_check(const struct tcp_sock
*tp
,
1613 const struct sk_buff
*skb
,
1614 unsigned int mss_now
, int nonagle
)
1616 return skb
->len
< mss_now
&&
1617 ((nonagle
& TCP_NAGLE_CORK
) ||
1618 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1621 /* Return true if the Nagle test allows this packet to be
1624 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1625 unsigned int cur_mss
, int nonagle
)
1627 /* Nagle rule does not apply to frames, which sit in the middle of the
1628 * write_queue (they have no chances to get new data).
1630 * This is implemented in the callers, where they modify the 'nonagle'
1631 * argument based upon the location of SKB in the send queue.
1633 if (nonagle
& TCP_NAGLE_PUSH
)
1636 /* Don't use the nagle rule for urgent data (or for the final FIN).
1637 * Nagle can be ignored during F-RTO too (see RFC4138).
1639 if (tcp_urg_mode(tp
) || (tp
->frto_counter
== 2) ||
1640 (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1643 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1649 /* Does at least the first segment of SKB fit into the send window? */
1650 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1651 const struct sk_buff
*skb
,
1652 unsigned int cur_mss
)
1654 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1656 if (skb
->len
> cur_mss
)
1657 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1659 return !after(end_seq
, tcp_wnd_end(tp
));
1662 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1663 * should be put on the wire right now. If so, it returns the number of
1664 * packets allowed by the congestion window.
1666 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1667 unsigned int cur_mss
, int nonagle
)
1669 const struct tcp_sock
*tp
= tcp_sk(sk
);
1670 unsigned int cwnd_quota
;
1672 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1674 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1677 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1678 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1684 /* Test if sending is allowed right now. */
1685 bool tcp_may_send_now(struct sock
*sk
)
1687 const struct tcp_sock
*tp
= tcp_sk(sk
);
1688 struct sk_buff
*skb
= tcp_send_head(sk
);
1691 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1692 (tcp_skb_is_last(sk
, skb
) ?
1693 tp
->nonagle
: TCP_NAGLE_PUSH
));
1696 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1697 * which is put after SKB on the list. It is very much like
1698 * tcp_fragment() except that it may make several kinds of assumptions
1699 * in order to speed up the splitting operation. In particular, we
1700 * know that all the data is in scatter-gather pages, and that the
1701 * packet has never been sent out before (and thus is not cloned).
1703 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1704 unsigned int mss_now
, gfp_t gfp
)
1706 struct sk_buff
*buff
;
1707 int nlen
= skb
->len
- len
;
1710 /* All of a TSO frame must be composed of paged data. */
1711 if (skb
->len
!= skb
->data_len
)
1712 return tcp_fragment(sk
, skb
, len
, mss_now
);
1714 buff
= sk_stream_alloc_skb(sk
, 0, gfp
);
1715 if (unlikely(buff
== NULL
))
1718 sk
->sk_wmem_queued
+= buff
->truesize
;
1719 sk_mem_charge(sk
, buff
->truesize
);
1720 buff
->truesize
+= nlen
;
1721 skb
->truesize
-= nlen
;
1723 /* Correct the sequence numbers. */
1724 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1725 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1726 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1728 /* PSH and FIN should only be set in the second packet. */
1729 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1730 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1731 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1733 /* This packet was never sent out yet, so no SACK bits. */
1734 TCP_SKB_CB(buff
)->sacked
= 0;
1736 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1737 skb_split(skb
, buff
, len
);
1739 /* Fix up tso_factor for both original and new SKB. */
1740 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1741 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1743 /* Link BUFF into the send queue. */
1744 skb_header_release(buff
);
1745 tcp_insert_write_queue_after(skb
, buff
, sk
);
1750 /* Try to defer sending, if possible, in order to minimize the amount
1751 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1753 * This algorithm is from John Heffner.
1755 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1757 struct tcp_sock
*tp
= tcp_sk(sk
);
1758 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1759 u32 send_win
, cong_win
, limit
, in_flight
;
1762 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1765 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1768 /* Defer for less than two clock ticks. */
1769 if (tp
->tso_deferred
&&
1770 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1773 in_flight
= tcp_packets_in_flight(tp
);
1775 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1777 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1779 /* From in_flight test above, we know that cwnd > in_flight. */
1780 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1782 limit
= min(send_win
, cong_win
);
1784 /* If a full-sized TSO skb can be sent, do it. */
1785 if (limit
>= min_t(unsigned int, sk
->sk_gso_max_size
,
1786 sk
->sk_gso_max_segs
* tp
->mss_cache
))
1789 /* Middle in queue won't get any more data, full sendable already? */
1790 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1793 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1795 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1797 /* If at least some fraction of a window is available,
1800 chunk
/= win_divisor
;
1804 /* Different approach, try not to defer past a single
1805 * ACK. Receiver should ACK every other full sized
1806 * frame, so if we have space for more than 3 frames
1809 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1813 /* Ok, it looks like it is advisable to defer. */
1814 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1819 tp
->tso_deferred
= 0;
1823 /* Create a new MTU probe if we are ready.
1824 * MTU probe is regularly attempting to increase the path MTU by
1825 * deliberately sending larger packets. This discovers routing
1826 * changes resulting in larger path MTUs.
1828 * Returns 0 if we should wait to probe (no cwnd available),
1829 * 1 if a probe was sent,
1832 static int tcp_mtu_probe(struct sock
*sk
)
1834 struct tcp_sock
*tp
= tcp_sk(sk
);
1835 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1836 struct sk_buff
*skb
, *nskb
, *next
;
1843 /* Not currently probing/verifying,
1845 * have enough cwnd, and
1846 * not SACKing (the variable headers throw things off) */
1847 if (!icsk
->icsk_mtup
.enabled
||
1848 icsk
->icsk_mtup
.probe_size
||
1849 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1850 tp
->snd_cwnd
< 11 ||
1851 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1854 /* Very simple search strategy: just double the MSS. */
1855 mss_now
= tcp_current_mss(sk
);
1856 probe_size
= 2 * tp
->mss_cache
;
1857 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1858 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1859 /* TODO: set timer for probe_converge_event */
1863 /* Have enough data in the send queue to probe? */
1864 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1867 if (tp
->snd_wnd
< size_needed
)
1869 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1872 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1873 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1874 if (!tcp_packets_in_flight(tp
))
1880 /* We're allowed to probe. Build it now. */
1881 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1883 sk
->sk_wmem_queued
+= nskb
->truesize
;
1884 sk_mem_charge(sk
, nskb
->truesize
);
1886 skb
= tcp_send_head(sk
);
1888 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1889 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1890 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
1891 TCP_SKB_CB(nskb
)->sacked
= 0;
1893 nskb
->ip_summed
= skb
->ip_summed
;
1895 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1898 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1899 copy
= min_t(int, skb
->len
, probe_size
- len
);
1900 if (nskb
->ip_summed
)
1901 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1903 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1904 skb_put(nskb
, copy
),
1907 if (skb
->len
<= copy
) {
1908 /* We've eaten all the data from this skb.
1910 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1911 tcp_unlink_write_queue(skb
, sk
);
1912 sk_wmem_free_skb(sk
, skb
);
1914 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
1915 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1916 if (!skb_shinfo(skb
)->nr_frags
) {
1917 skb_pull(skb
, copy
);
1918 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1919 skb
->csum
= csum_partial(skb
->data
,
1922 __pskb_trim_head(skb
, copy
);
1923 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1925 TCP_SKB_CB(skb
)->seq
+= copy
;
1930 if (len
>= probe_size
)
1933 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1935 /* We're ready to send. If this fails, the probe will
1936 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1937 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1938 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1939 /* Decrement cwnd here because we are sending
1940 * effectively two packets. */
1942 tcp_event_new_data_sent(sk
, nskb
);
1944 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1945 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1946 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1954 /* This routine writes packets to the network. It advances the
1955 * send_head. This happens as incoming acks open up the remote
1958 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1959 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1960 * account rare use of URG, this is not a big flaw.
1962 * Returns true, if no segments are in flight and we have queued segments,
1963 * but cannot send anything now because of SWS or another problem.
1965 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1966 int push_one
, gfp_t gfp
)
1968 struct tcp_sock
*tp
= tcp_sk(sk
);
1969 struct sk_buff
*skb
;
1970 unsigned int tso_segs
, sent_pkts
;
1977 /* Do MTU probing. */
1978 result
= tcp_mtu_probe(sk
);
1981 } else if (result
> 0) {
1986 while ((skb
= tcp_send_head(sk
))) {
1990 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1993 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
)
1994 goto repair
; /* Skip network transmission */
1996 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2000 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
2003 if (tso_segs
== 1) {
2004 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2005 (tcp_skb_is_last(sk
, skb
) ?
2006 nonagle
: TCP_NAGLE_PUSH
))))
2009 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
2013 /* TSQ : sk_wmem_alloc accounts skb truesize,
2014 * including skb overhead. But thats OK.
2016 if (atomic_read(&sk
->sk_wmem_alloc
) >= sysctl_tcp_limit_output_bytes
) {
2017 set_bit(TSQ_THROTTLED
, &tp
->tsq_flags
);
2021 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
2022 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2025 sk
->sk_gso_max_segs
));
2027 if (skb
->len
> limit
&&
2028 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
2031 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2033 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2037 /* Advance the send_head. This one is sent out.
2038 * This call will increment packets_out.
2040 tcp_event_new_data_sent(sk
, skb
);
2042 tcp_minshall_update(tp
, mss_now
, skb
);
2043 sent_pkts
+= tcp_skb_pcount(skb
);
2049 if (likely(sent_pkts
)) {
2050 if (tcp_in_cwnd_reduction(sk
))
2051 tp
->prr_out
+= sent_pkts
;
2052 tcp_cwnd_validate(sk
);
2055 return !tp
->packets_out
&& tcp_send_head(sk
);
2058 /* Push out any pending frames which were held back due to
2059 * TCP_CORK or attempt at coalescing tiny packets.
2060 * The socket must be locked by the caller.
2062 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2065 /* If we are closed, the bytes will have to remain here.
2066 * In time closedown will finish, we empty the write queue and
2067 * all will be happy.
2069 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2072 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2073 sk_gfp_atomic(sk
, GFP_ATOMIC
)))
2074 tcp_check_probe_timer(sk
);
2077 /* Send _single_ skb sitting at the send head. This function requires
2078 * true push pending frames to setup probe timer etc.
2080 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2082 struct sk_buff
*skb
= tcp_send_head(sk
);
2084 BUG_ON(!skb
|| skb
->len
< mss_now
);
2086 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2089 /* This function returns the amount that we can raise the
2090 * usable window based on the following constraints
2092 * 1. The window can never be shrunk once it is offered (RFC 793)
2093 * 2. We limit memory per socket
2096 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2097 * RECV.NEXT + RCV.WIN fixed until:
2098 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2100 * i.e. don't raise the right edge of the window until you can raise
2101 * it at least MSS bytes.
2103 * Unfortunately, the recommended algorithm breaks header prediction,
2104 * since header prediction assumes th->window stays fixed.
2106 * Strictly speaking, keeping th->window fixed violates the receiver
2107 * side SWS prevention criteria. The problem is that under this rule
2108 * a stream of single byte packets will cause the right side of the
2109 * window to always advance by a single byte.
2111 * Of course, if the sender implements sender side SWS prevention
2112 * then this will not be a problem.
2114 * BSD seems to make the following compromise:
2116 * If the free space is less than the 1/4 of the maximum
2117 * space available and the free space is less than 1/2 mss,
2118 * then set the window to 0.
2119 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2120 * Otherwise, just prevent the window from shrinking
2121 * and from being larger than the largest representable value.
2123 * This prevents incremental opening of the window in the regime
2124 * where TCP is limited by the speed of the reader side taking
2125 * data out of the TCP receive queue. It does nothing about
2126 * those cases where the window is constrained on the sender side
2127 * because the pipeline is full.
2129 * BSD also seems to "accidentally" limit itself to windows that are a
2130 * multiple of MSS, at least until the free space gets quite small.
2131 * This would appear to be a side effect of the mbuf implementation.
2132 * Combining these two algorithms results in the observed behavior
2133 * of having a fixed window size at almost all times.
2135 * Below we obtain similar behavior by forcing the offered window to
2136 * a multiple of the mss when it is feasible to do so.
2138 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2139 * Regular options like TIMESTAMP are taken into account.
2141 u32
__tcp_select_window(struct sock
*sk
)
2143 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2144 struct tcp_sock
*tp
= tcp_sk(sk
);
2145 /* MSS for the peer's data. Previous versions used mss_clamp
2146 * here. I don't know if the value based on our guesses
2147 * of peer's MSS is better for the performance. It's more correct
2148 * but may be worse for the performance because of rcv_mss
2149 * fluctuations. --SAW 1998/11/1
2151 int mss
= icsk
->icsk_ack
.rcv_mss
;
2152 int free_space
= tcp_space(sk
);
2153 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
2156 if (mss
> full_space
)
2159 if (free_space
< (full_space
>> 1)) {
2160 icsk
->icsk_ack
.quick
= 0;
2162 if (sk_under_memory_pressure(sk
))
2163 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2166 if (free_space
< mss
)
2170 if (free_space
> tp
->rcv_ssthresh
)
2171 free_space
= tp
->rcv_ssthresh
;
2173 /* Don't do rounding if we are using window scaling, since the
2174 * scaled window will not line up with the MSS boundary anyway.
2176 window
= tp
->rcv_wnd
;
2177 if (tp
->rx_opt
.rcv_wscale
) {
2178 window
= free_space
;
2180 /* Advertise enough space so that it won't get scaled away.
2181 * Import case: prevent zero window announcement if
2182 * 1<<rcv_wscale > mss.
2184 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2185 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2186 << tp
->rx_opt
.rcv_wscale
);
2188 /* Get the largest window that is a nice multiple of mss.
2189 * Window clamp already applied above.
2190 * If our current window offering is within 1 mss of the
2191 * free space we just keep it. This prevents the divide
2192 * and multiply from happening most of the time.
2193 * We also don't do any window rounding when the free space
2196 if (window
<= free_space
- mss
|| window
> free_space
)
2197 window
= (free_space
/ mss
) * mss
;
2198 else if (mss
== full_space
&&
2199 free_space
> window
+ (full_space
>> 1))
2200 window
= free_space
;
2206 /* Collapses two adjacent SKB's during retransmission. */
2207 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2209 struct tcp_sock
*tp
= tcp_sk(sk
);
2210 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2211 int skb_size
, next_skb_size
;
2213 skb_size
= skb
->len
;
2214 next_skb_size
= next_skb
->len
;
2216 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2218 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2220 tcp_unlink_write_queue(next_skb
, sk
);
2222 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
2225 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2226 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2228 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2229 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2231 /* Update sequence range on original skb. */
2232 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2234 /* Merge over control information. This moves PSH/FIN etc. over */
2235 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2237 /* All done, get rid of second SKB and account for it so
2238 * packet counting does not break.
2240 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2242 /* changed transmit queue under us so clear hints */
2243 tcp_clear_retrans_hints_partial(tp
);
2244 if (next_skb
== tp
->retransmit_skb_hint
)
2245 tp
->retransmit_skb_hint
= skb
;
2247 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2249 sk_wmem_free_skb(sk
, next_skb
);
2252 /* Check if coalescing SKBs is legal. */
2253 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2255 if (tcp_skb_pcount(skb
) > 1)
2257 /* TODO: SACK collapsing could be used to remove this condition */
2258 if (skb_shinfo(skb
)->nr_frags
!= 0)
2260 if (skb_cloned(skb
))
2262 if (skb
== tcp_send_head(sk
))
2264 /* Some heurestics for collapsing over SACK'd could be invented */
2265 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2271 /* Collapse packets in the retransmit queue to make to create
2272 * less packets on the wire. This is only done on retransmission.
2274 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2277 struct tcp_sock
*tp
= tcp_sk(sk
);
2278 struct sk_buff
*skb
= to
, *tmp
;
2281 if (!sysctl_tcp_retrans_collapse
)
2283 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2286 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2287 if (!tcp_can_collapse(sk
, skb
))
2299 /* Punt if not enough space exists in the first SKB for
2300 * the data in the second
2302 if (skb
->len
> skb_availroom(to
))
2305 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2308 tcp_collapse_retrans(sk
, to
);
2312 /* This retransmits one SKB. Policy decisions and retransmit queue
2313 * state updates are done by the caller. Returns non-zero if an
2314 * error occurred which prevented the send.
2316 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2318 struct tcp_sock
*tp
= tcp_sk(sk
);
2319 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2320 unsigned int cur_mss
;
2322 /* Inconslusive MTU probe */
2323 if (icsk
->icsk_mtup
.probe_size
) {
2324 icsk
->icsk_mtup
.probe_size
= 0;
2327 /* Do not sent more than we queued. 1/4 is reserved for possible
2328 * copying overhead: fragmentation, tunneling, mangling etc.
2330 if (atomic_read(&sk
->sk_wmem_alloc
) >
2331 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2334 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2335 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2337 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2341 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2342 return -EHOSTUNREACH
; /* Routing failure or similar. */
2344 cur_mss
= tcp_current_mss(sk
);
2346 /* If receiver has shrunk his window, and skb is out of
2347 * new window, do not retransmit it. The exception is the
2348 * case, when window is shrunk to zero. In this case
2349 * our retransmit serves as a zero window probe.
2351 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2352 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2355 if (skb
->len
> cur_mss
) {
2356 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
2357 return -ENOMEM
; /* We'll try again later. */
2359 int oldpcount
= tcp_skb_pcount(skb
);
2361 if (unlikely(oldpcount
> 1)) {
2362 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2363 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2367 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2369 /* Some Solaris stacks overoptimize and ignore the FIN on a
2370 * retransmit when old data is attached. So strip it off
2371 * since it is cheap to do so and saves bytes on the network.
2374 (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
2375 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
2376 if (!pskb_trim(skb
, 0)) {
2377 /* Reuse, even though it does some unnecessary work */
2378 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
2379 TCP_SKB_CB(skb
)->tcp_flags
);
2380 skb
->ip_summed
= CHECKSUM_NONE
;
2384 /* Make a copy, if the first transmission SKB clone we made
2385 * is still in somebody's hands, else make a clone.
2387 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2389 /* make sure skb->data is aligned on arches that require it */
2390 if (unlikely(NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3))) {
2391 struct sk_buff
*nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
,
2393 return nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2396 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2400 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2402 struct tcp_sock
*tp
= tcp_sk(sk
);
2403 int err
= __tcp_retransmit_skb(sk
, skb
);
2406 /* Update global TCP statistics. */
2407 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2409 tp
->total_retrans
++;
2411 #if FASTRETRANS_DEBUG > 0
2412 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2413 net_dbg_ratelimited("retrans_out leaked\n");
2416 if (!tp
->retrans_out
)
2417 tp
->lost_retrans_low
= tp
->snd_nxt
;
2418 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2419 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2421 /* Save stamp of the first retransmit. */
2422 if (!tp
->retrans_stamp
)
2423 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2425 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2427 /* snd_nxt is stored to detect loss of retransmitted segment,
2428 * see tcp_input.c tcp_sacktag_write_queue().
2430 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2435 /* Check if we forward retransmits are possible in the current
2436 * window/congestion state.
2438 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2440 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2441 const struct tcp_sock
*tp
= tcp_sk(sk
);
2443 /* Forward retransmissions are possible only during Recovery. */
2444 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2447 /* No forward retransmissions in Reno are possible. */
2448 if (tcp_is_reno(tp
))
2451 /* Yeah, we have to make difficult choice between forward transmission
2452 * and retransmission... Both ways have their merits...
2454 * For now we do not retransmit anything, while we have some new
2455 * segments to send. In the other cases, follow rule 3 for
2456 * NextSeg() specified in RFC3517.
2459 if (tcp_may_send_now(sk
))
2465 /* This gets called after a retransmit timeout, and the initially
2466 * retransmitted data is acknowledged. It tries to continue
2467 * resending the rest of the retransmit queue, until either
2468 * we've sent it all or the congestion window limit is reached.
2469 * If doing SACK, the first ACK which comes back for a timeout
2470 * based retransmit packet might feed us FACK information again.
2471 * If so, we use it to avoid unnecessarily retransmissions.
2473 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2475 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2476 struct tcp_sock
*tp
= tcp_sk(sk
);
2477 struct sk_buff
*skb
;
2478 struct sk_buff
*hole
= NULL
;
2481 int fwd_rexmitting
= 0;
2483 if (!tp
->packets_out
)
2487 tp
->retransmit_high
= tp
->snd_una
;
2489 if (tp
->retransmit_skb_hint
) {
2490 skb
= tp
->retransmit_skb_hint
;
2491 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2492 if (after(last_lost
, tp
->retransmit_high
))
2493 last_lost
= tp
->retransmit_high
;
2495 skb
= tcp_write_queue_head(sk
);
2496 last_lost
= tp
->snd_una
;
2499 tcp_for_write_queue_from(skb
, sk
) {
2500 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2502 if (skb
== tcp_send_head(sk
))
2504 /* we could do better than to assign each time */
2506 tp
->retransmit_skb_hint
= skb
;
2508 /* Assume this retransmit will generate
2509 * only one packet for congestion window
2510 * calculation purposes. This works because
2511 * tcp_retransmit_skb() will chop up the
2512 * packet to be MSS sized and all the
2513 * packet counting works out.
2515 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2518 if (fwd_rexmitting
) {
2520 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2522 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2524 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2525 tp
->retransmit_high
= last_lost
;
2526 if (!tcp_can_forward_retransmit(sk
))
2528 /* Backtrack if necessary to non-L'ed skb */
2536 } else if (!(sacked
& TCPCB_LOST
)) {
2537 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2542 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2543 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2544 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2546 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2549 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2552 if (tcp_retransmit_skb(sk
, skb
)) {
2553 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2556 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2558 if (tcp_in_cwnd_reduction(sk
))
2559 tp
->prr_out
+= tcp_skb_pcount(skb
);
2561 if (skb
== tcp_write_queue_head(sk
))
2562 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2563 inet_csk(sk
)->icsk_rto
,
2568 /* Send a fin. The caller locks the socket for us. This cannot be
2569 * allowed to fail queueing a FIN frame under any circumstances.
2571 void tcp_send_fin(struct sock
*sk
)
2573 struct tcp_sock
*tp
= tcp_sk(sk
);
2574 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2577 /* Optimization, tack on the FIN if we have a queue of
2578 * unsent frames. But be careful about outgoing SACKS
2581 mss_now
= tcp_current_mss(sk
);
2583 if (tcp_send_head(sk
) != NULL
) {
2584 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_FIN
;
2585 TCP_SKB_CB(skb
)->end_seq
++;
2588 /* Socket is locked, keep trying until memory is available. */
2590 skb
= alloc_skb_fclone(MAX_TCP_HEADER
,
2597 /* Reserve space for headers and prepare control bits. */
2598 skb_reserve(skb
, MAX_TCP_HEADER
);
2599 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2600 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2601 TCPHDR_ACK
| TCPHDR_FIN
);
2602 tcp_queue_skb(sk
, skb
);
2604 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2607 /* We get here when a process closes a file descriptor (either due to
2608 * an explicit close() or as a byproduct of exit()'ing) and there
2609 * was unread data in the receive queue. This behavior is recommended
2610 * by RFC 2525, section 2.17. -DaveM
2612 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2614 struct sk_buff
*skb
;
2616 /* NOTE: No TCP options attached and we never retransmit this. */
2617 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2619 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2623 /* Reserve space for headers and prepare control bits. */
2624 skb_reserve(skb
, MAX_TCP_HEADER
);
2625 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2626 TCPHDR_ACK
| TCPHDR_RST
);
2628 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2629 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2630 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2632 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2635 /* Send a crossed SYN-ACK during socket establishment.
2636 * WARNING: This routine must only be called when we have already sent
2637 * a SYN packet that crossed the incoming SYN that caused this routine
2638 * to get called. If this assumption fails then the initial rcv_wnd
2639 * and rcv_wscale values will not be correct.
2641 int tcp_send_synack(struct sock
*sk
)
2643 struct sk_buff
*skb
;
2645 skb
= tcp_write_queue_head(sk
);
2646 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2647 pr_debug("%s: wrong queue state\n", __func__
);
2650 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
2651 if (skb_cloned(skb
)) {
2652 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2655 tcp_unlink_write_queue(skb
, sk
);
2656 skb_header_release(nskb
);
2657 __tcp_add_write_queue_head(sk
, nskb
);
2658 sk_wmem_free_skb(sk
, skb
);
2659 sk
->sk_wmem_queued
+= nskb
->truesize
;
2660 sk_mem_charge(sk
, nskb
->truesize
);
2664 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
2665 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2667 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2668 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2672 * tcp_make_synack - Prepare a SYN-ACK.
2673 * sk: listener socket
2674 * dst: dst entry attached to the SYNACK
2675 * req: request_sock pointer
2676 * rvp: request_values pointer
2678 * Allocate one skb and build a SYNACK packet.
2679 * @dst is consumed : Caller should not use it again.
2681 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2682 struct request_sock
*req
,
2683 struct request_values
*rvp
,
2684 struct tcp_fastopen_cookie
*foc
)
2686 struct tcp_out_options opts
;
2687 struct tcp_extend_values
*xvp
= tcp_xv(rvp
);
2688 struct inet_request_sock
*ireq
= inet_rsk(req
);
2689 struct tcp_sock
*tp
= tcp_sk(sk
);
2690 const struct tcp_cookie_values
*cvp
= tp
->cookie_values
;
2692 struct sk_buff
*skb
;
2693 struct tcp_md5sig_key
*md5
;
2694 int tcp_header_size
;
2696 int s_data_desired
= 0;
2698 if (cvp
!= NULL
&& cvp
->s_data_constant
&& cvp
->s_data_desired
)
2699 s_data_desired
= cvp
->s_data_desired
;
2700 skb
= alloc_skb(MAX_TCP_HEADER
+ 15 + s_data_desired
,
2701 sk_gfp_atomic(sk
, GFP_ATOMIC
));
2702 if (unlikely(!skb
)) {
2706 /* Reserve space for headers. */
2707 skb_reserve(skb
, MAX_TCP_HEADER
);
2709 skb_dst_set(skb
, dst
);
2711 mss
= dst_metric_advmss(dst
);
2712 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2713 mss
= tp
->rx_opt
.user_mss
;
2715 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2717 /* Set this up on the first call only */
2718 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2720 /* limit the window selection if the user enforce a smaller rx buffer */
2721 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2722 (req
->window_clamp
> tcp_full_space(sk
) || req
->window_clamp
== 0))
2723 req
->window_clamp
= tcp_full_space(sk
);
2725 /* tcp_full_space because it is guaranteed to be the first packet */
2726 tcp_select_initial_window(tcp_full_space(sk
),
2727 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2732 dst_metric(dst
, RTAX_INITRWND
));
2733 ireq
->rcv_wscale
= rcv_wscale
;
2736 memset(&opts
, 0, sizeof(opts
));
2737 #ifdef CONFIG_SYN_COOKIES
2738 if (unlikely(req
->cookie_ts
))
2739 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2742 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2743 tcp_header_size
= tcp_synack_options(sk
, req
, mss
,
2744 skb
, &opts
, &md5
, xvp
, foc
)
2747 skb_push(skb
, tcp_header_size
);
2748 skb_reset_transport_header(skb
);
2751 memset(th
, 0, sizeof(struct tcphdr
));
2754 TCP_ECN_make_synack(req
, th
);
2755 th
->source
= ireq
->loc_port
;
2756 th
->dest
= ireq
->rmt_port
;
2757 /* Setting of flags are superfluous here for callers (and ECE is
2758 * not even correctly set)
2760 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2761 TCPHDR_SYN
| TCPHDR_ACK
);
2763 if (OPTION_COOKIE_EXTENSION
& opts
.options
) {
2764 if (s_data_desired
) {
2765 u8
*buf
= skb_put(skb
, s_data_desired
);
2767 /* copy data directly from the listening socket. */
2768 memcpy(buf
, cvp
->s_data_payload
, s_data_desired
);
2769 TCP_SKB_CB(skb
)->end_seq
+= s_data_desired
;
2772 if (opts
.hash_size
> 0) {
2773 __u32 workspace
[SHA_WORKSPACE_WORDS
];
2774 u32
*mess
= &xvp
->cookie_bakery
[COOKIE_DIGEST_WORDS
];
2775 u32
*tail
= &mess
[COOKIE_MESSAGE_WORDS
-1];
2777 /* Secret recipe depends on the Timestamp, (future)
2778 * Sequence and Acknowledgment Numbers, Initiator
2779 * Cookie, and others handled by IP variant caller.
2781 *tail
-- ^= opts
.tsval
;
2782 *tail
-- ^= tcp_rsk(req
)->rcv_isn
+ 1;
2783 *tail
-- ^= TCP_SKB_CB(skb
)->seq
+ 1;
2786 *tail
-- ^= (((__force u32
)th
->dest
<< 16) | (__force u32
)th
->source
);
2787 *tail
-- ^= (u32
)(unsigned long)cvp
; /* per sockopt */
2789 sha_transform((__u32
*)&xvp
->cookie_bakery
[0],
2792 opts
.hash_location
=
2793 (__u8
*)&xvp
->cookie_bakery
[0];
2797 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2798 /* XXX data is queued and acked as is. No buffer/window check */
2799 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
2801 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2802 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2803 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2804 th
->doff
= (tcp_header_size
>> 2);
2805 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
, tcp_skb_pcount(skb
));
2807 #ifdef CONFIG_TCP_MD5SIG
2808 /* Okay, we have all we need - do the md5 hash if needed */
2810 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2811 md5
, NULL
, req
, skb
);
2817 EXPORT_SYMBOL(tcp_make_synack
);
2819 /* Do all connect socket setups that can be done AF independent. */
2820 void tcp_connect_init(struct sock
*sk
)
2822 const struct dst_entry
*dst
= __sk_dst_get(sk
);
2823 struct tcp_sock
*tp
= tcp_sk(sk
);
2826 /* We'll fix this up when we get a response from the other end.
2827 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2829 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2830 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2832 #ifdef CONFIG_TCP_MD5SIG
2833 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2834 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2837 /* If user gave his TCP_MAXSEG, record it to clamp */
2838 if (tp
->rx_opt
.user_mss
)
2839 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2842 tcp_sync_mss(sk
, dst_mtu(dst
));
2844 if (!tp
->window_clamp
)
2845 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2846 tp
->advmss
= dst_metric_advmss(dst
);
2847 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2848 tp
->advmss
= tp
->rx_opt
.user_mss
;
2850 tcp_initialize_rcv_mss(sk
);
2852 /* limit the window selection if the user enforce a smaller rx buffer */
2853 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2854 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
2855 tp
->window_clamp
= tcp_full_space(sk
);
2857 tcp_select_initial_window(tcp_full_space(sk
),
2858 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2861 sysctl_tcp_window_scaling
,
2863 dst_metric(dst
, RTAX_INITRWND
));
2865 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2866 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2869 sock_reset_flag(sk
, SOCK_DONE
);
2872 tp
->snd_una
= tp
->write_seq
;
2873 tp
->snd_sml
= tp
->write_seq
;
2874 tp
->snd_up
= tp
->write_seq
;
2875 tp
->snd_nxt
= tp
->write_seq
;
2877 if (likely(!tp
->repair
))
2879 tp
->rcv_wup
= tp
->rcv_nxt
;
2880 tp
->copied_seq
= tp
->rcv_nxt
;
2882 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2883 inet_csk(sk
)->icsk_retransmits
= 0;
2884 tcp_clear_retrans(tp
);
2887 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
2889 struct tcp_sock
*tp
= tcp_sk(sk
);
2890 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
2892 tcb
->end_seq
+= skb
->len
;
2893 skb_header_release(skb
);
2894 __tcp_add_write_queue_tail(sk
, skb
);
2895 sk
->sk_wmem_queued
+= skb
->truesize
;
2896 sk_mem_charge(sk
, skb
->truesize
);
2897 tp
->write_seq
= tcb
->end_seq
;
2898 tp
->packets_out
+= tcp_skb_pcount(skb
);
2901 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2902 * queue a data-only packet after the regular SYN, such that regular SYNs
2903 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2904 * only the SYN sequence, the data are retransmitted in the first ACK.
2905 * If cookie is not cached or other error occurs, falls back to send a
2906 * regular SYN with Fast Open cookie request option.
2908 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
2910 struct tcp_sock
*tp
= tcp_sk(sk
);
2911 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
2912 int syn_loss
= 0, space
, i
, err
= 0, iovlen
= fo
->data
->msg_iovlen
;
2913 struct sk_buff
*syn_data
= NULL
, *data
;
2914 unsigned long last_syn_loss
= 0;
2916 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
2917 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
2918 &syn_loss
, &last_syn_loss
);
2919 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2921 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
2922 fo
->cookie
.len
= -1;
2926 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
2927 fo
->cookie
.len
= -1;
2928 else if (fo
->cookie
.len
<= 0)
2931 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2932 * user-MSS. Reserve maximum option space for middleboxes that add
2933 * private TCP options. The cost is reduced data space in SYN :(
2935 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
2936 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2937 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
2938 MAX_TCP_OPTION_SPACE
;
2940 syn_data
= skb_copy_expand(syn
, skb_headroom(syn
), space
,
2942 if (syn_data
== NULL
)
2945 for (i
= 0; i
< iovlen
&& syn_data
->len
< space
; ++i
) {
2946 struct iovec
*iov
= &fo
->data
->msg_iov
[i
];
2947 unsigned char __user
*from
= iov
->iov_base
;
2948 int len
= iov
->iov_len
;
2950 if (syn_data
->len
+ len
> space
)
2951 len
= space
- syn_data
->len
;
2952 else if (i
+ 1 == iovlen
)
2953 /* No more data pending in inet_wait_for_connect() */
2956 if (skb_add_data(syn_data
, from
, len
))
2960 /* Queue a data-only packet after the regular SYN for retransmission */
2961 data
= pskb_copy(syn_data
, sk
->sk_allocation
);
2964 TCP_SKB_CB(data
)->seq
++;
2965 TCP_SKB_CB(data
)->tcp_flags
&= ~TCPHDR_SYN
;
2966 TCP_SKB_CB(data
)->tcp_flags
= (TCPHDR_ACK
|TCPHDR_PSH
);
2967 tcp_connect_queue_skb(sk
, data
);
2968 fo
->copied
= data
->len
;
2970 if (tcp_transmit_skb(sk
, syn_data
, 0, sk
->sk_allocation
) == 0) {
2971 tp
->syn_data
= (fo
->copied
> 0);
2972 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
2978 /* Send a regular SYN with Fast Open cookie request option */
2979 if (fo
->cookie
.len
> 0)
2981 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
2983 tp
->syn_fastopen
= 0;
2984 kfree_skb(syn_data
);
2986 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
2990 /* Build a SYN and send it off. */
2991 int tcp_connect(struct sock
*sk
)
2993 struct tcp_sock
*tp
= tcp_sk(sk
);
2994 struct sk_buff
*buff
;
2997 tcp_connect_init(sk
);
2999 if (unlikely(tp
->repair
)) {
3000 tcp_finish_connect(sk
, NULL
);
3004 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
3005 if (unlikely(buff
== NULL
))
3008 /* Reserve space for headers. */
3009 skb_reserve(buff
, MAX_TCP_HEADER
);
3011 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3012 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
3013 tcp_connect_queue_skb(sk
, buff
);
3014 TCP_ECN_send_syn(sk
, buff
);
3016 /* Send off SYN; include data in Fast Open. */
3017 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3018 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3019 if (err
== -ECONNREFUSED
)
3022 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3023 * in order to make this packet get counted in tcpOutSegs.
3025 tp
->snd_nxt
= tp
->write_seq
;
3026 tp
->pushed_seq
= tp
->write_seq
;
3027 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3029 /* Timer for repeating the SYN until an answer. */
3030 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3031 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3034 EXPORT_SYMBOL(tcp_connect
);
3036 /* Send out a delayed ack, the caller does the policy checking
3037 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3040 void tcp_send_delayed_ack(struct sock
*sk
)
3042 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3043 int ato
= icsk
->icsk_ack
.ato
;
3044 unsigned long timeout
;
3046 if (ato
> TCP_DELACK_MIN
) {
3047 const struct tcp_sock
*tp
= tcp_sk(sk
);
3048 int max_ato
= HZ
/ 2;
3050 if (icsk
->icsk_ack
.pingpong
||
3051 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3052 max_ato
= TCP_DELACK_MAX
;
3054 /* Slow path, intersegment interval is "high". */
3056 /* If some rtt estimate is known, use it to bound delayed ack.
3057 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3061 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
3067 ato
= min(ato
, max_ato
);
3070 /* Stay within the limit we were given */
3071 timeout
= jiffies
+ ato
;
3073 /* Use new timeout only if there wasn't a older one earlier. */
3074 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3075 /* If delack timer was blocked or is about to expire,
3078 if (icsk
->icsk_ack
.blocked
||
3079 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3084 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3085 timeout
= icsk
->icsk_ack
.timeout
;
3087 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3088 icsk
->icsk_ack
.timeout
= timeout
;
3089 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3092 /* This routine sends an ack and also updates the window. */
3093 void tcp_send_ack(struct sock
*sk
)
3095 struct sk_buff
*buff
;
3097 /* If we have been reset, we may not send again. */
3098 if (sk
->sk_state
== TCP_CLOSE
)
3101 /* We are not putting this on the write queue, so
3102 * tcp_transmit_skb() will set the ownership to this
3105 buff
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3107 inet_csk_schedule_ack(sk
);
3108 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3109 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3110 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3114 /* Reserve space for headers and prepare control bits. */
3115 skb_reserve(buff
, MAX_TCP_HEADER
);
3116 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3118 /* Send it off, this clears delayed acks for us. */
3119 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
3120 tcp_transmit_skb(sk
, buff
, 0, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3123 /* This routine sends a packet with an out of date sequence
3124 * number. It assumes the other end will try to ack it.
3126 * Question: what should we make while urgent mode?
3127 * 4.4BSD forces sending single byte of data. We cannot send
3128 * out of window data, because we have SND.NXT==SND.MAX...
3130 * Current solution: to send TWO zero-length segments in urgent mode:
3131 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3132 * out-of-date with SND.UNA-1 to probe window.
3134 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
3136 struct tcp_sock
*tp
= tcp_sk(sk
);
3137 struct sk_buff
*skb
;
3139 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3140 skb
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3144 /* Reserve space for headers and set control bits. */
3145 skb_reserve(skb
, MAX_TCP_HEADER
);
3146 /* Use a previous sequence. This should cause the other
3147 * end to send an ack. Don't queue or clone SKB, just
3150 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3151 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3152 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
3155 void tcp_send_window_probe(struct sock
*sk
)
3157 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3158 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3159 tcp_sk(sk
)->snd_nxt
= tcp_sk(sk
)->write_seq
;
3160 tcp_xmit_probe_skb(sk
, 0);
3164 /* Initiate keepalive or window probe from timer. */
3165 int tcp_write_wakeup(struct sock
*sk
)
3167 struct tcp_sock
*tp
= tcp_sk(sk
);
3168 struct sk_buff
*skb
;
3170 if (sk
->sk_state
== TCP_CLOSE
)
3173 if ((skb
= tcp_send_head(sk
)) != NULL
&&
3174 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3176 unsigned int mss
= tcp_current_mss(sk
);
3177 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3179 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3180 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3182 /* We are probing the opening of a window
3183 * but the window size is != 0
3184 * must have been a result SWS avoidance ( sender )
3186 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3188 seg_size
= min(seg_size
, mss
);
3189 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3190 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
3192 } else if (!tcp_skb_pcount(skb
))
3193 tcp_set_skb_tso_segs(sk
, skb
, mss
);
3195 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3196 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3197 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3199 tcp_event_new_data_sent(sk
, skb
);
3202 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3203 tcp_xmit_probe_skb(sk
, 1);
3204 return tcp_xmit_probe_skb(sk
, 0);
3208 /* A window probe timeout has occurred. If window is not closed send
3209 * a partial packet else a zero probe.
3211 void tcp_send_probe0(struct sock
*sk
)
3213 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3214 struct tcp_sock
*tp
= tcp_sk(sk
);
3217 err
= tcp_write_wakeup(sk
);
3219 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3220 /* Cancel probe timer, if it is not required. */
3221 icsk
->icsk_probes_out
= 0;
3222 icsk
->icsk_backoff
= 0;
3227 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
3228 icsk
->icsk_backoff
++;
3229 icsk
->icsk_probes_out
++;
3230 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3231 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3234 /* If packet was not sent due to local congestion,
3235 * do not backoff and do not remember icsk_probes_out.
3236 * Let local senders to fight for local resources.
3238 * Use accumulated backoff yet.
3240 if (!icsk
->icsk_probes_out
)
3241 icsk
->icsk_probes_out
= 1;
3242 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3243 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
3244 TCP_RESOURCE_PROBE_INTERVAL
),