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 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
69 int push_one
, gfp_t gfp
);
71 /* Account for new data that has been sent to the network. */
72 static void tcp_event_new_data_sent(struct sock
*sk
, const struct sk_buff
*skb
)
74 struct inet_connection_sock
*icsk
= inet_csk(sk
);
75 struct tcp_sock
*tp
= tcp_sk(sk
);
76 unsigned int prior_packets
= tp
->packets_out
;
78 tcp_advance_send_head(sk
, skb
);
79 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
81 tp
->packets_out
+= tcp_skb_pcount(skb
);
82 if (!prior_packets
|| icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
83 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
88 /* SND.NXT, if window was not shrunk.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
94 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
96 const struct tcp_sock
*tp
= tcp_sk(sk
);
98 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
101 return tcp_wnd_end(tp
);
104 /* Calculate mss to advertise in SYN segment.
105 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
107 * 1. It is independent of path mtu.
108 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
109 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
110 * attached devices, because some buggy hosts are confused by
112 * 4. We do not make 3, we advertise MSS, calculated from first
113 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
114 * This may be overridden via information stored in routing table.
115 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
116 * probably even Jumbo".
118 static __u16
tcp_advertise_mss(struct sock
*sk
)
120 struct tcp_sock
*tp
= tcp_sk(sk
);
121 const struct dst_entry
*dst
= __sk_dst_get(sk
);
122 int mss
= tp
->advmss
;
125 unsigned int metric
= dst_metric_advmss(dst
);
136 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
137 * This is the first part of cwnd validation mechanism. */
138 static void tcp_cwnd_restart(struct sock
*sk
, const struct dst_entry
*dst
)
140 struct tcp_sock
*tp
= tcp_sk(sk
);
141 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
142 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
143 u32 cwnd
= tp
->snd_cwnd
;
145 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
147 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
148 restart_cwnd
= min(restart_cwnd
, cwnd
);
150 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
152 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
153 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
154 tp
->snd_cwnd_used
= 0;
157 /* Congestion state accounting after a packet has been sent. */
158 static void tcp_event_data_sent(struct tcp_sock
*tp
,
161 struct inet_connection_sock
*icsk
= inet_csk(sk
);
162 const u32 now
= tcp_time_stamp
;
164 if (sysctl_tcp_slow_start_after_idle
&&
165 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
166 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
170 /* If it is a reply for ato after last received
171 * packet, enter pingpong mode.
173 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
174 icsk
->icsk_ack
.pingpong
= 1;
177 /* Account for an ACK we sent. */
178 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
180 tcp_dec_quickack_mode(sk
, pkts
);
181 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
184 /* Determine a window scaling and initial window to offer.
185 * Based on the assumption that the given amount of space
186 * will be offered. Store the results in the tp structure.
187 * NOTE: for smooth operation initial space offering should
188 * be a multiple of mss if possible. We assume here that mss >= 1.
189 * This MUST be enforced by all callers.
191 void tcp_select_initial_window(int __space
, __u32 mss
,
192 __u32
*rcv_wnd
, __u32
*window_clamp
,
193 int wscale_ok
, __u8
*rcv_wscale
,
196 unsigned int space
= (__space
< 0 ? 0 : __space
);
198 /* If no clamp set the clamp to the max possible scaled window */
199 if (*window_clamp
== 0)
200 (*window_clamp
) = (65535 << 14);
201 space
= min(*window_clamp
, space
);
203 /* Quantize space offering to a multiple of mss if possible. */
205 space
= (space
/ mss
) * mss
;
207 /* NOTE: offering an initial window larger than 32767
208 * will break some buggy TCP stacks. If the admin tells us
209 * it is likely we could be speaking with such a buggy stack
210 * we will truncate our initial window offering to 32K-1
211 * unless the remote has sent us a window scaling option,
212 * which we interpret as a sign the remote TCP is not
213 * misinterpreting the window field as a signed quantity.
215 if (sysctl_tcp_workaround_signed_windows
)
216 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
222 /* Set window scaling on max possible window
223 * See RFC1323 for an explanation of the limit to 14
225 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
226 space
= min_t(u32
, space
, *window_clamp
);
227 while (space
> 65535 && (*rcv_wscale
) < 14) {
233 /* Set initial window to a value enough for senders starting with
234 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
235 * a limit on the initial window when mss is larger than 1460.
237 if (mss
> (1 << *rcv_wscale
)) {
238 int init_cwnd
= TCP_DEFAULT_INIT_RCVWND
;
241 max_t(u32
, (1460 * TCP_DEFAULT_INIT_RCVWND
) / mss
, 2);
242 /* when initializing use the value from init_rcv_wnd
243 * rather than the default from above
246 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
248 *rcv_wnd
= min(*rcv_wnd
, init_cwnd
* mss
);
251 /* Set the clamp no higher than max representable value */
252 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
254 EXPORT_SYMBOL(tcp_select_initial_window
);
256 /* Chose a new window to advertise, update state in tcp_sock for the
257 * socket, and return result with RFC1323 scaling applied. The return
258 * value can be stuffed directly into th->window for an outgoing
261 static u16
tcp_select_window(struct sock
*sk
)
263 struct tcp_sock
*tp
= tcp_sk(sk
);
264 u32 cur_win
= tcp_receive_window(tp
);
265 u32 new_win
= __tcp_select_window(sk
);
267 /* Never shrink the offered window */
268 if (new_win
< cur_win
) {
269 /* Danger Will Robinson!
270 * Don't update rcv_wup/rcv_wnd here or else
271 * we will not be able to advertise a zero
272 * window in time. --DaveM
274 * Relax Will Robinson.
276 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
278 tp
->rcv_wnd
= new_win
;
279 tp
->rcv_wup
= tp
->rcv_nxt
;
281 /* Make sure we do not exceed the maximum possible
284 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
285 new_win
= min(new_win
, MAX_TCP_WINDOW
);
287 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
289 /* RFC1323 scaling applied */
290 new_win
>>= tp
->rx_opt
.rcv_wscale
;
292 /* If we advertise zero window, disable fast path. */
299 /* Packet ECN state for a SYN-ACK */
300 static inline void TCP_ECN_send_synack(const struct tcp_sock
*tp
, struct sk_buff
*skb
)
302 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
303 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
304 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
307 /* Packet ECN state for a SYN. */
308 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
310 struct tcp_sock
*tp
= tcp_sk(sk
);
313 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1) {
314 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
315 tp
->ecn_flags
= TCP_ECN_OK
;
319 static __inline__
void
320 TCP_ECN_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
322 if (inet_rsk(req
)->ecn_ok
)
326 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
329 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
332 struct tcp_sock
*tp
= tcp_sk(sk
);
334 if (tp
->ecn_flags
& TCP_ECN_OK
) {
335 /* Not-retransmitted data segment: set ECT and inject CWR. */
336 if (skb
->len
!= tcp_header_len
&&
337 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
339 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
340 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
341 tcp_hdr(skb
)->cwr
= 1;
342 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
345 /* ACK or retransmitted segment: clear ECT|CE */
346 INET_ECN_dontxmit(sk
);
348 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
349 tcp_hdr(skb
)->ece
= 1;
353 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
354 * auto increment end seqno.
356 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
358 skb
->ip_summed
= CHECKSUM_PARTIAL
;
361 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
362 TCP_SKB_CB(skb
)->sacked
= 0;
364 skb_shinfo(skb
)->gso_segs
= 1;
365 skb_shinfo(skb
)->gso_size
= 0;
366 skb_shinfo(skb
)->gso_type
= 0;
368 TCP_SKB_CB(skb
)->seq
= seq
;
369 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
371 TCP_SKB_CB(skb
)->end_seq
= seq
;
374 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
376 return tp
->snd_una
!= tp
->snd_up
;
379 #define OPTION_SACK_ADVERTISE (1 << 0)
380 #define OPTION_TS (1 << 1)
381 #define OPTION_MD5 (1 << 2)
382 #define OPTION_WSCALE (1 << 3)
383 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
385 struct tcp_out_options
{
386 u16 options
; /* bit field of OPTION_* */
387 u16 mss
; /* 0 to disable */
388 u8 ws
; /* window scale, 0 to disable */
389 u8 num_sack_blocks
; /* number of SACK blocks to include */
390 u8 hash_size
; /* bytes in hash_location */
391 __u8
*hash_location
; /* temporary pointer, overloaded */
392 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
393 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
396 /* Write previously computed TCP options to the packet.
398 * Beware: Something in the Internet is very sensitive to the ordering of
399 * TCP options, we learned this through the hard way, so be careful here.
400 * Luckily we can at least blame others for their non-compliance but from
401 * inter-operatibility perspective it seems that we're somewhat stuck with
402 * the ordering which we have been using if we want to keep working with
403 * those broken things (not that it currently hurts anybody as there isn't
404 * particular reason why the ordering would need to be changed).
406 * At least SACK_PERM as the first option is known to lead to a disaster
407 * (but it may well be that other scenarios fail similarly).
409 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
410 struct tcp_out_options
*opts
)
412 u16 options
= opts
->options
; /* mungable copy */
414 if (unlikely(OPTION_MD5
& options
)) {
415 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
416 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
417 /* overload cookie hash location */
418 opts
->hash_location
= (__u8
*)ptr
;
422 if (unlikely(opts
->mss
)) {
423 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
424 (TCPOLEN_MSS
<< 16) |
428 if (likely(OPTION_TS
& options
)) {
429 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
430 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
431 (TCPOLEN_SACK_PERM
<< 16) |
432 (TCPOPT_TIMESTAMP
<< 8) |
434 options
&= ~OPTION_SACK_ADVERTISE
;
436 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
438 (TCPOPT_TIMESTAMP
<< 8) |
441 *ptr
++ = htonl(opts
->tsval
);
442 *ptr
++ = htonl(opts
->tsecr
);
445 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
446 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
448 (TCPOPT_SACK_PERM
<< 8) |
452 if (unlikely(OPTION_WSCALE
& options
)) {
453 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
454 (TCPOPT_WINDOW
<< 16) |
455 (TCPOLEN_WINDOW
<< 8) |
459 if (unlikely(opts
->num_sack_blocks
)) {
460 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
461 tp
->duplicate_sack
: tp
->selective_acks
;
464 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
467 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
468 TCPOLEN_SACK_PERBLOCK
)));
470 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
472 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
473 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
476 tp
->rx_opt
.dsack
= 0;
479 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
480 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
482 *ptr
++ = htonl((TCPOPT_EXP
<< 24) |
483 ((TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
) << 16) |
484 TCPOPT_FASTOPEN_MAGIC
);
486 memcpy(ptr
, foc
->val
, foc
->len
);
487 if ((foc
->len
& 3) == 2) {
488 u8
*align
= ((u8
*)ptr
) + foc
->len
;
489 align
[0] = align
[1] = TCPOPT_NOP
;
491 ptr
+= (foc
->len
+ 3) >> 2;
495 /* Compute TCP options for SYN packets. This is not the final
496 * network wire format yet.
498 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
499 struct tcp_out_options
*opts
,
500 struct tcp_md5sig_key
**md5
)
502 struct tcp_sock
*tp
= tcp_sk(sk
);
503 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
504 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
506 #ifdef CONFIG_TCP_MD5SIG
507 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
509 opts
->options
|= OPTION_MD5
;
510 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
516 /* We always get an MSS option. The option bytes which will be seen in
517 * normal data packets should timestamps be used, must be in the MSS
518 * advertised. But we subtract them from tp->mss_cache so that
519 * calculations in tcp_sendmsg are simpler etc. So account for this
520 * fact here if necessary. If we don't do this correctly, as a
521 * receiver we won't recognize data packets as being full sized when we
522 * should, and thus we won't abide by the delayed ACK rules correctly.
523 * SACKs don't matter, we never delay an ACK when we have any of those
525 opts
->mss
= tcp_advertise_mss(sk
);
526 remaining
-= TCPOLEN_MSS_ALIGNED
;
528 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
529 opts
->options
|= OPTION_TS
;
530 opts
->tsval
= TCP_SKB_CB(skb
)->when
+ tp
->tsoffset
;
531 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
532 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
534 if (likely(sysctl_tcp_window_scaling
)) {
535 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
536 opts
->options
|= OPTION_WSCALE
;
537 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
539 if (likely(sysctl_tcp_sack
)) {
540 opts
->options
|= OPTION_SACK_ADVERTISE
;
541 if (unlikely(!(OPTION_TS
& opts
->options
)))
542 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
545 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
546 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ fastopen
->cookie
.len
;
547 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
548 if (remaining
>= need
) {
549 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
550 opts
->fastopen_cookie
= &fastopen
->cookie
;
552 tp
->syn_fastopen
= 1;
556 return MAX_TCP_OPTION_SPACE
- remaining
;
559 /* Set up TCP options for SYN-ACKs. */
560 static unsigned int tcp_synack_options(struct sock
*sk
,
561 struct request_sock
*req
,
562 unsigned int mss
, struct sk_buff
*skb
,
563 struct tcp_out_options
*opts
,
564 struct tcp_md5sig_key
**md5
,
565 struct tcp_fastopen_cookie
*foc
)
567 struct inet_request_sock
*ireq
= inet_rsk(req
);
568 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
570 #ifdef CONFIG_TCP_MD5SIG
571 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
573 opts
->options
|= OPTION_MD5
;
574 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
576 /* We can't fit any SACK blocks in a packet with MD5 + TS
577 * options. There was discussion about disabling SACK
578 * rather than TS in order to fit in better with old,
579 * buggy kernels, but that was deemed to be unnecessary.
581 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
587 /* We always send an MSS option. */
589 remaining
-= TCPOLEN_MSS_ALIGNED
;
591 if (likely(ireq
->wscale_ok
)) {
592 opts
->ws
= ireq
->rcv_wscale
;
593 opts
->options
|= OPTION_WSCALE
;
594 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
596 if (likely(ireq
->tstamp_ok
)) {
597 opts
->options
|= OPTION_TS
;
598 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
599 opts
->tsecr
= req
->ts_recent
;
600 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
602 if (likely(ireq
->sack_ok
)) {
603 opts
->options
|= OPTION_SACK_ADVERTISE
;
604 if (unlikely(!ireq
->tstamp_ok
))
605 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
608 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
609 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
610 if (remaining
>= need
) {
611 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
612 opts
->fastopen_cookie
= foc
;
617 return MAX_TCP_OPTION_SPACE
- remaining
;
620 /* Compute TCP options for ESTABLISHED sockets. This is not the
621 * final wire format yet.
623 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
624 struct tcp_out_options
*opts
,
625 struct tcp_md5sig_key
**md5
)
627 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
628 struct tcp_sock
*tp
= tcp_sk(sk
);
629 unsigned int size
= 0;
630 unsigned int eff_sacks
;
632 #ifdef CONFIG_TCP_MD5SIG
633 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
634 if (unlikely(*md5
)) {
635 opts
->options
|= OPTION_MD5
;
636 size
+= TCPOLEN_MD5SIG_ALIGNED
;
642 if (likely(tp
->rx_opt
.tstamp_ok
)) {
643 opts
->options
|= OPTION_TS
;
644 opts
->tsval
= tcb
? tcb
->when
+ tp
->tsoffset
: 0;
645 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
646 size
+= TCPOLEN_TSTAMP_ALIGNED
;
649 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
650 if (unlikely(eff_sacks
)) {
651 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
652 opts
->num_sack_blocks
=
653 min_t(unsigned int, eff_sacks
,
654 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
655 TCPOLEN_SACK_PERBLOCK
);
656 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
657 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
664 /* TCP SMALL QUEUES (TSQ)
666 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
667 * to reduce RTT and bufferbloat.
668 * We do this using a special skb destructor (tcp_wfree).
670 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
671 * needs to be reallocated in a driver.
672 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
674 * Since transmit from skb destructor is forbidden, we use a tasklet
675 * to process all sockets that eventually need to send more skbs.
676 * We use one tasklet per cpu, with its own queue of sockets.
679 struct tasklet_struct tasklet
;
680 struct list_head head
; /* queue of tcp sockets */
682 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
684 static void tcp_tsq_handler(struct sock
*sk
)
686 if ((1 << sk
->sk_state
) &
687 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
688 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
))
689 tcp_write_xmit(sk
, tcp_current_mss(sk
), 0, 0, GFP_ATOMIC
);
692 * One tasklest per cpu tries to send more skbs.
693 * We run in tasklet context but need to disable irqs when
694 * transfering tsq->head because tcp_wfree() might
695 * interrupt us (non NAPI drivers)
697 static void tcp_tasklet_func(unsigned long data
)
699 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
702 struct list_head
*q
, *n
;
706 local_irq_save(flags
);
707 list_splice_init(&tsq
->head
, &list
);
708 local_irq_restore(flags
);
710 list_for_each_safe(q
, n
, &list
) {
711 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
712 list_del(&tp
->tsq_node
);
714 sk
= (struct sock
*)tp
;
717 if (!sock_owned_by_user(sk
)) {
720 /* defer the work to tcp_release_cb() */
721 set_bit(TCP_TSQ_DEFERRED
, &tp
->tsq_flags
);
725 clear_bit(TSQ_QUEUED
, &tp
->tsq_flags
);
730 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
731 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
732 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
733 (1UL << TCP_MTU_REDUCED_DEFERRED))
735 * tcp_release_cb - tcp release_sock() callback
738 * called from release_sock() to perform protocol dependent
739 * actions before socket release.
741 void tcp_release_cb(struct sock
*sk
)
743 struct tcp_sock
*tp
= tcp_sk(sk
);
744 unsigned long flags
, nflags
;
746 /* perform an atomic operation only if at least one flag is set */
748 flags
= tp
->tsq_flags
;
749 if (!(flags
& TCP_DEFERRED_ALL
))
751 nflags
= flags
& ~TCP_DEFERRED_ALL
;
752 } while (cmpxchg(&tp
->tsq_flags
, flags
, nflags
) != flags
);
754 if (flags
& (1UL << TCP_TSQ_DEFERRED
))
757 if (flags
& (1UL << TCP_WRITE_TIMER_DEFERRED
)) {
758 tcp_write_timer_handler(sk
);
761 if (flags
& (1UL << TCP_DELACK_TIMER_DEFERRED
)) {
762 tcp_delack_timer_handler(sk
);
765 if (flags
& (1UL << TCP_MTU_REDUCED_DEFERRED
)) {
766 sk
->sk_prot
->mtu_reduced(sk
);
770 EXPORT_SYMBOL(tcp_release_cb
);
772 void __init
tcp_tasklet_init(void)
776 for_each_possible_cpu(i
) {
777 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
779 INIT_LIST_HEAD(&tsq
->head
);
780 tasklet_init(&tsq
->tasklet
,
787 * Write buffer destructor automatically called from kfree_skb.
788 * We cant xmit new skbs from this context, as we might already
791 void tcp_wfree(struct sk_buff
*skb
)
793 struct sock
*sk
= skb
->sk
;
794 struct tcp_sock
*tp
= tcp_sk(sk
);
796 if (test_and_clear_bit(TSQ_THROTTLED
, &tp
->tsq_flags
) &&
797 !test_and_set_bit(TSQ_QUEUED
, &tp
->tsq_flags
)) {
799 struct tsq_tasklet
*tsq
;
801 /* Keep a ref on socket.
802 * This last ref will be released in tcp_tasklet_func()
804 atomic_sub(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
806 /* queue this socket to tasklet queue */
807 local_irq_save(flags
);
808 tsq
= &__get_cpu_var(tsq_tasklet
);
809 list_add(&tp
->tsq_node
, &tsq
->head
);
810 tasklet_schedule(&tsq
->tasklet
);
811 local_irq_restore(flags
);
817 /* This routine actually transmits TCP packets queued in by
818 * tcp_do_sendmsg(). This is used by both the initial
819 * transmission and possible later retransmissions.
820 * All SKB's seen here are completely headerless. It is our
821 * job to build the TCP header, and pass the packet down to
822 * IP so it can do the same plus pass the packet off to the
825 * We are working here with either a clone of the original
826 * SKB, or a fresh unique copy made by the retransmit engine.
828 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
831 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
832 struct inet_sock
*inet
;
834 struct tcp_skb_cb
*tcb
;
835 struct tcp_out_options opts
;
836 unsigned int tcp_options_size
, tcp_header_size
;
837 struct tcp_md5sig_key
*md5
;
841 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
843 /* If congestion control is doing timestamping, we must
844 * take such a timestamp before we potentially clone/copy.
846 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
847 __net_timestamp(skb
);
849 if (likely(clone_it
)) {
850 const struct sk_buff
*fclone
= skb
+ 1;
852 if (unlikely(skb
->fclone
== SKB_FCLONE_ORIG
&&
853 fclone
->fclone
== SKB_FCLONE_CLONE
))
854 NET_INC_STATS_BH(sock_net(sk
),
855 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
857 if (unlikely(skb_cloned(skb
)))
858 skb
= pskb_copy(skb
, gfp_mask
);
860 skb
= skb_clone(skb
, gfp_mask
);
867 tcb
= TCP_SKB_CB(skb
);
868 memset(&opts
, 0, sizeof(opts
));
870 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
871 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
873 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
875 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
877 if (tcp_packets_in_flight(tp
) == 0)
878 tcp_ca_event(sk
, CA_EVENT_TX_START
);
880 /* if no packet is in qdisc/device queue, then allow XPS to select
883 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) == 0;
885 skb_push(skb
, tcp_header_size
);
886 skb_reset_transport_header(skb
);
890 skb
->destructor
= tcp_wfree
;
891 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
893 /* Build TCP header and checksum it. */
895 th
->source
= inet
->inet_sport
;
896 th
->dest
= inet
->inet_dport
;
897 th
->seq
= htonl(tcb
->seq
);
898 th
->ack_seq
= htonl(tp
->rcv_nxt
);
899 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
902 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
903 /* RFC1323: The window in SYN & SYN/ACK segments
906 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
908 th
->window
= htons(tcp_select_window(sk
));
913 /* The urg_mode check is necessary during a below snd_una win probe */
914 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
915 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
916 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
918 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
919 th
->urg_ptr
= htons(0xFFFF);
924 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
925 if (likely((tcb
->tcp_flags
& TCPHDR_SYN
) == 0))
926 TCP_ECN_send(sk
, skb
, tcp_header_size
);
928 #ifdef CONFIG_TCP_MD5SIG
929 /* Calculate the MD5 hash, as we have all we need now */
931 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
932 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
937 icsk
->icsk_af_ops
->send_check(sk
, skb
);
939 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
940 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
942 if (skb
->len
!= tcp_header_size
)
943 tcp_event_data_sent(tp
, sk
);
945 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
946 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
947 tcp_skb_pcount(skb
));
949 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, &inet
->cork
.fl
);
950 if (likely(err
<= 0))
953 tcp_enter_cwr(sk
, 1);
955 return net_xmit_eval(err
);
958 /* This routine just queues the buffer for sending.
960 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
961 * otherwise socket can stall.
963 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
965 struct tcp_sock
*tp
= tcp_sk(sk
);
967 /* Advance write_seq and place onto the write_queue. */
968 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
969 skb_header_release(skb
);
970 tcp_add_write_queue_tail(sk
, skb
);
971 sk
->sk_wmem_queued
+= skb
->truesize
;
972 sk_mem_charge(sk
, skb
->truesize
);
975 /* Initialize TSO segments for a packet. */
976 static void tcp_set_skb_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
977 unsigned int mss_now
)
979 /* Make sure we own this skb before messing gso_size/gso_segs */
980 WARN_ON_ONCE(skb_cloned(skb
));
982 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
) ||
983 skb
->ip_summed
== CHECKSUM_NONE
) {
984 /* Avoid the costly divide in the normal
987 skb_shinfo(skb
)->gso_segs
= 1;
988 skb_shinfo(skb
)->gso_size
= 0;
989 skb_shinfo(skb
)->gso_type
= 0;
991 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
992 skb_shinfo(skb
)->gso_size
= mss_now
;
993 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
997 /* When a modification to fackets out becomes necessary, we need to check
998 * skb is counted to fackets_out or not.
1000 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1003 struct tcp_sock
*tp
= tcp_sk(sk
);
1005 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1008 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1009 tp
->fackets_out
-= decr
;
1012 /* Pcount in the middle of the write queue got changed, we need to do various
1013 * tweaks to fix counters
1015 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1017 struct tcp_sock
*tp
= tcp_sk(sk
);
1019 tp
->packets_out
-= decr
;
1021 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1022 tp
->sacked_out
-= decr
;
1023 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1024 tp
->retrans_out
-= decr
;
1025 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1026 tp
->lost_out
-= decr
;
1028 /* Reno case is special. Sigh... */
1029 if (tcp_is_reno(tp
) && decr
> 0)
1030 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1032 tcp_adjust_fackets_out(sk
, skb
, decr
);
1034 if (tp
->lost_skb_hint
&&
1035 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1036 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1037 tp
->lost_cnt_hint
-= decr
;
1039 tcp_verify_left_out(tp
);
1042 /* Function to create two new TCP segments. Shrinks the given segment
1043 * to the specified size and appends a new segment with the rest of the
1044 * packet to the list. This won't be called frequently, I hope.
1045 * Remember, these are still headerless SKBs at this point.
1047 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1048 unsigned int mss_now
)
1050 struct tcp_sock
*tp
= tcp_sk(sk
);
1051 struct sk_buff
*buff
;
1052 int nsize
, old_factor
;
1056 if (WARN_ON(len
> skb
->len
))
1059 nsize
= skb_headlen(skb
) - len
;
1063 if (skb_unclone(skb
, GFP_ATOMIC
))
1066 /* Get a new skb... force flag on. */
1067 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
1069 return -ENOMEM
; /* We'll just try again later. */
1071 sk
->sk_wmem_queued
+= buff
->truesize
;
1072 sk_mem_charge(sk
, buff
->truesize
);
1073 nlen
= skb
->len
- len
- nsize
;
1074 buff
->truesize
+= nlen
;
1075 skb
->truesize
-= nlen
;
1077 /* Correct the sequence numbers. */
1078 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1079 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1080 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1082 /* PSH and FIN should only be set in the second packet. */
1083 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1084 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1085 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1086 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1088 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1089 /* Copy and checksum data tail into the new buffer. */
1090 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1091 skb_put(buff
, nsize
),
1096 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1098 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1099 skb_split(skb
, buff
, len
);
1102 buff
->ip_summed
= skb
->ip_summed
;
1104 /* Looks stupid, but our code really uses when of
1105 * skbs, which it never sent before. --ANK
1107 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
1108 buff
->tstamp
= skb
->tstamp
;
1110 old_factor
= tcp_skb_pcount(skb
);
1112 /* Fix up tso_factor for both original and new SKB. */
1113 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1114 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1116 /* If this packet has been sent out already, we must
1117 * adjust the various packet counters.
1119 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1120 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1121 tcp_skb_pcount(buff
);
1124 tcp_adjust_pcount(sk
, skb
, diff
);
1127 /* Link BUFF into the send queue. */
1128 skb_header_release(buff
);
1129 tcp_insert_write_queue_after(skb
, buff
, sk
);
1134 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1135 * eventually). The difference is that pulled data not copied, but
1136 * immediately discarded.
1138 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1142 eat
= min_t(int, len
, skb_headlen(skb
));
1144 __skb_pull(skb
, eat
);
1151 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
1152 int size
= skb_frag_size(&skb_shinfo(skb
)->frags
[i
]);
1155 skb_frag_unref(skb
, i
);
1158 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
1160 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
1161 skb_frag_size_sub(&skb_shinfo(skb
)->frags
[k
], eat
);
1167 skb_shinfo(skb
)->nr_frags
= k
;
1169 skb_reset_tail_pointer(skb
);
1170 skb
->data_len
-= len
;
1171 skb
->len
= skb
->data_len
;
1174 /* Remove acked data from a packet in the transmit queue. */
1175 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1177 if (skb_unclone(skb
, GFP_ATOMIC
))
1180 __pskb_trim_head(skb
, len
);
1182 TCP_SKB_CB(skb
)->seq
+= len
;
1183 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1185 skb
->truesize
-= len
;
1186 sk
->sk_wmem_queued
-= len
;
1187 sk_mem_uncharge(sk
, len
);
1188 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1190 /* Any change of skb->len requires recalculation of tso factor. */
1191 if (tcp_skb_pcount(skb
) > 1)
1192 tcp_set_skb_tso_segs(sk
, skb
, tcp_skb_mss(skb
));
1197 /* Calculate MSS not accounting any TCP options. */
1198 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1200 const struct tcp_sock
*tp
= tcp_sk(sk
);
1201 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1204 /* Calculate base mss without TCP options:
1205 It is MMS_S - sizeof(tcphdr) of rfc1122
1207 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1209 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1210 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1211 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1213 if (dst
&& dst_allfrag(dst
))
1214 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1217 /* Clamp it (mss_clamp does not include tcp options) */
1218 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1219 mss_now
= tp
->rx_opt
.mss_clamp
;
1221 /* Now subtract optional transport overhead */
1222 mss_now
-= icsk
->icsk_ext_hdr_len
;
1224 /* Then reserve room for full set of TCP options and 8 bytes of data */
1230 /* Calculate MSS. Not accounting for SACKs here. */
1231 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1233 /* Subtract TCP options size, not including SACKs */
1234 return __tcp_mtu_to_mss(sk
, pmtu
) -
1235 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1238 /* Inverse of above */
1239 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1241 const struct tcp_sock
*tp
= tcp_sk(sk
);
1242 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1246 tp
->tcp_header_len
+
1247 icsk
->icsk_ext_hdr_len
+
1248 icsk
->icsk_af_ops
->net_header_len
;
1250 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1251 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1252 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1254 if (dst
&& dst_allfrag(dst
))
1255 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1260 /* MTU probing init per socket */
1261 void tcp_mtup_init(struct sock
*sk
)
1263 struct tcp_sock
*tp
= tcp_sk(sk
);
1264 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1266 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
1267 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1268 icsk
->icsk_af_ops
->net_header_len
;
1269 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
1270 icsk
->icsk_mtup
.probe_size
= 0;
1272 EXPORT_SYMBOL(tcp_mtup_init
);
1274 /* This function synchronize snd mss to current pmtu/exthdr set.
1276 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1277 for TCP options, but includes only bare TCP header.
1279 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1280 It is minimum of user_mss and mss received with SYN.
1281 It also does not include TCP options.
1283 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1285 tp->mss_cache is current effective sending mss, including
1286 all tcp options except for SACKs. It is evaluated,
1287 taking into account current pmtu, but never exceeds
1288 tp->rx_opt.mss_clamp.
1290 NOTE1. rfc1122 clearly states that advertised MSS
1291 DOES NOT include either tcp or ip options.
1293 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1294 are READ ONLY outside this function. --ANK (980731)
1296 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1298 struct tcp_sock
*tp
= tcp_sk(sk
);
1299 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1302 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1303 icsk
->icsk_mtup
.search_high
= pmtu
;
1305 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1306 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1308 /* And store cached results */
1309 icsk
->icsk_pmtu_cookie
= pmtu
;
1310 if (icsk
->icsk_mtup
.enabled
)
1311 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1312 tp
->mss_cache
= mss_now
;
1316 EXPORT_SYMBOL(tcp_sync_mss
);
1318 /* Compute the current effective MSS, taking SACKs and IP options,
1319 * and even PMTU discovery events into account.
1321 unsigned int tcp_current_mss(struct sock
*sk
)
1323 const struct tcp_sock
*tp
= tcp_sk(sk
);
1324 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1326 unsigned int header_len
;
1327 struct tcp_out_options opts
;
1328 struct tcp_md5sig_key
*md5
;
1330 mss_now
= tp
->mss_cache
;
1333 u32 mtu
= dst_mtu(dst
);
1334 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1335 mss_now
= tcp_sync_mss(sk
, mtu
);
1338 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1339 sizeof(struct tcphdr
);
1340 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1341 * some common options. If this is an odd packet (because we have SACK
1342 * blocks etc) then our calculated header_len will be different, and
1343 * we have to adjust mss_now correspondingly */
1344 if (header_len
!= tp
->tcp_header_len
) {
1345 int delta
= (int) header_len
- tp
->tcp_header_len
;
1352 /* Congestion window validation. (RFC2861) */
1353 static void tcp_cwnd_validate(struct sock
*sk
)
1355 struct tcp_sock
*tp
= tcp_sk(sk
);
1357 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1358 /* Network is feed fully. */
1359 tp
->snd_cwnd_used
= 0;
1360 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1362 /* Network starves. */
1363 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1364 tp
->snd_cwnd_used
= tp
->packets_out
;
1366 if (sysctl_tcp_slow_start_after_idle
&&
1367 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1368 tcp_cwnd_application_limited(sk
);
1372 /* Returns the portion of skb which can be sent right away without
1373 * introducing MSS oddities to segment boundaries. In rare cases where
1374 * mss_now != mss_cache, we will request caller to create a small skb
1375 * per input skb which could be mostly avoided here (if desired).
1377 * We explicitly want to create a request for splitting write queue tail
1378 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1379 * thus all the complexity (cwnd_len is always MSS multiple which we
1380 * return whenever allowed by the other factors). Basically we need the
1381 * modulo only when the receiver window alone is the limiting factor or
1382 * when we would be allowed to send the split-due-to-Nagle skb fully.
1384 static unsigned int tcp_mss_split_point(const struct sock
*sk
, const struct sk_buff
*skb
,
1385 unsigned int mss_now
, unsigned int max_segs
)
1387 const struct tcp_sock
*tp
= tcp_sk(sk
);
1388 u32 needed
, window
, max_len
;
1390 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1391 max_len
= mss_now
* max_segs
;
1393 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1396 needed
= min(skb
->len
, window
);
1398 if (max_len
<= needed
)
1401 return needed
- needed
% mss_now
;
1404 /* Can at least one segment of SKB be sent right now, according to the
1405 * congestion window rules? If so, return how many segments are allowed.
1407 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1408 const struct sk_buff
*skb
)
1410 u32 in_flight
, cwnd
;
1412 /* Don't be strict about the congestion window for the final FIN. */
1413 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1414 tcp_skb_pcount(skb
) == 1)
1417 in_flight
= tcp_packets_in_flight(tp
);
1418 cwnd
= tp
->snd_cwnd
;
1419 if (in_flight
< cwnd
)
1420 return (cwnd
- in_flight
);
1425 /* Initialize TSO state of a skb.
1426 * This must be invoked the first time we consider transmitting
1427 * SKB onto the wire.
1429 static int tcp_init_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
1430 unsigned int mss_now
)
1432 int tso_segs
= tcp_skb_pcount(skb
);
1434 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1435 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1436 tso_segs
= tcp_skb_pcount(skb
);
1441 /* Minshall's variant of the Nagle send check. */
1442 static inline bool tcp_minshall_check(const struct tcp_sock
*tp
)
1444 return after(tp
->snd_sml
, tp
->snd_una
) &&
1445 !after(tp
->snd_sml
, tp
->snd_nxt
);
1448 /* Return false, if packet can be sent now without violation Nagle's rules:
1449 * 1. It is full sized.
1450 * 2. Or it contains FIN. (already checked by caller)
1451 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1452 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1453 * With Minshall's modification: all sent small packets are ACKed.
1455 static inline bool tcp_nagle_check(const struct tcp_sock
*tp
,
1456 const struct sk_buff
*skb
,
1457 unsigned int mss_now
, int nonagle
)
1459 return skb
->len
< mss_now
&&
1460 ((nonagle
& TCP_NAGLE_CORK
) ||
1461 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1464 /* Return true if the Nagle test allows this packet to be
1467 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1468 unsigned int cur_mss
, int nonagle
)
1470 /* Nagle rule does not apply to frames, which sit in the middle of the
1471 * write_queue (they have no chances to get new data).
1473 * This is implemented in the callers, where they modify the 'nonagle'
1474 * argument based upon the location of SKB in the send queue.
1476 if (nonagle
& TCP_NAGLE_PUSH
)
1479 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1480 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1483 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1489 /* Does at least the first segment of SKB fit into the send window? */
1490 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1491 const struct sk_buff
*skb
,
1492 unsigned int cur_mss
)
1494 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1496 if (skb
->len
> cur_mss
)
1497 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1499 return !after(end_seq
, tcp_wnd_end(tp
));
1502 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1503 * should be put on the wire right now. If so, it returns the number of
1504 * packets allowed by the congestion window.
1506 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1507 unsigned int cur_mss
, int nonagle
)
1509 const struct tcp_sock
*tp
= tcp_sk(sk
);
1510 unsigned int cwnd_quota
;
1512 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1514 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1517 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1518 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1524 /* Test if sending is allowed right now. */
1525 bool tcp_may_send_now(struct sock
*sk
)
1527 const struct tcp_sock
*tp
= tcp_sk(sk
);
1528 struct sk_buff
*skb
= tcp_send_head(sk
);
1531 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1532 (tcp_skb_is_last(sk
, skb
) ?
1533 tp
->nonagle
: TCP_NAGLE_PUSH
));
1536 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1537 * which is put after SKB on the list. It is very much like
1538 * tcp_fragment() except that it may make several kinds of assumptions
1539 * in order to speed up the splitting operation. In particular, we
1540 * know that all the data is in scatter-gather pages, and that the
1541 * packet has never been sent out before (and thus is not cloned).
1543 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1544 unsigned int mss_now
, gfp_t gfp
)
1546 struct sk_buff
*buff
;
1547 int nlen
= skb
->len
- len
;
1550 /* All of a TSO frame must be composed of paged data. */
1551 if (skb
->len
!= skb
->data_len
)
1552 return tcp_fragment(sk
, skb
, len
, mss_now
);
1554 buff
= sk_stream_alloc_skb(sk
, 0, gfp
);
1555 if (unlikely(buff
== NULL
))
1558 sk
->sk_wmem_queued
+= buff
->truesize
;
1559 sk_mem_charge(sk
, buff
->truesize
);
1560 buff
->truesize
+= nlen
;
1561 skb
->truesize
-= nlen
;
1563 /* Correct the sequence numbers. */
1564 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1565 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1566 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1568 /* PSH and FIN should only be set in the second packet. */
1569 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1570 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1571 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1573 /* This packet was never sent out yet, so no SACK bits. */
1574 TCP_SKB_CB(buff
)->sacked
= 0;
1576 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1577 skb_split(skb
, buff
, len
);
1579 /* Fix up tso_factor for both original and new SKB. */
1580 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1581 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1583 /* Link BUFF into the send queue. */
1584 skb_header_release(buff
);
1585 tcp_insert_write_queue_after(skb
, buff
, sk
);
1590 /* Try to defer sending, if possible, in order to minimize the amount
1591 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1593 * This algorithm is from John Heffner.
1595 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1597 struct tcp_sock
*tp
= tcp_sk(sk
);
1598 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1599 u32 send_win
, cong_win
, limit
, in_flight
;
1602 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1605 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1608 /* Defer for less than two clock ticks. */
1609 if (tp
->tso_deferred
&&
1610 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1613 in_flight
= tcp_packets_in_flight(tp
);
1615 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1617 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1619 /* From in_flight test above, we know that cwnd > in_flight. */
1620 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1622 limit
= min(send_win
, cong_win
);
1624 /* If a full-sized TSO skb can be sent, do it. */
1625 if (limit
>= min_t(unsigned int, sk
->sk_gso_max_size
,
1626 tp
->xmit_size_goal_segs
* tp
->mss_cache
))
1629 /* Middle in queue won't get any more data, full sendable already? */
1630 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1633 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1635 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1637 /* If at least some fraction of a window is available,
1640 chunk
/= win_divisor
;
1644 /* Different approach, try not to defer past a single
1645 * ACK. Receiver should ACK every other full sized
1646 * frame, so if we have space for more than 3 frames
1649 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1653 /* Ok, it looks like it is advisable to defer.
1654 * Do not rearm the timer if already set to not break TCP ACK clocking.
1656 if (!tp
->tso_deferred
)
1657 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1662 tp
->tso_deferred
= 0;
1666 /* Create a new MTU probe if we are ready.
1667 * MTU probe is regularly attempting to increase the path MTU by
1668 * deliberately sending larger packets. This discovers routing
1669 * changes resulting in larger path MTUs.
1671 * Returns 0 if we should wait to probe (no cwnd available),
1672 * 1 if a probe was sent,
1675 static int tcp_mtu_probe(struct sock
*sk
)
1677 struct tcp_sock
*tp
= tcp_sk(sk
);
1678 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1679 struct sk_buff
*skb
, *nskb
, *next
;
1686 /* Not currently probing/verifying,
1688 * have enough cwnd, and
1689 * not SACKing (the variable headers throw things off) */
1690 if (!icsk
->icsk_mtup
.enabled
||
1691 icsk
->icsk_mtup
.probe_size
||
1692 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1693 tp
->snd_cwnd
< 11 ||
1694 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1697 /* Very simple search strategy: just double the MSS. */
1698 mss_now
= tcp_current_mss(sk
);
1699 probe_size
= 2 * tp
->mss_cache
;
1700 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1701 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1702 /* TODO: set timer for probe_converge_event */
1706 /* Have enough data in the send queue to probe? */
1707 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1710 if (tp
->snd_wnd
< size_needed
)
1712 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1715 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1716 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1717 if (!tcp_packets_in_flight(tp
))
1723 /* We're allowed to probe. Build it now. */
1724 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1726 sk
->sk_wmem_queued
+= nskb
->truesize
;
1727 sk_mem_charge(sk
, nskb
->truesize
);
1729 skb
= tcp_send_head(sk
);
1731 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1732 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1733 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
1734 TCP_SKB_CB(nskb
)->sacked
= 0;
1736 nskb
->ip_summed
= skb
->ip_summed
;
1738 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1741 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1742 copy
= min_t(int, skb
->len
, probe_size
- len
);
1743 if (nskb
->ip_summed
)
1744 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1746 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1747 skb_put(nskb
, copy
),
1750 if (skb
->len
<= copy
) {
1751 /* We've eaten all the data from this skb.
1753 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1754 tcp_unlink_write_queue(skb
, sk
);
1755 sk_wmem_free_skb(sk
, skb
);
1757 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
1758 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1759 if (!skb_shinfo(skb
)->nr_frags
) {
1760 skb_pull(skb
, copy
);
1761 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1762 skb
->csum
= csum_partial(skb
->data
,
1765 __pskb_trim_head(skb
, copy
);
1766 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1768 TCP_SKB_CB(skb
)->seq
+= copy
;
1773 if (len
>= probe_size
)
1776 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1778 /* We're ready to send. If this fails, the probe will
1779 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1780 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1781 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1782 /* Decrement cwnd here because we are sending
1783 * effectively two packets. */
1785 tcp_event_new_data_sent(sk
, nskb
);
1787 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1788 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1789 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1797 /* This routine writes packets to the network. It advances the
1798 * send_head. This happens as incoming acks open up the remote
1801 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1802 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1803 * account rare use of URG, this is not a big flaw.
1805 * Send at most one packet when push_one > 0. Temporarily ignore
1806 * cwnd limit to force at most one packet out when push_one == 2.
1808 * Returns true, if no segments are in flight and we have queued segments,
1809 * but cannot send anything now because of SWS or another problem.
1811 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1812 int push_one
, gfp_t gfp
)
1814 struct tcp_sock
*tp
= tcp_sk(sk
);
1815 struct sk_buff
*skb
;
1816 unsigned int tso_segs
, sent_pkts
;
1823 /* Do MTU probing. */
1824 result
= tcp_mtu_probe(sk
);
1827 } else if (result
> 0) {
1832 while ((skb
= tcp_send_head(sk
))) {
1835 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1838 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
)
1839 goto repair
; /* Skip network transmission */
1841 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1844 /* Force out a loss probe pkt. */
1850 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1853 if (tso_segs
== 1) {
1854 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1855 (tcp_skb_is_last(sk
, skb
) ?
1856 nonagle
: TCP_NAGLE_PUSH
))))
1859 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
1863 /* TCP Small Queues :
1864 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
1866 * - better RTT estimation and ACK scheduling
1870 limit
= max(skb
->truesize
, sk
->sk_pacing_rate
>> 10);
1872 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
) {
1873 set_bit(TSQ_THROTTLED
, &tp
->tsq_flags
);
1878 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
1879 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1882 sk
->sk_gso_max_segs
));
1884 if (skb
->len
> limit
&&
1885 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
1888 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1890 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
1894 /* Advance the send_head. This one is sent out.
1895 * This call will increment packets_out.
1897 tcp_event_new_data_sent(sk
, skb
);
1899 tcp_minshall_update(tp
, mss_now
, skb
);
1900 sent_pkts
+= tcp_skb_pcount(skb
);
1906 if (likely(sent_pkts
)) {
1907 if (tcp_in_cwnd_reduction(sk
))
1908 tp
->prr_out
+= sent_pkts
;
1910 /* Send one loss probe per tail loss episode. */
1912 tcp_schedule_loss_probe(sk
);
1913 tcp_cwnd_validate(sk
);
1916 return (push_one
== 2) || (!tp
->packets_out
&& tcp_send_head(sk
));
1919 bool tcp_schedule_loss_probe(struct sock
*sk
)
1921 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1922 struct tcp_sock
*tp
= tcp_sk(sk
);
1923 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
1924 u32 rtt
= tp
->srtt
>> 3;
1926 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
))
1928 /* No consecutive loss probes. */
1929 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
1933 /* Don't do any loss probe on a Fast Open connection before 3WHS
1936 if (sk
->sk_state
== TCP_SYN_RECV
)
1939 /* TLP is only scheduled when next timer event is RTO. */
1940 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
1943 /* Schedule a loss probe in 2*RTT for SACK capable connections
1944 * in Open state, that are either limited by cwnd or application.
1946 if (sysctl_tcp_early_retrans
< 3 || !rtt
|| !tp
->packets_out
||
1947 !tcp_is_sack(tp
) || inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
1950 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
1954 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1955 * for delayed ack when there's one outstanding packet.
1958 if (tp
->packets_out
== 1)
1959 timeout
= max_t(u32
, timeout
,
1960 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
1961 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
1963 /* If RTO is shorter, just schedule TLP in its place. */
1964 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
1965 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
1966 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
1967 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
1972 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
1977 /* When probe timeout (PTO) fires, send a new segment if one exists, else
1978 * retransmit the last segment.
1980 void tcp_send_loss_probe(struct sock
*sk
)
1982 struct tcp_sock
*tp
= tcp_sk(sk
);
1983 struct sk_buff
*skb
;
1985 int mss
= tcp_current_mss(sk
);
1988 if (tcp_send_head(sk
) != NULL
) {
1989 err
= tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
1993 /* At most one outstanding TLP retransmission. */
1994 if (tp
->tlp_high_seq
)
1997 /* Retransmit last segment. */
1998 skb
= tcp_write_queue_tail(sk
);
2002 pcount
= tcp_skb_pcount(skb
);
2003 if (WARN_ON(!pcount
))
2006 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2007 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
)))
2009 skb
= tcp_write_queue_tail(sk
);
2012 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2015 /* Probe with zero data doesn't trigger fast recovery. */
2017 err
= __tcp_retransmit_skb(sk
, skb
);
2019 /* Record snd_nxt for loss detection. */
2021 tp
->tlp_high_seq
= tp
->snd_nxt
;
2024 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2025 inet_csk(sk
)->icsk_rto
,
2029 NET_INC_STATS_BH(sock_net(sk
),
2030 LINUX_MIB_TCPLOSSPROBES
);
2034 /* Push out any pending frames which were held back due to
2035 * TCP_CORK or attempt at coalescing tiny packets.
2036 * The socket must be locked by the caller.
2038 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2041 /* If we are closed, the bytes will have to remain here.
2042 * In time closedown will finish, we empty the write queue and
2043 * all will be happy.
2045 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2048 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2049 sk_gfp_atomic(sk
, GFP_ATOMIC
)))
2050 tcp_check_probe_timer(sk
);
2053 /* Send _single_ skb sitting at the send head. This function requires
2054 * true push pending frames to setup probe timer etc.
2056 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2058 struct sk_buff
*skb
= tcp_send_head(sk
);
2060 BUG_ON(!skb
|| skb
->len
< mss_now
);
2062 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2065 /* This function returns the amount that we can raise the
2066 * usable window based on the following constraints
2068 * 1. The window can never be shrunk once it is offered (RFC 793)
2069 * 2. We limit memory per socket
2072 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2073 * RECV.NEXT + RCV.WIN fixed until:
2074 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2076 * i.e. don't raise the right edge of the window until you can raise
2077 * it at least MSS bytes.
2079 * Unfortunately, the recommended algorithm breaks header prediction,
2080 * since header prediction assumes th->window stays fixed.
2082 * Strictly speaking, keeping th->window fixed violates the receiver
2083 * side SWS prevention criteria. The problem is that under this rule
2084 * a stream of single byte packets will cause the right side of the
2085 * window to always advance by a single byte.
2087 * Of course, if the sender implements sender side SWS prevention
2088 * then this will not be a problem.
2090 * BSD seems to make the following compromise:
2092 * If the free space is less than the 1/4 of the maximum
2093 * space available and the free space is less than 1/2 mss,
2094 * then set the window to 0.
2095 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2096 * Otherwise, just prevent the window from shrinking
2097 * and from being larger than the largest representable value.
2099 * This prevents incremental opening of the window in the regime
2100 * where TCP is limited by the speed of the reader side taking
2101 * data out of the TCP receive queue. It does nothing about
2102 * those cases where the window is constrained on the sender side
2103 * because the pipeline is full.
2105 * BSD also seems to "accidentally" limit itself to windows that are a
2106 * multiple of MSS, at least until the free space gets quite small.
2107 * This would appear to be a side effect of the mbuf implementation.
2108 * Combining these two algorithms results in the observed behavior
2109 * of having a fixed window size at almost all times.
2111 * Below we obtain similar behavior by forcing the offered window to
2112 * a multiple of the mss when it is feasible to do so.
2114 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2115 * Regular options like TIMESTAMP are taken into account.
2117 u32
__tcp_select_window(struct sock
*sk
)
2119 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2120 struct tcp_sock
*tp
= tcp_sk(sk
);
2121 /* MSS for the peer's data. Previous versions used mss_clamp
2122 * here. I don't know if the value based on our guesses
2123 * of peer's MSS is better for the performance. It's more correct
2124 * but may be worse for the performance because of rcv_mss
2125 * fluctuations. --SAW 1998/11/1
2127 int mss
= icsk
->icsk_ack
.rcv_mss
;
2128 int free_space
= tcp_space(sk
);
2129 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
2132 if (mss
> full_space
)
2135 if (free_space
< (full_space
>> 1)) {
2136 icsk
->icsk_ack
.quick
= 0;
2138 if (sk_under_memory_pressure(sk
))
2139 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2142 if (free_space
< mss
)
2146 if (free_space
> tp
->rcv_ssthresh
)
2147 free_space
= tp
->rcv_ssthresh
;
2149 /* Don't do rounding if we are using window scaling, since the
2150 * scaled window will not line up with the MSS boundary anyway.
2152 window
= tp
->rcv_wnd
;
2153 if (tp
->rx_opt
.rcv_wscale
) {
2154 window
= free_space
;
2156 /* Advertise enough space so that it won't get scaled away.
2157 * Import case: prevent zero window announcement if
2158 * 1<<rcv_wscale > mss.
2160 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2161 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2162 << tp
->rx_opt
.rcv_wscale
);
2164 /* Get the largest window that is a nice multiple of mss.
2165 * Window clamp already applied above.
2166 * If our current window offering is within 1 mss of the
2167 * free space we just keep it. This prevents the divide
2168 * and multiply from happening most of the time.
2169 * We also don't do any window rounding when the free space
2172 if (window
<= free_space
- mss
|| window
> free_space
)
2173 window
= (free_space
/ mss
) * mss
;
2174 else if (mss
== full_space
&&
2175 free_space
> window
+ (full_space
>> 1))
2176 window
= free_space
;
2182 /* Collapses two adjacent SKB's during retransmission. */
2183 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2185 struct tcp_sock
*tp
= tcp_sk(sk
);
2186 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2187 int skb_size
, next_skb_size
;
2189 skb_size
= skb
->len
;
2190 next_skb_size
= next_skb
->len
;
2192 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2194 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2196 tcp_unlink_write_queue(next_skb
, sk
);
2198 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
2201 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2202 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2204 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2205 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2207 /* Update sequence range on original skb. */
2208 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2210 /* Merge over control information. This moves PSH/FIN etc. over */
2211 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2213 /* All done, get rid of second SKB and account for it so
2214 * packet counting does not break.
2216 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2218 /* changed transmit queue under us so clear hints */
2219 tcp_clear_retrans_hints_partial(tp
);
2220 if (next_skb
== tp
->retransmit_skb_hint
)
2221 tp
->retransmit_skb_hint
= skb
;
2223 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2225 sk_wmem_free_skb(sk
, next_skb
);
2228 /* Check if coalescing SKBs is legal. */
2229 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2231 if (tcp_skb_pcount(skb
) > 1)
2233 /* TODO: SACK collapsing could be used to remove this condition */
2234 if (skb_shinfo(skb
)->nr_frags
!= 0)
2236 if (skb_cloned(skb
))
2238 if (skb
== tcp_send_head(sk
))
2240 /* Some heurestics for collapsing over SACK'd could be invented */
2241 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2247 /* Collapse packets in the retransmit queue to make to create
2248 * less packets on the wire. This is only done on retransmission.
2250 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2253 struct tcp_sock
*tp
= tcp_sk(sk
);
2254 struct sk_buff
*skb
= to
, *tmp
;
2257 if (!sysctl_tcp_retrans_collapse
)
2259 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2262 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2263 if (!tcp_can_collapse(sk
, skb
))
2275 /* Punt if not enough space exists in the first SKB for
2276 * the data in the second
2278 if (skb
->len
> skb_availroom(to
))
2281 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2284 tcp_collapse_retrans(sk
, to
);
2288 /* This retransmits one SKB. Policy decisions and retransmit queue
2289 * state updates are done by the caller. Returns non-zero if an
2290 * error occurred which prevented the send.
2292 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2294 struct tcp_sock
*tp
= tcp_sk(sk
);
2295 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2296 unsigned int cur_mss
;
2298 /* Inconslusive MTU probe */
2299 if (icsk
->icsk_mtup
.probe_size
) {
2300 icsk
->icsk_mtup
.probe_size
= 0;
2303 /* Do not sent more than we queued. 1/4 is reserved for possible
2304 * copying overhead: fragmentation, tunneling, mangling etc.
2306 if (atomic_read(&sk
->sk_wmem_alloc
) >
2307 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2310 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2311 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2313 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2317 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2318 return -EHOSTUNREACH
; /* Routing failure or similar. */
2320 cur_mss
= tcp_current_mss(sk
);
2322 /* If receiver has shrunk his window, and skb is out of
2323 * new window, do not retransmit it. The exception is the
2324 * case, when window is shrunk to zero. In this case
2325 * our retransmit serves as a zero window probe.
2327 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2328 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2331 if (skb
->len
> cur_mss
) {
2332 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
2333 return -ENOMEM
; /* We'll try again later. */
2335 int oldpcount
= tcp_skb_pcount(skb
);
2337 if (unlikely(oldpcount
> 1)) {
2338 if (skb_unclone(skb
, GFP_ATOMIC
))
2340 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2341 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2345 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2347 /* Some Solaris stacks overoptimize and ignore the FIN on a
2348 * retransmit when old data is attached. So strip it off
2349 * since it is cheap to do so and saves bytes on the network.
2352 (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
2353 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
2354 if (!pskb_trim(skb
, 0)) {
2355 /* Reuse, even though it does some unnecessary work */
2356 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
2357 TCP_SKB_CB(skb
)->tcp_flags
);
2358 skb
->ip_summed
= CHECKSUM_NONE
;
2362 /* Make a copy, if the first transmission SKB clone we made
2363 * is still in somebody's hands, else make a clone.
2365 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2367 /* make sure skb->data is aligned on arches that require it
2368 * and check if ack-trimming & collapsing extended the headroom
2369 * beyond what csum_start can cover.
2371 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2372 skb_headroom(skb
) >= 0xFFFF)) {
2373 struct sk_buff
*nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
,
2375 return nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2378 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2382 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2384 struct tcp_sock
*tp
= tcp_sk(sk
);
2385 int err
= __tcp_retransmit_skb(sk
, skb
);
2388 /* Update global TCP statistics. */
2389 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2391 tp
->total_retrans
++;
2393 #if FASTRETRANS_DEBUG > 0
2394 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2395 net_dbg_ratelimited("retrans_out leaked\n");
2398 if (!tp
->retrans_out
)
2399 tp
->lost_retrans_low
= tp
->snd_nxt
;
2400 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2401 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2403 /* Save stamp of the first retransmit. */
2404 if (!tp
->retrans_stamp
)
2405 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2407 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2409 /* snd_nxt is stored to detect loss of retransmitted segment,
2410 * see tcp_input.c tcp_sacktag_write_queue().
2412 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2417 /* Check if we forward retransmits are possible in the current
2418 * window/congestion state.
2420 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2422 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2423 const struct tcp_sock
*tp
= tcp_sk(sk
);
2425 /* Forward retransmissions are possible only during Recovery. */
2426 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2429 /* No forward retransmissions in Reno are possible. */
2430 if (tcp_is_reno(tp
))
2433 /* Yeah, we have to make difficult choice between forward transmission
2434 * and retransmission... Both ways have their merits...
2436 * For now we do not retransmit anything, while we have some new
2437 * segments to send. In the other cases, follow rule 3 for
2438 * NextSeg() specified in RFC3517.
2441 if (tcp_may_send_now(sk
))
2447 /* This gets called after a retransmit timeout, and the initially
2448 * retransmitted data is acknowledged. It tries to continue
2449 * resending the rest of the retransmit queue, until either
2450 * we've sent it all or the congestion window limit is reached.
2451 * If doing SACK, the first ACK which comes back for a timeout
2452 * based retransmit packet might feed us FACK information again.
2453 * If so, we use it to avoid unnecessarily retransmissions.
2455 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2457 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2458 struct tcp_sock
*tp
= tcp_sk(sk
);
2459 struct sk_buff
*skb
;
2460 struct sk_buff
*hole
= NULL
;
2463 int fwd_rexmitting
= 0;
2465 if (!tp
->packets_out
)
2469 tp
->retransmit_high
= tp
->snd_una
;
2471 if (tp
->retransmit_skb_hint
) {
2472 skb
= tp
->retransmit_skb_hint
;
2473 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2474 if (after(last_lost
, tp
->retransmit_high
))
2475 last_lost
= tp
->retransmit_high
;
2477 skb
= tcp_write_queue_head(sk
);
2478 last_lost
= tp
->snd_una
;
2481 tcp_for_write_queue_from(skb
, sk
) {
2482 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2484 if (skb
== tcp_send_head(sk
))
2486 /* we could do better than to assign each time */
2488 tp
->retransmit_skb_hint
= skb
;
2490 /* Assume this retransmit will generate
2491 * only one packet for congestion window
2492 * calculation purposes. This works because
2493 * tcp_retransmit_skb() will chop up the
2494 * packet to be MSS sized and all the
2495 * packet counting works out.
2497 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2500 if (fwd_rexmitting
) {
2502 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2504 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2506 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2507 tp
->retransmit_high
= last_lost
;
2508 if (!tcp_can_forward_retransmit(sk
))
2510 /* Backtrack if necessary to non-L'ed skb */
2518 } else if (!(sacked
& TCPCB_LOST
)) {
2519 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2524 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2525 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2526 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2528 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2531 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2534 if (tcp_retransmit_skb(sk
, skb
)) {
2535 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2538 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2540 if (tcp_in_cwnd_reduction(sk
))
2541 tp
->prr_out
+= tcp_skb_pcount(skb
);
2543 if (skb
== tcp_write_queue_head(sk
))
2544 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2545 inet_csk(sk
)->icsk_rto
,
2550 /* Send a fin. The caller locks the socket for us. This cannot be
2551 * allowed to fail queueing a FIN frame under any circumstances.
2553 void tcp_send_fin(struct sock
*sk
)
2555 struct tcp_sock
*tp
= tcp_sk(sk
);
2556 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2559 /* Optimization, tack on the FIN if we have a queue of
2560 * unsent frames. But be careful about outgoing SACKS
2563 mss_now
= tcp_current_mss(sk
);
2565 if (tcp_send_head(sk
) != NULL
) {
2566 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_FIN
;
2567 TCP_SKB_CB(skb
)->end_seq
++;
2570 /* Socket is locked, keep trying until memory is available. */
2572 skb
= alloc_skb_fclone(MAX_TCP_HEADER
,
2579 /* Reserve space for headers and prepare control bits. */
2580 skb_reserve(skb
, MAX_TCP_HEADER
);
2581 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2582 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2583 TCPHDR_ACK
| TCPHDR_FIN
);
2584 tcp_queue_skb(sk
, skb
);
2586 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2589 /* We get here when a process closes a file descriptor (either due to
2590 * an explicit close() or as a byproduct of exit()'ing) and there
2591 * was unread data in the receive queue. This behavior is recommended
2592 * by RFC 2525, section 2.17. -DaveM
2594 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2596 struct sk_buff
*skb
;
2598 /* NOTE: No TCP options attached and we never retransmit this. */
2599 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2601 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2605 /* Reserve space for headers and prepare control bits. */
2606 skb_reserve(skb
, MAX_TCP_HEADER
);
2607 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2608 TCPHDR_ACK
| TCPHDR_RST
);
2610 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2611 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2612 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2614 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2617 /* Send a crossed SYN-ACK during socket establishment.
2618 * WARNING: This routine must only be called when we have already sent
2619 * a SYN packet that crossed the incoming SYN that caused this routine
2620 * to get called. If this assumption fails then the initial rcv_wnd
2621 * and rcv_wscale values will not be correct.
2623 int tcp_send_synack(struct sock
*sk
)
2625 struct sk_buff
*skb
;
2627 skb
= tcp_write_queue_head(sk
);
2628 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2629 pr_debug("%s: wrong queue state\n", __func__
);
2632 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
2633 if (skb_cloned(skb
)) {
2634 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2637 tcp_unlink_write_queue(skb
, sk
);
2638 skb_header_release(nskb
);
2639 __tcp_add_write_queue_head(sk
, nskb
);
2640 sk_wmem_free_skb(sk
, skb
);
2641 sk
->sk_wmem_queued
+= nskb
->truesize
;
2642 sk_mem_charge(sk
, nskb
->truesize
);
2646 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
2647 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2649 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2650 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2654 * tcp_make_synack - Prepare a SYN-ACK.
2655 * sk: listener socket
2656 * dst: dst entry attached to the SYNACK
2657 * req: request_sock pointer
2659 * Allocate one skb and build a SYNACK packet.
2660 * @dst is consumed : Caller should not use it again.
2662 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2663 struct request_sock
*req
,
2664 struct tcp_fastopen_cookie
*foc
)
2666 struct tcp_out_options opts
;
2667 struct inet_request_sock
*ireq
= inet_rsk(req
);
2668 struct tcp_sock
*tp
= tcp_sk(sk
);
2670 struct sk_buff
*skb
;
2671 struct tcp_md5sig_key
*md5
;
2672 int tcp_header_size
;
2675 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2676 if (unlikely(!skb
)) {
2680 /* Reserve space for headers. */
2681 skb_reserve(skb
, MAX_TCP_HEADER
);
2683 skb_dst_set(skb
, dst
);
2684 security_skb_owned_by(skb
, sk
);
2686 mss
= dst_metric_advmss(dst
);
2687 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2688 mss
= tp
->rx_opt
.user_mss
;
2690 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2692 /* Set this up on the first call only */
2693 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2695 /* limit the window selection if the user enforce a smaller rx buffer */
2696 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2697 (req
->window_clamp
> tcp_full_space(sk
) || req
->window_clamp
== 0))
2698 req
->window_clamp
= tcp_full_space(sk
);
2700 /* tcp_full_space because it is guaranteed to be the first packet */
2701 tcp_select_initial_window(tcp_full_space(sk
),
2702 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2707 dst_metric(dst
, RTAX_INITRWND
));
2708 ireq
->rcv_wscale
= rcv_wscale
;
2711 memset(&opts
, 0, sizeof(opts
));
2712 #ifdef CONFIG_SYN_COOKIES
2713 if (unlikely(req
->cookie_ts
))
2714 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2717 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2718 tcp_header_size
= tcp_synack_options(sk
, req
, mss
, skb
, &opts
, &md5
,
2721 skb_push(skb
, tcp_header_size
);
2722 skb_reset_transport_header(skb
);
2725 memset(th
, 0, sizeof(struct tcphdr
));
2728 TCP_ECN_make_synack(req
, th
);
2729 th
->source
= ireq
->loc_port
;
2730 th
->dest
= ireq
->rmt_port
;
2731 /* Setting of flags are superfluous here for callers (and ECE is
2732 * not even correctly set)
2734 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2735 TCPHDR_SYN
| TCPHDR_ACK
);
2737 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2738 /* XXX data is queued and acked as is. No buffer/window check */
2739 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
2741 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2742 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2743 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2744 th
->doff
= (tcp_header_size
>> 2);
2745 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
, tcp_skb_pcount(skb
));
2747 #ifdef CONFIG_TCP_MD5SIG
2748 /* Okay, we have all we need - do the md5 hash if needed */
2750 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2751 md5
, NULL
, req
, skb
);
2757 EXPORT_SYMBOL(tcp_make_synack
);
2759 /* Do all connect socket setups that can be done AF independent. */
2760 void tcp_connect_init(struct sock
*sk
)
2762 const struct dst_entry
*dst
= __sk_dst_get(sk
);
2763 struct tcp_sock
*tp
= tcp_sk(sk
);
2766 /* We'll fix this up when we get a response from the other end.
2767 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2769 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2770 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2772 #ifdef CONFIG_TCP_MD5SIG
2773 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2774 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2777 /* If user gave his TCP_MAXSEG, record it to clamp */
2778 if (tp
->rx_opt
.user_mss
)
2779 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2782 tcp_sync_mss(sk
, dst_mtu(dst
));
2784 if (!tp
->window_clamp
)
2785 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2786 tp
->advmss
= dst_metric_advmss(dst
);
2787 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2788 tp
->advmss
= tp
->rx_opt
.user_mss
;
2790 tcp_initialize_rcv_mss(sk
);
2792 /* limit the window selection if the user enforce a smaller rx buffer */
2793 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2794 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
2795 tp
->window_clamp
= tcp_full_space(sk
);
2797 tcp_select_initial_window(tcp_full_space(sk
),
2798 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2801 sysctl_tcp_window_scaling
,
2803 dst_metric(dst
, RTAX_INITRWND
));
2805 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2806 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2809 sock_reset_flag(sk
, SOCK_DONE
);
2812 tp
->snd_una
= tp
->write_seq
;
2813 tp
->snd_sml
= tp
->write_seq
;
2814 tp
->snd_up
= tp
->write_seq
;
2815 tp
->snd_nxt
= tp
->write_seq
;
2817 if (likely(!tp
->repair
))
2820 tp
->rcv_tstamp
= tcp_time_stamp
;
2821 tp
->rcv_wup
= tp
->rcv_nxt
;
2822 tp
->copied_seq
= tp
->rcv_nxt
;
2824 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2825 inet_csk(sk
)->icsk_retransmits
= 0;
2826 tcp_clear_retrans(tp
);
2829 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
2831 struct tcp_sock
*tp
= tcp_sk(sk
);
2832 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
2834 tcb
->end_seq
+= skb
->len
;
2835 skb_header_release(skb
);
2836 __tcp_add_write_queue_tail(sk
, skb
);
2837 sk
->sk_wmem_queued
+= skb
->truesize
;
2838 sk_mem_charge(sk
, skb
->truesize
);
2839 tp
->write_seq
= tcb
->end_seq
;
2840 tp
->packets_out
+= tcp_skb_pcount(skb
);
2843 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2844 * queue a data-only packet after the regular SYN, such that regular SYNs
2845 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2846 * only the SYN sequence, the data are retransmitted in the first ACK.
2847 * If cookie is not cached or other error occurs, falls back to send a
2848 * regular SYN with Fast Open cookie request option.
2850 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
2852 struct tcp_sock
*tp
= tcp_sk(sk
);
2853 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
2854 int syn_loss
= 0, space
, i
, err
= 0, iovlen
= fo
->data
->msg_iovlen
;
2855 struct sk_buff
*syn_data
= NULL
, *data
;
2856 unsigned long last_syn_loss
= 0;
2858 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
2859 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
2860 &syn_loss
, &last_syn_loss
);
2861 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2863 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
2864 fo
->cookie
.len
= -1;
2868 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
2869 fo
->cookie
.len
= -1;
2870 else if (fo
->cookie
.len
<= 0)
2873 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2874 * user-MSS. Reserve maximum option space for middleboxes that add
2875 * private TCP options. The cost is reduced data space in SYN :(
2877 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
2878 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2879 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
2880 MAX_TCP_OPTION_SPACE
;
2882 syn_data
= skb_copy_expand(syn
, skb_headroom(syn
), space
,
2884 if (syn_data
== NULL
)
2887 for (i
= 0; i
< iovlen
&& syn_data
->len
< space
; ++i
) {
2888 struct iovec
*iov
= &fo
->data
->msg_iov
[i
];
2889 unsigned char __user
*from
= iov
->iov_base
;
2890 int len
= iov
->iov_len
;
2892 if (syn_data
->len
+ len
> space
)
2893 len
= space
- syn_data
->len
;
2894 else if (i
+ 1 == iovlen
)
2895 /* No more data pending in inet_wait_for_connect() */
2898 if (skb_add_data(syn_data
, from
, len
))
2902 /* Queue a data-only packet after the regular SYN for retransmission */
2903 data
= pskb_copy(syn_data
, sk
->sk_allocation
);
2906 TCP_SKB_CB(data
)->seq
++;
2907 TCP_SKB_CB(data
)->tcp_flags
&= ~TCPHDR_SYN
;
2908 TCP_SKB_CB(data
)->tcp_flags
= (TCPHDR_ACK
|TCPHDR_PSH
);
2909 tcp_connect_queue_skb(sk
, data
);
2910 fo
->copied
= data
->len
;
2912 if (tcp_transmit_skb(sk
, syn_data
, 0, sk
->sk_allocation
) == 0) {
2913 tp
->syn_data
= (fo
->copied
> 0);
2914 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
2920 /* Send a regular SYN with Fast Open cookie request option */
2921 if (fo
->cookie
.len
> 0)
2923 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
2925 tp
->syn_fastopen
= 0;
2926 kfree_skb(syn_data
);
2928 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
2932 /* Build a SYN and send it off. */
2933 int tcp_connect(struct sock
*sk
)
2935 struct tcp_sock
*tp
= tcp_sk(sk
);
2936 struct sk_buff
*buff
;
2939 tcp_connect_init(sk
);
2941 if (unlikely(tp
->repair
)) {
2942 tcp_finish_connect(sk
, NULL
);
2946 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
2947 if (unlikely(buff
== NULL
))
2950 /* Reserve space for headers. */
2951 skb_reserve(buff
, MAX_TCP_HEADER
);
2953 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
2954 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
2955 tcp_connect_queue_skb(sk
, buff
);
2956 TCP_ECN_send_syn(sk
, buff
);
2958 /* Send off SYN; include data in Fast Open. */
2959 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
2960 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
2961 if (err
== -ECONNREFUSED
)
2964 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2965 * in order to make this packet get counted in tcpOutSegs.
2967 tp
->snd_nxt
= tp
->write_seq
;
2968 tp
->pushed_seq
= tp
->write_seq
;
2969 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
2971 /* Timer for repeating the SYN until an answer. */
2972 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2973 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
2976 EXPORT_SYMBOL(tcp_connect
);
2978 /* Send out a delayed ack, the caller does the policy checking
2979 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2982 void tcp_send_delayed_ack(struct sock
*sk
)
2984 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2985 int ato
= icsk
->icsk_ack
.ato
;
2986 unsigned long timeout
;
2988 if (ato
> TCP_DELACK_MIN
) {
2989 const struct tcp_sock
*tp
= tcp_sk(sk
);
2990 int max_ato
= HZ
/ 2;
2992 if (icsk
->icsk_ack
.pingpong
||
2993 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
2994 max_ato
= TCP_DELACK_MAX
;
2996 /* Slow path, intersegment interval is "high". */
2998 /* If some rtt estimate is known, use it to bound delayed ack.
2999 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3003 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
3009 ato
= min(ato
, max_ato
);
3012 /* Stay within the limit we were given */
3013 timeout
= jiffies
+ ato
;
3015 /* Use new timeout only if there wasn't a older one earlier. */
3016 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3017 /* If delack timer was blocked or is about to expire,
3020 if (icsk
->icsk_ack
.blocked
||
3021 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3026 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3027 timeout
= icsk
->icsk_ack
.timeout
;
3029 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3030 icsk
->icsk_ack
.timeout
= timeout
;
3031 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3034 /* This routine sends an ack and also updates the window. */
3035 void tcp_send_ack(struct sock
*sk
)
3037 struct sk_buff
*buff
;
3039 /* If we have been reset, we may not send again. */
3040 if (sk
->sk_state
== TCP_CLOSE
)
3043 /* We are not putting this on the write queue, so
3044 * tcp_transmit_skb() will set the ownership to this
3047 buff
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3049 inet_csk_schedule_ack(sk
);
3050 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3051 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3052 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3056 /* Reserve space for headers and prepare control bits. */
3057 skb_reserve(buff
, MAX_TCP_HEADER
);
3058 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3060 /* Send it off, this clears delayed acks for us. */
3061 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
3062 tcp_transmit_skb(sk
, buff
, 0, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3065 /* This routine sends a packet with an out of date sequence
3066 * number. It assumes the other end will try to ack it.
3068 * Question: what should we make while urgent mode?
3069 * 4.4BSD forces sending single byte of data. We cannot send
3070 * out of window data, because we have SND.NXT==SND.MAX...
3072 * Current solution: to send TWO zero-length segments in urgent mode:
3073 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3074 * out-of-date with SND.UNA-1 to probe window.
3076 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
3078 struct tcp_sock
*tp
= tcp_sk(sk
);
3079 struct sk_buff
*skb
;
3081 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3082 skb
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3086 /* Reserve space for headers and set control bits. */
3087 skb_reserve(skb
, MAX_TCP_HEADER
);
3088 /* Use a previous sequence. This should cause the other
3089 * end to send an ack. Don't queue or clone SKB, just
3092 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3093 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3094 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
3097 void tcp_send_window_probe(struct sock
*sk
)
3099 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3100 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3101 tcp_sk(sk
)->snd_nxt
= tcp_sk(sk
)->write_seq
;
3102 tcp_xmit_probe_skb(sk
, 0);
3106 /* Initiate keepalive or window probe from timer. */
3107 int tcp_write_wakeup(struct sock
*sk
)
3109 struct tcp_sock
*tp
= tcp_sk(sk
);
3110 struct sk_buff
*skb
;
3112 if (sk
->sk_state
== TCP_CLOSE
)
3115 if ((skb
= tcp_send_head(sk
)) != NULL
&&
3116 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3118 unsigned int mss
= tcp_current_mss(sk
);
3119 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3121 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3122 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3124 /* We are probing the opening of a window
3125 * but the window size is != 0
3126 * must have been a result SWS avoidance ( sender )
3128 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3130 seg_size
= min(seg_size
, mss
);
3131 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3132 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
3134 } else if (!tcp_skb_pcount(skb
))
3135 tcp_set_skb_tso_segs(sk
, skb
, mss
);
3137 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3138 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3139 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3141 tcp_event_new_data_sent(sk
, skb
);
3144 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3145 tcp_xmit_probe_skb(sk
, 1);
3146 return tcp_xmit_probe_skb(sk
, 0);
3150 /* A window probe timeout has occurred. If window is not closed send
3151 * a partial packet else a zero probe.
3153 void tcp_send_probe0(struct sock
*sk
)
3155 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3156 struct tcp_sock
*tp
= tcp_sk(sk
);
3159 err
= tcp_write_wakeup(sk
);
3161 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3162 /* Cancel probe timer, if it is not required. */
3163 icsk
->icsk_probes_out
= 0;
3164 icsk
->icsk_backoff
= 0;
3169 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
3170 icsk
->icsk_backoff
++;
3171 icsk
->icsk_probes_out
++;
3172 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3173 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3176 /* If packet was not sent due to local congestion,
3177 * do not backoff and do not remember icsk_probes_out.
3178 * Let local senders to fight for local resources.
3180 * Use accumulated backoff yet.
3182 if (!icsk
->icsk_probes_out
)
3183 icsk
->icsk_probes_out
= 1;
3184 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3185 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
3186 TCP_RESOURCE_PROBE_INTERVAL
),