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
, space
, sysctl_tcp_rmem
[2]);
226 space
= max_t(u32
, space
, sysctl_rmem_max
);
227 space
= min_t(u32
, space
, *window_clamp
);
228 while (space
> 65535 && (*rcv_wscale
) < 14) {
234 /* Set initial window to a value enough for senders starting with
235 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
236 * a limit on the initial window when mss is larger than 1460.
238 if (mss
> (1 << *rcv_wscale
)) {
239 int init_cwnd
= TCP_DEFAULT_INIT_RCVWND
;
242 max_t(u32
, (1460 * TCP_DEFAULT_INIT_RCVWND
) / mss
, 2);
243 /* when initializing use the value from init_rcv_wnd
244 * rather than the default from above
247 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
249 *rcv_wnd
= min(*rcv_wnd
, init_cwnd
* mss
);
252 /* Set the clamp no higher than max representable value */
253 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
255 EXPORT_SYMBOL(tcp_select_initial_window
);
257 /* Chose a new window to advertise, update state in tcp_sock for the
258 * socket, and return result with RFC1323 scaling applied. The return
259 * value can be stuffed directly into th->window for an outgoing
262 static u16
tcp_select_window(struct sock
*sk
)
264 struct tcp_sock
*tp
= tcp_sk(sk
);
265 u32 cur_win
= tcp_receive_window(tp
);
266 u32 new_win
= __tcp_select_window(sk
);
268 /* Never shrink the offered window */
269 if (new_win
< cur_win
) {
270 /* Danger Will Robinson!
271 * Don't update rcv_wup/rcv_wnd here or else
272 * we will not be able to advertise a zero
273 * window in time. --DaveM
275 * Relax Will Robinson.
277 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
279 tp
->rcv_wnd
= new_win
;
280 tp
->rcv_wup
= tp
->rcv_nxt
;
282 /* Make sure we do not exceed the maximum possible
285 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
286 new_win
= min(new_win
, MAX_TCP_WINDOW
);
288 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
290 /* RFC1323 scaling applied */
291 new_win
>>= tp
->rx_opt
.rcv_wscale
;
293 /* If we advertise zero window, disable fast path. */
300 /* Packet ECN state for a SYN-ACK */
301 static inline void TCP_ECN_send_synack(const struct tcp_sock
*tp
, struct sk_buff
*skb
)
303 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
304 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
305 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
308 /* Packet ECN state for a SYN. */
309 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
311 struct tcp_sock
*tp
= tcp_sk(sk
);
314 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1) {
315 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
316 tp
->ecn_flags
= TCP_ECN_OK
;
320 static __inline__
void
321 TCP_ECN_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
323 if (inet_rsk(req
)->ecn_ok
)
327 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
330 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
333 struct tcp_sock
*tp
= tcp_sk(sk
);
335 if (tp
->ecn_flags
& TCP_ECN_OK
) {
336 /* Not-retransmitted data segment: set ECT and inject CWR. */
337 if (skb
->len
!= tcp_header_len
&&
338 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
340 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
341 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
342 tcp_hdr(skb
)->cwr
= 1;
343 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
346 /* ACK or retransmitted segment: clear ECT|CE */
347 INET_ECN_dontxmit(sk
);
349 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
350 tcp_hdr(skb
)->ece
= 1;
354 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
355 * auto increment end seqno.
357 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
359 skb
->ip_summed
= CHECKSUM_PARTIAL
;
362 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
363 TCP_SKB_CB(skb
)->sacked
= 0;
365 skb_shinfo(skb
)->gso_segs
= 1;
366 skb_shinfo(skb
)->gso_size
= 0;
367 skb_shinfo(skb
)->gso_type
= 0;
369 TCP_SKB_CB(skb
)->seq
= seq
;
370 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
372 TCP_SKB_CB(skb
)->end_seq
= seq
;
375 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
377 return tp
->snd_una
!= tp
->snd_up
;
380 #define OPTION_SACK_ADVERTISE (1 << 0)
381 #define OPTION_TS (1 << 1)
382 #define OPTION_MD5 (1 << 2)
383 #define OPTION_WSCALE (1 << 3)
384 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
386 struct tcp_out_options
{
387 u16 options
; /* bit field of OPTION_* */
388 u16 mss
; /* 0 to disable */
389 u8 ws
; /* window scale, 0 to disable */
390 u8 num_sack_blocks
; /* number of SACK blocks to include */
391 u8 hash_size
; /* bytes in hash_location */
392 __u8
*hash_location
; /* temporary pointer, overloaded */
393 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
394 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
397 /* Write previously computed TCP options to the packet.
399 * Beware: Something in the Internet is very sensitive to the ordering of
400 * TCP options, we learned this through the hard way, so be careful here.
401 * Luckily we can at least blame others for their non-compliance but from
402 * inter-operatibility perspective it seems that we're somewhat stuck with
403 * the ordering which we have been using if we want to keep working with
404 * those broken things (not that it currently hurts anybody as there isn't
405 * particular reason why the ordering would need to be changed).
407 * At least SACK_PERM as the first option is known to lead to a disaster
408 * (but it may well be that other scenarios fail similarly).
410 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
411 struct tcp_out_options
*opts
)
413 u16 options
= opts
->options
; /* mungable copy */
415 if (unlikely(OPTION_MD5
& options
)) {
416 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
417 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
418 /* overload cookie hash location */
419 opts
->hash_location
= (__u8
*)ptr
;
423 if (unlikely(opts
->mss
)) {
424 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
425 (TCPOLEN_MSS
<< 16) |
429 if (likely(OPTION_TS
& options
)) {
430 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
431 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
432 (TCPOLEN_SACK_PERM
<< 16) |
433 (TCPOPT_TIMESTAMP
<< 8) |
435 options
&= ~OPTION_SACK_ADVERTISE
;
437 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
439 (TCPOPT_TIMESTAMP
<< 8) |
442 *ptr
++ = htonl(opts
->tsval
);
443 *ptr
++ = htonl(opts
->tsecr
);
446 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
447 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
449 (TCPOPT_SACK_PERM
<< 8) |
453 if (unlikely(OPTION_WSCALE
& options
)) {
454 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
455 (TCPOPT_WINDOW
<< 16) |
456 (TCPOLEN_WINDOW
<< 8) |
460 if (unlikely(opts
->num_sack_blocks
)) {
461 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
462 tp
->duplicate_sack
: tp
->selective_acks
;
465 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
468 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
469 TCPOLEN_SACK_PERBLOCK
)));
471 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
473 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
474 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
477 tp
->rx_opt
.dsack
= 0;
480 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
481 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
483 *ptr
++ = htonl((TCPOPT_EXP
<< 24) |
484 ((TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
) << 16) |
485 TCPOPT_FASTOPEN_MAGIC
);
487 memcpy(ptr
, foc
->val
, foc
->len
);
488 if ((foc
->len
& 3) == 2) {
489 u8
*align
= ((u8
*)ptr
) + foc
->len
;
490 align
[0] = align
[1] = TCPOPT_NOP
;
492 ptr
+= (foc
->len
+ 3) >> 2;
496 /* Compute TCP options for SYN packets. This is not the final
497 * network wire format yet.
499 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
500 struct tcp_out_options
*opts
,
501 struct tcp_md5sig_key
**md5
)
503 struct tcp_sock
*tp
= tcp_sk(sk
);
504 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
505 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
507 #ifdef CONFIG_TCP_MD5SIG
508 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
510 opts
->options
|= OPTION_MD5
;
511 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
517 /* We always get an MSS option. The option bytes which will be seen in
518 * normal data packets should timestamps be used, must be in the MSS
519 * advertised. But we subtract them from tp->mss_cache so that
520 * calculations in tcp_sendmsg are simpler etc. So account for this
521 * fact here if necessary. If we don't do this correctly, as a
522 * receiver we won't recognize data packets as being full sized when we
523 * should, and thus we won't abide by the delayed ACK rules correctly.
524 * SACKs don't matter, we never delay an ACK when we have any of those
526 opts
->mss
= tcp_advertise_mss(sk
);
527 remaining
-= TCPOLEN_MSS_ALIGNED
;
529 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
530 opts
->options
|= OPTION_TS
;
531 opts
->tsval
= TCP_SKB_CB(skb
)->when
+ tp
->tsoffset
;
532 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
533 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
535 if (likely(sysctl_tcp_window_scaling
)) {
536 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
537 opts
->options
|= OPTION_WSCALE
;
538 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
540 if (likely(sysctl_tcp_sack
)) {
541 opts
->options
|= OPTION_SACK_ADVERTISE
;
542 if (unlikely(!(OPTION_TS
& opts
->options
)))
543 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
546 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
547 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ fastopen
->cookie
.len
;
548 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
549 if (remaining
>= need
) {
550 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
551 opts
->fastopen_cookie
= &fastopen
->cookie
;
553 tp
->syn_fastopen
= 1;
557 return MAX_TCP_OPTION_SPACE
- remaining
;
560 /* Set up TCP options for SYN-ACKs. */
561 static unsigned int tcp_synack_options(struct sock
*sk
,
562 struct request_sock
*req
,
563 unsigned int mss
, struct sk_buff
*skb
,
564 struct tcp_out_options
*opts
,
565 struct tcp_md5sig_key
**md5
,
566 struct tcp_fastopen_cookie
*foc
)
568 struct inet_request_sock
*ireq
= inet_rsk(req
);
569 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
571 #ifdef CONFIG_TCP_MD5SIG
572 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
574 opts
->options
|= OPTION_MD5
;
575 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
577 /* We can't fit any SACK blocks in a packet with MD5 + TS
578 * options. There was discussion about disabling SACK
579 * rather than TS in order to fit in better with old,
580 * buggy kernels, but that was deemed to be unnecessary.
582 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
588 /* We always send an MSS option. */
590 remaining
-= TCPOLEN_MSS_ALIGNED
;
592 if (likely(ireq
->wscale_ok
)) {
593 opts
->ws
= ireq
->rcv_wscale
;
594 opts
->options
|= OPTION_WSCALE
;
595 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
597 if (likely(ireq
->tstamp_ok
)) {
598 opts
->options
|= OPTION_TS
;
599 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
600 opts
->tsecr
= req
->ts_recent
;
601 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
603 if (likely(ireq
->sack_ok
)) {
604 opts
->options
|= OPTION_SACK_ADVERTISE
;
605 if (unlikely(!ireq
->tstamp_ok
))
606 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
609 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
610 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
611 if (remaining
>= need
) {
612 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
613 opts
->fastopen_cookie
= foc
;
618 return MAX_TCP_OPTION_SPACE
- remaining
;
621 /* Compute TCP options for ESTABLISHED sockets. This is not the
622 * final wire format yet.
624 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
625 struct tcp_out_options
*opts
,
626 struct tcp_md5sig_key
**md5
)
628 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
629 struct tcp_sock
*tp
= tcp_sk(sk
);
630 unsigned int size
= 0;
631 unsigned int eff_sacks
;
633 #ifdef CONFIG_TCP_MD5SIG
634 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
635 if (unlikely(*md5
)) {
636 opts
->options
|= OPTION_MD5
;
637 size
+= TCPOLEN_MD5SIG_ALIGNED
;
643 if (likely(tp
->rx_opt
.tstamp_ok
)) {
644 opts
->options
|= OPTION_TS
;
645 opts
->tsval
= tcb
? tcb
->when
+ tp
->tsoffset
: 0;
646 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
647 size
+= TCPOLEN_TSTAMP_ALIGNED
;
650 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
651 if (unlikely(eff_sacks
)) {
652 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
653 opts
->num_sack_blocks
=
654 min_t(unsigned int, eff_sacks
,
655 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
656 TCPOLEN_SACK_PERBLOCK
);
657 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
658 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
665 /* TCP SMALL QUEUES (TSQ)
667 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
668 * to reduce RTT and bufferbloat.
669 * We do this using a special skb destructor (tcp_wfree).
671 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
672 * needs to be reallocated in a driver.
673 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
675 * Since transmit from skb destructor is forbidden, we use a tasklet
676 * to process all sockets that eventually need to send more skbs.
677 * We use one tasklet per cpu, with its own queue of sockets.
680 struct tasklet_struct tasklet
;
681 struct list_head head
; /* queue of tcp sockets */
683 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
685 static void tcp_tsq_handler(struct sock
*sk
)
687 if ((1 << sk
->sk_state
) &
688 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
689 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
))
690 tcp_write_xmit(sk
, tcp_current_mss(sk
), tcp_sk(sk
)->nonagle
,
694 * One tasklest per cpu tries to send more skbs.
695 * We run in tasklet context but need to disable irqs when
696 * transfering tsq->head because tcp_wfree() might
697 * interrupt us (non NAPI drivers)
699 static void tcp_tasklet_func(unsigned long data
)
701 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
704 struct list_head
*q
, *n
;
708 local_irq_save(flags
);
709 list_splice_init(&tsq
->head
, &list
);
710 local_irq_restore(flags
);
712 list_for_each_safe(q
, n
, &list
) {
713 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
714 list_del(&tp
->tsq_node
);
716 sk
= (struct sock
*)tp
;
719 if (!sock_owned_by_user(sk
)) {
722 /* defer the work to tcp_release_cb() */
723 set_bit(TCP_TSQ_DEFERRED
, &tp
->tsq_flags
);
727 clear_bit(TSQ_QUEUED
, &tp
->tsq_flags
);
732 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
733 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
734 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
735 (1UL << TCP_MTU_REDUCED_DEFERRED))
737 * tcp_release_cb - tcp release_sock() callback
740 * called from release_sock() to perform protocol dependent
741 * actions before socket release.
743 void tcp_release_cb(struct sock
*sk
)
745 struct tcp_sock
*tp
= tcp_sk(sk
);
746 unsigned long flags
, nflags
;
748 /* perform an atomic operation only if at least one flag is set */
750 flags
= tp
->tsq_flags
;
751 if (!(flags
& TCP_DEFERRED_ALL
))
753 nflags
= flags
& ~TCP_DEFERRED_ALL
;
754 } while (cmpxchg(&tp
->tsq_flags
, flags
, nflags
) != flags
);
756 if (flags
& (1UL << TCP_TSQ_DEFERRED
))
759 /* Here begins the tricky part :
760 * We are called from release_sock() with :
762 * 2) sk_lock.slock spinlock held
763 * 3) socket owned by us (sk->sk_lock.owned == 1)
765 * But following code is meant to be called from BH handlers,
766 * so we should keep BH disabled, but early release socket ownership
768 sock_release_ownership(sk
);
770 if (flags
& (1UL << TCP_WRITE_TIMER_DEFERRED
)) {
771 tcp_write_timer_handler(sk
);
774 if (flags
& (1UL << TCP_DELACK_TIMER_DEFERRED
)) {
775 tcp_delack_timer_handler(sk
);
778 if (flags
& (1UL << TCP_MTU_REDUCED_DEFERRED
)) {
779 inet_csk(sk
)->icsk_af_ops
->mtu_reduced(sk
);
783 EXPORT_SYMBOL(tcp_release_cb
);
785 void __init
tcp_tasklet_init(void)
789 for_each_possible_cpu(i
) {
790 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
792 INIT_LIST_HEAD(&tsq
->head
);
793 tasklet_init(&tsq
->tasklet
,
800 * Write buffer destructor automatically called from kfree_skb.
801 * We cant xmit new skbs from this context, as we might already
804 void tcp_wfree(struct sk_buff
*skb
)
806 struct sock
*sk
= skb
->sk
;
807 struct tcp_sock
*tp
= tcp_sk(sk
);
809 if (test_and_clear_bit(TSQ_THROTTLED
, &tp
->tsq_flags
) &&
810 !test_and_set_bit(TSQ_QUEUED
, &tp
->tsq_flags
)) {
812 struct tsq_tasklet
*tsq
;
814 /* Keep a ref on socket.
815 * This last ref will be released in tcp_tasklet_func()
817 atomic_sub(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
819 /* queue this socket to tasklet queue */
820 local_irq_save(flags
);
821 tsq
= &__get_cpu_var(tsq_tasklet
);
822 list_add(&tp
->tsq_node
, &tsq
->head
);
823 tasklet_schedule(&tsq
->tasklet
);
824 local_irq_restore(flags
);
830 /* This routine actually transmits TCP packets queued in by
831 * tcp_do_sendmsg(). This is used by both the initial
832 * transmission and possible later retransmissions.
833 * All SKB's seen here are completely headerless. It is our
834 * job to build the TCP header, and pass the packet down to
835 * IP so it can do the same plus pass the packet off to the
838 * We are working here with either a clone of the original
839 * SKB, or a fresh unique copy made by the retransmit engine.
841 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
844 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
845 struct inet_sock
*inet
;
847 struct tcp_skb_cb
*tcb
;
848 struct tcp_out_options opts
;
849 unsigned int tcp_options_size
, tcp_header_size
;
850 struct tcp_md5sig_key
*md5
;
854 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
856 /* If congestion control is doing timestamping, we must
857 * take such a timestamp before we potentially clone/copy.
859 if (icsk
->icsk_ca_ops
->flags
& TCP_CONG_RTT_STAMP
)
860 __net_timestamp(skb
);
862 if (likely(clone_it
)) {
863 const struct sk_buff
*fclone
= skb
+ 1;
865 if (unlikely(skb
->fclone
== SKB_FCLONE_ORIG
&&
866 fclone
->fclone
== SKB_FCLONE_CLONE
))
867 NET_INC_STATS_BH(sock_net(sk
),
868 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
870 if (unlikely(skb_cloned(skb
)))
871 skb
= pskb_copy(skb
, gfp_mask
);
873 skb
= skb_clone(skb
, gfp_mask
);
880 tcb
= TCP_SKB_CB(skb
);
881 memset(&opts
, 0, sizeof(opts
));
883 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
884 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
886 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
888 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
890 if (tcp_packets_in_flight(tp
) == 0)
891 tcp_ca_event(sk
, CA_EVENT_TX_START
);
893 /* if no packet is in qdisc/device queue, then allow XPS to select
896 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) == 0;
898 skb_push(skb
, tcp_header_size
);
899 skb_reset_transport_header(skb
);
903 skb
->destructor
= tcp_wfree
;
904 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
906 /* Build TCP header and checksum it. */
908 th
->source
= inet
->inet_sport
;
909 th
->dest
= inet
->inet_dport
;
910 th
->seq
= htonl(tcb
->seq
);
911 th
->ack_seq
= htonl(tp
->rcv_nxt
);
912 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
915 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
916 /* RFC1323: The window in SYN & SYN/ACK segments
919 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
921 th
->window
= htons(tcp_select_window(sk
));
926 /* The urg_mode check is necessary during a below snd_una win probe */
927 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
928 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
929 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
931 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
932 th
->urg_ptr
= htons(0xFFFF);
937 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
938 if (likely((tcb
->tcp_flags
& TCPHDR_SYN
) == 0))
939 TCP_ECN_send(sk
, skb
, tcp_header_size
);
941 #ifdef CONFIG_TCP_MD5SIG
942 /* Calculate the MD5 hash, as we have all we need now */
944 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
945 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
950 icsk
->icsk_af_ops
->send_check(sk
, skb
);
952 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
953 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
955 if (skb
->len
!= tcp_header_size
)
956 tcp_event_data_sent(tp
, sk
);
958 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
959 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
960 tcp_skb_pcount(skb
));
962 err
= icsk
->icsk_af_ops
->queue_xmit(skb
, &inet
->cork
.fl
);
963 if (likely(err
<= 0))
966 tcp_enter_cwr(sk
, 1);
968 return net_xmit_eval(err
);
971 /* This routine just queues the buffer for sending.
973 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
974 * otherwise socket can stall.
976 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
978 struct tcp_sock
*tp
= tcp_sk(sk
);
980 /* Advance write_seq and place onto the write_queue. */
981 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
982 skb_header_release(skb
);
983 tcp_add_write_queue_tail(sk
, skb
);
984 sk
->sk_wmem_queued
+= skb
->truesize
;
985 sk_mem_charge(sk
, skb
->truesize
);
988 /* Initialize TSO segments for a packet. */
989 static void tcp_set_skb_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
990 unsigned int mss_now
)
992 /* Make sure we own this skb before messing gso_size/gso_segs */
993 WARN_ON_ONCE(skb_cloned(skb
));
995 if (skb
->len
<= mss_now
|| !sk_can_gso(sk
) ||
996 skb
->ip_summed
== CHECKSUM_NONE
) {
997 /* Avoid the costly divide in the normal
1000 skb_shinfo(skb
)->gso_segs
= 1;
1001 skb_shinfo(skb
)->gso_size
= 0;
1002 skb_shinfo(skb
)->gso_type
= 0;
1004 skb_shinfo(skb
)->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
1005 skb_shinfo(skb
)->gso_size
= mss_now
;
1006 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
1010 /* When a modification to fackets out becomes necessary, we need to check
1011 * skb is counted to fackets_out or not.
1013 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1016 struct tcp_sock
*tp
= tcp_sk(sk
);
1018 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1021 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1022 tp
->fackets_out
-= decr
;
1025 /* Pcount in the middle of the write queue got changed, we need to do various
1026 * tweaks to fix counters
1028 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1030 struct tcp_sock
*tp
= tcp_sk(sk
);
1032 tp
->packets_out
-= decr
;
1034 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1035 tp
->sacked_out
-= decr
;
1036 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1037 tp
->retrans_out
-= decr
;
1038 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1039 tp
->lost_out
-= decr
;
1041 /* Reno case is special. Sigh... */
1042 if (tcp_is_reno(tp
) && decr
> 0)
1043 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1045 tcp_adjust_fackets_out(sk
, skb
, decr
);
1047 if (tp
->lost_skb_hint
&&
1048 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1049 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1050 tp
->lost_cnt_hint
-= decr
;
1052 tcp_verify_left_out(tp
);
1055 /* Function to create two new TCP segments. Shrinks the given segment
1056 * to the specified size and appends a new segment with the rest of the
1057 * packet to the list. This won't be called frequently, I hope.
1058 * Remember, these are still headerless SKBs at this point.
1060 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1061 unsigned int mss_now
)
1063 struct tcp_sock
*tp
= tcp_sk(sk
);
1064 struct sk_buff
*buff
;
1065 int nsize
, old_factor
;
1069 if (WARN_ON(len
> skb
->len
))
1072 nsize
= skb_headlen(skb
) - len
;
1076 if (skb_unclone(skb
, GFP_ATOMIC
))
1079 /* Get a new skb... force flag on. */
1080 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
1082 return -ENOMEM
; /* We'll just try again later. */
1084 sk
->sk_wmem_queued
+= buff
->truesize
;
1085 sk_mem_charge(sk
, buff
->truesize
);
1086 nlen
= skb
->len
- len
- nsize
;
1087 buff
->truesize
+= nlen
;
1088 skb
->truesize
-= nlen
;
1090 /* Correct the sequence numbers. */
1091 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1092 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1093 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1095 /* PSH and FIN should only be set in the second packet. */
1096 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1097 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1098 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1099 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1101 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1102 /* Copy and checksum data tail into the new buffer. */
1103 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1104 skb_put(buff
, nsize
),
1109 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1111 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1112 skb_split(skb
, buff
, len
);
1115 buff
->ip_summed
= skb
->ip_summed
;
1117 /* Looks stupid, but our code really uses when of
1118 * skbs, which it never sent before. --ANK
1120 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
1121 buff
->tstamp
= skb
->tstamp
;
1123 old_factor
= tcp_skb_pcount(skb
);
1125 /* Fix up tso_factor for both original and new SKB. */
1126 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1127 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1129 /* If this packet has been sent out already, we must
1130 * adjust the various packet counters.
1132 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1133 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1134 tcp_skb_pcount(buff
);
1137 tcp_adjust_pcount(sk
, skb
, diff
);
1140 /* Link BUFF into the send queue. */
1141 skb_header_release(buff
);
1142 tcp_insert_write_queue_after(skb
, buff
, sk
);
1147 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1148 * eventually). The difference is that pulled data not copied, but
1149 * immediately discarded.
1151 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1155 eat
= min_t(int, len
, skb_headlen(skb
));
1157 __skb_pull(skb
, eat
);
1164 for (i
= 0; i
< skb_shinfo(skb
)->nr_frags
; i
++) {
1165 int size
= skb_frag_size(&skb_shinfo(skb
)->frags
[i
]);
1168 skb_frag_unref(skb
, i
);
1171 skb_shinfo(skb
)->frags
[k
] = skb_shinfo(skb
)->frags
[i
];
1173 skb_shinfo(skb
)->frags
[k
].page_offset
+= eat
;
1174 skb_frag_size_sub(&skb_shinfo(skb
)->frags
[k
], eat
);
1180 skb_shinfo(skb
)->nr_frags
= k
;
1182 skb_reset_tail_pointer(skb
);
1183 skb
->data_len
-= len
;
1184 skb
->len
= skb
->data_len
;
1187 /* Remove acked data from a packet in the transmit queue. */
1188 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1190 if (skb_unclone(skb
, GFP_ATOMIC
))
1193 __pskb_trim_head(skb
, len
);
1195 TCP_SKB_CB(skb
)->seq
+= len
;
1196 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1198 skb
->truesize
-= len
;
1199 sk
->sk_wmem_queued
-= len
;
1200 sk_mem_uncharge(sk
, len
);
1201 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1203 /* Any change of skb->len requires recalculation of tso factor. */
1204 if (tcp_skb_pcount(skb
) > 1)
1205 tcp_set_skb_tso_segs(sk
, skb
, tcp_skb_mss(skb
));
1210 /* Calculate MSS not accounting any TCP options. */
1211 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1213 const struct tcp_sock
*tp
= tcp_sk(sk
);
1214 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1217 /* Calculate base mss without TCP options:
1218 It is MMS_S - sizeof(tcphdr) of rfc1122
1220 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1222 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1223 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1224 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1226 if (dst
&& dst_allfrag(dst
))
1227 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1230 /* Clamp it (mss_clamp does not include tcp options) */
1231 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1232 mss_now
= tp
->rx_opt
.mss_clamp
;
1234 /* Now subtract optional transport overhead */
1235 mss_now
-= icsk
->icsk_ext_hdr_len
;
1237 /* Then reserve room for full set of TCP options and 8 bytes of data */
1243 /* Calculate MSS. Not accounting for SACKs here. */
1244 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1246 /* Subtract TCP options size, not including SACKs */
1247 return __tcp_mtu_to_mss(sk
, pmtu
) -
1248 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1251 /* Inverse of above */
1252 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1254 const struct tcp_sock
*tp
= tcp_sk(sk
);
1255 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1259 tp
->tcp_header_len
+
1260 icsk
->icsk_ext_hdr_len
+
1261 icsk
->icsk_af_ops
->net_header_len
;
1263 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1264 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1265 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1267 if (dst
&& dst_allfrag(dst
))
1268 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1273 /* MTU probing init per socket */
1274 void tcp_mtup_init(struct sock
*sk
)
1276 struct tcp_sock
*tp
= tcp_sk(sk
);
1277 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1279 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
1280 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1281 icsk
->icsk_af_ops
->net_header_len
;
1282 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
1283 icsk
->icsk_mtup
.probe_size
= 0;
1285 EXPORT_SYMBOL(tcp_mtup_init
);
1287 /* This function synchronize snd mss to current pmtu/exthdr set.
1289 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1290 for TCP options, but includes only bare TCP header.
1292 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1293 It is minimum of user_mss and mss received with SYN.
1294 It also does not include TCP options.
1296 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1298 tp->mss_cache is current effective sending mss, including
1299 all tcp options except for SACKs. It is evaluated,
1300 taking into account current pmtu, but never exceeds
1301 tp->rx_opt.mss_clamp.
1303 NOTE1. rfc1122 clearly states that advertised MSS
1304 DOES NOT include either tcp or ip options.
1306 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1307 are READ ONLY outside this function. --ANK (980731)
1309 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1311 struct tcp_sock
*tp
= tcp_sk(sk
);
1312 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1315 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1316 icsk
->icsk_mtup
.search_high
= pmtu
;
1318 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1319 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1321 /* And store cached results */
1322 icsk
->icsk_pmtu_cookie
= pmtu
;
1323 if (icsk
->icsk_mtup
.enabled
)
1324 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1325 tp
->mss_cache
= mss_now
;
1329 EXPORT_SYMBOL(tcp_sync_mss
);
1331 /* Compute the current effective MSS, taking SACKs and IP options,
1332 * and even PMTU discovery events into account.
1334 unsigned int tcp_current_mss(struct sock
*sk
)
1336 const struct tcp_sock
*tp
= tcp_sk(sk
);
1337 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1339 unsigned int header_len
;
1340 struct tcp_out_options opts
;
1341 struct tcp_md5sig_key
*md5
;
1343 mss_now
= tp
->mss_cache
;
1346 u32 mtu
= dst_mtu(dst
);
1347 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1348 mss_now
= tcp_sync_mss(sk
, mtu
);
1351 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1352 sizeof(struct tcphdr
);
1353 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1354 * some common options. If this is an odd packet (because we have SACK
1355 * blocks etc) then our calculated header_len will be different, and
1356 * we have to adjust mss_now correspondingly */
1357 if (header_len
!= tp
->tcp_header_len
) {
1358 int delta
= (int) header_len
- tp
->tcp_header_len
;
1365 /* Congestion window validation. (RFC2861) */
1366 static void tcp_cwnd_validate(struct sock
*sk
)
1368 struct tcp_sock
*tp
= tcp_sk(sk
);
1370 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1371 /* Network is feed fully. */
1372 tp
->snd_cwnd_used
= 0;
1373 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1375 /* Network starves. */
1376 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1377 tp
->snd_cwnd_used
= tp
->packets_out
;
1379 if (sysctl_tcp_slow_start_after_idle
&&
1380 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1381 tcp_cwnd_application_limited(sk
);
1385 /* Returns the portion of skb which can be sent right away without
1386 * introducing MSS oddities to segment boundaries. In rare cases where
1387 * mss_now != mss_cache, we will request caller to create a small skb
1388 * per input skb which could be mostly avoided here (if desired).
1390 * We explicitly want to create a request for splitting write queue tail
1391 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1392 * thus all the complexity (cwnd_len is always MSS multiple which we
1393 * return whenever allowed by the other factors). Basically we need the
1394 * modulo only when the receiver window alone is the limiting factor or
1395 * when we would be allowed to send the split-due-to-Nagle skb fully.
1397 static unsigned int tcp_mss_split_point(const struct sock
*sk
, const struct sk_buff
*skb
,
1398 unsigned int mss_now
, unsigned int max_segs
)
1400 const struct tcp_sock
*tp
= tcp_sk(sk
);
1401 u32 needed
, window
, max_len
;
1403 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1404 max_len
= mss_now
* max_segs
;
1406 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1409 needed
= min(skb
->len
, window
);
1411 if (max_len
<= needed
)
1414 return needed
- needed
% mss_now
;
1417 /* Can at least one segment of SKB be sent right now, according to the
1418 * congestion window rules? If so, return how many segments are allowed.
1420 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1421 const struct sk_buff
*skb
)
1423 u32 in_flight
, cwnd
;
1425 /* Don't be strict about the congestion window for the final FIN. */
1426 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1427 tcp_skb_pcount(skb
) == 1)
1430 in_flight
= tcp_packets_in_flight(tp
);
1431 cwnd
= tp
->snd_cwnd
;
1432 if (in_flight
< cwnd
)
1433 return (cwnd
- in_flight
);
1438 /* Initialize TSO state of a skb.
1439 * This must be invoked the first time we consider transmitting
1440 * SKB onto the wire.
1442 static int tcp_init_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
1443 unsigned int mss_now
)
1445 int tso_segs
= tcp_skb_pcount(skb
);
1447 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1448 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1449 tso_segs
= tcp_skb_pcount(skb
);
1454 /* Minshall's variant of the Nagle send check. */
1455 static inline bool tcp_minshall_check(const struct tcp_sock
*tp
)
1457 return after(tp
->snd_sml
, tp
->snd_una
) &&
1458 !after(tp
->snd_sml
, tp
->snd_nxt
);
1461 /* Return false, if packet can be sent now without violation Nagle's rules:
1462 * 1. It is full sized.
1463 * 2. Or it contains FIN. (already checked by caller)
1464 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1465 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1466 * With Minshall's modification: all sent small packets are ACKed.
1468 static inline bool tcp_nagle_check(const struct tcp_sock
*tp
,
1469 const struct sk_buff
*skb
,
1470 unsigned int mss_now
, int nonagle
)
1472 return skb
->len
< mss_now
&&
1473 ((nonagle
& TCP_NAGLE_CORK
) ||
1474 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1477 /* Return true if the Nagle test allows this packet to be
1480 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1481 unsigned int cur_mss
, int nonagle
)
1483 /* Nagle rule does not apply to frames, which sit in the middle of the
1484 * write_queue (they have no chances to get new data).
1486 * This is implemented in the callers, where they modify the 'nonagle'
1487 * argument based upon the location of SKB in the send queue.
1489 if (nonagle
& TCP_NAGLE_PUSH
)
1492 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1493 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1496 if (!tcp_nagle_check(tp
, skb
, cur_mss
, nonagle
))
1502 /* Does at least the first segment of SKB fit into the send window? */
1503 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1504 const struct sk_buff
*skb
,
1505 unsigned int cur_mss
)
1507 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1509 if (skb
->len
> cur_mss
)
1510 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1512 return !after(end_seq
, tcp_wnd_end(tp
));
1515 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1516 * should be put on the wire right now. If so, it returns the number of
1517 * packets allowed by the congestion window.
1519 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1520 unsigned int cur_mss
, int nonagle
)
1522 const struct tcp_sock
*tp
= tcp_sk(sk
);
1523 unsigned int cwnd_quota
;
1525 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1527 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1530 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1531 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1537 /* Test if sending is allowed right now. */
1538 bool tcp_may_send_now(struct sock
*sk
)
1540 const struct tcp_sock
*tp
= tcp_sk(sk
);
1541 struct sk_buff
*skb
= tcp_send_head(sk
);
1544 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1545 (tcp_skb_is_last(sk
, skb
) ?
1546 tp
->nonagle
: TCP_NAGLE_PUSH
));
1549 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1550 * which is put after SKB on the list. It is very much like
1551 * tcp_fragment() except that it may make several kinds of assumptions
1552 * in order to speed up the splitting operation. In particular, we
1553 * know that all the data is in scatter-gather pages, and that the
1554 * packet has never been sent out before (and thus is not cloned).
1556 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1557 unsigned int mss_now
, gfp_t gfp
)
1559 struct sk_buff
*buff
;
1560 int nlen
= skb
->len
- len
;
1563 /* All of a TSO frame must be composed of paged data. */
1564 if (skb
->len
!= skb
->data_len
)
1565 return tcp_fragment(sk
, skb
, len
, mss_now
);
1567 buff
= sk_stream_alloc_skb(sk
, 0, gfp
);
1568 if (unlikely(buff
== NULL
))
1571 sk
->sk_wmem_queued
+= buff
->truesize
;
1572 sk_mem_charge(sk
, buff
->truesize
);
1573 buff
->truesize
+= nlen
;
1574 skb
->truesize
-= nlen
;
1576 /* Correct the sequence numbers. */
1577 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1578 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1579 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1581 /* PSH and FIN should only be set in the second packet. */
1582 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1583 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1584 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1586 /* This packet was never sent out yet, so no SACK bits. */
1587 TCP_SKB_CB(buff
)->sacked
= 0;
1589 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1590 skb_split(skb
, buff
, len
);
1592 /* Fix up tso_factor for both original and new SKB. */
1593 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1594 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1596 /* Link BUFF into the send queue. */
1597 skb_header_release(buff
);
1598 tcp_insert_write_queue_after(skb
, buff
, sk
);
1603 /* Try to defer sending, if possible, in order to minimize the amount
1604 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1606 * This algorithm is from John Heffner.
1608 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1610 struct tcp_sock
*tp
= tcp_sk(sk
);
1611 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1612 u32 send_win
, cong_win
, limit
, in_flight
;
1615 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1618 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1621 /* Defer for less than two clock ticks. */
1622 if (tp
->tso_deferred
&&
1623 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1626 in_flight
= tcp_packets_in_flight(tp
);
1628 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1630 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1632 /* From in_flight test above, we know that cwnd > in_flight. */
1633 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1635 limit
= min(send_win
, cong_win
);
1637 /* If a full-sized TSO skb can be sent, do it. */
1638 if (limit
>= min_t(unsigned int, sk
->sk_gso_max_size
,
1639 tp
->xmit_size_goal_segs
* tp
->mss_cache
))
1642 /* Middle in queue won't get any more data, full sendable already? */
1643 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1646 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1648 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1650 /* If at least some fraction of a window is available,
1653 chunk
/= win_divisor
;
1657 /* Different approach, try not to defer past a single
1658 * ACK. Receiver should ACK every other full sized
1659 * frame, so if we have space for more than 3 frames
1662 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1666 /* Ok, it looks like it is advisable to defer.
1667 * Do not rearm the timer if already set to not break TCP ACK clocking.
1669 if (!tp
->tso_deferred
)
1670 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1675 tp
->tso_deferred
= 0;
1679 /* Create a new MTU probe if we are ready.
1680 * MTU probe is regularly attempting to increase the path MTU by
1681 * deliberately sending larger packets. This discovers routing
1682 * changes resulting in larger path MTUs.
1684 * Returns 0 if we should wait to probe (no cwnd available),
1685 * 1 if a probe was sent,
1688 static int tcp_mtu_probe(struct sock
*sk
)
1690 struct tcp_sock
*tp
= tcp_sk(sk
);
1691 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1692 struct sk_buff
*skb
, *nskb
, *next
;
1699 /* Not currently probing/verifying,
1701 * have enough cwnd, and
1702 * not SACKing (the variable headers throw things off) */
1703 if (!icsk
->icsk_mtup
.enabled
||
1704 icsk
->icsk_mtup
.probe_size
||
1705 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1706 tp
->snd_cwnd
< 11 ||
1707 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1710 /* Very simple search strategy: just double the MSS. */
1711 mss_now
= tcp_current_mss(sk
);
1712 probe_size
= 2 * tp
->mss_cache
;
1713 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1714 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1715 /* TODO: set timer for probe_converge_event */
1719 /* Have enough data in the send queue to probe? */
1720 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1723 if (tp
->snd_wnd
< size_needed
)
1725 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1728 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1729 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1730 if (!tcp_packets_in_flight(tp
))
1736 /* We're allowed to probe. Build it now. */
1737 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1739 sk
->sk_wmem_queued
+= nskb
->truesize
;
1740 sk_mem_charge(sk
, nskb
->truesize
);
1742 skb
= tcp_send_head(sk
);
1744 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1745 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1746 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
1747 TCP_SKB_CB(nskb
)->sacked
= 0;
1749 nskb
->ip_summed
= skb
->ip_summed
;
1751 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1754 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1755 copy
= min_t(int, skb
->len
, probe_size
- len
);
1756 if (nskb
->ip_summed
) {
1757 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1759 __wsum csum
= skb_copy_and_csum_bits(skb
, 0,
1760 skb_put(nskb
, copy
),
1762 nskb
->csum
= csum_block_add(nskb
->csum
, csum
, len
);
1765 if (skb
->len
<= copy
) {
1766 /* We've eaten all the data from this skb.
1768 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1769 tcp_unlink_write_queue(skb
, sk
);
1770 sk_wmem_free_skb(sk
, skb
);
1772 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
1773 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1774 if (!skb_shinfo(skb
)->nr_frags
) {
1775 skb_pull(skb
, copy
);
1776 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1777 skb
->csum
= csum_partial(skb
->data
,
1780 __pskb_trim_head(skb
, copy
);
1781 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1783 TCP_SKB_CB(skb
)->seq
+= copy
;
1788 if (len
>= probe_size
)
1791 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1793 /* We're ready to send. If this fails, the probe will
1794 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1795 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1796 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1797 /* Decrement cwnd here because we are sending
1798 * effectively two packets. */
1800 tcp_event_new_data_sent(sk
, nskb
);
1802 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1803 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1804 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1812 /* This routine writes packets to the network. It advances the
1813 * send_head. This happens as incoming acks open up the remote
1816 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1817 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1818 * account rare use of URG, this is not a big flaw.
1820 * Send at most one packet when push_one > 0. Temporarily ignore
1821 * cwnd limit to force at most one packet out when push_one == 2.
1823 * Returns true, if no segments are in flight and we have queued segments,
1824 * but cannot send anything now because of SWS or another problem.
1826 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1827 int push_one
, gfp_t gfp
)
1829 struct tcp_sock
*tp
= tcp_sk(sk
);
1830 struct sk_buff
*skb
;
1831 unsigned int tso_segs
, sent_pkts
;
1838 /* Do MTU probing. */
1839 result
= tcp_mtu_probe(sk
);
1842 } else if (result
> 0) {
1847 while ((skb
= tcp_send_head(sk
))) {
1850 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1853 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
)
1854 goto repair
; /* Skip network transmission */
1856 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1859 /* Force out a loss probe pkt. */
1865 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1868 if (tso_segs
== 1 || !sk
->sk_gso_max_segs
) {
1869 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1870 (tcp_skb_is_last(sk
, skb
) ?
1871 nonagle
: TCP_NAGLE_PUSH
))))
1874 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
1878 /* TCP Small Queues :
1879 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
1881 * - better RTT estimation and ACK scheduling
1884 * Alas, some drivers / subsystems require a fair amount
1885 * of queued bytes to ensure line rate.
1886 * One example is wifi aggregation (802.11 AMPDU)
1888 limit
= max_t(unsigned int, sysctl_tcp_limit_output_bytes
,
1889 sk
->sk_pacing_rate
>> 10);
1891 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
) {
1892 set_bit(TSQ_THROTTLED
, &tp
->tsq_flags
);
1893 /* It is possible TX completion already happened
1894 * before we set TSQ_THROTTLED, so we must
1895 * test again the condition.
1896 * We abuse smp_mb__after_clear_bit() because
1897 * there is no smp_mb__after_set_bit() yet
1899 smp_mb__after_clear_bit();
1900 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
)
1905 if (tso_segs
> 1 && sk
->sk_gso_max_segs
&& !tcp_urg_mode(tp
))
1906 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1909 sk
->sk_gso_max_segs
));
1911 if (skb
->len
> limit
&&
1912 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
1915 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1917 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
1921 /* Advance the send_head. This one is sent out.
1922 * This call will increment packets_out.
1924 tcp_event_new_data_sent(sk
, skb
);
1926 tcp_minshall_update(tp
, mss_now
, skb
);
1927 sent_pkts
+= tcp_skb_pcount(skb
);
1933 if (likely(sent_pkts
)) {
1934 if (tcp_in_cwnd_reduction(sk
))
1935 tp
->prr_out
+= sent_pkts
;
1937 /* Send one loss probe per tail loss episode. */
1939 tcp_schedule_loss_probe(sk
);
1940 tcp_cwnd_validate(sk
);
1943 return (push_one
== 2) || (!tp
->packets_out
&& tcp_send_head(sk
));
1946 bool tcp_schedule_loss_probe(struct sock
*sk
)
1948 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1949 struct tcp_sock
*tp
= tcp_sk(sk
);
1950 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
1951 u32 rtt
= tp
->srtt
>> 3;
1953 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
))
1955 /* No consecutive loss probes. */
1956 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
1960 /* Don't do any loss probe on a Fast Open connection before 3WHS
1963 if (sk
->sk_state
== TCP_SYN_RECV
)
1966 /* TLP is only scheduled when next timer event is RTO. */
1967 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
1970 /* Schedule a loss probe in 2*RTT for SACK capable connections
1971 * in Open state, that are either limited by cwnd or application.
1973 if (sysctl_tcp_early_retrans
< 3 || !rtt
|| !tp
->packets_out
||
1974 !tcp_is_sack(tp
) || inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
1977 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
1981 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1982 * for delayed ack when there's one outstanding packet.
1985 if (tp
->packets_out
== 1)
1986 timeout
= max_t(u32
, timeout
,
1987 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
1988 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
1990 /* If RTO is shorter, just schedule TLP in its place. */
1991 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
1992 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
1993 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
1994 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
1999 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2004 /* When probe timeout (PTO) fires, send a new segment if one exists, else
2005 * retransmit the last segment.
2007 void tcp_send_loss_probe(struct sock
*sk
)
2009 struct tcp_sock
*tp
= tcp_sk(sk
);
2010 struct sk_buff
*skb
;
2012 int mss
= tcp_current_mss(sk
);
2015 if (tcp_send_head(sk
) != NULL
) {
2016 err
= tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2020 /* At most one outstanding TLP retransmission. */
2021 if (tp
->tlp_high_seq
)
2024 /* Retransmit last segment. */
2025 skb
= tcp_write_queue_tail(sk
);
2029 pcount
= tcp_skb_pcount(skb
);
2030 if (WARN_ON(!pcount
))
2033 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2034 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
)))
2036 skb
= tcp_write_queue_tail(sk
);
2039 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2042 err
= __tcp_retransmit_skb(sk
, skb
);
2044 /* Record snd_nxt for loss detection. */
2046 tp
->tlp_high_seq
= tp
->snd_nxt
;
2049 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2050 inet_csk(sk
)->icsk_rto
,
2054 NET_INC_STATS_BH(sock_net(sk
),
2055 LINUX_MIB_TCPLOSSPROBES
);
2059 /* Push out any pending frames which were held back due to
2060 * TCP_CORK or attempt at coalescing tiny packets.
2061 * The socket must be locked by the caller.
2063 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2066 /* If we are closed, the bytes will have to remain here.
2067 * In time closedown will finish, we empty the write queue and
2068 * all will be happy.
2070 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2073 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2074 sk_gfp_atomic(sk
, GFP_ATOMIC
)))
2075 tcp_check_probe_timer(sk
);
2078 /* Send _single_ skb sitting at the send head. This function requires
2079 * true push pending frames to setup probe timer etc.
2081 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2083 struct sk_buff
*skb
= tcp_send_head(sk
);
2085 BUG_ON(!skb
|| skb
->len
< mss_now
);
2087 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2090 /* This function returns the amount that we can raise the
2091 * usable window based on the following constraints
2093 * 1. The window can never be shrunk once it is offered (RFC 793)
2094 * 2. We limit memory per socket
2097 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2098 * RECV.NEXT + RCV.WIN fixed until:
2099 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2101 * i.e. don't raise the right edge of the window until you can raise
2102 * it at least MSS bytes.
2104 * Unfortunately, the recommended algorithm breaks header prediction,
2105 * since header prediction assumes th->window stays fixed.
2107 * Strictly speaking, keeping th->window fixed violates the receiver
2108 * side SWS prevention criteria. The problem is that under this rule
2109 * a stream of single byte packets will cause the right side of the
2110 * window to always advance by a single byte.
2112 * Of course, if the sender implements sender side SWS prevention
2113 * then this will not be a problem.
2115 * BSD seems to make the following compromise:
2117 * If the free space is less than the 1/4 of the maximum
2118 * space available and the free space is less than 1/2 mss,
2119 * then set the window to 0.
2120 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2121 * Otherwise, just prevent the window from shrinking
2122 * and from being larger than the largest representable value.
2124 * This prevents incremental opening of the window in the regime
2125 * where TCP is limited by the speed of the reader side taking
2126 * data out of the TCP receive queue. It does nothing about
2127 * those cases where the window is constrained on the sender side
2128 * because the pipeline is full.
2130 * BSD also seems to "accidentally" limit itself to windows that are a
2131 * multiple of MSS, at least until the free space gets quite small.
2132 * This would appear to be a side effect of the mbuf implementation.
2133 * Combining these two algorithms results in the observed behavior
2134 * of having a fixed window size at almost all times.
2136 * Below we obtain similar behavior by forcing the offered window to
2137 * a multiple of the mss when it is feasible to do so.
2139 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2140 * Regular options like TIMESTAMP are taken into account.
2142 u32
__tcp_select_window(struct sock
*sk
)
2144 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2145 struct tcp_sock
*tp
= tcp_sk(sk
);
2146 /* MSS for the peer's data. Previous versions used mss_clamp
2147 * here. I don't know if the value based on our guesses
2148 * of peer's MSS is better for the performance. It's more correct
2149 * but may be worse for the performance because of rcv_mss
2150 * fluctuations. --SAW 1998/11/1
2152 int mss
= icsk
->icsk_ack
.rcv_mss
;
2153 int free_space
= tcp_space(sk
);
2154 int full_space
= min_t(int, tp
->window_clamp
, tcp_full_space(sk
));
2157 if (unlikely(mss
> full_space
)) {
2162 if (free_space
< (full_space
>> 1)) {
2163 icsk
->icsk_ack
.quick
= 0;
2165 if (sk_under_memory_pressure(sk
))
2166 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2169 if (free_space
< mss
)
2173 if (free_space
> tp
->rcv_ssthresh
)
2174 free_space
= tp
->rcv_ssthresh
;
2176 /* Don't do rounding if we are using window scaling, since the
2177 * scaled window will not line up with the MSS boundary anyway.
2179 window
= tp
->rcv_wnd
;
2180 if (tp
->rx_opt
.rcv_wscale
) {
2181 window
= free_space
;
2183 /* Advertise enough space so that it won't get scaled away.
2184 * Import case: prevent zero window announcement if
2185 * 1<<rcv_wscale > mss.
2187 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2188 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2189 << tp
->rx_opt
.rcv_wscale
);
2191 /* Get the largest window that is a nice multiple of mss.
2192 * Window clamp already applied above.
2193 * If our current window offering is within 1 mss of the
2194 * free space we just keep it. This prevents the divide
2195 * and multiply from happening most of the time.
2196 * We also don't do any window rounding when the free space
2199 if (window
<= free_space
- mss
|| window
> free_space
)
2200 window
= (free_space
/ mss
) * mss
;
2201 else if (mss
== full_space
&&
2202 free_space
> window
+ (full_space
>> 1))
2203 window
= free_space
;
2209 /* Collapses two adjacent SKB's during retransmission. */
2210 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2212 struct tcp_sock
*tp
= tcp_sk(sk
);
2213 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2214 int skb_size
, next_skb_size
;
2216 skb_size
= skb
->len
;
2217 next_skb_size
= next_skb
->len
;
2219 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2221 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2223 tcp_unlink_write_queue(next_skb
, sk
);
2225 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
2228 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2229 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2231 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2232 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2234 /* Update sequence range on original skb. */
2235 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2237 /* Merge over control information. This moves PSH/FIN etc. over */
2238 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2240 /* All done, get rid of second SKB and account for it so
2241 * packet counting does not break.
2243 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2245 /* changed transmit queue under us so clear hints */
2246 tcp_clear_retrans_hints_partial(tp
);
2247 if (next_skb
== tp
->retransmit_skb_hint
)
2248 tp
->retransmit_skb_hint
= skb
;
2250 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2252 sk_wmem_free_skb(sk
, next_skb
);
2255 /* Check if coalescing SKBs is legal. */
2256 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2258 if (tcp_skb_pcount(skb
) > 1)
2260 /* TODO: SACK collapsing could be used to remove this condition */
2261 if (skb_shinfo(skb
)->nr_frags
!= 0)
2263 if (skb_cloned(skb
))
2265 if (skb
== tcp_send_head(sk
))
2267 /* Some heurestics for collapsing over SACK'd could be invented */
2268 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2274 /* Collapse packets in the retransmit queue to make to create
2275 * less packets on the wire. This is only done on retransmission.
2277 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2280 struct tcp_sock
*tp
= tcp_sk(sk
);
2281 struct sk_buff
*skb
= to
, *tmp
;
2284 if (!sysctl_tcp_retrans_collapse
)
2286 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2289 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2290 if (!tcp_can_collapse(sk
, skb
))
2302 /* Punt if not enough space exists in the first SKB for
2303 * the data in the second
2305 if (skb
->len
> skb_availroom(to
))
2308 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2311 tcp_collapse_retrans(sk
, to
);
2315 /* This retransmits one SKB. Policy decisions and retransmit queue
2316 * state updates are done by the caller. Returns non-zero if an
2317 * error occurred which prevented the send.
2319 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2321 struct tcp_sock
*tp
= tcp_sk(sk
);
2322 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2323 unsigned int cur_mss
;
2325 /* Inconslusive MTU probe */
2326 if (icsk
->icsk_mtup
.probe_size
) {
2327 icsk
->icsk_mtup
.probe_size
= 0;
2330 /* Do not sent more than we queued. 1/4 is reserved for possible
2331 * copying overhead: fragmentation, tunneling, mangling etc.
2333 if (atomic_read(&sk
->sk_wmem_alloc
) >
2334 min_t(u32
, sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2),
2338 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2339 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2341 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2345 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2346 return -EHOSTUNREACH
; /* Routing failure or similar. */
2348 cur_mss
= tcp_current_mss(sk
);
2350 /* If receiver has shrunk his window, and skb is out of
2351 * new window, do not retransmit it. The exception is the
2352 * case, when window is shrunk to zero. In this case
2353 * our retransmit serves as a zero window probe.
2355 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2356 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2359 if (skb
->len
> cur_mss
) {
2360 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
2361 return -ENOMEM
; /* We'll try again later. */
2363 int oldpcount
= tcp_skb_pcount(skb
);
2365 if (unlikely(oldpcount
> 1)) {
2366 if (skb_unclone(skb
, GFP_ATOMIC
))
2368 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2369 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2373 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2375 /* Some Solaris stacks overoptimize and ignore the FIN on a
2376 * retransmit when old data is attached. So strip it off
2377 * since it is cheap to do so and saves bytes on the network.
2380 (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
2381 tp
->snd_una
== (TCP_SKB_CB(skb
)->end_seq
- 1)) {
2382 if (!pskb_trim(skb
, 0)) {
2383 /* Reuse, even though it does some unnecessary work */
2384 tcp_init_nondata_skb(skb
, TCP_SKB_CB(skb
)->end_seq
- 1,
2385 TCP_SKB_CB(skb
)->tcp_flags
);
2386 skb
->ip_summed
= CHECKSUM_NONE
;
2390 /* Make a copy, if the first transmission SKB clone we made
2391 * is still in somebody's hands, else make a clone.
2393 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2395 /* make sure skb->data is aligned on arches that require it
2396 * and check if ack-trimming & collapsing extended the headroom
2397 * beyond what csum_start can cover.
2399 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2400 skb_headroom(skb
) >= 0xFFFF)) {
2401 struct sk_buff
*nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
,
2403 return nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2406 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2410 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2412 struct tcp_sock
*tp
= tcp_sk(sk
);
2413 int err
= __tcp_retransmit_skb(sk
, skb
);
2416 /* Update global TCP statistics. */
2417 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2419 tp
->total_retrans
++;
2421 #if FASTRETRANS_DEBUG > 0
2422 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2423 net_dbg_ratelimited("retrans_out leaked\n");
2426 if (!tp
->retrans_out
)
2427 tp
->lost_retrans_low
= tp
->snd_nxt
;
2428 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2429 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2431 /* Save stamp of the first retransmit. */
2432 if (!tp
->retrans_stamp
)
2433 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2435 /* snd_nxt is stored to detect loss of retransmitted segment,
2436 * see tcp_input.c tcp_sacktag_write_queue().
2438 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2441 if (tp
->undo_retrans
< 0)
2442 tp
->undo_retrans
= 0;
2443 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2447 /* Check if we forward retransmits are possible in the current
2448 * window/congestion state.
2450 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2452 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2453 const struct tcp_sock
*tp
= tcp_sk(sk
);
2455 /* Forward retransmissions are possible only during Recovery. */
2456 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2459 /* No forward retransmissions in Reno are possible. */
2460 if (tcp_is_reno(tp
))
2463 /* Yeah, we have to make difficult choice between forward transmission
2464 * and retransmission... Both ways have their merits...
2466 * For now we do not retransmit anything, while we have some new
2467 * segments to send. In the other cases, follow rule 3 for
2468 * NextSeg() specified in RFC3517.
2471 if (tcp_may_send_now(sk
))
2477 /* This gets called after a retransmit timeout, and the initially
2478 * retransmitted data is acknowledged. It tries to continue
2479 * resending the rest of the retransmit queue, until either
2480 * we've sent it all or the congestion window limit is reached.
2481 * If doing SACK, the first ACK which comes back for a timeout
2482 * based retransmit packet might feed us FACK information again.
2483 * If so, we use it to avoid unnecessarily retransmissions.
2485 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2487 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2488 struct tcp_sock
*tp
= tcp_sk(sk
);
2489 struct sk_buff
*skb
;
2490 struct sk_buff
*hole
= NULL
;
2493 int fwd_rexmitting
= 0;
2495 if (!tp
->packets_out
)
2499 tp
->retransmit_high
= tp
->snd_una
;
2501 if (tp
->retransmit_skb_hint
) {
2502 skb
= tp
->retransmit_skb_hint
;
2503 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2504 if (after(last_lost
, tp
->retransmit_high
))
2505 last_lost
= tp
->retransmit_high
;
2507 skb
= tcp_write_queue_head(sk
);
2508 last_lost
= tp
->snd_una
;
2511 tcp_for_write_queue_from(skb
, sk
) {
2512 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2514 if (skb
== tcp_send_head(sk
))
2516 /* we could do better than to assign each time */
2518 tp
->retransmit_skb_hint
= skb
;
2520 /* Assume this retransmit will generate
2521 * only one packet for congestion window
2522 * calculation purposes. This works because
2523 * tcp_retransmit_skb() will chop up the
2524 * packet to be MSS sized and all the
2525 * packet counting works out.
2527 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2530 if (fwd_rexmitting
) {
2532 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2534 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2536 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2537 tp
->retransmit_high
= last_lost
;
2538 if (!tcp_can_forward_retransmit(sk
))
2540 /* Backtrack if necessary to non-L'ed skb */
2548 } else if (!(sacked
& TCPCB_LOST
)) {
2549 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2554 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2555 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2556 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2558 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2561 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2564 if (tcp_retransmit_skb(sk
, skb
)) {
2565 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2568 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2570 if (tcp_in_cwnd_reduction(sk
))
2571 tp
->prr_out
+= tcp_skb_pcount(skb
);
2573 if (skb
== tcp_write_queue_head(sk
))
2574 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2575 inet_csk(sk
)->icsk_rto
,
2580 /* We allow to exceed memory limits for FIN packets to expedite
2581 * connection tear down and (memory) recovery.
2582 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2583 * or even be forced to close flow without any FIN.
2585 static void sk_forced_wmem_schedule(struct sock
*sk
, int size
)
2589 if (size
<= sk
->sk_forward_alloc
)
2591 amt
= sk_mem_pages(size
);
2592 sk
->sk_forward_alloc
+= amt
* SK_MEM_QUANTUM
;
2593 sk_memory_allocated_add(sk
, amt
, &status
);
2596 /* Send a FIN. The caller locks the socket for us.
2597 * We should try to send a FIN packet really hard, but eventually give up.
2599 void tcp_send_fin(struct sock
*sk
)
2601 struct sk_buff
*skb
, *tskb
= tcp_write_queue_tail(sk
);
2602 struct tcp_sock
*tp
= tcp_sk(sk
);
2604 /* Optimization, tack on the FIN if we have one skb in write queue and
2605 * this skb was not yet sent, or we are under memory pressure.
2606 * Note: in the latter case, FIN packet will be sent after a timeout,
2607 * as TCP stack thinks it has already been transmitted.
2609 if (tskb
&& (tcp_send_head(sk
) || sk_under_memory_pressure(sk
))) {
2611 TCP_SKB_CB(tskb
)->tcp_flags
|= TCPHDR_FIN
;
2612 TCP_SKB_CB(tskb
)->end_seq
++;
2614 if (!tcp_send_head(sk
)) {
2615 /* This means tskb was already sent.
2616 * Pretend we included the FIN on previous transmit.
2617 * We need to set tp->snd_nxt to the value it would have
2618 * if FIN had been sent. This is because retransmit path
2619 * does not change tp->snd_nxt.
2625 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, sk
->sk_allocation
);
2626 if (unlikely(!skb
)) {
2631 skb_reserve(skb
, MAX_TCP_HEADER
);
2632 sk_forced_wmem_schedule(sk
, skb
->truesize
);
2633 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2634 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2635 TCPHDR_ACK
| TCPHDR_FIN
);
2636 tcp_queue_skb(sk
, skb
);
2638 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
), TCP_NAGLE_OFF
);
2641 /* We get here when a process closes a file descriptor (either due to
2642 * an explicit close() or as a byproduct of exit()'ing) and there
2643 * was unread data in the receive queue. This behavior is recommended
2644 * by RFC 2525, section 2.17. -DaveM
2646 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2648 struct sk_buff
*skb
;
2650 /* NOTE: No TCP options attached and we never retransmit this. */
2651 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2653 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2657 /* Reserve space for headers and prepare control bits. */
2658 skb_reserve(skb
, MAX_TCP_HEADER
);
2659 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2660 TCPHDR_ACK
| TCPHDR_RST
);
2662 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2663 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2664 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2666 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2669 /* Send a crossed SYN-ACK during socket establishment.
2670 * WARNING: This routine must only be called when we have already sent
2671 * a SYN packet that crossed the incoming SYN that caused this routine
2672 * to get called. If this assumption fails then the initial rcv_wnd
2673 * and rcv_wscale values will not be correct.
2675 int tcp_send_synack(struct sock
*sk
)
2677 struct sk_buff
*skb
;
2679 skb
= tcp_write_queue_head(sk
);
2680 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2681 pr_debug("%s: wrong queue state\n", __func__
);
2684 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
2685 if (skb_cloned(skb
)) {
2686 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2689 tcp_unlink_write_queue(skb
, sk
);
2690 skb_header_release(nskb
);
2691 __tcp_add_write_queue_head(sk
, nskb
);
2692 sk_wmem_free_skb(sk
, skb
);
2693 sk
->sk_wmem_queued
+= nskb
->truesize
;
2694 sk_mem_charge(sk
, nskb
->truesize
);
2698 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
2699 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2701 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2702 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2706 * tcp_make_synack - Prepare a SYN-ACK.
2707 * sk: listener socket
2708 * dst: dst entry attached to the SYNACK
2709 * req: request_sock pointer
2711 * Allocate one skb and build a SYNACK packet.
2712 * @dst is consumed : Caller should not use it again.
2714 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2715 struct request_sock
*req
,
2716 struct tcp_fastopen_cookie
*foc
)
2718 struct tcp_out_options opts
;
2719 struct inet_request_sock
*ireq
= inet_rsk(req
);
2720 struct tcp_sock
*tp
= tcp_sk(sk
);
2722 struct sk_buff
*skb
;
2723 struct tcp_md5sig_key
*md5
;
2724 int tcp_header_size
;
2727 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
+ 15, 1, GFP_ATOMIC
);
2728 if (unlikely(!skb
)) {
2732 /* Reserve space for headers. */
2733 skb_reserve(skb
, MAX_TCP_HEADER
);
2735 skb_dst_set(skb
, dst
);
2736 security_skb_owned_by(skb
, sk
);
2738 mss
= dst_metric_advmss(dst
);
2739 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2740 mss
= tp
->rx_opt
.user_mss
;
2742 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2744 /* Set this up on the first call only */
2745 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2747 /* limit the window selection if the user enforce a smaller rx buffer */
2748 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2749 (req
->window_clamp
> tcp_full_space(sk
) || req
->window_clamp
== 0))
2750 req
->window_clamp
= tcp_full_space(sk
);
2752 /* tcp_full_space because it is guaranteed to be the first packet */
2753 tcp_select_initial_window(tcp_full_space(sk
),
2754 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2759 dst_metric(dst
, RTAX_INITRWND
));
2760 ireq
->rcv_wscale
= rcv_wscale
;
2763 memset(&opts
, 0, sizeof(opts
));
2764 #ifdef CONFIG_SYN_COOKIES
2765 if (unlikely(req
->cookie_ts
))
2766 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2769 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2770 tcp_header_size
= tcp_synack_options(sk
, req
, mss
, skb
, &opts
, &md5
,
2773 skb_push(skb
, tcp_header_size
);
2774 skb_reset_transport_header(skb
);
2777 memset(th
, 0, sizeof(struct tcphdr
));
2780 TCP_ECN_make_synack(req
, th
);
2781 th
->source
= ireq
->loc_port
;
2782 th
->dest
= ireq
->rmt_port
;
2783 /* Setting of flags are superfluous here for callers (and ECE is
2784 * not even correctly set)
2786 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2787 TCPHDR_SYN
| TCPHDR_ACK
);
2789 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2790 /* XXX data is queued and acked as is. No buffer/window check */
2791 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
2793 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2794 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2795 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2796 th
->doff
= (tcp_header_size
>> 2);
2797 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
, tcp_skb_pcount(skb
));
2799 #ifdef CONFIG_TCP_MD5SIG
2800 /* Okay, we have all we need - do the md5 hash if needed */
2802 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2803 md5
, NULL
, req
, skb
);
2807 /* Do not fool tcpdump (if any), clean our debris */
2808 skb
->tstamp
.tv64
= 0;
2811 EXPORT_SYMBOL(tcp_make_synack
);
2813 /* Do all connect socket setups that can be done AF independent. */
2814 void tcp_connect_init(struct sock
*sk
)
2816 const struct dst_entry
*dst
= __sk_dst_get(sk
);
2817 struct tcp_sock
*tp
= tcp_sk(sk
);
2820 /* We'll fix this up when we get a response from the other end.
2821 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2823 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2824 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2826 #ifdef CONFIG_TCP_MD5SIG
2827 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2828 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2831 /* If user gave his TCP_MAXSEG, record it to clamp */
2832 if (tp
->rx_opt
.user_mss
)
2833 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2836 tcp_sync_mss(sk
, dst_mtu(dst
));
2838 if (!tp
->window_clamp
)
2839 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2840 tp
->advmss
= dst_metric_advmss(dst
);
2841 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2842 tp
->advmss
= tp
->rx_opt
.user_mss
;
2844 tcp_initialize_rcv_mss(sk
);
2846 /* limit the window selection if the user enforce a smaller rx buffer */
2847 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2848 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
2849 tp
->window_clamp
= tcp_full_space(sk
);
2851 tcp_select_initial_window(tcp_full_space(sk
),
2852 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2855 sysctl_tcp_window_scaling
,
2857 dst_metric(dst
, RTAX_INITRWND
));
2859 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2860 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2863 sock_reset_flag(sk
, SOCK_DONE
);
2866 tp
->snd_una
= tp
->write_seq
;
2867 tp
->snd_sml
= tp
->write_seq
;
2868 tp
->snd_up
= tp
->write_seq
;
2869 tp
->snd_nxt
= tp
->write_seq
;
2871 if (likely(!tp
->repair
))
2874 tp
->rcv_tstamp
= tcp_time_stamp
;
2875 tp
->rcv_wup
= tp
->rcv_nxt
;
2876 tp
->copied_seq
= tp
->rcv_nxt
;
2878 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2879 inet_csk(sk
)->icsk_retransmits
= 0;
2880 tcp_clear_retrans(tp
);
2883 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
2885 struct tcp_sock
*tp
= tcp_sk(sk
);
2886 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
2888 tcb
->end_seq
+= skb
->len
;
2889 skb_header_release(skb
);
2890 __tcp_add_write_queue_tail(sk
, skb
);
2891 sk
->sk_wmem_queued
+= skb
->truesize
;
2892 sk_mem_charge(sk
, skb
->truesize
);
2893 tp
->write_seq
= tcb
->end_seq
;
2894 tp
->packets_out
+= tcp_skb_pcount(skb
);
2897 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2898 * queue a data-only packet after the regular SYN, such that regular SYNs
2899 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2900 * only the SYN sequence, the data are retransmitted in the first ACK.
2901 * If cookie is not cached or other error occurs, falls back to send a
2902 * regular SYN with Fast Open cookie request option.
2904 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
2906 struct tcp_sock
*tp
= tcp_sk(sk
);
2907 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
2908 int syn_loss
= 0, space
, err
= 0;
2909 unsigned long last_syn_loss
= 0;
2910 struct sk_buff
*syn_data
;
2912 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
2913 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
2914 &syn_loss
, &last_syn_loss
);
2915 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2917 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
2918 fo
->cookie
.len
= -1;
2922 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
2923 fo
->cookie
.len
= -1;
2924 else if (fo
->cookie
.len
<= 0)
2927 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2928 * user-MSS. Reserve maximum option space for middleboxes that add
2929 * private TCP options. The cost is reduced data space in SYN :(
2931 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
2932 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2933 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
2934 MAX_TCP_OPTION_SPACE
;
2936 space
= min_t(size_t, space
, fo
->size
);
2938 /* limit to order-0 allocations */
2939 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
2941 syn_data
= sk_stream_alloc_skb(sk
, space
, sk
->sk_allocation
);
2944 syn_data
->ip_summed
= CHECKSUM_PARTIAL
;
2945 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
2946 skb_shinfo(syn_data
)->gso_segs
= 1;
2947 if (unlikely(memcpy_fromiovecend(skb_put(syn_data
, space
),
2948 fo
->data
->msg_iov
, 0, space
))) {
2949 kfree_skb(syn_data
);
2953 /* No more data pending in inet_wait_for_connect() */
2954 if (space
== fo
->size
)
2958 tcp_connect_queue_skb(sk
, syn_data
);
2960 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
2962 /* Now full SYN+DATA was cloned and sent (or not),
2963 * remove the SYN from the original skb (syn_data)
2964 * we keep in write queue in case of a retransmit, as we
2965 * also have the SYN packet (with no data) in the same queue.
2967 TCP_SKB_CB(syn_data
)->seq
++;
2968 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
2970 tp
->syn_data
= (fo
->copied
> 0);
2971 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFASTOPENACTIVE
);
2976 /* Send a regular SYN with Fast Open cookie request option */
2977 if (fo
->cookie
.len
> 0)
2979 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
2981 tp
->syn_fastopen
= 0;
2983 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
2987 /* Build a SYN and send it off. */
2988 int tcp_connect(struct sock
*sk
)
2990 struct tcp_sock
*tp
= tcp_sk(sk
);
2991 struct sk_buff
*buff
;
2994 tcp_connect_init(sk
);
2996 if (unlikely(tp
->repair
)) {
2997 tcp_finish_connect(sk
, NULL
);
3001 buff
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
);
3002 if (unlikely(!buff
))
3005 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3006 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
3007 tcp_connect_queue_skb(sk
, buff
);
3008 TCP_ECN_send_syn(sk
, buff
);
3010 /* Send off SYN; include data in Fast Open. */
3011 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3012 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3013 if (err
== -ECONNREFUSED
)
3016 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3017 * in order to make this packet get counted in tcpOutSegs.
3019 tp
->snd_nxt
= tp
->write_seq
;
3020 tp
->pushed_seq
= tp
->write_seq
;
3021 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3023 /* Timer for repeating the SYN until an answer. */
3024 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3025 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3028 EXPORT_SYMBOL(tcp_connect
);
3030 /* Send out a delayed ack, the caller does the policy checking
3031 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3034 void tcp_send_delayed_ack(struct sock
*sk
)
3036 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3037 int ato
= icsk
->icsk_ack
.ato
;
3038 unsigned long timeout
;
3040 if (ato
> TCP_DELACK_MIN
) {
3041 const struct tcp_sock
*tp
= tcp_sk(sk
);
3042 int max_ato
= HZ
/ 2;
3044 if (icsk
->icsk_ack
.pingpong
||
3045 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3046 max_ato
= TCP_DELACK_MAX
;
3048 /* Slow path, intersegment interval is "high". */
3050 /* If some rtt estimate is known, use it to bound delayed ack.
3051 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3055 int rtt
= max(tp
->srtt
>> 3, TCP_DELACK_MIN
);
3061 ato
= min(ato
, max_ato
);
3064 /* Stay within the limit we were given */
3065 timeout
= jiffies
+ ato
;
3067 /* Use new timeout only if there wasn't a older one earlier. */
3068 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3069 /* If delack timer was blocked or is about to expire,
3072 if (icsk
->icsk_ack
.blocked
||
3073 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3078 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3079 timeout
= icsk
->icsk_ack
.timeout
;
3081 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3082 icsk
->icsk_ack
.timeout
= timeout
;
3083 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3086 /* This routine sends an ack and also updates the window. */
3087 void tcp_send_ack(struct sock
*sk
)
3089 struct sk_buff
*buff
;
3091 /* If we have been reset, we may not send again. */
3092 if (sk
->sk_state
== TCP_CLOSE
)
3095 /* We are not putting this on the write queue, so
3096 * tcp_transmit_skb() will set the ownership to this
3099 buff
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3101 inet_csk_schedule_ack(sk
);
3102 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3103 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3104 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3108 /* Reserve space for headers and prepare control bits. */
3109 skb_reserve(buff
, MAX_TCP_HEADER
);
3110 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3112 /* Send it off, this clears delayed acks for us. */
3113 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
3114 tcp_transmit_skb(sk
, buff
, 0, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3117 /* This routine sends a packet with an out of date sequence
3118 * number. It assumes the other end will try to ack it.
3120 * Question: what should we make while urgent mode?
3121 * 4.4BSD forces sending single byte of data. We cannot send
3122 * out of window data, because we have SND.NXT==SND.MAX...
3124 * Current solution: to send TWO zero-length segments in urgent mode:
3125 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3126 * out-of-date with SND.UNA-1 to probe window.
3128 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
3130 struct tcp_sock
*tp
= tcp_sk(sk
);
3131 struct sk_buff
*skb
;
3133 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3134 skb
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3138 /* Reserve space for headers and set control bits. */
3139 skb_reserve(skb
, MAX_TCP_HEADER
);
3140 /* Use a previous sequence. This should cause the other
3141 * end to send an ack. Don't queue or clone SKB, just
3144 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3145 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3146 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
3149 void tcp_send_window_probe(struct sock
*sk
)
3151 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3152 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3153 tcp_xmit_probe_skb(sk
, 0);
3157 /* Initiate keepalive or window probe from timer. */
3158 int tcp_write_wakeup(struct sock
*sk
)
3160 struct tcp_sock
*tp
= tcp_sk(sk
);
3161 struct sk_buff
*skb
;
3163 if (sk
->sk_state
== TCP_CLOSE
)
3166 if ((skb
= tcp_send_head(sk
)) != NULL
&&
3167 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3169 unsigned int mss
= tcp_current_mss(sk
);
3170 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3172 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3173 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3175 /* We are probing the opening of a window
3176 * but the window size is != 0
3177 * must have been a result SWS avoidance ( sender )
3179 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3181 seg_size
= min(seg_size
, mss
);
3182 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3183 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
3185 } else if (!tcp_skb_pcount(skb
))
3186 tcp_set_skb_tso_segs(sk
, skb
, mss
);
3188 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3189 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3190 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3192 tcp_event_new_data_sent(sk
, skb
);
3195 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3196 tcp_xmit_probe_skb(sk
, 1);
3197 return tcp_xmit_probe_skb(sk
, 0);
3201 /* A window probe timeout has occurred. If window is not closed send
3202 * a partial packet else a zero probe.
3204 void tcp_send_probe0(struct sock
*sk
)
3206 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3207 struct tcp_sock
*tp
= tcp_sk(sk
);
3210 err
= tcp_write_wakeup(sk
);
3212 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3213 /* Cancel probe timer, if it is not required. */
3214 icsk
->icsk_probes_out
= 0;
3215 icsk
->icsk_backoff
= 0;
3220 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
3221 icsk
->icsk_backoff
++;
3222 icsk
->icsk_probes_out
++;
3223 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3224 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3227 /* If packet was not sent due to local congestion,
3228 * do not backoff and do not remember icsk_probes_out.
3229 * Let local senders to fight for local resources.
3231 * Use accumulated backoff yet.
3233 if (!icsk
->icsk_probes_out
)
3234 icsk
->icsk_probes_out
= 1;
3235 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3236 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
3237 TCP_RESOURCE_PROBE_INTERVAL
),