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 unsigned int sysctl_tcp_notsent_lowat __read_mostly
= UINT_MAX
;
69 EXPORT_SYMBOL(sysctl_tcp_notsent_lowat
);
71 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
72 int push_one
, gfp_t gfp
);
74 /* Account for new data that has been sent to the network. */
75 static void tcp_event_new_data_sent(struct sock
*sk
, const struct sk_buff
*skb
)
77 struct inet_connection_sock
*icsk
= inet_csk(sk
);
78 struct tcp_sock
*tp
= tcp_sk(sk
);
79 unsigned int prior_packets
= tp
->packets_out
;
81 tcp_advance_send_head(sk
, skb
);
82 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
84 tp
->packets_out
+= tcp_skb_pcount(skb
);
85 if (!prior_packets
|| icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
86 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
90 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
,
94 /* SND.NXT, if window was not shrunk.
95 * If window has been shrunk, what should we make? It is not clear at all.
96 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
97 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
98 * invalid. OK, let's make this for now:
100 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
102 const struct tcp_sock
*tp
= tcp_sk(sk
);
104 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
107 return tcp_wnd_end(tp
);
110 /* Calculate mss to advertise in SYN segment.
111 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
113 * 1. It is independent of path mtu.
114 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
115 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
116 * attached devices, because some buggy hosts are confused by
118 * 4. We do not make 3, we advertise MSS, calculated from first
119 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
120 * This may be overridden via information stored in routing table.
121 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
122 * probably even Jumbo".
124 static __u16
tcp_advertise_mss(struct sock
*sk
)
126 struct tcp_sock
*tp
= tcp_sk(sk
);
127 const struct dst_entry
*dst
= __sk_dst_get(sk
);
128 int mss
= tp
->advmss
;
131 unsigned int metric
= dst_metric_advmss(dst
);
142 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
143 * This is the first part of cwnd validation mechanism. */
144 static void tcp_cwnd_restart(struct sock
*sk
, const struct dst_entry
*dst
)
146 struct tcp_sock
*tp
= tcp_sk(sk
);
147 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
148 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
149 u32 cwnd
= tp
->snd_cwnd
;
151 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
153 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
154 restart_cwnd
= min(restart_cwnd
, cwnd
);
156 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
158 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
159 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
160 tp
->snd_cwnd_used
= 0;
163 /* Congestion state accounting after a packet has been sent. */
164 static void tcp_event_data_sent(struct tcp_sock
*tp
,
167 struct inet_connection_sock
*icsk
= inet_csk(sk
);
168 const u32 now
= tcp_time_stamp
;
169 const struct dst_entry
*dst
= __sk_dst_get(sk
);
171 if (sysctl_tcp_slow_start_after_idle
&&
172 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
173 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
177 /* If it is a reply for ato after last received
178 * packet, enter pingpong mode.
180 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
&&
181 (!dst
|| !dst_metric(dst
, RTAX_QUICKACK
)))
182 icsk
->icsk_ack
.pingpong
= 1;
185 /* Account for an ACK we sent. */
186 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
188 tcp_dec_quickack_mode(sk
, pkts
);
189 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
193 u32
tcp_default_init_rwnd(u32 mss
)
195 /* Initial receive window should be twice of TCP_INIT_CWND to
196 * enable proper sending of new unsent data during fast recovery
197 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
198 * limit when mss is larger than 1460.
200 u32 init_rwnd
= TCP_INIT_CWND
* 2;
203 init_rwnd
= max((1460 * init_rwnd
) / mss
, 2U);
207 /* Determine a window scaling and initial window to offer.
208 * Based on the assumption that the given amount of space
209 * will be offered. Store the results in the tp structure.
210 * NOTE: for smooth operation initial space offering should
211 * be a multiple of mss if possible. We assume here that mss >= 1.
212 * This MUST be enforced by all callers.
214 void tcp_select_initial_window(int __space
, __u32 mss
,
215 __u32
*rcv_wnd
, __u32
*window_clamp
,
216 int wscale_ok
, __u8
*rcv_wscale
,
219 unsigned int space
= (__space
< 0 ? 0 : __space
);
221 /* If no clamp set the clamp to the max possible scaled window */
222 if (*window_clamp
== 0)
223 (*window_clamp
) = (65535 << 14);
224 space
= min(*window_clamp
, space
);
226 /* Quantize space offering to a multiple of mss if possible. */
228 space
= (space
/ mss
) * mss
;
230 /* NOTE: offering an initial window larger than 32767
231 * will break some buggy TCP stacks. If the admin tells us
232 * it is likely we could be speaking with such a buggy stack
233 * we will truncate our initial window offering to 32K-1
234 * unless the remote has sent us a window scaling option,
235 * which we interpret as a sign the remote TCP is not
236 * misinterpreting the window field as a signed quantity.
238 if (sysctl_tcp_workaround_signed_windows
)
239 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
245 /* Set window scaling on max possible window
246 * See RFC1323 for an explanation of the limit to 14
248 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
249 space
= min_t(u32
, space
, *window_clamp
);
250 while (space
> 65535 && (*rcv_wscale
) < 14) {
256 if (mss
> (1 << *rcv_wscale
)) {
257 if (!init_rcv_wnd
) /* Use default unless specified otherwise */
258 init_rcv_wnd
= tcp_default_init_rwnd(mss
);
259 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
262 /* Set the clamp no higher than max representable value */
263 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
265 EXPORT_SYMBOL(tcp_select_initial_window
);
267 /* Chose a new window to advertise, update state in tcp_sock for the
268 * socket, and return result with RFC1323 scaling applied. The return
269 * value can be stuffed directly into th->window for an outgoing
272 static u16
tcp_select_window(struct sock
*sk
)
274 struct tcp_sock
*tp
= tcp_sk(sk
);
275 u32 old_win
= tp
->rcv_wnd
;
276 u32 cur_win
= tcp_receive_window(tp
);
277 u32 new_win
= __tcp_select_window(sk
);
279 /* Never shrink the offered window */
280 if (new_win
< cur_win
) {
281 /* Danger Will Robinson!
282 * Don't update rcv_wup/rcv_wnd here or else
283 * we will not be able to advertise a zero
284 * window in time. --DaveM
286 * Relax Will Robinson.
289 NET_INC_STATS(sock_net(sk
),
290 LINUX_MIB_TCPWANTZEROWINDOWADV
);
291 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
293 tp
->rcv_wnd
= new_win
;
294 tp
->rcv_wup
= tp
->rcv_nxt
;
296 /* Make sure we do not exceed the maximum possible
299 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
300 new_win
= min(new_win
, MAX_TCP_WINDOW
);
302 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
304 /* RFC1323 scaling applied */
305 new_win
>>= tp
->rx_opt
.rcv_wscale
;
307 /* If we advertise zero window, disable fast path. */
311 NET_INC_STATS(sock_net(sk
),
312 LINUX_MIB_TCPTOZEROWINDOWADV
);
313 } else if (old_win
== 0) {
314 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFROMZEROWINDOWADV
);
320 /* Packet ECN state for a SYN-ACK */
321 static inline void TCP_ECN_send_synack(const struct tcp_sock
*tp
, struct sk_buff
*skb
)
323 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
324 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
325 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
328 /* Packet ECN state for a SYN. */
329 static inline void TCP_ECN_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
331 struct tcp_sock
*tp
= tcp_sk(sk
);
334 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1) {
335 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
336 tp
->ecn_flags
= TCP_ECN_OK
;
340 static __inline__
void
341 TCP_ECN_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
343 if (inet_rsk(req
)->ecn_ok
)
347 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
350 static inline void TCP_ECN_send(struct sock
*sk
, struct sk_buff
*skb
,
353 struct tcp_sock
*tp
= tcp_sk(sk
);
355 if (tp
->ecn_flags
& TCP_ECN_OK
) {
356 /* Not-retransmitted data segment: set ECT and inject CWR. */
357 if (skb
->len
!= tcp_header_len
&&
358 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
360 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
361 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
362 tcp_hdr(skb
)->cwr
= 1;
363 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
366 /* ACK or retransmitted segment: clear ECT|CE */
367 INET_ECN_dontxmit(sk
);
369 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
370 tcp_hdr(skb
)->ece
= 1;
374 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
375 * auto increment end seqno.
377 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
379 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
381 skb
->ip_summed
= CHECKSUM_PARTIAL
;
384 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
385 TCP_SKB_CB(skb
)->sacked
= 0;
387 shinfo
->gso_segs
= 1;
388 shinfo
->gso_size
= 0;
389 shinfo
->gso_type
= 0;
391 TCP_SKB_CB(skb
)->seq
= seq
;
392 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
394 TCP_SKB_CB(skb
)->end_seq
= seq
;
397 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
399 return tp
->snd_una
!= tp
->snd_up
;
402 #define OPTION_SACK_ADVERTISE (1 << 0)
403 #define OPTION_TS (1 << 1)
404 #define OPTION_MD5 (1 << 2)
405 #define OPTION_WSCALE (1 << 3)
406 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
408 struct tcp_out_options
{
409 u16 options
; /* bit field of OPTION_* */
410 u16 mss
; /* 0 to disable */
411 u8 ws
; /* window scale, 0 to disable */
412 u8 num_sack_blocks
; /* number of SACK blocks to include */
413 u8 hash_size
; /* bytes in hash_location */
414 __u8
*hash_location
; /* temporary pointer, overloaded */
415 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
416 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
419 /* Write previously computed TCP options to the packet.
421 * Beware: Something in the Internet is very sensitive to the ordering of
422 * TCP options, we learned this through the hard way, so be careful here.
423 * Luckily we can at least blame others for their non-compliance but from
424 * inter-operability perspective it seems that we're somewhat stuck with
425 * the ordering which we have been using if we want to keep working with
426 * those broken things (not that it currently hurts anybody as there isn't
427 * particular reason why the ordering would need to be changed).
429 * At least SACK_PERM as the first option is known to lead to a disaster
430 * (but it may well be that other scenarios fail similarly).
432 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
433 struct tcp_out_options
*opts
)
435 u16 options
= opts
->options
; /* mungable copy */
437 if (unlikely(OPTION_MD5
& options
)) {
438 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
439 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
440 /* overload cookie hash location */
441 opts
->hash_location
= (__u8
*)ptr
;
445 if (unlikely(opts
->mss
)) {
446 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
447 (TCPOLEN_MSS
<< 16) |
451 if (likely(OPTION_TS
& options
)) {
452 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
453 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
454 (TCPOLEN_SACK_PERM
<< 16) |
455 (TCPOPT_TIMESTAMP
<< 8) |
457 options
&= ~OPTION_SACK_ADVERTISE
;
459 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
461 (TCPOPT_TIMESTAMP
<< 8) |
464 *ptr
++ = htonl(opts
->tsval
);
465 *ptr
++ = htonl(opts
->tsecr
);
468 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
469 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
471 (TCPOPT_SACK_PERM
<< 8) |
475 if (unlikely(OPTION_WSCALE
& options
)) {
476 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
477 (TCPOPT_WINDOW
<< 16) |
478 (TCPOLEN_WINDOW
<< 8) |
482 if (unlikely(opts
->num_sack_blocks
)) {
483 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
484 tp
->duplicate_sack
: tp
->selective_acks
;
487 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
490 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
491 TCPOLEN_SACK_PERBLOCK
)));
493 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
495 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
496 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
499 tp
->rx_opt
.dsack
= 0;
502 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
503 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
505 *ptr
++ = htonl((TCPOPT_EXP
<< 24) |
506 ((TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
) << 16) |
507 TCPOPT_FASTOPEN_MAGIC
);
509 memcpy(ptr
, foc
->val
, foc
->len
);
510 if ((foc
->len
& 3) == 2) {
511 u8
*align
= ((u8
*)ptr
) + foc
->len
;
512 align
[0] = align
[1] = TCPOPT_NOP
;
514 ptr
+= (foc
->len
+ 3) >> 2;
518 /* Compute TCP options for SYN packets. This is not the final
519 * network wire format yet.
521 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
522 struct tcp_out_options
*opts
,
523 struct tcp_md5sig_key
**md5
)
525 struct tcp_sock
*tp
= tcp_sk(sk
);
526 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
527 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
529 #ifdef CONFIG_TCP_MD5SIG
530 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
532 opts
->options
|= OPTION_MD5
;
533 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
539 /* We always get an MSS option. The option bytes which will be seen in
540 * normal data packets should timestamps be used, must be in the MSS
541 * advertised. But we subtract them from tp->mss_cache so that
542 * calculations in tcp_sendmsg are simpler etc. So account for this
543 * fact here if necessary. If we don't do this correctly, as a
544 * receiver we won't recognize data packets as being full sized when we
545 * should, and thus we won't abide by the delayed ACK rules correctly.
546 * SACKs don't matter, we never delay an ACK when we have any of those
548 opts
->mss
= tcp_advertise_mss(sk
);
549 remaining
-= TCPOLEN_MSS_ALIGNED
;
551 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
552 opts
->options
|= OPTION_TS
;
553 opts
->tsval
= TCP_SKB_CB(skb
)->when
+ tp
->tsoffset
;
554 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
555 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
557 if (likely(sysctl_tcp_window_scaling
)) {
558 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
559 opts
->options
|= OPTION_WSCALE
;
560 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
562 if (likely(sysctl_tcp_sack
)) {
563 opts
->options
|= OPTION_SACK_ADVERTISE
;
564 if (unlikely(!(OPTION_TS
& opts
->options
)))
565 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
568 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
569 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ fastopen
->cookie
.len
;
570 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
571 if (remaining
>= need
) {
572 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
573 opts
->fastopen_cookie
= &fastopen
->cookie
;
575 tp
->syn_fastopen
= 1;
579 return MAX_TCP_OPTION_SPACE
- remaining
;
582 /* Set up TCP options for SYN-ACKs. */
583 static unsigned int tcp_synack_options(struct sock
*sk
,
584 struct request_sock
*req
,
585 unsigned int mss
, struct sk_buff
*skb
,
586 struct tcp_out_options
*opts
,
587 struct tcp_md5sig_key
**md5
,
588 struct tcp_fastopen_cookie
*foc
)
590 struct inet_request_sock
*ireq
= inet_rsk(req
);
591 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
593 #ifdef CONFIG_TCP_MD5SIG
594 *md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
596 opts
->options
|= OPTION_MD5
;
597 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
599 /* We can't fit any SACK blocks in a packet with MD5 + TS
600 * options. There was discussion about disabling SACK
601 * rather than TS in order to fit in better with old,
602 * buggy kernels, but that was deemed to be unnecessary.
604 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
610 /* We always send an MSS option. */
612 remaining
-= TCPOLEN_MSS_ALIGNED
;
614 if (likely(ireq
->wscale_ok
)) {
615 opts
->ws
= ireq
->rcv_wscale
;
616 opts
->options
|= OPTION_WSCALE
;
617 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
619 if (likely(ireq
->tstamp_ok
)) {
620 opts
->options
|= OPTION_TS
;
621 opts
->tsval
= TCP_SKB_CB(skb
)->when
;
622 opts
->tsecr
= req
->ts_recent
;
623 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
625 if (likely(ireq
->sack_ok
)) {
626 opts
->options
|= OPTION_SACK_ADVERTISE
;
627 if (unlikely(!ireq
->tstamp_ok
))
628 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
631 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
632 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
633 if (remaining
>= need
) {
634 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
635 opts
->fastopen_cookie
= foc
;
640 return MAX_TCP_OPTION_SPACE
- remaining
;
643 /* Compute TCP options for ESTABLISHED sockets. This is not the
644 * final wire format yet.
646 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
647 struct tcp_out_options
*opts
,
648 struct tcp_md5sig_key
**md5
)
650 struct tcp_skb_cb
*tcb
= skb
? TCP_SKB_CB(skb
) : NULL
;
651 struct tcp_sock
*tp
= tcp_sk(sk
);
652 unsigned int size
= 0;
653 unsigned int eff_sacks
;
657 #ifdef CONFIG_TCP_MD5SIG
658 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
659 if (unlikely(*md5
)) {
660 opts
->options
|= OPTION_MD5
;
661 size
+= TCPOLEN_MD5SIG_ALIGNED
;
667 if (likely(tp
->rx_opt
.tstamp_ok
)) {
668 opts
->options
|= OPTION_TS
;
669 opts
->tsval
= tcb
? tcb
->when
+ tp
->tsoffset
: 0;
670 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
671 size
+= TCPOLEN_TSTAMP_ALIGNED
;
674 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
675 if (unlikely(eff_sacks
)) {
676 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
677 opts
->num_sack_blocks
=
678 min_t(unsigned int, eff_sacks
,
679 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
680 TCPOLEN_SACK_PERBLOCK
);
681 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
682 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
689 /* TCP SMALL QUEUES (TSQ)
691 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
692 * to reduce RTT and bufferbloat.
693 * We do this using a special skb destructor (tcp_wfree).
695 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
696 * needs to be reallocated in a driver.
697 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
699 * Since transmit from skb destructor is forbidden, we use a tasklet
700 * to process all sockets that eventually need to send more skbs.
701 * We use one tasklet per cpu, with its own queue of sockets.
704 struct tasklet_struct tasklet
;
705 struct list_head head
; /* queue of tcp sockets */
707 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
709 static void tcp_tsq_handler(struct sock
*sk
)
711 if ((1 << sk
->sk_state
) &
712 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
713 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
))
714 tcp_write_xmit(sk
, tcp_current_mss(sk
), tcp_sk(sk
)->nonagle
,
718 * One tasklet per cpu tries to send more skbs.
719 * We run in tasklet context but need to disable irqs when
720 * transferring tsq->head because tcp_wfree() might
721 * interrupt us (non NAPI drivers)
723 static void tcp_tasklet_func(unsigned long data
)
725 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
728 struct list_head
*q
, *n
;
732 local_irq_save(flags
);
733 list_splice_init(&tsq
->head
, &list
);
734 local_irq_restore(flags
);
736 list_for_each_safe(q
, n
, &list
) {
737 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
738 list_del(&tp
->tsq_node
);
740 sk
= (struct sock
*)tp
;
743 if (!sock_owned_by_user(sk
)) {
746 /* defer the work to tcp_release_cb() */
747 set_bit(TCP_TSQ_DEFERRED
, &tp
->tsq_flags
);
751 clear_bit(TSQ_QUEUED
, &tp
->tsq_flags
);
756 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
757 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
758 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
759 (1UL << TCP_MTU_REDUCED_DEFERRED))
761 * tcp_release_cb - tcp release_sock() callback
764 * called from release_sock() to perform protocol dependent
765 * actions before socket release.
767 void tcp_release_cb(struct sock
*sk
)
769 struct tcp_sock
*tp
= tcp_sk(sk
);
770 unsigned long flags
, nflags
;
772 /* perform an atomic operation only if at least one flag is set */
774 flags
= tp
->tsq_flags
;
775 if (!(flags
& TCP_DEFERRED_ALL
))
777 nflags
= flags
& ~TCP_DEFERRED_ALL
;
778 } while (cmpxchg(&tp
->tsq_flags
, flags
, nflags
) != flags
);
780 if (flags
& (1UL << TCP_TSQ_DEFERRED
))
783 /* Here begins the tricky part :
784 * We are called from release_sock() with :
786 * 2) sk_lock.slock spinlock held
787 * 3) socket owned by us (sk->sk_lock.owned == 1)
789 * But following code is meant to be called from BH handlers,
790 * so we should keep BH disabled, but early release socket ownership
792 sock_release_ownership(sk
);
794 if (flags
& (1UL << TCP_WRITE_TIMER_DEFERRED
)) {
795 tcp_write_timer_handler(sk
);
798 if (flags
& (1UL << TCP_DELACK_TIMER_DEFERRED
)) {
799 tcp_delack_timer_handler(sk
);
802 if (flags
& (1UL << TCP_MTU_REDUCED_DEFERRED
)) {
803 sk
->sk_prot
->mtu_reduced(sk
);
807 EXPORT_SYMBOL(tcp_release_cb
);
809 void __init
tcp_tasklet_init(void)
813 for_each_possible_cpu(i
) {
814 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
816 INIT_LIST_HEAD(&tsq
->head
);
817 tasklet_init(&tsq
->tasklet
,
824 * Write buffer destructor automatically called from kfree_skb.
825 * We can't xmit new skbs from this context, as we might already
828 void tcp_wfree(struct sk_buff
*skb
)
830 struct sock
*sk
= skb
->sk
;
831 struct tcp_sock
*tp
= tcp_sk(sk
);
833 if (test_and_clear_bit(TSQ_THROTTLED
, &tp
->tsq_flags
) &&
834 !test_and_set_bit(TSQ_QUEUED
, &tp
->tsq_flags
)) {
836 struct tsq_tasklet
*tsq
;
838 /* Keep a ref on socket.
839 * This last ref will be released in tcp_tasklet_func()
841 atomic_sub(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
843 /* queue this socket to tasklet queue */
844 local_irq_save(flags
);
845 tsq
= &__get_cpu_var(tsq_tasklet
);
846 list_add(&tp
->tsq_node
, &tsq
->head
);
847 tasklet_schedule(&tsq
->tasklet
);
848 local_irq_restore(flags
);
854 /* This routine actually transmits TCP packets queued in by
855 * tcp_do_sendmsg(). This is used by both the initial
856 * transmission and possible later retransmissions.
857 * All SKB's seen here are completely headerless. It is our
858 * job to build the TCP header, and pass the packet down to
859 * IP so it can do the same plus pass the packet off to the
862 * We are working here with either a clone of the original
863 * SKB, or a fresh unique copy made by the retransmit engine.
865 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
868 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
869 struct inet_sock
*inet
;
871 struct tcp_skb_cb
*tcb
;
872 struct tcp_out_options opts
;
873 unsigned int tcp_options_size
, tcp_header_size
;
874 struct tcp_md5sig_key
*md5
;
878 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
881 const struct sk_buff
*fclone
= skb
+ 1;
883 skb_mstamp_get(&skb
->skb_mstamp
);
885 if (unlikely(skb
->fclone
== SKB_FCLONE_ORIG
&&
886 fclone
->fclone
== SKB_FCLONE_CLONE
))
887 NET_INC_STATS(sock_net(sk
),
888 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
890 if (unlikely(skb_cloned(skb
)))
891 skb
= pskb_copy(skb
, gfp_mask
);
893 skb
= skb_clone(skb
, gfp_mask
);
896 /* Our usage of tstamp should remain private */
897 skb
->tstamp
.tv64
= 0;
902 tcb
= TCP_SKB_CB(skb
);
903 memset(&opts
, 0, sizeof(opts
));
905 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
906 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
908 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
910 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
912 if (tcp_packets_in_flight(tp
) == 0)
913 tcp_ca_event(sk
, CA_EVENT_TX_START
);
915 /* if no packet is in qdisc/device queue, then allow XPS to select
918 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) == 0;
920 skb_push(skb
, tcp_header_size
);
921 skb_reset_transport_header(skb
);
925 skb
->destructor
= tcp_wfree
;
926 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
928 /* Build TCP header and checksum it. */
930 th
->source
= inet
->inet_sport
;
931 th
->dest
= inet
->inet_dport
;
932 th
->seq
= htonl(tcb
->seq
);
933 th
->ack_seq
= htonl(tp
->rcv_nxt
);
934 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
937 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
938 /* RFC1323: The window in SYN & SYN/ACK segments
941 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
943 th
->window
= htons(tcp_select_window(sk
));
948 /* The urg_mode check is necessary during a below snd_una win probe */
949 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
950 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
951 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
953 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
954 th
->urg_ptr
= htons(0xFFFF);
959 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
960 if (likely((tcb
->tcp_flags
& TCPHDR_SYN
) == 0))
961 TCP_ECN_send(sk
, skb
, tcp_header_size
);
963 #ifdef CONFIG_TCP_MD5SIG
964 /* Calculate the MD5 hash, as we have all we need now */
966 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
967 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
972 icsk
->icsk_af_ops
->send_check(sk
, skb
);
974 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
975 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
977 if (skb
->len
!= tcp_header_size
)
978 tcp_event_data_sent(tp
, sk
);
980 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
981 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
982 tcp_skb_pcount(skb
));
984 err
= icsk
->icsk_af_ops
->queue_xmit(sk
, skb
, &inet
->cork
.fl
);
985 if (likely(err
<= 0))
988 tcp_enter_cwr(sk
, 1);
990 return net_xmit_eval(err
);
993 /* This routine just queues the buffer for sending.
995 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
996 * otherwise socket can stall.
998 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1000 struct tcp_sock
*tp
= tcp_sk(sk
);
1002 /* Advance write_seq and place onto the write_queue. */
1003 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
1004 skb_header_release(skb
);
1005 tcp_add_write_queue_tail(sk
, skb
);
1006 sk
->sk_wmem_queued
+= skb
->truesize
;
1007 sk_mem_charge(sk
, skb
->truesize
);
1010 /* Initialize TSO segments for a packet. */
1011 static void tcp_set_skb_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
1012 unsigned int mss_now
)
1014 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1016 /* Make sure we own this skb before messing gso_size/gso_segs */
1017 WARN_ON_ONCE(skb_cloned(skb
));
1019 if (skb
->len
<= mss_now
|| skb
->ip_summed
== CHECKSUM_NONE
) {
1020 /* Avoid the costly divide in the normal
1023 shinfo
->gso_segs
= 1;
1024 shinfo
->gso_size
= 0;
1025 shinfo
->gso_type
= 0;
1027 shinfo
->gso_segs
= DIV_ROUND_UP(skb
->len
, mss_now
);
1028 shinfo
->gso_size
= mss_now
;
1029 shinfo
->gso_type
= sk
->sk_gso_type
;
1033 /* When a modification to fackets out becomes necessary, we need to check
1034 * skb is counted to fackets_out or not.
1036 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1039 struct tcp_sock
*tp
= tcp_sk(sk
);
1041 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1044 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1045 tp
->fackets_out
-= decr
;
1048 /* Pcount in the middle of the write queue got changed, we need to do various
1049 * tweaks to fix counters
1051 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1053 struct tcp_sock
*tp
= tcp_sk(sk
);
1055 tp
->packets_out
-= decr
;
1057 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1058 tp
->sacked_out
-= decr
;
1059 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1060 tp
->retrans_out
-= decr
;
1061 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1062 tp
->lost_out
-= decr
;
1064 /* Reno case is special. Sigh... */
1065 if (tcp_is_reno(tp
) && decr
> 0)
1066 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1068 tcp_adjust_fackets_out(sk
, skb
, decr
);
1070 if (tp
->lost_skb_hint
&&
1071 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1072 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1073 tp
->lost_cnt_hint
-= decr
;
1075 tcp_verify_left_out(tp
);
1078 /* Function to create two new TCP segments. Shrinks the given segment
1079 * to the specified size and appends a new segment with the rest of the
1080 * packet to the list. This won't be called frequently, I hope.
1081 * Remember, these are still headerless SKBs at this point.
1083 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1084 unsigned int mss_now
)
1086 struct tcp_sock
*tp
= tcp_sk(sk
);
1087 struct sk_buff
*buff
;
1088 int nsize
, old_factor
;
1092 if (WARN_ON(len
> skb
->len
))
1095 nsize
= skb_headlen(skb
) - len
;
1099 if (skb_unclone(skb
, GFP_ATOMIC
))
1102 /* Get a new skb... force flag on. */
1103 buff
= sk_stream_alloc_skb(sk
, nsize
, GFP_ATOMIC
);
1105 return -ENOMEM
; /* We'll just try again later. */
1107 sk
->sk_wmem_queued
+= buff
->truesize
;
1108 sk_mem_charge(sk
, buff
->truesize
);
1109 nlen
= skb
->len
- len
- nsize
;
1110 buff
->truesize
+= nlen
;
1111 skb
->truesize
-= nlen
;
1113 /* Correct the sequence numbers. */
1114 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1115 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1116 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1118 /* PSH and FIN should only be set in the second packet. */
1119 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1120 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1121 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1122 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1124 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1125 /* Copy and checksum data tail into the new buffer. */
1126 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1127 skb_put(buff
, nsize
),
1132 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1134 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1135 skb_split(skb
, buff
, len
);
1138 buff
->ip_summed
= skb
->ip_summed
;
1140 /* Looks stupid, but our code really uses when of
1141 * skbs, which it never sent before. --ANK
1143 TCP_SKB_CB(buff
)->when
= TCP_SKB_CB(skb
)->when
;
1144 buff
->tstamp
= skb
->tstamp
;
1146 old_factor
= tcp_skb_pcount(skb
);
1148 /* Fix up tso_factor for both original and new SKB. */
1149 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1150 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1152 /* If this packet has been sent out already, we must
1153 * adjust the various packet counters.
1155 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1156 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1157 tcp_skb_pcount(buff
);
1160 tcp_adjust_pcount(sk
, skb
, diff
);
1163 /* Link BUFF into the send queue. */
1164 skb_header_release(buff
);
1165 tcp_insert_write_queue_after(skb
, buff
, sk
);
1170 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1171 * eventually). The difference is that pulled data not copied, but
1172 * immediately discarded.
1174 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1176 struct skb_shared_info
*shinfo
;
1179 eat
= min_t(int, len
, skb_headlen(skb
));
1181 __skb_pull(skb
, eat
);
1188 shinfo
= skb_shinfo(skb
);
1189 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1190 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1193 skb_frag_unref(skb
, i
);
1196 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1198 shinfo
->frags
[k
].page_offset
+= eat
;
1199 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1205 shinfo
->nr_frags
= k
;
1207 skb_reset_tail_pointer(skb
);
1208 skb
->data_len
-= len
;
1209 skb
->len
= skb
->data_len
;
1212 /* Remove acked data from a packet in the transmit queue. */
1213 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1215 if (skb_unclone(skb
, GFP_ATOMIC
))
1218 __pskb_trim_head(skb
, len
);
1220 TCP_SKB_CB(skb
)->seq
+= len
;
1221 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1223 skb
->truesize
-= len
;
1224 sk
->sk_wmem_queued
-= len
;
1225 sk_mem_uncharge(sk
, len
);
1226 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1228 /* Any change of skb->len requires recalculation of tso factor. */
1229 if (tcp_skb_pcount(skb
) > 1)
1230 tcp_set_skb_tso_segs(sk
, skb
, tcp_skb_mss(skb
));
1235 /* Calculate MSS not accounting any TCP options. */
1236 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1238 const struct tcp_sock
*tp
= tcp_sk(sk
);
1239 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1242 /* Calculate base mss without TCP options:
1243 It is MMS_S - sizeof(tcphdr) of rfc1122
1245 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1247 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1248 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1249 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1251 if (dst
&& dst_allfrag(dst
))
1252 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1255 /* Clamp it (mss_clamp does not include tcp options) */
1256 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1257 mss_now
= tp
->rx_opt
.mss_clamp
;
1259 /* Now subtract optional transport overhead */
1260 mss_now
-= icsk
->icsk_ext_hdr_len
;
1262 /* Then reserve room for full set of TCP options and 8 bytes of data */
1268 /* Calculate MSS. Not accounting for SACKs here. */
1269 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1271 /* Subtract TCP options size, not including SACKs */
1272 return __tcp_mtu_to_mss(sk
, pmtu
) -
1273 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1276 /* Inverse of above */
1277 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1279 const struct tcp_sock
*tp
= tcp_sk(sk
);
1280 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1284 tp
->tcp_header_len
+
1285 icsk
->icsk_ext_hdr_len
+
1286 icsk
->icsk_af_ops
->net_header_len
;
1288 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1289 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1290 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1292 if (dst
&& dst_allfrag(dst
))
1293 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1298 /* MTU probing init per socket */
1299 void tcp_mtup_init(struct sock
*sk
)
1301 struct tcp_sock
*tp
= tcp_sk(sk
);
1302 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1304 icsk
->icsk_mtup
.enabled
= sysctl_tcp_mtu_probing
> 1;
1305 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1306 icsk
->icsk_af_ops
->net_header_len
;
1307 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, sysctl_tcp_base_mss
);
1308 icsk
->icsk_mtup
.probe_size
= 0;
1310 EXPORT_SYMBOL(tcp_mtup_init
);
1312 /* This function synchronize snd mss to current pmtu/exthdr set.
1314 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1315 for TCP options, but includes only bare TCP header.
1317 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1318 It is minimum of user_mss and mss received with SYN.
1319 It also does not include TCP options.
1321 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1323 tp->mss_cache is current effective sending mss, including
1324 all tcp options except for SACKs. It is evaluated,
1325 taking into account current pmtu, but never exceeds
1326 tp->rx_opt.mss_clamp.
1328 NOTE1. rfc1122 clearly states that advertised MSS
1329 DOES NOT include either tcp or ip options.
1331 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1332 are READ ONLY outside this function. --ANK (980731)
1334 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1336 struct tcp_sock
*tp
= tcp_sk(sk
);
1337 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1340 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1341 icsk
->icsk_mtup
.search_high
= pmtu
;
1343 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1344 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1346 /* And store cached results */
1347 icsk
->icsk_pmtu_cookie
= pmtu
;
1348 if (icsk
->icsk_mtup
.enabled
)
1349 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1350 tp
->mss_cache
= mss_now
;
1354 EXPORT_SYMBOL(tcp_sync_mss
);
1356 /* Compute the current effective MSS, taking SACKs and IP options,
1357 * and even PMTU discovery events into account.
1359 unsigned int tcp_current_mss(struct sock
*sk
)
1361 const struct tcp_sock
*tp
= tcp_sk(sk
);
1362 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1364 unsigned int header_len
;
1365 struct tcp_out_options opts
;
1366 struct tcp_md5sig_key
*md5
;
1368 mss_now
= tp
->mss_cache
;
1371 u32 mtu
= dst_mtu(dst
);
1372 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1373 mss_now
= tcp_sync_mss(sk
, mtu
);
1376 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1377 sizeof(struct tcphdr
);
1378 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1379 * some common options. If this is an odd packet (because we have SACK
1380 * blocks etc) then our calculated header_len will be different, and
1381 * we have to adjust mss_now correspondingly */
1382 if (header_len
!= tp
->tcp_header_len
) {
1383 int delta
= (int) header_len
- tp
->tcp_header_len
;
1390 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1391 * As additional protections, we do not touch cwnd in retransmission phases,
1392 * and if application hit its sndbuf limit recently.
1394 static void tcp_cwnd_application_limited(struct sock
*sk
)
1396 struct tcp_sock
*tp
= tcp_sk(sk
);
1398 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1399 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1400 /* Limited by application or receiver window. */
1401 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1402 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1403 if (win_used
< tp
->snd_cwnd
) {
1404 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1405 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
1407 tp
->snd_cwnd_used
= 0;
1409 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1412 /* Congestion window validation. (RFC2861) */
1413 static void tcp_cwnd_validate(struct sock
*sk
)
1415 struct tcp_sock
*tp
= tcp_sk(sk
);
1417 if (tp
->packets_out
>= tp
->snd_cwnd
) {
1418 /* Network is feed fully. */
1419 tp
->snd_cwnd_used
= 0;
1420 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1422 /* Network starves. */
1423 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1424 tp
->snd_cwnd_used
= tp
->packets_out
;
1426 if (sysctl_tcp_slow_start_after_idle
&&
1427 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1428 tcp_cwnd_application_limited(sk
);
1432 /* Minshall's variant of the Nagle send check. */
1433 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1435 return after(tp
->snd_sml
, tp
->snd_una
) &&
1436 !after(tp
->snd_sml
, tp
->snd_nxt
);
1439 /* Update snd_sml if this skb is under mss
1440 * Note that a TSO packet might end with a sub-mss segment
1441 * The test is really :
1442 * if ((skb->len % mss) != 0)
1443 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1444 * But we can avoid doing the divide again given we already have
1445 * skb_pcount = skb->len / mss_now
1447 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1448 const struct sk_buff
*skb
)
1450 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1451 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1454 /* Return false, if packet can be sent now without violation Nagle's rules:
1455 * 1. It is full sized. (provided by caller in %partial bool)
1456 * 2. Or it contains FIN. (already checked by caller)
1457 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1458 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1459 * With Minshall's modification: all sent small packets are ACKed.
1461 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1465 ((nonagle
& TCP_NAGLE_CORK
) ||
1466 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1468 /* Returns the portion of skb which can be sent right away */
1469 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1470 const struct sk_buff
*skb
,
1471 unsigned int mss_now
,
1472 unsigned int max_segs
,
1475 const struct tcp_sock
*tp
= tcp_sk(sk
);
1476 u32 partial
, needed
, window
, max_len
;
1478 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1479 max_len
= mss_now
* max_segs
;
1481 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1484 needed
= min(skb
->len
, window
);
1486 if (max_len
<= needed
)
1489 partial
= needed
% mss_now
;
1490 /* If last segment is not a full MSS, check if Nagle rules allow us
1491 * to include this last segment in this skb.
1492 * Otherwise, we'll split the skb at last MSS boundary
1494 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
1495 return needed
- partial
;
1500 /* Can at least one segment of SKB be sent right now, according to the
1501 * congestion window rules? If so, return how many segments are allowed.
1503 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1504 const struct sk_buff
*skb
)
1506 u32 in_flight
, cwnd
;
1508 /* Don't be strict about the congestion window for the final FIN. */
1509 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1510 tcp_skb_pcount(skb
) == 1)
1513 in_flight
= tcp_packets_in_flight(tp
);
1514 cwnd
= tp
->snd_cwnd
;
1515 if (in_flight
< cwnd
)
1516 return (cwnd
- in_flight
);
1521 /* Initialize TSO state of a skb.
1522 * This must be invoked the first time we consider transmitting
1523 * SKB onto the wire.
1525 static int tcp_init_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
1526 unsigned int mss_now
)
1528 int tso_segs
= tcp_skb_pcount(skb
);
1530 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1531 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1532 tso_segs
= tcp_skb_pcount(skb
);
1538 /* Return true if the Nagle test allows this packet to be
1541 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1542 unsigned int cur_mss
, int nonagle
)
1544 /* Nagle rule does not apply to frames, which sit in the middle of the
1545 * write_queue (they have no chances to get new data).
1547 * This is implemented in the callers, where they modify the 'nonagle'
1548 * argument based upon the location of SKB in the send queue.
1550 if (nonagle
& TCP_NAGLE_PUSH
)
1553 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1554 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1557 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
1563 /* Does at least the first segment of SKB fit into the send window? */
1564 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1565 const struct sk_buff
*skb
,
1566 unsigned int cur_mss
)
1568 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1570 if (skb
->len
> cur_mss
)
1571 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1573 return !after(end_seq
, tcp_wnd_end(tp
));
1576 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1577 * should be put on the wire right now. If so, it returns the number of
1578 * packets allowed by the congestion window.
1580 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1581 unsigned int cur_mss
, int nonagle
)
1583 const struct tcp_sock
*tp
= tcp_sk(sk
);
1584 unsigned int cwnd_quota
;
1586 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1588 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1591 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1592 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1598 /* Test if sending is allowed right now. */
1599 bool tcp_may_send_now(struct sock
*sk
)
1601 const struct tcp_sock
*tp
= tcp_sk(sk
);
1602 struct sk_buff
*skb
= tcp_send_head(sk
);
1605 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1606 (tcp_skb_is_last(sk
, skb
) ?
1607 tp
->nonagle
: TCP_NAGLE_PUSH
));
1610 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1611 * which is put after SKB on the list. It is very much like
1612 * tcp_fragment() except that it may make several kinds of assumptions
1613 * in order to speed up the splitting operation. In particular, we
1614 * know that all the data is in scatter-gather pages, and that the
1615 * packet has never been sent out before (and thus is not cloned).
1617 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1618 unsigned int mss_now
, gfp_t gfp
)
1620 struct sk_buff
*buff
;
1621 int nlen
= skb
->len
- len
;
1624 /* All of a TSO frame must be composed of paged data. */
1625 if (skb
->len
!= skb
->data_len
)
1626 return tcp_fragment(sk
, skb
, len
, mss_now
);
1628 buff
= sk_stream_alloc_skb(sk
, 0, gfp
);
1629 if (unlikely(buff
== NULL
))
1632 sk
->sk_wmem_queued
+= buff
->truesize
;
1633 sk_mem_charge(sk
, buff
->truesize
);
1634 buff
->truesize
+= nlen
;
1635 skb
->truesize
-= nlen
;
1637 /* Correct the sequence numbers. */
1638 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1639 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1640 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1642 /* PSH and FIN should only be set in the second packet. */
1643 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1644 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1645 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1647 /* This packet was never sent out yet, so no SACK bits. */
1648 TCP_SKB_CB(buff
)->sacked
= 0;
1650 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1651 skb_split(skb
, buff
, len
);
1653 /* Fix up tso_factor for both original and new SKB. */
1654 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1655 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1657 /* Link BUFF into the send queue. */
1658 skb_header_release(buff
);
1659 tcp_insert_write_queue_after(skb
, buff
, sk
);
1664 /* Try to defer sending, if possible, in order to minimize the amount
1665 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1667 * This algorithm is from John Heffner.
1669 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
)
1671 struct tcp_sock
*tp
= tcp_sk(sk
);
1672 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1673 u32 send_win
, cong_win
, limit
, in_flight
;
1676 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1679 if (icsk
->icsk_ca_state
!= TCP_CA_Open
)
1682 /* Defer for less than two clock ticks. */
1683 if (tp
->tso_deferred
&&
1684 (((u32
)jiffies
<< 1) >> 1) - (tp
->tso_deferred
>> 1) > 1)
1687 in_flight
= tcp_packets_in_flight(tp
);
1689 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1691 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1693 /* From in_flight test above, we know that cwnd > in_flight. */
1694 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1696 limit
= min(send_win
, cong_win
);
1698 /* If a full-sized TSO skb can be sent, do it. */
1699 if (limit
>= min_t(unsigned int, sk
->sk_gso_max_size
,
1700 tp
->xmit_size_goal_segs
* tp
->mss_cache
))
1703 /* Middle in queue won't get any more data, full sendable already? */
1704 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1707 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1709 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1711 /* If at least some fraction of a window is available,
1714 chunk
/= win_divisor
;
1718 /* Different approach, try not to defer past a single
1719 * ACK. Receiver should ACK every other full sized
1720 * frame, so if we have space for more than 3 frames
1723 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1727 /* Ok, it looks like it is advisable to defer.
1728 * Do not rearm the timer if already set to not break TCP ACK clocking.
1730 if (!tp
->tso_deferred
)
1731 tp
->tso_deferred
= 1 | (jiffies
<< 1);
1736 tp
->tso_deferred
= 0;
1740 /* Create a new MTU probe if we are ready.
1741 * MTU probe is regularly attempting to increase the path MTU by
1742 * deliberately sending larger packets. This discovers routing
1743 * changes resulting in larger path MTUs.
1745 * Returns 0 if we should wait to probe (no cwnd available),
1746 * 1 if a probe was sent,
1749 static int tcp_mtu_probe(struct sock
*sk
)
1751 struct tcp_sock
*tp
= tcp_sk(sk
);
1752 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1753 struct sk_buff
*skb
, *nskb
, *next
;
1760 /* Not currently probing/verifying,
1762 * have enough cwnd, and
1763 * not SACKing (the variable headers throw things off) */
1764 if (!icsk
->icsk_mtup
.enabled
||
1765 icsk
->icsk_mtup
.probe_size
||
1766 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1767 tp
->snd_cwnd
< 11 ||
1768 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1771 /* Very simple search strategy: just double the MSS. */
1772 mss_now
= tcp_current_mss(sk
);
1773 probe_size
= 2 * tp
->mss_cache
;
1774 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1775 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
)) {
1776 /* TODO: set timer for probe_converge_event */
1780 /* Have enough data in the send queue to probe? */
1781 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1784 if (tp
->snd_wnd
< size_needed
)
1786 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1789 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1790 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1791 if (!tcp_packets_in_flight(tp
))
1797 /* We're allowed to probe. Build it now. */
1798 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1800 sk
->sk_wmem_queued
+= nskb
->truesize
;
1801 sk_mem_charge(sk
, nskb
->truesize
);
1803 skb
= tcp_send_head(sk
);
1805 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1806 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1807 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
1808 TCP_SKB_CB(nskb
)->sacked
= 0;
1810 nskb
->ip_summed
= skb
->ip_summed
;
1812 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1815 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1816 copy
= min_t(int, skb
->len
, probe_size
- len
);
1817 if (nskb
->ip_summed
)
1818 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1820 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1821 skb_put(nskb
, copy
),
1824 if (skb
->len
<= copy
) {
1825 /* We've eaten all the data from this skb.
1827 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1828 tcp_unlink_write_queue(skb
, sk
);
1829 sk_wmem_free_skb(sk
, skb
);
1831 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
1832 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1833 if (!skb_shinfo(skb
)->nr_frags
) {
1834 skb_pull(skb
, copy
);
1835 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1836 skb
->csum
= csum_partial(skb
->data
,
1839 __pskb_trim_head(skb
, copy
);
1840 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1842 TCP_SKB_CB(skb
)->seq
+= copy
;
1847 if (len
>= probe_size
)
1850 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1852 /* We're ready to send. If this fails, the probe will
1853 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1854 TCP_SKB_CB(nskb
)->when
= tcp_time_stamp
;
1855 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1856 /* Decrement cwnd here because we are sending
1857 * effectively two packets. */
1859 tcp_event_new_data_sent(sk
, nskb
);
1861 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1862 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1863 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
1871 /* This routine writes packets to the network. It advances the
1872 * send_head. This happens as incoming acks open up the remote
1875 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1876 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1877 * account rare use of URG, this is not a big flaw.
1879 * Send at most one packet when push_one > 0. Temporarily ignore
1880 * cwnd limit to force at most one packet out when push_one == 2.
1882 * Returns true, if no segments are in flight and we have queued segments,
1883 * but cannot send anything now because of SWS or another problem.
1885 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
1886 int push_one
, gfp_t gfp
)
1888 struct tcp_sock
*tp
= tcp_sk(sk
);
1889 struct sk_buff
*skb
;
1890 unsigned int tso_segs
, sent_pkts
;
1897 /* Do MTU probing. */
1898 result
= tcp_mtu_probe(sk
);
1901 } else if (result
> 0) {
1906 while ((skb
= tcp_send_head(sk
))) {
1909 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
1912 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
)
1913 goto repair
; /* Skip network transmission */
1915 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1918 /* Force out a loss probe pkt. */
1924 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
1927 if (tso_segs
== 1) {
1928 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
1929 (tcp_skb_is_last(sk
, skb
) ?
1930 nonagle
: TCP_NAGLE_PUSH
))))
1933 if (!push_one
&& tcp_tso_should_defer(sk
, skb
))
1937 /* TCP Small Queues :
1938 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
1940 * - better RTT estimation and ACK scheduling
1943 * Alas, some drivers / subsystems require a fair amount
1944 * of queued bytes to ensure line rate.
1945 * One example is wifi aggregation (802.11 AMPDU)
1947 limit
= max_t(unsigned int, sysctl_tcp_limit_output_bytes
,
1948 sk
->sk_pacing_rate
>> 10);
1950 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
) {
1951 set_bit(TSQ_THROTTLED
, &tp
->tsq_flags
);
1952 /* It is possible TX completion already happened
1953 * before we set TSQ_THROTTLED, so we must
1954 * test again the condition.
1955 * We abuse smp_mb__after_clear_bit() because
1956 * there is no smp_mb__after_set_bit() yet
1958 smp_mb__after_clear_bit();
1959 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
)
1964 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
1965 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
1968 sk
->sk_gso_max_segs
),
1971 if (skb
->len
> limit
&&
1972 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
1975 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
1977 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
1981 /* Advance the send_head. This one is sent out.
1982 * This call will increment packets_out.
1984 tcp_event_new_data_sent(sk
, skb
);
1986 tcp_minshall_update(tp
, mss_now
, skb
);
1987 sent_pkts
+= tcp_skb_pcount(skb
);
1993 if (likely(sent_pkts
)) {
1994 if (tcp_in_cwnd_reduction(sk
))
1995 tp
->prr_out
+= sent_pkts
;
1997 /* Send one loss probe per tail loss episode. */
1999 tcp_schedule_loss_probe(sk
);
2000 tcp_cwnd_validate(sk
);
2003 return (push_one
== 2) || (!tp
->packets_out
&& tcp_send_head(sk
));
2006 bool tcp_schedule_loss_probe(struct sock
*sk
)
2008 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2009 struct tcp_sock
*tp
= tcp_sk(sk
);
2010 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
2011 u32 rtt
= usecs_to_jiffies(tp
->srtt_us
>> 3);
2013 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
))
2015 /* No consecutive loss probes. */
2016 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
2020 /* Don't do any loss probe on a Fast Open connection before 3WHS
2023 if (sk
->sk_state
== TCP_SYN_RECV
)
2026 /* TLP is only scheduled when next timer event is RTO. */
2027 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
2030 /* Schedule a loss probe in 2*RTT for SACK capable connections
2031 * in Open state, that are either limited by cwnd or application.
2033 if (sysctl_tcp_early_retrans
< 3 || !tp
->srtt_us
|| !tp
->packets_out
||
2034 !tcp_is_sack(tp
) || inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
2037 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
2041 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2042 * for delayed ack when there's one outstanding packet.
2045 if (tp
->packets_out
== 1)
2046 timeout
= max_t(u32
, timeout
,
2047 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
2048 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
2050 /* If RTO is shorter, just schedule TLP in its place. */
2051 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
2052 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
2053 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
2054 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
2059 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2064 /* When probe timeout (PTO) fires, send a new segment if one exists, else
2065 * retransmit the last segment.
2067 void tcp_send_loss_probe(struct sock
*sk
)
2069 struct tcp_sock
*tp
= tcp_sk(sk
);
2070 struct sk_buff
*skb
;
2072 int mss
= tcp_current_mss(sk
);
2075 if (tcp_send_head(sk
) != NULL
) {
2076 err
= tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2080 /* At most one outstanding TLP retransmission. */
2081 if (tp
->tlp_high_seq
)
2084 /* Retransmit last segment. */
2085 skb
= tcp_write_queue_tail(sk
);
2089 pcount
= tcp_skb_pcount(skb
);
2090 if (WARN_ON(!pcount
))
2093 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2094 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
)))
2096 skb
= tcp_write_queue_tail(sk
);
2099 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2102 /* Probe with zero data doesn't trigger fast recovery. */
2104 err
= __tcp_retransmit_skb(sk
, skb
);
2106 /* Record snd_nxt for loss detection. */
2108 tp
->tlp_high_seq
= tp
->snd_nxt
;
2111 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2112 inet_csk(sk
)->icsk_rto
,
2116 NET_INC_STATS_BH(sock_net(sk
),
2117 LINUX_MIB_TCPLOSSPROBES
);
2120 /* Push out any pending frames which were held back due to
2121 * TCP_CORK or attempt at coalescing tiny packets.
2122 * The socket must be locked by the caller.
2124 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2127 /* If we are closed, the bytes will have to remain here.
2128 * In time closedown will finish, we empty the write queue and
2129 * all will be happy.
2131 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2134 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2135 sk_gfp_atomic(sk
, GFP_ATOMIC
)))
2136 tcp_check_probe_timer(sk
);
2139 /* Send _single_ skb sitting at the send head. This function requires
2140 * true push pending frames to setup probe timer etc.
2142 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2144 struct sk_buff
*skb
= tcp_send_head(sk
);
2146 BUG_ON(!skb
|| skb
->len
< mss_now
);
2148 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2151 /* This function returns the amount that we can raise the
2152 * usable window based on the following constraints
2154 * 1. The window can never be shrunk once it is offered (RFC 793)
2155 * 2. We limit memory per socket
2158 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2159 * RECV.NEXT + RCV.WIN fixed until:
2160 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2162 * i.e. don't raise the right edge of the window until you can raise
2163 * it at least MSS bytes.
2165 * Unfortunately, the recommended algorithm breaks header prediction,
2166 * since header prediction assumes th->window stays fixed.
2168 * Strictly speaking, keeping th->window fixed violates the receiver
2169 * side SWS prevention criteria. The problem is that under this rule
2170 * a stream of single byte packets will cause the right side of the
2171 * window to always advance by a single byte.
2173 * Of course, if the sender implements sender side SWS prevention
2174 * then this will not be a problem.
2176 * BSD seems to make the following compromise:
2178 * If the free space is less than the 1/4 of the maximum
2179 * space available and the free space is less than 1/2 mss,
2180 * then set the window to 0.
2181 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2182 * Otherwise, just prevent the window from shrinking
2183 * and from being larger than the largest representable value.
2185 * This prevents incremental opening of the window in the regime
2186 * where TCP is limited by the speed of the reader side taking
2187 * data out of the TCP receive queue. It does nothing about
2188 * those cases where the window is constrained on the sender side
2189 * because the pipeline is full.
2191 * BSD also seems to "accidentally" limit itself to windows that are a
2192 * multiple of MSS, at least until the free space gets quite small.
2193 * This would appear to be a side effect of the mbuf implementation.
2194 * Combining these two algorithms results in the observed behavior
2195 * of having a fixed window size at almost all times.
2197 * Below we obtain similar behavior by forcing the offered window to
2198 * a multiple of the mss when it is feasible to do so.
2200 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2201 * Regular options like TIMESTAMP are taken into account.
2203 u32
__tcp_select_window(struct sock
*sk
)
2205 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2206 struct tcp_sock
*tp
= tcp_sk(sk
);
2207 /* MSS for the peer's data. Previous versions used mss_clamp
2208 * here. I don't know if the value based on our guesses
2209 * of peer's MSS is better for the performance. It's more correct
2210 * but may be worse for the performance because of rcv_mss
2211 * fluctuations. --SAW 1998/11/1
2213 int mss
= icsk
->icsk_ack
.rcv_mss
;
2214 int free_space
= tcp_space(sk
);
2215 int allowed_space
= tcp_full_space(sk
);
2216 int full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2219 if (mss
> full_space
)
2222 if (free_space
< (full_space
>> 1)) {
2223 icsk
->icsk_ack
.quick
= 0;
2225 if (sk_under_memory_pressure(sk
))
2226 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2229 /* free_space might become our new window, make sure we don't
2230 * increase it due to wscale.
2232 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2234 /* if free space is less than mss estimate, or is below 1/16th
2235 * of the maximum allowed, try to move to zero-window, else
2236 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2237 * new incoming data is dropped due to memory limits.
2238 * With large window, mss test triggers way too late in order
2239 * to announce zero window in time before rmem limit kicks in.
2241 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2245 if (free_space
> tp
->rcv_ssthresh
)
2246 free_space
= tp
->rcv_ssthresh
;
2248 /* Don't do rounding if we are using window scaling, since the
2249 * scaled window will not line up with the MSS boundary anyway.
2251 window
= tp
->rcv_wnd
;
2252 if (tp
->rx_opt
.rcv_wscale
) {
2253 window
= free_space
;
2255 /* Advertise enough space so that it won't get scaled away.
2256 * Import case: prevent zero window announcement if
2257 * 1<<rcv_wscale > mss.
2259 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2260 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2261 << tp
->rx_opt
.rcv_wscale
);
2263 /* Get the largest window that is a nice multiple of mss.
2264 * Window clamp already applied above.
2265 * If our current window offering is within 1 mss of the
2266 * free space we just keep it. This prevents the divide
2267 * and multiply from happening most of the time.
2268 * We also don't do any window rounding when the free space
2271 if (window
<= free_space
- mss
|| window
> free_space
)
2272 window
= (free_space
/ mss
) * mss
;
2273 else if (mss
== full_space
&&
2274 free_space
> window
+ (full_space
>> 1))
2275 window
= free_space
;
2281 /* Collapses two adjacent SKB's during retransmission. */
2282 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2284 struct tcp_sock
*tp
= tcp_sk(sk
);
2285 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2286 int skb_size
, next_skb_size
;
2288 skb_size
= skb
->len
;
2289 next_skb_size
= next_skb
->len
;
2291 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2293 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2295 tcp_unlink_write_queue(next_skb
, sk
);
2297 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
2300 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2301 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2303 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2304 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2306 /* Update sequence range on original skb. */
2307 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2309 /* Merge over control information. This moves PSH/FIN etc. over */
2310 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2312 /* All done, get rid of second SKB and account for it so
2313 * packet counting does not break.
2315 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2317 /* changed transmit queue under us so clear hints */
2318 tcp_clear_retrans_hints_partial(tp
);
2319 if (next_skb
== tp
->retransmit_skb_hint
)
2320 tp
->retransmit_skb_hint
= skb
;
2322 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2324 sk_wmem_free_skb(sk
, next_skb
);
2327 /* Check if coalescing SKBs is legal. */
2328 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2330 if (tcp_skb_pcount(skb
) > 1)
2332 /* TODO: SACK collapsing could be used to remove this condition */
2333 if (skb_shinfo(skb
)->nr_frags
!= 0)
2335 if (skb_cloned(skb
))
2337 if (skb
== tcp_send_head(sk
))
2339 /* Some heurestics for collapsing over SACK'd could be invented */
2340 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2346 /* Collapse packets in the retransmit queue to make to create
2347 * less packets on the wire. This is only done on retransmission.
2349 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2352 struct tcp_sock
*tp
= tcp_sk(sk
);
2353 struct sk_buff
*skb
= to
, *tmp
;
2356 if (!sysctl_tcp_retrans_collapse
)
2358 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2361 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2362 if (!tcp_can_collapse(sk
, skb
))
2374 /* Punt if not enough space exists in the first SKB for
2375 * the data in the second
2377 if (skb
->len
> skb_availroom(to
))
2380 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2383 tcp_collapse_retrans(sk
, to
);
2387 /* This retransmits one SKB. Policy decisions and retransmit queue
2388 * state updates are done by the caller. Returns non-zero if an
2389 * error occurred which prevented the send.
2391 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2393 struct tcp_sock
*tp
= tcp_sk(sk
);
2394 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2395 unsigned int cur_mss
;
2398 /* Inconslusive MTU probe */
2399 if (icsk
->icsk_mtup
.probe_size
) {
2400 icsk
->icsk_mtup
.probe_size
= 0;
2403 /* Do not sent more than we queued. 1/4 is reserved for possible
2404 * copying overhead: fragmentation, tunneling, mangling etc.
2406 if (atomic_read(&sk
->sk_wmem_alloc
) >
2407 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2410 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2411 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2413 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2417 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2418 return -EHOSTUNREACH
; /* Routing failure or similar. */
2420 cur_mss
= tcp_current_mss(sk
);
2422 /* If receiver has shrunk his window, and skb is out of
2423 * new window, do not retransmit it. The exception is the
2424 * case, when window is shrunk to zero. In this case
2425 * our retransmit serves as a zero window probe.
2427 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2428 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2431 if (skb
->len
> cur_mss
) {
2432 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
))
2433 return -ENOMEM
; /* We'll try again later. */
2435 int oldpcount
= tcp_skb_pcount(skb
);
2437 if (unlikely(oldpcount
> 1)) {
2438 if (skb_unclone(skb
, GFP_ATOMIC
))
2440 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2441 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2445 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2447 /* Make a copy, if the first transmission SKB clone we made
2448 * is still in somebody's hands, else make a clone.
2450 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2452 /* make sure skb->data is aligned on arches that require it
2453 * and check if ack-trimming & collapsing extended the headroom
2454 * beyond what csum_start can cover.
2456 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2457 skb_headroom(skb
) >= 0xFFFF)) {
2458 struct sk_buff
*nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
,
2460 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2463 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2467 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
2471 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2473 struct tcp_sock
*tp
= tcp_sk(sk
);
2474 int err
= __tcp_retransmit_skb(sk
, skb
);
2477 /* Update global TCP statistics. */
2478 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2479 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2480 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
2481 tp
->total_retrans
++;
2483 #if FASTRETRANS_DEBUG > 0
2484 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2485 net_dbg_ratelimited("retrans_out leaked\n");
2488 if (!tp
->retrans_out
)
2489 tp
->lost_retrans_low
= tp
->snd_nxt
;
2490 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2491 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2493 /* Save stamp of the first retransmit. */
2494 if (!tp
->retrans_stamp
)
2495 tp
->retrans_stamp
= TCP_SKB_CB(skb
)->when
;
2497 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2499 /* snd_nxt is stored to detect loss of retransmitted segment,
2500 * see tcp_input.c tcp_sacktag_write_queue().
2502 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2504 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2509 /* Check if we forward retransmits are possible in the current
2510 * window/congestion state.
2512 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2514 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2515 const struct tcp_sock
*tp
= tcp_sk(sk
);
2517 /* Forward retransmissions are possible only during Recovery. */
2518 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2521 /* No forward retransmissions in Reno are possible. */
2522 if (tcp_is_reno(tp
))
2525 /* Yeah, we have to make difficult choice between forward transmission
2526 * and retransmission... Both ways have their merits...
2528 * For now we do not retransmit anything, while we have some new
2529 * segments to send. In the other cases, follow rule 3 for
2530 * NextSeg() specified in RFC3517.
2533 if (tcp_may_send_now(sk
))
2539 /* This gets called after a retransmit timeout, and the initially
2540 * retransmitted data is acknowledged. It tries to continue
2541 * resending the rest of the retransmit queue, until either
2542 * we've sent it all or the congestion window limit is reached.
2543 * If doing SACK, the first ACK which comes back for a timeout
2544 * based retransmit packet might feed us FACK information again.
2545 * If so, we use it to avoid unnecessarily retransmissions.
2547 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2549 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2550 struct tcp_sock
*tp
= tcp_sk(sk
);
2551 struct sk_buff
*skb
;
2552 struct sk_buff
*hole
= NULL
;
2555 int fwd_rexmitting
= 0;
2557 if (!tp
->packets_out
)
2561 tp
->retransmit_high
= tp
->snd_una
;
2563 if (tp
->retransmit_skb_hint
) {
2564 skb
= tp
->retransmit_skb_hint
;
2565 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2566 if (after(last_lost
, tp
->retransmit_high
))
2567 last_lost
= tp
->retransmit_high
;
2569 skb
= tcp_write_queue_head(sk
);
2570 last_lost
= tp
->snd_una
;
2573 tcp_for_write_queue_from(skb
, sk
) {
2574 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2576 if (skb
== tcp_send_head(sk
))
2578 /* we could do better than to assign each time */
2580 tp
->retransmit_skb_hint
= skb
;
2582 /* Assume this retransmit will generate
2583 * only one packet for congestion window
2584 * calculation purposes. This works because
2585 * tcp_retransmit_skb() will chop up the
2586 * packet to be MSS sized and all the
2587 * packet counting works out.
2589 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2592 if (fwd_rexmitting
) {
2594 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2596 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2598 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2599 tp
->retransmit_high
= last_lost
;
2600 if (!tcp_can_forward_retransmit(sk
))
2602 /* Backtrack if necessary to non-L'ed skb */
2610 } else if (!(sacked
& TCPCB_LOST
)) {
2611 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2616 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2617 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2618 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2620 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2623 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2626 if (tcp_retransmit_skb(sk
, skb
))
2629 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2631 if (tcp_in_cwnd_reduction(sk
))
2632 tp
->prr_out
+= tcp_skb_pcount(skb
);
2634 if (skb
== tcp_write_queue_head(sk
))
2635 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2636 inet_csk(sk
)->icsk_rto
,
2641 /* Send a fin. The caller locks the socket for us. This cannot be
2642 * allowed to fail queueing a FIN frame under any circumstances.
2644 void tcp_send_fin(struct sock
*sk
)
2646 struct tcp_sock
*tp
= tcp_sk(sk
);
2647 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2650 /* Optimization, tack on the FIN if we have a queue of
2651 * unsent frames. But be careful about outgoing SACKS
2654 mss_now
= tcp_current_mss(sk
);
2656 if (tcp_send_head(sk
) != NULL
) {
2657 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_FIN
;
2658 TCP_SKB_CB(skb
)->end_seq
++;
2661 /* Socket is locked, keep trying until memory is available. */
2663 skb
= alloc_skb_fclone(MAX_TCP_HEADER
,
2670 /* Reserve space for headers and prepare control bits. */
2671 skb_reserve(skb
, MAX_TCP_HEADER
);
2672 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2673 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2674 TCPHDR_ACK
| TCPHDR_FIN
);
2675 tcp_queue_skb(sk
, skb
);
2677 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2680 /* We get here when a process closes a file descriptor (either due to
2681 * an explicit close() or as a byproduct of exit()'ing) and there
2682 * was unread data in the receive queue. This behavior is recommended
2683 * by RFC 2525, section 2.17. -DaveM
2685 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2687 struct sk_buff
*skb
;
2689 /* NOTE: No TCP options attached and we never retransmit this. */
2690 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2692 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2696 /* Reserve space for headers and prepare control bits. */
2697 skb_reserve(skb
, MAX_TCP_HEADER
);
2698 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2699 TCPHDR_ACK
| TCPHDR_RST
);
2701 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2702 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2703 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2705 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2708 /* Send a crossed SYN-ACK during socket establishment.
2709 * WARNING: This routine must only be called when we have already sent
2710 * a SYN packet that crossed the incoming SYN that caused this routine
2711 * to get called. If this assumption fails then the initial rcv_wnd
2712 * and rcv_wscale values will not be correct.
2714 int tcp_send_synack(struct sock
*sk
)
2716 struct sk_buff
*skb
;
2718 skb
= tcp_write_queue_head(sk
);
2719 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2720 pr_debug("%s: wrong queue state\n", __func__
);
2723 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
2724 if (skb_cloned(skb
)) {
2725 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2728 tcp_unlink_write_queue(skb
, sk
);
2729 skb_header_release(nskb
);
2730 __tcp_add_write_queue_head(sk
, nskb
);
2731 sk_wmem_free_skb(sk
, skb
);
2732 sk
->sk_wmem_queued
+= nskb
->truesize
;
2733 sk_mem_charge(sk
, nskb
->truesize
);
2737 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
2738 TCP_ECN_send_synack(tcp_sk(sk
), skb
);
2740 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2741 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2745 * tcp_make_synack - Prepare a SYN-ACK.
2746 * sk: listener socket
2747 * dst: dst entry attached to the SYNACK
2748 * req: request_sock pointer
2750 * Allocate one skb and build a SYNACK packet.
2751 * @dst is consumed : Caller should not use it again.
2753 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2754 struct request_sock
*req
,
2755 struct tcp_fastopen_cookie
*foc
)
2757 struct tcp_out_options opts
;
2758 struct inet_request_sock
*ireq
= inet_rsk(req
);
2759 struct tcp_sock
*tp
= tcp_sk(sk
);
2761 struct sk_buff
*skb
;
2762 struct tcp_md5sig_key
*md5
;
2763 int tcp_header_size
;
2766 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
, 1, GFP_ATOMIC
);
2767 if (unlikely(!skb
)) {
2771 /* Reserve space for headers. */
2772 skb_reserve(skb
, MAX_TCP_HEADER
);
2774 skb_dst_set(skb
, dst
);
2775 security_skb_owned_by(skb
, sk
);
2777 mss
= dst_metric_advmss(dst
);
2778 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2779 mss
= tp
->rx_opt
.user_mss
;
2781 if (req
->rcv_wnd
== 0) { /* ignored for retransmitted syns */
2783 /* Set this up on the first call only */
2784 req
->window_clamp
= tp
->window_clamp
? : dst_metric(dst
, RTAX_WINDOW
);
2786 /* limit the window selection if the user enforce a smaller rx buffer */
2787 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2788 (req
->window_clamp
> tcp_full_space(sk
) || req
->window_clamp
== 0))
2789 req
->window_clamp
= tcp_full_space(sk
);
2791 /* tcp_full_space because it is guaranteed to be the first packet */
2792 tcp_select_initial_window(tcp_full_space(sk
),
2793 mss
- (ireq
->tstamp_ok
? TCPOLEN_TSTAMP_ALIGNED
: 0),
2798 dst_metric(dst
, RTAX_INITRWND
));
2799 ireq
->rcv_wscale
= rcv_wscale
;
2802 memset(&opts
, 0, sizeof(opts
));
2803 #ifdef CONFIG_SYN_COOKIES
2804 if (unlikely(req
->cookie_ts
))
2805 TCP_SKB_CB(skb
)->when
= cookie_init_timestamp(req
);
2808 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
2809 tcp_header_size
= tcp_synack_options(sk
, req
, mss
, skb
, &opts
, &md5
,
2812 skb_push(skb
, tcp_header_size
);
2813 skb_reset_transport_header(skb
);
2816 memset(th
, 0, sizeof(struct tcphdr
));
2819 TCP_ECN_make_synack(req
, th
);
2820 th
->source
= htons(ireq
->ir_num
);
2821 th
->dest
= ireq
->ir_rmt_port
;
2822 /* Setting of flags are superfluous here for callers (and ECE is
2823 * not even correctly set)
2825 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2826 TCPHDR_SYN
| TCPHDR_ACK
);
2828 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2829 /* XXX data is queued and acked as is. No buffer/window check */
2830 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
2832 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2833 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2834 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2835 th
->doff
= (tcp_header_size
>> 2);
2836 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_OUTSEGS
);
2838 #ifdef CONFIG_TCP_MD5SIG
2839 /* Okay, we have all we need - do the md5 hash if needed */
2841 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2842 md5
, NULL
, req
, skb
);
2848 EXPORT_SYMBOL(tcp_make_synack
);
2850 /* Do all connect socket setups that can be done AF independent. */
2851 static void tcp_connect_init(struct sock
*sk
)
2853 const struct dst_entry
*dst
= __sk_dst_get(sk
);
2854 struct tcp_sock
*tp
= tcp_sk(sk
);
2857 /* We'll fix this up when we get a response from the other end.
2858 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2860 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
2861 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
2863 #ifdef CONFIG_TCP_MD5SIG
2864 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
2865 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
2868 /* If user gave his TCP_MAXSEG, record it to clamp */
2869 if (tp
->rx_opt
.user_mss
)
2870 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2873 tcp_sync_mss(sk
, dst_mtu(dst
));
2875 if (!tp
->window_clamp
)
2876 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
2877 tp
->advmss
= dst_metric_advmss(dst
);
2878 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
2879 tp
->advmss
= tp
->rx_opt
.user_mss
;
2881 tcp_initialize_rcv_mss(sk
);
2883 /* limit the window selection if the user enforce a smaller rx buffer */
2884 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
2885 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
2886 tp
->window_clamp
= tcp_full_space(sk
);
2888 tcp_select_initial_window(tcp_full_space(sk
),
2889 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
2892 sysctl_tcp_window_scaling
,
2894 dst_metric(dst
, RTAX_INITRWND
));
2896 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
2897 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
2900 sock_reset_flag(sk
, SOCK_DONE
);
2903 tp
->snd_una
= tp
->write_seq
;
2904 tp
->snd_sml
= tp
->write_seq
;
2905 tp
->snd_up
= tp
->write_seq
;
2906 tp
->snd_nxt
= tp
->write_seq
;
2908 if (likely(!tp
->repair
))
2911 tp
->rcv_tstamp
= tcp_time_stamp
;
2912 tp
->rcv_wup
= tp
->rcv_nxt
;
2913 tp
->copied_seq
= tp
->rcv_nxt
;
2915 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
2916 inet_csk(sk
)->icsk_retransmits
= 0;
2917 tcp_clear_retrans(tp
);
2920 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
2922 struct tcp_sock
*tp
= tcp_sk(sk
);
2923 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
2925 tcb
->end_seq
+= skb
->len
;
2926 skb_header_release(skb
);
2927 __tcp_add_write_queue_tail(sk
, skb
);
2928 sk
->sk_wmem_queued
+= skb
->truesize
;
2929 sk_mem_charge(sk
, skb
->truesize
);
2930 tp
->write_seq
= tcb
->end_seq
;
2931 tp
->packets_out
+= tcp_skb_pcount(skb
);
2934 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2935 * queue a data-only packet after the regular SYN, such that regular SYNs
2936 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2937 * only the SYN sequence, the data are retransmitted in the first ACK.
2938 * If cookie is not cached or other error occurs, falls back to send a
2939 * regular SYN with Fast Open cookie request option.
2941 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
2943 struct tcp_sock
*tp
= tcp_sk(sk
);
2944 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
2945 int syn_loss
= 0, space
, i
, err
= 0, iovlen
= fo
->data
->msg_iovlen
;
2946 struct sk_buff
*syn_data
= NULL
, *data
;
2947 unsigned long last_syn_loss
= 0;
2949 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
2950 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
2951 &syn_loss
, &last_syn_loss
);
2952 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2954 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
2955 fo
->cookie
.len
= -1;
2959 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
2960 fo
->cookie
.len
= -1;
2961 else if (fo
->cookie
.len
<= 0)
2964 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2965 * user-MSS. Reserve maximum option space for middleboxes that add
2966 * private TCP options. The cost is reduced data space in SYN :(
2968 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
2969 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
2970 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
2971 MAX_TCP_OPTION_SPACE
;
2973 space
= min_t(size_t, space
, fo
->size
);
2975 /* limit to order-0 allocations */
2976 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
2978 syn_data
= skb_copy_expand(syn
, MAX_TCP_HEADER
, space
,
2980 if (syn_data
== NULL
)
2983 for (i
= 0; i
< iovlen
&& syn_data
->len
< space
; ++i
) {
2984 struct iovec
*iov
= &fo
->data
->msg_iov
[i
];
2985 unsigned char __user
*from
= iov
->iov_base
;
2986 int len
= iov
->iov_len
;
2988 if (syn_data
->len
+ len
> space
)
2989 len
= space
- syn_data
->len
;
2990 else if (i
+ 1 == iovlen
)
2991 /* No more data pending in inet_wait_for_connect() */
2994 if (skb_add_data(syn_data
, from
, len
))
2998 /* Queue a data-only packet after the regular SYN for retransmission */
2999 data
= pskb_copy(syn_data
, sk
->sk_allocation
);
3002 TCP_SKB_CB(data
)->seq
++;
3003 TCP_SKB_CB(data
)->tcp_flags
&= ~TCPHDR_SYN
;
3004 TCP_SKB_CB(data
)->tcp_flags
= (TCPHDR_ACK
|TCPHDR_PSH
);
3005 tcp_connect_queue_skb(sk
, data
);
3006 fo
->copied
= data
->len
;
3008 /* syn_data is about to be sent, we need to take current time stamps
3009 * for the packets that are in write queue : SYN packet and DATA
3011 skb_mstamp_get(&syn
->skb_mstamp
);
3012 data
->skb_mstamp
= syn
->skb_mstamp
;
3014 if (tcp_transmit_skb(sk
, syn_data
, 0, sk
->sk_allocation
) == 0) {
3015 tp
->syn_data
= (fo
->copied
> 0);
3016 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3022 /* Send a regular SYN with Fast Open cookie request option */
3023 if (fo
->cookie
.len
> 0)
3025 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3027 tp
->syn_fastopen
= 0;
3028 kfree_skb(syn_data
);
3030 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3034 /* Build a SYN and send it off. */
3035 int tcp_connect(struct sock
*sk
)
3037 struct tcp_sock
*tp
= tcp_sk(sk
);
3038 struct sk_buff
*buff
;
3041 tcp_connect_init(sk
);
3043 if (unlikely(tp
->repair
)) {
3044 tcp_finish_connect(sk
, NULL
);
3048 buff
= alloc_skb_fclone(MAX_TCP_HEADER
+ 15, sk
->sk_allocation
);
3049 if (unlikely(buff
== NULL
))
3052 /* Reserve space for headers. */
3053 skb_reserve(buff
, MAX_TCP_HEADER
);
3055 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3056 tp
->retrans_stamp
= TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
3057 tcp_connect_queue_skb(sk
, buff
);
3058 TCP_ECN_send_syn(sk
, buff
);
3060 /* Send off SYN; include data in Fast Open. */
3061 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3062 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3063 if (err
== -ECONNREFUSED
)
3066 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3067 * in order to make this packet get counted in tcpOutSegs.
3069 tp
->snd_nxt
= tp
->write_seq
;
3070 tp
->pushed_seq
= tp
->write_seq
;
3071 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3073 /* Timer for repeating the SYN until an answer. */
3074 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3075 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3078 EXPORT_SYMBOL(tcp_connect
);
3080 /* Send out a delayed ack, the caller does the policy checking
3081 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3084 void tcp_send_delayed_ack(struct sock
*sk
)
3086 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3087 int ato
= icsk
->icsk_ack
.ato
;
3088 unsigned long timeout
;
3090 if (ato
> TCP_DELACK_MIN
) {
3091 const struct tcp_sock
*tp
= tcp_sk(sk
);
3092 int max_ato
= HZ
/ 2;
3094 if (icsk
->icsk_ack
.pingpong
||
3095 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3096 max_ato
= TCP_DELACK_MAX
;
3098 /* Slow path, intersegment interval is "high". */
3100 /* If some rtt estimate is known, use it to bound delayed ack.
3101 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3105 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3112 ato
= min(ato
, max_ato
);
3115 /* Stay within the limit we were given */
3116 timeout
= jiffies
+ ato
;
3118 /* Use new timeout only if there wasn't a older one earlier. */
3119 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3120 /* If delack timer was blocked or is about to expire,
3123 if (icsk
->icsk_ack
.blocked
||
3124 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3129 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3130 timeout
= icsk
->icsk_ack
.timeout
;
3132 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3133 icsk
->icsk_ack
.timeout
= timeout
;
3134 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3137 /* This routine sends an ack and also updates the window. */
3138 void tcp_send_ack(struct sock
*sk
)
3140 struct sk_buff
*buff
;
3142 /* If we have been reset, we may not send again. */
3143 if (sk
->sk_state
== TCP_CLOSE
)
3146 /* We are not putting this on the write queue, so
3147 * tcp_transmit_skb() will set the ownership to this
3150 buff
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3152 inet_csk_schedule_ack(sk
);
3153 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3154 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3155 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3159 /* Reserve space for headers and prepare control bits. */
3160 skb_reserve(buff
, MAX_TCP_HEADER
);
3161 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3163 /* Send it off, this clears delayed acks for us. */
3164 TCP_SKB_CB(buff
)->when
= tcp_time_stamp
;
3165 tcp_transmit_skb(sk
, buff
, 0, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3168 /* This routine sends a packet with an out of date sequence
3169 * number. It assumes the other end will try to ack it.
3171 * Question: what should we make while urgent mode?
3172 * 4.4BSD forces sending single byte of data. We cannot send
3173 * out of window data, because we have SND.NXT==SND.MAX...
3175 * Current solution: to send TWO zero-length segments in urgent mode:
3176 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3177 * out-of-date with SND.UNA-1 to probe window.
3179 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
3181 struct tcp_sock
*tp
= tcp_sk(sk
);
3182 struct sk_buff
*skb
;
3184 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3185 skb
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3189 /* Reserve space for headers and set control bits. */
3190 skb_reserve(skb
, MAX_TCP_HEADER
);
3191 /* Use a previous sequence. This should cause the other
3192 * end to send an ack. Don't queue or clone SKB, just
3195 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3196 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3197 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
3200 void tcp_send_window_probe(struct sock
*sk
)
3202 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3203 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3204 tcp_xmit_probe_skb(sk
, 0);
3208 /* Initiate keepalive or window probe from timer. */
3209 int tcp_write_wakeup(struct sock
*sk
)
3211 struct tcp_sock
*tp
= tcp_sk(sk
);
3212 struct sk_buff
*skb
;
3214 if (sk
->sk_state
== TCP_CLOSE
)
3217 if ((skb
= tcp_send_head(sk
)) != NULL
&&
3218 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3220 unsigned int mss
= tcp_current_mss(sk
);
3221 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3223 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3224 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3226 /* We are probing the opening of a window
3227 * but the window size is != 0
3228 * must have been a result SWS avoidance ( sender )
3230 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3232 seg_size
= min(seg_size
, mss
);
3233 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3234 if (tcp_fragment(sk
, skb
, seg_size
, mss
))
3236 } else if (!tcp_skb_pcount(skb
))
3237 tcp_set_skb_tso_segs(sk
, skb
, mss
);
3239 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3240 TCP_SKB_CB(skb
)->when
= tcp_time_stamp
;
3241 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3243 tcp_event_new_data_sent(sk
, skb
);
3246 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3247 tcp_xmit_probe_skb(sk
, 1);
3248 return tcp_xmit_probe_skb(sk
, 0);
3252 /* A window probe timeout has occurred. If window is not closed send
3253 * a partial packet else a zero probe.
3255 void tcp_send_probe0(struct sock
*sk
)
3257 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3258 struct tcp_sock
*tp
= tcp_sk(sk
);
3261 err
= tcp_write_wakeup(sk
);
3263 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3264 /* Cancel probe timer, if it is not required. */
3265 icsk
->icsk_probes_out
= 0;
3266 icsk
->icsk_backoff
= 0;
3271 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
3272 icsk
->icsk_backoff
++;
3273 icsk
->icsk_probes_out
++;
3274 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3275 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
, TCP_RTO_MAX
),
3278 /* If packet was not sent due to local congestion,
3279 * do not backoff and do not remember icsk_probes_out.
3280 * Let local senders to fight for local resources.
3282 * Use accumulated backoff yet.
3284 if (!icsk
->icsk_probes_out
)
3285 icsk
->icsk_probes_out
= 1;
3286 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3287 min(icsk
->icsk_rto
<< icsk
->icsk_backoff
,
3288 TCP_RESOURCE_PROBE_INTERVAL
),