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 four TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly
= 262144;
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 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
65 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
66 int push_one
, gfp_t gfp
);
68 /* Account for new data that has been sent to the network. */
69 static void tcp_event_new_data_sent(struct sock
*sk
, const struct sk_buff
*skb
)
71 struct inet_connection_sock
*icsk
= inet_csk(sk
);
72 struct tcp_sock
*tp
= tcp_sk(sk
);
73 unsigned int prior_packets
= tp
->packets_out
;
75 tcp_advance_send_head(sk
, skb
);
76 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
78 tp
->packets_out
+= tcp_skb_pcount(skb
);
79 if (!prior_packets
|| icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
80 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
84 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
,
88 /* SND.NXT, if window was not shrunk.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
94 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
96 const struct tcp_sock
*tp
= tcp_sk(sk
);
98 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
101 return tcp_wnd_end(tp
);
104 /* Calculate mss to advertise in SYN segment.
105 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
107 * 1. It is independent of path mtu.
108 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
109 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
110 * attached devices, because some buggy hosts are confused by
112 * 4. We do not make 3, we advertise MSS, calculated from first
113 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
114 * This may be overridden via information stored in routing table.
115 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
116 * probably even Jumbo".
118 static __u16
tcp_advertise_mss(struct sock
*sk
)
120 struct tcp_sock
*tp
= tcp_sk(sk
);
121 const struct dst_entry
*dst
= __sk_dst_get(sk
);
122 int mss
= tp
->advmss
;
125 unsigned int metric
= dst_metric_advmss(dst
);
136 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
137 * This is the first part of cwnd validation mechanism.
139 void tcp_cwnd_restart(struct sock
*sk
, s32 delta
)
141 struct tcp_sock
*tp
= tcp_sk(sk
);
142 u32 restart_cwnd
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
143 u32 cwnd
= tp
->snd_cwnd
;
145 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
147 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
148 restart_cwnd
= min(restart_cwnd
, cwnd
);
150 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
152 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
153 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
154 tp
->snd_cwnd_used
= 0;
157 /* Congestion state accounting after a packet has been sent. */
158 static void tcp_event_data_sent(struct tcp_sock
*tp
,
161 struct inet_connection_sock
*icsk
= inet_csk(sk
);
162 const u32 now
= tcp_time_stamp
;
164 if (tcp_packets_in_flight(tp
) == 0)
165 tcp_ca_event(sk
, CA_EVENT_TX_START
);
169 /* If it is a reply for ato after last received
170 * packet, enter pingpong mode.
172 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
173 icsk
->icsk_ack
.pingpong
= 1;
176 /* Account for an ACK we sent. */
177 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
179 tcp_dec_quickack_mode(sk
, pkts
);
180 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
184 u32
tcp_default_init_rwnd(u32 mss
)
186 /* Initial receive window should be twice of TCP_INIT_CWND to
187 * enable proper sending of new unsent data during fast recovery
188 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
189 * limit when mss is larger than 1460.
191 u32 init_rwnd
= TCP_INIT_CWND
* 2;
194 init_rwnd
= max((1460 * init_rwnd
) / mss
, 2U);
198 /* Determine a window scaling and initial window to offer.
199 * Based on the assumption that the given amount of space
200 * will be offered. Store the results in the tp structure.
201 * NOTE: for smooth operation initial space offering should
202 * be a multiple of mss if possible. We assume here that mss >= 1.
203 * This MUST be enforced by all callers.
205 void tcp_select_initial_window(int __space
, __u32 mss
,
206 __u32
*rcv_wnd
, __u32
*window_clamp
,
207 int wscale_ok
, __u8
*rcv_wscale
,
210 unsigned int space
= (__space
< 0 ? 0 : __space
);
212 /* If no clamp set the clamp to the max possible scaled window */
213 if (*window_clamp
== 0)
214 (*window_clamp
) = (65535 << 14);
215 space
= min(*window_clamp
, space
);
217 /* Quantize space offering to a multiple of mss if possible. */
219 space
= (space
/ mss
) * mss
;
221 /* NOTE: offering an initial window larger than 32767
222 * will break some buggy TCP stacks. If the admin tells us
223 * it is likely we could be speaking with such a buggy stack
224 * we will truncate our initial window offering to 32K-1
225 * unless the remote has sent us a window scaling option,
226 * which we interpret as a sign the remote TCP is not
227 * misinterpreting the window field as a signed quantity.
229 if (sysctl_tcp_workaround_signed_windows
)
230 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
236 /* Set window scaling on max possible window
237 * See RFC1323 for an explanation of the limit to 14
239 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
240 space
= min_t(u32
, space
, *window_clamp
);
241 while (space
> 65535 && (*rcv_wscale
) < 14) {
247 if (mss
> (1 << *rcv_wscale
)) {
248 if (!init_rcv_wnd
) /* Use default unless specified otherwise */
249 init_rcv_wnd
= tcp_default_init_rwnd(mss
);
250 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
253 /* Set the clamp no higher than max representable value */
254 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
256 EXPORT_SYMBOL(tcp_select_initial_window
);
258 /* Chose a new window to advertise, update state in tcp_sock for the
259 * socket, and return result with RFC1323 scaling applied. The return
260 * value can be stuffed directly into th->window for an outgoing
263 static u16
tcp_select_window(struct sock
*sk
)
265 struct tcp_sock
*tp
= tcp_sk(sk
);
266 u32 old_win
= tp
->rcv_wnd
;
267 u32 cur_win
= tcp_receive_window(tp
);
268 u32 new_win
= __tcp_select_window(sk
);
270 /* Never shrink the offered window */
271 if (new_win
< cur_win
) {
272 /* Danger Will Robinson!
273 * Don't update rcv_wup/rcv_wnd here or else
274 * we will not be able to advertise a zero
275 * window in time. --DaveM
277 * Relax Will Robinson.
280 NET_INC_STATS(sock_net(sk
),
281 LINUX_MIB_TCPWANTZEROWINDOWADV
);
282 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
284 tp
->rcv_wnd
= new_win
;
285 tp
->rcv_wup
= tp
->rcv_nxt
;
287 /* Make sure we do not exceed the maximum possible
290 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
291 new_win
= min(new_win
, MAX_TCP_WINDOW
);
293 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
295 /* RFC1323 scaling applied */
296 new_win
>>= tp
->rx_opt
.rcv_wscale
;
298 /* If we advertise zero window, disable fast path. */
302 NET_INC_STATS(sock_net(sk
),
303 LINUX_MIB_TCPTOZEROWINDOWADV
);
304 } else if (old_win
== 0) {
305 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFROMZEROWINDOWADV
);
311 /* Packet ECN state for a SYN-ACK */
312 static void tcp_ecn_send_synack(struct sock
*sk
, struct sk_buff
*skb
)
314 const struct tcp_sock
*tp
= tcp_sk(sk
);
316 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
317 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
318 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
319 else if (tcp_ca_needs_ecn(sk
))
323 /* Packet ECN state for a SYN. */
324 static void tcp_ecn_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
326 struct tcp_sock
*tp
= tcp_sk(sk
);
327 bool use_ecn
= sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1 ||
328 tcp_ca_needs_ecn(sk
);
331 const struct dst_entry
*dst
= __sk_dst_get(sk
);
333 if (dst
&& dst_feature(dst
, RTAX_FEATURE_ECN
))
340 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
341 tp
->ecn_flags
= TCP_ECN_OK
;
342 if (tcp_ca_needs_ecn(sk
))
347 static void tcp_ecn_clear_syn(struct sock
*sk
, struct sk_buff
*skb
)
349 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn_fallback
)
350 /* tp->ecn_flags are cleared at a later point in time when
351 * SYN ACK is ultimatively being received.
353 TCP_SKB_CB(skb
)->tcp_flags
&= ~(TCPHDR_ECE
| TCPHDR_CWR
);
357 tcp_ecn_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
359 if (inet_rsk(req
)->ecn_ok
)
363 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
366 static void tcp_ecn_send(struct sock
*sk
, struct sk_buff
*skb
,
369 struct tcp_sock
*tp
= tcp_sk(sk
);
371 if (tp
->ecn_flags
& TCP_ECN_OK
) {
372 /* Not-retransmitted data segment: set ECT and inject CWR. */
373 if (skb
->len
!= tcp_header_len
&&
374 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
376 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
377 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
378 tcp_hdr(skb
)->cwr
= 1;
379 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
381 } else if (!tcp_ca_needs_ecn(sk
)) {
382 /* ACK or retransmitted segment: clear ECT|CE */
383 INET_ECN_dontxmit(sk
);
385 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
386 tcp_hdr(skb
)->ece
= 1;
390 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
391 * auto increment end seqno.
393 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
395 skb
->ip_summed
= CHECKSUM_PARTIAL
;
398 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
399 TCP_SKB_CB(skb
)->sacked
= 0;
401 tcp_skb_pcount_set(skb
, 1);
403 TCP_SKB_CB(skb
)->seq
= seq
;
404 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
406 TCP_SKB_CB(skb
)->end_seq
= seq
;
409 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
411 return tp
->snd_una
!= tp
->snd_up
;
414 #define OPTION_SACK_ADVERTISE (1 << 0)
415 #define OPTION_TS (1 << 1)
416 #define OPTION_MD5 (1 << 2)
417 #define OPTION_WSCALE (1 << 3)
418 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
420 struct tcp_out_options
{
421 u16 options
; /* bit field of OPTION_* */
422 u16 mss
; /* 0 to disable */
423 u8 ws
; /* window scale, 0 to disable */
424 u8 num_sack_blocks
; /* number of SACK blocks to include */
425 u8 hash_size
; /* bytes in hash_location */
426 __u8
*hash_location
; /* temporary pointer, overloaded */
427 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
428 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
431 /* Write previously computed TCP options to the packet.
433 * Beware: Something in the Internet is very sensitive to the ordering of
434 * TCP options, we learned this through the hard way, so be careful here.
435 * Luckily we can at least blame others for their non-compliance but from
436 * inter-operability perspective it seems that we're somewhat stuck with
437 * the ordering which we have been using if we want to keep working with
438 * those broken things (not that it currently hurts anybody as there isn't
439 * particular reason why the ordering would need to be changed).
441 * At least SACK_PERM as the first option is known to lead to a disaster
442 * (but it may well be that other scenarios fail similarly).
444 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
445 struct tcp_out_options
*opts
)
447 u16 options
= opts
->options
; /* mungable copy */
449 if (unlikely(OPTION_MD5
& options
)) {
450 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
451 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
452 /* overload cookie hash location */
453 opts
->hash_location
= (__u8
*)ptr
;
457 if (unlikely(opts
->mss
)) {
458 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
459 (TCPOLEN_MSS
<< 16) |
463 if (likely(OPTION_TS
& options
)) {
464 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
465 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
466 (TCPOLEN_SACK_PERM
<< 16) |
467 (TCPOPT_TIMESTAMP
<< 8) |
469 options
&= ~OPTION_SACK_ADVERTISE
;
471 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
473 (TCPOPT_TIMESTAMP
<< 8) |
476 *ptr
++ = htonl(opts
->tsval
);
477 *ptr
++ = htonl(opts
->tsecr
);
480 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
481 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
483 (TCPOPT_SACK_PERM
<< 8) |
487 if (unlikely(OPTION_WSCALE
& options
)) {
488 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
489 (TCPOPT_WINDOW
<< 16) |
490 (TCPOLEN_WINDOW
<< 8) |
494 if (unlikely(opts
->num_sack_blocks
)) {
495 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
496 tp
->duplicate_sack
: tp
->selective_acks
;
499 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
502 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
503 TCPOLEN_SACK_PERBLOCK
)));
505 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
507 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
508 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
511 tp
->rx_opt
.dsack
= 0;
514 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
515 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
517 u32 len
; /* Fast Open option length */
520 len
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
521 *ptr
= htonl((TCPOPT_EXP
<< 24) | (len
<< 16) |
522 TCPOPT_FASTOPEN_MAGIC
);
523 p
+= TCPOLEN_EXP_FASTOPEN_BASE
;
525 len
= TCPOLEN_FASTOPEN_BASE
+ foc
->len
;
526 *p
++ = TCPOPT_FASTOPEN
;
530 memcpy(p
, foc
->val
, foc
->len
);
531 if ((len
& 3) == 2) {
532 p
[foc
->len
] = TCPOPT_NOP
;
533 p
[foc
->len
+ 1] = TCPOPT_NOP
;
535 ptr
+= (len
+ 3) >> 2;
539 /* Compute TCP options for SYN packets. This is not the final
540 * network wire format yet.
542 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
543 struct tcp_out_options
*opts
,
544 struct tcp_md5sig_key
**md5
)
546 struct tcp_sock
*tp
= tcp_sk(sk
);
547 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
548 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
550 #ifdef CONFIG_TCP_MD5SIG
551 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
553 opts
->options
|= OPTION_MD5
;
554 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
560 /* We always get an MSS option. The option bytes which will be seen in
561 * normal data packets should timestamps be used, must be in the MSS
562 * advertised. But we subtract them from tp->mss_cache so that
563 * calculations in tcp_sendmsg are simpler etc. So account for this
564 * fact here if necessary. If we don't do this correctly, as a
565 * receiver we won't recognize data packets as being full sized when we
566 * should, and thus we won't abide by the delayed ACK rules correctly.
567 * SACKs don't matter, we never delay an ACK when we have any of those
569 opts
->mss
= tcp_advertise_mss(sk
);
570 remaining
-= TCPOLEN_MSS_ALIGNED
;
572 if (likely(sysctl_tcp_timestamps
&& !*md5
)) {
573 opts
->options
|= OPTION_TS
;
574 opts
->tsval
= tcp_skb_timestamp(skb
) + tp
->tsoffset
;
575 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
576 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
578 if (likely(sysctl_tcp_window_scaling
)) {
579 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
580 opts
->options
|= OPTION_WSCALE
;
581 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
583 if (likely(sysctl_tcp_sack
)) {
584 opts
->options
|= OPTION_SACK_ADVERTISE
;
585 if (unlikely(!(OPTION_TS
& opts
->options
)))
586 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
589 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
590 u32 need
= fastopen
->cookie
.len
;
592 need
+= fastopen
->cookie
.exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
593 TCPOLEN_FASTOPEN_BASE
;
594 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
595 if (remaining
>= need
) {
596 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
597 opts
->fastopen_cookie
= &fastopen
->cookie
;
599 tp
->syn_fastopen
= 1;
600 tp
->syn_fastopen_exp
= fastopen
->cookie
.exp
? 1 : 0;
604 return MAX_TCP_OPTION_SPACE
- remaining
;
607 /* Set up TCP options for SYN-ACKs. */
608 static unsigned int tcp_synack_options(struct request_sock
*req
,
609 unsigned int mss
, struct sk_buff
*skb
,
610 struct tcp_out_options
*opts
,
611 const struct tcp_md5sig_key
*md5
,
612 struct tcp_fastopen_cookie
*foc
)
614 struct inet_request_sock
*ireq
= inet_rsk(req
);
615 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
617 #ifdef CONFIG_TCP_MD5SIG
619 opts
->options
|= OPTION_MD5
;
620 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
622 /* We can't fit any SACK blocks in a packet with MD5 + TS
623 * options. There was discussion about disabling SACK
624 * rather than TS in order to fit in better with old,
625 * buggy kernels, but that was deemed to be unnecessary.
627 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
631 /* We always send an MSS option. */
633 remaining
-= TCPOLEN_MSS_ALIGNED
;
635 if (likely(ireq
->wscale_ok
)) {
636 opts
->ws
= ireq
->rcv_wscale
;
637 opts
->options
|= OPTION_WSCALE
;
638 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
640 if (likely(ireq
->tstamp_ok
)) {
641 opts
->options
|= OPTION_TS
;
642 opts
->tsval
= tcp_skb_timestamp(skb
);
643 opts
->tsecr
= req
->ts_recent
;
644 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
646 if (likely(ireq
->sack_ok
)) {
647 opts
->options
|= OPTION_SACK_ADVERTISE
;
648 if (unlikely(!ireq
->tstamp_ok
))
649 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
651 if (foc
!= NULL
&& foc
->len
>= 0) {
654 need
+= foc
->exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
655 TCPOLEN_FASTOPEN_BASE
;
656 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
657 if (remaining
>= need
) {
658 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
659 opts
->fastopen_cookie
= foc
;
664 return MAX_TCP_OPTION_SPACE
- remaining
;
667 /* Compute TCP options for ESTABLISHED sockets. This is not the
668 * final wire format yet.
670 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
671 struct tcp_out_options
*opts
,
672 struct tcp_md5sig_key
**md5
)
674 struct tcp_sock
*tp
= tcp_sk(sk
);
675 unsigned int size
= 0;
676 unsigned int eff_sacks
;
680 #ifdef CONFIG_TCP_MD5SIG
681 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
682 if (unlikely(*md5
)) {
683 opts
->options
|= OPTION_MD5
;
684 size
+= TCPOLEN_MD5SIG_ALIGNED
;
690 if (likely(tp
->rx_opt
.tstamp_ok
)) {
691 opts
->options
|= OPTION_TS
;
692 opts
->tsval
= skb
? tcp_skb_timestamp(skb
) + tp
->tsoffset
: 0;
693 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
694 size
+= TCPOLEN_TSTAMP_ALIGNED
;
697 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
698 if (unlikely(eff_sacks
)) {
699 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
700 opts
->num_sack_blocks
=
701 min_t(unsigned int, eff_sacks
,
702 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
703 TCPOLEN_SACK_PERBLOCK
);
704 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
705 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
712 /* TCP SMALL QUEUES (TSQ)
714 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
715 * to reduce RTT and bufferbloat.
716 * We do this using a special skb destructor (tcp_wfree).
718 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
719 * needs to be reallocated in a driver.
720 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
722 * Since transmit from skb destructor is forbidden, we use a tasklet
723 * to process all sockets that eventually need to send more skbs.
724 * We use one tasklet per cpu, with its own queue of sockets.
727 struct tasklet_struct tasklet
;
728 struct list_head head
; /* queue of tcp sockets */
730 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
732 static void tcp_tsq_handler(struct sock
*sk
)
734 if ((1 << sk
->sk_state
) &
735 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
736 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
))
737 tcp_write_xmit(sk
, tcp_current_mss(sk
), tcp_sk(sk
)->nonagle
,
741 * One tasklet per cpu tries to send more skbs.
742 * We run in tasklet context but need to disable irqs when
743 * transferring tsq->head because tcp_wfree() might
744 * interrupt us (non NAPI drivers)
746 static void tcp_tasklet_func(unsigned long data
)
748 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
751 struct list_head
*q
, *n
;
755 local_irq_save(flags
);
756 list_splice_init(&tsq
->head
, &list
);
757 local_irq_restore(flags
);
759 list_for_each_safe(q
, n
, &list
) {
760 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
761 list_del(&tp
->tsq_node
);
763 sk
= (struct sock
*)tp
;
766 if (!sock_owned_by_user(sk
)) {
769 /* defer the work to tcp_release_cb() */
770 set_bit(TCP_TSQ_DEFERRED
, &tp
->tsq_flags
);
774 clear_bit(TSQ_QUEUED
, &tp
->tsq_flags
);
779 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
780 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
781 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
782 (1UL << TCP_MTU_REDUCED_DEFERRED))
784 * tcp_release_cb - tcp release_sock() callback
787 * called from release_sock() to perform protocol dependent
788 * actions before socket release.
790 void tcp_release_cb(struct sock
*sk
)
792 struct tcp_sock
*tp
= tcp_sk(sk
);
793 unsigned long flags
, nflags
;
795 /* perform an atomic operation only if at least one flag is set */
797 flags
= tp
->tsq_flags
;
798 if (!(flags
& TCP_DEFERRED_ALL
))
800 nflags
= flags
& ~TCP_DEFERRED_ALL
;
801 } while (cmpxchg(&tp
->tsq_flags
, flags
, nflags
) != flags
);
803 if (flags
& (1UL << TCP_TSQ_DEFERRED
))
806 /* Here begins the tricky part :
807 * We are called from release_sock() with :
809 * 2) sk_lock.slock spinlock held
810 * 3) socket owned by us (sk->sk_lock.owned == 1)
812 * But following code is meant to be called from BH handlers,
813 * so we should keep BH disabled, but early release socket ownership
815 sock_release_ownership(sk
);
817 if (flags
& (1UL << TCP_WRITE_TIMER_DEFERRED
)) {
818 tcp_write_timer_handler(sk
);
821 if (flags
& (1UL << TCP_DELACK_TIMER_DEFERRED
)) {
822 tcp_delack_timer_handler(sk
);
825 if (flags
& (1UL << TCP_MTU_REDUCED_DEFERRED
)) {
826 inet_csk(sk
)->icsk_af_ops
->mtu_reduced(sk
);
830 EXPORT_SYMBOL(tcp_release_cb
);
832 void __init
tcp_tasklet_init(void)
836 for_each_possible_cpu(i
) {
837 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
839 INIT_LIST_HEAD(&tsq
->head
);
840 tasklet_init(&tsq
->tasklet
,
847 * Write buffer destructor automatically called from kfree_skb.
848 * We can't xmit new skbs from this context, as we might already
851 void tcp_wfree(struct sk_buff
*skb
)
853 struct sock
*sk
= skb
->sk
;
854 struct tcp_sock
*tp
= tcp_sk(sk
);
857 /* Keep one reference on sk_wmem_alloc.
858 * Will be released by sk_free() from here or tcp_tasklet_func()
860 wmem
= atomic_sub_return(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
862 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
863 * Wait until our queues (qdisc + devices) are drained.
865 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
866 * - chance for incoming ACK (processed by another cpu maybe)
867 * to migrate this flow (skb->ooo_okay will be eventually set)
869 if (wmem
>= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current
)
872 if (test_and_clear_bit(TSQ_THROTTLED
, &tp
->tsq_flags
) &&
873 !test_and_set_bit(TSQ_QUEUED
, &tp
->tsq_flags
)) {
875 struct tsq_tasklet
*tsq
;
877 /* queue this socket to tasklet queue */
878 local_irq_save(flags
);
879 tsq
= this_cpu_ptr(&tsq_tasklet
);
880 list_add(&tp
->tsq_node
, &tsq
->head
);
881 tasklet_schedule(&tsq
->tasklet
);
882 local_irq_restore(flags
);
889 /* This routine actually transmits TCP packets queued in by
890 * tcp_do_sendmsg(). This is used by both the initial
891 * transmission and possible later retransmissions.
892 * All SKB's seen here are completely headerless. It is our
893 * job to build the TCP header, and pass the packet down to
894 * IP so it can do the same plus pass the packet off to the
897 * We are working here with either a clone of the original
898 * SKB, or a fresh unique copy made by the retransmit engine.
900 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
903 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
904 struct inet_sock
*inet
;
906 struct tcp_skb_cb
*tcb
;
907 struct tcp_out_options opts
;
908 unsigned int tcp_options_size
, tcp_header_size
;
909 struct tcp_md5sig_key
*md5
;
913 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
916 skb_mstamp_get(&skb
->skb_mstamp
);
918 if (unlikely(skb_cloned(skb
)))
919 skb
= pskb_copy(skb
, gfp_mask
);
921 skb
= skb_clone(skb
, gfp_mask
);
928 tcb
= TCP_SKB_CB(skb
);
929 memset(&opts
, 0, sizeof(opts
));
931 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
932 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
934 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
936 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
938 /* if no packet is in qdisc/device queue, then allow XPS to select
939 * another queue. We can be called from tcp_tsq_handler()
940 * which holds one reference to sk_wmem_alloc.
942 * TODO: Ideally, in-flight pure ACK packets should not matter here.
943 * One way to get this would be to set skb->truesize = 2 on them.
945 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) < SKB_TRUESIZE(1);
947 skb_push(skb
, tcp_header_size
);
948 skb_reset_transport_header(skb
);
952 skb
->destructor
= skb_is_tcp_pure_ack(skb
) ? sock_wfree
: tcp_wfree
;
953 skb_set_hash_from_sk(skb
, sk
);
954 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
956 /* Build TCP header and checksum it. */
958 th
->source
= inet
->inet_sport
;
959 th
->dest
= inet
->inet_dport
;
960 th
->seq
= htonl(tcb
->seq
);
961 th
->ack_seq
= htonl(tp
->rcv_nxt
);
962 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
965 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
966 /* RFC1323: The window in SYN & SYN/ACK segments
969 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
971 th
->window
= htons(tcp_select_window(sk
));
976 /* The urg_mode check is necessary during a below snd_una win probe */
977 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
978 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
979 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
981 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
982 th
->urg_ptr
= htons(0xFFFF);
987 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
988 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
989 if (likely((tcb
->tcp_flags
& TCPHDR_SYN
) == 0))
990 tcp_ecn_send(sk
, skb
, tcp_header_size
);
992 #ifdef CONFIG_TCP_MD5SIG
993 /* Calculate the MD5 hash, as we have all we need now */
995 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
996 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
1001 icsk
->icsk_af_ops
->send_check(sk
, skb
);
1003 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
1004 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
1006 if (skb
->len
!= tcp_header_size
) {
1007 tcp_event_data_sent(tp
, sk
);
1008 tp
->data_segs_out
+= tcp_skb_pcount(skb
);
1011 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
1012 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
1013 tcp_skb_pcount(skb
));
1015 tp
->segs_out
+= tcp_skb_pcount(skb
);
1016 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1017 skb_shinfo(skb
)->gso_segs
= tcp_skb_pcount(skb
);
1018 skb_shinfo(skb
)->gso_size
= tcp_skb_mss(skb
);
1020 /* Our usage of tstamp should remain private */
1021 skb
->tstamp
.tv64
= 0;
1023 /* Cleanup our debris for IP stacks */
1024 memset(skb
->cb
, 0, max(sizeof(struct inet_skb_parm
),
1025 sizeof(struct inet6_skb_parm
)));
1027 err
= icsk
->icsk_af_ops
->queue_xmit(sk
, skb
, &inet
->cork
.fl
);
1029 if (likely(err
<= 0))
1034 return net_xmit_eval(err
);
1037 /* This routine just queues the buffer for sending.
1039 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1040 * otherwise socket can stall.
1042 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1044 struct tcp_sock
*tp
= tcp_sk(sk
);
1046 /* Advance write_seq and place onto the write_queue. */
1047 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
1048 __skb_header_release(skb
);
1049 tcp_add_write_queue_tail(sk
, skb
);
1050 sk
->sk_wmem_queued
+= skb
->truesize
;
1051 sk_mem_charge(sk
, skb
->truesize
);
1054 /* Initialize TSO segments for a packet. */
1055 static void tcp_set_skb_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1057 if (skb
->len
<= mss_now
|| skb
->ip_summed
== CHECKSUM_NONE
) {
1058 /* Avoid the costly divide in the normal
1061 tcp_skb_pcount_set(skb
, 1);
1062 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1064 tcp_skb_pcount_set(skb
, DIV_ROUND_UP(skb
->len
, mss_now
));
1065 TCP_SKB_CB(skb
)->tcp_gso_size
= mss_now
;
1069 /* When a modification to fackets out becomes necessary, we need to check
1070 * skb is counted to fackets_out or not.
1072 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1075 struct tcp_sock
*tp
= tcp_sk(sk
);
1077 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1080 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1081 tp
->fackets_out
-= decr
;
1084 /* Pcount in the middle of the write queue got changed, we need to do various
1085 * tweaks to fix counters
1087 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1089 struct tcp_sock
*tp
= tcp_sk(sk
);
1091 tp
->packets_out
-= decr
;
1093 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1094 tp
->sacked_out
-= decr
;
1095 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1096 tp
->retrans_out
-= decr
;
1097 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1098 tp
->lost_out
-= decr
;
1100 /* Reno case is special. Sigh... */
1101 if (tcp_is_reno(tp
) && decr
> 0)
1102 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1104 tcp_adjust_fackets_out(sk
, skb
, decr
);
1106 if (tp
->lost_skb_hint
&&
1107 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1108 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1109 tp
->lost_cnt_hint
-= decr
;
1111 tcp_verify_left_out(tp
);
1114 static bool tcp_has_tx_tstamp(const struct sk_buff
*skb
)
1116 return TCP_SKB_CB(skb
)->txstamp_ack
||
1117 (skb_shinfo(skb
)->tx_flags
& SKBTX_ANY_TSTAMP
);
1120 static void tcp_fragment_tstamp(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1122 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1124 if (unlikely(tcp_has_tx_tstamp(skb
)) &&
1125 !before(shinfo
->tskey
, TCP_SKB_CB(skb2
)->seq
)) {
1126 struct skb_shared_info
*shinfo2
= skb_shinfo(skb2
);
1127 u8 tsflags
= shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
1129 shinfo
->tx_flags
&= ~tsflags
;
1130 shinfo2
->tx_flags
|= tsflags
;
1131 swap(shinfo
->tskey
, shinfo2
->tskey
);
1132 TCP_SKB_CB(skb2
)->txstamp_ack
= TCP_SKB_CB(skb
)->txstamp_ack
;
1133 TCP_SKB_CB(skb
)->txstamp_ack
= 0;
1137 /* Function to create two new TCP segments. Shrinks the given segment
1138 * to the specified size and appends a new segment with the rest of the
1139 * packet to the list. This won't be called frequently, I hope.
1140 * Remember, these are still headerless SKBs at this point.
1142 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1143 unsigned int mss_now
, gfp_t gfp
)
1145 struct tcp_sock
*tp
= tcp_sk(sk
);
1146 struct sk_buff
*buff
;
1147 int nsize
, old_factor
;
1151 if (WARN_ON(len
> skb
->len
))
1154 nsize
= skb_headlen(skb
) - len
;
1158 if (skb_unclone(skb
, gfp
))
1161 /* Get a new skb... force flag on. */
1162 buff
= sk_stream_alloc_skb(sk
, nsize
, gfp
, true);
1164 return -ENOMEM
; /* We'll just try again later. */
1166 sk
->sk_wmem_queued
+= buff
->truesize
;
1167 sk_mem_charge(sk
, buff
->truesize
);
1168 nlen
= skb
->len
- len
- nsize
;
1169 buff
->truesize
+= nlen
;
1170 skb
->truesize
-= nlen
;
1172 /* Correct the sequence numbers. */
1173 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1174 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1175 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1177 /* PSH and FIN should only be set in the second packet. */
1178 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1179 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1180 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1181 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1183 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1184 /* Copy and checksum data tail into the new buffer. */
1185 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1186 skb_put(buff
, nsize
),
1191 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1193 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1194 skb_split(skb
, buff
, len
);
1197 buff
->ip_summed
= skb
->ip_summed
;
1199 buff
->tstamp
= skb
->tstamp
;
1200 tcp_fragment_tstamp(skb
, buff
);
1202 old_factor
= tcp_skb_pcount(skb
);
1204 /* Fix up tso_factor for both original and new SKB. */
1205 tcp_set_skb_tso_segs(skb
, mss_now
);
1206 tcp_set_skb_tso_segs(buff
, mss_now
);
1208 /* If this packet has been sent out already, we must
1209 * adjust the various packet counters.
1211 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1212 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1213 tcp_skb_pcount(buff
);
1216 tcp_adjust_pcount(sk
, skb
, diff
);
1219 /* Link BUFF into the send queue. */
1220 __skb_header_release(buff
);
1221 tcp_insert_write_queue_after(skb
, buff
, sk
);
1226 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1227 * eventually). The difference is that pulled data not copied, but
1228 * immediately discarded.
1230 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1232 struct skb_shared_info
*shinfo
;
1235 eat
= min_t(int, len
, skb_headlen(skb
));
1237 __skb_pull(skb
, eat
);
1244 shinfo
= skb_shinfo(skb
);
1245 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1246 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1249 skb_frag_unref(skb
, i
);
1252 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1254 shinfo
->frags
[k
].page_offset
+= eat
;
1255 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1261 shinfo
->nr_frags
= k
;
1263 skb_reset_tail_pointer(skb
);
1264 skb
->data_len
-= len
;
1265 skb
->len
= skb
->data_len
;
1268 /* Remove acked data from a packet in the transmit queue. */
1269 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1271 if (skb_unclone(skb
, GFP_ATOMIC
))
1274 __pskb_trim_head(skb
, len
);
1276 TCP_SKB_CB(skb
)->seq
+= len
;
1277 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1279 skb
->truesize
-= len
;
1280 sk
->sk_wmem_queued
-= len
;
1281 sk_mem_uncharge(sk
, len
);
1282 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1284 /* Any change of skb->len requires recalculation of tso factor. */
1285 if (tcp_skb_pcount(skb
) > 1)
1286 tcp_set_skb_tso_segs(skb
, tcp_skb_mss(skb
));
1291 /* Calculate MSS not accounting any TCP options. */
1292 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1294 const struct tcp_sock
*tp
= tcp_sk(sk
);
1295 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1298 /* Calculate base mss without TCP options:
1299 It is MMS_S - sizeof(tcphdr) of rfc1122
1301 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1303 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1304 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1305 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1307 if (dst
&& dst_allfrag(dst
))
1308 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1311 /* Clamp it (mss_clamp does not include tcp options) */
1312 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1313 mss_now
= tp
->rx_opt
.mss_clamp
;
1315 /* Now subtract optional transport overhead */
1316 mss_now
-= icsk
->icsk_ext_hdr_len
;
1318 /* Then reserve room for full set of TCP options and 8 bytes of data */
1324 /* Calculate MSS. Not accounting for SACKs here. */
1325 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1327 /* Subtract TCP options size, not including SACKs */
1328 return __tcp_mtu_to_mss(sk
, pmtu
) -
1329 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1332 /* Inverse of above */
1333 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1335 const struct tcp_sock
*tp
= tcp_sk(sk
);
1336 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1340 tp
->tcp_header_len
+
1341 icsk
->icsk_ext_hdr_len
+
1342 icsk
->icsk_af_ops
->net_header_len
;
1344 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1345 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1346 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1348 if (dst
&& dst_allfrag(dst
))
1349 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1354 /* MTU probing init per socket */
1355 void tcp_mtup_init(struct sock
*sk
)
1357 struct tcp_sock
*tp
= tcp_sk(sk
);
1358 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1359 struct net
*net
= sock_net(sk
);
1361 icsk
->icsk_mtup
.enabled
= net
->ipv4
.sysctl_tcp_mtu_probing
> 1;
1362 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1363 icsk
->icsk_af_ops
->net_header_len
;
1364 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, net
->ipv4
.sysctl_tcp_base_mss
);
1365 icsk
->icsk_mtup
.probe_size
= 0;
1366 if (icsk
->icsk_mtup
.enabled
)
1367 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1369 EXPORT_SYMBOL(tcp_mtup_init
);
1371 /* This function synchronize snd mss to current pmtu/exthdr set.
1373 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1374 for TCP options, but includes only bare TCP header.
1376 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1377 It is minimum of user_mss and mss received with SYN.
1378 It also does not include TCP options.
1380 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1382 tp->mss_cache is current effective sending mss, including
1383 all tcp options except for SACKs. It is evaluated,
1384 taking into account current pmtu, but never exceeds
1385 tp->rx_opt.mss_clamp.
1387 NOTE1. rfc1122 clearly states that advertised MSS
1388 DOES NOT include either tcp or ip options.
1390 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1391 are READ ONLY outside this function. --ANK (980731)
1393 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1395 struct tcp_sock
*tp
= tcp_sk(sk
);
1396 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1399 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1400 icsk
->icsk_mtup
.search_high
= pmtu
;
1402 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1403 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1405 /* And store cached results */
1406 icsk
->icsk_pmtu_cookie
= pmtu
;
1407 if (icsk
->icsk_mtup
.enabled
)
1408 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1409 tp
->mss_cache
= mss_now
;
1413 EXPORT_SYMBOL(tcp_sync_mss
);
1415 /* Compute the current effective MSS, taking SACKs and IP options,
1416 * and even PMTU discovery events into account.
1418 unsigned int tcp_current_mss(struct sock
*sk
)
1420 const struct tcp_sock
*tp
= tcp_sk(sk
);
1421 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1423 unsigned int header_len
;
1424 struct tcp_out_options opts
;
1425 struct tcp_md5sig_key
*md5
;
1427 mss_now
= tp
->mss_cache
;
1430 u32 mtu
= dst_mtu(dst
);
1431 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1432 mss_now
= tcp_sync_mss(sk
, mtu
);
1435 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1436 sizeof(struct tcphdr
);
1437 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1438 * some common options. If this is an odd packet (because we have SACK
1439 * blocks etc) then our calculated header_len will be different, and
1440 * we have to adjust mss_now correspondingly */
1441 if (header_len
!= tp
->tcp_header_len
) {
1442 int delta
= (int) header_len
- tp
->tcp_header_len
;
1449 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1450 * As additional protections, we do not touch cwnd in retransmission phases,
1451 * and if application hit its sndbuf limit recently.
1453 static void tcp_cwnd_application_limited(struct sock
*sk
)
1455 struct tcp_sock
*tp
= tcp_sk(sk
);
1457 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1458 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1459 /* Limited by application or receiver window. */
1460 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1461 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1462 if (win_used
< tp
->snd_cwnd
) {
1463 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1464 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
1466 tp
->snd_cwnd_used
= 0;
1468 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1471 static void tcp_cwnd_validate(struct sock
*sk
, bool is_cwnd_limited
)
1473 struct tcp_sock
*tp
= tcp_sk(sk
);
1475 /* Track the maximum number of outstanding packets in each
1476 * window, and remember whether we were cwnd-limited then.
1478 if (!before(tp
->snd_una
, tp
->max_packets_seq
) ||
1479 tp
->packets_out
> tp
->max_packets_out
) {
1480 tp
->max_packets_out
= tp
->packets_out
;
1481 tp
->max_packets_seq
= tp
->snd_nxt
;
1482 tp
->is_cwnd_limited
= is_cwnd_limited
;
1485 if (tcp_is_cwnd_limited(sk
)) {
1486 /* Network is feed fully. */
1487 tp
->snd_cwnd_used
= 0;
1488 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1490 /* Network starves. */
1491 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1492 tp
->snd_cwnd_used
= tp
->packets_out
;
1494 if (sysctl_tcp_slow_start_after_idle
&&
1495 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1496 tcp_cwnd_application_limited(sk
);
1500 /* Minshall's variant of the Nagle send check. */
1501 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1503 return after(tp
->snd_sml
, tp
->snd_una
) &&
1504 !after(tp
->snd_sml
, tp
->snd_nxt
);
1507 /* Update snd_sml if this skb is under mss
1508 * Note that a TSO packet might end with a sub-mss segment
1509 * The test is really :
1510 * if ((skb->len % mss) != 0)
1511 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1512 * But we can avoid doing the divide again given we already have
1513 * skb_pcount = skb->len / mss_now
1515 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1516 const struct sk_buff
*skb
)
1518 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1519 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1522 /* Return false, if packet can be sent now without violation Nagle's rules:
1523 * 1. It is full sized. (provided by caller in %partial bool)
1524 * 2. Or it contains FIN. (already checked by caller)
1525 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1526 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1527 * With Minshall's modification: all sent small packets are ACKed.
1529 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1533 ((nonagle
& TCP_NAGLE_CORK
) ||
1534 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1537 /* Return how many segs we'd like on a TSO packet,
1538 * to send one TSO packet per ms
1540 static u32
tcp_tso_autosize(const struct sock
*sk
, unsigned int mss_now
)
1544 bytes
= min(sk
->sk_pacing_rate
>> 10,
1545 sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
);
1547 /* Goal is to send at least one packet per ms,
1548 * not one big TSO packet every 100 ms.
1549 * This preserves ACK clocking and is consistent
1550 * with tcp_tso_should_defer() heuristic.
1552 segs
= max_t(u32
, bytes
/ mss_now
, sysctl_tcp_min_tso_segs
);
1554 return min_t(u32
, segs
, sk
->sk_gso_max_segs
);
1557 /* Returns the portion of skb which can be sent right away */
1558 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1559 const struct sk_buff
*skb
,
1560 unsigned int mss_now
,
1561 unsigned int max_segs
,
1564 const struct tcp_sock
*tp
= tcp_sk(sk
);
1565 u32 partial
, needed
, window
, max_len
;
1567 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1568 max_len
= mss_now
* max_segs
;
1570 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1573 needed
= min(skb
->len
, window
);
1575 if (max_len
<= needed
)
1578 partial
= needed
% mss_now
;
1579 /* If last segment is not a full MSS, check if Nagle rules allow us
1580 * to include this last segment in this skb.
1581 * Otherwise, we'll split the skb at last MSS boundary
1583 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
1584 return needed
- partial
;
1589 /* Can at least one segment of SKB be sent right now, according to the
1590 * congestion window rules? If so, return how many segments are allowed.
1592 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1593 const struct sk_buff
*skb
)
1595 u32 in_flight
, cwnd
, halfcwnd
;
1597 /* Don't be strict about the congestion window for the final FIN. */
1598 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1599 tcp_skb_pcount(skb
) == 1)
1602 in_flight
= tcp_packets_in_flight(tp
);
1603 cwnd
= tp
->snd_cwnd
;
1604 if (in_flight
>= cwnd
)
1607 /* For better scheduling, ensure we have at least
1608 * 2 GSO packets in flight.
1610 halfcwnd
= max(cwnd
>> 1, 1U);
1611 return min(halfcwnd
, cwnd
- in_flight
);
1614 /* Initialize TSO state of a skb.
1615 * This must be invoked the first time we consider transmitting
1616 * SKB onto the wire.
1618 static int tcp_init_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1620 int tso_segs
= tcp_skb_pcount(skb
);
1622 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1623 tcp_set_skb_tso_segs(skb
, mss_now
);
1624 tso_segs
= tcp_skb_pcount(skb
);
1630 /* Return true if the Nagle test allows this packet to be
1633 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1634 unsigned int cur_mss
, int nonagle
)
1636 /* Nagle rule does not apply to frames, which sit in the middle of the
1637 * write_queue (they have no chances to get new data).
1639 * This is implemented in the callers, where they modify the 'nonagle'
1640 * argument based upon the location of SKB in the send queue.
1642 if (nonagle
& TCP_NAGLE_PUSH
)
1645 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1646 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1649 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
1655 /* Does at least the first segment of SKB fit into the send window? */
1656 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1657 const struct sk_buff
*skb
,
1658 unsigned int cur_mss
)
1660 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1662 if (skb
->len
> cur_mss
)
1663 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1665 return !after(end_seq
, tcp_wnd_end(tp
));
1668 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1669 * should be put on the wire right now. If so, it returns the number of
1670 * packets allowed by the congestion window.
1672 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1673 unsigned int cur_mss
, int nonagle
)
1675 const struct tcp_sock
*tp
= tcp_sk(sk
);
1676 unsigned int cwnd_quota
;
1678 tcp_init_tso_segs(skb
, cur_mss
);
1680 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1683 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1684 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1690 /* Test if sending is allowed right now. */
1691 bool tcp_may_send_now(struct sock
*sk
)
1693 const struct tcp_sock
*tp
= tcp_sk(sk
);
1694 struct sk_buff
*skb
= tcp_send_head(sk
);
1697 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1698 (tcp_skb_is_last(sk
, skb
) ?
1699 tp
->nonagle
: TCP_NAGLE_PUSH
));
1702 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1703 * which is put after SKB on the list. It is very much like
1704 * tcp_fragment() except that it may make several kinds of assumptions
1705 * in order to speed up the splitting operation. In particular, we
1706 * know that all the data is in scatter-gather pages, and that the
1707 * packet has never been sent out before (and thus is not cloned).
1709 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1710 unsigned int mss_now
, gfp_t gfp
)
1712 struct sk_buff
*buff
;
1713 int nlen
= skb
->len
- len
;
1716 /* All of a TSO frame must be composed of paged data. */
1717 if (skb
->len
!= skb
->data_len
)
1718 return tcp_fragment(sk
, skb
, len
, mss_now
, gfp
);
1720 buff
= sk_stream_alloc_skb(sk
, 0, gfp
, true);
1721 if (unlikely(!buff
))
1724 sk
->sk_wmem_queued
+= buff
->truesize
;
1725 sk_mem_charge(sk
, buff
->truesize
);
1726 buff
->truesize
+= nlen
;
1727 skb
->truesize
-= nlen
;
1729 /* Correct the sequence numbers. */
1730 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1731 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1732 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1734 /* PSH and FIN should only be set in the second packet. */
1735 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1736 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1737 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1739 /* This packet was never sent out yet, so no SACK bits. */
1740 TCP_SKB_CB(buff
)->sacked
= 0;
1742 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1743 skb_split(skb
, buff
, len
);
1744 tcp_fragment_tstamp(skb
, buff
);
1746 /* Fix up tso_factor for both original and new SKB. */
1747 tcp_set_skb_tso_segs(skb
, mss_now
);
1748 tcp_set_skb_tso_segs(buff
, mss_now
);
1750 /* Link BUFF into the send queue. */
1751 __skb_header_release(buff
);
1752 tcp_insert_write_queue_after(skb
, buff
, sk
);
1757 /* Try to defer sending, if possible, in order to minimize the amount
1758 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1760 * This algorithm is from John Heffner.
1762 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
,
1763 bool *is_cwnd_limited
, u32 max_segs
)
1765 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1766 u32 age
, send_win
, cong_win
, limit
, in_flight
;
1767 struct tcp_sock
*tp
= tcp_sk(sk
);
1768 struct skb_mstamp now
;
1769 struct sk_buff
*head
;
1772 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1775 if (icsk
->icsk_ca_state
>= TCP_CA_Recovery
)
1778 /* Avoid bursty behavior by allowing defer
1779 * only if the last write was recent.
1781 if ((s32
)(tcp_time_stamp
- tp
->lsndtime
) > 0)
1784 in_flight
= tcp_packets_in_flight(tp
);
1786 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1788 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1790 /* From in_flight test above, we know that cwnd > in_flight. */
1791 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1793 limit
= min(send_win
, cong_win
);
1795 /* If a full-sized TSO skb can be sent, do it. */
1796 if (limit
>= max_segs
* tp
->mss_cache
)
1799 /* Middle in queue won't get any more data, full sendable already? */
1800 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1803 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1805 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1807 /* If at least some fraction of a window is available,
1810 chunk
/= win_divisor
;
1814 /* Different approach, try not to defer past a single
1815 * ACK. Receiver should ACK every other full sized
1816 * frame, so if we have space for more than 3 frames
1819 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1823 head
= tcp_write_queue_head(sk
);
1824 skb_mstamp_get(&now
);
1825 age
= skb_mstamp_us_delta(&now
, &head
->skb_mstamp
);
1826 /* If next ACK is likely to come too late (half srtt), do not defer */
1827 if (age
< (tp
->srtt_us
>> 4))
1830 /* Ok, it looks like it is advisable to defer. */
1832 if (cong_win
< send_win
&& cong_win
<= skb
->len
)
1833 *is_cwnd_limited
= true;
1841 static inline void tcp_mtu_check_reprobe(struct sock
*sk
)
1843 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1844 struct tcp_sock
*tp
= tcp_sk(sk
);
1845 struct net
*net
= sock_net(sk
);
1849 interval
= net
->ipv4
.sysctl_tcp_probe_interval
;
1850 delta
= tcp_time_stamp
- icsk
->icsk_mtup
.probe_timestamp
;
1851 if (unlikely(delta
>= interval
* HZ
)) {
1852 int mss
= tcp_current_mss(sk
);
1854 /* Update current search range */
1855 icsk
->icsk_mtup
.probe_size
= 0;
1856 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+
1857 sizeof(struct tcphdr
) +
1858 icsk
->icsk_af_ops
->net_header_len
;
1859 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
1861 /* Update probe time stamp */
1862 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1866 /* Create a new MTU probe if we are ready.
1867 * MTU probe is regularly attempting to increase the path MTU by
1868 * deliberately sending larger packets. This discovers routing
1869 * changes resulting in larger path MTUs.
1871 * Returns 0 if we should wait to probe (no cwnd available),
1872 * 1 if a probe was sent,
1875 static int tcp_mtu_probe(struct sock
*sk
)
1877 struct tcp_sock
*tp
= tcp_sk(sk
);
1878 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1879 struct sk_buff
*skb
, *nskb
, *next
;
1880 struct net
*net
= sock_net(sk
);
1888 /* Not currently probing/verifying,
1890 * have enough cwnd, and
1891 * not SACKing (the variable headers throw things off) */
1892 if (!icsk
->icsk_mtup
.enabled
||
1893 icsk
->icsk_mtup
.probe_size
||
1894 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1895 tp
->snd_cwnd
< 11 ||
1896 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1899 /* Use binary search for probe_size between tcp_mss_base,
1900 * and current mss_clamp. if (search_high - search_low)
1901 * smaller than a threshold, backoff from probing.
1903 mss_now
= tcp_current_mss(sk
);
1904 probe_size
= tcp_mtu_to_mss(sk
, (icsk
->icsk_mtup
.search_high
+
1905 icsk
->icsk_mtup
.search_low
) >> 1);
1906 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1907 interval
= icsk
->icsk_mtup
.search_high
- icsk
->icsk_mtup
.search_low
;
1908 /* When misfortune happens, we are reprobing actively,
1909 * and then reprobe timer has expired. We stick with current
1910 * probing process by not resetting search range to its orignal.
1912 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
) ||
1913 interval
< net
->ipv4
.sysctl_tcp_probe_threshold
) {
1914 /* Check whether enough time has elaplased for
1915 * another round of probing.
1917 tcp_mtu_check_reprobe(sk
);
1921 /* Have enough data in the send queue to probe? */
1922 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1925 if (tp
->snd_wnd
< size_needed
)
1927 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1930 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1931 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1932 if (!tcp_packets_in_flight(tp
))
1938 /* We're allowed to probe. Build it now. */
1939 nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
, false);
1942 sk
->sk_wmem_queued
+= nskb
->truesize
;
1943 sk_mem_charge(sk
, nskb
->truesize
);
1945 skb
= tcp_send_head(sk
);
1947 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1948 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1949 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
1950 TCP_SKB_CB(nskb
)->sacked
= 0;
1952 nskb
->ip_summed
= skb
->ip_summed
;
1954 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1957 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1958 copy
= min_t(int, skb
->len
, probe_size
- len
);
1959 if (nskb
->ip_summed
)
1960 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1962 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1963 skb_put(nskb
, copy
),
1966 if (skb
->len
<= copy
) {
1967 /* We've eaten all the data from this skb.
1969 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1970 tcp_unlink_write_queue(skb
, sk
);
1971 sk_wmem_free_skb(sk
, skb
);
1973 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
1974 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1975 if (!skb_shinfo(skb
)->nr_frags
) {
1976 skb_pull(skb
, copy
);
1977 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1978 skb
->csum
= csum_partial(skb
->data
,
1981 __pskb_trim_head(skb
, copy
);
1982 tcp_set_skb_tso_segs(skb
, mss_now
);
1984 TCP_SKB_CB(skb
)->seq
+= copy
;
1989 if (len
>= probe_size
)
1992 tcp_init_tso_segs(nskb
, nskb
->len
);
1994 /* We're ready to send. If this fails, the probe will
1995 * be resegmented into mss-sized pieces by tcp_write_xmit().
1997 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1998 /* Decrement cwnd here because we are sending
1999 * effectively two packets. */
2001 tcp_event_new_data_sent(sk
, nskb
);
2003 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
2004 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
2005 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
2013 /* This routine writes packets to the network. It advances the
2014 * send_head. This happens as incoming acks open up the remote
2017 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2018 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2019 * account rare use of URG, this is not a big flaw.
2021 * Send at most one packet when push_one > 0. Temporarily ignore
2022 * cwnd limit to force at most one packet out when push_one == 2.
2024 * Returns true, if no segments are in flight and we have queued segments,
2025 * but cannot send anything now because of SWS or another problem.
2027 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
2028 int push_one
, gfp_t gfp
)
2030 struct tcp_sock
*tp
= tcp_sk(sk
);
2031 struct sk_buff
*skb
;
2032 unsigned int tso_segs
, sent_pkts
;
2035 bool is_cwnd_limited
= false;
2041 /* Do MTU probing. */
2042 result
= tcp_mtu_probe(sk
);
2045 } else if (result
> 0) {
2050 max_segs
= tcp_tso_autosize(sk
, mss_now
);
2051 while ((skb
= tcp_send_head(sk
))) {
2054 tso_segs
= tcp_init_tso_segs(skb
, mss_now
);
2057 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
) {
2058 /* "skb_mstamp" is used as a start point for the retransmit timer */
2059 skb_mstamp_get(&skb
->skb_mstamp
);
2060 goto repair
; /* Skip network transmission */
2063 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2066 /* Force out a loss probe pkt. */
2072 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
2075 if (tso_segs
== 1) {
2076 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2077 (tcp_skb_is_last(sk
, skb
) ?
2078 nonagle
: TCP_NAGLE_PUSH
))))
2082 tcp_tso_should_defer(sk
, skb
, &is_cwnd_limited
,
2088 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
2089 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2095 if (skb
->len
> limit
&&
2096 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
2099 /* TCP Small Queues :
2100 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2102 * - better RTT estimation and ACK scheduling
2105 * Alas, some drivers / subsystems require a fair amount
2106 * of queued bytes to ensure line rate.
2107 * One example is wifi aggregation (802.11 AMPDU)
2109 limit
= max(2 * skb
->truesize
, sk
->sk_pacing_rate
>> 10);
2110 limit
= min_t(u32
, limit
, sysctl_tcp_limit_output_bytes
);
2112 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
) {
2113 set_bit(TSQ_THROTTLED
, &tp
->tsq_flags
);
2114 /* It is possible TX completion already happened
2115 * before we set TSQ_THROTTLED, so we must
2116 * test again the condition.
2118 smp_mb__after_atomic();
2119 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
)
2123 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2127 /* Advance the send_head. This one is sent out.
2128 * This call will increment packets_out.
2130 tcp_event_new_data_sent(sk
, skb
);
2132 tcp_minshall_update(tp
, mss_now
, skb
);
2133 sent_pkts
+= tcp_skb_pcount(skb
);
2139 if (likely(sent_pkts
)) {
2140 if (tcp_in_cwnd_reduction(sk
))
2141 tp
->prr_out
+= sent_pkts
;
2143 /* Send one loss probe per tail loss episode. */
2145 tcp_schedule_loss_probe(sk
);
2146 is_cwnd_limited
|= (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
);
2147 tcp_cwnd_validate(sk
, is_cwnd_limited
);
2150 return !tp
->packets_out
&& tcp_send_head(sk
);
2153 bool tcp_schedule_loss_probe(struct sock
*sk
)
2155 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2156 struct tcp_sock
*tp
= tcp_sk(sk
);
2157 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
2158 u32 rtt
= usecs_to_jiffies(tp
->srtt_us
>> 3);
2160 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
))
2162 /* No consecutive loss probes. */
2163 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
2167 /* Don't do any loss probe on a Fast Open connection before 3WHS
2170 if (tp
->fastopen_rsk
)
2173 /* TLP is only scheduled when next timer event is RTO. */
2174 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
2177 /* Schedule a loss probe in 2*RTT for SACK capable connections
2178 * in Open state, that are either limited by cwnd or application.
2180 if (sysctl_tcp_early_retrans
< 3 || !tp
->packets_out
||
2181 !tcp_is_sack(tp
) || inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
2184 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
2188 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2189 * for delayed ack when there's one outstanding packet. If no RTT
2190 * sample is available then probe after TCP_TIMEOUT_INIT.
2192 timeout
= rtt
<< 1 ? : TCP_TIMEOUT_INIT
;
2193 if (tp
->packets_out
== 1)
2194 timeout
= max_t(u32
, timeout
,
2195 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
2196 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
2198 /* If RTO is shorter, just schedule TLP in its place. */
2199 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
2200 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
2201 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
2202 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
2207 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2212 /* Thanks to skb fast clones, we can detect if a prior transmit of
2213 * a packet is still in a qdisc or driver queue.
2214 * In this case, there is very little point doing a retransmit !
2215 * Note: This is called from BH context only.
2217 static bool skb_still_in_host_queue(const struct sock
*sk
,
2218 const struct sk_buff
*skb
)
2220 if (unlikely(skb_fclone_busy(sk
, skb
))) {
2221 __NET_INC_STATS(sock_net(sk
),
2222 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
2228 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2229 * retransmit the last segment.
2231 void tcp_send_loss_probe(struct sock
*sk
)
2233 struct tcp_sock
*tp
= tcp_sk(sk
);
2234 struct sk_buff
*skb
;
2236 int mss
= tcp_current_mss(sk
);
2238 skb
= tcp_send_head(sk
);
2240 if (tcp_snd_wnd_test(tp
, skb
, mss
)) {
2241 pcount
= tp
->packets_out
;
2242 tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2243 if (tp
->packets_out
> pcount
)
2247 skb
= tcp_write_queue_prev(sk
, skb
);
2249 skb
= tcp_write_queue_tail(sk
);
2252 /* At most one outstanding TLP retransmission. */
2253 if (tp
->tlp_high_seq
)
2256 /* Retransmit last segment. */
2260 if (skb_still_in_host_queue(sk
, skb
))
2263 pcount
= tcp_skb_pcount(skb
);
2264 if (WARN_ON(!pcount
))
2267 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2268 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
,
2271 skb
= tcp_write_queue_next(sk
, skb
);
2274 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2277 if (__tcp_retransmit_skb(sk
, skb
, 1))
2280 /* Record snd_nxt for loss detection. */
2281 tp
->tlp_high_seq
= tp
->snd_nxt
;
2284 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSPROBES
);
2285 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2286 inet_csk(sk
)->icsk_pending
= 0;
2291 /* Push out any pending frames which were held back due to
2292 * TCP_CORK or attempt at coalescing tiny packets.
2293 * The socket must be locked by the caller.
2295 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2298 /* If we are closed, the bytes will have to remain here.
2299 * In time closedown will finish, we empty the write queue and
2300 * all will be happy.
2302 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2305 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2306 sk_gfp_mask(sk
, GFP_ATOMIC
)))
2307 tcp_check_probe_timer(sk
);
2310 /* Send _single_ skb sitting at the send head. This function requires
2311 * true push pending frames to setup probe timer etc.
2313 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2315 struct sk_buff
*skb
= tcp_send_head(sk
);
2317 BUG_ON(!skb
|| skb
->len
< mss_now
);
2319 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2322 /* This function returns the amount that we can raise the
2323 * usable window based on the following constraints
2325 * 1. The window can never be shrunk once it is offered (RFC 793)
2326 * 2. We limit memory per socket
2329 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2330 * RECV.NEXT + RCV.WIN fixed until:
2331 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2333 * i.e. don't raise the right edge of the window until you can raise
2334 * it at least MSS bytes.
2336 * Unfortunately, the recommended algorithm breaks header prediction,
2337 * since header prediction assumes th->window stays fixed.
2339 * Strictly speaking, keeping th->window fixed violates the receiver
2340 * side SWS prevention criteria. The problem is that under this rule
2341 * a stream of single byte packets will cause the right side of the
2342 * window to always advance by a single byte.
2344 * Of course, if the sender implements sender side SWS prevention
2345 * then this will not be a problem.
2347 * BSD seems to make the following compromise:
2349 * If the free space is less than the 1/4 of the maximum
2350 * space available and the free space is less than 1/2 mss,
2351 * then set the window to 0.
2352 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2353 * Otherwise, just prevent the window from shrinking
2354 * and from being larger than the largest representable value.
2356 * This prevents incremental opening of the window in the regime
2357 * where TCP is limited by the speed of the reader side taking
2358 * data out of the TCP receive queue. It does nothing about
2359 * those cases where the window is constrained on the sender side
2360 * because the pipeline is full.
2362 * BSD also seems to "accidentally" limit itself to windows that are a
2363 * multiple of MSS, at least until the free space gets quite small.
2364 * This would appear to be a side effect of the mbuf implementation.
2365 * Combining these two algorithms results in the observed behavior
2366 * of having a fixed window size at almost all times.
2368 * Below we obtain similar behavior by forcing the offered window to
2369 * a multiple of the mss when it is feasible to do so.
2371 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2372 * Regular options like TIMESTAMP are taken into account.
2374 u32
__tcp_select_window(struct sock
*sk
)
2376 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2377 struct tcp_sock
*tp
= tcp_sk(sk
);
2378 /* MSS for the peer's data. Previous versions used mss_clamp
2379 * here. I don't know if the value based on our guesses
2380 * of peer's MSS is better for the performance. It's more correct
2381 * but may be worse for the performance because of rcv_mss
2382 * fluctuations. --SAW 1998/11/1
2384 int mss
= icsk
->icsk_ack
.rcv_mss
;
2385 int free_space
= tcp_space(sk
);
2386 int allowed_space
= tcp_full_space(sk
);
2387 int full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2390 if (mss
> full_space
)
2393 if (free_space
< (full_space
>> 1)) {
2394 icsk
->icsk_ack
.quick
= 0;
2396 if (tcp_under_memory_pressure(sk
))
2397 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2400 /* free_space might become our new window, make sure we don't
2401 * increase it due to wscale.
2403 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2405 /* if free space is less than mss estimate, or is below 1/16th
2406 * of the maximum allowed, try to move to zero-window, else
2407 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2408 * new incoming data is dropped due to memory limits.
2409 * With large window, mss test triggers way too late in order
2410 * to announce zero window in time before rmem limit kicks in.
2412 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2416 if (free_space
> tp
->rcv_ssthresh
)
2417 free_space
= tp
->rcv_ssthresh
;
2419 /* Don't do rounding if we are using window scaling, since the
2420 * scaled window will not line up with the MSS boundary anyway.
2422 window
= tp
->rcv_wnd
;
2423 if (tp
->rx_opt
.rcv_wscale
) {
2424 window
= free_space
;
2426 /* Advertise enough space so that it won't get scaled away.
2427 * Import case: prevent zero window announcement if
2428 * 1<<rcv_wscale > mss.
2430 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2431 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2432 << tp
->rx_opt
.rcv_wscale
);
2434 /* Get the largest window that is a nice multiple of mss.
2435 * Window clamp already applied above.
2436 * If our current window offering is within 1 mss of the
2437 * free space we just keep it. This prevents the divide
2438 * and multiply from happening most of the time.
2439 * We also don't do any window rounding when the free space
2442 if (window
<= free_space
- mss
|| window
> free_space
)
2443 window
= (free_space
/ mss
) * mss
;
2444 else if (mss
== full_space
&&
2445 free_space
> window
+ (full_space
>> 1))
2446 window
= free_space
;
2452 void tcp_skb_collapse_tstamp(struct sk_buff
*skb
,
2453 const struct sk_buff
*next_skb
)
2455 if (unlikely(tcp_has_tx_tstamp(next_skb
))) {
2456 const struct skb_shared_info
*next_shinfo
=
2457 skb_shinfo(next_skb
);
2458 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
2460 shinfo
->tx_flags
|= next_shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
2461 shinfo
->tskey
= next_shinfo
->tskey
;
2462 TCP_SKB_CB(skb
)->txstamp_ack
|=
2463 TCP_SKB_CB(next_skb
)->txstamp_ack
;
2467 /* Collapses two adjacent SKB's during retransmission. */
2468 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2470 struct tcp_sock
*tp
= tcp_sk(sk
);
2471 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2472 int skb_size
, next_skb_size
;
2474 skb_size
= skb
->len
;
2475 next_skb_size
= next_skb
->len
;
2477 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2479 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2481 tcp_unlink_write_queue(next_skb
, sk
);
2483 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
2486 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2487 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2489 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2490 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2492 /* Update sequence range on original skb. */
2493 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2495 /* Merge over control information. This moves PSH/FIN etc. over */
2496 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2498 /* All done, get rid of second SKB and account for it so
2499 * packet counting does not break.
2501 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2502 TCP_SKB_CB(skb
)->eor
= TCP_SKB_CB(next_skb
)->eor
;
2504 /* changed transmit queue under us so clear hints */
2505 tcp_clear_retrans_hints_partial(tp
);
2506 if (next_skb
== tp
->retransmit_skb_hint
)
2507 tp
->retransmit_skb_hint
= skb
;
2509 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2511 tcp_skb_collapse_tstamp(skb
, next_skb
);
2513 sk_wmem_free_skb(sk
, next_skb
);
2516 /* Check if coalescing SKBs is legal. */
2517 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2519 if (tcp_skb_pcount(skb
) > 1)
2521 /* TODO: SACK collapsing could be used to remove this condition */
2522 if (skb_shinfo(skb
)->nr_frags
!= 0)
2524 if (skb_cloned(skb
))
2526 if (skb
== tcp_send_head(sk
))
2528 /* Some heurestics for collapsing over SACK'd could be invented */
2529 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2535 /* Collapse packets in the retransmit queue to make to create
2536 * less packets on the wire. This is only done on retransmission.
2538 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2541 struct tcp_sock
*tp
= tcp_sk(sk
);
2542 struct sk_buff
*skb
= to
, *tmp
;
2545 if (!sysctl_tcp_retrans_collapse
)
2547 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2550 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2551 if (!tcp_can_collapse(sk
, skb
))
2554 if (!tcp_skb_can_collapse_to(to
))
2566 /* Punt if not enough space exists in the first SKB for
2567 * the data in the second
2569 if (skb
->len
> skb_availroom(to
))
2572 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2575 tcp_collapse_retrans(sk
, to
);
2579 /* This retransmits one SKB. Policy decisions and retransmit queue
2580 * state updates are done by the caller. Returns non-zero if an
2581 * error occurred which prevented the send.
2583 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2585 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2586 struct tcp_sock
*tp
= tcp_sk(sk
);
2587 unsigned int cur_mss
;
2591 /* Inconclusive MTU probe */
2592 if (icsk
->icsk_mtup
.probe_size
)
2593 icsk
->icsk_mtup
.probe_size
= 0;
2595 /* Do not sent more than we queued. 1/4 is reserved for possible
2596 * copying overhead: fragmentation, tunneling, mangling etc.
2598 if (atomic_read(&sk
->sk_wmem_alloc
) >
2599 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2602 if (skb_still_in_host_queue(sk
, skb
))
2605 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2606 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2608 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2612 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2613 return -EHOSTUNREACH
; /* Routing failure or similar. */
2615 cur_mss
= tcp_current_mss(sk
);
2617 /* If receiver has shrunk his window, and skb is out of
2618 * new window, do not retransmit it. The exception is the
2619 * case, when window is shrunk to zero. In this case
2620 * our retransmit serves as a zero window probe.
2622 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2623 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2626 len
= cur_mss
* segs
;
2627 if (skb
->len
> len
) {
2628 if (tcp_fragment(sk
, skb
, len
, cur_mss
, GFP_ATOMIC
))
2629 return -ENOMEM
; /* We'll try again later. */
2631 if (skb_unclone(skb
, GFP_ATOMIC
))
2634 diff
= tcp_skb_pcount(skb
);
2635 tcp_set_skb_tso_segs(skb
, cur_mss
);
2636 diff
-= tcp_skb_pcount(skb
);
2638 tcp_adjust_pcount(sk
, skb
, diff
);
2639 if (skb
->len
< cur_mss
)
2640 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2643 /* RFC3168, section 6.1.1.1. ECN fallback */
2644 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN_ECN
) == TCPHDR_SYN_ECN
)
2645 tcp_ecn_clear_syn(sk
, skb
);
2647 /* make sure skb->data is aligned on arches that require it
2648 * and check if ack-trimming & collapsing extended the headroom
2649 * beyond what csum_start can cover.
2651 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2652 skb_headroom(skb
) >= 0xFFFF)) {
2653 struct sk_buff
*nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
,
2655 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2658 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2662 segs
= tcp_skb_pcount(skb
);
2664 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
2665 /* Update global TCP statistics. */
2666 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
, segs
);
2667 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2668 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
2669 tp
->total_retrans
+= segs
;
2674 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2676 struct tcp_sock
*tp
= tcp_sk(sk
);
2677 int err
= __tcp_retransmit_skb(sk
, skb
, segs
);
2680 #if FASTRETRANS_DEBUG > 0
2681 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2682 net_dbg_ratelimited("retrans_out leaked\n");
2685 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2686 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2688 /* Save stamp of the first retransmit. */
2689 if (!tp
->retrans_stamp
)
2690 tp
->retrans_stamp
= tcp_skb_timestamp(skb
);
2692 } else if (err
!= -EBUSY
) {
2693 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2696 if (tp
->undo_retrans
< 0)
2697 tp
->undo_retrans
= 0;
2698 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2702 /* Check if we forward retransmits are possible in the current
2703 * window/congestion state.
2705 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2707 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2708 const struct tcp_sock
*tp
= tcp_sk(sk
);
2710 /* Forward retransmissions are possible only during Recovery. */
2711 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2714 /* No forward retransmissions in Reno are possible. */
2715 if (tcp_is_reno(tp
))
2718 /* Yeah, we have to make difficult choice between forward transmission
2719 * and retransmission... Both ways have their merits...
2721 * For now we do not retransmit anything, while we have some new
2722 * segments to send. In the other cases, follow rule 3 for
2723 * NextSeg() specified in RFC3517.
2726 if (tcp_may_send_now(sk
))
2732 /* This gets called after a retransmit timeout, and the initially
2733 * retransmitted data is acknowledged. It tries to continue
2734 * resending the rest of the retransmit queue, until either
2735 * we've sent it all or the congestion window limit is reached.
2736 * If doing SACK, the first ACK which comes back for a timeout
2737 * based retransmit packet might feed us FACK information again.
2738 * If so, we use it to avoid unnecessarily retransmissions.
2740 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2742 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2743 struct tcp_sock
*tp
= tcp_sk(sk
);
2744 struct sk_buff
*skb
;
2745 struct sk_buff
*hole
= NULL
;
2748 int fwd_rexmitting
= 0;
2750 if (!tp
->packets_out
)
2754 tp
->retransmit_high
= tp
->snd_una
;
2756 if (tp
->retransmit_skb_hint
) {
2757 skb
= tp
->retransmit_skb_hint
;
2758 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2759 if (after(last_lost
, tp
->retransmit_high
))
2760 last_lost
= tp
->retransmit_high
;
2762 skb
= tcp_write_queue_head(sk
);
2763 last_lost
= tp
->snd_una
;
2766 tcp_for_write_queue_from(skb
, sk
) {
2767 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2770 if (skb
== tcp_send_head(sk
))
2772 /* we could do better than to assign each time */
2774 tp
->retransmit_skb_hint
= skb
;
2776 segs
= tp
->snd_cwnd
- tcp_packets_in_flight(tp
);
2780 if (fwd_rexmitting
) {
2782 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2784 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2786 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2787 tp
->retransmit_high
= last_lost
;
2788 if (!tcp_can_forward_retransmit(sk
))
2790 /* Backtrack if necessary to non-L'ed skb */
2798 } else if (!(sacked
& TCPCB_LOST
)) {
2799 if (!hole
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2804 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2805 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2806 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2808 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2811 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2814 if (tcp_retransmit_skb(sk
, skb
, segs
))
2817 __NET_INC_STATS(sock_net(sk
), mib_idx
);
2819 if (tcp_in_cwnd_reduction(sk
))
2820 tp
->prr_out
+= tcp_skb_pcount(skb
);
2822 if (skb
== tcp_write_queue_head(sk
))
2823 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2824 inet_csk(sk
)->icsk_rto
,
2829 /* We allow to exceed memory limits for FIN packets to expedite
2830 * connection tear down and (memory) recovery.
2831 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2832 * or even be forced to close flow without any FIN.
2833 * In general, we want to allow one skb per socket to avoid hangs
2834 * with edge trigger epoll()
2836 void sk_forced_mem_schedule(struct sock
*sk
, int size
)
2840 if (size
<= sk
->sk_forward_alloc
)
2842 amt
= sk_mem_pages(size
);
2843 sk
->sk_forward_alloc
+= amt
* SK_MEM_QUANTUM
;
2844 sk_memory_allocated_add(sk
, amt
);
2846 if (mem_cgroup_sockets_enabled
&& sk
->sk_memcg
)
2847 mem_cgroup_charge_skmem(sk
->sk_memcg
, amt
);
2850 /* Send a FIN. The caller locks the socket for us.
2851 * We should try to send a FIN packet really hard, but eventually give up.
2853 void tcp_send_fin(struct sock
*sk
)
2855 struct sk_buff
*skb
, *tskb
= tcp_write_queue_tail(sk
);
2856 struct tcp_sock
*tp
= tcp_sk(sk
);
2858 /* Optimization, tack on the FIN if we have one skb in write queue and
2859 * this skb was not yet sent, or we are under memory pressure.
2860 * Note: in the latter case, FIN packet will be sent after a timeout,
2861 * as TCP stack thinks it has already been transmitted.
2863 if (tskb
&& (tcp_send_head(sk
) || tcp_under_memory_pressure(sk
))) {
2865 TCP_SKB_CB(tskb
)->tcp_flags
|= TCPHDR_FIN
;
2866 TCP_SKB_CB(tskb
)->end_seq
++;
2868 if (!tcp_send_head(sk
)) {
2869 /* This means tskb was already sent.
2870 * Pretend we included the FIN on previous transmit.
2871 * We need to set tp->snd_nxt to the value it would have
2872 * if FIN had been sent. This is because retransmit path
2873 * does not change tp->snd_nxt.
2879 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, sk
->sk_allocation
);
2880 if (unlikely(!skb
)) {
2885 skb_reserve(skb
, MAX_TCP_HEADER
);
2886 sk_forced_mem_schedule(sk
, skb
->truesize
);
2887 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2888 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2889 TCPHDR_ACK
| TCPHDR_FIN
);
2890 tcp_queue_skb(sk
, skb
);
2892 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
), TCP_NAGLE_OFF
);
2895 /* We get here when a process closes a file descriptor (either due to
2896 * an explicit close() or as a byproduct of exit()'ing) and there
2897 * was unread data in the receive queue. This behavior is recommended
2898 * by RFC 2525, section 2.17. -DaveM
2900 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2902 struct sk_buff
*skb
;
2904 /* NOTE: No TCP options attached and we never retransmit this. */
2905 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2907 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2911 /* Reserve space for headers and prepare control bits. */
2912 skb_reserve(skb
, MAX_TCP_HEADER
);
2913 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2914 TCPHDR_ACK
| TCPHDR_RST
);
2915 skb_mstamp_get(&skb
->skb_mstamp
);
2917 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2918 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2920 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2923 /* Send a crossed SYN-ACK during socket establishment.
2924 * WARNING: This routine must only be called when we have already sent
2925 * a SYN packet that crossed the incoming SYN that caused this routine
2926 * to get called. If this assumption fails then the initial rcv_wnd
2927 * and rcv_wscale values will not be correct.
2929 int tcp_send_synack(struct sock
*sk
)
2931 struct sk_buff
*skb
;
2933 skb
= tcp_write_queue_head(sk
);
2934 if (!skb
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2935 pr_debug("%s: wrong queue state\n", __func__
);
2938 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
2939 if (skb_cloned(skb
)) {
2940 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2943 tcp_unlink_write_queue(skb
, sk
);
2944 __skb_header_release(nskb
);
2945 __tcp_add_write_queue_head(sk
, nskb
);
2946 sk_wmem_free_skb(sk
, skb
);
2947 sk
->sk_wmem_queued
+= nskb
->truesize
;
2948 sk_mem_charge(sk
, nskb
->truesize
);
2952 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
2953 tcp_ecn_send_synack(sk
, skb
);
2955 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2959 * tcp_make_synack - Prepare a SYN-ACK.
2960 * sk: listener socket
2961 * dst: dst entry attached to the SYNACK
2962 * req: request_sock pointer
2964 * Allocate one skb and build a SYNACK packet.
2965 * @dst is consumed : Caller should not use it again.
2967 struct sk_buff
*tcp_make_synack(const struct sock
*sk
, struct dst_entry
*dst
,
2968 struct request_sock
*req
,
2969 struct tcp_fastopen_cookie
*foc
,
2970 enum tcp_synack_type synack_type
)
2972 struct inet_request_sock
*ireq
= inet_rsk(req
);
2973 const struct tcp_sock
*tp
= tcp_sk(sk
);
2974 struct tcp_md5sig_key
*md5
= NULL
;
2975 struct tcp_out_options opts
;
2976 struct sk_buff
*skb
;
2977 int tcp_header_size
;
2982 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2983 if (unlikely(!skb
)) {
2987 /* Reserve space for headers. */
2988 skb_reserve(skb
, MAX_TCP_HEADER
);
2990 switch (synack_type
) {
2991 case TCP_SYNACK_NORMAL
:
2992 skb_set_owner_w(skb
, req_to_sk(req
));
2994 case TCP_SYNACK_COOKIE
:
2995 /* Under synflood, we do not attach skb to a socket,
2996 * to avoid false sharing.
2999 case TCP_SYNACK_FASTOPEN
:
3000 /* sk is a const pointer, because we want to express multiple
3001 * cpu might call us concurrently.
3002 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3004 skb_set_owner_w(skb
, (struct sock
*)sk
);
3007 skb_dst_set(skb
, dst
);
3009 mss
= dst_metric_advmss(dst
);
3010 user_mss
= READ_ONCE(tp
->rx_opt
.user_mss
);
3011 if (user_mss
&& user_mss
< mss
)
3014 memset(&opts
, 0, sizeof(opts
));
3015 #ifdef CONFIG_SYN_COOKIES
3016 if (unlikely(req
->cookie_ts
))
3017 skb
->skb_mstamp
.stamp_jiffies
= cookie_init_timestamp(req
);
3020 skb_mstamp_get(&skb
->skb_mstamp
);
3022 #ifdef CONFIG_TCP_MD5SIG
3024 md5
= tcp_rsk(req
)->af_specific
->req_md5_lookup(sk
, req_to_sk(req
));
3026 skb_set_hash(skb
, tcp_rsk(req
)->txhash
, PKT_HASH_TYPE_L4
);
3027 tcp_header_size
= tcp_synack_options(req
, mss
, skb
, &opts
, md5
, foc
) +
3030 skb_push(skb
, tcp_header_size
);
3031 skb_reset_transport_header(skb
);
3034 memset(th
, 0, sizeof(struct tcphdr
));
3037 tcp_ecn_make_synack(req
, th
);
3038 th
->source
= htons(ireq
->ir_num
);
3039 th
->dest
= ireq
->ir_rmt_port
;
3040 /* Setting of flags are superfluous here for callers (and ECE is
3041 * not even correctly set)
3043 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
3044 TCPHDR_SYN
| TCPHDR_ACK
);
3046 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
3047 /* XXX data is queued and acked as is. No buffer/window check */
3048 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
3050 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3051 th
->window
= htons(min(req
->rsk_rcv_wnd
, 65535U));
3052 tcp_options_write((__be32
*)(th
+ 1), NULL
, &opts
);
3053 th
->doff
= (tcp_header_size
>> 2);
3054 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
3056 #ifdef CONFIG_TCP_MD5SIG
3057 /* Okay, we have all we need - do the md5 hash if needed */
3059 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
3060 md5
, req_to_sk(req
), skb
);
3064 /* Do not fool tcpdump (if any), clean our debris */
3065 skb
->tstamp
.tv64
= 0;
3068 EXPORT_SYMBOL(tcp_make_synack
);
3070 static void tcp_ca_dst_init(struct sock
*sk
, const struct dst_entry
*dst
)
3072 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3073 const struct tcp_congestion_ops
*ca
;
3074 u32 ca_key
= dst_metric(dst
, RTAX_CC_ALGO
);
3076 if (ca_key
== TCP_CA_UNSPEC
)
3080 ca
= tcp_ca_find_key(ca_key
);
3081 if (likely(ca
&& try_module_get(ca
->owner
))) {
3082 module_put(icsk
->icsk_ca_ops
->owner
);
3083 icsk
->icsk_ca_dst_locked
= tcp_ca_dst_locked(dst
);
3084 icsk
->icsk_ca_ops
= ca
;
3089 /* Do all connect socket setups that can be done AF independent. */
3090 static void tcp_connect_init(struct sock
*sk
)
3092 const struct dst_entry
*dst
= __sk_dst_get(sk
);
3093 struct tcp_sock
*tp
= tcp_sk(sk
);
3096 /* We'll fix this up when we get a response from the other end.
3097 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3099 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
3100 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
3102 #ifdef CONFIG_TCP_MD5SIG
3103 if (tp
->af_specific
->md5_lookup(sk
, sk
))
3104 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
3107 /* If user gave his TCP_MAXSEG, record it to clamp */
3108 if (tp
->rx_opt
.user_mss
)
3109 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3112 tcp_sync_mss(sk
, dst_mtu(dst
));
3114 tcp_ca_dst_init(sk
, dst
);
3116 if (!tp
->window_clamp
)
3117 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
3118 tp
->advmss
= dst_metric_advmss(dst
);
3119 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
3120 tp
->advmss
= tp
->rx_opt
.user_mss
;
3122 tcp_initialize_rcv_mss(sk
);
3124 /* limit the window selection if the user enforce a smaller rx buffer */
3125 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
3126 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
3127 tp
->window_clamp
= tcp_full_space(sk
);
3129 tcp_select_initial_window(tcp_full_space(sk
),
3130 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
3133 sysctl_tcp_window_scaling
,
3135 dst_metric(dst
, RTAX_INITRWND
));
3137 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
3138 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
3141 sock_reset_flag(sk
, SOCK_DONE
);
3144 tp
->snd_una
= tp
->write_seq
;
3145 tp
->snd_sml
= tp
->write_seq
;
3146 tp
->snd_up
= tp
->write_seq
;
3147 tp
->snd_nxt
= tp
->write_seq
;
3149 if (likely(!tp
->repair
))
3152 tp
->rcv_tstamp
= tcp_time_stamp
;
3153 tp
->rcv_wup
= tp
->rcv_nxt
;
3154 tp
->copied_seq
= tp
->rcv_nxt
;
3156 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
3157 inet_csk(sk
)->icsk_retransmits
= 0;
3158 tcp_clear_retrans(tp
);
3161 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
3163 struct tcp_sock
*tp
= tcp_sk(sk
);
3164 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
3166 tcb
->end_seq
+= skb
->len
;
3167 __skb_header_release(skb
);
3168 __tcp_add_write_queue_tail(sk
, skb
);
3169 sk
->sk_wmem_queued
+= skb
->truesize
;
3170 sk_mem_charge(sk
, skb
->truesize
);
3171 tp
->write_seq
= tcb
->end_seq
;
3172 tp
->packets_out
+= tcp_skb_pcount(skb
);
3175 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3176 * queue a data-only packet after the regular SYN, such that regular SYNs
3177 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3178 * only the SYN sequence, the data are retransmitted in the first ACK.
3179 * If cookie is not cached or other error occurs, falls back to send a
3180 * regular SYN with Fast Open cookie request option.
3182 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
3184 struct tcp_sock
*tp
= tcp_sk(sk
);
3185 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
3186 int syn_loss
= 0, space
, err
= 0;
3187 unsigned long last_syn_loss
= 0;
3188 struct sk_buff
*syn_data
;
3190 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
3191 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
3192 &syn_loss
, &last_syn_loss
);
3193 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3195 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
3196 fo
->cookie
.len
= -1;
3200 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
3201 fo
->cookie
.len
= -1;
3202 else if (fo
->cookie
.len
<= 0)
3205 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3206 * user-MSS. Reserve maximum option space for middleboxes that add
3207 * private TCP options. The cost is reduced data space in SYN :(
3209 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
3210 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3211 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
3212 MAX_TCP_OPTION_SPACE
;
3214 space
= min_t(size_t, space
, fo
->size
);
3216 /* limit to order-0 allocations */
3217 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
3219 syn_data
= sk_stream_alloc_skb(sk
, space
, sk
->sk_allocation
, false);
3222 syn_data
->ip_summed
= CHECKSUM_PARTIAL
;
3223 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
3225 int copied
= copy_from_iter(skb_put(syn_data
, space
), space
,
3226 &fo
->data
->msg_iter
);
3227 if (unlikely(!copied
)) {
3228 kfree_skb(syn_data
);
3231 if (copied
!= space
) {
3232 skb_trim(syn_data
, copied
);
3236 /* No more data pending in inet_wait_for_connect() */
3237 if (space
== fo
->size
)
3241 tcp_connect_queue_skb(sk
, syn_data
);
3243 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
3245 syn
->skb_mstamp
= syn_data
->skb_mstamp
;
3247 /* Now full SYN+DATA was cloned and sent (or not),
3248 * remove the SYN from the original skb (syn_data)
3249 * we keep in write queue in case of a retransmit, as we
3250 * also have the SYN packet (with no data) in the same queue.
3252 TCP_SKB_CB(syn_data
)->seq
++;
3253 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
3255 tp
->syn_data
= (fo
->copied
> 0);
3256 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3261 /* Send a regular SYN with Fast Open cookie request option */
3262 if (fo
->cookie
.len
> 0)
3264 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3266 tp
->syn_fastopen
= 0;
3268 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3272 /* Build a SYN and send it off. */
3273 int tcp_connect(struct sock
*sk
)
3275 struct tcp_sock
*tp
= tcp_sk(sk
);
3276 struct sk_buff
*buff
;
3279 tcp_connect_init(sk
);
3281 if (unlikely(tp
->repair
)) {
3282 tcp_finish_connect(sk
, NULL
);
3286 buff
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
, true);
3287 if (unlikely(!buff
))
3290 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3291 tp
->retrans_stamp
= tcp_time_stamp
;
3292 tcp_connect_queue_skb(sk
, buff
);
3293 tcp_ecn_send_syn(sk
, buff
);
3295 /* Send off SYN; include data in Fast Open. */
3296 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3297 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3298 if (err
== -ECONNREFUSED
)
3301 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3302 * in order to make this packet get counted in tcpOutSegs.
3304 tp
->snd_nxt
= tp
->write_seq
;
3305 tp
->pushed_seq
= tp
->write_seq
;
3306 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3308 /* Timer for repeating the SYN until an answer. */
3309 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3310 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3313 EXPORT_SYMBOL(tcp_connect
);
3315 /* Send out a delayed ack, the caller does the policy checking
3316 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3319 void tcp_send_delayed_ack(struct sock
*sk
)
3321 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3322 int ato
= icsk
->icsk_ack
.ato
;
3323 unsigned long timeout
;
3325 tcp_ca_event(sk
, CA_EVENT_DELAYED_ACK
);
3327 if (ato
> TCP_DELACK_MIN
) {
3328 const struct tcp_sock
*tp
= tcp_sk(sk
);
3329 int max_ato
= HZ
/ 2;
3331 if (icsk
->icsk_ack
.pingpong
||
3332 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3333 max_ato
= TCP_DELACK_MAX
;
3335 /* Slow path, intersegment interval is "high". */
3337 /* If some rtt estimate is known, use it to bound delayed ack.
3338 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3342 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3349 ato
= min(ato
, max_ato
);
3352 /* Stay within the limit we were given */
3353 timeout
= jiffies
+ ato
;
3355 /* Use new timeout only if there wasn't a older one earlier. */
3356 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3357 /* If delack timer was blocked or is about to expire,
3360 if (icsk
->icsk_ack
.blocked
||
3361 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3366 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3367 timeout
= icsk
->icsk_ack
.timeout
;
3369 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3370 icsk
->icsk_ack
.timeout
= timeout
;
3371 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3374 /* This routine sends an ack and also updates the window. */
3375 void tcp_send_ack(struct sock
*sk
)
3377 struct sk_buff
*buff
;
3379 /* If we have been reset, we may not send again. */
3380 if (sk
->sk_state
== TCP_CLOSE
)
3383 tcp_ca_event(sk
, CA_EVENT_NON_DELAYED_ACK
);
3385 /* We are not putting this on the write queue, so
3386 * tcp_transmit_skb() will set the ownership to this
3389 buff
= alloc_skb(MAX_TCP_HEADER
,
3390 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3391 if (unlikely(!buff
)) {
3392 inet_csk_schedule_ack(sk
);
3393 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3394 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3395 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3399 /* Reserve space for headers and prepare control bits. */
3400 skb_reserve(buff
, MAX_TCP_HEADER
);
3401 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3403 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3405 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3406 * We also avoid tcp_wfree() overhead (cache line miss accessing
3407 * tp->tsq_flags) by using regular sock_wfree()
3409 skb_set_tcp_pure_ack(buff
);
3411 /* Send it off, this clears delayed acks for us. */
3412 skb_mstamp_get(&buff
->skb_mstamp
);
3413 tcp_transmit_skb(sk
, buff
, 0, (__force gfp_t
)0);
3415 EXPORT_SYMBOL_GPL(tcp_send_ack
);
3417 /* This routine sends a packet with an out of date sequence
3418 * number. It assumes the other end will try to ack it.
3420 * Question: what should we make while urgent mode?
3421 * 4.4BSD forces sending single byte of data. We cannot send
3422 * out of window data, because we have SND.NXT==SND.MAX...
3424 * Current solution: to send TWO zero-length segments in urgent mode:
3425 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3426 * out-of-date with SND.UNA-1 to probe window.
3428 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
, int mib
)
3430 struct tcp_sock
*tp
= tcp_sk(sk
);
3431 struct sk_buff
*skb
;
3433 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3434 skb
= alloc_skb(MAX_TCP_HEADER
,
3435 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3439 /* Reserve space for headers and set control bits. */
3440 skb_reserve(skb
, MAX_TCP_HEADER
);
3441 /* Use a previous sequence. This should cause the other
3442 * end to send an ack. Don't queue or clone SKB, just
3445 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3446 skb_mstamp_get(&skb
->skb_mstamp
);
3447 NET_INC_STATS(sock_net(sk
), mib
);
3448 return tcp_transmit_skb(sk
, skb
, 0, (__force gfp_t
)0);
3451 void tcp_send_window_probe(struct sock
*sk
)
3453 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3454 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3455 tcp_xmit_probe_skb(sk
, 0, LINUX_MIB_TCPWINPROBE
);
3459 /* Initiate keepalive or window probe from timer. */
3460 int tcp_write_wakeup(struct sock
*sk
, int mib
)
3462 struct tcp_sock
*tp
= tcp_sk(sk
);
3463 struct sk_buff
*skb
;
3465 if (sk
->sk_state
== TCP_CLOSE
)
3468 skb
= tcp_send_head(sk
);
3469 if (skb
&& before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3471 unsigned int mss
= tcp_current_mss(sk
);
3472 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3474 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3475 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3477 /* We are probing the opening of a window
3478 * but the window size is != 0
3479 * must have been a result SWS avoidance ( sender )
3481 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3483 seg_size
= min(seg_size
, mss
);
3484 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3485 if (tcp_fragment(sk
, skb
, seg_size
, mss
, GFP_ATOMIC
))
3487 } else if (!tcp_skb_pcount(skb
))
3488 tcp_set_skb_tso_segs(skb
, mss
);
3490 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3491 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3493 tcp_event_new_data_sent(sk
, skb
);
3496 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3497 tcp_xmit_probe_skb(sk
, 1, mib
);
3498 return tcp_xmit_probe_skb(sk
, 0, mib
);
3502 /* A window probe timeout has occurred. If window is not closed send
3503 * a partial packet else a zero probe.
3505 void tcp_send_probe0(struct sock
*sk
)
3507 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3508 struct tcp_sock
*tp
= tcp_sk(sk
);
3509 struct net
*net
= sock_net(sk
);
3510 unsigned long probe_max
;
3513 err
= tcp_write_wakeup(sk
, LINUX_MIB_TCPWINPROBE
);
3515 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3516 /* Cancel probe timer, if it is not required. */
3517 icsk
->icsk_probes_out
= 0;
3518 icsk
->icsk_backoff
= 0;
3523 if (icsk
->icsk_backoff
< net
->ipv4
.sysctl_tcp_retries2
)
3524 icsk
->icsk_backoff
++;
3525 icsk
->icsk_probes_out
++;
3526 probe_max
= TCP_RTO_MAX
;
3528 /* If packet was not sent due to local congestion,
3529 * do not backoff and do not remember icsk_probes_out.
3530 * Let local senders to fight for local resources.
3532 * Use accumulated backoff yet.
3534 if (!icsk
->icsk_probes_out
)
3535 icsk
->icsk_probes_out
= 1;
3536 probe_max
= TCP_RESOURCE_PROBE_INTERVAL
;
3538 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3539 tcp_probe0_when(sk
, probe_max
),
3543 int tcp_rtx_synack(const struct sock
*sk
, struct request_sock
*req
)
3545 const struct tcp_request_sock_ops
*af_ops
= tcp_rsk(req
)->af_specific
;
3549 tcp_rsk(req
)->txhash
= net_tx_rndhash();
3550 res
= af_ops
->send_synack(sk
, NULL
, &fl
, req
, NULL
, TCP_SYNACK_NORMAL
);
3552 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
3553 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
3557 EXPORT_SYMBOL(tcp_rtx_synack
);