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 unsigned int sysctl_tcp_notsent_lowat __read_mostly
= UINT_MAX
;
66 EXPORT_SYMBOL(sysctl_tcp_notsent_lowat
);
68 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
69 int push_one
, gfp_t gfp
);
71 /* Account for new data that has been sent to the network. */
72 static void tcp_event_new_data_sent(struct sock
*sk
, const struct sk_buff
*skb
)
74 struct inet_connection_sock
*icsk
= inet_csk(sk
);
75 struct tcp_sock
*tp
= tcp_sk(sk
);
76 unsigned int prior_packets
= tp
->packets_out
;
78 tcp_advance_send_head(sk
, skb
);
79 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
81 tp
->packets_out
+= tcp_skb_pcount(skb
);
82 if (!prior_packets
|| icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
83 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
87 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
,
91 /* SND.NXT, if window was not shrunk.
92 * If window has been shrunk, what should we make? It is not clear at all.
93 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
94 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
95 * invalid. OK, let's make this for now:
97 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
99 const struct tcp_sock
*tp
= tcp_sk(sk
);
101 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
104 return tcp_wnd_end(tp
);
107 /* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
121 static __u16
tcp_advertise_mss(struct sock
*sk
)
123 struct tcp_sock
*tp
= tcp_sk(sk
);
124 const struct dst_entry
*dst
= __sk_dst_get(sk
);
125 int mss
= tp
->advmss
;
128 unsigned int metric
= dst_metric_advmss(dst
);
139 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism.
142 void tcp_cwnd_restart(struct sock
*sk
, s32 delta
)
144 struct tcp_sock
*tp
= tcp_sk(sk
);
145 u32 restart_cwnd
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
146 u32 cwnd
= tp
->snd_cwnd
;
148 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
150 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
151 restart_cwnd
= min(restart_cwnd
, cwnd
);
153 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
155 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
156 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
157 tp
->snd_cwnd_used
= 0;
160 /* Congestion state accounting after a packet has been sent. */
161 static void tcp_event_data_sent(struct tcp_sock
*tp
,
164 struct inet_connection_sock
*icsk
= inet_csk(sk
);
165 const u32 now
= tcp_time_stamp
;
167 if (tcp_packets_in_flight(tp
) == 0)
168 tcp_ca_event(sk
, CA_EVENT_TX_START
);
172 /* If it is a reply for ato after last received
173 * packet, enter pingpong mode.
175 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
176 icsk
->icsk_ack
.pingpong
= 1;
179 /* Account for an ACK we sent. */
180 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
182 tcp_dec_quickack_mode(sk
, pkts
);
183 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
187 u32
tcp_default_init_rwnd(u32 mss
)
189 /* Initial receive window should be twice of TCP_INIT_CWND to
190 * enable proper sending of new unsent data during fast recovery
191 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
192 * limit when mss is larger than 1460.
194 u32 init_rwnd
= sysctl_tcp_default_init_rwnd
;
197 init_rwnd
= max((1460 * init_rwnd
) / mss
, 2U);
201 /* Determine a window scaling and initial window to offer.
202 * Based on the assumption that the given amount of space
203 * will be offered. Store the results in the tp structure.
204 * NOTE: for smooth operation initial space offering should
205 * be a multiple of mss if possible. We assume here that mss >= 1.
206 * This MUST be enforced by all callers.
208 void tcp_select_initial_window(int __space
, __u32 mss
,
209 __u32
*rcv_wnd
, __u32
*window_clamp
,
210 int wscale_ok
, __u8
*rcv_wscale
,
213 unsigned int space
= (__space
< 0 ? 0 : __space
);
215 /* If no clamp set the clamp to the max possible scaled window */
216 if (*window_clamp
== 0)
217 (*window_clamp
) = (65535 << 14);
218 space
= min(*window_clamp
, space
);
220 /* Quantize space offering to a multiple of mss if possible. */
222 space
= (space
/ mss
) * mss
;
224 /* NOTE: offering an initial window larger than 32767
225 * will break some buggy TCP stacks. If the admin tells us
226 * it is likely we could be speaking with such a buggy stack
227 * we will truncate our initial window offering to 32K-1
228 * unless the remote has sent us a window scaling option,
229 * which we interpret as a sign the remote TCP is not
230 * misinterpreting the window field as a signed quantity.
232 if (sysctl_tcp_workaround_signed_windows
)
233 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
239 /* Set window scaling on max possible window
240 * See RFC1323 for an explanation of the limit to 14
242 space
= max_t(u32
, space
, sysctl_tcp_rmem
[2]);
243 space
= max_t(u32
, space
, sysctl_rmem_max
);
244 space
= min_t(u32
, space
, *window_clamp
);
245 while (space
> 65535 && (*rcv_wscale
) < 14) {
251 if (mss
> (1 << *rcv_wscale
)) {
252 if (!init_rcv_wnd
) /* Use default unless specified otherwise */
253 init_rcv_wnd
= tcp_default_init_rwnd(mss
);
254 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
257 /* Set the clamp no higher than max representable value */
258 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
260 EXPORT_SYMBOL(tcp_select_initial_window
);
262 /* Chose a new window to advertise, update state in tcp_sock for the
263 * socket, and return result with RFC1323 scaling applied. The return
264 * value can be stuffed directly into th->window for an outgoing
267 static u16
tcp_select_window(struct sock
*sk
)
269 struct tcp_sock
*tp
= tcp_sk(sk
);
270 u32 old_win
= tp
->rcv_wnd
;
271 u32 cur_win
= tcp_receive_window(tp
);
272 u32 new_win
= __tcp_select_window(sk
);
274 /* Never shrink the offered window */
275 if (new_win
< cur_win
) {
276 /* Danger Will Robinson!
277 * Don't update rcv_wup/rcv_wnd here or else
278 * we will not be able to advertise a zero
279 * window in time. --DaveM
281 * Relax Will Robinson.
284 NET_INC_STATS(sock_net(sk
),
285 LINUX_MIB_TCPWANTZEROWINDOWADV
);
286 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
288 tp
->rcv_wnd
= new_win
;
289 tp
->rcv_wup
= tp
->rcv_nxt
;
291 /* Make sure we do not exceed the maximum possible
294 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
295 new_win
= min(new_win
, MAX_TCP_WINDOW
);
297 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
299 /* RFC1323 scaling applied */
300 new_win
>>= tp
->rx_opt
.rcv_wscale
;
302 /* If we advertise zero window, disable fast path. */
306 NET_INC_STATS(sock_net(sk
),
307 LINUX_MIB_TCPTOZEROWINDOWADV
);
308 } else if (old_win
== 0) {
309 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFROMZEROWINDOWADV
);
315 /* Packet ECN state for a SYN-ACK */
316 static void tcp_ecn_send_synack(struct sock
*sk
, struct sk_buff
*skb
)
318 const struct tcp_sock
*tp
= tcp_sk(sk
);
320 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
321 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
322 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
323 else if (tcp_ca_needs_ecn(sk
))
327 /* Packet ECN state for a SYN. */
328 static void tcp_ecn_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
330 struct tcp_sock
*tp
= tcp_sk(sk
);
331 bool use_ecn
= sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1 ||
332 tcp_ca_needs_ecn(sk
);
335 const struct dst_entry
*dst
= __sk_dst_get(sk
);
337 if (dst
&& dst_feature(dst
, RTAX_FEATURE_ECN
))
344 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
345 tp
->ecn_flags
= TCP_ECN_OK
;
346 if (tcp_ca_needs_ecn(sk
))
351 static void tcp_ecn_clear_syn(struct sock
*sk
, struct sk_buff
*skb
)
353 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn_fallback
)
354 /* tp->ecn_flags are cleared at a later point in time when
355 * SYN ACK is ultimatively being received.
357 TCP_SKB_CB(skb
)->tcp_flags
&= ~(TCPHDR_ECE
| TCPHDR_CWR
);
361 tcp_ecn_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
363 if (inet_rsk(req
)->ecn_ok
)
367 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
370 static void tcp_ecn_send(struct sock
*sk
, struct sk_buff
*skb
,
373 struct tcp_sock
*tp
= tcp_sk(sk
);
375 if (tp
->ecn_flags
& TCP_ECN_OK
) {
376 /* Not-retransmitted data segment: set ECT and inject CWR. */
377 if (skb
->len
!= tcp_header_len
&&
378 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
380 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
381 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
382 tcp_hdr(skb
)->cwr
= 1;
383 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
385 } else if (!tcp_ca_needs_ecn(sk
)) {
386 /* ACK or retransmitted segment: clear ECT|CE */
387 INET_ECN_dontxmit(sk
);
389 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
390 tcp_hdr(skb
)->ece
= 1;
394 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
395 * auto increment end seqno.
397 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
399 skb
->ip_summed
= CHECKSUM_PARTIAL
;
402 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
403 TCP_SKB_CB(skb
)->sacked
= 0;
405 tcp_skb_pcount_set(skb
, 1);
407 TCP_SKB_CB(skb
)->seq
= seq
;
408 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
410 TCP_SKB_CB(skb
)->end_seq
= seq
;
413 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
415 return tp
->snd_una
!= tp
->snd_up
;
418 #define OPTION_SACK_ADVERTISE (1 << 0)
419 #define OPTION_TS (1 << 1)
420 #define OPTION_MD5 (1 << 2)
421 #define OPTION_WSCALE (1 << 3)
422 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
424 struct tcp_out_options
{
425 u16 options
; /* bit field of OPTION_* */
426 u16 mss
; /* 0 to disable */
427 u8 ws
; /* window scale, 0 to disable */
428 u8 num_sack_blocks
; /* number of SACK blocks to include */
429 u8 hash_size
; /* bytes in hash_location */
430 __u8
*hash_location
; /* temporary pointer, overloaded */
431 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
432 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
435 /* Write previously computed TCP options to the packet.
437 * Beware: Something in the Internet is very sensitive to the ordering of
438 * TCP options, we learned this through the hard way, so be careful here.
439 * Luckily we can at least blame others for their non-compliance but from
440 * inter-operability perspective it seems that we're somewhat stuck with
441 * the ordering which we have been using if we want to keep working with
442 * those broken things (not that it currently hurts anybody as there isn't
443 * particular reason why the ordering would need to be changed).
445 * At least SACK_PERM as the first option is known to lead to a disaster
446 * (but it may well be that other scenarios fail similarly).
448 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
449 struct tcp_out_options
*opts
)
451 u16 options
= opts
->options
; /* mungable copy */
453 if (unlikely(OPTION_MD5
& options
)) {
454 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
455 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
456 /* overload cookie hash location */
457 opts
->hash_location
= (__u8
*)ptr
;
461 if (unlikely(opts
->mss
)) {
462 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
463 (TCPOLEN_MSS
<< 16) |
467 if (likely(OPTION_TS
& options
)) {
468 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
469 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
470 (TCPOLEN_SACK_PERM
<< 16) |
471 (TCPOPT_TIMESTAMP
<< 8) |
473 options
&= ~OPTION_SACK_ADVERTISE
;
475 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
477 (TCPOPT_TIMESTAMP
<< 8) |
480 *ptr
++ = htonl(opts
->tsval
);
481 *ptr
++ = htonl(opts
->tsecr
);
484 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
485 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
487 (TCPOPT_SACK_PERM
<< 8) |
491 if (unlikely(OPTION_WSCALE
& options
)) {
492 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
493 (TCPOPT_WINDOW
<< 16) |
494 (TCPOLEN_WINDOW
<< 8) |
498 if (unlikely(opts
->num_sack_blocks
)) {
499 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
500 tp
->duplicate_sack
: tp
->selective_acks
;
503 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
506 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
507 TCPOLEN_SACK_PERBLOCK
)));
509 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
511 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
512 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
515 tp
->rx_opt
.dsack
= 0;
518 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
519 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
521 u32 len
; /* Fast Open option length */
524 len
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
525 *ptr
= htonl((TCPOPT_EXP
<< 24) | (len
<< 16) |
526 TCPOPT_FASTOPEN_MAGIC
);
527 p
+= TCPOLEN_EXP_FASTOPEN_BASE
;
529 len
= TCPOLEN_FASTOPEN_BASE
+ foc
->len
;
530 *p
++ = TCPOPT_FASTOPEN
;
534 memcpy(p
, foc
->val
, foc
->len
);
535 if ((len
& 3) == 2) {
536 p
[foc
->len
] = TCPOPT_NOP
;
537 p
[foc
->len
+ 1] = TCPOPT_NOP
;
539 ptr
+= (len
+ 3) >> 2;
543 /* Compute TCP options for SYN packets. This is not the final
544 * network wire format yet.
546 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
547 struct tcp_out_options
*opts
,
548 struct tcp_md5sig_key
**md5
)
550 struct tcp_sock
*tp
= tcp_sk(sk
);
551 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
552 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
554 #ifdef CONFIG_TCP_MD5SIG
555 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
557 opts
->options
|= OPTION_MD5
;
558 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
564 /* We always get an MSS option. The option bytes which will be seen in
565 * normal data packets should timestamps be used, must be in the MSS
566 * advertised. But we subtract them from tp->mss_cache so that
567 * calculations in tcp_sendmsg are simpler etc. So account for this
568 * fact here if necessary. If we don't do this correctly, as a
569 * receiver we won't recognize data packets as being full sized when we
570 * should, and thus we won't abide by the delayed ACK rules correctly.
571 * SACKs don't matter, we never delay an ACK when we have any of those
573 opts
->mss
= tcp_advertise_mss(sk
);
574 remaining
-= TCPOLEN_MSS_ALIGNED
;
576 if (likely(sysctl_tcp_timestamps
&& !*md5
)) {
577 opts
->options
|= OPTION_TS
;
578 opts
->tsval
= tcp_skb_timestamp(skb
) + tp
->tsoffset
;
579 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
580 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
582 if (likely(sysctl_tcp_window_scaling
)) {
583 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
584 opts
->options
|= OPTION_WSCALE
;
585 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
587 if (likely(sysctl_tcp_sack
)) {
588 opts
->options
|= OPTION_SACK_ADVERTISE
;
589 if (unlikely(!(OPTION_TS
& opts
->options
)))
590 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
593 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
594 u32 need
= fastopen
->cookie
.len
;
596 need
+= fastopen
->cookie
.exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
597 TCPOLEN_FASTOPEN_BASE
;
598 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
599 if (remaining
>= need
) {
600 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
601 opts
->fastopen_cookie
= &fastopen
->cookie
;
603 tp
->syn_fastopen
= 1;
604 tp
->syn_fastopen_exp
= fastopen
->cookie
.exp
? 1 : 0;
608 return MAX_TCP_OPTION_SPACE
- remaining
;
611 /* Set up TCP options for SYN-ACKs. */
612 static unsigned int tcp_synack_options(struct request_sock
*req
,
613 unsigned int mss
, struct sk_buff
*skb
,
614 struct tcp_out_options
*opts
,
615 const struct tcp_md5sig_key
*md5
,
616 struct tcp_fastopen_cookie
*foc
)
618 struct inet_request_sock
*ireq
= inet_rsk(req
);
619 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
621 #ifdef CONFIG_TCP_MD5SIG
623 opts
->options
|= OPTION_MD5
;
624 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
626 /* We can't fit any SACK blocks in a packet with MD5 + TS
627 * options. There was discussion about disabling SACK
628 * rather than TS in order to fit in better with old,
629 * buggy kernels, but that was deemed to be unnecessary.
631 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
635 /* We always send an MSS option. */
637 remaining
-= TCPOLEN_MSS_ALIGNED
;
639 if (likely(ireq
->wscale_ok
)) {
640 opts
->ws
= ireq
->rcv_wscale
;
641 opts
->options
|= OPTION_WSCALE
;
642 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
644 if (likely(ireq
->tstamp_ok
)) {
645 opts
->options
|= OPTION_TS
;
646 opts
->tsval
= tcp_skb_timestamp(skb
);
647 opts
->tsecr
= req
->ts_recent
;
648 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
650 if (likely(ireq
->sack_ok
)) {
651 opts
->options
|= OPTION_SACK_ADVERTISE
;
652 if (unlikely(!ireq
->tstamp_ok
))
653 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
655 if (foc
!= NULL
&& foc
->len
>= 0) {
658 need
+= foc
->exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
659 TCPOLEN_FASTOPEN_BASE
;
660 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
661 if (remaining
>= need
) {
662 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
663 opts
->fastopen_cookie
= foc
;
668 return MAX_TCP_OPTION_SPACE
- remaining
;
671 /* Compute TCP options for ESTABLISHED sockets. This is not the
672 * final wire format yet.
674 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
675 struct tcp_out_options
*opts
,
676 struct tcp_md5sig_key
**md5
)
678 struct tcp_sock
*tp
= tcp_sk(sk
);
679 unsigned int size
= 0;
680 unsigned int eff_sacks
;
684 #ifdef CONFIG_TCP_MD5SIG
685 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
686 if (unlikely(*md5
)) {
687 opts
->options
|= OPTION_MD5
;
688 size
+= TCPOLEN_MD5SIG_ALIGNED
;
694 if (likely(tp
->rx_opt
.tstamp_ok
)) {
695 opts
->options
|= OPTION_TS
;
696 opts
->tsval
= skb
? tcp_skb_timestamp(skb
) + tp
->tsoffset
: 0;
697 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
698 size
+= TCPOLEN_TSTAMP_ALIGNED
;
701 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
702 if (unlikely(eff_sacks
)) {
703 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
704 opts
->num_sack_blocks
=
705 min_t(unsigned int, eff_sacks
,
706 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
707 TCPOLEN_SACK_PERBLOCK
);
708 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
709 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
716 /* TCP SMALL QUEUES (TSQ)
718 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
719 * to reduce RTT and bufferbloat.
720 * We do this using a special skb destructor (tcp_wfree).
722 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
723 * needs to be reallocated in a driver.
724 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
726 * Since transmit from skb destructor is forbidden, we use a tasklet
727 * to process all sockets that eventually need to send more skbs.
728 * We use one tasklet per cpu, with its own queue of sockets.
731 struct tasklet_struct tasklet
;
732 struct list_head head
; /* queue of tcp sockets */
734 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
736 static void tcp_tsq_handler(struct sock
*sk
)
738 if ((1 << sk
->sk_state
) &
739 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
740 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
))
741 tcp_write_xmit(sk
, tcp_current_mss(sk
), tcp_sk(sk
)->nonagle
,
745 * One tasklet per cpu tries to send more skbs.
746 * We run in tasklet context but need to disable irqs when
747 * transferring tsq->head because tcp_wfree() might
748 * interrupt us (non NAPI drivers)
750 static void tcp_tasklet_func(unsigned long data
)
752 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
755 struct list_head
*q
, *n
;
759 local_irq_save(flags
);
760 list_splice_init(&tsq
->head
, &list
);
761 local_irq_restore(flags
);
763 list_for_each_safe(q
, n
, &list
) {
764 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
765 list_del(&tp
->tsq_node
);
767 sk
= (struct sock
*)tp
;
770 if (!sock_owned_by_user(sk
)) {
773 /* defer the work to tcp_release_cb() */
774 set_bit(TCP_TSQ_DEFERRED
, &tp
->tsq_flags
);
778 clear_bit(TSQ_QUEUED
, &tp
->tsq_flags
);
783 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
784 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
785 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
786 (1UL << TCP_MTU_REDUCED_DEFERRED))
788 * tcp_release_cb - tcp release_sock() callback
791 * called from release_sock() to perform protocol dependent
792 * actions before socket release.
794 void tcp_release_cb(struct sock
*sk
)
796 struct tcp_sock
*tp
= tcp_sk(sk
);
797 unsigned long flags
, nflags
;
799 /* perform an atomic operation only if at least one flag is set */
801 flags
= tp
->tsq_flags
;
802 if (!(flags
& TCP_DEFERRED_ALL
))
804 nflags
= flags
& ~TCP_DEFERRED_ALL
;
805 } while (cmpxchg(&tp
->tsq_flags
, flags
, nflags
) != flags
);
807 if (flags
& (1UL << TCP_TSQ_DEFERRED
))
810 /* Here begins the tricky part :
811 * We are called from release_sock() with :
813 * 2) sk_lock.slock spinlock held
814 * 3) socket owned by us (sk->sk_lock.owned == 1)
816 * But following code is meant to be called from BH handlers,
817 * so we should keep BH disabled, but early release socket ownership
819 sock_release_ownership(sk
);
821 if (flags
& (1UL << TCP_WRITE_TIMER_DEFERRED
)) {
822 tcp_write_timer_handler(sk
);
825 if (flags
& (1UL << TCP_DELACK_TIMER_DEFERRED
)) {
826 tcp_delack_timer_handler(sk
);
829 if (flags
& (1UL << TCP_MTU_REDUCED_DEFERRED
)) {
830 inet_csk(sk
)->icsk_af_ops
->mtu_reduced(sk
);
834 EXPORT_SYMBOL(tcp_release_cb
);
836 void __init
tcp_tasklet_init(void)
840 for_each_possible_cpu(i
) {
841 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
843 INIT_LIST_HEAD(&tsq
->head
);
844 tasklet_init(&tsq
->tasklet
,
851 * Write buffer destructor automatically called from kfree_skb.
852 * We can't xmit new skbs from this context, as we might already
855 void tcp_wfree(struct sk_buff
*skb
)
857 struct sock
*sk
= skb
->sk
;
858 struct tcp_sock
*tp
= tcp_sk(sk
);
861 /* Keep one reference on sk_wmem_alloc.
862 * Will be released by sk_free() from here or tcp_tasklet_func()
864 wmem
= atomic_sub_return(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
866 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
867 * Wait until our queues (qdisc + devices) are drained.
869 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
870 * - chance for incoming ACK (processed by another cpu maybe)
871 * to migrate this flow (skb->ooo_okay will be eventually set)
873 if (wmem
>= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current
)
876 if (test_and_clear_bit(TSQ_THROTTLED
, &tp
->tsq_flags
) &&
877 !test_and_set_bit(TSQ_QUEUED
, &tp
->tsq_flags
)) {
879 struct tsq_tasklet
*tsq
;
881 /* queue this socket to tasklet queue */
882 local_irq_save(flags
);
883 tsq
= this_cpu_ptr(&tsq_tasklet
);
884 list_add(&tp
->tsq_node
, &tsq
->head
);
885 tasklet_schedule(&tsq
->tasklet
);
886 local_irq_restore(flags
);
893 /* This routine actually transmits TCP packets queued in by
894 * tcp_do_sendmsg(). This is used by both the initial
895 * transmission and possible later retransmissions.
896 * All SKB's seen here are completely headerless. It is our
897 * job to build the TCP header, and pass the packet down to
898 * IP so it can do the same plus pass the packet off to the
901 * We are working here with either a clone of the original
902 * SKB, or a fresh unique copy made by the retransmit engine.
904 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
907 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
908 struct inet_sock
*inet
;
910 struct tcp_skb_cb
*tcb
;
911 struct tcp_out_options opts
;
912 unsigned int tcp_options_size
, tcp_header_size
;
913 struct tcp_md5sig_key
*md5
;
917 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
920 skb_mstamp_get(&skb
->skb_mstamp
);
922 if (unlikely(skb_cloned(skb
)))
923 skb
= pskb_copy(skb
, gfp_mask
);
925 skb
= skb_clone(skb
, gfp_mask
);
932 tcb
= TCP_SKB_CB(skb
);
933 memset(&opts
, 0, sizeof(opts
));
935 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
936 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
938 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
940 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
942 /* if no packet is in qdisc/device queue, then allow XPS to select
943 * another queue. We can be called from tcp_tsq_handler()
944 * which holds one reference to sk_wmem_alloc.
946 * TODO: Ideally, in-flight pure ACK packets should not matter here.
947 * One way to get this would be to set skb->truesize = 2 on them.
949 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) < SKB_TRUESIZE(1);
951 skb_push(skb
, tcp_header_size
);
952 skb_reset_transport_header(skb
);
956 skb
->destructor
= skb_is_tcp_pure_ack(skb
) ? sock_wfree
: tcp_wfree
;
957 skb_set_hash_from_sk(skb
, sk
);
958 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
960 /* Build TCP header and checksum it. */
962 th
->source
= inet
->inet_sport
;
963 th
->dest
= inet
->inet_dport
;
964 th
->seq
= htonl(tcb
->seq
);
965 th
->ack_seq
= htonl(tp
->rcv_nxt
);
966 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
969 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
970 /* RFC1323: The window in SYN & SYN/ACK segments
973 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
975 th
->window
= htons(tcp_select_window(sk
));
980 /* The urg_mode check is necessary during a below snd_una win probe */
981 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
982 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
983 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
985 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
986 th
->urg_ptr
= htons(0xFFFF);
991 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
992 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
993 if (likely((tcb
->tcp_flags
& TCPHDR_SYN
) == 0))
994 tcp_ecn_send(sk
, skb
, tcp_header_size
);
996 #ifdef CONFIG_TCP_MD5SIG
997 /* Calculate the MD5 hash, as we have all we need now */
999 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
1000 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
1005 icsk
->icsk_af_ops
->send_check(sk
, skb
);
1007 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
1008 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
1010 if (skb
->len
!= tcp_header_size
)
1011 tcp_event_data_sent(tp
, sk
);
1013 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
1014 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
1015 tcp_skb_pcount(skb
));
1017 tp
->segs_out
+= tcp_skb_pcount(skb
);
1018 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1019 skb_shinfo(skb
)->gso_segs
= tcp_skb_pcount(skb
);
1020 skb_shinfo(skb
)->gso_size
= tcp_skb_mss(skb
);
1022 /* Our usage of tstamp should remain private */
1023 skb
->tstamp
.tv64
= 0;
1025 /* Cleanup our debris for IP stacks */
1026 memset(skb
->cb
, 0, max(sizeof(struct inet_skb_parm
),
1027 sizeof(struct inet6_skb_parm
)));
1029 err
= icsk
->icsk_af_ops
->queue_xmit(sk
, skb
, &inet
->cork
.fl
);
1031 if (likely(err
<= 0))
1036 return net_xmit_eval(err
);
1039 /* This routine just queues the buffer for sending.
1041 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1042 * otherwise socket can stall.
1044 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1046 struct tcp_sock
*tp
= tcp_sk(sk
);
1048 /* Advance write_seq and place onto the write_queue. */
1049 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
1050 __skb_header_release(skb
);
1051 tcp_add_write_queue_tail(sk
, skb
);
1052 sk
->sk_wmem_queued
+= skb
->truesize
;
1053 sk_mem_charge(sk
, skb
->truesize
);
1056 /* Initialize TSO segments for a packet. */
1057 static void tcp_set_skb_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1059 if (skb
->len
<= mss_now
|| skb
->ip_summed
== CHECKSUM_NONE
) {
1060 /* Avoid the costly divide in the normal
1063 tcp_skb_pcount_set(skb
, 1);
1064 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1066 tcp_skb_pcount_set(skb
, DIV_ROUND_UP(skb
->len
, mss_now
));
1067 TCP_SKB_CB(skb
)->tcp_gso_size
= mss_now
;
1071 /* When a modification to fackets out becomes necessary, we need to check
1072 * skb is counted to fackets_out or not.
1074 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1077 struct tcp_sock
*tp
= tcp_sk(sk
);
1079 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1082 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1083 tp
->fackets_out
-= decr
;
1086 /* Pcount in the middle of the write queue got changed, we need to do various
1087 * tweaks to fix counters
1089 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1091 struct tcp_sock
*tp
= tcp_sk(sk
);
1093 tp
->packets_out
-= decr
;
1095 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1096 tp
->sacked_out
-= decr
;
1097 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1098 tp
->retrans_out
-= decr
;
1099 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1100 tp
->lost_out
-= decr
;
1102 /* Reno case is special. Sigh... */
1103 if (tcp_is_reno(tp
) && decr
> 0)
1104 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1106 tcp_adjust_fackets_out(sk
, skb
, decr
);
1108 if (tp
->lost_skb_hint
&&
1109 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1110 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1111 tp
->lost_cnt_hint
-= decr
;
1113 tcp_verify_left_out(tp
);
1116 static void tcp_fragment_tstamp(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1118 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1120 if (unlikely(shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
) &&
1121 !before(shinfo
->tskey
, TCP_SKB_CB(skb2
)->seq
)) {
1122 struct skb_shared_info
*shinfo2
= skb_shinfo(skb2
);
1123 u8 tsflags
= shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
1125 shinfo
->tx_flags
&= ~tsflags
;
1126 shinfo2
->tx_flags
|= tsflags
;
1127 swap(shinfo
->tskey
, shinfo2
->tskey
);
1131 /* Function to create two new TCP segments. Shrinks the given segment
1132 * to the specified size and appends a new segment with the rest of the
1133 * packet to the list. This won't be called frequently, I hope.
1134 * Remember, these are still headerless SKBs at this point.
1136 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1137 unsigned int mss_now
, gfp_t gfp
)
1139 struct tcp_sock
*tp
= tcp_sk(sk
);
1140 struct sk_buff
*buff
;
1141 int nsize
, old_factor
;
1145 if (WARN_ON(len
> skb
->len
))
1148 nsize
= skb_headlen(skb
) - len
;
1152 if (skb_unclone(skb
, gfp
))
1155 /* Get a new skb... force flag on. */
1156 buff
= sk_stream_alloc_skb(sk
, nsize
, gfp
, true);
1158 return -ENOMEM
; /* We'll just try again later. */
1160 sk
->sk_wmem_queued
+= buff
->truesize
;
1161 sk_mem_charge(sk
, buff
->truesize
);
1162 nlen
= skb
->len
- len
- nsize
;
1163 buff
->truesize
+= nlen
;
1164 skb
->truesize
-= nlen
;
1166 /* Correct the sequence numbers. */
1167 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1168 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1169 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1171 /* PSH and FIN should only be set in the second packet. */
1172 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1173 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1174 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1175 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1177 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1178 /* Copy and checksum data tail into the new buffer. */
1179 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1180 skb_put(buff
, nsize
),
1185 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1187 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1188 skb_split(skb
, buff
, len
);
1191 buff
->ip_summed
= skb
->ip_summed
;
1193 buff
->tstamp
= skb
->tstamp
;
1194 tcp_fragment_tstamp(skb
, buff
);
1196 old_factor
= tcp_skb_pcount(skb
);
1198 /* Fix up tso_factor for both original and new SKB. */
1199 tcp_set_skb_tso_segs(skb
, mss_now
);
1200 tcp_set_skb_tso_segs(buff
, mss_now
);
1202 /* If this packet has been sent out already, we must
1203 * adjust the various packet counters.
1205 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1206 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1207 tcp_skb_pcount(buff
);
1210 tcp_adjust_pcount(sk
, skb
, diff
);
1213 /* Link BUFF into the send queue. */
1214 __skb_header_release(buff
);
1215 tcp_insert_write_queue_after(skb
, buff
, sk
);
1220 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1221 * eventually). The difference is that pulled data not copied, but
1222 * immediately discarded.
1224 static int __pskb_trim_head(struct sk_buff
*skb
, int len
)
1226 struct skb_shared_info
*shinfo
;
1229 eat
= min_t(int, len
, skb_headlen(skb
));
1231 __skb_pull(skb
, eat
);
1238 shinfo
= skb_shinfo(skb
);
1239 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1240 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1243 skb_frag_unref(skb
, i
);
1246 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1248 shinfo
->frags
[k
].page_offset
+= eat
;
1249 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1255 shinfo
->nr_frags
= k
;
1257 skb_reset_tail_pointer(skb
);
1258 skb
->data_len
-= len
;
1259 skb
->len
= skb
->data_len
;
1263 /* Remove acked data from a packet in the transmit queue. */
1264 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1268 if (skb_unclone(skb
, GFP_ATOMIC
))
1271 delta_truesize
= __pskb_trim_head(skb
, len
);
1273 TCP_SKB_CB(skb
)->seq
+= len
;
1274 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1276 if (delta_truesize
) {
1277 skb
->truesize
-= delta_truesize
;
1278 sk
->sk_wmem_queued
-= delta_truesize
;
1279 sk_mem_uncharge(sk
, delta_truesize
);
1280 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1283 /* Any change of skb->len requires recalculation of tso factor. */
1284 if (tcp_skb_pcount(skb
) > 1)
1285 tcp_set_skb_tso_segs(skb
, tcp_skb_mss(skb
));
1290 /* Calculate MSS not accounting any TCP options. */
1291 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1293 const struct tcp_sock
*tp
= tcp_sk(sk
);
1294 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1297 /* Calculate base mss without TCP options:
1298 It is MMS_S - sizeof(tcphdr) of rfc1122
1300 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1302 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1303 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1304 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1306 if (dst
&& dst_allfrag(dst
))
1307 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1310 /* Clamp it (mss_clamp does not include tcp options) */
1311 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1312 mss_now
= tp
->rx_opt
.mss_clamp
;
1314 /* Now subtract optional transport overhead */
1315 mss_now
-= icsk
->icsk_ext_hdr_len
;
1317 /* Then reserve room for full set of TCP options and 8 bytes of data */
1323 /* Calculate MSS. Not accounting for SACKs here. */
1324 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1326 /* Subtract TCP options size, not including SACKs */
1327 return __tcp_mtu_to_mss(sk
, pmtu
) -
1328 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1331 /* Inverse of above */
1332 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1334 const struct tcp_sock
*tp
= tcp_sk(sk
);
1335 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1339 tp
->tcp_header_len
+
1340 icsk
->icsk_ext_hdr_len
+
1341 icsk
->icsk_af_ops
->net_header_len
;
1343 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1344 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1345 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1347 if (dst
&& dst_allfrag(dst
))
1348 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1353 /* MTU probing init per socket */
1354 void tcp_mtup_init(struct sock
*sk
)
1356 struct tcp_sock
*tp
= tcp_sk(sk
);
1357 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1358 struct net
*net
= sock_net(sk
);
1360 icsk
->icsk_mtup
.enabled
= net
->ipv4
.sysctl_tcp_mtu_probing
> 1;
1361 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1362 icsk
->icsk_af_ops
->net_header_len
;
1363 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, net
->ipv4
.sysctl_tcp_base_mss
);
1364 icsk
->icsk_mtup
.probe_size
= 0;
1365 if (icsk
->icsk_mtup
.enabled
)
1366 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1368 EXPORT_SYMBOL(tcp_mtup_init
);
1370 /* This function synchronize snd mss to current pmtu/exthdr set.
1372 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1373 for TCP options, but includes only bare TCP header.
1375 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1376 It is minimum of user_mss and mss received with SYN.
1377 It also does not include TCP options.
1379 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1381 tp->mss_cache is current effective sending mss, including
1382 all tcp options except for SACKs. It is evaluated,
1383 taking into account current pmtu, but never exceeds
1384 tp->rx_opt.mss_clamp.
1386 NOTE1. rfc1122 clearly states that advertised MSS
1387 DOES NOT include either tcp or ip options.
1389 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1390 are READ ONLY outside this function. --ANK (980731)
1392 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1394 struct tcp_sock
*tp
= tcp_sk(sk
);
1395 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1398 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1399 icsk
->icsk_mtup
.search_high
= pmtu
;
1401 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1402 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1404 /* And store cached results */
1405 icsk
->icsk_pmtu_cookie
= pmtu
;
1406 if (icsk
->icsk_mtup
.enabled
)
1407 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1408 tp
->mss_cache
= mss_now
;
1412 EXPORT_SYMBOL(tcp_sync_mss
);
1414 /* Compute the current effective MSS, taking SACKs and IP options,
1415 * and even PMTU discovery events into account.
1417 unsigned int tcp_current_mss(struct sock
*sk
)
1419 const struct tcp_sock
*tp
= tcp_sk(sk
);
1420 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1422 unsigned int header_len
;
1423 struct tcp_out_options opts
;
1424 struct tcp_md5sig_key
*md5
;
1426 mss_now
= tp
->mss_cache
;
1429 u32 mtu
= dst_mtu(dst
);
1430 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1431 mss_now
= tcp_sync_mss(sk
, mtu
);
1434 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1435 sizeof(struct tcphdr
);
1436 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1437 * some common options. If this is an odd packet (because we have SACK
1438 * blocks etc) then our calculated header_len will be different, and
1439 * we have to adjust mss_now correspondingly */
1440 if (header_len
!= tp
->tcp_header_len
) {
1441 int delta
= (int) header_len
- tp
->tcp_header_len
;
1448 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1449 * As additional protections, we do not touch cwnd in retransmission phases,
1450 * and if application hit its sndbuf limit recently.
1452 static void tcp_cwnd_application_limited(struct sock
*sk
)
1454 struct tcp_sock
*tp
= tcp_sk(sk
);
1456 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1457 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1458 /* Limited by application or receiver window. */
1459 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1460 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1461 if (win_used
< tp
->snd_cwnd
) {
1462 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1463 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
1465 tp
->snd_cwnd_used
= 0;
1467 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1470 static void tcp_cwnd_validate(struct sock
*sk
, bool is_cwnd_limited
)
1472 struct tcp_sock
*tp
= tcp_sk(sk
);
1474 /* Track the maximum number of outstanding packets in each
1475 * window, and remember whether we were cwnd-limited then.
1477 if (!before(tp
->snd_una
, tp
->max_packets_seq
) ||
1478 tp
->packets_out
> tp
->max_packets_out
) {
1479 tp
->max_packets_out
= tp
->packets_out
;
1480 tp
->max_packets_seq
= tp
->snd_nxt
;
1481 tp
->is_cwnd_limited
= is_cwnd_limited
;
1484 if (tcp_is_cwnd_limited(sk
)) {
1485 /* Network is feed fully. */
1486 tp
->snd_cwnd_used
= 0;
1487 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1489 /* Network starves. */
1490 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1491 tp
->snd_cwnd_used
= tp
->packets_out
;
1493 if (sysctl_tcp_slow_start_after_idle
&&
1494 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1495 tcp_cwnd_application_limited(sk
);
1499 /* Minshall's variant of the Nagle send check. */
1500 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1502 return after(tp
->snd_sml
, tp
->snd_una
) &&
1503 !after(tp
->snd_sml
, tp
->snd_nxt
);
1506 /* Update snd_sml if this skb is under mss
1507 * Note that a TSO packet might end with a sub-mss segment
1508 * The test is really :
1509 * if ((skb->len % mss) != 0)
1510 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1511 * But we can avoid doing the divide again given we already have
1512 * skb_pcount = skb->len / mss_now
1514 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1515 const struct sk_buff
*skb
)
1517 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1518 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1521 /* Return false, if packet can be sent now without violation Nagle's rules:
1522 * 1. It is full sized. (provided by caller in %partial bool)
1523 * 2. Or it contains FIN. (already checked by caller)
1524 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1525 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1526 * With Minshall's modification: all sent small packets are ACKed.
1528 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1532 ((nonagle
& TCP_NAGLE_CORK
) ||
1533 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1536 /* Return how many segs we'd like on a TSO packet,
1537 * to send one TSO packet per ms
1539 static u32
tcp_tso_autosize(const struct sock
*sk
, unsigned int mss_now
)
1543 bytes
= min(sk
->sk_pacing_rate
>> 10,
1544 sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
);
1546 /* Goal is to send at least one packet per ms,
1547 * not one big TSO packet every 100 ms.
1548 * This preserves ACK clocking and is consistent
1549 * with tcp_tso_should_defer() heuristic.
1551 segs
= max_t(u32
, bytes
/ mss_now
, sysctl_tcp_min_tso_segs
);
1553 return min_t(u32
, segs
, sk
->sk_gso_max_segs
);
1556 /* Returns the portion of skb which can be sent right away */
1557 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1558 const struct sk_buff
*skb
,
1559 unsigned int mss_now
,
1560 unsigned int max_segs
,
1563 const struct tcp_sock
*tp
= tcp_sk(sk
);
1564 u32 partial
, needed
, window
, max_len
;
1566 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1567 max_len
= mss_now
* max_segs
;
1569 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1572 needed
= min(skb
->len
, window
);
1574 if (max_len
<= needed
)
1577 partial
= needed
% mss_now
;
1578 /* If last segment is not a full MSS, check if Nagle rules allow us
1579 * to include this last segment in this skb.
1580 * Otherwise, we'll split the skb at last MSS boundary
1582 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
1583 return needed
- partial
;
1588 /* Can at least one segment of SKB be sent right now, according to the
1589 * congestion window rules? If so, return how many segments are allowed.
1591 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1592 const struct sk_buff
*skb
)
1594 u32 in_flight
, cwnd
, halfcwnd
;
1596 /* Don't be strict about the congestion window for the final FIN. */
1597 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1598 tcp_skb_pcount(skb
) == 1)
1601 in_flight
= tcp_packets_in_flight(tp
);
1602 cwnd
= tp
->snd_cwnd
;
1603 if (in_flight
>= cwnd
)
1606 /* For better scheduling, ensure we have at least
1607 * 2 GSO packets in flight.
1609 halfcwnd
= max(cwnd
>> 1, 1U);
1610 return min(halfcwnd
, cwnd
- in_flight
);
1613 /* Initialize TSO state of a skb.
1614 * This must be invoked the first time we consider transmitting
1615 * SKB onto the wire.
1617 static int tcp_init_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1619 int tso_segs
= tcp_skb_pcount(skb
);
1621 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1622 tcp_set_skb_tso_segs(skb
, mss_now
);
1623 tso_segs
= tcp_skb_pcount(skb
);
1629 /* Return true if the Nagle test allows this packet to be
1632 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1633 unsigned int cur_mss
, int nonagle
)
1635 /* Nagle rule does not apply to frames, which sit in the middle of the
1636 * write_queue (they have no chances to get new data).
1638 * This is implemented in the callers, where they modify the 'nonagle'
1639 * argument based upon the location of SKB in the send queue.
1641 if (nonagle
& TCP_NAGLE_PUSH
)
1644 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1645 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1648 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
1654 /* Does at least the first segment of SKB fit into the send window? */
1655 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1656 const struct sk_buff
*skb
,
1657 unsigned int cur_mss
)
1659 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1661 if (skb
->len
> cur_mss
)
1662 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1664 return !after(end_seq
, tcp_wnd_end(tp
));
1667 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1668 * should be put on the wire right now. If so, it returns the number of
1669 * packets allowed by the congestion window.
1671 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1672 unsigned int cur_mss
, int nonagle
)
1674 const struct tcp_sock
*tp
= tcp_sk(sk
);
1675 unsigned int cwnd_quota
;
1677 tcp_init_tso_segs(skb
, cur_mss
);
1679 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1682 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1683 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1689 /* Test if sending is allowed right now. */
1690 bool tcp_may_send_now(struct sock
*sk
)
1692 const struct tcp_sock
*tp
= tcp_sk(sk
);
1693 struct sk_buff
*skb
= tcp_send_head(sk
);
1696 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1697 (tcp_skb_is_last(sk
, skb
) ?
1698 tp
->nonagle
: TCP_NAGLE_PUSH
));
1701 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1702 * which is put after SKB on the list. It is very much like
1703 * tcp_fragment() except that it may make several kinds of assumptions
1704 * in order to speed up the splitting operation. In particular, we
1705 * know that all the data is in scatter-gather pages, and that the
1706 * packet has never been sent out before (and thus is not cloned).
1708 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1709 unsigned int mss_now
, gfp_t gfp
)
1711 struct sk_buff
*buff
;
1712 int nlen
= skb
->len
- len
;
1715 /* All of a TSO frame must be composed of paged data. */
1716 if (skb
->len
!= skb
->data_len
)
1717 return tcp_fragment(sk
, skb
, len
, mss_now
, gfp
);
1719 buff
= sk_stream_alloc_skb(sk
, 0, gfp
, true);
1720 if (unlikely(!buff
))
1723 sk
->sk_wmem_queued
+= buff
->truesize
;
1724 sk_mem_charge(sk
, buff
->truesize
);
1725 buff
->truesize
+= nlen
;
1726 skb
->truesize
-= nlen
;
1728 /* Correct the sequence numbers. */
1729 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1730 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1731 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1733 /* PSH and FIN should only be set in the second packet. */
1734 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1735 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1736 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1738 /* This packet was never sent out yet, so no SACK bits. */
1739 TCP_SKB_CB(buff
)->sacked
= 0;
1741 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1742 skb_split(skb
, buff
, len
);
1743 tcp_fragment_tstamp(skb
, buff
);
1745 /* Fix up tso_factor for both original and new SKB. */
1746 tcp_set_skb_tso_segs(skb
, mss_now
);
1747 tcp_set_skb_tso_segs(buff
, mss_now
);
1749 /* Link BUFF into the send queue. */
1750 __skb_header_release(buff
);
1751 tcp_insert_write_queue_after(skb
, buff
, sk
);
1756 /* Try to defer sending, if possible, in order to minimize the amount
1757 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1759 * This algorithm is from John Heffner.
1761 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
,
1762 bool *is_cwnd_limited
, u32 max_segs
)
1764 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1765 u32 age
, send_win
, cong_win
, limit
, in_flight
;
1766 struct tcp_sock
*tp
= tcp_sk(sk
);
1767 struct skb_mstamp now
;
1768 struct sk_buff
*head
;
1771 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1774 if (icsk
->icsk_ca_state
>= TCP_CA_Recovery
)
1777 /* Avoid bursty behavior by allowing defer
1778 * only if the last write was recent.
1780 if ((s32
)(tcp_time_stamp
- tp
->lsndtime
) > 0)
1783 in_flight
= tcp_packets_in_flight(tp
);
1785 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1787 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1789 /* From in_flight test above, we know that cwnd > in_flight. */
1790 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1792 limit
= min(send_win
, cong_win
);
1794 /* If a full-sized TSO skb can be sent, do it. */
1795 if (limit
>= max_segs
* tp
->mss_cache
)
1798 /* Middle in queue won't get any more data, full sendable already? */
1799 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1802 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1804 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1806 /* If at least some fraction of a window is available,
1809 chunk
/= win_divisor
;
1813 /* Different approach, try not to defer past a single
1814 * ACK. Receiver should ACK every other full sized
1815 * frame, so if we have space for more than 3 frames
1818 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1822 head
= tcp_write_queue_head(sk
);
1823 skb_mstamp_get(&now
);
1824 age
= skb_mstamp_us_delta(&now
, &head
->skb_mstamp
);
1825 /* If next ACK is likely to come too late (half srtt), do not defer */
1826 if (age
< (tp
->srtt_us
>> 4))
1829 /* Ok, it looks like it is advisable to defer. */
1831 if (cong_win
< send_win
&& cong_win
<= skb
->len
)
1832 *is_cwnd_limited
= true;
1840 static inline void tcp_mtu_check_reprobe(struct sock
*sk
)
1842 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1843 struct tcp_sock
*tp
= tcp_sk(sk
);
1844 struct net
*net
= sock_net(sk
);
1848 interval
= net
->ipv4
.sysctl_tcp_probe_interval
;
1849 delta
= tcp_time_stamp
- icsk
->icsk_mtup
.probe_timestamp
;
1850 if (unlikely(delta
>= interval
* HZ
)) {
1851 int mss
= tcp_current_mss(sk
);
1853 /* Update current search range */
1854 icsk
->icsk_mtup
.probe_size
= 0;
1855 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+
1856 sizeof(struct tcphdr
) +
1857 icsk
->icsk_af_ops
->net_header_len
;
1858 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
1860 /* Update probe time stamp */
1861 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1865 /* Create a new MTU probe if we are ready.
1866 * MTU probe is regularly attempting to increase the path MTU by
1867 * deliberately sending larger packets. This discovers routing
1868 * changes resulting in larger path MTUs.
1870 * Returns 0 if we should wait to probe (no cwnd available),
1871 * 1 if a probe was sent,
1874 static int tcp_mtu_probe(struct sock
*sk
)
1876 struct tcp_sock
*tp
= tcp_sk(sk
);
1877 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1878 struct sk_buff
*skb
, *nskb
, *next
;
1879 struct net
*net
= sock_net(sk
);
1887 /* Not currently probing/verifying,
1889 * have enough cwnd, and
1890 * not SACKing (the variable headers throw things off) */
1891 if (!icsk
->icsk_mtup
.enabled
||
1892 icsk
->icsk_mtup
.probe_size
||
1893 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1894 tp
->snd_cwnd
< 11 ||
1895 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1898 /* Use binary search for probe_size between tcp_mss_base,
1899 * and current mss_clamp. if (search_high - search_low)
1900 * smaller than a threshold, backoff from probing.
1902 mss_now
= tcp_current_mss(sk
);
1903 probe_size
= tcp_mtu_to_mss(sk
, (icsk
->icsk_mtup
.search_high
+
1904 icsk
->icsk_mtup
.search_low
) >> 1);
1905 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1906 interval
= icsk
->icsk_mtup
.search_high
- icsk
->icsk_mtup
.search_low
;
1907 /* When misfortune happens, we are reprobing actively,
1908 * and then reprobe timer has expired. We stick with current
1909 * probing process by not resetting search range to its orignal.
1911 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
) ||
1912 interval
< net
->ipv4
.sysctl_tcp_probe_threshold
) {
1913 /* Check whether enough time has elaplased for
1914 * another round of probing.
1916 tcp_mtu_check_reprobe(sk
);
1920 /* Have enough data in the send queue to probe? */
1921 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1924 if (tp
->snd_wnd
< size_needed
)
1926 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1929 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1930 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1931 if (!tcp_packets_in_flight(tp
))
1937 /* We're allowed to probe. Build it now. */
1938 nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
, false);
1941 sk
->sk_wmem_queued
+= nskb
->truesize
;
1942 sk_mem_charge(sk
, nskb
->truesize
);
1944 skb
= tcp_send_head(sk
);
1946 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1947 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1948 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
1949 TCP_SKB_CB(nskb
)->sacked
= 0;
1951 nskb
->ip_summed
= skb
->ip_summed
;
1953 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1956 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1957 copy
= min_t(int, skb
->len
, probe_size
- len
);
1958 if (nskb
->ip_summed
) {
1959 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1961 __wsum csum
= skb_copy_and_csum_bits(skb
, 0,
1962 skb_put(nskb
, copy
),
1964 nskb
->csum
= csum_block_add(nskb
->csum
, csum
, len
);
1967 if (skb
->len
<= copy
) {
1968 /* We've eaten all the data from this skb.
1970 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1971 tcp_unlink_write_queue(skb
, sk
);
1972 sk_wmem_free_skb(sk
, skb
);
1974 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
1975 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1976 if (!skb_shinfo(skb
)->nr_frags
) {
1977 skb_pull(skb
, copy
);
1978 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1979 skb
->csum
= csum_partial(skb
->data
,
1982 __pskb_trim_head(skb
, copy
);
1983 tcp_set_skb_tso_segs(skb
, mss_now
);
1985 TCP_SKB_CB(skb
)->seq
+= copy
;
1990 if (len
>= probe_size
)
1993 tcp_init_tso_segs(nskb
, nskb
->len
);
1995 /* We're ready to send. If this fails, the probe will
1996 * be resegmented into mss-sized pieces by tcp_write_xmit().
1998 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1999 /* Decrement cwnd here because we are sending
2000 * effectively two packets. */
2002 tcp_event_new_data_sent(sk
, nskb
);
2004 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
2005 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
2006 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
2014 /* This routine writes packets to the network. It advances the
2015 * send_head. This happens as incoming acks open up the remote
2018 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2019 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2020 * account rare use of URG, this is not a big flaw.
2022 * Send at most one packet when push_one > 0. Temporarily ignore
2023 * cwnd limit to force at most one packet out when push_one == 2.
2025 * Returns true, if no segments are in flight and we have queued segments,
2026 * but cannot send anything now because of SWS or another problem.
2028 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
2029 int push_one
, gfp_t gfp
)
2031 struct tcp_sock
*tp
= tcp_sk(sk
);
2032 struct sk_buff
*skb
;
2033 unsigned int tso_segs
, sent_pkts
;
2036 bool is_cwnd_limited
= false;
2042 /* Do MTU probing. */
2043 result
= tcp_mtu_probe(sk
);
2046 } else if (result
> 0) {
2051 max_segs
= tcp_tso_autosize(sk
, mss_now
);
2052 while ((skb
= tcp_send_head(sk
))) {
2055 tso_segs
= tcp_init_tso_segs(skb
, mss_now
);
2058 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
) {
2059 /* "skb_mstamp" is used as a start point for the retransmit timer */
2060 skb_mstamp_get(&skb
->skb_mstamp
);
2061 goto repair
; /* Skip network transmission */
2064 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2067 /* Force out a loss probe pkt. */
2073 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
2076 if (tso_segs
== 1) {
2077 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2078 (tcp_skb_is_last(sk
, skb
) ?
2079 nonagle
: TCP_NAGLE_PUSH
))))
2083 tcp_tso_should_defer(sk
, skb
, &is_cwnd_limited
,
2089 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
2090 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2096 if (skb
->len
> limit
&&
2097 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
2100 /* TCP Small Queues :
2101 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2103 * - better RTT estimation and ACK scheduling
2106 * Alas, some drivers / subsystems require a fair amount
2107 * of queued bytes to ensure line rate.
2108 * One example is wifi aggregation (802.11 AMPDU)
2110 limit
= max(2 * skb
->truesize
, sk
->sk_pacing_rate
>> 10);
2111 limit
= min_t(u32
, limit
, sysctl_tcp_limit_output_bytes
);
2113 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
) {
2114 set_bit(TSQ_THROTTLED
, &tp
->tsq_flags
);
2115 /* It is possible TX completion already happened
2116 * before we set TSQ_THROTTLED, so we must
2117 * test again the condition.
2119 smp_mb__after_atomic();
2120 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
)
2124 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2128 /* Advance the send_head. This one is sent out.
2129 * This call will increment packets_out.
2131 tcp_event_new_data_sent(sk
, skb
);
2133 tcp_minshall_update(tp
, mss_now
, skb
);
2134 sent_pkts
+= tcp_skb_pcount(skb
);
2140 if (likely(sent_pkts
)) {
2141 if (tcp_in_cwnd_reduction(sk
))
2142 tp
->prr_out
+= sent_pkts
;
2144 /* Send one loss probe per tail loss episode. */
2146 tcp_schedule_loss_probe(sk
);
2147 is_cwnd_limited
|= (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
);
2148 tcp_cwnd_validate(sk
, is_cwnd_limited
);
2151 return !tp
->packets_out
&& tcp_send_head(sk
);
2154 bool tcp_schedule_loss_probe(struct sock
*sk
)
2156 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2157 struct tcp_sock
*tp
= tcp_sk(sk
);
2158 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
2159 u32 rtt
= usecs_to_jiffies(tp
->srtt_us
>> 3);
2161 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
))
2163 /* No consecutive loss probes. */
2164 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
2168 /* Don't do any loss probe on a Fast Open connection before 3WHS
2171 if (tp
->fastopen_rsk
)
2174 /* TLP is only scheduled when next timer event is RTO. */
2175 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
2178 /* Schedule a loss probe in 2*RTT for SACK capable connections
2179 * in Open state, that are either limited by cwnd or application.
2181 if (sysctl_tcp_early_retrans
< 3 || !tp
->packets_out
||
2182 !tcp_is_sack(tp
) || inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
2185 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
2189 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2190 * for delayed ack when there's one outstanding packet. If no RTT
2191 * sample is available then probe after TCP_TIMEOUT_INIT.
2193 timeout
= rtt
<< 1 ? : TCP_TIMEOUT_INIT
;
2194 if (tp
->packets_out
== 1)
2195 timeout
= max_t(u32
, timeout
,
2196 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
2197 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
2199 /* If RTO is shorter, just schedule TLP in its place. */
2200 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
2201 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
2202 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
2203 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
2208 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2213 /* Thanks to skb fast clones, we can detect if a prior transmit of
2214 * a packet is still in a qdisc or driver queue.
2215 * In this case, there is very little point doing a retransmit !
2216 * Note: This is called from BH context only.
2218 static bool skb_still_in_host_queue(const struct sock
*sk
,
2219 const struct sk_buff
*skb
)
2221 if (unlikely(skb_fclone_busy(sk
, skb
))) {
2222 NET_INC_STATS_BH(sock_net(sk
),
2223 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
2229 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2230 * retransmit the last segment.
2232 void tcp_send_loss_probe(struct sock
*sk
)
2234 struct tcp_sock
*tp
= tcp_sk(sk
);
2235 struct sk_buff
*skb
;
2237 int mss
= tcp_current_mss(sk
);
2239 skb
= tcp_send_head(sk
);
2241 if (tcp_snd_wnd_test(tp
, skb
, mss
)) {
2242 pcount
= tp
->packets_out
;
2243 tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2244 if (tp
->packets_out
> pcount
)
2248 skb
= tcp_write_queue_prev(sk
, skb
);
2250 skb
= tcp_write_queue_tail(sk
);
2253 /* At most one outstanding TLP retransmission. */
2254 if (tp
->tlp_high_seq
)
2257 /* Retransmit last segment. */
2261 if (skb_still_in_host_queue(sk
, skb
))
2264 pcount
= tcp_skb_pcount(skb
);
2265 if (WARN_ON(!pcount
))
2268 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2269 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
,
2272 skb
= tcp_write_queue_next(sk
, skb
);
2275 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2278 if (__tcp_retransmit_skb(sk
, skb
))
2281 /* Record snd_nxt for loss detection. */
2282 tp
->tlp_high_seq
= tp
->snd_nxt
;
2285 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSPROBES
);
2286 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2287 inet_csk(sk
)->icsk_pending
= 0;
2292 /* Push out any pending frames which were held back due to
2293 * TCP_CORK or attempt at coalescing tiny packets.
2294 * The socket must be locked by the caller.
2296 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2299 /* If we are closed, the bytes will have to remain here.
2300 * In time closedown will finish, we empty the write queue and
2301 * all will be happy.
2303 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2306 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2307 sk_gfp_atomic(sk
, GFP_ATOMIC
)))
2308 tcp_check_probe_timer(sk
);
2311 /* Send _single_ skb sitting at the send head. This function requires
2312 * true push pending frames to setup probe timer etc.
2314 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2316 struct sk_buff
*skb
= tcp_send_head(sk
);
2318 BUG_ON(!skb
|| skb
->len
< mss_now
);
2320 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2323 /* This function returns the amount that we can raise the
2324 * usable window based on the following constraints
2326 * 1. The window can never be shrunk once it is offered (RFC 793)
2327 * 2. We limit memory per socket
2330 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2331 * RECV.NEXT + RCV.WIN fixed until:
2332 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2334 * i.e. don't raise the right edge of the window until you can raise
2335 * it at least MSS bytes.
2337 * Unfortunately, the recommended algorithm breaks header prediction,
2338 * since header prediction assumes th->window stays fixed.
2340 * Strictly speaking, keeping th->window fixed violates the receiver
2341 * side SWS prevention criteria. The problem is that under this rule
2342 * a stream of single byte packets will cause the right side of the
2343 * window to always advance by a single byte.
2345 * Of course, if the sender implements sender side SWS prevention
2346 * then this will not be a problem.
2348 * BSD seems to make the following compromise:
2350 * If the free space is less than the 1/4 of the maximum
2351 * space available and the free space is less than 1/2 mss,
2352 * then set the window to 0.
2353 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2354 * Otherwise, just prevent the window from shrinking
2355 * and from being larger than the largest representable value.
2357 * This prevents incremental opening of the window in the regime
2358 * where TCP is limited by the speed of the reader side taking
2359 * data out of the TCP receive queue. It does nothing about
2360 * those cases where the window is constrained on the sender side
2361 * because the pipeline is full.
2363 * BSD also seems to "accidentally" limit itself to windows that are a
2364 * multiple of MSS, at least until the free space gets quite small.
2365 * This would appear to be a side effect of the mbuf implementation.
2366 * Combining these two algorithms results in the observed behavior
2367 * of having a fixed window size at almost all times.
2369 * Below we obtain similar behavior by forcing the offered window to
2370 * a multiple of the mss when it is feasible to do so.
2372 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2373 * Regular options like TIMESTAMP are taken into account.
2375 u32
__tcp_select_window(struct sock
*sk
)
2377 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2378 struct tcp_sock
*tp
= tcp_sk(sk
);
2379 /* MSS for the peer's data. Previous versions used mss_clamp
2380 * here. I don't know if the value based on our guesses
2381 * of peer's MSS is better for the performance. It's more correct
2382 * but may be worse for the performance because of rcv_mss
2383 * fluctuations. --SAW 1998/11/1
2385 int mss
= icsk
->icsk_ack
.rcv_mss
;
2386 int free_space
= tcp_space(sk
);
2387 int allowed_space
= tcp_full_space(sk
);
2388 int full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2391 if (unlikely(mss
> full_space
)) {
2396 if (free_space
< (full_space
>> 1)) {
2397 icsk
->icsk_ack
.quick
= 0;
2399 if (tcp_under_memory_pressure(sk
))
2400 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2403 /* free_space might become our new window, make sure we don't
2404 * increase it due to wscale.
2406 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2408 /* if free space is less than mss estimate, or is below 1/16th
2409 * of the maximum allowed, try to move to zero-window, else
2410 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2411 * new incoming data is dropped due to memory limits.
2412 * With large window, mss test triggers way too late in order
2413 * to announce zero window in time before rmem limit kicks in.
2415 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2419 if (free_space
> tp
->rcv_ssthresh
)
2420 free_space
= tp
->rcv_ssthresh
;
2422 /* Don't do rounding if we are using window scaling, since the
2423 * scaled window will not line up with the MSS boundary anyway.
2425 window
= tp
->rcv_wnd
;
2426 if (tp
->rx_opt
.rcv_wscale
) {
2427 window
= free_space
;
2429 /* Advertise enough space so that it won't get scaled away.
2430 * Import case: prevent zero window announcement if
2431 * 1<<rcv_wscale > mss.
2433 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2434 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2435 << tp
->rx_opt
.rcv_wscale
);
2437 /* Get the largest window that is a nice multiple of mss.
2438 * Window clamp already applied above.
2439 * If our current window offering is within 1 mss of the
2440 * free space we just keep it. This prevents the divide
2441 * and multiply from happening most of the time.
2442 * We also don't do any window rounding when the free space
2445 if (window
<= free_space
- mss
|| window
> free_space
)
2446 window
= (free_space
/ mss
) * mss
;
2447 else if (mss
== full_space
&&
2448 free_space
> window
+ (full_space
>> 1))
2449 window
= free_space
;
2455 /* Collapses two adjacent SKB's during retransmission. */
2456 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2458 struct tcp_sock
*tp
= tcp_sk(sk
);
2459 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2460 int skb_size
, next_skb_size
;
2462 skb_size
= skb
->len
;
2463 next_skb_size
= next_skb
->len
;
2465 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2467 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2469 tcp_unlink_write_queue(next_skb
, sk
);
2471 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
2474 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2475 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2477 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2478 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2480 /* Update sequence range on original skb. */
2481 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2483 /* Merge over control information. This moves PSH/FIN etc. over */
2484 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2486 /* All done, get rid of second SKB and account for it so
2487 * packet counting does not break.
2489 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2491 /* changed transmit queue under us so clear hints */
2492 tcp_clear_retrans_hints_partial(tp
);
2493 if (next_skb
== tp
->retransmit_skb_hint
)
2494 tp
->retransmit_skb_hint
= skb
;
2496 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2498 sk_wmem_free_skb(sk
, next_skb
);
2501 /* Check if coalescing SKBs is legal. */
2502 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2504 if (tcp_skb_pcount(skb
) > 1)
2506 /* TODO: SACK collapsing could be used to remove this condition */
2507 if (skb_shinfo(skb
)->nr_frags
!= 0)
2509 if (skb_cloned(skb
))
2511 if (skb
== tcp_send_head(sk
))
2513 /* Some heurestics for collapsing over SACK'd could be invented */
2514 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2520 /* Collapse packets in the retransmit queue to make to create
2521 * less packets on the wire. This is only done on retransmission.
2523 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2526 struct tcp_sock
*tp
= tcp_sk(sk
);
2527 struct sk_buff
*skb
= to
, *tmp
;
2530 if (!sysctl_tcp_retrans_collapse
)
2532 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2535 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2536 if (!tcp_can_collapse(sk
, skb
))
2548 /* Punt if not enough space exists in the first SKB for
2549 * the data in the second
2551 if (skb
->len
> skb_availroom(to
))
2554 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2557 tcp_collapse_retrans(sk
, to
);
2561 /* This retransmits one SKB. Policy decisions and retransmit queue
2562 * state updates are done by the caller. Returns non-zero if an
2563 * error occurred which prevented the send.
2565 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2567 struct tcp_sock
*tp
= tcp_sk(sk
);
2568 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2569 unsigned int cur_mss
;
2572 /* Inconslusive MTU probe */
2573 if (icsk
->icsk_mtup
.probe_size
) {
2574 icsk
->icsk_mtup
.probe_size
= 0;
2577 /* Do not sent more than we queued. 1/4 is reserved for possible
2578 * copying overhead: fragmentation, tunneling, mangling etc.
2580 if (atomic_read(&sk
->sk_wmem_alloc
) >
2581 min_t(u32
, sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2),
2585 if (skb_still_in_host_queue(sk
, skb
))
2588 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2589 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2591 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2595 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2596 return -EHOSTUNREACH
; /* Routing failure or similar. */
2598 cur_mss
= tcp_current_mss(sk
);
2600 /* If receiver has shrunk his window, and skb is out of
2601 * new window, do not retransmit it. The exception is the
2602 * case, when window is shrunk to zero. In this case
2603 * our retransmit serves as a zero window probe.
2605 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2606 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2609 if (skb
->len
> cur_mss
) {
2610 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
, GFP_ATOMIC
))
2611 return -ENOMEM
; /* We'll try again later. */
2613 int oldpcount
= tcp_skb_pcount(skb
);
2615 if (unlikely(oldpcount
> 1)) {
2616 if (skb_unclone(skb
, GFP_ATOMIC
))
2618 tcp_init_tso_segs(skb
, cur_mss
);
2619 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2623 /* RFC3168, section 6.1.1.1. ECN fallback */
2624 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN_ECN
) == TCPHDR_SYN_ECN
)
2625 tcp_ecn_clear_syn(sk
, skb
);
2627 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2629 /* Make a copy, if the first transmission SKB clone we made
2630 * is still in somebody's hands, else make a clone.
2633 /* make sure skb->data is aligned on arches that require it
2634 * and check if ack-trimming & collapsing extended the headroom
2635 * beyond what csum_start can cover.
2637 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2638 skb_headroom(skb
) >= 0xFFFF)) {
2639 struct sk_buff
*nskb
;
2641 skb_mstamp_get(&skb
->skb_mstamp
);
2642 nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
, GFP_ATOMIC
);
2643 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2646 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2650 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
2651 /* Update global TCP statistics. */
2652 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2653 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2654 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
2655 tp
->total_retrans
++;
2660 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2662 struct tcp_sock
*tp
= tcp_sk(sk
);
2663 int err
= __tcp_retransmit_skb(sk
, skb
);
2666 #if FASTRETRANS_DEBUG > 0
2667 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2668 net_dbg_ratelimited("retrans_out leaked\n");
2671 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2672 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2674 /* Save stamp of the first retransmit. */
2675 if (!tp
->retrans_stamp
)
2676 tp
->retrans_stamp
= tcp_skb_timestamp(skb
);
2678 } else if (err
!= -EBUSY
) {
2679 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2682 if (tp
->undo_retrans
< 0)
2683 tp
->undo_retrans
= 0;
2684 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2688 /* Check if we forward retransmits are possible in the current
2689 * window/congestion state.
2691 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2693 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2694 const struct tcp_sock
*tp
= tcp_sk(sk
);
2696 /* Forward retransmissions are possible only during Recovery. */
2697 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2700 /* No forward retransmissions in Reno are possible. */
2701 if (tcp_is_reno(tp
))
2704 /* Yeah, we have to make difficult choice between forward transmission
2705 * and retransmission... Both ways have their merits...
2707 * For now we do not retransmit anything, while we have some new
2708 * segments to send. In the other cases, follow rule 3 for
2709 * NextSeg() specified in RFC3517.
2712 if (tcp_may_send_now(sk
))
2718 /* This gets called after a retransmit timeout, and the initially
2719 * retransmitted data is acknowledged. It tries to continue
2720 * resending the rest of the retransmit queue, until either
2721 * we've sent it all or the congestion window limit is reached.
2722 * If doing SACK, the first ACK which comes back for a timeout
2723 * based retransmit packet might feed us FACK information again.
2724 * If so, we use it to avoid unnecessarily retransmissions.
2726 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2728 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2729 struct tcp_sock
*tp
= tcp_sk(sk
);
2730 struct sk_buff
*skb
;
2731 struct sk_buff
*hole
= NULL
;
2734 int fwd_rexmitting
= 0;
2736 if (!tp
->packets_out
)
2740 tp
->retransmit_high
= tp
->snd_una
;
2742 if (tp
->retransmit_skb_hint
) {
2743 skb
= tp
->retransmit_skb_hint
;
2744 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2745 if (after(last_lost
, tp
->retransmit_high
))
2746 last_lost
= tp
->retransmit_high
;
2748 skb
= tcp_write_queue_head(sk
);
2749 last_lost
= tp
->snd_una
;
2752 tcp_for_write_queue_from(skb
, sk
) {
2753 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2755 if (skb
== tcp_send_head(sk
))
2757 /* we could do better than to assign each time */
2759 tp
->retransmit_skb_hint
= skb
;
2761 /* Assume this retransmit will generate
2762 * only one packet for congestion window
2763 * calculation purposes. This works because
2764 * tcp_retransmit_skb() will chop up the
2765 * packet to be MSS sized and all the
2766 * packet counting works out.
2768 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2771 if (fwd_rexmitting
) {
2773 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2775 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2777 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2778 tp
->retransmit_high
= last_lost
;
2779 if (!tcp_can_forward_retransmit(sk
))
2781 /* Backtrack if necessary to non-L'ed skb */
2789 } else if (!(sacked
& TCPCB_LOST
)) {
2790 if (!hole
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2795 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2796 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2797 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2799 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2802 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2805 if (tcp_retransmit_skb(sk
, skb
))
2808 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2810 if (tcp_in_cwnd_reduction(sk
))
2811 tp
->prr_out
+= tcp_skb_pcount(skb
);
2813 if (skb
== tcp_write_queue_head(sk
))
2814 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2815 inet_csk(sk
)->icsk_rto
,
2820 /* We allow to exceed memory limits for FIN packets to expedite
2821 * connection tear down and (memory) recovery.
2822 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2823 * or even be forced to close flow without any FIN.
2824 * In general, we want to allow one skb per socket to avoid hangs
2825 * with edge trigger epoll()
2827 void sk_forced_mem_schedule(struct sock
*sk
, int size
)
2831 if (size
<= sk
->sk_forward_alloc
)
2833 amt
= sk_mem_pages(size
);
2834 sk
->sk_forward_alloc
+= amt
* SK_MEM_QUANTUM
;
2835 sk_memory_allocated_add(sk
, amt
, &status
);
2838 /* Send a FIN. The caller locks the socket for us.
2839 * We should try to send a FIN packet really hard, but eventually give up.
2841 void tcp_send_fin(struct sock
*sk
)
2843 struct sk_buff
*skb
, *tskb
= tcp_write_queue_tail(sk
);
2844 struct tcp_sock
*tp
= tcp_sk(sk
);
2846 /* Optimization, tack on the FIN if we have one skb in write queue and
2847 * this skb was not yet sent, or we are under memory pressure.
2848 * Note: in the latter case, FIN packet will be sent after a timeout,
2849 * as TCP stack thinks it has already been transmitted.
2851 if (tskb
&& (tcp_send_head(sk
) || tcp_under_memory_pressure(sk
))) {
2853 TCP_SKB_CB(tskb
)->tcp_flags
|= TCPHDR_FIN
;
2854 TCP_SKB_CB(tskb
)->end_seq
++;
2856 if (!tcp_send_head(sk
)) {
2857 /* This means tskb was already sent.
2858 * Pretend we included the FIN on previous transmit.
2859 * We need to set tp->snd_nxt to the value it would have
2860 * if FIN had been sent. This is because retransmit path
2861 * does not change tp->snd_nxt.
2867 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, sk
->sk_allocation
);
2868 if (unlikely(!skb
)) {
2873 skb_reserve(skb
, MAX_TCP_HEADER
);
2874 sk_forced_mem_schedule(sk
, skb
->truesize
);
2875 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2876 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2877 TCPHDR_ACK
| TCPHDR_FIN
);
2878 tcp_queue_skb(sk
, skb
);
2880 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
), TCP_NAGLE_OFF
);
2883 /* We get here when a process closes a file descriptor (either due to
2884 * an explicit close() or as a byproduct of exit()'ing) and there
2885 * was unread data in the receive queue. This behavior is recommended
2886 * by RFC 2525, section 2.17. -DaveM
2888 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2890 struct sk_buff
*skb
;
2892 /* NOTE: No TCP options attached and we never retransmit this. */
2893 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2895 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2899 /* Reserve space for headers and prepare control bits. */
2900 skb_reserve(skb
, MAX_TCP_HEADER
);
2901 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2902 TCPHDR_ACK
| TCPHDR_RST
);
2903 skb_mstamp_get(&skb
->skb_mstamp
);
2905 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2906 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2908 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2911 /* Send a crossed SYN-ACK during socket establishment.
2912 * WARNING: This routine must only be called when we have already sent
2913 * a SYN packet that crossed the incoming SYN that caused this routine
2914 * to get called. If this assumption fails then the initial rcv_wnd
2915 * and rcv_wscale values will not be correct.
2917 int tcp_send_synack(struct sock
*sk
)
2919 struct sk_buff
*skb
;
2921 skb
= tcp_write_queue_head(sk
);
2922 if (!skb
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2923 pr_debug("%s: wrong queue state\n", __func__
);
2926 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
2927 if (skb_cloned(skb
)) {
2928 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2931 tcp_unlink_write_queue(skb
, sk
);
2932 __skb_header_release(nskb
);
2933 __tcp_add_write_queue_head(sk
, nskb
);
2934 sk_wmem_free_skb(sk
, skb
);
2935 sk
->sk_wmem_queued
+= nskb
->truesize
;
2936 sk_mem_charge(sk
, nskb
->truesize
);
2940 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
2941 tcp_ecn_send_synack(sk
, skb
);
2943 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2947 * tcp_make_synack - Prepare a SYN-ACK.
2948 * sk: listener socket
2949 * dst: dst entry attached to the SYNACK
2950 * req: request_sock pointer
2952 * Allocate one skb and build a SYNACK packet.
2953 * @dst is consumed : Caller should not use it again.
2955 struct sk_buff
*tcp_make_synack(const struct sock
*sk
, struct dst_entry
*dst
,
2956 struct request_sock
*req
,
2957 struct tcp_fastopen_cookie
*foc
,
2960 struct inet_request_sock
*ireq
= inet_rsk(req
);
2961 const struct tcp_sock
*tp
= tcp_sk(sk
);
2962 struct tcp_md5sig_key
*md5
= NULL
;
2963 struct tcp_out_options opts
;
2964 struct sk_buff
*skb
;
2965 int tcp_header_size
;
2970 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
2971 if (unlikely(!skb
)) {
2975 /* Reserve space for headers. */
2976 skb_reserve(skb
, MAX_TCP_HEADER
);
2979 skb_set_owner_w(skb
, req_to_sk(req
));
2981 /* sk is a const pointer, because we want to express multiple
2982 * cpu might call us concurrently.
2983 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
2985 skb_set_owner_w(skb
, (struct sock
*)sk
);
2987 skb_dst_set(skb
, dst
);
2989 mss
= dst_metric_advmss(dst
);
2990 user_mss
= READ_ONCE(tp
->rx_opt
.user_mss
);
2991 if (user_mss
&& user_mss
< mss
)
2994 memset(&opts
, 0, sizeof(opts
));
2995 #ifdef CONFIG_SYN_COOKIES
2996 if (unlikely(req
->cookie_ts
))
2997 skb
->skb_mstamp
.stamp_jiffies
= cookie_init_timestamp(req
);
3000 skb_mstamp_get(&skb
->skb_mstamp
);
3002 #ifdef CONFIG_TCP_MD5SIG
3004 md5
= tcp_rsk(req
)->af_specific
->req_md5_lookup(sk
, req_to_sk(req
));
3006 skb_set_hash(skb
, tcp_rsk(req
)->txhash
, PKT_HASH_TYPE_L4
);
3007 tcp_header_size
= tcp_synack_options(req
, mss
, skb
, &opts
, md5
, foc
) +
3010 skb_push(skb
, tcp_header_size
);
3011 skb_reset_transport_header(skb
);
3014 memset(th
, 0, sizeof(struct tcphdr
));
3017 tcp_ecn_make_synack(req
, th
);
3018 th
->source
= htons(ireq
->ir_num
);
3019 th
->dest
= ireq
->ir_rmt_port
;
3020 /* Setting of flags are superfluous here for callers (and ECE is
3021 * not even correctly set)
3023 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
3024 TCPHDR_SYN
| TCPHDR_ACK
);
3026 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
3027 /* XXX data is queued and acked as is. No buffer/window check */
3028 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
3030 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3031 th
->window
= htons(min(req
->rsk_rcv_wnd
, 65535U));
3032 tcp_options_write((__be32
*)(th
+ 1), NULL
, &opts
);
3033 th
->doff
= (tcp_header_size
>> 2);
3034 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_OUTSEGS
);
3036 #ifdef CONFIG_TCP_MD5SIG
3037 /* Okay, we have all we need - do the md5 hash if needed */
3039 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
3040 md5
, req_to_sk(req
), skb
);
3044 /* Do not fool tcpdump (if any), clean our debris */
3045 skb
->tstamp
.tv64
= 0;
3048 EXPORT_SYMBOL(tcp_make_synack
);
3050 static void tcp_ca_dst_init(struct sock
*sk
, const struct dst_entry
*dst
)
3052 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3053 const struct tcp_congestion_ops
*ca
;
3054 u32 ca_key
= dst_metric(dst
, RTAX_CC_ALGO
);
3056 if (ca_key
== TCP_CA_UNSPEC
)
3060 ca
= tcp_ca_find_key(ca_key
);
3061 if (likely(ca
&& try_module_get(ca
->owner
))) {
3062 module_put(icsk
->icsk_ca_ops
->owner
);
3063 icsk
->icsk_ca_dst_locked
= tcp_ca_dst_locked(dst
);
3064 icsk
->icsk_ca_ops
= ca
;
3069 /* Do all connect socket setups that can be done AF independent. */
3070 static void tcp_connect_init(struct sock
*sk
)
3072 const struct dst_entry
*dst
= __sk_dst_get(sk
);
3073 struct tcp_sock
*tp
= tcp_sk(sk
);
3076 /* We'll fix this up when we get a response from the other end.
3077 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3079 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
3080 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
3082 #ifdef CONFIG_TCP_MD5SIG
3083 if (tp
->af_specific
->md5_lookup(sk
, sk
))
3084 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
3087 /* If user gave his TCP_MAXSEG, record it to clamp */
3088 if (tp
->rx_opt
.user_mss
)
3089 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3092 tcp_sync_mss(sk
, dst_mtu(dst
));
3094 tcp_ca_dst_init(sk
, dst
);
3096 if (!tp
->window_clamp
)
3097 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
3098 tp
->advmss
= dst_metric_advmss(dst
);
3099 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
3100 tp
->advmss
= tp
->rx_opt
.user_mss
;
3102 tcp_initialize_rcv_mss(sk
);
3104 /* limit the window selection if the user enforce a smaller rx buffer */
3105 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
3106 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
3107 tp
->window_clamp
= tcp_full_space(sk
);
3109 tcp_select_initial_window(tcp_full_space(sk
),
3110 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
3113 sysctl_tcp_window_scaling
,
3115 dst_metric(dst
, RTAX_INITRWND
));
3117 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
3118 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
3121 sock_reset_flag(sk
, SOCK_DONE
);
3124 tp
->snd_una
= tp
->write_seq
;
3125 tp
->snd_sml
= tp
->write_seq
;
3126 tp
->snd_up
= tp
->write_seq
;
3127 tp
->snd_nxt
= tp
->write_seq
;
3129 if (likely(!tp
->repair
))
3132 tp
->rcv_tstamp
= tcp_time_stamp
;
3133 tp
->rcv_wup
= tp
->rcv_nxt
;
3134 tp
->copied_seq
= tp
->rcv_nxt
;
3136 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
3137 inet_csk(sk
)->icsk_retransmits
= 0;
3138 tcp_clear_retrans(tp
);
3141 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
3143 struct tcp_sock
*tp
= tcp_sk(sk
);
3144 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
3146 tcb
->end_seq
+= skb
->len
;
3147 __skb_header_release(skb
);
3148 __tcp_add_write_queue_tail(sk
, skb
);
3149 sk
->sk_wmem_queued
+= skb
->truesize
;
3150 sk_mem_charge(sk
, skb
->truesize
);
3151 tp
->write_seq
= tcb
->end_seq
;
3152 tp
->packets_out
+= tcp_skb_pcount(skb
);
3155 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3156 * queue a data-only packet after the regular SYN, such that regular SYNs
3157 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3158 * only the SYN sequence, the data are retransmitted in the first ACK.
3159 * If cookie is not cached or other error occurs, falls back to send a
3160 * regular SYN with Fast Open cookie request option.
3162 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
3164 struct tcp_sock
*tp
= tcp_sk(sk
);
3165 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
3166 int syn_loss
= 0, space
, err
= 0;
3167 unsigned long last_syn_loss
= 0;
3168 struct sk_buff
*syn_data
;
3170 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
3171 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
3172 &syn_loss
, &last_syn_loss
);
3173 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3175 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
3176 fo
->cookie
.len
= -1;
3180 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
3181 fo
->cookie
.len
= -1;
3182 else if (fo
->cookie
.len
<= 0)
3185 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3186 * user-MSS. Reserve maximum option space for middleboxes that add
3187 * private TCP options. The cost is reduced data space in SYN :(
3189 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
3190 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3191 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
3192 MAX_TCP_OPTION_SPACE
;
3194 space
= min_t(size_t, space
, fo
->size
);
3196 /* limit to order-0 allocations */
3197 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
3199 syn_data
= sk_stream_alloc_skb(sk
, space
, sk
->sk_allocation
, false);
3202 syn_data
->ip_summed
= CHECKSUM_PARTIAL
;
3203 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
3205 int copied
= copy_from_iter(skb_put(syn_data
, space
), space
,
3206 &fo
->data
->msg_iter
);
3207 if (unlikely(!copied
)) {
3208 kfree_skb(syn_data
);
3211 if (copied
!= space
) {
3212 skb_trim(syn_data
, copied
);
3216 /* No more data pending in inet_wait_for_connect() */
3217 if (space
== fo
->size
)
3221 tcp_connect_queue_skb(sk
, syn_data
);
3223 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
3225 syn
->skb_mstamp
= syn_data
->skb_mstamp
;
3227 /* Now full SYN+DATA was cloned and sent (or not),
3228 * remove the SYN from the original skb (syn_data)
3229 * we keep in write queue in case of a retransmit, as we
3230 * also have the SYN packet (with no data) in the same queue.
3232 TCP_SKB_CB(syn_data
)->seq
++;
3233 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
3235 tp
->syn_data
= (fo
->copied
> 0);
3236 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3241 /* Send a regular SYN with Fast Open cookie request option */
3242 if (fo
->cookie
.len
> 0)
3244 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3246 tp
->syn_fastopen
= 0;
3248 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3252 /* Build a SYN and send it off. */
3253 int tcp_connect(struct sock
*sk
)
3255 struct tcp_sock
*tp
= tcp_sk(sk
);
3256 struct sk_buff
*buff
;
3259 tcp_connect_init(sk
);
3261 if (unlikely(tp
->repair
)) {
3262 tcp_finish_connect(sk
, NULL
);
3266 buff
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
, true);
3267 if (unlikely(!buff
))
3270 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3271 tp
->retrans_stamp
= tcp_time_stamp
;
3272 tcp_connect_queue_skb(sk
, buff
);
3273 tcp_ecn_send_syn(sk
, buff
);
3275 /* Send off SYN; include data in Fast Open. */
3276 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3277 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3278 if (err
== -ECONNREFUSED
)
3281 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3282 * in order to make this packet get counted in tcpOutSegs.
3284 tp
->snd_nxt
= tp
->write_seq
;
3285 tp
->pushed_seq
= tp
->write_seq
;
3286 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3288 /* Timer for repeating the SYN until an answer. */
3289 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3290 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3293 EXPORT_SYMBOL(tcp_connect
);
3295 /* Send out a delayed ack, the caller does the policy checking
3296 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3299 void tcp_send_delayed_ack(struct sock
*sk
)
3301 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3302 int ato
= icsk
->icsk_ack
.ato
;
3303 unsigned long timeout
;
3305 tcp_ca_event(sk
, CA_EVENT_DELAYED_ACK
);
3307 if (ato
> TCP_DELACK_MIN
) {
3308 const struct tcp_sock
*tp
= tcp_sk(sk
);
3309 int max_ato
= HZ
/ 2;
3311 if (icsk
->icsk_ack
.pingpong
||
3312 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3313 max_ato
= TCP_DELACK_MAX
;
3315 /* Slow path, intersegment interval is "high". */
3317 /* If some rtt estimate is known, use it to bound delayed ack.
3318 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3322 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3329 ato
= min(ato
, max_ato
);
3332 /* Stay within the limit we were given */
3333 timeout
= jiffies
+ ato
;
3335 /* Use new timeout only if there wasn't a older one earlier. */
3336 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3337 /* If delack timer was blocked or is about to expire,
3340 if (icsk
->icsk_ack
.blocked
||
3341 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3346 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3347 timeout
= icsk
->icsk_ack
.timeout
;
3349 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3350 icsk
->icsk_ack
.timeout
= timeout
;
3351 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3354 /* This routine sends an ack and also updates the window. */
3355 void tcp_send_ack(struct sock
*sk
)
3357 struct sk_buff
*buff
;
3359 /* If we have been reset, we may not send again. */
3360 if (sk
->sk_state
== TCP_CLOSE
)
3363 tcp_ca_event(sk
, CA_EVENT_NON_DELAYED_ACK
);
3365 /* We are not putting this on the write queue, so
3366 * tcp_transmit_skb() will set the ownership to this
3369 buff
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3371 inet_csk_schedule_ack(sk
);
3372 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3373 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3374 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3378 /* Reserve space for headers and prepare control bits. */
3379 skb_reserve(buff
, MAX_TCP_HEADER
);
3380 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3382 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3384 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3385 * We also avoid tcp_wfree() overhead (cache line miss accessing
3386 * tp->tsq_flags) by using regular sock_wfree()
3388 skb_set_tcp_pure_ack(buff
);
3390 /* Send it off, this clears delayed acks for us. */
3391 skb_mstamp_get(&buff
->skb_mstamp
);
3392 tcp_transmit_skb(sk
, buff
, 0, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3394 EXPORT_SYMBOL_GPL(tcp_send_ack
);
3396 /* This routine sends a packet with an out of date sequence
3397 * number. It assumes the other end will try to ack it.
3399 * Question: what should we make while urgent mode?
3400 * 4.4BSD forces sending single byte of data. We cannot send
3401 * out of window data, because we have SND.NXT==SND.MAX...
3403 * Current solution: to send TWO zero-length segments in urgent mode:
3404 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3405 * out-of-date with SND.UNA-1 to probe window.
3407 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
, int mib
)
3409 struct tcp_sock
*tp
= tcp_sk(sk
);
3410 struct sk_buff
*skb
;
3412 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3413 skb
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3417 /* Reserve space for headers and set control bits. */
3418 skb_reserve(skb
, MAX_TCP_HEADER
);
3419 /* Use a previous sequence. This should cause the other
3420 * end to send an ack. Don't queue or clone SKB, just
3423 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3424 skb_mstamp_get(&skb
->skb_mstamp
);
3425 NET_INC_STATS(sock_net(sk
), mib
);
3426 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
3429 void tcp_send_window_probe(struct sock
*sk
)
3431 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3432 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3433 tcp_xmit_probe_skb(sk
, 0, LINUX_MIB_TCPWINPROBE
);
3437 /* Initiate keepalive or window probe from timer. */
3438 int tcp_write_wakeup(struct sock
*sk
, int mib
)
3440 struct tcp_sock
*tp
= tcp_sk(sk
);
3441 struct sk_buff
*skb
;
3443 if (sk
->sk_state
== TCP_CLOSE
)
3446 skb
= tcp_send_head(sk
);
3447 if (skb
&& before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3449 unsigned int mss
= tcp_current_mss(sk
);
3450 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3452 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3453 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3455 /* We are probing the opening of a window
3456 * but the window size is != 0
3457 * must have been a result SWS avoidance ( sender )
3459 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3461 seg_size
= min(seg_size
, mss
);
3462 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3463 if (tcp_fragment(sk
, skb
, seg_size
, mss
, GFP_ATOMIC
))
3465 } else if (!tcp_skb_pcount(skb
))
3466 tcp_set_skb_tso_segs(skb
, mss
);
3468 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3469 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3471 tcp_event_new_data_sent(sk
, skb
);
3474 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3475 tcp_xmit_probe_skb(sk
, 1, mib
);
3476 return tcp_xmit_probe_skb(sk
, 0, mib
);
3480 /* A window probe timeout has occurred. If window is not closed send
3481 * a partial packet else a zero probe.
3483 void tcp_send_probe0(struct sock
*sk
)
3485 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3486 struct tcp_sock
*tp
= tcp_sk(sk
);
3487 unsigned long probe_max
;
3490 err
= tcp_write_wakeup(sk
, LINUX_MIB_TCPWINPROBE
);
3492 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3493 /* Cancel probe timer, if it is not required. */
3494 icsk
->icsk_probes_out
= 0;
3495 icsk
->icsk_backoff
= 0;
3500 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
3501 icsk
->icsk_backoff
++;
3502 icsk
->icsk_probes_out
++;
3503 probe_max
= TCP_RTO_MAX
;
3505 /* If packet was not sent due to local congestion,
3506 * do not backoff and do not remember icsk_probes_out.
3507 * Let local senders to fight for local resources.
3509 * Use accumulated backoff yet.
3511 if (!icsk
->icsk_probes_out
)
3512 icsk
->icsk_probes_out
= 1;
3513 probe_max
= TCP_RESOURCE_PROBE_INTERVAL
;
3515 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3516 tcp_probe0_when(sk
, probe_max
),
3520 int tcp_rtx_synack(const struct sock
*sk
, struct request_sock
*req
)
3522 const struct tcp_request_sock_ops
*af_ops
= tcp_rsk(req
)->af_specific
;
3526 tcp_rsk(req
)->txhash
= net_tx_rndhash();
3527 res
= af_ops
->send_synack(sk
, NULL
, &fl
, req
, NULL
, true);
3529 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
3530 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
3534 EXPORT_SYMBOL(tcp_rtx_synack
);