2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
53 /* Default TSQ limit of four TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly
= 262144;
56 /* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
60 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
65 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
66 int push_one
, gfp_t gfp
);
68 /* Account for new data that has been sent to the network. */
69 static void tcp_event_new_data_sent(struct sock
*sk
, const struct sk_buff
*skb
)
71 struct inet_connection_sock
*icsk
= inet_csk(sk
);
72 struct tcp_sock
*tp
= tcp_sk(sk
);
73 unsigned int prior_packets
= tp
->packets_out
;
75 tcp_advance_send_head(sk
, skb
);
76 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
78 tp
->packets_out
+= tcp_skb_pcount(skb
);
79 if (!prior_packets
|| icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
82 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
,
86 /* SND.NXT, if window was not shrunk.
87 * If window has been shrunk, what should we make? It is not clear at all.
88 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
89 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
90 * invalid. OK, let's make this for now:
92 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
94 const struct tcp_sock
*tp
= tcp_sk(sk
);
96 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
99 return tcp_wnd_end(tp
);
102 /* Calculate mss to advertise in SYN segment.
103 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
105 * 1. It is independent of path mtu.
106 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
107 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
108 * attached devices, because some buggy hosts are confused by
110 * 4. We do not make 3, we advertise MSS, calculated from first
111 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
112 * This may be overridden via information stored in routing table.
113 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
114 * probably even Jumbo".
116 static __u16
tcp_advertise_mss(struct sock
*sk
)
118 struct tcp_sock
*tp
= tcp_sk(sk
);
119 const struct dst_entry
*dst
= __sk_dst_get(sk
);
120 int mss
= tp
->advmss
;
123 unsigned int metric
= dst_metric_advmss(dst
);
134 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
135 * This is the first part of cwnd validation mechanism.
137 void tcp_cwnd_restart(struct sock
*sk
, s32 delta
)
139 struct tcp_sock
*tp
= tcp_sk(sk
);
140 u32 restart_cwnd
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
141 u32 cwnd
= tp
->snd_cwnd
;
143 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
145 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
146 restart_cwnd
= min(restart_cwnd
, cwnd
);
148 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
150 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
151 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
152 tp
->snd_cwnd_used
= 0;
155 /* Congestion state accounting after a packet has been sent. */
156 static void tcp_event_data_sent(struct tcp_sock
*tp
,
159 struct inet_connection_sock
*icsk
= inet_csk(sk
);
160 const u32 now
= tcp_time_stamp
;
162 if (tcp_packets_in_flight(tp
) == 0)
163 tcp_ca_event(sk
, CA_EVENT_TX_START
);
167 /* If it is a reply for ato after last received
168 * packet, enter pingpong mode.
170 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
171 icsk
->icsk_ack
.pingpong
= 1;
174 /* Account for an ACK we sent. */
175 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
177 tcp_dec_quickack_mode(sk
, pkts
);
178 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
182 u32
tcp_default_init_rwnd(u32 mss
)
184 /* Initial receive window should be twice of TCP_INIT_CWND to
185 * enable proper sending of new unsent data during fast recovery
186 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
187 * limit when mss is larger than 1460.
189 u32 init_rwnd
= TCP_INIT_CWND
* 2;
192 init_rwnd
= max((1460 * init_rwnd
) / mss
, 2U);
196 /* Determine a window scaling and initial window to offer.
197 * Based on the assumption that the given amount of space
198 * will be offered. Store the results in the tp structure.
199 * NOTE: for smooth operation initial space offering should
200 * be a multiple of mss if possible. We assume here that mss >= 1.
201 * This MUST be enforced by all callers.
203 void tcp_select_initial_window(int __space
, __u32 mss
,
204 __u32
*rcv_wnd
, __u32
*window_clamp
,
205 int wscale_ok
, __u8
*rcv_wscale
,
208 unsigned int space
= (__space
< 0 ? 0 : __space
);
210 /* If no clamp set the clamp to the max possible scaled window */
211 if (*window_clamp
== 0)
212 (*window_clamp
) = (65535 << 14);
213 space
= min(*window_clamp
, space
);
215 /* Quantize space offering to a multiple of mss if possible. */
217 space
= (space
/ mss
) * mss
;
219 /* NOTE: offering an initial window larger than 32767
220 * will break some buggy TCP stacks. If the admin tells us
221 * it is likely we could be speaking with such a buggy stack
222 * we will truncate our initial window offering to 32K-1
223 * unless the remote has sent us a window scaling option,
224 * which we interpret as a sign the remote TCP is not
225 * misinterpreting the window field as a signed quantity.
227 if (sysctl_tcp_workaround_signed_windows
)
228 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
234 /* Set window scaling on max possible window
235 * See RFC1323 for an explanation of the limit to 14
237 space
= max_t(u32
, space
, sysctl_tcp_rmem
[2]);
238 space
= max_t(u32
, space
, sysctl_rmem_max
);
239 space
= min_t(u32
, space
, *window_clamp
);
240 while (space
> 65535 && (*rcv_wscale
) < 14) {
246 if (mss
> (1 << *rcv_wscale
)) {
247 if (!init_rcv_wnd
) /* Use default unless specified otherwise */
248 init_rcv_wnd
= tcp_default_init_rwnd(mss
);
249 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
252 /* Set the clamp no higher than max representable value */
253 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
255 EXPORT_SYMBOL(tcp_select_initial_window
);
257 /* Chose a new window to advertise, update state in tcp_sock for the
258 * socket, and return result with RFC1323 scaling applied. The return
259 * value can be stuffed directly into th->window for an outgoing
262 static u16
tcp_select_window(struct sock
*sk
)
264 struct tcp_sock
*tp
= tcp_sk(sk
);
265 u32 old_win
= tp
->rcv_wnd
;
266 u32 cur_win
= tcp_receive_window(tp
);
267 u32 new_win
= __tcp_select_window(sk
);
269 /* Never shrink the offered window */
270 if (new_win
< cur_win
) {
271 /* Danger Will Robinson!
272 * Don't update rcv_wup/rcv_wnd here or else
273 * we will not be able to advertise a zero
274 * window in time. --DaveM
276 * Relax Will Robinson.
279 NET_INC_STATS(sock_net(sk
),
280 LINUX_MIB_TCPWANTZEROWINDOWADV
);
281 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
283 tp
->rcv_wnd
= new_win
;
284 tp
->rcv_wup
= tp
->rcv_nxt
;
286 /* Make sure we do not exceed the maximum possible
289 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
290 new_win
= min(new_win
, MAX_TCP_WINDOW
);
292 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
294 /* RFC1323 scaling applied */
295 new_win
>>= tp
->rx_opt
.rcv_wscale
;
297 /* If we advertise zero window, disable fast path. */
301 NET_INC_STATS(sock_net(sk
),
302 LINUX_MIB_TCPTOZEROWINDOWADV
);
303 } else if (old_win
== 0) {
304 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFROMZEROWINDOWADV
);
310 /* Packet ECN state for a SYN-ACK */
311 static void tcp_ecn_send_synack(struct sock
*sk
, struct sk_buff
*skb
)
313 const struct tcp_sock
*tp
= tcp_sk(sk
);
315 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
316 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
317 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
318 else if (tcp_ca_needs_ecn(sk
))
322 /* Packet ECN state for a SYN. */
323 static void tcp_ecn_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
325 struct tcp_sock
*tp
= tcp_sk(sk
);
326 bool use_ecn
= sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1 ||
327 tcp_ca_needs_ecn(sk
);
330 const struct dst_entry
*dst
= __sk_dst_get(sk
);
332 if (dst
&& dst_feature(dst
, RTAX_FEATURE_ECN
))
339 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
340 tp
->ecn_flags
= TCP_ECN_OK
;
341 if (tcp_ca_needs_ecn(sk
))
346 static void tcp_ecn_clear_syn(struct sock
*sk
, struct sk_buff
*skb
)
348 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn_fallback
)
349 /* tp->ecn_flags are cleared at a later point in time when
350 * SYN ACK is ultimatively being received.
352 TCP_SKB_CB(skb
)->tcp_flags
&= ~(TCPHDR_ECE
| TCPHDR_CWR
);
356 tcp_ecn_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
358 if (inet_rsk(req
)->ecn_ok
)
362 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
365 static void tcp_ecn_send(struct sock
*sk
, struct sk_buff
*skb
,
366 struct tcphdr
*th
, int tcp_header_len
)
368 struct tcp_sock
*tp
= tcp_sk(sk
);
370 if (tp
->ecn_flags
& TCP_ECN_OK
) {
371 /* Not-retransmitted data segment: set ECT and inject CWR. */
372 if (skb
->len
!= tcp_header_len
&&
373 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
375 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
376 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
378 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
380 } else if (!tcp_ca_needs_ecn(sk
)) {
381 /* ACK or retransmitted segment: clear ECT|CE */
382 INET_ECN_dontxmit(sk
);
384 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
389 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
390 * auto increment end seqno.
392 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
394 skb
->ip_summed
= CHECKSUM_PARTIAL
;
397 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
398 TCP_SKB_CB(skb
)->sacked
= 0;
400 tcp_skb_pcount_set(skb
, 1);
402 TCP_SKB_CB(skb
)->seq
= seq
;
403 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
405 TCP_SKB_CB(skb
)->end_seq
= seq
;
408 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
410 return tp
->snd_una
!= tp
->snd_up
;
413 #define OPTION_SACK_ADVERTISE (1 << 0)
414 #define OPTION_TS (1 << 1)
415 #define OPTION_MD5 (1 << 2)
416 #define OPTION_WSCALE (1 << 3)
417 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
419 struct tcp_out_options
{
420 u16 options
; /* bit field of OPTION_* */
421 u16 mss
; /* 0 to disable */
422 u8 ws
; /* window scale, 0 to disable */
423 u8 num_sack_blocks
; /* number of SACK blocks to include */
424 u8 hash_size
; /* bytes in hash_location */
425 __u8
*hash_location
; /* temporary pointer, overloaded */
426 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
427 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
430 /* Write previously computed TCP options to the packet.
432 * Beware: Something in the Internet is very sensitive to the ordering of
433 * TCP options, we learned this through the hard way, so be careful here.
434 * Luckily we can at least blame others for their non-compliance but from
435 * inter-operability perspective it seems that we're somewhat stuck with
436 * the ordering which we have been using if we want to keep working with
437 * those broken things (not that it currently hurts anybody as there isn't
438 * particular reason why the ordering would need to be changed).
440 * At least SACK_PERM as the first option is known to lead to a disaster
441 * (but it may well be that other scenarios fail similarly).
443 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
444 struct tcp_out_options
*opts
)
446 u16 options
= opts
->options
; /* mungable copy */
448 if (unlikely(OPTION_MD5
& options
)) {
449 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
450 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
451 /* overload cookie hash location */
452 opts
->hash_location
= (__u8
*)ptr
;
456 if (unlikely(opts
->mss
)) {
457 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
458 (TCPOLEN_MSS
<< 16) |
462 if (likely(OPTION_TS
& options
)) {
463 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
464 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
465 (TCPOLEN_SACK_PERM
<< 16) |
466 (TCPOPT_TIMESTAMP
<< 8) |
468 options
&= ~OPTION_SACK_ADVERTISE
;
470 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
472 (TCPOPT_TIMESTAMP
<< 8) |
475 *ptr
++ = htonl(opts
->tsval
);
476 *ptr
++ = htonl(opts
->tsecr
);
479 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
480 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
482 (TCPOPT_SACK_PERM
<< 8) |
486 if (unlikely(OPTION_WSCALE
& options
)) {
487 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
488 (TCPOPT_WINDOW
<< 16) |
489 (TCPOLEN_WINDOW
<< 8) |
493 if (unlikely(opts
->num_sack_blocks
)) {
494 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
495 tp
->duplicate_sack
: tp
->selective_acks
;
498 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
501 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
502 TCPOLEN_SACK_PERBLOCK
)));
504 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
506 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
507 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
510 tp
->rx_opt
.dsack
= 0;
513 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
514 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
516 u32 len
; /* Fast Open option length */
519 len
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
520 *ptr
= htonl((TCPOPT_EXP
<< 24) | (len
<< 16) |
521 TCPOPT_FASTOPEN_MAGIC
);
522 p
+= TCPOLEN_EXP_FASTOPEN_BASE
;
524 len
= TCPOLEN_FASTOPEN_BASE
+ foc
->len
;
525 *p
++ = TCPOPT_FASTOPEN
;
529 memcpy(p
, foc
->val
, foc
->len
);
530 if ((len
& 3) == 2) {
531 p
[foc
->len
] = TCPOPT_NOP
;
532 p
[foc
->len
+ 1] = TCPOPT_NOP
;
534 ptr
+= (len
+ 3) >> 2;
538 /* Compute TCP options for SYN packets. This is not the final
539 * network wire format yet.
541 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
542 struct tcp_out_options
*opts
,
543 struct tcp_md5sig_key
**md5
)
545 struct tcp_sock
*tp
= tcp_sk(sk
);
546 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
547 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
549 #ifdef CONFIG_TCP_MD5SIG
550 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
552 opts
->options
|= OPTION_MD5
;
553 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
559 /* We always get an MSS option. The option bytes which will be seen in
560 * normal data packets should timestamps be used, must be in the MSS
561 * advertised. But we subtract them from tp->mss_cache so that
562 * calculations in tcp_sendmsg are simpler etc. So account for this
563 * fact here if necessary. If we don't do this correctly, as a
564 * receiver we won't recognize data packets as being full sized when we
565 * should, and thus we won't abide by the delayed ACK rules correctly.
566 * SACKs don't matter, we never delay an ACK when we have any of those
568 opts
->mss
= tcp_advertise_mss(sk
);
569 remaining
-= TCPOLEN_MSS_ALIGNED
;
571 if (likely(sysctl_tcp_timestamps
&& !*md5
)) {
572 opts
->options
|= OPTION_TS
;
573 opts
->tsval
= tcp_skb_timestamp(skb
) + tp
->tsoffset
;
574 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
575 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
577 if (likely(sysctl_tcp_window_scaling
)) {
578 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
579 opts
->options
|= OPTION_WSCALE
;
580 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
582 if (likely(sysctl_tcp_sack
)) {
583 opts
->options
|= OPTION_SACK_ADVERTISE
;
584 if (unlikely(!(OPTION_TS
& opts
->options
)))
585 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
588 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
589 u32 need
= fastopen
->cookie
.len
;
591 need
+= fastopen
->cookie
.exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
592 TCPOLEN_FASTOPEN_BASE
;
593 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
594 if (remaining
>= need
) {
595 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
596 opts
->fastopen_cookie
= &fastopen
->cookie
;
598 tp
->syn_fastopen
= 1;
599 tp
->syn_fastopen_exp
= fastopen
->cookie
.exp
? 1 : 0;
603 return MAX_TCP_OPTION_SPACE
- remaining
;
606 /* Set up TCP options for SYN-ACKs. */
607 static unsigned int tcp_synack_options(struct request_sock
*req
,
608 unsigned int mss
, struct sk_buff
*skb
,
609 struct tcp_out_options
*opts
,
610 const struct tcp_md5sig_key
*md5
,
611 struct tcp_fastopen_cookie
*foc
)
613 struct inet_request_sock
*ireq
= inet_rsk(req
);
614 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
616 #ifdef CONFIG_TCP_MD5SIG
618 opts
->options
|= OPTION_MD5
;
619 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
621 /* We can't fit any SACK blocks in a packet with MD5 + TS
622 * options. There was discussion about disabling SACK
623 * rather than TS in order to fit in better with old,
624 * buggy kernels, but that was deemed to be unnecessary.
626 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
630 /* We always send an MSS option. */
632 remaining
-= TCPOLEN_MSS_ALIGNED
;
634 if (likely(ireq
->wscale_ok
)) {
635 opts
->ws
= ireq
->rcv_wscale
;
636 opts
->options
|= OPTION_WSCALE
;
637 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
639 if (likely(ireq
->tstamp_ok
)) {
640 opts
->options
|= OPTION_TS
;
641 opts
->tsval
= tcp_skb_timestamp(skb
) + tcp_rsk(req
)->ts_off
;
642 opts
->tsecr
= req
->ts_recent
;
643 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
645 if (likely(ireq
->sack_ok
)) {
646 opts
->options
|= OPTION_SACK_ADVERTISE
;
647 if (unlikely(!ireq
->tstamp_ok
))
648 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
650 if (foc
!= NULL
&& foc
->len
>= 0) {
653 need
+= foc
->exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
654 TCPOLEN_FASTOPEN_BASE
;
655 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
656 if (remaining
>= need
) {
657 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
658 opts
->fastopen_cookie
= foc
;
663 return MAX_TCP_OPTION_SPACE
- remaining
;
666 /* Compute TCP options for ESTABLISHED sockets. This is not the
667 * final wire format yet.
669 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
670 struct tcp_out_options
*opts
,
671 struct tcp_md5sig_key
**md5
)
673 struct tcp_sock
*tp
= tcp_sk(sk
);
674 unsigned int size
= 0;
675 unsigned int eff_sacks
;
679 #ifdef CONFIG_TCP_MD5SIG
680 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
681 if (unlikely(*md5
)) {
682 opts
->options
|= OPTION_MD5
;
683 size
+= TCPOLEN_MD5SIG_ALIGNED
;
689 if (likely(tp
->rx_opt
.tstamp_ok
)) {
690 opts
->options
|= OPTION_TS
;
691 opts
->tsval
= skb
? tcp_skb_timestamp(skb
) + tp
->tsoffset
: 0;
692 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
693 size
+= TCPOLEN_TSTAMP_ALIGNED
;
696 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
697 if (unlikely(eff_sacks
)) {
698 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
699 opts
->num_sack_blocks
=
700 min_t(unsigned int, eff_sacks
,
701 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
702 TCPOLEN_SACK_PERBLOCK
);
703 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
704 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
711 /* TCP SMALL QUEUES (TSQ)
713 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
714 * to reduce RTT and bufferbloat.
715 * We do this using a special skb destructor (tcp_wfree).
717 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
718 * needs to be reallocated in a driver.
719 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
721 * Since transmit from skb destructor is forbidden, we use a tasklet
722 * to process all sockets that eventually need to send more skbs.
723 * We use one tasklet per cpu, with its own queue of sockets.
726 struct tasklet_struct tasklet
;
727 struct list_head head
; /* queue of tcp sockets */
729 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
731 static void tcp_tsq_handler(struct sock
*sk
)
733 if ((1 << sk
->sk_state
) &
734 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
735 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
)) {
736 struct tcp_sock
*tp
= tcp_sk(sk
);
738 if (tp
->lost_out
> tp
->retrans_out
&&
739 tp
->snd_cwnd
> tcp_packets_in_flight(tp
))
740 tcp_xmit_retransmit_queue(sk
);
742 tcp_write_xmit(sk
, tcp_current_mss(sk
), tp
->nonagle
,
747 * One tasklet per cpu tries to send more skbs.
748 * We run in tasklet context but need to disable irqs when
749 * transferring tsq->head because tcp_wfree() might
750 * interrupt us (non NAPI drivers)
752 static void tcp_tasklet_func(unsigned long data
)
754 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
757 struct list_head
*q
, *n
;
761 local_irq_save(flags
);
762 list_splice_init(&tsq
->head
, &list
);
763 local_irq_restore(flags
);
765 list_for_each_safe(q
, n
, &list
) {
766 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
767 list_del(&tp
->tsq_node
);
769 sk
= (struct sock
*)tp
;
770 smp_mb__before_atomic();
771 clear_bit(TSQ_QUEUED
, &sk
->sk_tsq_flags
);
773 if (!sk
->sk_lock
.owned
&&
774 test_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
)) {
776 if (!sock_owned_by_user(sk
)) {
777 clear_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
);
787 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
788 TCPF_WRITE_TIMER_DEFERRED | \
789 TCPF_DELACK_TIMER_DEFERRED | \
790 TCPF_MTU_REDUCED_DEFERRED)
792 * tcp_release_cb - tcp release_sock() callback
795 * called from release_sock() to perform protocol dependent
796 * actions before socket release.
798 void tcp_release_cb(struct sock
*sk
)
800 unsigned long flags
, nflags
;
802 /* perform an atomic operation only if at least one flag is set */
804 flags
= sk
->sk_tsq_flags
;
805 if (!(flags
& TCP_DEFERRED_ALL
))
807 nflags
= flags
& ~TCP_DEFERRED_ALL
;
808 } while (cmpxchg(&sk
->sk_tsq_flags
, flags
, nflags
) != flags
);
810 if (flags
& TCPF_TSQ_DEFERRED
)
813 /* Here begins the tricky part :
814 * We are called from release_sock() with :
816 * 2) sk_lock.slock spinlock held
817 * 3) socket owned by us (sk->sk_lock.owned == 1)
819 * But following code is meant to be called from BH handlers,
820 * so we should keep BH disabled, but early release socket ownership
822 sock_release_ownership(sk
);
824 if (flags
& TCPF_WRITE_TIMER_DEFERRED
) {
825 tcp_write_timer_handler(sk
);
828 if (flags
& TCPF_DELACK_TIMER_DEFERRED
) {
829 tcp_delack_timer_handler(sk
);
832 if (flags
& TCPF_MTU_REDUCED_DEFERRED
) {
833 inet_csk(sk
)->icsk_af_ops
->mtu_reduced(sk
);
837 EXPORT_SYMBOL(tcp_release_cb
);
839 void __init
tcp_tasklet_init(void)
843 for_each_possible_cpu(i
) {
844 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
846 INIT_LIST_HEAD(&tsq
->head
);
847 tasklet_init(&tsq
->tasklet
,
854 * Write buffer destructor automatically called from kfree_skb.
855 * We can't xmit new skbs from this context, as we might already
858 void tcp_wfree(struct sk_buff
*skb
)
860 struct sock
*sk
= skb
->sk
;
861 struct tcp_sock
*tp
= tcp_sk(sk
);
862 unsigned long flags
, nval
, oval
;
865 /* Keep one reference on sk_wmem_alloc.
866 * Will be released by sk_free() from here or tcp_tasklet_func()
868 wmem
= atomic_sub_return(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
870 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
871 * Wait until our queues (qdisc + devices) are drained.
873 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
874 * - chance for incoming ACK (processed by another cpu maybe)
875 * to migrate this flow (skb->ooo_okay will be eventually set)
877 if (wmem
>= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current
)
880 for (oval
= READ_ONCE(sk
->sk_tsq_flags
);; oval
= nval
) {
881 struct tsq_tasklet
*tsq
;
884 if (!(oval
& TSQF_THROTTLED
) || (oval
& TSQF_QUEUED
))
887 nval
= (oval
& ~TSQF_THROTTLED
) | TSQF_QUEUED
| TCPF_TSQ_DEFERRED
;
888 nval
= cmpxchg(&sk
->sk_tsq_flags
, oval
, nval
);
892 /* queue this socket to tasklet queue */
893 local_irq_save(flags
);
894 tsq
= this_cpu_ptr(&tsq_tasklet
);
895 empty
= list_empty(&tsq
->head
);
896 list_add(&tp
->tsq_node
, &tsq
->head
);
898 tasklet_schedule(&tsq
->tasklet
);
899 local_irq_restore(flags
);
906 /* This routine actually transmits TCP packets queued in by
907 * tcp_do_sendmsg(). This is used by both the initial
908 * transmission and possible later retransmissions.
909 * All SKB's seen here are completely headerless. It is our
910 * job to build the TCP header, and pass the packet down to
911 * IP so it can do the same plus pass the packet off to the
914 * We are working here with either a clone of the original
915 * SKB, or a fresh unique copy made by the retransmit engine.
917 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
920 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
921 struct inet_sock
*inet
;
923 struct tcp_skb_cb
*tcb
;
924 struct tcp_out_options opts
;
925 unsigned int tcp_options_size
, tcp_header_size
;
926 struct tcp_md5sig_key
*md5
;
930 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
934 skb_mstamp_get(&skb
->skb_mstamp
);
935 TCP_SKB_CB(skb
)->tx
.in_flight
= TCP_SKB_CB(skb
)->end_seq
937 tcp_rate_skb_sent(sk
, skb
);
939 if (unlikely(skb_cloned(skb
)))
940 skb
= pskb_copy(skb
, gfp_mask
);
942 skb
= skb_clone(skb
, gfp_mask
);
948 tcb
= TCP_SKB_CB(skb
);
949 memset(&opts
, 0, sizeof(opts
));
951 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
952 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
954 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
956 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
958 /* if no packet is in qdisc/device queue, then allow XPS to select
959 * another queue. We can be called from tcp_tsq_handler()
960 * which holds one reference to sk_wmem_alloc.
962 * TODO: Ideally, in-flight pure ACK packets should not matter here.
963 * One way to get this would be to set skb->truesize = 2 on them.
965 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) < SKB_TRUESIZE(1);
967 skb_push(skb
, tcp_header_size
);
968 skb_reset_transport_header(skb
);
972 skb
->destructor
= skb_is_tcp_pure_ack(skb
) ? __sock_wfree
: tcp_wfree
;
973 skb_set_hash_from_sk(skb
, sk
);
974 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
976 /* Build TCP header and checksum it. */
977 th
= (struct tcphdr
*)skb
->data
;
978 th
->source
= inet
->inet_sport
;
979 th
->dest
= inet
->inet_dport
;
980 th
->seq
= htonl(tcb
->seq
);
981 th
->ack_seq
= htonl(tp
->rcv_nxt
);
982 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
988 /* The urg_mode check is necessary during a below snd_una win probe */
989 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
990 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
991 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
993 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
994 th
->urg_ptr
= htons(0xFFFF);
999 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
1000 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
1001 if (likely(!(tcb
->tcp_flags
& TCPHDR_SYN
))) {
1002 th
->window
= htons(tcp_select_window(sk
));
1003 tcp_ecn_send(sk
, skb
, th
, tcp_header_size
);
1005 /* RFC1323: The window in SYN & SYN/ACK segments
1008 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
1010 #ifdef CONFIG_TCP_MD5SIG
1011 /* Calculate the MD5 hash, as we have all we need now */
1013 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
1014 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
1019 icsk
->icsk_af_ops
->send_check(sk
, skb
);
1021 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
1022 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
1024 if (skb
->len
!= tcp_header_size
) {
1025 tcp_event_data_sent(tp
, sk
);
1026 tp
->data_segs_out
+= tcp_skb_pcount(skb
);
1029 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
1030 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
1031 tcp_skb_pcount(skb
));
1033 tp
->segs_out
+= tcp_skb_pcount(skb
);
1034 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1035 skb_shinfo(skb
)->gso_segs
= tcp_skb_pcount(skb
);
1036 skb_shinfo(skb
)->gso_size
= tcp_skb_mss(skb
);
1038 /* Our usage of tstamp should remain private */
1041 /* Cleanup our debris for IP stacks */
1042 memset(skb
->cb
, 0, max(sizeof(struct inet_skb_parm
),
1043 sizeof(struct inet6_skb_parm
)));
1045 err
= icsk
->icsk_af_ops
->queue_xmit(sk
, skb
, &inet
->cork
.fl
);
1047 if (likely(err
<= 0))
1052 return net_xmit_eval(err
);
1055 /* This routine just queues the buffer for sending.
1057 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1058 * otherwise socket can stall.
1060 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1062 struct tcp_sock
*tp
= tcp_sk(sk
);
1064 /* Advance write_seq and place onto the write_queue. */
1065 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
1066 __skb_header_release(skb
);
1067 tcp_add_write_queue_tail(sk
, skb
);
1068 sk
->sk_wmem_queued
+= skb
->truesize
;
1069 sk_mem_charge(sk
, skb
->truesize
);
1072 /* Initialize TSO segments for a packet. */
1073 static void tcp_set_skb_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1075 if (skb
->len
<= mss_now
|| skb
->ip_summed
== CHECKSUM_NONE
) {
1076 /* Avoid the costly divide in the normal
1079 tcp_skb_pcount_set(skb
, 1);
1080 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1082 tcp_skb_pcount_set(skb
, DIV_ROUND_UP(skb
->len
, mss_now
));
1083 TCP_SKB_CB(skb
)->tcp_gso_size
= mss_now
;
1087 /* When a modification to fackets out becomes necessary, we need to check
1088 * skb is counted to fackets_out or not.
1090 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1093 struct tcp_sock
*tp
= tcp_sk(sk
);
1095 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1098 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1099 tp
->fackets_out
-= decr
;
1102 /* Pcount in the middle of the write queue got changed, we need to do various
1103 * tweaks to fix counters
1105 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1107 struct tcp_sock
*tp
= tcp_sk(sk
);
1109 tp
->packets_out
-= decr
;
1111 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1112 tp
->sacked_out
-= decr
;
1113 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1114 tp
->retrans_out
-= decr
;
1115 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1116 tp
->lost_out
-= decr
;
1118 /* Reno case is special. Sigh... */
1119 if (tcp_is_reno(tp
) && decr
> 0)
1120 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1122 tcp_adjust_fackets_out(sk
, skb
, decr
);
1124 if (tp
->lost_skb_hint
&&
1125 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1126 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1127 tp
->lost_cnt_hint
-= decr
;
1129 tcp_verify_left_out(tp
);
1132 static bool tcp_has_tx_tstamp(const struct sk_buff
*skb
)
1134 return TCP_SKB_CB(skb
)->txstamp_ack
||
1135 (skb_shinfo(skb
)->tx_flags
& SKBTX_ANY_TSTAMP
);
1138 static void tcp_fragment_tstamp(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1140 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1142 if (unlikely(tcp_has_tx_tstamp(skb
)) &&
1143 !before(shinfo
->tskey
, TCP_SKB_CB(skb2
)->seq
)) {
1144 struct skb_shared_info
*shinfo2
= skb_shinfo(skb2
);
1145 u8 tsflags
= shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
1147 shinfo
->tx_flags
&= ~tsflags
;
1148 shinfo2
->tx_flags
|= tsflags
;
1149 swap(shinfo
->tskey
, shinfo2
->tskey
);
1150 TCP_SKB_CB(skb2
)->txstamp_ack
= TCP_SKB_CB(skb
)->txstamp_ack
;
1151 TCP_SKB_CB(skb
)->txstamp_ack
= 0;
1155 static void tcp_skb_fragment_eor(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1157 TCP_SKB_CB(skb2
)->eor
= TCP_SKB_CB(skb
)->eor
;
1158 TCP_SKB_CB(skb
)->eor
= 0;
1161 /* Function to create two new TCP segments. Shrinks the given segment
1162 * to the specified size and appends a new segment with the rest of the
1163 * packet to the list. This won't be called frequently, I hope.
1164 * Remember, these are still headerless SKBs at this point.
1166 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1167 unsigned int mss_now
, gfp_t gfp
)
1169 struct tcp_sock
*tp
= tcp_sk(sk
);
1170 struct sk_buff
*buff
;
1171 int nsize
, old_factor
;
1175 if (WARN_ON(len
> skb
->len
))
1178 nsize
= skb_headlen(skb
) - len
;
1182 if (skb_unclone(skb
, gfp
))
1185 /* Get a new skb... force flag on. */
1186 buff
= sk_stream_alloc_skb(sk
, nsize
, gfp
, true);
1188 return -ENOMEM
; /* We'll just try again later. */
1190 sk
->sk_wmem_queued
+= buff
->truesize
;
1191 sk_mem_charge(sk
, buff
->truesize
);
1192 nlen
= skb
->len
- len
- nsize
;
1193 buff
->truesize
+= nlen
;
1194 skb
->truesize
-= nlen
;
1196 /* Correct the sequence numbers. */
1197 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1198 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1199 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1201 /* PSH and FIN should only be set in the second packet. */
1202 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1203 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1204 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1205 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1206 tcp_skb_fragment_eor(skb
, buff
);
1208 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1209 /* Copy and checksum data tail into the new buffer. */
1210 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1211 skb_put(buff
, nsize
),
1216 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1218 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1219 skb_split(skb
, buff
, len
);
1222 buff
->ip_summed
= skb
->ip_summed
;
1224 buff
->tstamp
= skb
->tstamp
;
1225 tcp_fragment_tstamp(skb
, buff
);
1227 old_factor
= tcp_skb_pcount(skb
);
1229 /* Fix up tso_factor for both original and new SKB. */
1230 tcp_set_skb_tso_segs(skb
, mss_now
);
1231 tcp_set_skb_tso_segs(buff
, mss_now
);
1233 /* Update delivered info for the new segment */
1234 TCP_SKB_CB(buff
)->tx
= TCP_SKB_CB(skb
)->tx
;
1236 /* If this packet has been sent out already, we must
1237 * adjust the various packet counters.
1239 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1240 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1241 tcp_skb_pcount(buff
);
1244 tcp_adjust_pcount(sk
, skb
, diff
);
1247 /* Link BUFF into the send queue. */
1248 __skb_header_release(buff
);
1249 tcp_insert_write_queue_after(skb
, buff
, sk
);
1254 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1255 * eventually). The difference is that pulled data not copied, but
1256 * immediately discarded.
1258 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1260 struct skb_shared_info
*shinfo
;
1263 eat
= min_t(int, len
, skb_headlen(skb
));
1265 __skb_pull(skb
, eat
);
1272 shinfo
= skb_shinfo(skb
);
1273 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1274 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1277 skb_frag_unref(skb
, i
);
1280 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1282 shinfo
->frags
[k
].page_offset
+= eat
;
1283 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1289 shinfo
->nr_frags
= k
;
1291 skb_reset_tail_pointer(skb
);
1292 skb
->data_len
-= len
;
1293 skb
->len
= skb
->data_len
;
1296 /* Remove acked data from a packet in the transmit queue. */
1297 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1299 if (skb_unclone(skb
, GFP_ATOMIC
))
1302 __pskb_trim_head(skb
, len
);
1304 TCP_SKB_CB(skb
)->seq
+= len
;
1305 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1307 skb
->truesize
-= len
;
1308 sk
->sk_wmem_queued
-= len
;
1309 sk_mem_uncharge(sk
, len
);
1310 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1312 /* Any change of skb->len requires recalculation of tso factor. */
1313 if (tcp_skb_pcount(skb
) > 1)
1314 tcp_set_skb_tso_segs(skb
, tcp_skb_mss(skb
));
1319 /* Calculate MSS not accounting any TCP options. */
1320 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1322 const struct tcp_sock
*tp
= tcp_sk(sk
);
1323 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1326 /* Calculate base mss without TCP options:
1327 It is MMS_S - sizeof(tcphdr) of rfc1122
1329 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1331 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1332 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1333 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1335 if (dst
&& dst_allfrag(dst
))
1336 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1339 /* Clamp it (mss_clamp does not include tcp options) */
1340 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1341 mss_now
= tp
->rx_opt
.mss_clamp
;
1343 /* Now subtract optional transport overhead */
1344 mss_now
-= icsk
->icsk_ext_hdr_len
;
1346 /* Then reserve room for full set of TCP options and 8 bytes of data */
1352 /* Calculate MSS. Not accounting for SACKs here. */
1353 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1355 /* Subtract TCP options size, not including SACKs */
1356 return __tcp_mtu_to_mss(sk
, pmtu
) -
1357 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1360 /* Inverse of above */
1361 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1363 const struct tcp_sock
*tp
= tcp_sk(sk
);
1364 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1368 tp
->tcp_header_len
+
1369 icsk
->icsk_ext_hdr_len
+
1370 icsk
->icsk_af_ops
->net_header_len
;
1372 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1373 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1374 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1376 if (dst
&& dst_allfrag(dst
))
1377 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1381 EXPORT_SYMBOL(tcp_mss_to_mtu
);
1383 /* MTU probing init per socket */
1384 void tcp_mtup_init(struct sock
*sk
)
1386 struct tcp_sock
*tp
= tcp_sk(sk
);
1387 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1388 struct net
*net
= sock_net(sk
);
1390 icsk
->icsk_mtup
.enabled
= net
->ipv4
.sysctl_tcp_mtu_probing
> 1;
1391 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1392 icsk
->icsk_af_ops
->net_header_len
;
1393 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, net
->ipv4
.sysctl_tcp_base_mss
);
1394 icsk
->icsk_mtup
.probe_size
= 0;
1395 if (icsk
->icsk_mtup
.enabled
)
1396 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1398 EXPORT_SYMBOL(tcp_mtup_init
);
1400 /* This function synchronize snd mss to current pmtu/exthdr set.
1402 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1403 for TCP options, but includes only bare TCP header.
1405 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1406 It is minimum of user_mss and mss received with SYN.
1407 It also does not include TCP options.
1409 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1411 tp->mss_cache is current effective sending mss, including
1412 all tcp options except for SACKs. It is evaluated,
1413 taking into account current pmtu, but never exceeds
1414 tp->rx_opt.mss_clamp.
1416 NOTE1. rfc1122 clearly states that advertised MSS
1417 DOES NOT include either tcp or ip options.
1419 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1420 are READ ONLY outside this function. --ANK (980731)
1422 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1424 struct tcp_sock
*tp
= tcp_sk(sk
);
1425 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1428 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1429 icsk
->icsk_mtup
.search_high
= pmtu
;
1431 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1432 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1434 /* And store cached results */
1435 icsk
->icsk_pmtu_cookie
= pmtu
;
1436 if (icsk
->icsk_mtup
.enabled
)
1437 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1438 tp
->mss_cache
= mss_now
;
1442 EXPORT_SYMBOL(tcp_sync_mss
);
1444 /* Compute the current effective MSS, taking SACKs and IP options,
1445 * and even PMTU discovery events into account.
1447 unsigned int tcp_current_mss(struct sock
*sk
)
1449 const struct tcp_sock
*tp
= tcp_sk(sk
);
1450 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1452 unsigned int header_len
;
1453 struct tcp_out_options opts
;
1454 struct tcp_md5sig_key
*md5
;
1456 mss_now
= tp
->mss_cache
;
1459 u32 mtu
= dst_mtu(dst
);
1460 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1461 mss_now
= tcp_sync_mss(sk
, mtu
);
1464 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1465 sizeof(struct tcphdr
);
1466 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1467 * some common options. If this is an odd packet (because we have SACK
1468 * blocks etc) then our calculated header_len will be different, and
1469 * we have to adjust mss_now correspondingly */
1470 if (header_len
!= tp
->tcp_header_len
) {
1471 int delta
= (int) header_len
- tp
->tcp_header_len
;
1478 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1479 * As additional protections, we do not touch cwnd in retransmission phases,
1480 * and if application hit its sndbuf limit recently.
1482 static void tcp_cwnd_application_limited(struct sock
*sk
)
1484 struct tcp_sock
*tp
= tcp_sk(sk
);
1486 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1487 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1488 /* Limited by application or receiver window. */
1489 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1490 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1491 if (win_used
< tp
->snd_cwnd
) {
1492 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1493 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
1495 tp
->snd_cwnd_used
= 0;
1497 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1500 static void tcp_cwnd_validate(struct sock
*sk
, bool is_cwnd_limited
)
1502 struct tcp_sock
*tp
= tcp_sk(sk
);
1504 /* Track the maximum number of outstanding packets in each
1505 * window, and remember whether we were cwnd-limited then.
1507 if (!before(tp
->snd_una
, tp
->max_packets_seq
) ||
1508 tp
->packets_out
> tp
->max_packets_out
) {
1509 tp
->max_packets_out
= tp
->packets_out
;
1510 tp
->max_packets_seq
= tp
->snd_nxt
;
1511 tp
->is_cwnd_limited
= is_cwnd_limited
;
1514 if (tcp_is_cwnd_limited(sk
)) {
1515 /* Network is feed fully. */
1516 tp
->snd_cwnd_used
= 0;
1517 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1519 /* Network starves. */
1520 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1521 tp
->snd_cwnd_used
= tp
->packets_out
;
1523 if (sysctl_tcp_slow_start_after_idle
&&
1524 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1525 tcp_cwnd_application_limited(sk
);
1527 /* The following conditions together indicate the starvation
1528 * is caused by insufficient sender buffer:
1529 * 1) just sent some data (see tcp_write_xmit)
1530 * 2) not cwnd limited (this else condition)
1531 * 3) no more data to send (null tcp_send_head )
1532 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1534 if (!tcp_send_head(sk
) && sk
->sk_socket
&&
1535 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
) &&
1536 (1 << sk
->sk_state
) & (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
1537 tcp_chrono_start(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1541 /* Minshall's variant of the Nagle send check. */
1542 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1544 return after(tp
->snd_sml
, tp
->snd_una
) &&
1545 !after(tp
->snd_sml
, tp
->snd_nxt
);
1548 /* Update snd_sml if this skb is under mss
1549 * Note that a TSO packet might end with a sub-mss segment
1550 * The test is really :
1551 * if ((skb->len % mss) != 0)
1552 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1553 * But we can avoid doing the divide again given we already have
1554 * skb_pcount = skb->len / mss_now
1556 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1557 const struct sk_buff
*skb
)
1559 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1560 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1563 /* Return false, if packet can be sent now without violation Nagle's rules:
1564 * 1. It is full sized. (provided by caller in %partial bool)
1565 * 2. Or it contains FIN. (already checked by caller)
1566 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1567 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1568 * With Minshall's modification: all sent small packets are ACKed.
1570 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1574 ((nonagle
& TCP_NAGLE_CORK
) ||
1575 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1578 /* Return how many segs we'd like on a TSO packet,
1579 * to send one TSO packet per ms
1581 u32
tcp_tso_autosize(const struct sock
*sk
, unsigned int mss_now
,
1586 bytes
= min(sk
->sk_pacing_rate
>> 10,
1587 sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
);
1589 /* Goal is to send at least one packet per ms,
1590 * not one big TSO packet every 100 ms.
1591 * This preserves ACK clocking and is consistent
1592 * with tcp_tso_should_defer() heuristic.
1594 segs
= max_t(u32
, bytes
/ mss_now
, min_tso_segs
);
1596 return min_t(u32
, segs
, sk
->sk_gso_max_segs
);
1598 EXPORT_SYMBOL(tcp_tso_autosize
);
1600 /* Return the number of segments we want in the skb we are transmitting.
1601 * See if congestion control module wants to decide; otherwise, autosize.
1603 static u32
tcp_tso_segs(struct sock
*sk
, unsigned int mss_now
)
1605 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1606 u32 tso_segs
= ca_ops
->tso_segs_goal
? ca_ops
->tso_segs_goal(sk
) : 0;
1609 tcp_tso_autosize(sk
, mss_now
, sysctl_tcp_min_tso_segs
);
1612 /* Returns the portion of skb which can be sent right away */
1613 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1614 const struct sk_buff
*skb
,
1615 unsigned int mss_now
,
1616 unsigned int max_segs
,
1619 const struct tcp_sock
*tp
= tcp_sk(sk
);
1620 u32 partial
, needed
, window
, max_len
;
1622 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1623 max_len
= mss_now
* max_segs
;
1625 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1628 needed
= min(skb
->len
, window
);
1630 if (max_len
<= needed
)
1633 partial
= needed
% mss_now
;
1634 /* If last segment is not a full MSS, check if Nagle rules allow us
1635 * to include this last segment in this skb.
1636 * Otherwise, we'll split the skb at last MSS boundary
1638 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
1639 return needed
- partial
;
1644 /* Can at least one segment of SKB be sent right now, according to the
1645 * congestion window rules? If so, return how many segments are allowed.
1647 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1648 const struct sk_buff
*skb
)
1650 u32 in_flight
, cwnd
, halfcwnd
;
1652 /* Don't be strict about the congestion window for the final FIN. */
1653 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1654 tcp_skb_pcount(skb
) == 1)
1657 in_flight
= tcp_packets_in_flight(tp
);
1658 cwnd
= tp
->snd_cwnd
;
1659 if (in_flight
>= cwnd
)
1662 /* For better scheduling, ensure we have at least
1663 * 2 GSO packets in flight.
1665 halfcwnd
= max(cwnd
>> 1, 1U);
1666 return min(halfcwnd
, cwnd
- in_flight
);
1669 /* Initialize TSO state of a skb.
1670 * This must be invoked the first time we consider transmitting
1671 * SKB onto the wire.
1673 static int tcp_init_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1675 int tso_segs
= tcp_skb_pcount(skb
);
1677 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1678 tcp_set_skb_tso_segs(skb
, mss_now
);
1679 tso_segs
= tcp_skb_pcount(skb
);
1685 /* Return true if the Nagle test allows this packet to be
1688 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1689 unsigned int cur_mss
, int nonagle
)
1691 /* Nagle rule does not apply to frames, which sit in the middle of the
1692 * write_queue (they have no chances to get new data).
1694 * This is implemented in the callers, where they modify the 'nonagle'
1695 * argument based upon the location of SKB in the send queue.
1697 if (nonagle
& TCP_NAGLE_PUSH
)
1700 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1701 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1704 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
1710 /* Does at least the first segment of SKB fit into the send window? */
1711 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1712 const struct sk_buff
*skb
,
1713 unsigned int cur_mss
)
1715 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1717 if (skb
->len
> cur_mss
)
1718 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1720 return !after(end_seq
, tcp_wnd_end(tp
));
1723 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1724 * should be put on the wire right now. If so, it returns the number of
1725 * packets allowed by the congestion window.
1727 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1728 unsigned int cur_mss
, int nonagle
)
1730 const struct tcp_sock
*tp
= tcp_sk(sk
);
1731 unsigned int cwnd_quota
;
1733 tcp_init_tso_segs(skb
, cur_mss
);
1735 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1738 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1739 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1745 /* Test if sending is allowed right now. */
1746 bool tcp_may_send_now(struct sock
*sk
)
1748 const struct tcp_sock
*tp
= tcp_sk(sk
);
1749 struct sk_buff
*skb
= tcp_send_head(sk
);
1752 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1753 (tcp_skb_is_last(sk
, skb
) ?
1754 tp
->nonagle
: TCP_NAGLE_PUSH
));
1757 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1758 * which is put after SKB on the list. It is very much like
1759 * tcp_fragment() except that it may make several kinds of assumptions
1760 * in order to speed up the splitting operation. In particular, we
1761 * know that all the data is in scatter-gather pages, and that the
1762 * packet has never been sent out before (and thus is not cloned).
1764 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1765 unsigned int mss_now
, gfp_t gfp
)
1767 struct sk_buff
*buff
;
1768 int nlen
= skb
->len
- len
;
1771 /* All of a TSO frame must be composed of paged data. */
1772 if (skb
->len
!= skb
->data_len
)
1773 return tcp_fragment(sk
, skb
, len
, mss_now
, gfp
);
1775 buff
= sk_stream_alloc_skb(sk
, 0, gfp
, true);
1776 if (unlikely(!buff
))
1779 sk
->sk_wmem_queued
+= buff
->truesize
;
1780 sk_mem_charge(sk
, buff
->truesize
);
1781 buff
->truesize
+= nlen
;
1782 skb
->truesize
-= nlen
;
1784 /* Correct the sequence numbers. */
1785 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1786 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1787 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1789 /* PSH and FIN should only be set in the second packet. */
1790 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1791 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1792 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1794 /* This packet was never sent out yet, so no SACK bits. */
1795 TCP_SKB_CB(buff
)->sacked
= 0;
1797 tcp_skb_fragment_eor(skb
, buff
);
1799 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1800 skb_split(skb
, buff
, len
);
1801 tcp_fragment_tstamp(skb
, buff
);
1803 /* Fix up tso_factor for both original and new SKB. */
1804 tcp_set_skb_tso_segs(skb
, mss_now
);
1805 tcp_set_skb_tso_segs(buff
, mss_now
);
1807 /* Link BUFF into the send queue. */
1808 __skb_header_release(buff
);
1809 tcp_insert_write_queue_after(skb
, buff
, sk
);
1814 /* Try to defer sending, if possible, in order to minimize the amount
1815 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1817 * This algorithm is from John Heffner.
1819 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
,
1820 bool *is_cwnd_limited
, u32 max_segs
)
1822 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1823 u32 age
, send_win
, cong_win
, limit
, in_flight
;
1824 struct tcp_sock
*tp
= tcp_sk(sk
);
1825 struct skb_mstamp now
;
1826 struct sk_buff
*head
;
1829 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1832 if (icsk
->icsk_ca_state
>= TCP_CA_Recovery
)
1835 /* Avoid bursty behavior by allowing defer
1836 * only if the last write was recent.
1838 if ((s32
)(tcp_time_stamp
- tp
->lsndtime
) > 0)
1841 in_flight
= tcp_packets_in_flight(tp
);
1843 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1845 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1847 /* From in_flight test above, we know that cwnd > in_flight. */
1848 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1850 limit
= min(send_win
, cong_win
);
1852 /* If a full-sized TSO skb can be sent, do it. */
1853 if (limit
>= max_segs
* tp
->mss_cache
)
1856 /* Middle in queue won't get any more data, full sendable already? */
1857 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1860 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1862 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1864 /* If at least some fraction of a window is available,
1867 chunk
/= win_divisor
;
1871 /* Different approach, try not to defer past a single
1872 * ACK. Receiver should ACK every other full sized
1873 * frame, so if we have space for more than 3 frames
1876 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1880 head
= tcp_write_queue_head(sk
);
1881 skb_mstamp_get(&now
);
1882 age
= skb_mstamp_us_delta(&now
, &head
->skb_mstamp
);
1883 /* If next ACK is likely to come too late (half srtt), do not defer */
1884 if (age
< (tp
->srtt_us
>> 4))
1887 /* Ok, it looks like it is advisable to defer. */
1889 if (cong_win
< send_win
&& cong_win
<= skb
->len
)
1890 *is_cwnd_limited
= true;
1898 static inline void tcp_mtu_check_reprobe(struct sock
*sk
)
1900 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1901 struct tcp_sock
*tp
= tcp_sk(sk
);
1902 struct net
*net
= sock_net(sk
);
1906 interval
= net
->ipv4
.sysctl_tcp_probe_interval
;
1907 delta
= tcp_time_stamp
- icsk
->icsk_mtup
.probe_timestamp
;
1908 if (unlikely(delta
>= interval
* HZ
)) {
1909 int mss
= tcp_current_mss(sk
);
1911 /* Update current search range */
1912 icsk
->icsk_mtup
.probe_size
= 0;
1913 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+
1914 sizeof(struct tcphdr
) +
1915 icsk
->icsk_af_ops
->net_header_len
;
1916 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
1918 /* Update probe time stamp */
1919 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1923 /* Create a new MTU probe if we are ready.
1924 * MTU probe is regularly attempting to increase the path MTU by
1925 * deliberately sending larger packets. This discovers routing
1926 * changes resulting in larger path MTUs.
1928 * Returns 0 if we should wait to probe (no cwnd available),
1929 * 1 if a probe was sent,
1932 static int tcp_mtu_probe(struct sock
*sk
)
1934 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1935 struct tcp_sock
*tp
= tcp_sk(sk
);
1936 struct sk_buff
*skb
, *nskb
, *next
;
1937 struct net
*net
= sock_net(sk
);
1944 /* Not currently probing/verifying,
1946 * have enough cwnd, and
1947 * not SACKing (the variable headers throw things off)
1949 if (likely(!icsk
->icsk_mtup
.enabled
||
1950 icsk
->icsk_mtup
.probe_size
||
1951 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1952 tp
->snd_cwnd
< 11 ||
1953 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
))
1956 /* Use binary search for probe_size between tcp_mss_base,
1957 * and current mss_clamp. if (search_high - search_low)
1958 * smaller than a threshold, backoff from probing.
1960 mss_now
= tcp_current_mss(sk
);
1961 probe_size
= tcp_mtu_to_mss(sk
, (icsk
->icsk_mtup
.search_high
+
1962 icsk
->icsk_mtup
.search_low
) >> 1);
1963 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1964 interval
= icsk
->icsk_mtup
.search_high
- icsk
->icsk_mtup
.search_low
;
1965 /* When misfortune happens, we are reprobing actively,
1966 * and then reprobe timer has expired. We stick with current
1967 * probing process by not resetting search range to its orignal.
1969 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
) ||
1970 interval
< net
->ipv4
.sysctl_tcp_probe_threshold
) {
1971 /* Check whether enough time has elaplased for
1972 * another round of probing.
1974 tcp_mtu_check_reprobe(sk
);
1978 /* Have enough data in the send queue to probe? */
1979 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1982 if (tp
->snd_wnd
< size_needed
)
1984 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1987 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1988 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1989 if (!tcp_packets_in_flight(tp
))
1995 /* We're allowed to probe. Build it now. */
1996 nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
, false);
1999 sk
->sk_wmem_queued
+= nskb
->truesize
;
2000 sk_mem_charge(sk
, nskb
->truesize
);
2002 skb
= tcp_send_head(sk
);
2004 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
2005 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
2006 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
2007 TCP_SKB_CB(nskb
)->sacked
= 0;
2009 nskb
->ip_summed
= skb
->ip_summed
;
2011 tcp_insert_write_queue_before(nskb
, skb
, sk
);
2014 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
2015 copy
= min_t(int, skb
->len
, probe_size
- len
);
2016 if (nskb
->ip_summed
) {
2017 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
2019 __wsum csum
= skb_copy_and_csum_bits(skb
, 0,
2020 skb_put(nskb
, copy
),
2022 nskb
->csum
= csum_block_add(nskb
->csum
, csum
, len
);
2025 if (skb
->len
<= copy
) {
2026 /* We've eaten all the data from this skb.
2028 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
2029 tcp_unlink_write_queue(skb
, sk
);
2030 sk_wmem_free_skb(sk
, skb
);
2032 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
2033 ~(TCPHDR_FIN
|TCPHDR_PSH
);
2034 if (!skb_shinfo(skb
)->nr_frags
) {
2035 skb_pull(skb
, copy
);
2036 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2037 skb
->csum
= csum_partial(skb
->data
,
2040 __pskb_trim_head(skb
, copy
);
2041 tcp_set_skb_tso_segs(skb
, mss_now
);
2043 TCP_SKB_CB(skb
)->seq
+= copy
;
2048 if (len
>= probe_size
)
2051 tcp_init_tso_segs(nskb
, nskb
->len
);
2053 /* We're ready to send. If this fails, the probe will
2054 * be resegmented into mss-sized pieces by tcp_write_xmit().
2056 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
2057 /* Decrement cwnd here because we are sending
2058 * effectively two packets. */
2060 tcp_event_new_data_sent(sk
, nskb
);
2062 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
2063 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
2064 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
2072 /* TCP Small Queues :
2073 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2074 * (These limits are doubled for retransmits)
2076 * - better RTT estimation and ACK scheduling
2079 * Alas, some drivers / subsystems require a fair amount
2080 * of queued bytes to ensure line rate.
2081 * One example is wifi aggregation (802.11 AMPDU)
2083 static bool tcp_small_queue_check(struct sock
*sk
, const struct sk_buff
*skb
,
2084 unsigned int factor
)
2088 limit
= max(2 * skb
->truesize
, sk
->sk_pacing_rate
>> 10);
2089 limit
= min_t(u32
, limit
, sysctl_tcp_limit_output_bytes
);
2092 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
) {
2093 /* Always send the 1st or 2nd skb in write queue.
2094 * No need to wait for TX completion to call us back,
2095 * after softirq/tasklet schedule.
2096 * This helps when TX completions are delayed too much.
2098 if (skb
== sk
->sk_write_queue
.next
||
2099 skb
->prev
== sk
->sk_write_queue
.next
)
2102 set_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
);
2103 /* It is possible TX completion already happened
2104 * before we set TSQ_THROTTLED, so we must
2105 * test again the condition.
2107 smp_mb__after_atomic();
2108 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
)
2114 static void tcp_chrono_set(struct tcp_sock
*tp
, const enum tcp_chrono
new)
2116 const u32 now
= tcp_time_stamp
;
2118 if (tp
->chrono_type
> TCP_CHRONO_UNSPEC
)
2119 tp
->chrono_stat
[tp
->chrono_type
- 1] += now
- tp
->chrono_start
;
2120 tp
->chrono_start
= now
;
2121 tp
->chrono_type
= new;
2124 void tcp_chrono_start(struct sock
*sk
, const enum tcp_chrono type
)
2126 struct tcp_sock
*tp
= tcp_sk(sk
);
2128 /* If there are multiple conditions worthy of tracking in a
2129 * chronograph then the highest priority enum takes precedence
2130 * over the other conditions. So that if something "more interesting"
2131 * starts happening, stop the previous chrono and start a new one.
2133 if (type
> tp
->chrono_type
)
2134 tcp_chrono_set(tp
, type
);
2137 void tcp_chrono_stop(struct sock
*sk
, const enum tcp_chrono type
)
2139 struct tcp_sock
*tp
= tcp_sk(sk
);
2142 /* There are multiple conditions worthy of tracking in a
2143 * chronograph, so that the highest priority enum takes
2144 * precedence over the other conditions (see tcp_chrono_start).
2145 * If a condition stops, we only stop chrono tracking if
2146 * it's the "most interesting" or current chrono we are
2147 * tracking and starts busy chrono if we have pending data.
2149 if (tcp_write_queue_empty(sk
))
2150 tcp_chrono_set(tp
, TCP_CHRONO_UNSPEC
);
2151 else if (type
== tp
->chrono_type
)
2152 tcp_chrono_set(tp
, TCP_CHRONO_BUSY
);
2155 /* This routine writes packets to the network. It advances the
2156 * send_head. This happens as incoming acks open up the remote
2159 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2160 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2161 * account rare use of URG, this is not a big flaw.
2163 * Send at most one packet when push_one > 0. Temporarily ignore
2164 * cwnd limit to force at most one packet out when push_one == 2.
2166 * Returns true, if no segments are in flight and we have queued segments,
2167 * but cannot send anything now because of SWS or another problem.
2169 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
2170 int push_one
, gfp_t gfp
)
2172 struct tcp_sock
*tp
= tcp_sk(sk
);
2173 struct sk_buff
*skb
;
2174 unsigned int tso_segs
, sent_pkts
;
2177 bool is_cwnd_limited
= false, is_rwnd_limited
= false;
2183 /* Do MTU probing. */
2184 result
= tcp_mtu_probe(sk
);
2187 } else if (result
> 0) {
2192 max_segs
= tcp_tso_segs(sk
, mss_now
);
2193 while ((skb
= tcp_send_head(sk
))) {
2196 tso_segs
= tcp_init_tso_segs(skb
, mss_now
);
2199 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
) {
2200 /* "skb_mstamp" is used as a start point for the retransmit timer */
2201 skb_mstamp_get(&skb
->skb_mstamp
);
2202 goto repair
; /* Skip network transmission */
2205 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2208 /* Force out a loss probe pkt. */
2214 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
))) {
2215 is_rwnd_limited
= true;
2219 if (tso_segs
== 1) {
2220 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2221 (tcp_skb_is_last(sk
, skb
) ?
2222 nonagle
: TCP_NAGLE_PUSH
))))
2226 tcp_tso_should_defer(sk
, skb
, &is_cwnd_limited
,
2232 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
2233 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2239 if (skb
->len
> limit
&&
2240 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
2243 if (test_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
))
2244 clear_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
);
2245 if (tcp_small_queue_check(sk
, skb
, 0))
2248 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2252 /* Advance the send_head. This one is sent out.
2253 * This call will increment packets_out.
2255 tcp_event_new_data_sent(sk
, skb
);
2257 tcp_minshall_update(tp
, mss_now
, skb
);
2258 sent_pkts
+= tcp_skb_pcount(skb
);
2264 if (is_rwnd_limited
)
2265 tcp_chrono_start(sk
, TCP_CHRONO_RWND_LIMITED
);
2267 tcp_chrono_stop(sk
, TCP_CHRONO_RWND_LIMITED
);
2269 if (likely(sent_pkts
)) {
2270 if (tcp_in_cwnd_reduction(sk
))
2271 tp
->prr_out
+= sent_pkts
;
2273 /* Send one loss probe per tail loss episode. */
2275 tcp_schedule_loss_probe(sk
);
2276 is_cwnd_limited
|= (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
);
2277 tcp_cwnd_validate(sk
, is_cwnd_limited
);
2280 return !tp
->packets_out
&& tcp_send_head(sk
);
2283 bool tcp_schedule_loss_probe(struct sock
*sk
)
2285 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2286 struct tcp_sock
*tp
= tcp_sk(sk
);
2287 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
2288 u32 rtt
= usecs_to_jiffies(tp
->srtt_us
>> 3);
2290 /* No consecutive loss probes. */
2291 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
2295 /* Don't do any loss probe on a Fast Open connection before 3WHS
2298 if (tp
->fastopen_rsk
)
2301 /* TLP is only scheduled when next timer event is RTO. */
2302 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
2305 /* Schedule a loss probe in 2*RTT for SACK capable connections
2306 * in Open state, that are either limited by cwnd or application.
2308 if ((sysctl_tcp_early_retrans
!= 3 && sysctl_tcp_early_retrans
!= 4) ||
2309 !tp
->packets_out
|| !tcp_is_sack(tp
) ||
2310 icsk
->icsk_ca_state
!= TCP_CA_Open
)
2313 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
2317 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2318 * for delayed ack when there's one outstanding packet. If no RTT
2319 * sample is available then probe after TCP_TIMEOUT_INIT.
2321 timeout
= rtt
<< 1 ? : TCP_TIMEOUT_INIT
;
2322 if (tp
->packets_out
== 1)
2323 timeout
= max_t(u32
, timeout
,
2324 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
2325 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
2327 /* If RTO is shorter, just schedule TLP in its place. */
2328 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
2329 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
2330 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
2331 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
2336 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2341 /* Thanks to skb fast clones, we can detect if a prior transmit of
2342 * a packet is still in a qdisc or driver queue.
2343 * In this case, there is very little point doing a retransmit !
2345 static bool skb_still_in_host_queue(const struct sock
*sk
,
2346 const struct sk_buff
*skb
)
2348 if (unlikely(skb_fclone_busy(sk
, skb
))) {
2349 NET_INC_STATS(sock_net(sk
),
2350 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
2356 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2357 * retransmit the last segment.
2359 void tcp_send_loss_probe(struct sock
*sk
)
2361 struct tcp_sock
*tp
= tcp_sk(sk
);
2362 struct sk_buff
*skb
;
2364 int mss
= tcp_current_mss(sk
);
2366 skb
= tcp_send_head(sk
);
2368 if (tcp_snd_wnd_test(tp
, skb
, mss
)) {
2369 pcount
= tp
->packets_out
;
2370 tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2371 if (tp
->packets_out
> pcount
)
2375 skb
= tcp_write_queue_prev(sk
, skb
);
2377 skb
= tcp_write_queue_tail(sk
);
2380 /* At most one outstanding TLP retransmission. */
2381 if (tp
->tlp_high_seq
)
2384 /* Retransmit last segment. */
2388 if (skb_still_in_host_queue(sk
, skb
))
2391 pcount
= tcp_skb_pcount(skb
);
2392 if (WARN_ON(!pcount
))
2395 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2396 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
,
2399 skb
= tcp_write_queue_next(sk
, skb
);
2402 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2405 if (__tcp_retransmit_skb(sk
, skb
, 1))
2408 /* Record snd_nxt for loss detection. */
2409 tp
->tlp_high_seq
= tp
->snd_nxt
;
2412 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSPROBES
);
2413 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2414 inet_csk(sk
)->icsk_pending
= 0;
2419 /* Push out any pending frames which were held back due to
2420 * TCP_CORK or attempt at coalescing tiny packets.
2421 * The socket must be locked by the caller.
2423 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2426 /* If we are closed, the bytes will have to remain here.
2427 * In time closedown will finish, we empty the write queue and
2428 * all will be happy.
2430 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2433 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2434 sk_gfp_mask(sk
, GFP_ATOMIC
)))
2435 tcp_check_probe_timer(sk
);
2438 /* Send _single_ skb sitting at the send head. This function requires
2439 * true push pending frames to setup probe timer etc.
2441 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2443 struct sk_buff
*skb
= tcp_send_head(sk
);
2445 BUG_ON(!skb
|| skb
->len
< mss_now
);
2447 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2450 /* This function returns the amount that we can raise the
2451 * usable window based on the following constraints
2453 * 1. The window can never be shrunk once it is offered (RFC 793)
2454 * 2. We limit memory per socket
2457 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2458 * RECV.NEXT + RCV.WIN fixed until:
2459 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2461 * i.e. don't raise the right edge of the window until you can raise
2462 * it at least MSS bytes.
2464 * Unfortunately, the recommended algorithm breaks header prediction,
2465 * since header prediction assumes th->window stays fixed.
2467 * Strictly speaking, keeping th->window fixed violates the receiver
2468 * side SWS prevention criteria. The problem is that under this rule
2469 * a stream of single byte packets will cause the right side of the
2470 * window to always advance by a single byte.
2472 * Of course, if the sender implements sender side SWS prevention
2473 * then this will not be a problem.
2475 * BSD seems to make the following compromise:
2477 * If the free space is less than the 1/4 of the maximum
2478 * space available and the free space is less than 1/2 mss,
2479 * then set the window to 0.
2480 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2481 * Otherwise, just prevent the window from shrinking
2482 * and from being larger than the largest representable value.
2484 * This prevents incremental opening of the window in the regime
2485 * where TCP is limited by the speed of the reader side taking
2486 * data out of the TCP receive queue. It does nothing about
2487 * those cases where the window is constrained on the sender side
2488 * because the pipeline is full.
2490 * BSD also seems to "accidentally" limit itself to windows that are a
2491 * multiple of MSS, at least until the free space gets quite small.
2492 * This would appear to be a side effect of the mbuf implementation.
2493 * Combining these two algorithms results in the observed behavior
2494 * of having a fixed window size at almost all times.
2496 * Below we obtain similar behavior by forcing the offered window to
2497 * a multiple of the mss when it is feasible to do so.
2499 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2500 * Regular options like TIMESTAMP are taken into account.
2502 u32
__tcp_select_window(struct sock
*sk
)
2504 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2505 struct tcp_sock
*tp
= tcp_sk(sk
);
2506 /* MSS for the peer's data. Previous versions used mss_clamp
2507 * here. I don't know if the value based on our guesses
2508 * of peer's MSS is better for the performance. It's more correct
2509 * but may be worse for the performance because of rcv_mss
2510 * fluctuations. --SAW 1998/11/1
2512 int mss
= icsk
->icsk_ack
.rcv_mss
;
2513 int free_space
= tcp_space(sk
);
2514 int allowed_space
= tcp_full_space(sk
);
2515 int full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2518 if (mss
> full_space
)
2521 if (free_space
< (full_space
>> 1)) {
2522 icsk
->icsk_ack
.quick
= 0;
2524 if (tcp_under_memory_pressure(sk
))
2525 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2528 /* free_space might become our new window, make sure we don't
2529 * increase it due to wscale.
2531 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2533 /* if free space is less than mss estimate, or is below 1/16th
2534 * of the maximum allowed, try to move to zero-window, else
2535 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2536 * new incoming data is dropped due to memory limits.
2537 * With large window, mss test triggers way too late in order
2538 * to announce zero window in time before rmem limit kicks in.
2540 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2544 if (free_space
> tp
->rcv_ssthresh
)
2545 free_space
= tp
->rcv_ssthresh
;
2547 /* Don't do rounding if we are using window scaling, since the
2548 * scaled window will not line up with the MSS boundary anyway.
2550 window
= tp
->rcv_wnd
;
2551 if (tp
->rx_opt
.rcv_wscale
) {
2552 window
= free_space
;
2554 /* Advertise enough space so that it won't get scaled away.
2555 * Import case: prevent zero window announcement if
2556 * 1<<rcv_wscale > mss.
2558 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2559 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2560 << tp
->rx_opt
.rcv_wscale
);
2562 /* Get the largest window that is a nice multiple of mss.
2563 * Window clamp already applied above.
2564 * If our current window offering is within 1 mss of the
2565 * free space we just keep it. This prevents the divide
2566 * and multiply from happening most of the time.
2567 * We also don't do any window rounding when the free space
2570 if (window
<= free_space
- mss
|| window
> free_space
)
2571 window
= (free_space
/ mss
) * mss
;
2572 else if (mss
== full_space
&&
2573 free_space
> window
+ (full_space
>> 1))
2574 window
= free_space
;
2580 void tcp_skb_collapse_tstamp(struct sk_buff
*skb
,
2581 const struct sk_buff
*next_skb
)
2583 if (unlikely(tcp_has_tx_tstamp(next_skb
))) {
2584 const struct skb_shared_info
*next_shinfo
=
2585 skb_shinfo(next_skb
);
2586 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
2588 shinfo
->tx_flags
|= next_shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
2589 shinfo
->tskey
= next_shinfo
->tskey
;
2590 TCP_SKB_CB(skb
)->txstamp_ack
|=
2591 TCP_SKB_CB(next_skb
)->txstamp_ack
;
2595 /* Collapses two adjacent SKB's during retransmission. */
2596 static bool tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2598 struct tcp_sock
*tp
= tcp_sk(sk
);
2599 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2600 int skb_size
, next_skb_size
;
2602 skb_size
= skb
->len
;
2603 next_skb_size
= next_skb
->len
;
2605 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2607 if (next_skb_size
) {
2608 if (next_skb_size
<= skb_availroom(skb
))
2609 skb_copy_bits(next_skb
, 0, skb_put(skb
, next_skb_size
),
2611 else if (!skb_shift(skb
, next_skb
, next_skb_size
))
2614 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2616 tcp_unlink_write_queue(next_skb
, sk
);
2618 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2619 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2621 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2622 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2624 /* Update sequence range on original skb. */
2625 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2627 /* Merge over control information. This moves PSH/FIN etc. over */
2628 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2630 /* All done, get rid of second SKB and account for it so
2631 * packet counting does not break.
2633 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2634 TCP_SKB_CB(skb
)->eor
= TCP_SKB_CB(next_skb
)->eor
;
2636 /* changed transmit queue under us so clear hints */
2637 tcp_clear_retrans_hints_partial(tp
);
2638 if (next_skb
== tp
->retransmit_skb_hint
)
2639 tp
->retransmit_skb_hint
= skb
;
2641 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2643 tcp_skb_collapse_tstamp(skb
, next_skb
);
2645 sk_wmem_free_skb(sk
, next_skb
);
2649 /* Check if coalescing SKBs is legal. */
2650 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2652 if (tcp_skb_pcount(skb
) > 1)
2654 if (skb_cloned(skb
))
2656 if (skb
== tcp_send_head(sk
))
2658 /* Some heuristics for collapsing over SACK'd could be invented */
2659 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2665 /* Collapse packets in the retransmit queue to make to create
2666 * less packets on the wire. This is only done on retransmission.
2668 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2671 struct tcp_sock
*tp
= tcp_sk(sk
);
2672 struct sk_buff
*skb
= to
, *tmp
;
2675 if (!sysctl_tcp_retrans_collapse
)
2677 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2680 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2681 if (!tcp_can_collapse(sk
, skb
))
2684 if (!tcp_skb_can_collapse_to(to
))
2697 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2700 if (!tcp_collapse_retrans(sk
, to
))
2705 /* This retransmits one SKB. Policy decisions and retransmit queue
2706 * state updates are done by the caller. Returns non-zero if an
2707 * error occurred which prevented the send.
2709 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2711 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2712 struct tcp_sock
*tp
= tcp_sk(sk
);
2713 unsigned int cur_mss
;
2717 /* Inconclusive MTU probe */
2718 if (icsk
->icsk_mtup
.probe_size
)
2719 icsk
->icsk_mtup
.probe_size
= 0;
2721 /* Do not sent more than we queued. 1/4 is reserved for possible
2722 * copying overhead: fragmentation, tunneling, mangling etc.
2724 if (atomic_read(&sk
->sk_wmem_alloc
) >
2725 min_t(u32
, sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2),
2729 if (skb_still_in_host_queue(sk
, skb
))
2732 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2733 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2735 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2739 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2740 return -EHOSTUNREACH
; /* Routing failure or similar. */
2742 cur_mss
= tcp_current_mss(sk
);
2744 /* If receiver has shrunk his window, and skb is out of
2745 * new window, do not retransmit it. The exception is the
2746 * case, when window is shrunk to zero. In this case
2747 * our retransmit serves as a zero window probe.
2749 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2750 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2753 len
= cur_mss
* segs
;
2754 if (skb
->len
> len
) {
2755 if (tcp_fragment(sk
, skb
, len
, cur_mss
, GFP_ATOMIC
))
2756 return -ENOMEM
; /* We'll try again later. */
2758 if (skb_unclone(skb
, GFP_ATOMIC
))
2761 diff
= tcp_skb_pcount(skb
);
2762 tcp_set_skb_tso_segs(skb
, cur_mss
);
2763 diff
-= tcp_skb_pcount(skb
);
2765 tcp_adjust_pcount(sk
, skb
, diff
);
2766 if (skb
->len
< cur_mss
)
2767 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2770 /* RFC3168, section 6.1.1.1. ECN fallback */
2771 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN_ECN
) == TCPHDR_SYN_ECN
)
2772 tcp_ecn_clear_syn(sk
, skb
);
2774 /* make sure skb->data is aligned on arches that require it
2775 * and check if ack-trimming & collapsing extended the headroom
2776 * beyond what csum_start can cover.
2778 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2779 skb_headroom(skb
) >= 0xFFFF)) {
2780 struct sk_buff
*nskb
;
2782 skb_mstamp_get(&skb
->skb_mstamp
);
2783 nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
, GFP_ATOMIC
);
2784 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2787 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2791 segs
= tcp_skb_pcount(skb
);
2793 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
2794 /* Update global TCP statistics. */
2795 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
, segs
);
2796 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2797 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
2798 tp
->total_retrans
+= segs
;
2803 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2805 struct tcp_sock
*tp
= tcp_sk(sk
);
2806 int err
= __tcp_retransmit_skb(sk
, skb
, segs
);
2809 #if FASTRETRANS_DEBUG > 0
2810 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2811 net_dbg_ratelimited("retrans_out leaked\n");
2814 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2815 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2817 /* Save stamp of the first retransmit. */
2818 if (!tp
->retrans_stamp
)
2819 tp
->retrans_stamp
= tcp_skb_timestamp(skb
);
2821 } else if (err
!= -EBUSY
) {
2822 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2825 if (tp
->undo_retrans
< 0)
2826 tp
->undo_retrans
= 0;
2827 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2831 /* This gets called after a retransmit timeout, and the initially
2832 * retransmitted data is acknowledged. It tries to continue
2833 * resending the rest of the retransmit queue, until either
2834 * we've sent it all or the congestion window limit is reached.
2835 * If doing SACK, the first ACK which comes back for a timeout
2836 * based retransmit packet might feed us FACK information again.
2837 * If so, we use it to avoid unnecessarily retransmissions.
2839 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2841 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2842 struct tcp_sock
*tp
= tcp_sk(sk
);
2843 struct sk_buff
*skb
;
2844 struct sk_buff
*hole
= NULL
;
2848 if (!tp
->packets_out
)
2851 if (tp
->retransmit_skb_hint
) {
2852 skb
= tp
->retransmit_skb_hint
;
2854 skb
= tcp_write_queue_head(sk
);
2857 max_segs
= tcp_tso_segs(sk
, tcp_current_mss(sk
));
2858 tcp_for_write_queue_from(skb
, sk
) {
2862 if (skb
== tcp_send_head(sk
))
2864 /* we could do better than to assign each time */
2866 tp
->retransmit_skb_hint
= skb
;
2868 segs
= tp
->snd_cwnd
- tcp_packets_in_flight(tp
);
2871 sacked
= TCP_SKB_CB(skb
)->sacked
;
2872 /* In case tcp_shift_skb_data() have aggregated large skbs,
2873 * we need to make sure not sending too bigs TSO packets
2875 segs
= min_t(int, segs
, max_segs
);
2877 if (tp
->retrans_out
>= tp
->lost_out
) {
2879 } else if (!(sacked
& TCPCB_LOST
)) {
2880 if (!hole
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2885 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2886 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2888 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2891 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2894 if (tcp_small_queue_check(sk
, skb
, 1))
2897 if (tcp_retransmit_skb(sk
, skb
, segs
))
2900 NET_ADD_STATS(sock_net(sk
), mib_idx
, tcp_skb_pcount(skb
));
2902 if (tcp_in_cwnd_reduction(sk
))
2903 tp
->prr_out
+= tcp_skb_pcount(skb
);
2905 if (skb
== tcp_write_queue_head(sk
) &&
2906 icsk
->icsk_pending
!= ICSK_TIME_REO_TIMEOUT
)
2907 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2908 inet_csk(sk
)->icsk_rto
,
2913 /* We allow to exceed memory limits for FIN packets to expedite
2914 * connection tear down and (memory) recovery.
2915 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2916 * or even be forced to close flow without any FIN.
2917 * In general, we want to allow one skb per socket to avoid hangs
2918 * with edge trigger epoll()
2920 void sk_forced_mem_schedule(struct sock
*sk
, int size
)
2924 if (size
<= sk
->sk_forward_alloc
)
2926 amt
= sk_mem_pages(size
);
2927 sk
->sk_forward_alloc
+= amt
* SK_MEM_QUANTUM
;
2928 sk_memory_allocated_add(sk
, amt
);
2930 if (mem_cgroup_sockets_enabled
&& sk
->sk_memcg
)
2931 mem_cgroup_charge_skmem(sk
->sk_memcg
, amt
);
2934 /* Send a FIN. The caller locks the socket for us.
2935 * We should try to send a FIN packet really hard, but eventually give up.
2937 void tcp_send_fin(struct sock
*sk
)
2939 struct sk_buff
*skb
, *tskb
= tcp_write_queue_tail(sk
);
2940 struct tcp_sock
*tp
= tcp_sk(sk
);
2942 /* Optimization, tack on the FIN if we have one skb in write queue and
2943 * this skb was not yet sent, or we are under memory pressure.
2944 * Note: in the latter case, FIN packet will be sent after a timeout,
2945 * as TCP stack thinks it has already been transmitted.
2947 if (tskb
&& (tcp_send_head(sk
) || tcp_under_memory_pressure(sk
))) {
2949 TCP_SKB_CB(tskb
)->tcp_flags
|= TCPHDR_FIN
;
2950 TCP_SKB_CB(tskb
)->end_seq
++;
2952 if (!tcp_send_head(sk
)) {
2953 /* This means tskb was already sent.
2954 * Pretend we included the FIN on previous transmit.
2955 * We need to set tp->snd_nxt to the value it would have
2956 * if FIN had been sent. This is because retransmit path
2957 * does not change tp->snd_nxt.
2963 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, sk
->sk_allocation
);
2964 if (unlikely(!skb
)) {
2969 skb_reserve(skb
, MAX_TCP_HEADER
);
2970 sk_forced_mem_schedule(sk
, skb
->truesize
);
2971 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2972 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2973 TCPHDR_ACK
| TCPHDR_FIN
);
2974 tcp_queue_skb(sk
, skb
);
2976 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
), TCP_NAGLE_OFF
);
2979 /* We get here when a process closes a file descriptor (either due to
2980 * an explicit close() or as a byproduct of exit()'ing) and there
2981 * was unread data in the receive queue. This behavior is recommended
2982 * by RFC 2525, section 2.17. -DaveM
2984 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2986 struct sk_buff
*skb
;
2988 /* NOTE: No TCP options attached and we never retransmit this. */
2989 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2991 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2995 /* Reserve space for headers and prepare control bits. */
2996 skb_reserve(skb
, MAX_TCP_HEADER
);
2997 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2998 TCPHDR_ACK
| TCPHDR_RST
);
2999 skb_mstamp_get(&skb
->skb_mstamp
);
3001 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
3002 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3004 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
3007 /* Send a crossed SYN-ACK during socket establishment.
3008 * WARNING: This routine must only be called when we have already sent
3009 * a SYN packet that crossed the incoming SYN that caused this routine
3010 * to get called. If this assumption fails then the initial rcv_wnd
3011 * and rcv_wscale values will not be correct.
3013 int tcp_send_synack(struct sock
*sk
)
3015 struct sk_buff
*skb
;
3017 skb
= tcp_write_queue_head(sk
);
3018 if (!skb
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
3019 pr_debug("%s: wrong queue state\n", __func__
);
3022 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
3023 if (skb_cloned(skb
)) {
3024 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
3027 tcp_unlink_write_queue(skb
, sk
);
3028 __skb_header_release(nskb
);
3029 __tcp_add_write_queue_head(sk
, nskb
);
3030 sk_wmem_free_skb(sk
, skb
);
3031 sk
->sk_wmem_queued
+= nskb
->truesize
;
3032 sk_mem_charge(sk
, nskb
->truesize
);
3036 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
3037 tcp_ecn_send_synack(sk
, skb
);
3039 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3043 * tcp_make_synack - Prepare a SYN-ACK.
3044 * sk: listener socket
3045 * dst: dst entry attached to the SYNACK
3046 * req: request_sock pointer
3048 * Allocate one skb and build a SYNACK packet.
3049 * @dst is consumed : Caller should not use it again.
3051 struct sk_buff
*tcp_make_synack(const struct sock
*sk
, struct dst_entry
*dst
,
3052 struct request_sock
*req
,
3053 struct tcp_fastopen_cookie
*foc
,
3054 enum tcp_synack_type synack_type
)
3056 struct inet_request_sock
*ireq
= inet_rsk(req
);
3057 const struct tcp_sock
*tp
= tcp_sk(sk
);
3058 struct tcp_md5sig_key
*md5
= NULL
;
3059 struct tcp_out_options opts
;
3060 struct sk_buff
*skb
;
3061 int tcp_header_size
;
3066 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
3067 if (unlikely(!skb
)) {
3071 /* Reserve space for headers. */
3072 skb_reserve(skb
, MAX_TCP_HEADER
);
3074 switch (synack_type
) {
3075 case TCP_SYNACK_NORMAL
:
3076 skb_set_owner_w(skb
, req_to_sk(req
));
3078 case TCP_SYNACK_COOKIE
:
3079 /* Under synflood, we do not attach skb to a socket,
3080 * to avoid false sharing.
3083 case TCP_SYNACK_FASTOPEN
:
3084 /* sk is a const pointer, because we want to express multiple
3085 * cpu might call us concurrently.
3086 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3088 skb_set_owner_w(skb
, (struct sock
*)sk
);
3091 skb_dst_set(skb
, dst
);
3093 mss
= dst_metric_advmss(dst
);
3094 user_mss
= READ_ONCE(tp
->rx_opt
.user_mss
);
3095 if (user_mss
&& user_mss
< mss
)
3098 memset(&opts
, 0, sizeof(opts
));
3099 #ifdef CONFIG_SYN_COOKIES
3100 if (unlikely(req
->cookie_ts
))
3101 skb
->skb_mstamp
.stamp_jiffies
= cookie_init_timestamp(req
);
3104 skb_mstamp_get(&skb
->skb_mstamp
);
3106 #ifdef CONFIG_TCP_MD5SIG
3108 md5
= tcp_rsk(req
)->af_specific
->req_md5_lookup(sk
, req_to_sk(req
));
3110 skb_set_hash(skb
, tcp_rsk(req
)->txhash
, PKT_HASH_TYPE_L4
);
3111 tcp_header_size
= tcp_synack_options(req
, mss
, skb
, &opts
, md5
, foc
) +
3114 skb_push(skb
, tcp_header_size
);
3115 skb_reset_transport_header(skb
);
3117 th
= (struct tcphdr
*)skb
->data
;
3118 memset(th
, 0, sizeof(struct tcphdr
));
3121 tcp_ecn_make_synack(req
, th
);
3122 th
->source
= htons(ireq
->ir_num
);
3123 th
->dest
= ireq
->ir_rmt_port
;
3124 /* Setting of flags are superfluous here for callers (and ECE is
3125 * not even correctly set)
3127 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
3128 TCPHDR_SYN
| TCPHDR_ACK
);
3130 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
3131 /* XXX data is queued and acked as is. No buffer/window check */
3132 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
3134 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3135 th
->window
= htons(min(req
->rsk_rcv_wnd
, 65535U));
3136 tcp_options_write((__be32
*)(th
+ 1), NULL
, &opts
);
3137 th
->doff
= (tcp_header_size
>> 2);
3138 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
3140 #ifdef CONFIG_TCP_MD5SIG
3141 /* Okay, we have all we need - do the md5 hash if needed */
3143 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
3144 md5
, req_to_sk(req
), skb
);
3148 /* Do not fool tcpdump (if any), clean our debris */
3152 EXPORT_SYMBOL(tcp_make_synack
);
3154 static void tcp_ca_dst_init(struct sock
*sk
, const struct dst_entry
*dst
)
3156 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3157 const struct tcp_congestion_ops
*ca
;
3158 u32 ca_key
= dst_metric(dst
, RTAX_CC_ALGO
);
3160 if (ca_key
== TCP_CA_UNSPEC
)
3164 ca
= tcp_ca_find_key(ca_key
);
3165 if (likely(ca
&& try_module_get(ca
->owner
))) {
3166 module_put(icsk
->icsk_ca_ops
->owner
);
3167 icsk
->icsk_ca_dst_locked
= tcp_ca_dst_locked(dst
);
3168 icsk
->icsk_ca_ops
= ca
;
3173 /* Do all connect socket setups that can be done AF independent. */
3174 static void tcp_connect_init(struct sock
*sk
)
3176 const struct dst_entry
*dst
= __sk_dst_get(sk
);
3177 struct tcp_sock
*tp
= tcp_sk(sk
);
3180 /* We'll fix this up when we get a response from the other end.
3181 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3183 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
3184 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
3186 #ifdef CONFIG_TCP_MD5SIG
3187 if (tp
->af_specific
->md5_lookup(sk
, sk
))
3188 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
3191 /* If user gave his TCP_MAXSEG, record it to clamp */
3192 if (tp
->rx_opt
.user_mss
)
3193 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3196 tcp_sync_mss(sk
, dst_mtu(dst
));
3198 tcp_ca_dst_init(sk
, dst
);
3200 if (!tp
->window_clamp
)
3201 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
3202 tp
->advmss
= dst_metric_advmss(dst
);
3203 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
3204 tp
->advmss
= tp
->rx_opt
.user_mss
;
3206 tcp_initialize_rcv_mss(sk
);
3208 /* limit the window selection if the user enforce a smaller rx buffer */
3209 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
3210 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
3211 tp
->window_clamp
= tcp_full_space(sk
);
3213 tcp_select_initial_window(tcp_full_space(sk
),
3214 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
3217 sysctl_tcp_window_scaling
,
3219 dst_metric(dst
, RTAX_INITRWND
));
3221 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
3222 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
3225 sock_reset_flag(sk
, SOCK_DONE
);
3228 tp
->snd_una
= tp
->write_seq
;
3229 tp
->snd_sml
= tp
->write_seq
;
3230 tp
->snd_up
= tp
->write_seq
;
3231 tp
->snd_nxt
= tp
->write_seq
;
3233 if (likely(!tp
->repair
))
3236 tp
->rcv_tstamp
= tcp_time_stamp
;
3237 tp
->rcv_wup
= tp
->rcv_nxt
;
3238 tp
->copied_seq
= tp
->rcv_nxt
;
3240 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
3241 inet_csk(sk
)->icsk_retransmits
= 0;
3242 tcp_clear_retrans(tp
);
3245 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
3247 struct tcp_sock
*tp
= tcp_sk(sk
);
3248 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
3250 tcb
->end_seq
+= skb
->len
;
3251 __skb_header_release(skb
);
3252 __tcp_add_write_queue_tail(sk
, skb
);
3253 sk
->sk_wmem_queued
+= skb
->truesize
;
3254 sk_mem_charge(sk
, skb
->truesize
);
3255 tp
->write_seq
= tcb
->end_seq
;
3256 tp
->packets_out
+= tcp_skb_pcount(skb
);
3259 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3260 * queue a data-only packet after the regular SYN, such that regular SYNs
3261 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3262 * only the SYN sequence, the data are retransmitted in the first ACK.
3263 * If cookie is not cached or other error occurs, falls back to send a
3264 * regular SYN with Fast Open cookie request option.
3266 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
3268 struct tcp_sock
*tp
= tcp_sk(sk
);
3269 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
3270 int syn_loss
= 0, space
, err
= 0;
3271 unsigned long last_syn_loss
= 0;
3272 struct sk_buff
*syn_data
;
3274 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
3275 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
3276 &syn_loss
, &last_syn_loss
);
3277 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3279 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
3280 fo
->cookie
.len
= -1;
3284 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
3285 fo
->cookie
.len
= -1;
3286 else if (fo
->cookie
.len
<= 0)
3289 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3290 * user-MSS. Reserve maximum option space for middleboxes that add
3291 * private TCP options. The cost is reduced data space in SYN :(
3293 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
3294 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3295 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
3296 MAX_TCP_OPTION_SPACE
;
3298 space
= min_t(size_t, space
, fo
->size
);
3300 /* limit to order-0 allocations */
3301 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
3303 syn_data
= sk_stream_alloc_skb(sk
, space
, sk
->sk_allocation
, false);
3306 syn_data
->ip_summed
= CHECKSUM_PARTIAL
;
3307 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
3309 int copied
= copy_from_iter(skb_put(syn_data
, space
), space
,
3310 &fo
->data
->msg_iter
);
3311 if (unlikely(!copied
)) {
3312 kfree_skb(syn_data
);
3315 if (copied
!= space
) {
3316 skb_trim(syn_data
, copied
);
3320 /* No more data pending in inet_wait_for_connect() */
3321 if (space
== fo
->size
)
3325 tcp_connect_queue_skb(sk
, syn_data
);
3327 tcp_chrono_start(sk
, TCP_CHRONO_BUSY
);
3329 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
3331 syn
->skb_mstamp
= syn_data
->skb_mstamp
;
3333 /* Now full SYN+DATA was cloned and sent (or not),
3334 * remove the SYN from the original skb (syn_data)
3335 * we keep in write queue in case of a retransmit, as we
3336 * also have the SYN packet (with no data) in the same queue.
3338 TCP_SKB_CB(syn_data
)->seq
++;
3339 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
3341 tp
->syn_data
= (fo
->copied
> 0);
3342 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3347 /* Send a regular SYN with Fast Open cookie request option */
3348 if (fo
->cookie
.len
> 0)
3350 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3352 tp
->syn_fastopen
= 0;
3354 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3358 /* Build a SYN and send it off. */
3359 int tcp_connect(struct sock
*sk
)
3361 struct tcp_sock
*tp
= tcp_sk(sk
);
3362 struct sk_buff
*buff
;
3365 tcp_connect_init(sk
);
3367 if (unlikely(tp
->repair
)) {
3368 tcp_finish_connect(sk
, NULL
);
3372 buff
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
, true);
3373 if (unlikely(!buff
))
3376 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3377 tp
->retrans_stamp
= tcp_time_stamp
;
3378 tcp_connect_queue_skb(sk
, buff
);
3379 tcp_ecn_send_syn(sk
, buff
);
3381 /* Send off SYN; include data in Fast Open. */
3382 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3383 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3384 if (err
== -ECONNREFUSED
)
3387 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3388 * in order to make this packet get counted in tcpOutSegs.
3390 tp
->snd_nxt
= tp
->write_seq
;
3391 tp
->pushed_seq
= tp
->write_seq
;
3392 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3394 /* Timer for repeating the SYN until an answer. */
3395 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3396 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3399 EXPORT_SYMBOL(tcp_connect
);
3401 /* Send out a delayed ack, the caller does the policy checking
3402 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3405 void tcp_send_delayed_ack(struct sock
*sk
)
3407 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3408 int ato
= icsk
->icsk_ack
.ato
;
3409 unsigned long timeout
;
3411 tcp_ca_event(sk
, CA_EVENT_DELAYED_ACK
);
3413 if (ato
> TCP_DELACK_MIN
) {
3414 const struct tcp_sock
*tp
= tcp_sk(sk
);
3415 int max_ato
= HZ
/ 2;
3417 if (icsk
->icsk_ack
.pingpong
||
3418 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3419 max_ato
= TCP_DELACK_MAX
;
3421 /* Slow path, intersegment interval is "high". */
3423 /* If some rtt estimate is known, use it to bound delayed ack.
3424 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3428 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3435 ato
= min(ato
, max_ato
);
3438 /* Stay within the limit we were given */
3439 timeout
= jiffies
+ ato
;
3441 /* Use new timeout only if there wasn't a older one earlier. */
3442 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3443 /* If delack timer was blocked or is about to expire,
3446 if (icsk
->icsk_ack
.blocked
||
3447 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3452 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3453 timeout
= icsk
->icsk_ack
.timeout
;
3455 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3456 icsk
->icsk_ack
.timeout
= timeout
;
3457 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3460 /* This routine sends an ack and also updates the window. */
3461 void tcp_send_ack(struct sock
*sk
)
3463 struct sk_buff
*buff
;
3465 /* If we have been reset, we may not send again. */
3466 if (sk
->sk_state
== TCP_CLOSE
)
3469 tcp_ca_event(sk
, CA_EVENT_NON_DELAYED_ACK
);
3471 /* We are not putting this on the write queue, so
3472 * tcp_transmit_skb() will set the ownership to this
3475 buff
= alloc_skb(MAX_TCP_HEADER
,
3476 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3477 if (unlikely(!buff
)) {
3478 inet_csk_schedule_ack(sk
);
3479 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3480 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3481 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3485 /* Reserve space for headers and prepare control bits. */
3486 skb_reserve(buff
, MAX_TCP_HEADER
);
3487 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3489 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3491 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3493 skb_set_tcp_pure_ack(buff
);
3495 /* Send it off, this clears delayed acks for us. */
3496 skb_mstamp_get(&buff
->skb_mstamp
);
3497 tcp_transmit_skb(sk
, buff
, 0, (__force gfp_t
)0);
3499 EXPORT_SYMBOL_GPL(tcp_send_ack
);
3501 /* This routine sends a packet with an out of date sequence
3502 * number. It assumes the other end will try to ack it.
3504 * Question: what should we make while urgent mode?
3505 * 4.4BSD forces sending single byte of data. We cannot send
3506 * out of window data, because we have SND.NXT==SND.MAX...
3508 * Current solution: to send TWO zero-length segments in urgent mode:
3509 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3510 * out-of-date with SND.UNA-1 to probe window.
3512 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
, int mib
)
3514 struct tcp_sock
*tp
= tcp_sk(sk
);
3515 struct sk_buff
*skb
;
3517 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3518 skb
= alloc_skb(MAX_TCP_HEADER
,
3519 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3523 /* Reserve space for headers and set control bits. */
3524 skb_reserve(skb
, MAX_TCP_HEADER
);
3525 /* Use a previous sequence. This should cause the other
3526 * end to send an ack. Don't queue or clone SKB, just
3529 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3530 skb_mstamp_get(&skb
->skb_mstamp
);
3531 NET_INC_STATS(sock_net(sk
), mib
);
3532 return tcp_transmit_skb(sk
, skb
, 0, (__force gfp_t
)0);
3535 void tcp_send_window_probe(struct sock
*sk
)
3537 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3538 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3539 tcp_xmit_probe_skb(sk
, 0, LINUX_MIB_TCPWINPROBE
);
3543 /* Initiate keepalive or window probe from timer. */
3544 int tcp_write_wakeup(struct sock
*sk
, int mib
)
3546 struct tcp_sock
*tp
= tcp_sk(sk
);
3547 struct sk_buff
*skb
;
3549 if (sk
->sk_state
== TCP_CLOSE
)
3552 skb
= tcp_send_head(sk
);
3553 if (skb
&& before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3555 unsigned int mss
= tcp_current_mss(sk
);
3556 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3558 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3559 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3561 /* We are probing the opening of a window
3562 * but the window size is != 0
3563 * must have been a result SWS avoidance ( sender )
3565 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3567 seg_size
= min(seg_size
, mss
);
3568 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3569 if (tcp_fragment(sk
, skb
, seg_size
, mss
, GFP_ATOMIC
))
3571 } else if (!tcp_skb_pcount(skb
))
3572 tcp_set_skb_tso_segs(skb
, mss
);
3574 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3575 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3577 tcp_event_new_data_sent(sk
, skb
);
3580 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3581 tcp_xmit_probe_skb(sk
, 1, mib
);
3582 return tcp_xmit_probe_skb(sk
, 0, mib
);
3586 /* A window probe timeout has occurred. If window is not closed send
3587 * a partial packet else a zero probe.
3589 void tcp_send_probe0(struct sock
*sk
)
3591 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3592 struct tcp_sock
*tp
= tcp_sk(sk
);
3593 struct net
*net
= sock_net(sk
);
3594 unsigned long probe_max
;
3597 err
= tcp_write_wakeup(sk
, LINUX_MIB_TCPWINPROBE
);
3599 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3600 /* Cancel probe timer, if it is not required. */
3601 icsk
->icsk_probes_out
= 0;
3602 icsk
->icsk_backoff
= 0;
3607 if (icsk
->icsk_backoff
< net
->ipv4
.sysctl_tcp_retries2
)
3608 icsk
->icsk_backoff
++;
3609 icsk
->icsk_probes_out
++;
3610 probe_max
= TCP_RTO_MAX
;
3612 /* If packet was not sent due to local congestion,
3613 * do not backoff and do not remember icsk_probes_out.
3614 * Let local senders to fight for local resources.
3616 * Use accumulated backoff yet.
3618 if (!icsk
->icsk_probes_out
)
3619 icsk
->icsk_probes_out
= 1;
3620 probe_max
= TCP_RESOURCE_PROBE_INTERVAL
;
3622 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3623 tcp_probe0_when(sk
, probe_max
),
3627 int tcp_rtx_synack(const struct sock
*sk
, struct request_sock
*req
)
3629 const struct tcp_request_sock_ops
*af_ops
= tcp_rsk(req
)->af_specific
;
3633 tcp_rsk(req
)->txhash
= net_tx_rndhash();
3634 res
= af_ops
->send_synack(sk
, NULL
, &fl
, req
, NULL
, TCP_SYNACK_NORMAL
);
3636 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
3637 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
3638 if (unlikely(tcp_passive_fastopen(sk
)))
3639 tcp_sk(sk
)->total_retrans
++;
3643 EXPORT_SYMBOL(tcp_rtx_synack
);