2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
53 /* Default TSQ limit of four TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly
= 262144;
56 /* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
60 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
65 unsigned int sysctl_tcp_notsent_lowat __read_mostly
= UINT_MAX
;
66 EXPORT_SYMBOL(sysctl_tcp_notsent_lowat
);
68 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
69 int push_one
, gfp_t gfp
);
71 /* Account for new data that has been sent to the network. */
72 static void tcp_event_new_data_sent(struct sock
*sk
, const struct sk_buff
*skb
)
74 struct inet_connection_sock
*icsk
= inet_csk(sk
);
75 struct tcp_sock
*tp
= tcp_sk(sk
);
76 unsigned int prior_packets
= tp
->packets_out
;
78 tcp_advance_send_head(sk
, skb
);
79 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
81 tp
->packets_out
+= tcp_skb_pcount(skb
);
82 if (!prior_packets
|| icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
83 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
87 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
,
91 /* SND.NXT, if window was not shrunk.
92 * If window has been shrunk, what should we make? It is not clear at all.
93 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
94 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
95 * invalid. OK, let's make this for now:
97 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
99 const struct tcp_sock
*tp
= tcp_sk(sk
);
101 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
104 return tcp_wnd_end(tp
);
107 /* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
121 static __u16
tcp_advertise_mss(struct sock
*sk
)
123 struct tcp_sock
*tp
= tcp_sk(sk
);
124 const struct dst_entry
*dst
= __sk_dst_get(sk
);
125 int mss
= tp
->advmss
;
128 unsigned int metric
= dst_metric_advmss(dst
);
139 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism. */
141 static void tcp_cwnd_restart(struct sock
*sk
, const struct dst_entry
*dst
)
143 struct tcp_sock
*tp
= tcp_sk(sk
);
144 s32 delta
= tcp_time_stamp
- tp
->lsndtime
;
145 u32 restart_cwnd
= tcp_init_cwnd(tp
, dst
);
146 u32 cwnd
= tp
->snd_cwnd
;
148 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
150 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
151 restart_cwnd
= min(restart_cwnd
, cwnd
);
153 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
155 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
156 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
157 tp
->snd_cwnd_used
= 0;
160 /* Congestion state accounting after a packet has been sent. */
161 static void tcp_event_data_sent(struct tcp_sock
*tp
,
164 struct inet_connection_sock
*icsk
= inet_csk(sk
);
165 const u32 now
= tcp_time_stamp
;
167 if (sysctl_tcp_slow_start_after_idle
&&
168 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
169 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
173 /* If it is a reply for ato after last received
174 * packet, enter pingpong mode.
176 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
177 icsk
->icsk_ack
.pingpong
= 1;
180 /* Account for an ACK we sent. */
181 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
183 tcp_dec_quickack_mode(sk
, pkts
);
184 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
188 u32
tcp_default_init_rwnd(u32 mss
)
190 /* Initial receive window should be twice of TCP_INIT_CWND to
191 * enable proper sending of new unsent data during fast recovery
192 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
193 * limit when mss is larger than 1460.
195 u32 init_rwnd
= TCP_INIT_CWND
* 2;
198 init_rwnd
= max((1460 * init_rwnd
) / mss
, 2U);
202 /* Determine a window scaling and initial window to offer.
203 * Based on the assumption that the given amount of space
204 * will be offered. Store the results in the tp structure.
205 * NOTE: for smooth operation initial space offering should
206 * be a multiple of mss if possible. We assume here that mss >= 1.
207 * This MUST be enforced by all callers.
209 void tcp_select_initial_window(int __space
, __u32 mss
,
210 __u32
*rcv_wnd
, __u32
*window_clamp
,
211 int wscale_ok
, __u8
*rcv_wscale
,
214 unsigned int space
= (__space
< 0 ? 0 : __space
);
216 /* If no clamp set the clamp to the max possible scaled window */
217 if (*window_clamp
== 0)
218 (*window_clamp
) = (65535 << 14);
219 space
= min(*window_clamp
, space
);
221 /* Quantize space offering to a multiple of mss if possible. */
223 space
= (space
/ mss
) * mss
;
225 /* NOTE: offering an initial window larger than 32767
226 * will break some buggy TCP stacks. If the admin tells us
227 * it is likely we could be speaking with such a buggy stack
228 * we will truncate our initial window offering to 32K-1
229 * unless the remote has sent us a window scaling option,
230 * which we interpret as a sign the remote TCP is not
231 * misinterpreting the window field as a signed quantity.
233 if (sysctl_tcp_workaround_signed_windows
)
234 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
240 /* Set window scaling on max possible window
241 * See RFC1323 for an explanation of the limit to 14
243 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
244 space
= min_t(u32
, space
, *window_clamp
);
245 while (space
> 65535 && (*rcv_wscale
) < 14) {
251 if (mss
> (1 << *rcv_wscale
)) {
252 if (!init_rcv_wnd
) /* Use default unless specified otherwise */
253 init_rcv_wnd
= tcp_default_init_rwnd(mss
);
254 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
257 /* Set the clamp no higher than max representable value */
258 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
260 EXPORT_SYMBOL(tcp_select_initial_window
);
262 /* Chose a new window to advertise, update state in tcp_sock for the
263 * socket, and return result with RFC1323 scaling applied. The return
264 * value can be stuffed directly into th->window for an outgoing
267 static u16
tcp_select_window(struct sock
*sk
)
269 struct tcp_sock
*tp
= tcp_sk(sk
);
270 u32 old_win
= tp
->rcv_wnd
;
271 u32 cur_win
= tcp_receive_window(tp
);
272 u32 new_win
= __tcp_select_window(sk
);
274 /* Never shrink the offered window */
275 if (new_win
< cur_win
) {
276 /* Danger Will Robinson!
277 * Don't update rcv_wup/rcv_wnd here or else
278 * we will not be able to advertise a zero
279 * window in time. --DaveM
281 * Relax Will Robinson.
284 NET_INC_STATS(sock_net(sk
),
285 LINUX_MIB_TCPWANTZEROWINDOWADV
);
286 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
288 tp
->rcv_wnd
= new_win
;
289 tp
->rcv_wup
= tp
->rcv_nxt
;
291 /* Make sure we do not exceed the maximum possible
294 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
295 new_win
= min(new_win
, MAX_TCP_WINDOW
);
297 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
299 /* RFC1323 scaling applied */
300 new_win
>>= tp
->rx_opt
.rcv_wscale
;
302 /* If we advertise zero window, disable fast path. */
306 NET_INC_STATS(sock_net(sk
),
307 LINUX_MIB_TCPTOZEROWINDOWADV
);
308 } else if (old_win
== 0) {
309 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFROMZEROWINDOWADV
);
315 /* Packet ECN state for a SYN-ACK */
316 static void tcp_ecn_send_synack(struct sock
*sk
, struct sk_buff
*skb
)
318 const struct tcp_sock
*tp
= tcp_sk(sk
);
320 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
321 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
322 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
323 else if (tcp_ca_needs_ecn(sk
))
327 /* Packet ECN state for a SYN. */
328 static void tcp_ecn_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
330 struct tcp_sock
*tp
= tcp_sk(sk
);
331 bool use_ecn
= sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1 ||
332 tcp_ca_needs_ecn(sk
);
335 const struct dst_entry
*dst
= __sk_dst_get(sk
);
337 if (dst
&& dst_feature(dst
, RTAX_FEATURE_ECN
))
344 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
345 tp
->ecn_flags
= TCP_ECN_OK
;
346 if (tcp_ca_needs_ecn(sk
))
351 static void tcp_ecn_clear_syn(struct sock
*sk
, struct sk_buff
*skb
)
353 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn_fallback
)
354 /* tp->ecn_flags are cleared at a later point in time when
355 * SYN ACK is ultimatively being received.
357 TCP_SKB_CB(skb
)->tcp_flags
&= ~(TCPHDR_ECE
| TCPHDR_CWR
);
361 tcp_ecn_make_synack(const struct request_sock
*req
, struct tcphdr
*th
,
364 if (inet_rsk(req
)->ecn_ok
) {
366 if (tcp_ca_needs_ecn(sk
))
371 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
374 static void tcp_ecn_send(struct sock
*sk
, struct sk_buff
*skb
,
377 struct tcp_sock
*tp
= tcp_sk(sk
);
379 if (tp
->ecn_flags
& TCP_ECN_OK
) {
380 /* Not-retransmitted data segment: set ECT and inject CWR. */
381 if (skb
->len
!= tcp_header_len
&&
382 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
384 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
385 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
386 tcp_hdr(skb
)->cwr
= 1;
387 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
389 } else if (!tcp_ca_needs_ecn(sk
)) {
390 /* ACK or retransmitted segment: clear ECT|CE */
391 INET_ECN_dontxmit(sk
);
393 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
394 tcp_hdr(skb
)->ece
= 1;
398 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
399 * auto increment end seqno.
401 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
403 skb
->ip_summed
= CHECKSUM_PARTIAL
;
406 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
407 TCP_SKB_CB(skb
)->sacked
= 0;
409 tcp_skb_pcount_set(skb
, 1);
411 TCP_SKB_CB(skb
)->seq
= seq
;
412 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
414 TCP_SKB_CB(skb
)->end_seq
= seq
;
417 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
419 return tp
->snd_una
!= tp
->snd_up
;
422 #define OPTION_SACK_ADVERTISE (1 << 0)
423 #define OPTION_TS (1 << 1)
424 #define OPTION_MD5 (1 << 2)
425 #define OPTION_WSCALE (1 << 3)
426 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
428 struct tcp_out_options
{
429 u16 options
; /* bit field of OPTION_* */
430 u16 mss
; /* 0 to disable */
431 u8 ws
; /* window scale, 0 to disable */
432 u8 num_sack_blocks
; /* number of SACK blocks to include */
433 u8 hash_size
; /* bytes in hash_location */
434 __u8
*hash_location
; /* temporary pointer, overloaded */
435 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
436 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
439 /* Write previously computed TCP options to the packet.
441 * Beware: Something in the Internet is very sensitive to the ordering of
442 * TCP options, we learned this through the hard way, so be careful here.
443 * Luckily we can at least blame others for their non-compliance but from
444 * inter-operability perspective it seems that we're somewhat stuck with
445 * the ordering which we have been using if we want to keep working with
446 * those broken things (not that it currently hurts anybody as there isn't
447 * particular reason why the ordering would need to be changed).
449 * At least SACK_PERM as the first option is known to lead to a disaster
450 * (but it may well be that other scenarios fail similarly).
452 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
453 struct tcp_out_options
*opts
)
455 u16 options
= opts
->options
; /* mungable copy */
457 if (unlikely(OPTION_MD5
& options
)) {
458 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
459 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
460 /* overload cookie hash location */
461 opts
->hash_location
= (__u8
*)ptr
;
465 if (unlikely(opts
->mss
)) {
466 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
467 (TCPOLEN_MSS
<< 16) |
471 if (likely(OPTION_TS
& options
)) {
472 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
473 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
474 (TCPOLEN_SACK_PERM
<< 16) |
475 (TCPOPT_TIMESTAMP
<< 8) |
477 options
&= ~OPTION_SACK_ADVERTISE
;
479 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
481 (TCPOPT_TIMESTAMP
<< 8) |
484 *ptr
++ = htonl(opts
->tsval
);
485 *ptr
++ = htonl(opts
->tsecr
);
488 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
489 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
491 (TCPOPT_SACK_PERM
<< 8) |
495 if (unlikely(OPTION_WSCALE
& options
)) {
496 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
497 (TCPOPT_WINDOW
<< 16) |
498 (TCPOLEN_WINDOW
<< 8) |
502 if (unlikely(opts
->num_sack_blocks
)) {
503 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
504 tp
->duplicate_sack
: tp
->selective_acks
;
507 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
510 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
511 TCPOLEN_SACK_PERBLOCK
)));
513 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
515 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
516 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
519 tp
->rx_opt
.dsack
= 0;
522 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
523 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
525 u32 len
; /* Fast Open option length */
528 len
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
529 *ptr
= htonl((TCPOPT_EXP
<< 24) | (len
<< 16) |
530 TCPOPT_FASTOPEN_MAGIC
);
531 p
+= TCPOLEN_EXP_FASTOPEN_BASE
;
533 len
= TCPOLEN_FASTOPEN_BASE
+ foc
->len
;
534 *p
++ = TCPOPT_FASTOPEN
;
538 memcpy(p
, foc
->val
, foc
->len
);
539 if ((len
& 3) == 2) {
540 p
[foc
->len
] = TCPOPT_NOP
;
541 p
[foc
->len
+ 1] = TCPOPT_NOP
;
543 ptr
+= (len
+ 3) >> 2;
547 /* Compute TCP options for SYN packets. This is not the final
548 * network wire format yet.
550 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
551 struct tcp_out_options
*opts
,
552 struct tcp_md5sig_key
**md5
)
554 struct tcp_sock
*tp
= tcp_sk(sk
);
555 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
556 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
558 #ifdef CONFIG_TCP_MD5SIG
559 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
561 opts
->options
|= OPTION_MD5
;
562 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
568 /* We always get an MSS option. The option bytes which will be seen in
569 * normal data packets should timestamps be used, must be in the MSS
570 * advertised. But we subtract them from tp->mss_cache so that
571 * calculations in tcp_sendmsg are simpler etc. So account for this
572 * fact here if necessary. If we don't do this correctly, as a
573 * receiver we won't recognize data packets as being full sized when we
574 * should, and thus we won't abide by the delayed ACK rules correctly.
575 * SACKs don't matter, we never delay an ACK when we have any of those
577 opts
->mss
= tcp_advertise_mss(sk
);
578 remaining
-= TCPOLEN_MSS_ALIGNED
;
580 if (likely(sysctl_tcp_timestamps
&& !*md5
)) {
581 opts
->options
|= OPTION_TS
;
582 opts
->tsval
= tcp_skb_timestamp(skb
) + tp
->tsoffset
;
583 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
584 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
586 if (likely(sysctl_tcp_window_scaling
)) {
587 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
588 opts
->options
|= OPTION_WSCALE
;
589 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
591 if (likely(sysctl_tcp_sack
)) {
592 opts
->options
|= OPTION_SACK_ADVERTISE
;
593 if (unlikely(!(OPTION_TS
& opts
->options
)))
594 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
597 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
598 u32 need
= fastopen
->cookie
.len
;
600 need
+= fastopen
->cookie
.exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
601 TCPOLEN_FASTOPEN_BASE
;
602 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
603 if (remaining
>= need
) {
604 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
605 opts
->fastopen_cookie
= &fastopen
->cookie
;
607 tp
->syn_fastopen
= 1;
608 tp
->syn_fastopen_exp
= fastopen
->cookie
.exp
? 1 : 0;
612 return MAX_TCP_OPTION_SPACE
- remaining
;
615 /* Set up TCP options for SYN-ACKs. */
616 static unsigned int tcp_synack_options(struct sock
*sk
,
617 struct request_sock
*req
,
618 unsigned int mss
, struct sk_buff
*skb
,
619 struct tcp_out_options
*opts
,
620 const struct tcp_md5sig_key
*md5
,
621 struct tcp_fastopen_cookie
*foc
)
623 struct inet_request_sock
*ireq
= inet_rsk(req
);
624 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
626 #ifdef CONFIG_TCP_MD5SIG
628 opts
->options
|= OPTION_MD5
;
629 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
631 /* We can't fit any SACK blocks in a packet with MD5 + TS
632 * options. There was discussion about disabling SACK
633 * rather than TS in order to fit in better with old,
634 * buggy kernels, but that was deemed to be unnecessary.
636 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
640 /* We always send an MSS option. */
642 remaining
-= TCPOLEN_MSS_ALIGNED
;
644 if (likely(ireq
->wscale_ok
)) {
645 opts
->ws
= ireq
->rcv_wscale
;
646 opts
->options
|= OPTION_WSCALE
;
647 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
649 if (likely(ireq
->tstamp_ok
)) {
650 opts
->options
|= OPTION_TS
;
651 opts
->tsval
= tcp_skb_timestamp(skb
);
652 opts
->tsecr
= req
->ts_recent
;
653 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
655 if (likely(ireq
->sack_ok
)) {
656 opts
->options
|= OPTION_SACK_ADVERTISE
;
657 if (unlikely(!ireq
->tstamp_ok
))
658 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
660 if (foc
!= NULL
&& foc
->len
>= 0) {
663 need
+= foc
->exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
664 TCPOLEN_FASTOPEN_BASE
;
665 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
666 if (remaining
>= need
) {
667 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
668 opts
->fastopen_cookie
= foc
;
673 return MAX_TCP_OPTION_SPACE
- remaining
;
676 /* Compute TCP options for ESTABLISHED sockets. This is not the
677 * final wire format yet.
679 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
680 struct tcp_out_options
*opts
,
681 struct tcp_md5sig_key
**md5
)
683 struct tcp_sock
*tp
= tcp_sk(sk
);
684 unsigned int size
= 0;
685 unsigned int eff_sacks
;
689 #ifdef CONFIG_TCP_MD5SIG
690 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
691 if (unlikely(*md5
)) {
692 opts
->options
|= OPTION_MD5
;
693 size
+= TCPOLEN_MD5SIG_ALIGNED
;
699 if (likely(tp
->rx_opt
.tstamp_ok
)) {
700 opts
->options
|= OPTION_TS
;
701 opts
->tsval
= skb
? tcp_skb_timestamp(skb
) + tp
->tsoffset
: 0;
702 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
703 size
+= TCPOLEN_TSTAMP_ALIGNED
;
706 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
707 if (unlikely(eff_sacks
)) {
708 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
709 opts
->num_sack_blocks
=
710 min_t(unsigned int, eff_sacks
,
711 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
712 TCPOLEN_SACK_PERBLOCK
);
713 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
714 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
721 /* TCP SMALL QUEUES (TSQ)
723 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
724 * to reduce RTT and bufferbloat.
725 * We do this using a special skb destructor (tcp_wfree).
727 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
728 * needs to be reallocated in a driver.
729 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
731 * Since transmit from skb destructor is forbidden, we use a tasklet
732 * to process all sockets that eventually need to send more skbs.
733 * We use one tasklet per cpu, with its own queue of sockets.
736 struct tasklet_struct tasklet
;
737 struct list_head head
; /* queue of tcp sockets */
739 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
741 static void tcp_tsq_handler(struct sock
*sk
)
743 if ((1 << sk
->sk_state
) &
744 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
745 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
))
746 tcp_write_xmit(sk
, tcp_current_mss(sk
), tcp_sk(sk
)->nonagle
,
750 * One tasklet per cpu tries to send more skbs.
751 * We run in tasklet context but need to disable irqs when
752 * transferring tsq->head because tcp_wfree() might
753 * interrupt us (non NAPI drivers)
755 static void tcp_tasklet_func(unsigned long data
)
757 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
760 struct list_head
*q
, *n
;
764 local_irq_save(flags
);
765 list_splice_init(&tsq
->head
, &list
);
766 local_irq_restore(flags
);
768 list_for_each_safe(q
, n
, &list
) {
769 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
770 list_del(&tp
->tsq_node
);
772 sk
= (struct sock
*)tp
;
775 if (!sock_owned_by_user(sk
)) {
778 /* defer the work to tcp_release_cb() */
779 set_bit(TCP_TSQ_DEFERRED
, &tp
->tsq_flags
);
783 clear_bit(TSQ_QUEUED
, &tp
->tsq_flags
);
788 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
789 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
790 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
791 (1UL << TCP_MTU_REDUCED_DEFERRED))
793 * tcp_release_cb - tcp release_sock() callback
796 * called from release_sock() to perform protocol dependent
797 * actions before socket release.
799 void tcp_release_cb(struct sock
*sk
)
801 struct tcp_sock
*tp
= tcp_sk(sk
);
802 unsigned long flags
, nflags
;
804 /* perform an atomic operation only if at least one flag is set */
806 flags
= tp
->tsq_flags
;
807 if (!(flags
& TCP_DEFERRED_ALL
))
809 nflags
= flags
& ~TCP_DEFERRED_ALL
;
810 } while (cmpxchg(&tp
->tsq_flags
, flags
, nflags
) != flags
);
812 if (flags
& (1UL << TCP_TSQ_DEFERRED
))
815 /* Here begins the tricky part :
816 * We are called from release_sock() with :
818 * 2) sk_lock.slock spinlock held
819 * 3) socket owned by us (sk->sk_lock.owned == 1)
821 * But following code is meant to be called from BH handlers,
822 * so we should keep BH disabled, but early release socket ownership
824 sock_release_ownership(sk
);
826 if (flags
& (1UL << TCP_WRITE_TIMER_DEFERRED
)) {
827 tcp_write_timer_handler(sk
);
830 if (flags
& (1UL << TCP_DELACK_TIMER_DEFERRED
)) {
831 tcp_delack_timer_handler(sk
);
834 if (flags
& (1UL << TCP_MTU_REDUCED_DEFERRED
)) {
835 inet_csk(sk
)->icsk_af_ops
->mtu_reduced(sk
);
839 EXPORT_SYMBOL(tcp_release_cb
);
841 void __init
tcp_tasklet_init(void)
845 for_each_possible_cpu(i
) {
846 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
848 INIT_LIST_HEAD(&tsq
->head
);
849 tasklet_init(&tsq
->tasklet
,
856 * Write buffer destructor automatically called from kfree_skb.
857 * We can't xmit new skbs from this context, as we might already
860 void tcp_wfree(struct sk_buff
*skb
)
862 struct sock
*sk
= skb
->sk
;
863 struct tcp_sock
*tp
= tcp_sk(sk
);
866 /* Keep one reference on sk_wmem_alloc.
867 * Will be released by sk_free() from here or tcp_tasklet_func()
869 wmem
= atomic_sub_return(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
871 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
872 * Wait until our queues (qdisc + devices) are drained.
874 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
875 * - chance for incoming ACK (processed by another cpu maybe)
876 * to migrate this flow (skb->ooo_okay will be eventually set)
878 if (wmem
>= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current
)
881 if (test_and_clear_bit(TSQ_THROTTLED
, &tp
->tsq_flags
) &&
882 !test_and_set_bit(TSQ_QUEUED
, &tp
->tsq_flags
)) {
884 struct tsq_tasklet
*tsq
;
886 /* queue this socket to tasklet queue */
887 local_irq_save(flags
);
888 tsq
= this_cpu_ptr(&tsq_tasklet
);
889 list_add(&tp
->tsq_node
, &tsq
->head
);
890 tasklet_schedule(&tsq
->tasklet
);
891 local_irq_restore(flags
);
898 /* This routine actually transmits TCP packets queued in by
899 * tcp_do_sendmsg(). This is used by both the initial
900 * transmission and possible later retransmissions.
901 * All SKB's seen here are completely headerless. It is our
902 * job to build the TCP header, and pass the packet down to
903 * IP so it can do the same plus pass the packet off to the
906 * We are working here with either a clone of the original
907 * SKB, or a fresh unique copy made by the retransmit engine.
909 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
912 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
913 struct inet_sock
*inet
;
915 struct tcp_skb_cb
*tcb
;
916 struct tcp_out_options opts
;
917 unsigned int tcp_options_size
, tcp_header_size
;
918 struct tcp_md5sig_key
*md5
;
922 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
925 skb_mstamp_get(&skb
->skb_mstamp
);
927 if (unlikely(skb_cloned(skb
)))
928 skb
= pskb_copy(skb
, gfp_mask
);
930 skb
= skb_clone(skb
, gfp_mask
);
937 tcb
= TCP_SKB_CB(skb
);
938 memset(&opts
, 0, sizeof(opts
));
940 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
941 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
943 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
945 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
947 if (tcp_packets_in_flight(tp
) == 0)
948 tcp_ca_event(sk
, CA_EVENT_TX_START
);
950 /* if no packet is in qdisc/device queue, then allow XPS to select
951 * another queue. We can be called from tcp_tsq_handler()
952 * which holds one reference to sk_wmem_alloc.
954 * TODO: Ideally, in-flight pure ACK packets should not matter here.
955 * One way to get this would be to set skb->truesize = 2 on them.
957 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) < SKB_TRUESIZE(1);
959 skb_push(skb
, tcp_header_size
);
960 skb_reset_transport_header(skb
);
964 skb
->destructor
= skb_is_tcp_pure_ack(skb
) ? sock_wfree
: tcp_wfree
;
965 skb_set_hash_from_sk(skb
, sk
);
966 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
968 /* Build TCP header and checksum it. */
970 th
->source
= inet
->inet_sport
;
971 th
->dest
= inet
->inet_dport
;
972 th
->seq
= htonl(tcb
->seq
);
973 th
->ack_seq
= htonl(tp
->rcv_nxt
);
974 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
977 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
978 /* RFC1323: The window in SYN & SYN/ACK segments
981 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
983 th
->window
= htons(tcp_select_window(sk
));
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
) == 0))
1002 tcp_ecn_send(sk
, skb
, tcp_header_size
);
1004 #ifdef CONFIG_TCP_MD5SIG
1005 /* Calculate the MD5 hash, as we have all we need now */
1007 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
1008 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
1013 icsk
->icsk_af_ops
->send_check(sk
, skb
);
1015 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
1016 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
1018 if (skb
->len
!= tcp_header_size
)
1019 tcp_event_data_sent(tp
, sk
);
1021 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
1022 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
1023 tcp_skb_pcount(skb
));
1025 tp
->segs_out
+= tcp_skb_pcount(skb
);
1026 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1027 skb_shinfo(skb
)->gso_segs
= tcp_skb_pcount(skb
);
1028 skb_shinfo(skb
)->gso_size
= tcp_skb_mss(skb
);
1030 /* Our usage of tstamp should remain private */
1031 skb
->tstamp
.tv64
= 0;
1033 /* Cleanup our debris for IP stacks */
1034 memset(skb
->cb
, 0, max(sizeof(struct inet_skb_parm
),
1035 sizeof(struct inet6_skb_parm
)));
1037 err
= icsk
->icsk_af_ops
->queue_xmit(sk
, skb
, &inet
->cork
.fl
);
1039 if (likely(err
<= 0))
1044 return net_xmit_eval(err
);
1047 /* This routine just queues the buffer for sending.
1049 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1050 * otherwise socket can stall.
1052 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1054 struct tcp_sock
*tp
= tcp_sk(sk
);
1056 /* Advance write_seq and place onto the write_queue. */
1057 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
1058 __skb_header_release(skb
);
1059 tcp_add_write_queue_tail(sk
, skb
);
1060 sk
->sk_wmem_queued
+= skb
->truesize
;
1061 sk_mem_charge(sk
, skb
->truesize
);
1064 /* Initialize TSO segments for a packet. */
1065 static void tcp_set_skb_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1067 if (skb
->len
<= mss_now
|| skb
->ip_summed
== CHECKSUM_NONE
) {
1068 /* Avoid the costly divide in the normal
1071 tcp_skb_pcount_set(skb
, 1);
1072 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1074 tcp_skb_pcount_set(skb
, DIV_ROUND_UP(skb
->len
, mss_now
));
1075 TCP_SKB_CB(skb
)->tcp_gso_size
= mss_now
;
1079 /* When a modification to fackets out becomes necessary, we need to check
1080 * skb is counted to fackets_out or not.
1082 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1085 struct tcp_sock
*tp
= tcp_sk(sk
);
1087 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1090 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1091 tp
->fackets_out
-= decr
;
1094 /* Pcount in the middle of the write queue got changed, we need to do various
1095 * tweaks to fix counters
1097 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1099 struct tcp_sock
*tp
= tcp_sk(sk
);
1101 tp
->packets_out
-= decr
;
1103 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1104 tp
->sacked_out
-= decr
;
1105 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1106 tp
->retrans_out
-= decr
;
1107 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1108 tp
->lost_out
-= decr
;
1110 /* Reno case is special. Sigh... */
1111 if (tcp_is_reno(tp
) && decr
> 0)
1112 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1114 tcp_adjust_fackets_out(sk
, skb
, decr
);
1116 if (tp
->lost_skb_hint
&&
1117 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1118 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1119 tp
->lost_cnt_hint
-= decr
;
1121 tcp_verify_left_out(tp
);
1124 static void tcp_fragment_tstamp(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1126 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1128 if (unlikely(shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
) &&
1129 !before(shinfo
->tskey
, TCP_SKB_CB(skb2
)->seq
)) {
1130 struct skb_shared_info
*shinfo2
= skb_shinfo(skb2
);
1131 u8 tsflags
= shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
1133 shinfo
->tx_flags
&= ~tsflags
;
1134 shinfo2
->tx_flags
|= tsflags
;
1135 swap(shinfo
->tskey
, shinfo2
->tskey
);
1139 /* Function to create two new TCP segments. Shrinks the given segment
1140 * to the specified size and appends a new segment with the rest of the
1141 * packet to the list. This won't be called frequently, I hope.
1142 * Remember, these are still headerless SKBs at this point.
1144 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1145 unsigned int mss_now
, gfp_t gfp
)
1147 struct tcp_sock
*tp
= tcp_sk(sk
);
1148 struct sk_buff
*buff
;
1149 int nsize
, old_factor
;
1153 if (WARN_ON(len
> skb
->len
))
1156 nsize
= skb_headlen(skb
) - len
;
1160 if (skb_unclone(skb
, gfp
))
1163 /* Get a new skb... force flag on. */
1164 buff
= sk_stream_alloc_skb(sk
, nsize
, gfp
, true);
1166 return -ENOMEM
; /* We'll just try again later. */
1168 sk
->sk_wmem_queued
+= buff
->truesize
;
1169 sk_mem_charge(sk
, buff
->truesize
);
1170 nlen
= skb
->len
- len
- nsize
;
1171 buff
->truesize
+= nlen
;
1172 skb
->truesize
-= nlen
;
1174 /* Correct the sequence numbers. */
1175 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1176 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1177 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1179 /* PSH and FIN should only be set in the second packet. */
1180 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1181 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1182 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1183 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1185 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1186 /* Copy and checksum data tail into the new buffer. */
1187 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1188 skb_put(buff
, nsize
),
1193 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1195 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1196 skb_split(skb
, buff
, len
);
1199 buff
->ip_summed
= skb
->ip_summed
;
1201 buff
->tstamp
= skb
->tstamp
;
1202 tcp_fragment_tstamp(skb
, buff
);
1204 old_factor
= tcp_skb_pcount(skb
);
1206 /* Fix up tso_factor for both original and new SKB. */
1207 tcp_set_skb_tso_segs(skb
, mss_now
);
1208 tcp_set_skb_tso_segs(buff
, mss_now
);
1210 /* If this packet has been sent out already, we must
1211 * adjust the various packet counters.
1213 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1214 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1215 tcp_skb_pcount(buff
);
1218 tcp_adjust_pcount(sk
, skb
, diff
);
1221 /* Link BUFF into the send queue. */
1222 __skb_header_release(buff
);
1223 tcp_insert_write_queue_after(skb
, buff
, sk
);
1228 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1229 * eventually). The difference is that pulled data not copied, but
1230 * immediately discarded.
1232 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1234 struct skb_shared_info
*shinfo
;
1237 eat
= min_t(int, len
, skb_headlen(skb
));
1239 __skb_pull(skb
, eat
);
1246 shinfo
= skb_shinfo(skb
);
1247 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1248 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1251 skb_frag_unref(skb
, i
);
1254 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1256 shinfo
->frags
[k
].page_offset
+= eat
;
1257 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1263 shinfo
->nr_frags
= k
;
1265 skb_reset_tail_pointer(skb
);
1266 skb
->data_len
-= len
;
1267 skb
->len
= skb
->data_len
;
1270 /* Remove acked data from a packet in the transmit queue. */
1271 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1273 if (skb_unclone(skb
, GFP_ATOMIC
))
1276 __pskb_trim_head(skb
, len
);
1278 TCP_SKB_CB(skb
)->seq
+= len
;
1279 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1281 skb
->truesize
-= len
;
1282 sk
->sk_wmem_queued
-= len
;
1283 sk_mem_uncharge(sk
, len
);
1284 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1286 /* Any change of skb->len requires recalculation of tso factor. */
1287 if (tcp_skb_pcount(skb
) > 1)
1288 tcp_set_skb_tso_segs(skb
, tcp_skb_mss(skb
));
1293 /* Calculate MSS not accounting any TCP options. */
1294 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1296 const struct tcp_sock
*tp
= tcp_sk(sk
);
1297 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1300 /* Calculate base mss without TCP options:
1301 It is MMS_S - sizeof(tcphdr) of rfc1122
1303 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1305 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1306 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1307 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1309 if (dst
&& dst_allfrag(dst
))
1310 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1313 /* Clamp it (mss_clamp does not include tcp options) */
1314 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1315 mss_now
= tp
->rx_opt
.mss_clamp
;
1317 /* Now subtract optional transport overhead */
1318 mss_now
-= icsk
->icsk_ext_hdr_len
;
1320 /* Then reserve room for full set of TCP options and 8 bytes of data */
1326 /* Calculate MSS. Not accounting for SACKs here. */
1327 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1329 /* Subtract TCP options size, not including SACKs */
1330 return __tcp_mtu_to_mss(sk
, pmtu
) -
1331 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1334 /* Inverse of above */
1335 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1337 const struct tcp_sock
*tp
= tcp_sk(sk
);
1338 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1342 tp
->tcp_header_len
+
1343 icsk
->icsk_ext_hdr_len
+
1344 icsk
->icsk_af_ops
->net_header_len
;
1346 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1347 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1348 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1350 if (dst
&& dst_allfrag(dst
))
1351 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1356 /* MTU probing init per socket */
1357 void tcp_mtup_init(struct sock
*sk
)
1359 struct tcp_sock
*tp
= tcp_sk(sk
);
1360 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1361 struct net
*net
= sock_net(sk
);
1363 icsk
->icsk_mtup
.enabled
= net
->ipv4
.sysctl_tcp_mtu_probing
> 1;
1364 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1365 icsk
->icsk_af_ops
->net_header_len
;
1366 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, net
->ipv4
.sysctl_tcp_base_mss
);
1367 icsk
->icsk_mtup
.probe_size
= 0;
1368 if (icsk
->icsk_mtup
.enabled
)
1369 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1371 EXPORT_SYMBOL(tcp_mtup_init
);
1373 /* This function synchronize snd mss to current pmtu/exthdr set.
1375 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1376 for TCP options, but includes only bare TCP header.
1378 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1379 It is minimum of user_mss and mss received with SYN.
1380 It also does not include TCP options.
1382 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1384 tp->mss_cache is current effective sending mss, including
1385 all tcp options except for SACKs. It is evaluated,
1386 taking into account current pmtu, but never exceeds
1387 tp->rx_opt.mss_clamp.
1389 NOTE1. rfc1122 clearly states that advertised MSS
1390 DOES NOT include either tcp or ip options.
1392 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1393 are READ ONLY outside this function. --ANK (980731)
1395 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1397 struct tcp_sock
*tp
= tcp_sk(sk
);
1398 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1401 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1402 icsk
->icsk_mtup
.search_high
= pmtu
;
1404 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1405 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1407 /* And store cached results */
1408 icsk
->icsk_pmtu_cookie
= pmtu
;
1409 if (icsk
->icsk_mtup
.enabled
)
1410 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1411 tp
->mss_cache
= mss_now
;
1415 EXPORT_SYMBOL(tcp_sync_mss
);
1417 /* Compute the current effective MSS, taking SACKs and IP options,
1418 * and even PMTU discovery events into account.
1420 unsigned int tcp_current_mss(struct sock
*sk
)
1422 const struct tcp_sock
*tp
= tcp_sk(sk
);
1423 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1425 unsigned int header_len
;
1426 struct tcp_out_options opts
;
1427 struct tcp_md5sig_key
*md5
;
1429 mss_now
= tp
->mss_cache
;
1432 u32 mtu
= dst_mtu(dst
);
1433 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1434 mss_now
= tcp_sync_mss(sk
, mtu
);
1437 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1438 sizeof(struct tcphdr
);
1439 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1440 * some common options. If this is an odd packet (because we have SACK
1441 * blocks etc) then our calculated header_len will be different, and
1442 * we have to adjust mss_now correspondingly */
1443 if (header_len
!= tp
->tcp_header_len
) {
1444 int delta
= (int) header_len
- tp
->tcp_header_len
;
1451 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1452 * As additional protections, we do not touch cwnd in retransmission phases,
1453 * and if application hit its sndbuf limit recently.
1455 static void tcp_cwnd_application_limited(struct sock
*sk
)
1457 struct tcp_sock
*tp
= tcp_sk(sk
);
1459 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1460 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1461 /* Limited by application or receiver window. */
1462 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1463 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1464 if (win_used
< tp
->snd_cwnd
) {
1465 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1466 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
1468 tp
->snd_cwnd_used
= 0;
1470 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1473 static void tcp_cwnd_validate(struct sock
*sk
, bool is_cwnd_limited
)
1475 struct tcp_sock
*tp
= tcp_sk(sk
);
1477 /* Track the maximum number of outstanding packets in each
1478 * window, and remember whether we were cwnd-limited then.
1480 if (!before(tp
->snd_una
, tp
->max_packets_seq
) ||
1481 tp
->packets_out
> tp
->max_packets_out
) {
1482 tp
->max_packets_out
= tp
->packets_out
;
1483 tp
->max_packets_seq
= tp
->snd_nxt
;
1484 tp
->is_cwnd_limited
= is_cwnd_limited
;
1487 if (tcp_is_cwnd_limited(sk
)) {
1488 /* Network is feed fully. */
1489 tp
->snd_cwnd_used
= 0;
1490 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1492 /* Network starves. */
1493 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1494 tp
->snd_cwnd_used
= tp
->packets_out
;
1496 if (sysctl_tcp_slow_start_after_idle
&&
1497 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1498 tcp_cwnd_application_limited(sk
);
1502 /* Minshall's variant of the Nagle send check. */
1503 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1505 return after(tp
->snd_sml
, tp
->snd_una
) &&
1506 !after(tp
->snd_sml
, tp
->snd_nxt
);
1509 /* Update snd_sml if this skb is under mss
1510 * Note that a TSO packet might end with a sub-mss segment
1511 * The test is really :
1512 * if ((skb->len % mss) != 0)
1513 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1514 * But we can avoid doing the divide again given we already have
1515 * skb_pcount = skb->len / mss_now
1517 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1518 const struct sk_buff
*skb
)
1520 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1521 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1524 /* Return false, if packet can be sent now without violation Nagle's rules:
1525 * 1. It is full sized. (provided by caller in %partial bool)
1526 * 2. Or it contains FIN. (already checked by caller)
1527 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1528 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1529 * With Minshall's modification: all sent small packets are ACKed.
1531 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1535 ((nonagle
& TCP_NAGLE_CORK
) ||
1536 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1539 /* Return how many segs we'd like on a TSO packet,
1540 * to send one TSO packet per ms
1542 static u32
tcp_tso_autosize(const struct sock
*sk
, unsigned int mss_now
)
1546 bytes
= min(sk
->sk_pacing_rate
>> 10,
1547 sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
);
1549 /* Goal is to send at least one packet per ms,
1550 * not one big TSO packet every 100 ms.
1551 * This preserves ACK clocking and is consistent
1552 * with tcp_tso_should_defer() heuristic.
1554 segs
= max_t(u32
, bytes
/ mss_now
, sysctl_tcp_min_tso_segs
);
1556 return min_t(u32
, segs
, sk
->sk_gso_max_segs
);
1559 /* Returns the portion of skb which can be sent right away */
1560 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1561 const struct sk_buff
*skb
,
1562 unsigned int mss_now
,
1563 unsigned int max_segs
,
1566 const struct tcp_sock
*tp
= tcp_sk(sk
);
1567 u32 partial
, needed
, window
, max_len
;
1569 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1570 max_len
= mss_now
* max_segs
;
1572 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1575 needed
= min(skb
->len
, window
);
1577 if (max_len
<= needed
)
1580 partial
= needed
% mss_now
;
1581 /* If last segment is not a full MSS, check if Nagle rules allow us
1582 * to include this last segment in this skb.
1583 * Otherwise, we'll split the skb at last MSS boundary
1585 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
1586 return needed
- partial
;
1591 /* Can at least one segment of SKB be sent right now, according to the
1592 * congestion window rules? If so, return how many segments are allowed.
1594 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1595 const struct sk_buff
*skb
)
1597 u32 in_flight
, cwnd
, halfcwnd
;
1599 /* Don't be strict about the congestion window for the final FIN. */
1600 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1601 tcp_skb_pcount(skb
) == 1)
1604 in_flight
= tcp_packets_in_flight(tp
);
1605 cwnd
= tp
->snd_cwnd
;
1606 if (in_flight
>= cwnd
)
1609 /* For better scheduling, ensure we have at least
1610 * 2 GSO packets in flight.
1612 halfcwnd
= max(cwnd
>> 1, 1U);
1613 return min(halfcwnd
, cwnd
- in_flight
);
1616 /* Initialize TSO state of a skb.
1617 * This must be invoked the first time we consider transmitting
1618 * SKB onto the wire.
1620 static int tcp_init_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1622 int tso_segs
= tcp_skb_pcount(skb
);
1624 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1625 tcp_set_skb_tso_segs(skb
, mss_now
);
1626 tso_segs
= tcp_skb_pcount(skb
);
1632 /* Return true if the Nagle test allows this packet to be
1635 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1636 unsigned int cur_mss
, int nonagle
)
1638 /* Nagle rule does not apply to frames, which sit in the middle of the
1639 * write_queue (they have no chances to get new data).
1641 * This is implemented in the callers, where they modify the 'nonagle'
1642 * argument based upon the location of SKB in the send queue.
1644 if (nonagle
& TCP_NAGLE_PUSH
)
1647 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1648 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1651 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
1657 /* Does at least the first segment of SKB fit into the send window? */
1658 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1659 const struct sk_buff
*skb
,
1660 unsigned int cur_mss
)
1662 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1664 if (skb
->len
> cur_mss
)
1665 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1667 return !after(end_seq
, tcp_wnd_end(tp
));
1670 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1671 * should be put on the wire right now. If so, it returns the number of
1672 * packets allowed by the congestion window.
1674 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1675 unsigned int cur_mss
, int nonagle
)
1677 const struct tcp_sock
*tp
= tcp_sk(sk
);
1678 unsigned int cwnd_quota
;
1680 tcp_init_tso_segs(skb
, cur_mss
);
1682 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1685 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1686 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1692 /* Test if sending is allowed right now. */
1693 bool tcp_may_send_now(struct sock
*sk
)
1695 const struct tcp_sock
*tp
= tcp_sk(sk
);
1696 struct sk_buff
*skb
= tcp_send_head(sk
);
1699 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1700 (tcp_skb_is_last(sk
, skb
) ?
1701 tp
->nonagle
: TCP_NAGLE_PUSH
));
1704 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1705 * which is put after SKB on the list. It is very much like
1706 * tcp_fragment() except that it may make several kinds of assumptions
1707 * in order to speed up the splitting operation. In particular, we
1708 * know that all the data is in scatter-gather pages, and that the
1709 * packet has never been sent out before (and thus is not cloned).
1711 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1712 unsigned int mss_now
, gfp_t gfp
)
1714 struct sk_buff
*buff
;
1715 int nlen
= skb
->len
- len
;
1718 /* All of a TSO frame must be composed of paged data. */
1719 if (skb
->len
!= skb
->data_len
)
1720 return tcp_fragment(sk
, skb
, len
, mss_now
, gfp
);
1722 buff
= sk_stream_alloc_skb(sk
, 0, gfp
, true);
1723 if (unlikely(!buff
))
1726 sk
->sk_wmem_queued
+= buff
->truesize
;
1727 sk_mem_charge(sk
, buff
->truesize
);
1728 buff
->truesize
+= nlen
;
1729 skb
->truesize
-= nlen
;
1731 /* Correct the sequence numbers. */
1732 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1733 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1734 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1736 /* PSH and FIN should only be set in the second packet. */
1737 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1738 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1739 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1741 /* This packet was never sent out yet, so no SACK bits. */
1742 TCP_SKB_CB(buff
)->sacked
= 0;
1744 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1745 skb_split(skb
, buff
, len
);
1746 tcp_fragment_tstamp(skb
, buff
);
1748 /* Fix up tso_factor for both original and new SKB. */
1749 tcp_set_skb_tso_segs(skb
, mss_now
);
1750 tcp_set_skb_tso_segs(buff
, mss_now
);
1752 /* Link BUFF into the send queue. */
1753 __skb_header_release(buff
);
1754 tcp_insert_write_queue_after(skb
, buff
, sk
);
1759 /* Try to defer sending, if possible, in order to minimize the amount
1760 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1762 * This algorithm is from John Heffner.
1764 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
,
1765 bool *is_cwnd_limited
, u32 max_segs
)
1767 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1768 u32 age
, send_win
, cong_win
, limit
, in_flight
;
1769 struct tcp_sock
*tp
= tcp_sk(sk
);
1770 struct skb_mstamp now
;
1771 struct sk_buff
*head
;
1774 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1777 if (!((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Open
| TCPF_CA_CWR
)))
1780 /* Avoid bursty behavior by allowing defer
1781 * only if the last write was recent.
1783 if ((s32
)(tcp_time_stamp
- tp
->lsndtime
) > 0)
1786 in_flight
= tcp_packets_in_flight(tp
);
1788 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1790 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1792 /* From in_flight test above, we know that cwnd > in_flight. */
1793 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1795 limit
= min(send_win
, cong_win
);
1797 /* If a full-sized TSO skb can be sent, do it. */
1798 if (limit
>= max_segs
* tp
->mss_cache
)
1801 /* Middle in queue won't get any more data, full sendable already? */
1802 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1805 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1807 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1809 /* If at least some fraction of a window is available,
1812 chunk
/= win_divisor
;
1816 /* Different approach, try not to defer past a single
1817 * ACK. Receiver should ACK every other full sized
1818 * frame, so if we have space for more than 3 frames
1821 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1825 head
= tcp_write_queue_head(sk
);
1826 skb_mstamp_get(&now
);
1827 age
= skb_mstamp_us_delta(&now
, &head
->skb_mstamp
);
1828 /* If next ACK is likely to come too late (half srtt), do not defer */
1829 if (age
< (tp
->srtt_us
>> 4))
1832 /* Ok, it looks like it is advisable to defer. */
1834 if (cong_win
< send_win
&& cong_win
< skb
->len
)
1835 *is_cwnd_limited
= true;
1843 static inline void tcp_mtu_check_reprobe(struct sock
*sk
)
1845 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1846 struct tcp_sock
*tp
= tcp_sk(sk
);
1847 struct net
*net
= sock_net(sk
);
1851 interval
= net
->ipv4
.sysctl_tcp_probe_interval
;
1852 delta
= tcp_time_stamp
- icsk
->icsk_mtup
.probe_timestamp
;
1853 if (unlikely(delta
>= interval
* HZ
)) {
1854 int mss
= tcp_current_mss(sk
);
1856 /* Update current search range */
1857 icsk
->icsk_mtup
.probe_size
= 0;
1858 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+
1859 sizeof(struct tcphdr
) +
1860 icsk
->icsk_af_ops
->net_header_len
;
1861 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
1863 /* Update probe time stamp */
1864 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1868 /* Create a new MTU probe if we are ready.
1869 * MTU probe is regularly attempting to increase the path MTU by
1870 * deliberately sending larger packets. This discovers routing
1871 * changes resulting in larger path MTUs.
1873 * Returns 0 if we should wait to probe (no cwnd available),
1874 * 1 if a probe was sent,
1877 static int tcp_mtu_probe(struct sock
*sk
)
1879 struct tcp_sock
*tp
= tcp_sk(sk
);
1880 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1881 struct sk_buff
*skb
, *nskb
, *next
;
1882 struct net
*net
= sock_net(sk
);
1890 /* Not currently probing/verifying,
1892 * have enough cwnd, and
1893 * not SACKing (the variable headers throw things off) */
1894 if (!icsk
->icsk_mtup
.enabled
||
1895 icsk
->icsk_mtup
.probe_size
||
1896 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1897 tp
->snd_cwnd
< 11 ||
1898 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1901 /* Use binary search for probe_size between tcp_mss_base,
1902 * and current mss_clamp. if (search_high - search_low)
1903 * smaller than a threshold, backoff from probing.
1905 mss_now
= tcp_current_mss(sk
);
1906 probe_size
= tcp_mtu_to_mss(sk
, (icsk
->icsk_mtup
.search_high
+
1907 icsk
->icsk_mtup
.search_low
) >> 1);
1908 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1909 interval
= icsk
->icsk_mtup
.search_high
- icsk
->icsk_mtup
.search_low
;
1910 /* When misfortune happens, we are reprobing actively,
1911 * and then reprobe timer has expired. We stick with current
1912 * probing process by not resetting search range to its orignal.
1914 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
) ||
1915 interval
< net
->ipv4
.sysctl_tcp_probe_threshold
) {
1916 /* Check whether enough time has elaplased for
1917 * another round of probing.
1919 tcp_mtu_check_reprobe(sk
);
1923 /* Have enough data in the send queue to probe? */
1924 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1927 if (tp
->snd_wnd
< size_needed
)
1929 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1932 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1933 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1934 if (!tcp_packets_in_flight(tp
))
1940 /* We're allowed to probe. Build it now. */
1941 nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
, false);
1944 sk
->sk_wmem_queued
+= nskb
->truesize
;
1945 sk_mem_charge(sk
, nskb
->truesize
);
1947 skb
= tcp_send_head(sk
);
1949 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1950 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1951 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
1952 TCP_SKB_CB(nskb
)->sacked
= 0;
1954 nskb
->ip_summed
= skb
->ip_summed
;
1956 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1959 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1960 copy
= min_t(int, skb
->len
, probe_size
- len
);
1961 if (nskb
->ip_summed
)
1962 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1964 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1965 skb_put(nskb
, copy
),
1968 if (skb
->len
<= copy
) {
1969 /* We've eaten all the data from this skb.
1971 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1972 tcp_unlink_write_queue(skb
, sk
);
1973 sk_wmem_free_skb(sk
, skb
);
1975 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
1976 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1977 if (!skb_shinfo(skb
)->nr_frags
) {
1978 skb_pull(skb
, copy
);
1979 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1980 skb
->csum
= csum_partial(skb
->data
,
1983 __pskb_trim_head(skb
, copy
);
1984 tcp_set_skb_tso_segs(skb
, mss_now
);
1986 TCP_SKB_CB(skb
)->seq
+= copy
;
1991 if (len
>= probe_size
)
1994 tcp_init_tso_segs(nskb
, nskb
->len
);
1996 /* We're ready to send. If this fails, the probe will
1997 * be resegmented into mss-sized pieces by tcp_write_xmit().
1999 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
2000 /* Decrement cwnd here because we are sending
2001 * effectively two packets. */
2003 tcp_event_new_data_sent(sk
, nskb
);
2005 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
2006 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
2007 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
2015 /* This routine writes packets to the network. It advances the
2016 * send_head. This happens as incoming acks open up the remote
2019 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2020 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2021 * account rare use of URG, this is not a big flaw.
2023 * Send at most one packet when push_one > 0. Temporarily ignore
2024 * cwnd limit to force at most one packet out when push_one == 2.
2026 * Returns true, if no segments are in flight and we have queued segments,
2027 * but cannot send anything now because of SWS or another problem.
2029 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
2030 int push_one
, gfp_t gfp
)
2032 struct tcp_sock
*tp
= tcp_sk(sk
);
2033 struct sk_buff
*skb
;
2034 unsigned int tso_segs
, sent_pkts
;
2037 bool is_cwnd_limited
= false;
2043 /* Do MTU probing. */
2044 result
= tcp_mtu_probe(sk
);
2047 } else if (result
> 0) {
2052 max_segs
= tcp_tso_autosize(sk
, mss_now
);
2053 while ((skb
= tcp_send_head(sk
))) {
2056 tso_segs
= tcp_init_tso_segs(skb
, mss_now
);
2059 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
) {
2060 /* "skb_mstamp" is used as a start point for the retransmit timer */
2061 skb_mstamp_get(&skb
->skb_mstamp
);
2062 goto repair
; /* Skip network transmission */
2065 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2067 is_cwnd_limited
= true;
2069 /* Force out a loss probe pkt. */
2075 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
2078 if (tso_segs
== 1) {
2079 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2080 (tcp_skb_is_last(sk
, skb
) ?
2081 nonagle
: TCP_NAGLE_PUSH
))))
2085 tcp_tso_should_defer(sk
, skb
, &is_cwnd_limited
,
2091 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
2092 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2098 if (skb
->len
> limit
&&
2099 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
2102 /* TCP Small Queues :
2103 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2105 * - better RTT estimation and ACK scheduling
2108 * Alas, some drivers / subsystems require a fair amount
2109 * of queued bytes to ensure line rate.
2110 * One example is wifi aggregation (802.11 AMPDU)
2112 limit
= max(2 * skb
->truesize
, sk
->sk_pacing_rate
>> 10);
2113 limit
= min_t(u32
, limit
, sysctl_tcp_limit_output_bytes
);
2115 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
) {
2116 set_bit(TSQ_THROTTLED
, &tp
->tsq_flags
);
2117 /* It is possible TX completion already happened
2118 * before we set TSQ_THROTTLED, so we must
2119 * test again the condition.
2121 smp_mb__after_atomic();
2122 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
)
2126 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2130 /* Advance the send_head. This one is sent out.
2131 * This call will increment packets_out.
2133 tcp_event_new_data_sent(sk
, skb
);
2135 tcp_minshall_update(tp
, mss_now
, skb
);
2136 sent_pkts
+= tcp_skb_pcount(skb
);
2142 if (likely(sent_pkts
)) {
2143 if (tcp_in_cwnd_reduction(sk
))
2144 tp
->prr_out
+= sent_pkts
;
2146 /* Send one loss probe per tail loss episode. */
2148 tcp_schedule_loss_probe(sk
);
2149 tcp_cwnd_validate(sk
, is_cwnd_limited
);
2152 return (push_one
== 2) || (!tp
->packets_out
&& tcp_send_head(sk
));
2155 bool tcp_schedule_loss_probe(struct sock
*sk
)
2157 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2158 struct tcp_sock
*tp
= tcp_sk(sk
);
2159 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
2160 u32 rtt
= usecs_to_jiffies(tp
->srtt_us
>> 3);
2162 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
))
2164 /* No consecutive loss probes. */
2165 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
2169 /* Don't do any loss probe on a Fast Open connection before 3WHS
2172 if (sk
->sk_state
== TCP_SYN_RECV
)
2175 /* TLP is only scheduled when next timer event is RTO. */
2176 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
2179 /* Schedule a loss probe in 2*RTT for SACK capable connections
2180 * in Open state, that are either limited by cwnd or application.
2182 if (sysctl_tcp_early_retrans
< 3 || !tp
->srtt_us
|| !tp
->packets_out
||
2183 !tcp_is_sack(tp
) || inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
2186 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
2190 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2191 * for delayed ack when there's one outstanding packet.
2194 if (tp
->packets_out
== 1)
2195 timeout
= max_t(u32
, timeout
,
2196 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
2197 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
2199 /* If RTO is shorter, just schedule TLP in its place. */
2200 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
2201 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
2202 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
2203 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
2208 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2213 /* Thanks to skb fast clones, we can detect if a prior transmit of
2214 * a packet is still in a qdisc or driver queue.
2215 * In this case, there is very little point doing a retransmit !
2216 * Note: This is called from BH context only.
2218 static bool skb_still_in_host_queue(const struct sock
*sk
,
2219 const struct sk_buff
*skb
)
2221 if (unlikely(skb_fclone_busy(sk
, skb
))) {
2222 NET_INC_STATS_BH(sock_net(sk
),
2223 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
2229 /* When probe timeout (PTO) fires, send a new segment if one exists, else
2230 * retransmit the last segment.
2232 void tcp_send_loss_probe(struct sock
*sk
)
2234 struct tcp_sock
*tp
= tcp_sk(sk
);
2235 struct sk_buff
*skb
;
2237 int mss
= tcp_current_mss(sk
);
2240 if (tcp_send_head(sk
)) {
2241 err
= tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2245 /* At most one outstanding TLP retransmission. */
2246 if (tp
->tlp_high_seq
)
2249 /* Retransmit last segment. */
2250 skb
= tcp_write_queue_tail(sk
);
2254 if (skb_still_in_host_queue(sk
, skb
))
2257 pcount
= tcp_skb_pcount(skb
);
2258 if (WARN_ON(!pcount
))
2261 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2262 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
,
2265 skb
= tcp_write_queue_tail(sk
);
2268 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2271 err
= __tcp_retransmit_skb(sk
, skb
);
2273 /* Record snd_nxt for loss detection. */
2275 tp
->tlp_high_seq
= tp
->snd_nxt
;
2278 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2279 inet_csk(sk
)->icsk_rto
,
2283 NET_INC_STATS_BH(sock_net(sk
),
2284 LINUX_MIB_TCPLOSSPROBES
);
2287 /* Push out any pending frames which were held back due to
2288 * TCP_CORK or attempt at coalescing tiny packets.
2289 * The socket must be locked by the caller.
2291 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2294 /* If we are closed, the bytes will have to remain here.
2295 * In time closedown will finish, we empty the write queue and
2296 * all will be happy.
2298 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2301 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2302 sk_gfp_atomic(sk
, GFP_ATOMIC
)))
2303 tcp_check_probe_timer(sk
);
2306 /* Send _single_ skb sitting at the send head. This function requires
2307 * true push pending frames to setup probe timer etc.
2309 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2311 struct sk_buff
*skb
= tcp_send_head(sk
);
2313 BUG_ON(!skb
|| skb
->len
< mss_now
);
2315 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2318 /* This function returns the amount that we can raise the
2319 * usable window based on the following constraints
2321 * 1. The window can never be shrunk once it is offered (RFC 793)
2322 * 2. We limit memory per socket
2325 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2326 * RECV.NEXT + RCV.WIN fixed until:
2327 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2329 * i.e. don't raise the right edge of the window until you can raise
2330 * it at least MSS bytes.
2332 * Unfortunately, the recommended algorithm breaks header prediction,
2333 * since header prediction assumes th->window stays fixed.
2335 * Strictly speaking, keeping th->window fixed violates the receiver
2336 * side SWS prevention criteria. The problem is that under this rule
2337 * a stream of single byte packets will cause the right side of the
2338 * window to always advance by a single byte.
2340 * Of course, if the sender implements sender side SWS prevention
2341 * then this will not be a problem.
2343 * BSD seems to make the following compromise:
2345 * If the free space is less than the 1/4 of the maximum
2346 * space available and the free space is less than 1/2 mss,
2347 * then set the window to 0.
2348 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2349 * Otherwise, just prevent the window from shrinking
2350 * and from being larger than the largest representable value.
2352 * This prevents incremental opening of the window in the regime
2353 * where TCP is limited by the speed of the reader side taking
2354 * data out of the TCP receive queue. It does nothing about
2355 * those cases where the window is constrained on the sender side
2356 * because the pipeline is full.
2358 * BSD also seems to "accidentally" limit itself to windows that are a
2359 * multiple of MSS, at least until the free space gets quite small.
2360 * This would appear to be a side effect of the mbuf implementation.
2361 * Combining these two algorithms results in the observed behavior
2362 * of having a fixed window size at almost all times.
2364 * Below we obtain similar behavior by forcing the offered window to
2365 * a multiple of the mss when it is feasible to do so.
2367 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2368 * Regular options like TIMESTAMP are taken into account.
2370 u32
__tcp_select_window(struct sock
*sk
)
2372 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2373 struct tcp_sock
*tp
= tcp_sk(sk
);
2374 /* MSS for the peer's data. Previous versions used mss_clamp
2375 * here. I don't know if the value based on our guesses
2376 * of peer's MSS is better for the performance. It's more correct
2377 * but may be worse for the performance because of rcv_mss
2378 * fluctuations. --SAW 1998/11/1
2380 int mss
= icsk
->icsk_ack
.rcv_mss
;
2381 int free_space
= tcp_space(sk
);
2382 int allowed_space
= tcp_full_space(sk
);
2383 int full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2386 if (mss
> full_space
)
2389 if (free_space
< (full_space
>> 1)) {
2390 icsk
->icsk_ack
.quick
= 0;
2392 if (tcp_under_memory_pressure(sk
))
2393 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2396 /* free_space might become our new window, make sure we don't
2397 * increase it due to wscale.
2399 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2401 /* if free space is less than mss estimate, or is below 1/16th
2402 * of the maximum allowed, try to move to zero-window, else
2403 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2404 * new incoming data is dropped due to memory limits.
2405 * With large window, mss test triggers way too late in order
2406 * to announce zero window in time before rmem limit kicks in.
2408 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2412 if (free_space
> tp
->rcv_ssthresh
)
2413 free_space
= tp
->rcv_ssthresh
;
2415 /* Don't do rounding if we are using window scaling, since the
2416 * scaled window will not line up with the MSS boundary anyway.
2418 window
= tp
->rcv_wnd
;
2419 if (tp
->rx_opt
.rcv_wscale
) {
2420 window
= free_space
;
2422 /* Advertise enough space so that it won't get scaled away.
2423 * Import case: prevent zero window announcement if
2424 * 1<<rcv_wscale > mss.
2426 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2427 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2428 << tp
->rx_opt
.rcv_wscale
);
2430 /* Get the largest window that is a nice multiple of mss.
2431 * Window clamp already applied above.
2432 * If our current window offering is within 1 mss of the
2433 * free space we just keep it. This prevents the divide
2434 * and multiply from happening most of the time.
2435 * We also don't do any window rounding when the free space
2438 if (window
<= free_space
- mss
|| window
> free_space
)
2439 window
= (free_space
/ mss
) * mss
;
2440 else if (mss
== full_space
&&
2441 free_space
> window
+ (full_space
>> 1))
2442 window
= free_space
;
2448 /* Collapses two adjacent SKB's during retransmission. */
2449 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2451 struct tcp_sock
*tp
= tcp_sk(sk
);
2452 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2453 int skb_size
, next_skb_size
;
2455 skb_size
= skb
->len
;
2456 next_skb_size
= next_skb
->len
;
2458 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2460 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2462 tcp_unlink_write_queue(next_skb
, sk
);
2464 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
2467 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2468 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2470 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2471 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2473 /* Update sequence range on original skb. */
2474 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2476 /* Merge over control information. This moves PSH/FIN etc. over */
2477 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2479 /* All done, get rid of second SKB and account for it so
2480 * packet counting does not break.
2482 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2484 /* changed transmit queue under us so clear hints */
2485 tcp_clear_retrans_hints_partial(tp
);
2486 if (next_skb
== tp
->retransmit_skb_hint
)
2487 tp
->retransmit_skb_hint
= skb
;
2489 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2491 sk_wmem_free_skb(sk
, next_skb
);
2494 /* Check if coalescing SKBs is legal. */
2495 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2497 if (tcp_skb_pcount(skb
) > 1)
2499 /* TODO: SACK collapsing could be used to remove this condition */
2500 if (skb_shinfo(skb
)->nr_frags
!= 0)
2502 if (skb_cloned(skb
))
2504 if (skb
== tcp_send_head(sk
))
2506 /* Some heurestics for collapsing over SACK'd could be invented */
2507 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2513 /* Collapse packets in the retransmit queue to make to create
2514 * less packets on the wire. This is only done on retransmission.
2516 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2519 struct tcp_sock
*tp
= tcp_sk(sk
);
2520 struct sk_buff
*skb
= to
, *tmp
;
2523 if (!sysctl_tcp_retrans_collapse
)
2525 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2528 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2529 if (!tcp_can_collapse(sk
, skb
))
2541 /* Punt if not enough space exists in the first SKB for
2542 * the data in the second
2544 if (skb
->len
> skb_availroom(to
))
2547 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2550 tcp_collapse_retrans(sk
, to
);
2554 /* This retransmits one SKB. Policy decisions and retransmit queue
2555 * state updates are done by the caller. Returns non-zero if an
2556 * error occurred which prevented the send.
2558 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2560 struct tcp_sock
*tp
= tcp_sk(sk
);
2561 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2562 unsigned int cur_mss
;
2565 /* Inconslusive MTU probe */
2566 if (icsk
->icsk_mtup
.probe_size
) {
2567 icsk
->icsk_mtup
.probe_size
= 0;
2570 /* Do not sent more than we queued. 1/4 is reserved for possible
2571 * copying overhead: fragmentation, tunneling, mangling etc.
2573 if (atomic_read(&sk
->sk_wmem_alloc
) >
2574 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2577 if (skb_still_in_host_queue(sk
, skb
))
2580 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2581 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2583 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2587 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2588 return -EHOSTUNREACH
; /* Routing failure or similar. */
2590 cur_mss
= tcp_current_mss(sk
);
2592 /* If receiver has shrunk his window, and skb is out of
2593 * new window, do not retransmit it. The exception is the
2594 * case, when window is shrunk to zero. In this case
2595 * our retransmit serves as a zero window probe.
2597 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2598 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2601 if (skb
->len
> cur_mss
) {
2602 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
, GFP_ATOMIC
))
2603 return -ENOMEM
; /* We'll try again later. */
2605 int oldpcount
= tcp_skb_pcount(skb
);
2607 if (unlikely(oldpcount
> 1)) {
2608 if (skb_unclone(skb
, GFP_ATOMIC
))
2610 tcp_init_tso_segs(skb
, cur_mss
);
2611 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2615 /* RFC3168, section 6.1.1.1. ECN fallback */
2616 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN_ECN
) == TCPHDR_SYN_ECN
)
2617 tcp_ecn_clear_syn(sk
, skb
);
2619 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2621 /* Make a copy, if the first transmission SKB clone we made
2622 * is still in somebody's hands, else make a clone.
2625 /* make sure skb->data is aligned on arches that require it
2626 * and check if ack-trimming & collapsing extended the headroom
2627 * beyond what csum_start can cover.
2629 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2630 skb_headroom(skb
) >= 0xFFFF)) {
2631 struct sk_buff
*nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
,
2633 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2636 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2640 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
2641 /* Update global TCP statistics. */
2642 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2643 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2644 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
2645 tp
->total_retrans
++;
2650 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2652 struct tcp_sock
*tp
= tcp_sk(sk
);
2653 int err
= __tcp_retransmit_skb(sk
, skb
);
2656 #if FASTRETRANS_DEBUG > 0
2657 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2658 net_dbg_ratelimited("retrans_out leaked\n");
2661 if (!tp
->retrans_out
)
2662 tp
->lost_retrans_low
= tp
->snd_nxt
;
2663 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2664 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2666 /* Save stamp of the first retransmit. */
2667 if (!tp
->retrans_stamp
)
2668 tp
->retrans_stamp
= tcp_skb_timestamp(skb
);
2670 /* snd_nxt is stored to detect loss of retransmitted segment,
2671 * see tcp_input.c tcp_sacktag_write_queue().
2673 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2674 } else if (err
!= -EBUSY
) {
2675 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2678 if (tp
->undo_retrans
< 0)
2679 tp
->undo_retrans
= 0;
2680 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2684 /* Check if we forward retransmits are possible in the current
2685 * window/congestion state.
2687 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2689 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2690 const struct tcp_sock
*tp
= tcp_sk(sk
);
2692 /* Forward retransmissions are possible only during Recovery. */
2693 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2696 /* No forward retransmissions in Reno are possible. */
2697 if (tcp_is_reno(tp
))
2700 /* Yeah, we have to make difficult choice between forward transmission
2701 * and retransmission... Both ways have their merits...
2703 * For now we do not retransmit anything, while we have some new
2704 * segments to send. In the other cases, follow rule 3 for
2705 * NextSeg() specified in RFC3517.
2708 if (tcp_may_send_now(sk
))
2714 /* This gets called after a retransmit timeout, and the initially
2715 * retransmitted data is acknowledged. It tries to continue
2716 * resending the rest of the retransmit queue, until either
2717 * we've sent it all or the congestion window limit is reached.
2718 * If doing SACK, the first ACK which comes back for a timeout
2719 * based retransmit packet might feed us FACK information again.
2720 * If so, we use it to avoid unnecessarily retransmissions.
2722 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2724 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2725 struct tcp_sock
*tp
= tcp_sk(sk
);
2726 struct sk_buff
*skb
;
2727 struct sk_buff
*hole
= NULL
;
2730 int fwd_rexmitting
= 0;
2732 if (!tp
->packets_out
)
2736 tp
->retransmit_high
= tp
->snd_una
;
2738 if (tp
->retransmit_skb_hint
) {
2739 skb
= tp
->retransmit_skb_hint
;
2740 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2741 if (after(last_lost
, tp
->retransmit_high
))
2742 last_lost
= tp
->retransmit_high
;
2744 skb
= tcp_write_queue_head(sk
);
2745 last_lost
= tp
->snd_una
;
2748 tcp_for_write_queue_from(skb
, sk
) {
2749 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2751 if (skb
== tcp_send_head(sk
))
2753 /* we could do better than to assign each time */
2755 tp
->retransmit_skb_hint
= skb
;
2757 /* Assume this retransmit will generate
2758 * only one packet for congestion window
2759 * calculation purposes. This works because
2760 * tcp_retransmit_skb() will chop up the
2761 * packet to be MSS sized and all the
2762 * packet counting works out.
2764 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2767 if (fwd_rexmitting
) {
2769 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2771 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2773 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2774 tp
->retransmit_high
= last_lost
;
2775 if (!tcp_can_forward_retransmit(sk
))
2777 /* Backtrack if necessary to non-L'ed skb */
2785 } else if (!(sacked
& TCPCB_LOST
)) {
2786 if (!hole
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2791 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2792 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2793 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2795 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2798 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2801 if (tcp_retransmit_skb(sk
, skb
))
2804 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2806 if (tcp_in_cwnd_reduction(sk
))
2807 tp
->prr_out
+= tcp_skb_pcount(skb
);
2809 if (skb
== tcp_write_queue_head(sk
))
2810 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2811 inet_csk(sk
)->icsk_rto
,
2816 /* We allow to exceed memory limits for FIN packets to expedite
2817 * connection tear down and (memory) recovery.
2818 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2819 * or even be forced to close flow without any FIN.
2820 * In general, we want to allow one skb per socket to avoid hangs
2821 * with edge trigger epoll()
2823 void sk_forced_mem_schedule(struct sock
*sk
, int size
)
2827 if (size
<= sk
->sk_forward_alloc
)
2829 amt
= sk_mem_pages(size
);
2830 sk
->sk_forward_alloc
+= amt
* SK_MEM_QUANTUM
;
2831 sk_memory_allocated_add(sk
, amt
, &status
);
2834 /* Send a FIN. The caller locks the socket for us.
2835 * We should try to send a FIN packet really hard, but eventually give up.
2837 void tcp_send_fin(struct sock
*sk
)
2839 struct sk_buff
*skb
, *tskb
= tcp_write_queue_tail(sk
);
2840 struct tcp_sock
*tp
= tcp_sk(sk
);
2842 /* Optimization, tack on the FIN if we have one skb in write queue and
2843 * this skb was not yet sent, or we are under memory pressure.
2844 * Note: in the latter case, FIN packet will be sent after a timeout,
2845 * as TCP stack thinks it has already been transmitted.
2847 if (tskb
&& (tcp_send_head(sk
) || tcp_under_memory_pressure(sk
))) {
2849 TCP_SKB_CB(tskb
)->tcp_flags
|= TCPHDR_FIN
;
2850 TCP_SKB_CB(tskb
)->end_seq
++;
2852 if (!tcp_send_head(sk
)) {
2853 /* This means tskb was already sent.
2854 * Pretend we included the FIN on previous transmit.
2855 * We need to set tp->snd_nxt to the value it would have
2856 * if FIN had been sent. This is because retransmit path
2857 * does not change tp->snd_nxt.
2863 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, sk
->sk_allocation
);
2864 if (unlikely(!skb
)) {
2869 skb_reserve(skb
, MAX_TCP_HEADER
);
2870 sk_forced_mem_schedule(sk
, skb
->truesize
);
2871 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2872 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2873 TCPHDR_ACK
| TCPHDR_FIN
);
2874 tcp_queue_skb(sk
, skb
);
2876 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
), TCP_NAGLE_OFF
);
2879 /* We get here when a process closes a file descriptor (either due to
2880 * an explicit close() or as a byproduct of exit()'ing) and there
2881 * was unread data in the receive queue. This behavior is recommended
2882 * by RFC 2525, section 2.17. -DaveM
2884 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2886 struct sk_buff
*skb
;
2888 /* NOTE: No TCP options attached and we never retransmit this. */
2889 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2891 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2895 /* Reserve space for headers and prepare control bits. */
2896 skb_reserve(skb
, MAX_TCP_HEADER
);
2897 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2898 TCPHDR_ACK
| TCPHDR_RST
);
2900 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2901 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2903 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2906 /* Send a crossed SYN-ACK during socket establishment.
2907 * WARNING: This routine must only be called when we have already sent
2908 * a SYN packet that crossed the incoming SYN that caused this routine
2909 * to get called. If this assumption fails then the initial rcv_wnd
2910 * and rcv_wscale values will not be correct.
2912 int tcp_send_synack(struct sock
*sk
)
2914 struct sk_buff
*skb
;
2916 skb
= tcp_write_queue_head(sk
);
2917 if (!skb
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2918 pr_debug("%s: wrong queue state\n", __func__
);
2921 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
2922 if (skb_cloned(skb
)) {
2923 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2926 tcp_unlink_write_queue(skb
, sk
);
2927 __skb_header_release(nskb
);
2928 __tcp_add_write_queue_head(sk
, nskb
);
2929 sk_wmem_free_skb(sk
, skb
);
2930 sk
->sk_wmem_queued
+= nskb
->truesize
;
2931 sk_mem_charge(sk
, nskb
->truesize
);
2935 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
2936 tcp_ecn_send_synack(sk
, skb
);
2938 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2942 * tcp_make_synack - Prepare a SYN-ACK.
2943 * sk: listener socket
2944 * dst: dst entry attached to the SYNACK
2945 * req: request_sock pointer
2947 * Allocate one skb and build a SYNACK packet.
2948 * @dst is consumed : Caller should not use it again.
2950 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2951 struct request_sock
*req
,
2952 struct tcp_fastopen_cookie
*foc
)
2954 struct tcp_out_options opts
;
2955 struct inet_request_sock
*ireq
= inet_rsk(req
);
2956 struct tcp_sock
*tp
= tcp_sk(sk
);
2958 struct sk_buff
*skb
;
2959 struct tcp_md5sig_key
*md5
= NULL
;
2960 int tcp_header_size
;
2963 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
, 1, GFP_ATOMIC
);
2964 if (unlikely(!skb
)) {
2968 /* Reserve space for headers. */
2969 skb_reserve(skb
, MAX_TCP_HEADER
);
2971 skb_dst_set(skb
, dst
);
2973 mss
= dst_metric_advmss(dst
);
2974 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2975 mss
= tp
->rx_opt
.user_mss
;
2977 memset(&opts
, 0, sizeof(opts
));
2978 #ifdef CONFIG_SYN_COOKIES
2979 if (unlikely(req
->cookie_ts
))
2980 skb
->skb_mstamp
.stamp_jiffies
= cookie_init_timestamp(req
);
2983 skb_mstamp_get(&skb
->skb_mstamp
);
2985 #ifdef CONFIG_TCP_MD5SIG
2987 md5
= tcp_rsk(req
)->af_specific
->req_md5_lookup(sk
, req_to_sk(req
));
2989 tcp_header_size
= tcp_synack_options(sk
, req
, mss
, skb
, &opts
, md5
,
2992 skb_push(skb
, tcp_header_size
);
2993 skb_reset_transport_header(skb
);
2996 memset(th
, 0, sizeof(struct tcphdr
));
2999 tcp_ecn_make_synack(req
, th
, sk
);
3000 th
->source
= htons(ireq
->ir_num
);
3001 th
->dest
= ireq
->ir_rmt_port
;
3002 /* Setting of flags are superfluous here for callers (and ECE is
3003 * not even correctly set)
3005 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
3006 TCPHDR_SYN
| TCPHDR_ACK
);
3008 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
3009 /* XXX data is queued and acked as is. No buffer/window check */
3010 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
3012 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3013 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
3014 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
3015 th
->doff
= (tcp_header_size
>> 2);
3016 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_OUTSEGS
);
3018 #ifdef CONFIG_TCP_MD5SIG
3019 /* Okay, we have all we need - do the md5 hash if needed */
3021 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
3022 md5
, req_to_sk(req
), skb
);
3026 /* Do not fool tcpdump (if any), clean our debris */
3027 skb
->tstamp
.tv64
= 0;
3030 EXPORT_SYMBOL(tcp_make_synack
);
3032 static void tcp_ca_dst_init(struct sock
*sk
, const struct dst_entry
*dst
)
3034 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3035 const struct tcp_congestion_ops
*ca
;
3036 u32 ca_key
= dst_metric(dst
, RTAX_CC_ALGO
);
3038 if (ca_key
== TCP_CA_UNSPEC
)
3042 ca
= tcp_ca_find_key(ca_key
);
3043 if (likely(ca
&& try_module_get(ca
->owner
))) {
3044 module_put(icsk
->icsk_ca_ops
->owner
);
3045 icsk
->icsk_ca_dst_locked
= tcp_ca_dst_locked(dst
);
3046 icsk
->icsk_ca_ops
= ca
;
3051 /* Do all connect socket setups that can be done AF independent. */
3052 static void tcp_connect_init(struct sock
*sk
)
3054 const struct dst_entry
*dst
= __sk_dst_get(sk
);
3055 struct tcp_sock
*tp
= tcp_sk(sk
);
3058 /* We'll fix this up when we get a response from the other end.
3059 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3061 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
3062 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
3064 #ifdef CONFIG_TCP_MD5SIG
3065 if (tp
->af_specific
->md5_lookup(sk
, sk
))
3066 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
3069 /* If user gave his TCP_MAXSEG, record it to clamp */
3070 if (tp
->rx_opt
.user_mss
)
3071 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3074 tcp_sync_mss(sk
, dst_mtu(dst
));
3076 tcp_ca_dst_init(sk
, dst
);
3078 if (!tp
->window_clamp
)
3079 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
3080 tp
->advmss
= dst_metric_advmss(dst
);
3081 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
3082 tp
->advmss
= tp
->rx_opt
.user_mss
;
3084 tcp_initialize_rcv_mss(sk
);
3086 /* limit the window selection if the user enforce a smaller rx buffer */
3087 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
3088 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
3089 tp
->window_clamp
= tcp_full_space(sk
);
3091 tcp_select_initial_window(tcp_full_space(sk
),
3092 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
3095 sysctl_tcp_window_scaling
,
3097 dst_metric(dst
, RTAX_INITRWND
));
3099 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
3100 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
3103 sock_reset_flag(sk
, SOCK_DONE
);
3106 tp
->snd_una
= tp
->write_seq
;
3107 tp
->snd_sml
= tp
->write_seq
;
3108 tp
->snd_up
= tp
->write_seq
;
3109 tp
->snd_nxt
= tp
->write_seq
;
3111 if (likely(!tp
->repair
))
3114 tp
->rcv_tstamp
= tcp_time_stamp
;
3115 tp
->rcv_wup
= tp
->rcv_nxt
;
3116 tp
->copied_seq
= tp
->rcv_nxt
;
3118 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
3119 inet_csk(sk
)->icsk_retransmits
= 0;
3120 tcp_clear_retrans(tp
);
3123 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
3125 struct tcp_sock
*tp
= tcp_sk(sk
);
3126 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
3128 tcb
->end_seq
+= skb
->len
;
3129 __skb_header_release(skb
);
3130 __tcp_add_write_queue_tail(sk
, skb
);
3131 sk
->sk_wmem_queued
+= skb
->truesize
;
3132 sk_mem_charge(sk
, skb
->truesize
);
3133 tp
->write_seq
= tcb
->end_seq
;
3134 tp
->packets_out
+= tcp_skb_pcount(skb
);
3137 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3138 * queue a data-only packet after the regular SYN, such that regular SYNs
3139 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3140 * only the SYN sequence, the data are retransmitted in the first ACK.
3141 * If cookie is not cached or other error occurs, falls back to send a
3142 * regular SYN with Fast Open cookie request option.
3144 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
3146 struct tcp_sock
*tp
= tcp_sk(sk
);
3147 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
3148 int syn_loss
= 0, space
, err
= 0, copied
;
3149 unsigned long last_syn_loss
= 0;
3150 struct sk_buff
*syn_data
;
3152 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
3153 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
3154 &syn_loss
, &last_syn_loss
);
3155 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3157 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
3158 fo
->cookie
.len
= -1;
3162 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
3163 fo
->cookie
.len
= -1;
3164 else if (fo
->cookie
.len
<= 0)
3167 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3168 * user-MSS. Reserve maximum option space for middleboxes that add
3169 * private TCP options. The cost is reduced data space in SYN :(
3171 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
3172 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3173 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
3174 MAX_TCP_OPTION_SPACE
;
3176 space
= min_t(size_t, space
, fo
->size
);
3178 /* limit to order-0 allocations */
3179 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
3181 syn_data
= sk_stream_alloc_skb(sk
, space
, sk
->sk_allocation
, false);
3184 syn_data
->ip_summed
= CHECKSUM_PARTIAL
;
3185 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
3186 copied
= copy_from_iter(skb_put(syn_data
, space
), space
,
3187 &fo
->data
->msg_iter
);
3188 if (unlikely(!copied
)) {
3189 kfree_skb(syn_data
);
3192 if (copied
!= space
) {
3193 skb_trim(syn_data
, copied
);
3197 /* No more data pending in inet_wait_for_connect() */
3198 if (space
== fo
->size
)
3202 tcp_connect_queue_skb(sk
, syn_data
);
3204 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
3206 syn
->skb_mstamp
= syn_data
->skb_mstamp
;
3208 /* Now full SYN+DATA was cloned and sent (or not),
3209 * remove the SYN from the original skb (syn_data)
3210 * we keep in write queue in case of a retransmit, as we
3211 * also have the SYN packet (with no data) in the same queue.
3213 TCP_SKB_CB(syn_data
)->seq
++;
3214 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
3216 tp
->syn_data
= (fo
->copied
> 0);
3217 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3222 /* Send a regular SYN with Fast Open cookie request option */
3223 if (fo
->cookie
.len
> 0)
3225 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3227 tp
->syn_fastopen
= 0;
3229 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3233 /* Build a SYN and send it off. */
3234 int tcp_connect(struct sock
*sk
)
3236 struct tcp_sock
*tp
= tcp_sk(sk
);
3237 struct sk_buff
*buff
;
3240 tcp_connect_init(sk
);
3242 if (unlikely(tp
->repair
)) {
3243 tcp_finish_connect(sk
, NULL
);
3247 buff
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
, true);
3248 if (unlikely(!buff
))
3251 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3252 tp
->retrans_stamp
= tcp_time_stamp
;
3253 tcp_connect_queue_skb(sk
, buff
);
3254 tcp_ecn_send_syn(sk
, buff
);
3256 /* Send off SYN; include data in Fast Open. */
3257 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3258 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3259 if (err
== -ECONNREFUSED
)
3262 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3263 * in order to make this packet get counted in tcpOutSegs.
3265 tp
->snd_nxt
= tp
->write_seq
;
3266 tp
->pushed_seq
= tp
->write_seq
;
3267 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3269 /* Timer for repeating the SYN until an answer. */
3270 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3271 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3274 EXPORT_SYMBOL(tcp_connect
);
3276 /* Send out a delayed ack, the caller does the policy checking
3277 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3280 void tcp_send_delayed_ack(struct sock
*sk
)
3282 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3283 int ato
= icsk
->icsk_ack
.ato
;
3284 unsigned long timeout
;
3286 tcp_ca_event(sk
, CA_EVENT_DELAYED_ACK
);
3288 if (ato
> TCP_DELACK_MIN
) {
3289 const struct tcp_sock
*tp
= tcp_sk(sk
);
3290 int max_ato
= HZ
/ 2;
3292 if (icsk
->icsk_ack
.pingpong
||
3293 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3294 max_ato
= TCP_DELACK_MAX
;
3296 /* Slow path, intersegment interval is "high". */
3298 /* If some rtt estimate is known, use it to bound delayed ack.
3299 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3303 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3310 ato
= min(ato
, max_ato
);
3313 /* Stay within the limit we were given */
3314 timeout
= jiffies
+ ato
;
3316 /* Use new timeout only if there wasn't a older one earlier. */
3317 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3318 /* If delack timer was blocked or is about to expire,
3321 if (icsk
->icsk_ack
.blocked
||
3322 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3327 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3328 timeout
= icsk
->icsk_ack
.timeout
;
3330 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3331 icsk
->icsk_ack
.timeout
= timeout
;
3332 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3335 /* This routine sends an ack and also updates the window. */
3336 void tcp_send_ack(struct sock
*sk
)
3338 struct sk_buff
*buff
;
3340 /* If we have been reset, we may not send again. */
3341 if (sk
->sk_state
== TCP_CLOSE
)
3344 tcp_ca_event(sk
, CA_EVENT_NON_DELAYED_ACK
);
3346 /* We are not putting this on the write queue, so
3347 * tcp_transmit_skb() will set the ownership to this
3350 buff
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3352 inet_csk_schedule_ack(sk
);
3353 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3354 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3355 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3359 /* Reserve space for headers and prepare control bits. */
3360 skb_reserve(buff
, MAX_TCP_HEADER
);
3361 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3363 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3365 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3366 * We also avoid tcp_wfree() overhead (cache line miss accessing
3367 * tp->tsq_flags) by using regular sock_wfree()
3369 skb_set_tcp_pure_ack(buff
);
3371 /* Send it off, this clears delayed acks for us. */
3372 skb_mstamp_get(&buff
->skb_mstamp
);
3373 tcp_transmit_skb(sk
, buff
, 0, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3375 EXPORT_SYMBOL_GPL(tcp_send_ack
);
3377 /* This routine sends a packet with an out of date sequence
3378 * number. It assumes the other end will try to ack it.
3380 * Question: what should we make while urgent mode?
3381 * 4.4BSD forces sending single byte of data. We cannot send
3382 * out of window data, because we have SND.NXT==SND.MAX...
3384 * Current solution: to send TWO zero-length segments in urgent mode:
3385 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3386 * out-of-date with SND.UNA-1 to probe window.
3388 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
, int mib
)
3390 struct tcp_sock
*tp
= tcp_sk(sk
);
3391 struct sk_buff
*skb
;
3393 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3394 skb
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3398 /* Reserve space for headers and set control bits. */
3399 skb_reserve(skb
, MAX_TCP_HEADER
);
3400 /* Use a previous sequence. This should cause the other
3401 * end to send an ack. Don't queue or clone SKB, just
3404 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3405 skb_mstamp_get(&skb
->skb_mstamp
);
3406 NET_INC_STATS_BH(sock_net(sk
), mib
);
3407 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
3410 void tcp_send_window_probe(struct sock
*sk
)
3412 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3413 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3414 tcp_xmit_probe_skb(sk
, 0, LINUX_MIB_TCPWINPROBE
);
3418 /* Initiate keepalive or window probe from timer. */
3419 int tcp_write_wakeup(struct sock
*sk
, int mib
)
3421 struct tcp_sock
*tp
= tcp_sk(sk
);
3422 struct sk_buff
*skb
;
3424 if (sk
->sk_state
== TCP_CLOSE
)
3427 skb
= tcp_send_head(sk
);
3428 if (skb
&& before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3430 unsigned int mss
= tcp_current_mss(sk
);
3431 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3433 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3434 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3436 /* We are probing the opening of a window
3437 * but the window size is != 0
3438 * must have been a result SWS avoidance ( sender )
3440 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3442 seg_size
= min(seg_size
, mss
);
3443 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3444 if (tcp_fragment(sk
, skb
, seg_size
, mss
, GFP_ATOMIC
))
3446 } else if (!tcp_skb_pcount(skb
))
3447 tcp_set_skb_tso_segs(skb
, mss
);
3449 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3450 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3452 tcp_event_new_data_sent(sk
, skb
);
3455 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3456 tcp_xmit_probe_skb(sk
, 1, mib
);
3457 return tcp_xmit_probe_skb(sk
, 0, mib
);
3461 /* A window probe timeout has occurred. If window is not closed send
3462 * a partial packet else a zero probe.
3464 void tcp_send_probe0(struct sock
*sk
)
3466 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3467 struct tcp_sock
*tp
= tcp_sk(sk
);
3468 unsigned long probe_max
;
3471 err
= tcp_write_wakeup(sk
, LINUX_MIB_TCPWINPROBE
);
3473 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3474 /* Cancel probe timer, if it is not required. */
3475 icsk
->icsk_probes_out
= 0;
3476 icsk
->icsk_backoff
= 0;
3481 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
3482 icsk
->icsk_backoff
++;
3483 icsk
->icsk_probes_out
++;
3484 probe_max
= TCP_RTO_MAX
;
3486 /* If packet was not sent due to local congestion,
3487 * do not backoff and do not remember icsk_probes_out.
3488 * Let local senders to fight for local resources.
3490 * Use accumulated backoff yet.
3492 if (!icsk
->icsk_probes_out
)
3493 icsk
->icsk_probes_out
= 1;
3494 probe_max
= TCP_RESOURCE_PROBE_INTERVAL
;
3496 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3497 tcp_probe0_when(sk
, probe_max
),
3501 int tcp_rtx_synack(struct sock
*sk
, struct request_sock
*req
)
3503 const struct tcp_request_sock_ops
*af_ops
= tcp_rsk(req
)->af_specific
;
3507 res
= af_ops
->send_synack(sk
, NULL
, &fl
, req
, 0, NULL
);
3509 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
3510 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
3514 EXPORT_SYMBOL(tcp_rtx_synack
);