2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
53 /* Default TSQ limit of two TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly
= 131072;
56 /* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
60 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
62 /* 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
;
166 const struct dst_entry
*dst
= __sk_dst_get(sk
);
168 if (sysctl_tcp_slow_start_after_idle
&&
169 (!tp
->packets_out
&& (s32
)(now
- tp
->lsndtime
) > icsk
->icsk_rto
))
170 tcp_cwnd_restart(sk
, __sk_dst_get(sk
));
174 /* If it is a reply for ato after last received
175 * packet, enter pingpong mode.
177 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
&&
178 (!dst
|| !dst_metric(dst
, RTAX_QUICKACK
)))
179 icsk
->icsk_ack
.pingpong
= 1;
182 /* Account for an ACK we sent. */
183 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
185 tcp_dec_quickack_mode(sk
, pkts
);
186 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
190 u32
tcp_default_init_rwnd(u32 mss
)
192 /* Initial receive window should be twice of TCP_INIT_CWND to
193 * enable proper sending of new unsent data during fast recovery
194 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
195 * limit when mss is larger than 1460.
197 u32 init_rwnd
= TCP_INIT_CWND
* 2;
200 init_rwnd
= max((1460 * init_rwnd
) / mss
, 2U);
204 /* Determine a window scaling and initial window to offer.
205 * Based on the assumption that the given amount of space
206 * will be offered. Store the results in the tp structure.
207 * NOTE: for smooth operation initial space offering should
208 * be a multiple of mss if possible. We assume here that mss >= 1.
209 * This MUST be enforced by all callers.
211 void tcp_select_initial_window(int __space
, __u32 mss
,
212 __u32
*rcv_wnd
, __u32
*window_clamp
,
213 int wscale_ok
, __u8
*rcv_wscale
,
216 unsigned int space
= (__space
< 0 ? 0 : __space
);
218 /* If no clamp set the clamp to the max possible scaled window */
219 if (*window_clamp
== 0)
220 (*window_clamp
) = (65535 << 14);
221 space
= min(*window_clamp
, space
);
223 /* Quantize space offering to a multiple of mss if possible. */
225 space
= (space
/ mss
) * mss
;
227 /* NOTE: offering an initial window larger than 32767
228 * will break some buggy TCP stacks. If the admin tells us
229 * it is likely we could be speaking with such a buggy stack
230 * we will truncate our initial window offering to 32K-1
231 * unless the remote has sent us a window scaling option,
232 * which we interpret as a sign the remote TCP is not
233 * misinterpreting the window field as a signed quantity.
235 if (sysctl_tcp_workaround_signed_windows
)
236 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
242 /* Set window scaling on max possible window
243 * See RFC1323 for an explanation of the limit to 14
245 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
246 space
= min_t(u32
, space
, *window_clamp
);
247 while (space
> 65535 && (*rcv_wscale
) < 14) {
253 if (mss
> (1 << *rcv_wscale
)) {
254 if (!init_rcv_wnd
) /* Use default unless specified otherwise */
255 init_rcv_wnd
= tcp_default_init_rwnd(mss
);
256 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
259 /* Set the clamp no higher than max representable value */
260 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
262 EXPORT_SYMBOL(tcp_select_initial_window
);
264 /* Chose a new window to advertise, update state in tcp_sock for the
265 * socket, and return result with RFC1323 scaling applied. The return
266 * value can be stuffed directly into th->window for an outgoing
269 static u16
tcp_select_window(struct sock
*sk
)
271 struct tcp_sock
*tp
= tcp_sk(sk
);
272 u32 old_win
= tp
->rcv_wnd
;
273 u32 cur_win
= tcp_receive_window(tp
);
274 u32 new_win
= __tcp_select_window(sk
);
276 /* Never shrink the offered window */
277 if (new_win
< cur_win
) {
278 /* Danger Will Robinson!
279 * Don't update rcv_wup/rcv_wnd here or else
280 * we will not be able to advertise a zero
281 * window in time. --DaveM
283 * Relax Will Robinson.
286 NET_INC_STATS(sock_net(sk
),
287 LINUX_MIB_TCPWANTZEROWINDOWADV
);
288 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
290 tp
->rcv_wnd
= new_win
;
291 tp
->rcv_wup
= tp
->rcv_nxt
;
293 /* Make sure we do not exceed the maximum possible
296 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
297 new_win
= min(new_win
, MAX_TCP_WINDOW
);
299 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
301 /* RFC1323 scaling applied */
302 new_win
>>= tp
->rx_opt
.rcv_wscale
;
304 /* If we advertise zero window, disable fast path. */
308 NET_INC_STATS(sock_net(sk
),
309 LINUX_MIB_TCPTOZEROWINDOWADV
);
310 } else if (old_win
== 0) {
311 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFROMZEROWINDOWADV
);
317 /* Packet ECN state for a SYN-ACK */
318 static void tcp_ecn_send_synack(struct sock
*sk
, struct sk_buff
*skb
)
320 const struct tcp_sock
*tp
= tcp_sk(sk
);
322 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
323 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
324 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
325 else if (tcp_ca_needs_ecn(sk
))
329 /* Packet ECN state for a SYN. */
330 static void tcp_ecn_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
332 struct tcp_sock
*tp
= tcp_sk(sk
);
333 bool use_ecn
= sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1 ||
334 tcp_ca_needs_ecn(sk
);
337 const struct dst_entry
*dst
= __sk_dst_get(sk
);
339 if (dst
&& dst_feature(dst
, RTAX_FEATURE_ECN
))
346 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
347 tp
->ecn_flags
= TCP_ECN_OK
;
348 if (tcp_ca_needs_ecn(sk
))
354 tcp_ecn_make_synack(const struct request_sock
*req
, struct tcphdr
*th
,
357 if (inet_rsk(req
)->ecn_ok
) {
359 if (tcp_ca_needs_ecn(sk
))
364 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
367 static void tcp_ecn_send(struct sock
*sk
, struct sk_buff
*skb
,
370 struct tcp_sock
*tp
= tcp_sk(sk
);
372 if (tp
->ecn_flags
& TCP_ECN_OK
) {
373 /* Not-retransmitted data segment: set ECT and inject CWR. */
374 if (skb
->len
!= tcp_header_len
&&
375 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
377 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
378 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
379 tcp_hdr(skb
)->cwr
= 1;
380 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
382 } else if (!tcp_ca_needs_ecn(sk
)) {
383 /* ACK or retransmitted segment: clear ECT|CE */
384 INET_ECN_dontxmit(sk
);
386 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
387 tcp_hdr(skb
)->ece
= 1;
391 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
392 * auto increment end seqno.
394 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
396 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
398 skb
->ip_summed
= CHECKSUM_PARTIAL
;
401 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
402 TCP_SKB_CB(skb
)->sacked
= 0;
404 tcp_skb_pcount_set(skb
, 1);
405 shinfo
->gso_size
= 0;
406 shinfo
->gso_type
= 0;
408 TCP_SKB_CB(skb
)->seq
= seq
;
409 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
411 TCP_SKB_CB(skb
)->end_seq
= seq
;
414 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
416 return tp
->snd_una
!= tp
->snd_up
;
419 #define OPTION_SACK_ADVERTISE (1 << 0)
420 #define OPTION_TS (1 << 1)
421 #define OPTION_MD5 (1 << 2)
422 #define OPTION_WSCALE (1 << 3)
423 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
425 struct tcp_out_options
{
426 u16 options
; /* bit field of OPTION_* */
427 u16 mss
; /* 0 to disable */
428 u8 ws
; /* window scale, 0 to disable */
429 u8 num_sack_blocks
; /* number of SACK blocks to include */
430 u8 hash_size
; /* bytes in hash_location */
431 __u8
*hash_location
; /* temporary pointer, overloaded */
432 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
433 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
436 /* Write previously computed TCP options to the packet.
438 * Beware: Something in the Internet is very sensitive to the ordering of
439 * TCP options, we learned this through the hard way, so be careful here.
440 * Luckily we can at least blame others for their non-compliance but from
441 * inter-operability perspective it seems that we're somewhat stuck with
442 * the ordering which we have been using if we want to keep working with
443 * those broken things (not that it currently hurts anybody as there isn't
444 * particular reason why the ordering would need to be changed).
446 * At least SACK_PERM as the first option is known to lead to a disaster
447 * (but it may well be that other scenarios fail similarly).
449 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
450 struct tcp_out_options
*opts
)
452 u16 options
= opts
->options
; /* mungable copy */
454 if (unlikely(OPTION_MD5
& options
)) {
455 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
456 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
457 /* overload cookie hash location */
458 opts
->hash_location
= (__u8
*)ptr
;
462 if (unlikely(opts
->mss
)) {
463 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
464 (TCPOLEN_MSS
<< 16) |
468 if (likely(OPTION_TS
& options
)) {
469 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
470 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
471 (TCPOLEN_SACK_PERM
<< 16) |
472 (TCPOPT_TIMESTAMP
<< 8) |
474 options
&= ~OPTION_SACK_ADVERTISE
;
476 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
478 (TCPOPT_TIMESTAMP
<< 8) |
481 *ptr
++ = htonl(opts
->tsval
);
482 *ptr
++ = htonl(opts
->tsecr
);
485 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
486 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
488 (TCPOPT_SACK_PERM
<< 8) |
492 if (unlikely(OPTION_WSCALE
& options
)) {
493 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
494 (TCPOPT_WINDOW
<< 16) |
495 (TCPOLEN_WINDOW
<< 8) |
499 if (unlikely(opts
->num_sack_blocks
)) {
500 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
501 tp
->duplicate_sack
: tp
->selective_acks
;
504 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
507 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
508 TCPOLEN_SACK_PERBLOCK
)));
510 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
512 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
513 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
516 tp
->rx_opt
.dsack
= 0;
519 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
520 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
522 *ptr
++ = htonl((TCPOPT_EXP
<< 24) |
523 ((TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
) << 16) |
524 TCPOPT_FASTOPEN_MAGIC
);
526 memcpy(ptr
, foc
->val
, foc
->len
);
527 if ((foc
->len
& 3) == 2) {
528 u8
*align
= ((u8
*)ptr
) + foc
->len
;
529 align
[0] = align
[1] = TCPOPT_NOP
;
531 ptr
+= (foc
->len
+ 3) >> 2;
535 /* Compute TCP options for SYN packets. This is not the final
536 * network wire format yet.
538 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
539 struct tcp_out_options
*opts
,
540 struct tcp_md5sig_key
**md5
)
542 struct tcp_sock
*tp
= tcp_sk(sk
);
543 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
544 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
546 #ifdef CONFIG_TCP_MD5SIG
547 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
549 opts
->options
|= OPTION_MD5
;
550 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
556 /* We always get an MSS option. The option bytes which will be seen in
557 * normal data packets should timestamps be used, must be in the MSS
558 * advertised. But we subtract them from tp->mss_cache so that
559 * calculations in tcp_sendmsg are simpler etc. So account for this
560 * fact here if necessary. If we don't do this correctly, as a
561 * receiver we won't recognize data packets as being full sized when we
562 * should, and thus we won't abide by the delayed ACK rules correctly.
563 * SACKs don't matter, we never delay an ACK when we have any of those
565 opts
->mss
= tcp_advertise_mss(sk
);
566 remaining
-= TCPOLEN_MSS_ALIGNED
;
568 if (likely(sysctl_tcp_timestamps
&& *md5
== NULL
)) {
569 opts
->options
|= OPTION_TS
;
570 opts
->tsval
= tcp_skb_timestamp(skb
) + tp
->tsoffset
;
571 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
572 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
574 if (likely(sysctl_tcp_window_scaling
)) {
575 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
576 opts
->options
|= OPTION_WSCALE
;
577 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
579 if (likely(sysctl_tcp_sack
)) {
580 opts
->options
|= OPTION_SACK_ADVERTISE
;
581 if (unlikely(!(OPTION_TS
& opts
->options
)))
582 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
585 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
586 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ fastopen
->cookie
.len
;
587 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
588 if (remaining
>= need
) {
589 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
590 opts
->fastopen_cookie
= &fastopen
->cookie
;
592 tp
->syn_fastopen
= 1;
596 return MAX_TCP_OPTION_SPACE
- remaining
;
599 /* Set up TCP options for SYN-ACKs. */
600 static unsigned int tcp_synack_options(struct sock
*sk
,
601 struct request_sock
*req
,
602 unsigned int mss
, struct sk_buff
*skb
,
603 struct tcp_out_options
*opts
,
604 const struct tcp_md5sig_key
*md5
,
605 struct tcp_fastopen_cookie
*foc
)
607 struct inet_request_sock
*ireq
= inet_rsk(req
);
608 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
610 #ifdef CONFIG_TCP_MD5SIG
612 opts
->options
|= OPTION_MD5
;
613 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
615 /* We can't fit any SACK blocks in a packet with MD5 + TS
616 * options. There was discussion about disabling SACK
617 * rather than TS in order to fit in better with old,
618 * buggy kernels, but that was deemed to be unnecessary.
620 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
624 /* We always send an MSS option. */
626 remaining
-= TCPOLEN_MSS_ALIGNED
;
628 if (likely(ireq
->wscale_ok
)) {
629 opts
->ws
= ireq
->rcv_wscale
;
630 opts
->options
|= OPTION_WSCALE
;
631 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
633 if (likely(ireq
->tstamp_ok
)) {
634 opts
->options
|= OPTION_TS
;
635 opts
->tsval
= tcp_skb_timestamp(skb
);
636 opts
->tsecr
= req
->ts_recent
;
637 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
639 if (likely(ireq
->sack_ok
)) {
640 opts
->options
|= OPTION_SACK_ADVERTISE
;
641 if (unlikely(!ireq
->tstamp_ok
))
642 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
644 if (foc
!= NULL
&& foc
->len
>= 0) {
645 u32 need
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
646 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
647 if (remaining
>= need
) {
648 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
649 opts
->fastopen_cookie
= foc
;
654 return MAX_TCP_OPTION_SPACE
- remaining
;
657 /* Compute TCP options for ESTABLISHED sockets. This is not the
658 * final wire format yet.
660 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
661 struct tcp_out_options
*opts
,
662 struct tcp_md5sig_key
**md5
)
664 struct tcp_sock
*tp
= tcp_sk(sk
);
665 unsigned int size
= 0;
666 unsigned int eff_sacks
;
670 #ifdef CONFIG_TCP_MD5SIG
671 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
672 if (unlikely(*md5
)) {
673 opts
->options
|= OPTION_MD5
;
674 size
+= TCPOLEN_MD5SIG_ALIGNED
;
680 if (likely(tp
->rx_opt
.tstamp_ok
)) {
681 opts
->options
|= OPTION_TS
;
682 opts
->tsval
= skb
? tcp_skb_timestamp(skb
) + tp
->tsoffset
: 0;
683 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
684 size
+= TCPOLEN_TSTAMP_ALIGNED
;
687 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
688 if (unlikely(eff_sacks
)) {
689 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
690 opts
->num_sack_blocks
=
691 min_t(unsigned int, eff_sacks
,
692 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
693 TCPOLEN_SACK_PERBLOCK
);
694 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
695 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
702 /* TCP SMALL QUEUES (TSQ)
704 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
705 * to reduce RTT and bufferbloat.
706 * We do this using a special skb destructor (tcp_wfree).
708 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
709 * needs to be reallocated in a driver.
710 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
712 * Since transmit from skb destructor is forbidden, we use a tasklet
713 * to process all sockets that eventually need to send more skbs.
714 * We use one tasklet per cpu, with its own queue of sockets.
717 struct tasklet_struct tasklet
;
718 struct list_head head
; /* queue of tcp sockets */
720 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
722 static void tcp_tsq_handler(struct sock
*sk
)
724 if ((1 << sk
->sk_state
) &
725 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
726 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
))
727 tcp_write_xmit(sk
, tcp_current_mss(sk
), tcp_sk(sk
)->nonagle
,
731 * One tasklet per cpu tries to send more skbs.
732 * We run in tasklet context but need to disable irqs when
733 * transferring tsq->head because tcp_wfree() might
734 * interrupt us (non NAPI drivers)
736 static void tcp_tasklet_func(unsigned long data
)
738 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
741 struct list_head
*q
, *n
;
745 local_irq_save(flags
);
746 list_splice_init(&tsq
->head
, &list
);
747 local_irq_restore(flags
);
749 list_for_each_safe(q
, n
, &list
) {
750 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
751 list_del(&tp
->tsq_node
);
753 sk
= (struct sock
*)tp
;
756 if (!sock_owned_by_user(sk
)) {
759 /* defer the work to tcp_release_cb() */
760 set_bit(TCP_TSQ_DEFERRED
, &tp
->tsq_flags
);
764 clear_bit(TSQ_QUEUED
, &tp
->tsq_flags
);
769 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
770 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
771 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
772 (1UL << TCP_MTU_REDUCED_DEFERRED))
774 * tcp_release_cb - tcp release_sock() callback
777 * called from release_sock() to perform protocol dependent
778 * actions before socket release.
780 void tcp_release_cb(struct sock
*sk
)
782 struct tcp_sock
*tp
= tcp_sk(sk
);
783 unsigned long flags
, nflags
;
785 /* perform an atomic operation only if at least one flag is set */
787 flags
= tp
->tsq_flags
;
788 if (!(flags
& TCP_DEFERRED_ALL
))
790 nflags
= flags
& ~TCP_DEFERRED_ALL
;
791 } while (cmpxchg(&tp
->tsq_flags
, flags
, nflags
) != flags
);
793 if (flags
& (1UL << TCP_TSQ_DEFERRED
))
796 /* Here begins the tricky part :
797 * We are called from release_sock() with :
799 * 2) sk_lock.slock spinlock held
800 * 3) socket owned by us (sk->sk_lock.owned == 1)
802 * But following code is meant to be called from BH handlers,
803 * so we should keep BH disabled, but early release socket ownership
805 sock_release_ownership(sk
);
807 if (flags
& (1UL << TCP_WRITE_TIMER_DEFERRED
)) {
808 tcp_write_timer_handler(sk
);
811 if (flags
& (1UL << TCP_DELACK_TIMER_DEFERRED
)) {
812 tcp_delack_timer_handler(sk
);
815 if (flags
& (1UL << TCP_MTU_REDUCED_DEFERRED
)) {
816 inet_csk(sk
)->icsk_af_ops
->mtu_reduced(sk
);
820 EXPORT_SYMBOL(tcp_release_cb
);
822 void __init
tcp_tasklet_init(void)
826 for_each_possible_cpu(i
) {
827 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
829 INIT_LIST_HEAD(&tsq
->head
);
830 tasklet_init(&tsq
->tasklet
,
837 * Write buffer destructor automatically called from kfree_skb.
838 * We can't xmit new skbs from this context, as we might already
841 void tcp_wfree(struct sk_buff
*skb
)
843 struct sock
*sk
= skb
->sk
;
844 struct tcp_sock
*tp
= tcp_sk(sk
);
847 /* Keep one reference on sk_wmem_alloc.
848 * Will be released by sk_free() from here or tcp_tasklet_func()
850 wmem
= atomic_sub_return(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
852 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
853 * Wait until our queues (qdisc + devices) are drained.
855 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
856 * - chance for incoming ACK (processed by another cpu maybe)
857 * to migrate this flow (skb->ooo_okay will be eventually set)
859 if (wmem
>= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current
)
862 if (test_and_clear_bit(TSQ_THROTTLED
, &tp
->tsq_flags
) &&
863 !test_and_set_bit(TSQ_QUEUED
, &tp
->tsq_flags
)) {
865 struct tsq_tasklet
*tsq
;
867 /* queue this socket to tasklet queue */
868 local_irq_save(flags
);
869 tsq
= this_cpu_ptr(&tsq_tasklet
);
870 list_add(&tp
->tsq_node
, &tsq
->head
);
871 tasklet_schedule(&tsq
->tasklet
);
872 local_irq_restore(flags
);
879 /* This routine actually transmits TCP packets queued in by
880 * tcp_do_sendmsg(). This is used by both the initial
881 * transmission and possible later retransmissions.
882 * All SKB's seen here are completely headerless. It is our
883 * job to build the TCP header, and pass the packet down to
884 * IP so it can do the same plus pass the packet off to the
887 * We are working here with either a clone of the original
888 * SKB, or a fresh unique copy made by the retransmit engine.
890 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
893 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
894 struct inet_sock
*inet
;
896 struct tcp_skb_cb
*tcb
;
897 struct tcp_out_options opts
;
898 unsigned int tcp_options_size
, tcp_header_size
;
899 struct tcp_md5sig_key
*md5
;
903 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
906 skb_mstamp_get(&skb
->skb_mstamp
);
908 if (unlikely(skb_cloned(skb
)))
909 skb
= pskb_copy(skb
, gfp_mask
);
911 skb
= skb_clone(skb
, gfp_mask
);
918 tcb
= TCP_SKB_CB(skb
);
919 memset(&opts
, 0, sizeof(opts
));
921 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
922 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
924 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
926 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
928 if (tcp_packets_in_flight(tp
) == 0)
929 tcp_ca_event(sk
, CA_EVENT_TX_START
);
931 /* if no packet is in qdisc/device queue, then allow XPS to select
932 * another queue. We can be called from tcp_tsq_handler()
933 * which holds one reference to sk_wmem_alloc.
935 * TODO: Ideally, in-flight pure ACK packets should not matter here.
936 * One way to get this would be to set skb->truesize = 2 on them.
938 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) < SKB_TRUESIZE(1);
940 skb_push(skb
, tcp_header_size
);
941 skb_reset_transport_header(skb
);
945 skb
->destructor
= skb_is_tcp_pure_ack(skb
) ? sock_wfree
: tcp_wfree
;
946 skb_set_hash_from_sk(skb
, sk
);
947 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
949 /* Build TCP header and checksum it. */
951 th
->source
= inet
->inet_sport
;
952 th
->dest
= inet
->inet_dport
;
953 th
->seq
= htonl(tcb
->seq
);
954 th
->ack_seq
= htonl(tp
->rcv_nxt
);
955 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
958 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
959 /* RFC1323: The window in SYN & SYN/ACK segments
962 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
964 th
->window
= htons(tcp_select_window(sk
));
969 /* The urg_mode check is necessary during a below snd_una win probe */
970 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
971 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
972 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
974 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
975 th
->urg_ptr
= htons(0xFFFF);
980 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
981 if (likely((tcb
->tcp_flags
& TCPHDR_SYN
) == 0))
982 tcp_ecn_send(sk
, skb
, tcp_header_size
);
984 #ifdef CONFIG_TCP_MD5SIG
985 /* Calculate the MD5 hash, as we have all we need now */
987 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
988 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
993 icsk
->icsk_af_ops
->send_check(sk
, skb
);
995 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
996 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
998 if (skb
->len
!= tcp_header_size
)
999 tcp_event_data_sent(tp
, sk
);
1001 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
1002 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
1003 tcp_skb_pcount(skb
));
1005 /* OK, its time to fill skb_shinfo(skb)->gso_segs */
1006 skb_shinfo(skb
)->gso_segs
= tcp_skb_pcount(skb
);
1008 /* Our usage of tstamp should remain private */
1009 skb
->tstamp
.tv64
= 0;
1011 /* Cleanup our debris for IP stacks */
1012 memset(skb
->cb
, 0, max(sizeof(struct inet_skb_parm
),
1013 sizeof(struct inet6_skb_parm
)));
1015 err
= icsk
->icsk_af_ops
->queue_xmit(sk
, skb
, &inet
->cork
.fl
);
1017 if (likely(err
<= 0))
1022 return net_xmit_eval(err
);
1025 /* This routine just queues the buffer for sending.
1027 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1028 * otherwise socket can stall.
1030 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1032 struct tcp_sock
*tp
= tcp_sk(sk
);
1034 /* Advance write_seq and place onto the write_queue. */
1035 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
1036 __skb_header_release(skb
);
1037 tcp_add_write_queue_tail(sk
, skb
);
1038 sk
->sk_wmem_queued
+= skb
->truesize
;
1039 sk_mem_charge(sk
, skb
->truesize
);
1042 /* Initialize TSO segments for a packet. */
1043 static void tcp_set_skb_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
1044 unsigned int mss_now
)
1046 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1048 /* Make sure we own this skb before messing gso_size/gso_segs */
1049 WARN_ON_ONCE(skb_cloned(skb
));
1051 if (skb
->len
<= mss_now
|| skb
->ip_summed
== CHECKSUM_NONE
) {
1052 /* Avoid the costly divide in the normal
1055 tcp_skb_pcount_set(skb
, 1);
1056 shinfo
->gso_size
= 0;
1057 shinfo
->gso_type
= 0;
1059 tcp_skb_pcount_set(skb
, DIV_ROUND_UP(skb
->len
, mss_now
));
1060 shinfo
->gso_size
= mss_now
;
1061 shinfo
->gso_type
= sk
->sk_gso_type
;
1065 /* When a modification to fackets out becomes necessary, we need to check
1066 * skb is counted to fackets_out or not.
1068 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1071 struct tcp_sock
*tp
= tcp_sk(sk
);
1073 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1076 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1077 tp
->fackets_out
-= decr
;
1080 /* Pcount in the middle of the write queue got changed, we need to do various
1081 * tweaks to fix counters
1083 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1085 struct tcp_sock
*tp
= tcp_sk(sk
);
1087 tp
->packets_out
-= decr
;
1089 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1090 tp
->sacked_out
-= decr
;
1091 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1092 tp
->retrans_out
-= decr
;
1093 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1094 tp
->lost_out
-= decr
;
1096 /* Reno case is special. Sigh... */
1097 if (tcp_is_reno(tp
) && decr
> 0)
1098 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1100 tcp_adjust_fackets_out(sk
, skb
, decr
);
1102 if (tp
->lost_skb_hint
&&
1103 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1104 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1105 tp
->lost_cnt_hint
-= decr
;
1107 tcp_verify_left_out(tp
);
1110 static void tcp_fragment_tstamp(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1112 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1114 if (unlikely(shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
) &&
1115 !before(shinfo
->tskey
, TCP_SKB_CB(skb2
)->seq
)) {
1116 struct skb_shared_info
*shinfo2
= skb_shinfo(skb2
);
1117 u8 tsflags
= shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
1119 shinfo
->tx_flags
&= ~tsflags
;
1120 shinfo2
->tx_flags
|= tsflags
;
1121 swap(shinfo
->tskey
, shinfo2
->tskey
);
1125 /* Function to create two new TCP segments. Shrinks the given segment
1126 * to the specified size and appends a new segment with the rest of the
1127 * packet to the list. This won't be called frequently, I hope.
1128 * Remember, these are still headerless SKBs at this point.
1130 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1131 unsigned int mss_now
, gfp_t gfp
)
1133 struct tcp_sock
*tp
= tcp_sk(sk
);
1134 struct sk_buff
*buff
;
1135 int nsize
, old_factor
;
1139 if (WARN_ON(len
> skb
->len
))
1142 nsize
= skb_headlen(skb
) - len
;
1146 if (skb_unclone(skb
, gfp
))
1149 /* Get a new skb... force flag on. */
1150 buff
= sk_stream_alloc_skb(sk
, nsize
, gfp
);
1152 return -ENOMEM
; /* We'll just try again later. */
1154 sk
->sk_wmem_queued
+= buff
->truesize
;
1155 sk_mem_charge(sk
, buff
->truesize
);
1156 nlen
= skb
->len
- len
- nsize
;
1157 buff
->truesize
+= nlen
;
1158 skb
->truesize
-= nlen
;
1160 /* Correct the sequence numbers. */
1161 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1162 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1163 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1165 /* PSH and FIN should only be set in the second packet. */
1166 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1167 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1168 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1169 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1171 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1172 /* Copy and checksum data tail into the new buffer. */
1173 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1174 skb_put(buff
, nsize
),
1179 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1181 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1182 skb_split(skb
, buff
, len
);
1185 buff
->ip_summed
= skb
->ip_summed
;
1187 buff
->tstamp
= skb
->tstamp
;
1188 tcp_fragment_tstamp(skb
, buff
);
1190 old_factor
= tcp_skb_pcount(skb
);
1192 /* Fix up tso_factor for both original and new SKB. */
1193 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1194 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1196 /* If this packet has been sent out already, we must
1197 * adjust the various packet counters.
1199 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1200 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1201 tcp_skb_pcount(buff
);
1204 tcp_adjust_pcount(sk
, skb
, diff
);
1207 /* Link BUFF into the send queue. */
1208 __skb_header_release(buff
);
1209 tcp_insert_write_queue_after(skb
, buff
, sk
);
1214 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1215 * eventually). The difference is that pulled data not copied, but
1216 * immediately discarded.
1218 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1220 struct skb_shared_info
*shinfo
;
1223 eat
= min_t(int, len
, skb_headlen(skb
));
1225 __skb_pull(skb
, eat
);
1232 shinfo
= skb_shinfo(skb
);
1233 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1234 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1237 skb_frag_unref(skb
, i
);
1240 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1242 shinfo
->frags
[k
].page_offset
+= eat
;
1243 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1249 shinfo
->nr_frags
= k
;
1251 skb_reset_tail_pointer(skb
);
1252 skb
->data_len
-= len
;
1253 skb
->len
= skb
->data_len
;
1256 /* Remove acked data from a packet in the transmit queue. */
1257 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1259 if (skb_unclone(skb
, GFP_ATOMIC
))
1262 __pskb_trim_head(skb
, len
);
1264 TCP_SKB_CB(skb
)->seq
+= len
;
1265 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1267 skb
->truesize
-= len
;
1268 sk
->sk_wmem_queued
-= len
;
1269 sk_mem_uncharge(sk
, len
);
1270 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1272 /* Any change of skb->len requires recalculation of tso factor. */
1273 if (tcp_skb_pcount(skb
) > 1)
1274 tcp_set_skb_tso_segs(sk
, skb
, tcp_skb_mss(skb
));
1279 /* Calculate MSS not accounting any TCP options. */
1280 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1282 const struct tcp_sock
*tp
= tcp_sk(sk
);
1283 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1286 /* Calculate base mss without TCP options:
1287 It is MMS_S - sizeof(tcphdr) of rfc1122
1289 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1291 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1292 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1293 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1295 if (dst
&& dst_allfrag(dst
))
1296 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1299 /* Clamp it (mss_clamp does not include tcp options) */
1300 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1301 mss_now
= tp
->rx_opt
.mss_clamp
;
1303 /* Now subtract optional transport overhead */
1304 mss_now
-= icsk
->icsk_ext_hdr_len
;
1306 /* Then reserve room for full set of TCP options and 8 bytes of data */
1312 /* Calculate MSS. Not accounting for SACKs here. */
1313 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1315 /* Subtract TCP options size, not including SACKs */
1316 return __tcp_mtu_to_mss(sk
, pmtu
) -
1317 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1320 /* Inverse of above */
1321 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1323 const struct tcp_sock
*tp
= tcp_sk(sk
);
1324 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1328 tp
->tcp_header_len
+
1329 icsk
->icsk_ext_hdr_len
+
1330 icsk
->icsk_af_ops
->net_header_len
;
1332 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1333 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1334 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1336 if (dst
&& dst_allfrag(dst
))
1337 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1342 /* MTU probing init per socket */
1343 void tcp_mtup_init(struct sock
*sk
)
1345 struct tcp_sock
*tp
= tcp_sk(sk
);
1346 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1347 struct net
*net
= sock_net(sk
);
1349 icsk
->icsk_mtup
.enabled
= net
->ipv4
.sysctl_tcp_mtu_probing
> 1;
1350 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1351 icsk
->icsk_af_ops
->net_header_len
;
1352 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, net
->ipv4
.sysctl_tcp_base_mss
);
1353 icsk
->icsk_mtup
.probe_size
= 0;
1354 if (icsk
->icsk_mtup
.enabled
)
1355 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1357 EXPORT_SYMBOL(tcp_mtup_init
);
1359 /* This function synchronize snd mss to current pmtu/exthdr set.
1361 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1362 for TCP options, but includes only bare TCP header.
1364 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1365 It is minimum of user_mss and mss received with SYN.
1366 It also does not include TCP options.
1368 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1370 tp->mss_cache is current effective sending mss, including
1371 all tcp options except for SACKs. It is evaluated,
1372 taking into account current pmtu, but never exceeds
1373 tp->rx_opt.mss_clamp.
1375 NOTE1. rfc1122 clearly states that advertised MSS
1376 DOES NOT include either tcp or ip options.
1378 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1379 are READ ONLY outside this function. --ANK (980731)
1381 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1383 struct tcp_sock
*tp
= tcp_sk(sk
);
1384 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1387 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1388 icsk
->icsk_mtup
.search_high
= pmtu
;
1390 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1391 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1393 /* And store cached results */
1394 icsk
->icsk_pmtu_cookie
= pmtu
;
1395 if (icsk
->icsk_mtup
.enabled
)
1396 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1397 tp
->mss_cache
= mss_now
;
1401 EXPORT_SYMBOL(tcp_sync_mss
);
1403 /* Compute the current effective MSS, taking SACKs and IP options,
1404 * and even PMTU discovery events into account.
1406 unsigned int tcp_current_mss(struct sock
*sk
)
1408 const struct tcp_sock
*tp
= tcp_sk(sk
);
1409 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1411 unsigned int header_len
;
1412 struct tcp_out_options opts
;
1413 struct tcp_md5sig_key
*md5
;
1415 mss_now
= tp
->mss_cache
;
1418 u32 mtu
= dst_mtu(dst
);
1419 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1420 mss_now
= tcp_sync_mss(sk
, mtu
);
1423 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1424 sizeof(struct tcphdr
);
1425 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1426 * some common options. If this is an odd packet (because we have SACK
1427 * blocks etc) then our calculated header_len will be different, and
1428 * we have to adjust mss_now correspondingly */
1429 if (header_len
!= tp
->tcp_header_len
) {
1430 int delta
= (int) header_len
- tp
->tcp_header_len
;
1437 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1438 * As additional protections, we do not touch cwnd in retransmission phases,
1439 * and if application hit its sndbuf limit recently.
1441 static void tcp_cwnd_application_limited(struct sock
*sk
)
1443 struct tcp_sock
*tp
= tcp_sk(sk
);
1445 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1446 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1447 /* Limited by application or receiver window. */
1448 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1449 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1450 if (win_used
< tp
->snd_cwnd
) {
1451 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1452 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
1454 tp
->snd_cwnd_used
= 0;
1456 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1459 static void tcp_cwnd_validate(struct sock
*sk
, bool is_cwnd_limited
)
1461 struct tcp_sock
*tp
= tcp_sk(sk
);
1463 /* Track the maximum number of outstanding packets in each
1464 * window, and remember whether we were cwnd-limited then.
1466 if (!before(tp
->snd_una
, tp
->max_packets_seq
) ||
1467 tp
->packets_out
> tp
->max_packets_out
) {
1468 tp
->max_packets_out
= tp
->packets_out
;
1469 tp
->max_packets_seq
= tp
->snd_nxt
;
1470 tp
->is_cwnd_limited
= is_cwnd_limited
;
1473 if (tcp_is_cwnd_limited(sk
)) {
1474 /* Network is feed fully. */
1475 tp
->snd_cwnd_used
= 0;
1476 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1478 /* Network starves. */
1479 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1480 tp
->snd_cwnd_used
= tp
->packets_out
;
1482 if (sysctl_tcp_slow_start_after_idle
&&
1483 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1484 tcp_cwnd_application_limited(sk
);
1488 /* Minshall's variant of the Nagle send check. */
1489 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1491 return after(tp
->snd_sml
, tp
->snd_una
) &&
1492 !after(tp
->snd_sml
, tp
->snd_nxt
);
1495 /* Update snd_sml if this skb is under mss
1496 * Note that a TSO packet might end with a sub-mss segment
1497 * The test is really :
1498 * if ((skb->len % mss) != 0)
1499 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1500 * But we can avoid doing the divide again given we already have
1501 * skb_pcount = skb->len / mss_now
1503 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1504 const struct sk_buff
*skb
)
1506 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1507 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1510 /* Return false, if packet can be sent now without violation Nagle's rules:
1511 * 1. It is full sized. (provided by caller in %partial bool)
1512 * 2. Or it contains FIN. (already checked by caller)
1513 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1514 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1515 * With Minshall's modification: all sent small packets are ACKed.
1517 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1521 ((nonagle
& TCP_NAGLE_CORK
) ||
1522 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1525 /* Return how many segs we'd like on a TSO packet,
1526 * to send one TSO packet per ms
1528 static u32
tcp_tso_autosize(const struct sock
*sk
, unsigned int mss_now
)
1532 bytes
= min(sk
->sk_pacing_rate
>> 10,
1533 sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
);
1535 /* Goal is to send at least one packet per ms,
1536 * not one big TSO packet every 100 ms.
1537 * This preserves ACK clocking and is consistent
1538 * with tcp_tso_should_defer() heuristic.
1540 segs
= max_t(u32
, bytes
/ mss_now
, sysctl_tcp_min_tso_segs
);
1542 return min_t(u32
, segs
, sk
->sk_gso_max_segs
);
1545 /* Returns the portion of skb which can be sent right away */
1546 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1547 const struct sk_buff
*skb
,
1548 unsigned int mss_now
,
1549 unsigned int max_segs
,
1552 const struct tcp_sock
*tp
= tcp_sk(sk
);
1553 u32 partial
, needed
, window
, max_len
;
1555 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1556 max_len
= mss_now
* max_segs
;
1558 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1561 needed
= min(skb
->len
, window
);
1563 if (max_len
<= needed
)
1566 partial
= needed
% mss_now
;
1567 /* If last segment is not a full MSS, check if Nagle rules allow us
1568 * to include this last segment in this skb.
1569 * Otherwise, we'll split the skb at last MSS boundary
1571 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
1572 return needed
- partial
;
1577 /* Can at least one segment of SKB be sent right now, according to the
1578 * congestion window rules? If so, return how many segments are allowed.
1580 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1581 const struct sk_buff
*skb
)
1583 u32 in_flight
, cwnd
, halfcwnd
;
1585 /* Don't be strict about the congestion window for the final FIN. */
1586 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1587 tcp_skb_pcount(skb
) == 1)
1590 in_flight
= tcp_packets_in_flight(tp
);
1591 cwnd
= tp
->snd_cwnd
;
1592 if (in_flight
>= cwnd
)
1595 /* For better scheduling, ensure we have at least
1596 * 2 GSO packets in flight.
1598 halfcwnd
= max(cwnd
>> 1, 1U);
1599 return min(halfcwnd
, cwnd
- in_flight
);
1602 /* Initialize TSO state of a skb.
1603 * This must be invoked the first time we consider transmitting
1604 * SKB onto the wire.
1606 static int tcp_init_tso_segs(const struct sock
*sk
, struct sk_buff
*skb
,
1607 unsigned int mss_now
)
1609 int tso_segs
= tcp_skb_pcount(skb
);
1611 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1612 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1613 tso_segs
= tcp_skb_pcount(skb
);
1619 /* Return true if the Nagle test allows this packet to be
1622 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1623 unsigned int cur_mss
, int nonagle
)
1625 /* Nagle rule does not apply to frames, which sit in the middle of the
1626 * write_queue (they have no chances to get new data).
1628 * This is implemented in the callers, where they modify the 'nonagle'
1629 * argument based upon the location of SKB in the send queue.
1631 if (nonagle
& TCP_NAGLE_PUSH
)
1634 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1635 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1638 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
1644 /* Does at least the first segment of SKB fit into the send window? */
1645 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1646 const struct sk_buff
*skb
,
1647 unsigned int cur_mss
)
1649 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1651 if (skb
->len
> cur_mss
)
1652 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1654 return !after(end_seq
, tcp_wnd_end(tp
));
1657 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1658 * should be put on the wire right now. If so, it returns the number of
1659 * packets allowed by the congestion window.
1661 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1662 unsigned int cur_mss
, int nonagle
)
1664 const struct tcp_sock
*tp
= tcp_sk(sk
);
1665 unsigned int cwnd_quota
;
1667 tcp_init_tso_segs(sk
, skb
, cur_mss
);
1669 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1672 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1673 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1679 /* Test if sending is allowed right now. */
1680 bool tcp_may_send_now(struct sock
*sk
)
1682 const struct tcp_sock
*tp
= tcp_sk(sk
);
1683 struct sk_buff
*skb
= tcp_send_head(sk
);
1686 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1687 (tcp_skb_is_last(sk
, skb
) ?
1688 tp
->nonagle
: TCP_NAGLE_PUSH
));
1691 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1692 * which is put after SKB on the list. It is very much like
1693 * tcp_fragment() except that it may make several kinds of assumptions
1694 * in order to speed up the splitting operation. In particular, we
1695 * know that all the data is in scatter-gather pages, and that the
1696 * packet has never been sent out before (and thus is not cloned).
1698 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1699 unsigned int mss_now
, gfp_t gfp
)
1701 struct sk_buff
*buff
;
1702 int nlen
= skb
->len
- len
;
1705 /* All of a TSO frame must be composed of paged data. */
1706 if (skb
->len
!= skb
->data_len
)
1707 return tcp_fragment(sk
, skb
, len
, mss_now
, gfp
);
1709 buff
= sk_stream_alloc_skb(sk
, 0, gfp
);
1710 if (unlikely(buff
== NULL
))
1713 sk
->sk_wmem_queued
+= buff
->truesize
;
1714 sk_mem_charge(sk
, buff
->truesize
);
1715 buff
->truesize
+= nlen
;
1716 skb
->truesize
-= nlen
;
1718 /* Correct the sequence numbers. */
1719 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1720 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1721 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1723 /* PSH and FIN should only be set in the second packet. */
1724 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1725 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1726 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1728 /* This packet was never sent out yet, so no SACK bits. */
1729 TCP_SKB_CB(buff
)->sacked
= 0;
1731 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1732 skb_split(skb
, buff
, len
);
1733 tcp_fragment_tstamp(skb
, buff
);
1735 /* Fix up tso_factor for both original and new SKB. */
1736 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1737 tcp_set_skb_tso_segs(sk
, buff
, mss_now
);
1739 /* Link BUFF into the send queue. */
1740 __skb_header_release(buff
);
1741 tcp_insert_write_queue_after(skb
, buff
, sk
);
1746 /* Try to defer sending, if possible, in order to minimize the amount
1747 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1749 * This algorithm is from John Heffner.
1751 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
,
1752 bool *is_cwnd_limited
, u32 max_segs
)
1754 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1755 u32 age
, send_win
, cong_win
, limit
, in_flight
;
1756 struct tcp_sock
*tp
= tcp_sk(sk
);
1757 struct skb_mstamp now
;
1758 struct sk_buff
*head
;
1761 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1764 if (!((1 << icsk
->icsk_ca_state
) & (TCPF_CA_Open
| TCPF_CA_CWR
)))
1767 /* Avoid bursty behavior by allowing defer
1768 * only if the last write was recent.
1770 if ((s32
)(tcp_time_stamp
- tp
->lsndtime
) > 0)
1773 in_flight
= tcp_packets_in_flight(tp
);
1775 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1777 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1779 /* From in_flight test above, we know that cwnd > in_flight. */
1780 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1782 limit
= min(send_win
, cong_win
);
1784 /* If a full-sized TSO skb can be sent, do it. */
1785 if (limit
>= max_segs
* tp
->mss_cache
)
1788 /* Middle in queue won't get any more data, full sendable already? */
1789 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1792 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1794 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1796 /* If at least some fraction of a window is available,
1799 chunk
/= win_divisor
;
1803 /* Different approach, try not to defer past a single
1804 * ACK. Receiver should ACK every other full sized
1805 * frame, so if we have space for more than 3 frames
1808 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1812 head
= tcp_write_queue_head(sk
);
1813 skb_mstamp_get(&now
);
1814 age
= skb_mstamp_us_delta(&now
, &head
->skb_mstamp
);
1815 /* If next ACK is likely to come too late (half srtt), do not defer */
1816 if (age
< (tp
->srtt_us
>> 4))
1819 /* Ok, it looks like it is advisable to defer. */
1821 if (cong_win
< send_win
&& cong_win
< skb
->len
)
1822 *is_cwnd_limited
= true;
1830 static inline void tcp_mtu_check_reprobe(struct sock
*sk
)
1832 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1833 struct tcp_sock
*tp
= tcp_sk(sk
);
1834 struct net
*net
= sock_net(sk
);
1838 interval
= net
->ipv4
.sysctl_tcp_probe_interval
;
1839 delta
= tcp_time_stamp
- icsk
->icsk_mtup
.probe_timestamp
;
1840 if (unlikely(delta
>= interval
* HZ
)) {
1841 int mss
= tcp_current_mss(sk
);
1843 /* Update current search range */
1844 icsk
->icsk_mtup
.probe_size
= 0;
1845 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+
1846 sizeof(struct tcphdr
) +
1847 icsk
->icsk_af_ops
->net_header_len
;
1848 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
1850 /* Update probe time stamp */
1851 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1855 /* Create a new MTU probe if we are ready.
1856 * MTU probe is regularly attempting to increase the path MTU by
1857 * deliberately sending larger packets. This discovers routing
1858 * changes resulting in larger path MTUs.
1860 * Returns 0 if we should wait to probe (no cwnd available),
1861 * 1 if a probe was sent,
1864 static int tcp_mtu_probe(struct sock
*sk
)
1866 struct tcp_sock
*tp
= tcp_sk(sk
);
1867 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1868 struct sk_buff
*skb
, *nskb
, *next
;
1869 struct net
*net
= sock_net(sk
);
1877 /* Not currently probing/verifying,
1879 * have enough cwnd, and
1880 * not SACKing (the variable headers throw things off) */
1881 if (!icsk
->icsk_mtup
.enabled
||
1882 icsk
->icsk_mtup
.probe_size
||
1883 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1884 tp
->snd_cwnd
< 11 ||
1885 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1888 /* Use binary search for probe_size between tcp_mss_base,
1889 * and current mss_clamp. if (search_high - search_low)
1890 * smaller than a threshold, backoff from probing.
1892 mss_now
= tcp_current_mss(sk
);
1893 probe_size
= tcp_mtu_to_mss(sk
, (icsk
->icsk_mtup
.search_high
+
1894 icsk
->icsk_mtup
.search_low
) >> 1);
1895 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1896 interval
= icsk
->icsk_mtup
.search_high
- icsk
->icsk_mtup
.search_low
;
1897 /* When misfortune happens, we are reprobing actively,
1898 * and then reprobe timer has expired. We stick with current
1899 * probing process by not resetting search range to its orignal.
1901 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
) ||
1902 interval
< net
->ipv4
.sysctl_tcp_probe_threshold
) {
1903 /* Check whether enough time has elaplased for
1904 * another round of probing.
1906 tcp_mtu_check_reprobe(sk
);
1910 /* Have enough data in the send queue to probe? */
1911 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1914 if (tp
->snd_wnd
< size_needed
)
1916 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1919 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1920 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1921 if (!tcp_packets_in_flight(tp
))
1927 /* We're allowed to probe. Build it now. */
1928 if ((nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
)) == NULL
)
1930 sk
->sk_wmem_queued
+= nskb
->truesize
;
1931 sk_mem_charge(sk
, nskb
->truesize
);
1933 skb
= tcp_send_head(sk
);
1935 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1936 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1937 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
1938 TCP_SKB_CB(nskb
)->sacked
= 0;
1940 nskb
->ip_summed
= skb
->ip_summed
;
1942 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1945 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1946 copy
= min_t(int, skb
->len
, probe_size
- len
);
1947 if (nskb
->ip_summed
)
1948 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1950 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1951 skb_put(nskb
, copy
),
1954 if (skb
->len
<= copy
) {
1955 /* We've eaten all the data from this skb.
1957 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1958 tcp_unlink_write_queue(skb
, sk
);
1959 sk_wmem_free_skb(sk
, skb
);
1961 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
1962 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1963 if (!skb_shinfo(skb
)->nr_frags
) {
1964 skb_pull(skb
, copy
);
1965 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1966 skb
->csum
= csum_partial(skb
->data
,
1969 __pskb_trim_head(skb
, copy
);
1970 tcp_set_skb_tso_segs(sk
, skb
, mss_now
);
1972 TCP_SKB_CB(skb
)->seq
+= copy
;
1977 if (len
>= probe_size
)
1980 tcp_init_tso_segs(sk
, nskb
, nskb
->len
);
1982 /* We're ready to send. If this fails, the probe will
1983 * be resegmented into mss-sized pieces by tcp_write_xmit().
1985 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1986 /* Decrement cwnd here because we are sending
1987 * effectively two packets. */
1989 tcp_event_new_data_sent(sk
, nskb
);
1991 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
1992 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
1993 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
2001 /* This routine writes packets to the network. It advances the
2002 * send_head. This happens as incoming acks open up the remote
2005 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2006 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2007 * account rare use of URG, this is not a big flaw.
2009 * Send at most one packet when push_one > 0. Temporarily ignore
2010 * cwnd limit to force at most one packet out when push_one == 2.
2012 * Returns true, if no segments are in flight and we have queued segments,
2013 * but cannot send anything now because of SWS or another problem.
2015 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
2016 int push_one
, gfp_t gfp
)
2018 struct tcp_sock
*tp
= tcp_sk(sk
);
2019 struct sk_buff
*skb
;
2020 unsigned int tso_segs
, sent_pkts
;
2023 bool is_cwnd_limited
= false;
2029 /* Do MTU probing. */
2030 result
= tcp_mtu_probe(sk
);
2033 } else if (result
> 0) {
2038 max_segs
= tcp_tso_autosize(sk
, mss_now
);
2039 while ((skb
= tcp_send_head(sk
))) {
2042 tso_segs
= tcp_init_tso_segs(sk
, skb
, mss_now
);
2045 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
) {
2046 /* "skb_mstamp" is used as a start point for the retransmit timer */
2047 skb_mstamp_get(&skb
->skb_mstamp
);
2048 goto repair
; /* Skip network transmission */
2051 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2053 is_cwnd_limited
= true;
2055 /* Force out a loss probe pkt. */
2061 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
2064 if (tso_segs
== 1 || !max_segs
) {
2065 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2066 (tcp_skb_is_last(sk
, skb
) ?
2067 nonagle
: TCP_NAGLE_PUSH
))))
2071 tcp_tso_should_defer(sk
, skb
, &is_cwnd_limited
,
2077 if (tso_segs
> 1 && max_segs
&& !tcp_urg_mode(tp
))
2078 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2084 if (skb
->len
> limit
&&
2085 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
2088 /* TCP Small Queues :
2089 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2091 * - better RTT estimation and ACK scheduling
2094 * Alas, some drivers / subsystems require a fair amount
2095 * of queued bytes to ensure line rate.
2096 * One example is wifi aggregation (802.11 AMPDU)
2098 limit
= max(2 * skb
->truesize
, sk
->sk_pacing_rate
>> 10);
2099 limit
= min_t(u32
, limit
, sysctl_tcp_limit_output_bytes
);
2101 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
) {
2102 set_bit(TSQ_THROTTLED
, &tp
->tsq_flags
);
2103 /* It is possible TX completion already happened
2104 * before we set TSQ_THROTTLED, so we must
2105 * test again the condition.
2107 smp_mb__after_atomic();
2108 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
)
2112 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2116 /* Advance the send_head. This one is sent out.
2117 * This call will increment packets_out.
2119 tcp_event_new_data_sent(sk
, skb
);
2121 tcp_minshall_update(tp
, mss_now
, skb
);
2122 sent_pkts
+= tcp_skb_pcount(skb
);
2128 if (likely(sent_pkts
)) {
2129 if (tcp_in_cwnd_reduction(sk
))
2130 tp
->prr_out
+= sent_pkts
;
2132 /* Send one loss probe per tail loss episode. */
2134 tcp_schedule_loss_probe(sk
);
2135 tcp_cwnd_validate(sk
, is_cwnd_limited
);
2138 return (push_one
== 2) || (!tp
->packets_out
&& tcp_send_head(sk
));
2141 bool tcp_schedule_loss_probe(struct sock
*sk
)
2143 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2144 struct tcp_sock
*tp
= tcp_sk(sk
);
2145 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
2146 u32 rtt
= usecs_to_jiffies(tp
->srtt_us
>> 3);
2148 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
))
2150 /* No consecutive loss probes. */
2151 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
2155 /* Don't do any loss probe on a Fast Open connection before 3WHS
2158 if (sk
->sk_state
== TCP_SYN_RECV
)
2161 /* TLP is only scheduled when next timer event is RTO. */
2162 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
2165 /* Schedule a loss probe in 2*RTT for SACK capable connections
2166 * in Open state, that are either limited by cwnd or application.
2168 if (sysctl_tcp_early_retrans
< 3 || !tp
->srtt_us
|| !tp
->packets_out
||
2169 !tcp_is_sack(tp
) || inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
2172 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
2176 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2177 * for delayed ack when there's one outstanding packet.
2180 if (tp
->packets_out
== 1)
2181 timeout
= max_t(u32
, timeout
,
2182 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
2183 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
2185 /* If RTO is shorter, just schedule TLP in its place. */
2186 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
2187 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
2188 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
2189 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
2194 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2199 /* Thanks to skb fast clones, we can detect if a prior transmit of
2200 * a packet is still in a qdisc or driver queue.
2201 * In this case, there is very little point doing a retransmit !
2202 * Note: This is called from BH context only.
2204 static bool skb_still_in_host_queue(const struct sock
*sk
,
2205 const struct sk_buff
*skb
)
2207 if (unlikely(skb_fclone_busy(sk
, skb
))) {
2208 NET_INC_STATS_BH(sock_net(sk
),
2209 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
2215 /* When probe timeout (PTO) fires, send a new segment if one exists, else
2216 * retransmit the last segment.
2218 void tcp_send_loss_probe(struct sock
*sk
)
2220 struct tcp_sock
*tp
= tcp_sk(sk
);
2221 struct sk_buff
*skb
;
2223 int mss
= tcp_current_mss(sk
);
2226 if (tcp_send_head(sk
) != NULL
) {
2227 err
= tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2231 /* At most one outstanding TLP retransmission. */
2232 if (tp
->tlp_high_seq
)
2235 /* Retransmit last segment. */
2236 skb
= tcp_write_queue_tail(sk
);
2240 if (skb_still_in_host_queue(sk
, skb
))
2243 pcount
= tcp_skb_pcount(skb
);
2244 if (WARN_ON(!pcount
))
2247 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2248 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
,
2251 skb
= tcp_write_queue_tail(sk
);
2254 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2257 err
= __tcp_retransmit_skb(sk
, skb
);
2259 /* Record snd_nxt for loss detection. */
2261 tp
->tlp_high_seq
= tp
->snd_nxt
;
2264 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2265 inet_csk(sk
)->icsk_rto
,
2269 NET_INC_STATS_BH(sock_net(sk
),
2270 LINUX_MIB_TCPLOSSPROBES
);
2273 /* Push out any pending frames which were held back due to
2274 * TCP_CORK or attempt at coalescing tiny packets.
2275 * The socket must be locked by the caller.
2277 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2280 /* If we are closed, the bytes will have to remain here.
2281 * In time closedown will finish, we empty the write queue and
2282 * all will be happy.
2284 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2287 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2288 sk_gfp_atomic(sk
, GFP_ATOMIC
)))
2289 tcp_check_probe_timer(sk
);
2292 /* Send _single_ skb sitting at the send head. This function requires
2293 * true push pending frames to setup probe timer etc.
2295 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2297 struct sk_buff
*skb
= tcp_send_head(sk
);
2299 BUG_ON(!skb
|| skb
->len
< mss_now
);
2301 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2304 /* This function returns the amount that we can raise the
2305 * usable window based on the following constraints
2307 * 1. The window can never be shrunk once it is offered (RFC 793)
2308 * 2. We limit memory per socket
2311 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2312 * RECV.NEXT + RCV.WIN fixed until:
2313 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2315 * i.e. don't raise the right edge of the window until you can raise
2316 * it at least MSS bytes.
2318 * Unfortunately, the recommended algorithm breaks header prediction,
2319 * since header prediction assumes th->window stays fixed.
2321 * Strictly speaking, keeping th->window fixed violates the receiver
2322 * side SWS prevention criteria. The problem is that under this rule
2323 * a stream of single byte packets will cause the right side of the
2324 * window to always advance by a single byte.
2326 * Of course, if the sender implements sender side SWS prevention
2327 * then this will not be a problem.
2329 * BSD seems to make the following compromise:
2331 * If the free space is less than the 1/4 of the maximum
2332 * space available and the free space is less than 1/2 mss,
2333 * then set the window to 0.
2334 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2335 * Otherwise, just prevent the window from shrinking
2336 * and from being larger than the largest representable value.
2338 * This prevents incremental opening of the window in the regime
2339 * where TCP is limited by the speed of the reader side taking
2340 * data out of the TCP receive queue. It does nothing about
2341 * those cases where the window is constrained on the sender side
2342 * because the pipeline is full.
2344 * BSD also seems to "accidentally" limit itself to windows that are a
2345 * multiple of MSS, at least until the free space gets quite small.
2346 * This would appear to be a side effect of the mbuf implementation.
2347 * Combining these two algorithms results in the observed behavior
2348 * of having a fixed window size at almost all times.
2350 * Below we obtain similar behavior by forcing the offered window to
2351 * a multiple of the mss when it is feasible to do so.
2353 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2354 * Regular options like TIMESTAMP are taken into account.
2356 u32
__tcp_select_window(struct sock
*sk
)
2358 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2359 struct tcp_sock
*tp
= tcp_sk(sk
);
2360 /* MSS for the peer's data. Previous versions used mss_clamp
2361 * here. I don't know if the value based on our guesses
2362 * of peer's MSS is better for the performance. It's more correct
2363 * but may be worse for the performance because of rcv_mss
2364 * fluctuations. --SAW 1998/11/1
2366 int mss
= icsk
->icsk_ack
.rcv_mss
;
2367 int free_space
= tcp_space(sk
);
2368 int allowed_space
= tcp_full_space(sk
);
2369 int full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2372 if (mss
> full_space
)
2375 if (free_space
< (full_space
>> 1)) {
2376 icsk
->icsk_ack
.quick
= 0;
2378 if (sk_under_memory_pressure(sk
))
2379 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2382 /* free_space might become our new window, make sure we don't
2383 * increase it due to wscale.
2385 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2387 /* if free space is less than mss estimate, or is below 1/16th
2388 * of the maximum allowed, try to move to zero-window, else
2389 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2390 * new incoming data is dropped due to memory limits.
2391 * With large window, mss test triggers way too late in order
2392 * to announce zero window in time before rmem limit kicks in.
2394 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2398 if (free_space
> tp
->rcv_ssthresh
)
2399 free_space
= tp
->rcv_ssthresh
;
2401 /* Don't do rounding if we are using window scaling, since the
2402 * scaled window will not line up with the MSS boundary anyway.
2404 window
= tp
->rcv_wnd
;
2405 if (tp
->rx_opt
.rcv_wscale
) {
2406 window
= free_space
;
2408 /* Advertise enough space so that it won't get scaled away.
2409 * Import case: prevent zero window announcement if
2410 * 1<<rcv_wscale > mss.
2412 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2413 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2414 << tp
->rx_opt
.rcv_wscale
);
2416 /* Get the largest window that is a nice multiple of mss.
2417 * Window clamp already applied above.
2418 * If our current window offering is within 1 mss of the
2419 * free space we just keep it. This prevents the divide
2420 * and multiply from happening most of the time.
2421 * We also don't do any window rounding when the free space
2424 if (window
<= free_space
- mss
|| window
> free_space
)
2425 window
= (free_space
/ mss
) * mss
;
2426 else if (mss
== full_space
&&
2427 free_space
> window
+ (full_space
>> 1))
2428 window
= free_space
;
2434 /* Collapses two adjacent SKB's during retransmission. */
2435 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2437 struct tcp_sock
*tp
= tcp_sk(sk
);
2438 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2439 int skb_size
, next_skb_size
;
2441 skb_size
= skb
->len
;
2442 next_skb_size
= next_skb
->len
;
2444 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2446 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2448 tcp_unlink_write_queue(next_skb
, sk
);
2450 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
2453 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2454 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2456 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2457 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2459 /* Update sequence range on original skb. */
2460 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2462 /* Merge over control information. This moves PSH/FIN etc. over */
2463 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2465 /* All done, get rid of second SKB and account for it so
2466 * packet counting does not break.
2468 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2470 /* changed transmit queue under us so clear hints */
2471 tcp_clear_retrans_hints_partial(tp
);
2472 if (next_skb
== tp
->retransmit_skb_hint
)
2473 tp
->retransmit_skb_hint
= skb
;
2475 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2477 sk_wmem_free_skb(sk
, next_skb
);
2480 /* Check if coalescing SKBs is legal. */
2481 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2483 if (tcp_skb_pcount(skb
) > 1)
2485 /* TODO: SACK collapsing could be used to remove this condition */
2486 if (skb_shinfo(skb
)->nr_frags
!= 0)
2488 if (skb_cloned(skb
))
2490 if (skb
== tcp_send_head(sk
))
2492 /* Some heurestics for collapsing over SACK'd could be invented */
2493 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2499 /* Collapse packets in the retransmit queue to make to create
2500 * less packets on the wire. This is only done on retransmission.
2502 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2505 struct tcp_sock
*tp
= tcp_sk(sk
);
2506 struct sk_buff
*skb
= to
, *tmp
;
2509 if (!sysctl_tcp_retrans_collapse
)
2511 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2514 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2515 if (!tcp_can_collapse(sk
, skb
))
2527 /* Punt if not enough space exists in the first SKB for
2528 * the data in the second
2530 if (skb
->len
> skb_availroom(to
))
2533 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2536 tcp_collapse_retrans(sk
, to
);
2540 /* This retransmits one SKB. Policy decisions and retransmit queue
2541 * state updates are done by the caller. Returns non-zero if an
2542 * error occurred which prevented the send.
2544 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2546 struct tcp_sock
*tp
= tcp_sk(sk
);
2547 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2548 unsigned int cur_mss
;
2551 /* Inconslusive MTU probe */
2552 if (icsk
->icsk_mtup
.probe_size
) {
2553 icsk
->icsk_mtup
.probe_size
= 0;
2556 /* Do not sent more than we queued. 1/4 is reserved for possible
2557 * copying overhead: fragmentation, tunneling, mangling etc.
2559 if (atomic_read(&sk
->sk_wmem_alloc
) >
2560 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2563 if (skb_still_in_host_queue(sk
, skb
))
2566 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2567 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2569 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2573 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2574 return -EHOSTUNREACH
; /* Routing failure or similar. */
2576 cur_mss
= tcp_current_mss(sk
);
2578 /* If receiver has shrunk his window, and skb is out of
2579 * new window, do not retransmit it. The exception is the
2580 * case, when window is shrunk to zero. In this case
2581 * our retransmit serves as a zero window probe.
2583 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2584 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2587 if (skb
->len
> cur_mss
) {
2588 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
, GFP_ATOMIC
))
2589 return -ENOMEM
; /* We'll try again later. */
2591 int oldpcount
= tcp_skb_pcount(skb
);
2593 if (unlikely(oldpcount
> 1)) {
2594 if (skb_unclone(skb
, GFP_ATOMIC
))
2596 tcp_init_tso_segs(sk
, skb
, cur_mss
);
2597 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2601 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2603 /* Make a copy, if the first transmission SKB clone we made
2604 * is still in somebody's hands, else make a clone.
2607 /* make sure skb->data is aligned on arches that require it
2608 * and check if ack-trimming & collapsing extended the headroom
2609 * beyond what csum_start can cover.
2611 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2612 skb_headroom(skb
) >= 0xFFFF)) {
2613 struct sk_buff
*nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
,
2615 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2618 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2622 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
2623 /* Update global TCP statistics. */
2624 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2625 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2626 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
2627 tp
->total_retrans
++;
2632 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2634 struct tcp_sock
*tp
= tcp_sk(sk
);
2635 int err
= __tcp_retransmit_skb(sk
, skb
);
2638 #if FASTRETRANS_DEBUG > 0
2639 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2640 net_dbg_ratelimited("retrans_out leaked\n");
2643 if (!tp
->retrans_out
)
2644 tp
->lost_retrans_low
= tp
->snd_nxt
;
2645 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2646 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2648 /* Save stamp of the first retransmit. */
2649 if (!tp
->retrans_stamp
)
2650 tp
->retrans_stamp
= tcp_skb_timestamp(skb
);
2652 /* snd_nxt is stored to detect loss of retransmitted segment,
2653 * see tcp_input.c tcp_sacktag_write_queue().
2655 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2656 } else if (err
!= -EBUSY
) {
2657 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2660 if (tp
->undo_retrans
< 0)
2661 tp
->undo_retrans
= 0;
2662 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2666 /* Check if we forward retransmits are possible in the current
2667 * window/congestion state.
2669 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2671 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2672 const struct tcp_sock
*tp
= tcp_sk(sk
);
2674 /* Forward retransmissions are possible only during Recovery. */
2675 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2678 /* No forward retransmissions in Reno are possible. */
2679 if (tcp_is_reno(tp
))
2682 /* Yeah, we have to make difficult choice between forward transmission
2683 * and retransmission... Both ways have their merits...
2685 * For now we do not retransmit anything, while we have some new
2686 * segments to send. In the other cases, follow rule 3 for
2687 * NextSeg() specified in RFC3517.
2690 if (tcp_may_send_now(sk
))
2696 /* This gets called after a retransmit timeout, and the initially
2697 * retransmitted data is acknowledged. It tries to continue
2698 * resending the rest of the retransmit queue, until either
2699 * we've sent it all or the congestion window limit is reached.
2700 * If doing SACK, the first ACK which comes back for a timeout
2701 * based retransmit packet might feed us FACK information again.
2702 * If so, we use it to avoid unnecessarily retransmissions.
2704 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2706 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2707 struct tcp_sock
*tp
= tcp_sk(sk
);
2708 struct sk_buff
*skb
;
2709 struct sk_buff
*hole
= NULL
;
2712 int fwd_rexmitting
= 0;
2714 if (!tp
->packets_out
)
2718 tp
->retransmit_high
= tp
->snd_una
;
2720 if (tp
->retransmit_skb_hint
) {
2721 skb
= tp
->retransmit_skb_hint
;
2722 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2723 if (after(last_lost
, tp
->retransmit_high
))
2724 last_lost
= tp
->retransmit_high
;
2726 skb
= tcp_write_queue_head(sk
);
2727 last_lost
= tp
->snd_una
;
2730 tcp_for_write_queue_from(skb
, sk
) {
2731 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2733 if (skb
== tcp_send_head(sk
))
2735 /* we could do better than to assign each time */
2737 tp
->retransmit_skb_hint
= skb
;
2739 /* Assume this retransmit will generate
2740 * only one packet for congestion window
2741 * calculation purposes. This works because
2742 * tcp_retransmit_skb() will chop up the
2743 * packet to be MSS sized and all the
2744 * packet counting works out.
2746 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2749 if (fwd_rexmitting
) {
2751 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2753 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2755 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2756 tp
->retransmit_high
= last_lost
;
2757 if (!tcp_can_forward_retransmit(sk
))
2759 /* Backtrack if necessary to non-L'ed skb */
2767 } else if (!(sacked
& TCPCB_LOST
)) {
2768 if (hole
== NULL
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2773 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2774 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2775 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2777 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2780 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2783 if (tcp_retransmit_skb(sk
, skb
))
2786 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2788 if (tcp_in_cwnd_reduction(sk
))
2789 tp
->prr_out
+= tcp_skb_pcount(skb
);
2791 if (skb
== tcp_write_queue_head(sk
))
2792 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2793 inet_csk(sk
)->icsk_rto
,
2798 /* Send a fin. The caller locks the socket for us. This cannot be
2799 * allowed to fail queueing a FIN frame under any circumstances.
2801 void tcp_send_fin(struct sock
*sk
)
2803 struct tcp_sock
*tp
= tcp_sk(sk
);
2804 struct sk_buff
*skb
= tcp_write_queue_tail(sk
);
2807 /* Optimization, tack on the FIN if we have a queue of
2808 * unsent frames. But be careful about outgoing SACKS
2811 mss_now
= tcp_current_mss(sk
);
2813 if (tcp_send_head(sk
) != NULL
) {
2814 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_FIN
;
2815 TCP_SKB_CB(skb
)->end_seq
++;
2818 /* Socket is locked, keep trying until memory is available. */
2820 skb
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
);
2825 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2826 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2827 TCPHDR_ACK
| TCPHDR_FIN
);
2828 tcp_queue_skb(sk
, skb
);
2830 __tcp_push_pending_frames(sk
, mss_now
, TCP_NAGLE_OFF
);
2833 /* We get here when a process closes a file descriptor (either due to
2834 * an explicit close() or as a byproduct of exit()'ing) and there
2835 * was unread data in the receive queue. This behavior is recommended
2836 * by RFC 2525, section 2.17. -DaveM
2838 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2840 struct sk_buff
*skb
;
2842 /* NOTE: No TCP options attached and we never retransmit this. */
2843 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2845 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2849 /* Reserve space for headers and prepare control bits. */
2850 skb_reserve(skb
, MAX_TCP_HEADER
);
2851 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2852 TCPHDR_ACK
| TCPHDR_RST
);
2854 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2855 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2857 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2860 /* Send a crossed SYN-ACK during socket establishment.
2861 * WARNING: This routine must only be called when we have already sent
2862 * a SYN packet that crossed the incoming SYN that caused this routine
2863 * to get called. If this assumption fails then the initial rcv_wnd
2864 * and rcv_wscale values will not be correct.
2866 int tcp_send_synack(struct sock
*sk
)
2868 struct sk_buff
*skb
;
2870 skb
= tcp_write_queue_head(sk
);
2871 if (skb
== NULL
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2872 pr_debug("%s: wrong queue state\n", __func__
);
2875 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
2876 if (skb_cloned(skb
)) {
2877 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2880 tcp_unlink_write_queue(skb
, sk
);
2881 __skb_header_release(nskb
);
2882 __tcp_add_write_queue_head(sk
, nskb
);
2883 sk_wmem_free_skb(sk
, skb
);
2884 sk
->sk_wmem_queued
+= nskb
->truesize
;
2885 sk_mem_charge(sk
, nskb
->truesize
);
2889 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
2890 tcp_ecn_send_synack(sk
, skb
);
2892 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2896 * tcp_make_synack - Prepare a SYN-ACK.
2897 * sk: listener socket
2898 * dst: dst entry attached to the SYNACK
2899 * req: request_sock pointer
2901 * Allocate one skb and build a SYNACK packet.
2902 * @dst is consumed : Caller should not use it again.
2904 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2905 struct request_sock
*req
,
2906 struct tcp_fastopen_cookie
*foc
)
2908 struct tcp_out_options opts
;
2909 struct inet_request_sock
*ireq
= inet_rsk(req
);
2910 struct tcp_sock
*tp
= tcp_sk(sk
);
2912 struct sk_buff
*skb
;
2913 struct tcp_md5sig_key
*md5
= NULL
;
2914 int tcp_header_size
;
2917 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
, 1, GFP_ATOMIC
);
2918 if (unlikely(!skb
)) {
2922 /* Reserve space for headers. */
2923 skb_reserve(skb
, MAX_TCP_HEADER
);
2925 skb_dst_set(skb
, dst
);
2927 mss
= dst_metric_advmss(dst
);
2928 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2929 mss
= tp
->rx_opt
.user_mss
;
2931 memset(&opts
, 0, sizeof(opts
));
2932 #ifdef CONFIG_SYN_COOKIES
2933 if (unlikely(req
->cookie_ts
))
2934 skb
->skb_mstamp
.stamp_jiffies
= cookie_init_timestamp(req
);
2937 skb_mstamp_get(&skb
->skb_mstamp
);
2939 #ifdef CONFIG_TCP_MD5SIG
2941 md5
= tcp_rsk(req
)->af_specific
->md5_lookup(sk
, req
);
2943 tcp_header_size
= tcp_synack_options(sk
, req
, mss
, skb
, &opts
, md5
,
2946 skb_push(skb
, tcp_header_size
);
2947 skb_reset_transport_header(skb
);
2950 memset(th
, 0, sizeof(struct tcphdr
));
2953 tcp_ecn_make_synack(req
, th
, sk
);
2954 th
->source
= htons(ireq
->ir_num
);
2955 th
->dest
= ireq
->ir_rmt_port
;
2956 /* Setting of flags are superfluous here for callers (and ECE is
2957 * not even correctly set)
2959 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
2960 TCPHDR_SYN
| TCPHDR_ACK
);
2962 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
2963 /* XXX data is queued and acked as is. No buffer/window check */
2964 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
2966 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2967 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
2968 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
2969 th
->doff
= (tcp_header_size
>> 2);
2970 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_OUTSEGS
);
2972 #ifdef CONFIG_TCP_MD5SIG
2973 /* Okay, we have all we need - do the md5 hash if needed */
2975 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
2976 md5
, req_to_sk(req
), skb
);
2982 EXPORT_SYMBOL(tcp_make_synack
);
2984 static void tcp_ca_dst_init(struct sock
*sk
, const struct dst_entry
*dst
)
2986 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2987 const struct tcp_congestion_ops
*ca
;
2988 u32 ca_key
= dst_metric(dst
, RTAX_CC_ALGO
);
2990 if (ca_key
== TCP_CA_UNSPEC
)
2994 ca
= tcp_ca_find_key(ca_key
);
2995 if (likely(ca
&& try_module_get(ca
->owner
))) {
2996 module_put(icsk
->icsk_ca_ops
->owner
);
2997 icsk
->icsk_ca_dst_locked
= tcp_ca_dst_locked(dst
);
2998 icsk
->icsk_ca_ops
= ca
;
3003 /* Do all connect socket setups that can be done AF independent. */
3004 static void tcp_connect_init(struct sock
*sk
)
3006 const struct dst_entry
*dst
= __sk_dst_get(sk
);
3007 struct tcp_sock
*tp
= tcp_sk(sk
);
3010 /* We'll fix this up when we get a response from the other end.
3011 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3013 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
3014 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
3016 #ifdef CONFIG_TCP_MD5SIG
3017 if (tp
->af_specific
->md5_lookup(sk
, sk
) != NULL
)
3018 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
3021 /* If user gave his TCP_MAXSEG, record it to clamp */
3022 if (tp
->rx_opt
.user_mss
)
3023 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3026 tcp_sync_mss(sk
, dst_mtu(dst
));
3028 tcp_ca_dst_init(sk
, dst
);
3030 if (!tp
->window_clamp
)
3031 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
3032 tp
->advmss
= dst_metric_advmss(dst
);
3033 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
3034 tp
->advmss
= tp
->rx_opt
.user_mss
;
3036 tcp_initialize_rcv_mss(sk
);
3038 /* limit the window selection if the user enforce a smaller rx buffer */
3039 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
3040 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
3041 tp
->window_clamp
= tcp_full_space(sk
);
3043 tcp_select_initial_window(tcp_full_space(sk
),
3044 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
3047 sysctl_tcp_window_scaling
,
3049 dst_metric(dst
, RTAX_INITRWND
));
3051 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
3052 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
3055 sock_reset_flag(sk
, SOCK_DONE
);
3058 tp
->snd_una
= tp
->write_seq
;
3059 tp
->snd_sml
= tp
->write_seq
;
3060 tp
->snd_up
= tp
->write_seq
;
3061 tp
->snd_nxt
= tp
->write_seq
;
3063 if (likely(!tp
->repair
))
3066 tp
->rcv_tstamp
= tcp_time_stamp
;
3067 tp
->rcv_wup
= tp
->rcv_nxt
;
3068 tp
->copied_seq
= tp
->rcv_nxt
;
3070 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
3071 inet_csk(sk
)->icsk_retransmits
= 0;
3072 tcp_clear_retrans(tp
);
3075 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
3077 struct tcp_sock
*tp
= tcp_sk(sk
);
3078 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
3080 tcb
->end_seq
+= skb
->len
;
3081 __skb_header_release(skb
);
3082 __tcp_add_write_queue_tail(sk
, skb
);
3083 sk
->sk_wmem_queued
+= skb
->truesize
;
3084 sk_mem_charge(sk
, skb
->truesize
);
3085 tp
->write_seq
= tcb
->end_seq
;
3086 tp
->packets_out
+= tcp_skb_pcount(skb
);
3089 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3090 * queue a data-only packet after the regular SYN, such that regular SYNs
3091 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3092 * only the SYN sequence, the data are retransmitted in the first ACK.
3093 * If cookie is not cached or other error occurs, falls back to send a
3094 * regular SYN with Fast Open cookie request option.
3096 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
3098 struct tcp_sock
*tp
= tcp_sk(sk
);
3099 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
3100 int syn_loss
= 0, space
, err
= 0, copied
;
3101 unsigned long last_syn_loss
= 0;
3102 struct sk_buff
*syn_data
;
3104 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
3105 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
3106 &syn_loss
, &last_syn_loss
);
3107 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3109 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
3110 fo
->cookie
.len
= -1;
3114 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
3115 fo
->cookie
.len
= -1;
3116 else if (fo
->cookie
.len
<= 0)
3119 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3120 * user-MSS. Reserve maximum option space for middleboxes that add
3121 * private TCP options. The cost is reduced data space in SYN :(
3123 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
3124 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3125 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
3126 MAX_TCP_OPTION_SPACE
;
3128 space
= min_t(size_t, space
, fo
->size
);
3130 /* limit to order-0 allocations */
3131 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
3133 syn_data
= sk_stream_alloc_skb(sk
, space
, sk
->sk_allocation
);
3136 syn_data
->ip_summed
= CHECKSUM_PARTIAL
;
3137 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
3138 copied
= copy_from_iter(skb_put(syn_data
, space
), space
,
3139 &fo
->data
->msg_iter
);
3140 if (unlikely(!copied
)) {
3141 kfree_skb(syn_data
);
3144 if (copied
!= space
) {
3145 skb_trim(syn_data
, copied
);
3149 /* No more data pending in inet_wait_for_connect() */
3150 if (space
== fo
->size
)
3154 tcp_connect_queue_skb(sk
, syn_data
);
3156 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
3158 syn
->skb_mstamp
= syn_data
->skb_mstamp
;
3160 /* Now full SYN+DATA was cloned and sent (or not),
3161 * remove the SYN from the original skb (syn_data)
3162 * we keep in write queue in case of a retransmit, as we
3163 * also have the SYN packet (with no data) in the same queue.
3165 TCP_SKB_CB(syn_data
)->seq
++;
3166 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
3168 tp
->syn_data
= (fo
->copied
> 0);
3169 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3174 /* Send a regular SYN with Fast Open cookie request option */
3175 if (fo
->cookie
.len
> 0)
3177 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3179 tp
->syn_fastopen
= 0;
3181 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3185 /* Build a SYN and send it off. */
3186 int tcp_connect(struct sock
*sk
)
3188 struct tcp_sock
*tp
= tcp_sk(sk
);
3189 struct sk_buff
*buff
;
3192 tcp_connect_init(sk
);
3194 if (unlikely(tp
->repair
)) {
3195 tcp_finish_connect(sk
, NULL
);
3199 buff
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
);
3200 if (unlikely(!buff
))
3203 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3204 tp
->retrans_stamp
= tcp_time_stamp
;
3205 tcp_connect_queue_skb(sk
, buff
);
3206 tcp_ecn_send_syn(sk
, buff
);
3208 /* Send off SYN; include data in Fast Open. */
3209 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3210 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3211 if (err
== -ECONNREFUSED
)
3214 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3215 * in order to make this packet get counted in tcpOutSegs.
3217 tp
->snd_nxt
= tp
->write_seq
;
3218 tp
->pushed_seq
= tp
->write_seq
;
3219 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3221 /* Timer for repeating the SYN until an answer. */
3222 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3223 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3226 EXPORT_SYMBOL(tcp_connect
);
3228 /* Send out a delayed ack, the caller does the policy checking
3229 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3232 void tcp_send_delayed_ack(struct sock
*sk
)
3234 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3235 int ato
= icsk
->icsk_ack
.ato
;
3236 unsigned long timeout
;
3238 tcp_ca_event(sk
, CA_EVENT_DELAYED_ACK
);
3240 if (ato
> TCP_DELACK_MIN
) {
3241 const struct tcp_sock
*tp
= tcp_sk(sk
);
3242 int max_ato
= HZ
/ 2;
3244 if (icsk
->icsk_ack
.pingpong
||
3245 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3246 max_ato
= TCP_DELACK_MAX
;
3248 /* Slow path, intersegment interval is "high". */
3250 /* If some rtt estimate is known, use it to bound delayed ack.
3251 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3255 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3262 ato
= min(ato
, max_ato
);
3265 /* Stay within the limit we were given */
3266 timeout
= jiffies
+ ato
;
3268 /* Use new timeout only if there wasn't a older one earlier. */
3269 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3270 /* If delack timer was blocked or is about to expire,
3273 if (icsk
->icsk_ack
.blocked
||
3274 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3279 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3280 timeout
= icsk
->icsk_ack
.timeout
;
3282 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3283 icsk
->icsk_ack
.timeout
= timeout
;
3284 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3287 /* This routine sends an ack and also updates the window. */
3288 void tcp_send_ack(struct sock
*sk
)
3290 struct sk_buff
*buff
;
3292 /* If we have been reset, we may not send again. */
3293 if (sk
->sk_state
== TCP_CLOSE
)
3296 tcp_ca_event(sk
, CA_EVENT_NON_DELAYED_ACK
);
3298 /* We are not putting this on the write queue, so
3299 * tcp_transmit_skb() will set the ownership to this
3302 buff
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3304 inet_csk_schedule_ack(sk
);
3305 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3306 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3307 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3311 /* Reserve space for headers and prepare control bits. */
3312 skb_reserve(buff
, MAX_TCP_HEADER
);
3313 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3315 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3317 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3318 * We also avoid tcp_wfree() overhead (cache line miss accessing
3319 * tp->tsq_flags) by using regular sock_wfree()
3321 skb_set_tcp_pure_ack(buff
);
3323 /* Send it off, this clears delayed acks for us. */
3324 skb_mstamp_get(&buff
->skb_mstamp
);
3325 tcp_transmit_skb(sk
, buff
, 0, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3327 EXPORT_SYMBOL_GPL(tcp_send_ack
);
3329 /* This routine sends a packet with an out of date sequence
3330 * number. It assumes the other end will try to ack it.
3332 * Question: what should we make while urgent mode?
3333 * 4.4BSD forces sending single byte of data. We cannot send
3334 * out of window data, because we have SND.NXT==SND.MAX...
3336 * Current solution: to send TWO zero-length segments in urgent mode:
3337 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3338 * out-of-date with SND.UNA-1 to probe window.
3340 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
)
3342 struct tcp_sock
*tp
= tcp_sk(sk
);
3343 struct sk_buff
*skb
;
3345 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3346 skb
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3350 /* Reserve space for headers and set control bits. */
3351 skb_reserve(skb
, MAX_TCP_HEADER
);
3352 /* Use a previous sequence. This should cause the other
3353 * end to send an ack. Don't queue or clone SKB, just
3356 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3357 skb_mstamp_get(&skb
->skb_mstamp
);
3358 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
3361 void tcp_send_window_probe(struct sock
*sk
)
3363 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3364 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3365 tcp_xmit_probe_skb(sk
, 0);
3369 /* Initiate keepalive or window probe from timer. */
3370 int tcp_write_wakeup(struct sock
*sk
)
3372 struct tcp_sock
*tp
= tcp_sk(sk
);
3373 struct sk_buff
*skb
;
3375 if (sk
->sk_state
== TCP_CLOSE
)
3378 if ((skb
= tcp_send_head(sk
)) != NULL
&&
3379 before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3381 unsigned int mss
= tcp_current_mss(sk
);
3382 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3384 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3385 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3387 /* We are probing the opening of a window
3388 * but the window size is != 0
3389 * must have been a result SWS avoidance ( sender )
3391 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3393 seg_size
= min(seg_size
, mss
);
3394 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3395 if (tcp_fragment(sk
, skb
, seg_size
, mss
, GFP_ATOMIC
))
3397 } else if (!tcp_skb_pcount(skb
))
3398 tcp_set_skb_tso_segs(sk
, skb
, mss
);
3400 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3401 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3403 tcp_event_new_data_sent(sk
, skb
);
3406 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3407 tcp_xmit_probe_skb(sk
, 1);
3408 return tcp_xmit_probe_skb(sk
, 0);
3412 /* A window probe timeout has occurred. If window is not closed send
3413 * a partial packet else a zero probe.
3415 void tcp_send_probe0(struct sock
*sk
)
3417 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3418 struct tcp_sock
*tp
= tcp_sk(sk
);
3419 unsigned long probe_max
;
3422 err
= tcp_write_wakeup(sk
);
3424 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3425 /* Cancel probe timer, if it is not required. */
3426 icsk
->icsk_probes_out
= 0;
3427 icsk
->icsk_backoff
= 0;
3432 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
3433 icsk
->icsk_backoff
++;
3434 icsk
->icsk_probes_out
++;
3435 probe_max
= TCP_RTO_MAX
;
3437 /* If packet was not sent due to local congestion,
3438 * do not backoff and do not remember icsk_probes_out.
3439 * Let local senders to fight for local resources.
3441 * Use accumulated backoff yet.
3443 if (!icsk
->icsk_probes_out
)
3444 icsk
->icsk_probes_out
= 1;
3445 probe_max
= TCP_RESOURCE_PROBE_INTERVAL
;
3447 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3448 inet_csk_rto_backoff(icsk
, probe_max
),
3452 int tcp_rtx_synack(struct sock
*sk
, struct request_sock
*req
)
3454 const struct tcp_request_sock_ops
*af_ops
= tcp_rsk(req
)->af_specific
;
3458 res
= af_ops
->send_synack(sk
, NULL
, &fl
, req
, 0, NULL
);
3460 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
3461 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
3465 EXPORT_SYMBOL(tcp_rtx_synack
);