2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
53 /* Default TSQ limit of four TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly
= 262144;
56 /* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
60 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
65 unsigned int sysctl_tcp_notsent_lowat __read_mostly
= UINT_MAX
;
66 EXPORT_SYMBOL(sysctl_tcp_notsent_lowat
);
68 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
69 int push_one
, gfp_t gfp
);
71 /* Account for new data that has been sent to the network. */
72 static void tcp_event_new_data_sent(struct sock
*sk
, const struct sk_buff
*skb
)
74 struct inet_connection_sock
*icsk
= inet_csk(sk
);
75 struct tcp_sock
*tp
= tcp_sk(sk
);
76 unsigned int prior_packets
= tp
->packets_out
;
78 tcp_advance_send_head(sk
, skb
);
79 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
81 tp
->packets_out
+= tcp_skb_pcount(skb
);
82 if (!prior_packets
|| icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
83 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
87 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
,
91 /* SND.NXT, if window was not shrunk.
92 * If window has been shrunk, what should we make? It is not clear at all.
93 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
94 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
95 * invalid. OK, let's make this for now:
97 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
99 const struct tcp_sock
*tp
= tcp_sk(sk
);
101 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
104 return tcp_wnd_end(tp
);
107 /* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
121 static __u16
tcp_advertise_mss(struct sock
*sk
)
123 struct tcp_sock
*tp
= tcp_sk(sk
);
124 const struct dst_entry
*dst
= __sk_dst_get(sk
);
125 int mss
= tp
->advmss
;
128 unsigned int metric
= dst_metric_advmss(dst
);
139 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism.
142 void tcp_cwnd_restart(struct sock
*sk
, s32 delta
)
144 struct tcp_sock
*tp
= tcp_sk(sk
);
145 u32 restart_cwnd
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
146 u32 cwnd
= tp
->snd_cwnd
;
148 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
150 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
151 restart_cwnd
= min(restart_cwnd
, cwnd
);
153 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
155 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
156 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
157 tp
->snd_cwnd_used
= 0;
160 /* Congestion state accounting after a packet has been sent. */
161 static void tcp_event_data_sent(struct tcp_sock
*tp
,
164 struct inet_connection_sock
*icsk
= inet_csk(sk
);
165 const u32 now
= tcp_time_stamp
;
167 if (tcp_packets_in_flight(tp
) == 0)
168 tcp_ca_event(sk
, CA_EVENT_TX_START
);
172 /* If it is a reply for ato after last received
173 * packet, enter pingpong mode.
175 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
176 icsk
->icsk_ack
.pingpong
= 1;
179 /* Account for an ACK we sent. */
180 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
182 tcp_dec_quickack_mode(sk
, pkts
);
183 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
187 u32
tcp_default_init_rwnd(u32 mss
)
189 /* Initial receive window should be twice of TCP_INIT_CWND to
190 * enable proper sending of new unsent data during fast recovery
191 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
192 * limit when mss is larger than 1460.
194 u32 init_rwnd
= TCP_INIT_CWND
* 2;
197 init_rwnd
= max((1460 * init_rwnd
) / mss
, 2U);
201 /* Determine a window scaling and initial window to offer.
202 * Based on the assumption that the given amount of space
203 * will be offered. Store the results in the tp structure.
204 * NOTE: for smooth operation initial space offering should
205 * be a multiple of mss if possible. We assume here that mss >= 1.
206 * This MUST be enforced by all callers.
208 void tcp_select_initial_window(int __space
, __u32 mss
,
209 __u32
*rcv_wnd
, __u32
*window_clamp
,
210 int wscale_ok
, __u8
*rcv_wscale
,
213 unsigned int space
= (__space
< 0 ? 0 : __space
);
215 /* If no clamp set the clamp to the max possible scaled window */
216 if (*window_clamp
== 0)
217 (*window_clamp
) = (65535 << 14);
218 space
= min(*window_clamp
, space
);
220 /* Quantize space offering to a multiple of mss if possible. */
222 space
= (space
/ mss
) * mss
;
224 /* NOTE: offering an initial window larger than 32767
225 * will break some buggy TCP stacks. If the admin tells us
226 * it is likely we could be speaking with such a buggy stack
227 * we will truncate our initial window offering to 32K-1
228 * unless the remote has sent us a window scaling option,
229 * which we interpret as a sign the remote TCP is not
230 * misinterpreting the window field as a signed quantity.
232 if (sysctl_tcp_workaround_signed_windows
)
233 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
239 /* Set window scaling on max possible window
240 * See RFC1323 for an explanation of the limit to 14
242 space
= max_t(u32
, sysctl_tcp_rmem
[2], sysctl_rmem_max
);
243 space
= min_t(u32
, space
, *window_clamp
);
244 while (space
> 65535 && (*rcv_wscale
) < 14) {
250 if (mss
> (1 << *rcv_wscale
)) {
251 if (!init_rcv_wnd
) /* Use default unless specified otherwise */
252 init_rcv_wnd
= tcp_default_init_rwnd(mss
);
253 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
256 /* Set the clamp no higher than max representable value */
257 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
259 EXPORT_SYMBOL(tcp_select_initial_window
);
261 /* Chose a new window to advertise, update state in tcp_sock for the
262 * socket, and return result with RFC1323 scaling applied. The return
263 * value can be stuffed directly into th->window for an outgoing
266 static u16
tcp_select_window(struct sock
*sk
)
268 struct tcp_sock
*tp
= tcp_sk(sk
);
269 u32 old_win
= tp
->rcv_wnd
;
270 u32 cur_win
= tcp_receive_window(tp
);
271 u32 new_win
= __tcp_select_window(sk
);
273 /* Never shrink the offered window */
274 if (new_win
< cur_win
) {
275 /* Danger Will Robinson!
276 * Don't update rcv_wup/rcv_wnd here or else
277 * we will not be able to advertise a zero
278 * window in time. --DaveM
280 * Relax Will Robinson.
283 NET_INC_STATS(sock_net(sk
),
284 LINUX_MIB_TCPWANTZEROWINDOWADV
);
285 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
287 tp
->rcv_wnd
= new_win
;
288 tp
->rcv_wup
= tp
->rcv_nxt
;
290 /* Make sure we do not exceed the maximum possible
293 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
294 new_win
= min(new_win
, MAX_TCP_WINDOW
);
296 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
298 /* RFC1323 scaling applied */
299 new_win
>>= tp
->rx_opt
.rcv_wscale
;
301 /* If we advertise zero window, disable fast path. */
305 NET_INC_STATS(sock_net(sk
),
306 LINUX_MIB_TCPTOZEROWINDOWADV
);
307 } else if (old_win
== 0) {
308 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFROMZEROWINDOWADV
);
314 /* Packet ECN state for a SYN-ACK */
315 static void tcp_ecn_send_synack(struct sock
*sk
, struct sk_buff
*skb
)
317 const struct tcp_sock
*tp
= tcp_sk(sk
);
319 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
320 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
321 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
322 else if (tcp_ca_needs_ecn(sk
))
326 /* Packet ECN state for a SYN. */
327 static void tcp_ecn_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
329 struct tcp_sock
*tp
= tcp_sk(sk
);
330 bool use_ecn
= sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1 ||
331 tcp_ca_needs_ecn(sk
);
334 const struct dst_entry
*dst
= __sk_dst_get(sk
);
336 if (dst
&& dst_feature(dst
, RTAX_FEATURE_ECN
))
343 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
344 tp
->ecn_flags
= TCP_ECN_OK
;
345 if (tcp_ca_needs_ecn(sk
))
350 static void tcp_ecn_clear_syn(struct sock
*sk
, struct sk_buff
*skb
)
352 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn_fallback
)
353 /* tp->ecn_flags are cleared at a later point in time when
354 * SYN ACK is ultimatively being received.
356 TCP_SKB_CB(skb
)->tcp_flags
&= ~(TCPHDR_ECE
| TCPHDR_CWR
);
360 tcp_ecn_make_synack(const struct request_sock
*req
, struct tcphdr
*th
,
363 if (inet_rsk(req
)->ecn_ok
) {
365 if (tcp_ca_needs_ecn(sk
))
370 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
373 static void tcp_ecn_send(struct sock
*sk
, struct sk_buff
*skb
,
376 struct tcp_sock
*tp
= tcp_sk(sk
);
378 if (tp
->ecn_flags
& TCP_ECN_OK
) {
379 /* Not-retransmitted data segment: set ECT and inject CWR. */
380 if (skb
->len
!= tcp_header_len
&&
381 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
383 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
384 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
385 tcp_hdr(skb
)->cwr
= 1;
386 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
388 } else if (!tcp_ca_needs_ecn(sk
)) {
389 /* ACK or retransmitted segment: clear ECT|CE */
390 INET_ECN_dontxmit(sk
);
392 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
393 tcp_hdr(skb
)->ece
= 1;
397 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
398 * auto increment end seqno.
400 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
402 skb
->ip_summed
= CHECKSUM_PARTIAL
;
405 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
406 TCP_SKB_CB(skb
)->sacked
= 0;
408 tcp_skb_pcount_set(skb
, 1);
410 TCP_SKB_CB(skb
)->seq
= seq
;
411 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
413 TCP_SKB_CB(skb
)->end_seq
= seq
;
416 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
418 return tp
->snd_una
!= tp
->snd_up
;
421 #define OPTION_SACK_ADVERTISE (1 << 0)
422 #define OPTION_TS (1 << 1)
423 #define OPTION_MD5 (1 << 2)
424 #define OPTION_WSCALE (1 << 3)
425 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
427 struct tcp_out_options
{
428 u16 options
; /* bit field of OPTION_* */
429 u16 mss
; /* 0 to disable */
430 u8 ws
; /* window scale, 0 to disable */
431 u8 num_sack_blocks
; /* number of SACK blocks to include */
432 u8 hash_size
; /* bytes in hash_location */
433 __u8
*hash_location
; /* temporary pointer, overloaded */
434 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
435 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
438 /* Write previously computed TCP options to the packet.
440 * Beware: Something in the Internet is very sensitive to the ordering of
441 * TCP options, we learned this through the hard way, so be careful here.
442 * Luckily we can at least blame others for their non-compliance but from
443 * inter-operability perspective it seems that we're somewhat stuck with
444 * the ordering which we have been using if we want to keep working with
445 * those broken things (not that it currently hurts anybody as there isn't
446 * particular reason why the ordering would need to be changed).
448 * At least SACK_PERM as the first option is known to lead to a disaster
449 * (but it may well be that other scenarios fail similarly).
451 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
452 struct tcp_out_options
*opts
)
454 u16 options
= opts
->options
; /* mungable copy */
456 if (unlikely(OPTION_MD5
& options
)) {
457 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
458 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
459 /* overload cookie hash location */
460 opts
->hash_location
= (__u8
*)ptr
;
464 if (unlikely(opts
->mss
)) {
465 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
466 (TCPOLEN_MSS
<< 16) |
470 if (likely(OPTION_TS
& options
)) {
471 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
472 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
473 (TCPOLEN_SACK_PERM
<< 16) |
474 (TCPOPT_TIMESTAMP
<< 8) |
476 options
&= ~OPTION_SACK_ADVERTISE
;
478 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
480 (TCPOPT_TIMESTAMP
<< 8) |
483 *ptr
++ = htonl(opts
->tsval
);
484 *ptr
++ = htonl(opts
->tsecr
);
487 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
488 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
490 (TCPOPT_SACK_PERM
<< 8) |
494 if (unlikely(OPTION_WSCALE
& options
)) {
495 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
496 (TCPOPT_WINDOW
<< 16) |
497 (TCPOLEN_WINDOW
<< 8) |
501 if (unlikely(opts
->num_sack_blocks
)) {
502 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
503 tp
->duplicate_sack
: tp
->selective_acks
;
506 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
509 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
510 TCPOLEN_SACK_PERBLOCK
)));
512 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
514 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
515 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
518 tp
->rx_opt
.dsack
= 0;
521 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
522 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
524 u32 len
; /* Fast Open option length */
527 len
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
528 *ptr
= htonl((TCPOPT_EXP
<< 24) | (len
<< 16) |
529 TCPOPT_FASTOPEN_MAGIC
);
530 p
+= TCPOLEN_EXP_FASTOPEN_BASE
;
532 len
= TCPOLEN_FASTOPEN_BASE
+ foc
->len
;
533 *p
++ = TCPOPT_FASTOPEN
;
537 memcpy(p
, foc
->val
, foc
->len
);
538 if ((len
& 3) == 2) {
539 p
[foc
->len
] = TCPOPT_NOP
;
540 p
[foc
->len
+ 1] = TCPOPT_NOP
;
542 ptr
+= (len
+ 3) >> 2;
546 /* Compute TCP options for SYN packets. This is not the final
547 * network wire format yet.
549 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
550 struct tcp_out_options
*opts
,
551 struct tcp_md5sig_key
**md5
)
553 struct tcp_sock
*tp
= tcp_sk(sk
);
554 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
555 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
557 #ifdef CONFIG_TCP_MD5SIG
558 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
560 opts
->options
|= OPTION_MD5
;
561 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
567 /* We always get an MSS option. The option bytes which will be seen in
568 * normal data packets should timestamps be used, must be in the MSS
569 * advertised. But we subtract them from tp->mss_cache so that
570 * calculations in tcp_sendmsg are simpler etc. So account for this
571 * fact here if necessary. If we don't do this correctly, as a
572 * receiver we won't recognize data packets as being full sized when we
573 * should, and thus we won't abide by the delayed ACK rules correctly.
574 * SACKs don't matter, we never delay an ACK when we have any of those
576 opts
->mss
= tcp_advertise_mss(sk
);
577 remaining
-= TCPOLEN_MSS_ALIGNED
;
579 if (likely(sysctl_tcp_timestamps
&& !*md5
)) {
580 opts
->options
|= OPTION_TS
;
581 opts
->tsval
= tcp_skb_timestamp(skb
) + tp
->tsoffset
;
582 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
583 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
585 if (likely(sysctl_tcp_window_scaling
)) {
586 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
587 opts
->options
|= OPTION_WSCALE
;
588 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
590 if (likely(sysctl_tcp_sack
)) {
591 opts
->options
|= OPTION_SACK_ADVERTISE
;
592 if (unlikely(!(OPTION_TS
& opts
->options
)))
593 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
596 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
597 u32 need
= fastopen
->cookie
.len
;
599 need
+= fastopen
->cookie
.exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
600 TCPOLEN_FASTOPEN_BASE
;
601 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
602 if (remaining
>= need
) {
603 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
604 opts
->fastopen_cookie
= &fastopen
->cookie
;
606 tp
->syn_fastopen
= 1;
607 tp
->syn_fastopen_exp
= fastopen
->cookie
.exp
? 1 : 0;
611 return MAX_TCP_OPTION_SPACE
- remaining
;
614 /* Set up TCP options for SYN-ACKs. */
615 static unsigned int tcp_synack_options(struct sock
*sk
,
616 struct request_sock
*req
,
617 unsigned int mss
, struct sk_buff
*skb
,
618 struct tcp_out_options
*opts
,
619 const struct tcp_md5sig_key
*md5
,
620 struct tcp_fastopen_cookie
*foc
)
622 struct inet_request_sock
*ireq
= inet_rsk(req
);
623 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
625 #ifdef CONFIG_TCP_MD5SIG
627 opts
->options
|= OPTION_MD5
;
628 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
630 /* We can't fit any SACK blocks in a packet with MD5 + TS
631 * options. There was discussion about disabling SACK
632 * rather than TS in order to fit in better with old,
633 * buggy kernels, but that was deemed to be unnecessary.
635 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
639 /* We always send an MSS option. */
641 remaining
-= TCPOLEN_MSS_ALIGNED
;
643 if (likely(ireq
->wscale_ok
)) {
644 opts
->ws
= ireq
->rcv_wscale
;
645 opts
->options
|= OPTION_WSCALE
;
646 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
648 if (likely(ireq
->tstamp_ok
)) {
649 opts
->options
|= OPTION_TS
;
650 opts
->tsval
= tcp_skb_timestamp(skb
);
651 opts
->tsecr
= req
->ts_recent
;
652 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
654 if (likely(ireq
->sack_ok
)) {
655 opts
->options
|= OPTION_SACK_ADVERTISE
;
656 if (unlikely(!ireq
->tstamp_ok
))
657 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
659 if (foc
!= NULL
&& foc
->len
>= 0) {
662 need
+= foc
->exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
663 TCPOLEN_FASTOPEN_BASE
;
664 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
665 if (remaining
>= need
) {
666 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
667 opts
->fastopen_cookie
= foc
;
672 return MAX_TCP_OPTION_SPACE
- remaining
;
675 /* Compute TCP options for ESTABLISHED sockets. This is not the
676 * final wire format yet.
678 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
679 struct tcp_out_options
*opts
,
680 struct tcp_md5sig_key
**md5
)
682 struct tcp_sock
*tp
= tcp_sk(sk
);
683 unsigned int size
= 0;
684 unsigned int eff_sacks
;
688 #ifdef CONFIG_TCP_MD5SIG
689 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
690 if (unlikely(*md5
)) {
691 opts
->options
|= OPTION_MD5
;
692 size
+= TCPOLEN_MD5SIG_ALIGNED
;
698 if (likely(tp
->rx_opt
.tstamp_ok
)) {
699 opts
->options
|= OPTION_TS
;
700 opts
->tsval
= skb
? tcp_skb_timestamp(skb
) + tp
->tsoffset
: 0;
701 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
702 size
+= TCPOLEN_TSTAMP_ALIGNED
;
705 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
706 if (unlikely(eff_sacks
)) {
707 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
708 opts
->num_sack_blocks
=
709 min_t(unsigned int, eff_sacks
,
710 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
711 TCPOLEN_SACK_PERBLOCK
);
712 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
713 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
720 /* TCP SMALL QUEUES (TSQ)
722 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
723 * to reduce RTT and bufferbloat.
724 * We do this using a special skb destructor (tcp_wfree).
726 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
727 * needs to be reallocated in a driver.
728 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
730 * Since transmit from skb destructor is forbidden, we use a tasklet
731 * to process all sockets that eventually need to send more skbs.
732 * We use one tasklet per cpu, with its own queue of sockets.
735 struct tasklet_struct tasklet
;
736 struct list_head head
; /* queue of tcp sockets */
738 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
740 static void tcp_tsq_handler(struct sock
*sk
)
742 if ((1 << sk
->sk_state
) &
743 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
744 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
))
745 tcp_write_xmit(sk
, tcp_current_mss(sk
), tcp_sk(sk
)->nonagle
,
749 * One tasklet per cpu tries to send more skbs.
750 * We run in tasklet context but need to disable irqs when
751 * transferring tsq->head because tcp_wfree() might
752 * interrupt us (non NAPI drivers)
754 static void tcp_tasklet_func(unsigned long data
)
756 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
759 struct list_head
*q
, *n
;
763 local_irq_save(flags
);
764 list_splice_init(&tsq
->head
, &list
);
765 local_irq_restore(flags
);
767 list_for_each_safe(q
, n
, &list
) {
768 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
769 list_del(&tp
->tsq_node
);
771 sk
= (struct sock
*)tp
;
774 if (!sock_owned_by_user(sk
)) {
777 /* defer the work to tcp_release_cb() */
778 set_bit(TCP_TSQ_DEFERRED
, &tp
->tsq_flags
);
782 clear_bit(TSQ_QUEUED
, &tp
->tsq_flags
);
787 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
788 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
789 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
790 (1UL << TCP_MTU_REDUCED_DEFERRED))
792 * tcp_release_cb - tcp release_sock() callback
795 * called from release_sock() to perform protocol dependent
796 * actions before socket release.
798 void tcp_release_cb(struct sock
*sk
)
800 struct tcp_sock
*tp
= tcp_sk(sk
);
801 unsigned long flags
, nflags
;
803 /* perform an atomic operation only if at least one flag is set */
805 flags
= tp
->tsq_flags
;
806 if (!(flags
& TCP_DEFERRED_ALL
))
808 nflags
= flags
& ~TCP_DEFERRED_ALL
;
809 } while (cmpxchg(&tp
->tsq_flags
, flags
, nflags
) != flags
);
811 if (flags
& (1UL << TCP_TSQ_DEFERRED
))
814 /* Here begins the tricky part :
815 * We are called from release_sock() with :
817 * 2) sk_lock.slock spinlock held
818 * 3) socket owned by us (sk->sk_lock.owned == 1)
820 * But following code is meant to be called from BH handlers,
821 * so we should keep BH disabled, but early release socket ownership
823 sock_release_ownership(sk
);
825 if (flags
& (1UL << TCP_WRITE_TIMER_DEFERRED
)) {
826 tcp_write_timer_handler(sk
);
829 if (flags
& (1UL << TCP_DELACK_TIMER_DEFERRED
)) {
830 tcp_delack_timer_handler(sk
);
833 if (flags
& (1UL << TCP_MTU_REDUCED_DEFERRED
)) {
834 inet_csk(sk
)->icsk_af_ops
->mtu_reduced(sk
);
838 EXPORT_SYMBOL(tcp_release_cb
);
840 void __init
tcp_tasklet_init(void)
844 for_each_possible_cpu(i
) {
845 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
847 INIT_LIST_HEAD(&tsq
->head
);
848 tasklet_init(&tsq
->tasklet
,
855 * Write buffer destructor automatically called from kfree_skb.
856 * We can't xmit new skbs from this context, as we might already
859 void tcp_wfree(struct sk_buff
*skb
)
861 struct sock
*sk
= skb
->sk
;
862 struct tcp_sock
*tp
= tcp_sk(sk
);
865 /* Keep one reference on sk_wmem_alloc.
866 * Will be released by sk_free() from here or tcp_tasklet_func()
868 wmem
= atomic_sub_return(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
870 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
871 * Wait until our queues (qdisc + devices) are drained.
873 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
874 * - chance for incoming ACK (processed by another cpu maybe)
875 * to migrate this flow (skb->ooo_okay will be eventually set)
877 if (wmem
>= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current
)
880 if (test_and_clear_bit(TSQ_THROTTLED
, &tp
->tsq_flags
) &&
881 !test_and_set_bit(TSQ_QUEUED
, &tp
->tsq_flags
)) {
883 struct tsq_tasklet
*tsq
;
885 /* queue this socket to tasklet queue */
886 local_irq_save(flags
);
887 tsq
= this_cpu_ptr(&tsq_tasklet
);
888 list_add(&tp
->tsq_node
, &tsq
->head
);
889 tasklet_schedule(&tsq
->tasklet
);
890 local_irq_restore(flags
);
897 /* This routine actually transmits TCP packets queued in by
898 * tcp_do_sendmsg(). This is used by both the initial
899 * transmission and possible later retransmissions.
900 * All SKB's seen here are completely headerless. It is our
901 * job to build the TCP header, and pass the packet down to
902 * IP so it can do the same plus pass the packet off to the
905 * We are working here with either a clone of the original
906 * SKB, or a fresh unique copy made by the retransmit engine.
908 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
911 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
912 struct inet_sock
*inet
;
914 struct tcp_skb_cb
*tcb
;
915 struct tcp_out_options opts
;
916 unsigned int tcp_options_size
, tcp_header_size
;
917 struct tcp_md5sig_key
*md5
;
921 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
924 skb_mstamp_get(&skb
->skb_mstamp
);
926 if (unlikely(skb_cloned(skb
)))
927 skb
= pskb_copy(skb
, gfp_mask
);
929 skb
= skb_clone(skb
, gfp_mask
);
936 tcb
= TCP_SKB_CB(skb
);
937 memset(&opts
, 0, sizeof(opts
));
939 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
940 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
942 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
944 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
946 /* if no packet is in qdisc/device queue, then allow XPS to select
947 * another queue. We can be called from tcp_tsq_handler()
948 * which holds one reference to sk_wmem_alloc.
950 * TODO: Ideally, in-flight pure ACK packets should not matter here.
951 * One way to get this would be to set skb->truesize = 2 on them.
953 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) < SKB_TRUESIZE(1);
955 skb_push(skb
, tcp_header_size
);
956 skb_reset_transport_header(skb
);
960 skb
->destructor
= skb_is_tcp_pure_ack(skb
) ? sock_wfree
: tcp_wfree
;
961 skb_set_hash_from_sk(skb
, sk
);
962 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
964 /* Build TCP header and checksum it. */
966 th
->source
= inet
->inet_sport
;
967 th
->dest
= inet
->inet_dport
;
968 th
->seq
= htonl(tcb
->seq
);
969 th
->ack_seq
= htonl(tp
->rcv_nxt
);
970 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
973 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
974 /* RFC1323: The window in SYN & SYN/ACK segments
977 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
979 th
->window
= htons(tcp_select_window(sk
));
984 /* The urg_mode check is necessary during a below snd_una win probe */
985 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
986 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
987 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
989 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
990 th
->urg_ptr
= htons(0xFFFF);
995 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
996 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
997 if (likely((tcb
->tcp_flags
& TCPHDR_SYN
) == 0))
998 tcp_ecn_send(sk
, skb
, tcp_header_size
);
1000 #ifdef CONFIG_TCP_MD5SIG
1001 /* Calculate the MD5 hash, as we have all we need now */
1003 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
1004 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
1009 icsk
->icsk_af_ops
->send_check(sk
, skb
);
1011 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
1012 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
1014 if (skb
->len
!= tcp_header_size
)
1015 tcp_event_data_sent(tp
, sk
);
1017 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
1018 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
1019 tcp_skb_pcount(skb
));
1021 tp
->segs_out
+= tcp_skb_pcount(skb
);
1022 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1023 skb_shinfo(skb
)->gso_segs
= tcp_skb_pcount(skb
);
1024 skb_shinfo(skb
)->gso_size
= tcp_skb_mss(skb
);
1026 /* Our usage of tstamp should remain private */
1027 skb
->tstamp
.tv64
= 0;
1029 /* Cleanup our debris for IP stacks */
1030 memset(skb
->cb
, 0, max(sizeof(struct inet_skb_parm
),
1031 sizeof(struct inet6_skb_parm
)));
1033 err
= icsk
->icsk_af_ops
->queue_xmit(sk
, skb
, &inet
->cork
.fl
);
1035 if (likely(err
<= 0))
1040 return net_xmit_eval(err
);
1043 /* This routine just queues the buffer for sending.
1045 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1046 * otherwise socket can stall.
1048 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1050 struct tcp_sock
*tp
= tcp_sk(sk
);
1052 /* Advance write_seq and place onto the write_queue. */
1053 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
1054 __skb_header_release(skb
);
1055 tcp_add_write_queue_tail(sk
, skb
);
1056 sk
->sk_wmem_queued
+= skb
->truesize
;
1057 sk_mem_charge(sk
, skb
->truesize
);
1060 /* Initialize TSO segments for a packet. */
1061 static void tcp_set_skb_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1063 if (skb
->len
<= mss_now
|| skb
->ip_summed
== CHECKSUM_NONE
) {
1064 /* Avoid the costly divide in the normal
1067 tcp_skb_pcount_set(skb
, 1);
1068 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1070 tcp_skb_pcount_set(skb
, DIV_ROUND_UP(skb
->len
, mss_now
));
1071 TCP_SKB_CB(skb
)->tcp_gso_size
= mss_now
;
1075 /* When a modification to fackets out becomes necessary, we need to check
1076 * skb is counted to fackets_out or not.
1078 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1081 struct tcp_sock
*tp
= tcp_sk(sk
);
1083 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1086 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1087 tp
->fackets_out
-= decr
;
1090 /* Pcount in the middle of the write queue got changed, we need to do various
1091 * tweaks to fix counters
1093 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1095 struct tcp_sock
*tp
= tcp_sk(sk
);
1097 tp
->packets_out
-= decr
;
1099 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1100 tp
->sacked_out
-= decr
;
1101 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1102 tp
->retrans_out
-= decr
;
1103 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1104 tp
->lost_out
-= decr
;
1106 /* Reno case is special. Sigh... */
1107 if (tcp_is_reno(tp
) && decr
> 0)
1108 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1110 tcp_adjust_fackets_out(sk
, skb
, decr
);
1112 if (tp
->lost_skb_hint
&&
1113 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1114 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1115 tp
->lost_cnt_hint
-= decr
;
1117 tcp_verify_left_out(tp
);
1120 static void tcp_fragment_tstamp(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1122 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1124 if (unlikely(shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
) &&
1125 !before(shinfo
->tskey
, TCP_SKB_CB(skb2
)->seq
)) {
1126 struct skb_shared_info
*shinfo2
= skb_shinfo(skb2
);
1127 u8 tsflags
= shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
1129 shinfo
->tx_flags
&= ~tsflags
;
1130 shinfo2
->tx_flags
|= tsflags
;
1131 swap(shinfo
->tskey
, shinfo2
->tskey
);
1135 /* Function to create two new TCP segments. Shrinks the given segment
1136 * to the specified size and appends a new segment with the rest of the
1137 * packet to the list. This won't be called frequently, I hope.
1138 * Remember, these are still headerless SKBs at this point.
1140 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1141 unsigned int mss_now
, gfp_t gfp
)
1143 struct tcp_sock
*tp
= tcp_sk(sk
);
1144 struct sk_buff
*buff
;
1145 int nsize
, old_factor
;
1149 if (WARN_ON(len
> skb
->len
))
1152 nsize
= skb_headlen(skb
) - len
;
1156 if (skb_unclone(skb
, gfp
))
1159 /* Get a new skb... force flag on. */
1160 buff
= sk_stream_alloc_skb(sk
, nsize
, gfp
, true);
1162 return -ENOMEM
; /* We'll just try again later. */
1164 sk
->sk_wmem_queued
+= buff
->truesize
;
1165 sk_mem_charge(sk
, buff
->truesize
);
1166 nlen
= skb
->len
- len
- nsize
;
1167 buff
->truesize
+= nlen
;
1168 skb
->truesize
-= nlen
;
1170 /* Correct the sequence numbers. */
1171 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1172 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1173 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1175 /* PSH and FIN should only be set in the second packet. */
1176 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1177 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1178 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1179 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1181 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1182 /* Copy and checksum data tail into the new buffer. */
1183 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1184 skb_put(buff
, nsize
),
1189 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1191 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1192 skb_split(skb
, buff
, len
);
1195 buff
->ip_summed
= skb
->ip_summed
;
1197 buff
->tstamp
= skb
->tstamp
;
1198 tcp_fragment_tstamp(skb
, buff
);
1200 old_factor
= tcp_skb_pcount(skb
);
1202 /* Fix up tso_factor for both original and new SKB. */
1203 tcp_set_skb_tso_segs(skb
, mss_now
);
1204 tcp_set_skb_tso_segs(buff
, mss_now
);
1206 /* If this packet has been sent out already, we must
1207 * adjust the various packet counters.
1209 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1210 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1211 tcp_skb_pcount(buff
);
1214 tcp_adjust_pcount(sk
, skb
, diff
);
1217 /* Link BUFF into the send queue. */
1218 __skb_header_release(buff
);
1219 tcp_insert_write_queue_after(skb
, buff
, sk
);
1224 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1225 * eventually). The difference is that pulled data not copied, but
1226 * immediately discarded.
1228 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1230 struct skb_shared_info
*shinfo
;
1233 eat
= min_t(int, len
, skb_headlen(skb
));
1235 __skb_pull(skb
, eat
);
1242 shinfo
= skb_shinfo(skb
);
1243 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1244 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1247 skb_frag_unref(skb
, i
);
1250 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1252 shinfo
->frags
[k
].page_offset
+= eat
;
1253 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1259 shinfo
->nr_frags
= k
;
1261 skb_reset_tail_pointer(skb
);
1262 skb
->data_len
-= len
;
1263 skb
->len
= skb
->data_len
;
1266 /* Remove acked data from a packet in the transmit queue. */
1267 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1269 if (skb_unclone(skb
, GFP_ATOMIC
))
1272 __pskb_trim_head(skb
, len
);
1274 TCP_SKB_CB(skb
)->seq
+= len
;
1275 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1277 skb
->truesize
-= len
;
1278 sk
->sk_wmem_queued
-= len
;
1279 sk_mem_uncharge(sk
, len
);
1280 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1282 /* Any change of skb->len requires recalculation of tso factor. */
1283 if (tcp_skb_pcount(skb
) > 1)
1284 tcp_set_skb_tso_segs(skb
, tcp_skb_mss(skb
));
1289 /* Calculate MSS not accounting any TCP options. */
1290 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1292 const struct tcp_sock
*tp
= tcp_sk(sk
);
1293 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1296 /* Calculate base mss without TCP options:
1297 It is MMS_S - sizeof(tcphdr) of rfc1122
1299 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1301 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1302 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1303 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1305 if (dst
&& dst_allfrag(dst
))
1306 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1309 /* Clamp it (mss_clamp does not include tcp options) */
1310 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1311 mss_now
= tp
->rx_opt
.mss_clamp
;
1313 /* Now subtract optional transport overhead */
1314 mss_now
-= icsk
->icsk_ext_hdr_len
;
1316 /* Then reserve room for full set of TCP options and 8 bytes of data */
1322 /* Calculate MSS. Not accounting for SACKs here. */
1323 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1325 /* Subtract TCP options size, not including SACKs */
1326 return __tcp_mtu_to_mss(sk
, pmtu
) -
1327 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1330 /* Inverse of above */
1331 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1333 const struct tcp_sock
*tp
= tcp_sk(sk
);
1334 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1338 tp
->tcp_header_len
+
1339 icsk
->icsk_ext_hdr_len
+
1340 icsk
->icsk_af_ops
->net_header_len
;
1342 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1343 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1344 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1346 if (dst
&& dst_allfrag(dst
))
1347 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1352 /* MTU probing init per socket */
1353 void tcp_mtup_init(struct sock
*sk
)
1355 struct tcp_sock
*tp
= tcp_sk(sk
);
1356 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1357 struct net
*net
= sock_net(sk
);
1359 icsk
->icsk_mtup
.enabled
= net
->ipv4
.sysctl_tcp_mtu_probing
> 1;
1360 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1361 icsk
->icsk_af_ops
->net_header_len
;
1362 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, net
->ipv4
.sysctl_tcp_base_mss
);
1363 icsk
->icsk_mtup
.probe_size
= 0;
1364 if (icsk
->icsk_mtup
.enabled
)
1365 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1367 EXPORT_SYMBOL(tcp_mtup_init
);
1369 /* This function synchronize snd mss to current pmtu/exthdr set.
1371 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1372 for TCP options, but includes only bare TCP header.
1374 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1375 It is minimum of user_mss and mss received with SYN.
1376 It also does not include TCP options.
1378 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1380 tp->mss_cache is current effective sending mss, including
1381 all tcp options except for SACKs. It is evaluated,
1382 taking into account current pmtu, but never exceeds
1383 tp->rx_opt.mss_clamp.
1385 NOTE1. rfc1122 clearly states that advertised MSS
1386 DOES NOT include either tcp or ip options.
1388 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1389 are READ ONLY outside this function. --ANK (980731)
1391 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1393 struct tcp_sock
*tp
= tcp_sk(sk
);
1394 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1397 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1398 icsk
->icsk_mtup
.search_high
= pmtu
;
1400 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1401 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1403 /* And store cached results */
1404 icsk
->icsk_pmtu_cookie
= pmtu
;
1405 if (icsk
->icsk_mtup
.enabled
)
1406 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1407 tp
->mss_cache
= mss_now
;
1411 EXPORT_SYMBOL(tcp_sync_mss
);
1413 /* Compute the current effective MSS, taking SACKs and IP options,
1414 * and even PMTU discovery events into account.
1416 unsigned int tcp_current_mss(struct sock
*sk
)
1418 const struct tcp_sock
*tp
= tcp_sk(sk
);
1419 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1421 unsigned int header_len
;
1422 struct tcp_out_options opts
;
1423 struct tcp_md5sig_key
*md5
;
1425 mss_now
= tp
->mss_cache
;
1428 u32 mtu
= dst_mtu(dst
);
1429 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1430 mss_now
= tcp_sync_mss(sk
, mtu
);
1433 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1434 sizeof(struct tcphdr
);
1435 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1436 * some common options. If this is an odd packet (because we have SACK
1437 * blocks etc) then our calculated header_len will be different, and
1438 * we have to adjust mss_now correspondingly */
1439 if (header_len
!= tp
->tcp_header_len
) {
1440 int delta
= (int) header_len
- tp
->tcp_header_len
;
1447 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1448 * As additional protections, we do not touch cwnd in retransmission phases,
1449 * and if application hit its sndbuf limit recently.
1451 static void tcp_cwnd_application_limited(struct sock
*sk
)
1453 struct tcp_sock
*tp
= tcp_sk(sk
);
1455 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1456 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1457 /* Limited by application or receiver window. */
1458 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1459 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1460 if (win_used
< tp
->snd_cwnd
) {
1461 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1462 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
1464 tp
->snd_cwnd_used
= 0;
1466 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1469 static void tcp_cwnd_validate(struct sock
*sk
, bool is_cwnd_limited
)
1471 struct tcp_sock
*tp
= tcp_sk(sk
);
1473 /* Track the maximum number of outstanding packets in each
1474 * window, and remember whether we were cwnd-limited then.
1476 if (!before(tp
->snd_una
, tp
->max_packets_seq
) ||
1477 tp
->packets_out
> tp
->max_packets_out
) {
1478 tp
->max_packets_out
= tp
->packets_out
;
1479 tp
->max_packets_seq
= tp
->snd_nxt
;
1480 tp
->is_cwnd_limited
= is_cwnd_limited
;
1483 if (tcp_is_cwnd_limited(sk
)) {
1484 /* Network is feed fully. */
1485 tp
->snd_cwnd_used
= 0;
1486 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1488 /* Network starves. */
1489 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1490 tp
->snd_cwnd_used
= tp
->packets_out
;
1492 if (sysctl_tcp_slow_start_after_idle
&&
1493 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1494 tcp_cwnd_application_limited(sk
);
1498 /* Minshall's variant of the Nagle send check. */
1499 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1501 return after(tp
->snd_sml
, tp
->snd_una
) &&
1502 !after(tp
->snd_sml
, tp
->snd_nxt
);
1505 /* Update snd_sml if this skb is under mss
1506 * Note that a TSO packet might end with a sub-mss segment
1507 * The test is really :
1508 * if ((skb->len % mss) != 0)
1509 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1510 * But we can avoid doing the divide again given we already have
1511 * skb_pcount = skb->len / mss_now
1513 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1514 const struct sk_buff
*skb
)
1516 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1517 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1520 /* Return false, if packet can be sent now without violation Nagle's rules:
1521 * 1. It is full sized. (provided by caller in %partial bool)
1522 * 2. Or it contains FIN. (already checked by caller)
1523 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1524 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1525 * With Minshall's modification: all sent small packets are ACKed.
1527 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1531 ((nonagle
& TCP_NAGLE_CORK
) ||
1532 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1535 /* Return how many segs we'd like on a TSO packet,
1536 * to send one TSO packet per ms
1538 static u32
tcp_tso_autosize(const struct sock
*sk
, unsigned int mss_now
)
1542 bytes
= min(sk
->sk_pacing_rate
>> 10,
1543 sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
);
1545 /* Goal is to send at least one packet per ms,
1546 * not one big TSO packet every 100 ms.
1547 * This preserves ACK clocking and is consistent
1548 * with tcp_tso_should_defer() heuristic.
1550 segs
= max_t(u32
, bytes
/ mss_now
, sysctl_tcp_min_tso_segs
);
1552 return min_t(u32
, segs
, sk
->sk_gso_max_segs
);
1555 /* Returns the portion of skb which can be sent right away */
1556 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1557 const struct sk_buff
*skb
,
1558 unsigned int mss_now
,
1559 unsigned int max_segs
,
1562 const struct tcp_sock
*tp
= tcp_sk(sk
);
1563 u32 partial
, needed
, window
, max_len
;
1565 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1566 max_len
= mss_now
* max_segs
;
1568 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1571 needed
= min(skb
->len
, window
);
1573 if (max_len
<= needed
)
1576 partial
= needed
% mss_now
;
1577 /* If last segment is not a full MSS, check if Nagle rules allow us
1578 * to include this last segment in this skb.
1579 * Otherwise, we'll split the skb at last MSS boundary
1581 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
1582 return needed
- partial
;
1587 /* Can at least one segment of SKB be sent right now, according to the
1588 * congestion window rules? If so, return how many segments are allowed.
1590 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1591 const struct sk_buff
*skb
)
1593 u32 in_flight
, cwnd
, halfcwnd
;
1595 /* Don't be strict about the congestion window for the final FIN. */
1596 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1597 tcp_skb_pcount(skb
) == 1)
1600 in_flight
= tcp_packets_in_flight(tp
);
1601 cwnd
= tp
->snd_cwnd
;
1602 if (in_flight
>= cwnd
)
1605 /* For better scheduling, ensure we have at least
1606 * 2 GSO packets in flight.
1608 halfcwnd
= max(cwnd
>> 1, 1U);
1609 return min(halfcwnd
, cwnd
- in_flight
);
1612 /* Initialize TSO state of a skb.
1613 * This must be invoked the first time we consider transmitting
1614 * SKB onto the wire.
1616 static int tcp_init_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1618 int tso_segs
= tcp_skb_pcount(skb
);
1620 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1621 tcp_set_skb_tso_segs(skb
, mss_now
);
1622 tso_segs
= tcp_skb_pcount(skb
);
1628 /* Return true if the Nagle test allows this packet to be
1631 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1632 unsigned int cur_mss
, int nonagle
)
1634 /* Nagle rule does not apply to frames, which sit in the middle of the
1635 * write_queue (they have no chances to get new data).
1637 * This is implemented in the callers, where they modify the 'nonagle'
1638 * argument based upon the location of SKB in the send queue.
1640 if (nonagle
& TCP_NAGLE_PUSH
)
1643 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1644 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1647 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
1653 /* Does at least the first segment of SKB fit into the send window? */
1654 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1655 const struct sk_buff
*skb
,
1656 unsigned int cur_mss
)
1658 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1660 if (skb
->len
> cur_mss
)
1661 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1663 return !after(end_seq
, tcp_wnd_end(tp
));
1666 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1667 * should be put on the wire right now. If so, it returns the number of
1668 * packets allowed by the congestion window.
1670 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1671 unsigned int cur_mss
, int nonagle
)
1673 const struct tcp_sock
*tp
= tcp_sk(sk
);
1674 unsigned int cwnd_quota
;
1676 tcp_init_tso_segs(skb
, cur_mss
);
1678 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1681 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1682 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1688 /* Test if sending is allowed right now. */
1689 bool tcp_may_send_now(struct sock
*sk
)
1691 const struct tcp_sock
*tp
= tcp_sk(sk
);
1692 struct sk_buff
*skb
= tcp_send_head(sk
);
1695 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1696 (tcp_skb_is_last(sk
, skb
) ?
1697 tp
->nonagle
: TCP_NAGLE_PUSH
));
1700 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1701 * which is put after SKB on the list. It is very much like
1702 * tcp_fragment() except that it may make several kinds of assumptions
1703 * in order to speed up the splitting operation. In particular, we
1704 * know that all the data is in scatter-gather pages, and that the
1705 * packet has never been sent out before (and thus is not cloned).
1707 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1708 unsigned int mss_now
, gfp_t gfp
)
1710 struct sk_buff
*buff
;
1711 int nlen
= skb
->len
- len
;
1714 /* All of a TSO frame must be composed of paged data. */
1715 if (skb
->len
!= skb
->data_len
)
1716 return tcp_fragment(sk
, skb
, len
, mss_now
, gfp
);
1718 buff
= sk_stream_alloc_skb(sk
, 0, gfp
, true);
1719 if (unlikely(!buff
))
1722 sk
->sk_wmem_queued
+= buff
->truesize
;
1723 sk_mem_charge(sk
, buff
->truesize
);
1724 buff
->truesize
+= nlen
;
1725 skb
->truesize
-= nlen
;
1727 /* Correct the sequence numbers. */
1728 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1729 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1730 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1732 /* PSH and FIN should only be set in the second packet. */
1733 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1734 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1735 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1737 /* This packet was never sent out yet, so no SACK bits. */
1738 TCP_SKB_CB(buff
)->sacked
= 0;
1740 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1741 skb_split(skb
, buff
, len
);
1742 tcp_fragment_tstamp(skb
, buff
);
1744 /* Fix up tso_factor for both original and new SKB. */
1745 tcp_set_skb_tso_segs(skb
, mss_now
);
1746 tcp_set_skb_tso_segs(buff
, mss_now
);
1748 /* Link BUFF into the send queue. */
1749 __skb_header_release(buff
);
1750 tcp_insert_write_queue_after(skb
, buff
, sk
);
1755 /* Try to defer sending, if possible, in order to minimize the amount
1756 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1758 * This algorithm is from John Heffner.
1760 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
,
1761 bool *is_cwnd_limited
, u32 max_segs
)
1763 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1764 u32 age
, send_win
, cong_win
, limit
, in_flight
;
1765 struct tcp_sock
*tp
= tcp_sk(sk
);
1766 struct skb_mstamp now
;
1767 struct sk_buff
*head
;
1770 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1773 if (icsk
->icsk_ca_state
>= TCP_CA_Recovery
)
1776 /* Avoid bursty behavior by allowing defer
1777 * only if the last write was recent.
1779 if ((s32
)(tcp_time_stamp
- tp
->lsndtime
) > 0)
1782 in_flight
= tcp_packets_in_flight(tp
);
1784 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1786 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1788 /* From in_flight test above, we know that cwnd > in_flight. */
1789 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1791 limit
= min(send_win
, cong_win
);
1793 /* If a full-sized TSO skb can be sent, do it. */
1794 if (limit
>= max_segs
* tp
->mss_cache
)
1797 /* Middle in queue won't get any more data, full sendable already? */
1798 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1801 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1803 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1805 /* If at least some fraction of a window is available,
1808 chunk
/= win_divisor
;
1812 /* Different approach, try not to defer past a single
1813 * ACK. Receiver should ACK every other full sized
1814 * frame, so if we have space for more than 3 frames
1817 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1821 head
= tcp_write_queue_head(sk
);
1822 skb_mstamp_get(&now
);
1823 age
= skb_mstamp_us_delta(&now
, &head
->skb_mstamp
);
1824 /* If next ACK is likely to come too late (half srtt), do not defer */
1825 if (age
< (tp
->srtt_us
>> 4))
1828 /* Ok, it looks like it is advisable to defer. */
1830 if (cong_win
< send_win
&& cong_win
< skb
->len
)
1831 *is_cwnd_limited
= true;
1839 static inline void tcp_mtu_check_reprobe(struct sock
*sk
)
1841 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1842 struct tcp_sock
*tp
= tcp_sk(sk
);
1843 struct net
*net
= sock_net(sk
);
1847 interval
= net
->ipv4
.sysctl_tcp_probe_interval
;
1848 delta
= tcp_time_stamp
- icsk
->icsk_mtup
.probe_timestamp
;
1849 if (unlikely(delta
>= interval
* HZ
)) {
1850 int mss
= tcp_current_mss(sk
);
1852 /* Update current search range */
1853 icsk
->icsk_mtup
.probe_size
= 0;
1854 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+
1855 sizeof(struct tcphdr
) +
1856 icsk
->icsk_af_ops
->net_header_len
;
1857 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
1859 /* Update probe time stamp */
1860 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1864 /* Create a new MTU probe if we are ready.
1865 * MTU probe is regularly attempting to increase the path MTU by
1866 * deliberately sending larger packets. This discovers routing
1867 * changes resulting in larger path MTUs.
1869 * Returns 0 if we should wait to probe (no cwnd available),
1870 * 1 if a probe was sent,
1873 static int tcp_mtu_probe(struct sock
*sk
)
1875 struct tcp_sock
*tp
= tcp_sk(sk
);
1876 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1877 struct sk_buff
*skb
, *nskb
, *next
;
1878 struct net
*net
= sock_net(sk
);
1886 /* Not currently probing/verifying,
1888 * have enough cwnd, and
1889 * not SACKing (the variable headers throw things off) */
1890 if (!icsk
->icsk_mtup
.enabled
||
1891 icsk
->icsk_mtup
.probe_size
||
1892 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1893 tp
->snd_cwnd
< 11 ||
1894 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
)
1897 /* Use binary search for probe_size between tcp_mss_base,
1898 * and current mss_clamp. if (search_high - search_low)
1899 * smaller than a threshold, backoff from probing.
1901 mss_now
= tcp_current_mss(sk
);
1902 probe_size
= tcp_mtu_to_mss(sk
, (icsk
->icsk_mtup
.search_high
+
1903 icsk
->icsk_mtup
.search_low
) >> 1);
1904 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1905 interval
= icsk
->icsk_mtup
.search_high
- icsk
->icsk_mtup
.search_low
;
1906 /* When misfortune happens, we are reprobing actively,
1907 * and then reprobe timer has expired. We stick with current
1908 * probing process by not resetting search range to its orignal.
1910 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
) ||
1911 interval
< net
->ipv4
.sysctl_tcp_probe_threshold
) {
1912 /* Check whether enough time has elaplased for
1913 * another round of probing.
1915 tcp_mtu_check_reprobe(sk
);
1919 /* Have enough data in the send queue to probe? */
1920 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1923 if (tp
->snd_wnd
< size_needed
)
1925 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1928 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1929 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1930 if (!tcp_packets_in_flight(tp
))
1936 /* We're allowed to probe. Build it now. */
1937 nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
, false);
1940 sk
->sk_wmem_queued
+= nskb
->truesize
;
1941 sk_mem_charge(sk
, nskb
->truesize
);
1943 skb
= tcp_send_head(sk
);
1945 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
1946 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
1947 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
1948 TCP_SKB_CB(nskb
)->sacked
= 0;
1950 nskb
->ip_summed
= skb
->ip_summed
;
1952 tcp_insert_write_queue_before(nskb
, skb
, sk
);
1955 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
1956 copy
= min_t(int, skb
->len
, probe_size
- len
);
1957 if (nskb
->ip_summed
)
1958 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
1960 nskb
->csum
= skb_copy_and_csum_bits(skb
, 0,
1961 skb_put(nskb
, copy
),
1964 if (skb
->len
<= copy
) {
1965 /* We've eaten all the data from this skb.
1967 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
1968 tcp_unlink_write_queue(skb
, sk
);
1969 sk_wmem_free_skb(sk
, skb
);
1971 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
1972 ~(TCPHDR_FIN
|TCPHDR_PSH
);
1973 if (!skb_shinfo(skb
)->nr_frags
) {
1974 skb_pull(skb
, copy
);
1975 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
1976 skb
->csum
= csum_partial(skb
->data
,
1979 __pskb_trim_head(skb
, copy
);
1980 tcp_set_skb_tso_segs(skb
, mss_now
);
1982 TCP_SKB_CB(skb
)->seq
+= copy
;
1987 if (len
>= probe_size
)
1990 tcp_init_tso_segs(nskb
, nskb
->len
);
1992 /* We're ready to send. If this fails, the probe will
1993 * be resegmented into mss-sized pieces by tcp_write_xmit().
1995 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
1996 /* Decrement cwnd here because we are sending
1997 * effectively two packets. */
1999 tcp_event_new_data_sent(sk
, nskb
);
2001 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
2002 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
2003 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
2011 /* This routine writes packets to the network. It advances the
2012 * send_head. This happens as incoming acks open up the remote
2015 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2016 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2017 * account rare use of URG, this is not a big flaw.
2019 * Send at most one packet when push_one > 0. Temporarily ignore
2020 * cwnd limit to force at most one packet out when push_one == 2.
2022 * Returns true, if no segments are in flight and we have queued segments,
2023 * but cannot send anything now because of SWS or another problem.
2025 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
2026 int push_one
, gfp_t gfp
)
2028 struct tcp_sock
*tp
= tcp_sk(sk
);
2029 struct sk_buff
*skb
;
2030 unsigned int tso_segs
, sent_pkts
;
2033 bool is_cwnd_limited
= false;
2039 /* Do MTU probing. */
2040 result
= tcp_mtu_probe(sk
);
2043 } else if (result
> 0) {
2048 max_segs
= tcp_tso_autosize(sk
, mss_now
);
2049 while ((skb
= tcp_send_head(sk
))) {
2052 tso_segs
= tcp_init_tso_segs(skb
, mss_now
);
2055 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
) {
2056 /* "skb_mstamp" is used as a start point for the retransmit timer */
2057 skb_mstamp_get(&skb
->skb_mstamp
);
2058 goto repair
; /* Skip network transmission */
2061 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2063 is_cwnd_limited
= true;
2065 /* Force out a loss probe pkt. */
2071 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
)))
2074 if (tso_segs
== 1) {
2075 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2076 (tcp_skb_is_last(sk
, skb
) ?
2077 nonagle
: TCP_NAGLE_PUSH
))))
2081 tcp_tso_should_defer(sk
, skb
, &is_cwnd_limited
,
2087 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
2088 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2094 if (skb
->len
> limit
&&
2095 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
2098 /* TCP Small Queues :
2099 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2101 * - better RTT estimation and ACK scheduling
2104 * Alas, some drivers / subsystems require a fair amount
2105 * of queued bytes to ensure line rate.
2106 * One example is wifi aggregation (802.11 AMPDU)
2108 limit
= max(2 * skb
->truesize
, sk
->sk_pacing_rate
>> 10);
2109 limit
= min_t(u32
, limit
, sysctl_tcp_limit_output_bytes
);
2111 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
) {
2112 set_bit(TSQ_THROTTLED
, &tp
->tsq_flags
);
2113 /* It is possible TX completion already happened
2114 * before we set TSQ_THROTTLED, so we must
2115 * test again the condition.
2117 smp_mb__after_atomic();
2118 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
)
2122 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2126 /* Advance the send_head. This one is sent out.
2127 * This call will increment packets_out.
2129 tcp_event_new_data_sent(sk
, skb
);
2131 tcp_minshall_update(tp
, mss_now
, skb
);
2132 sent_pkts
+= tcp_skb_pcount(skb
);
2138 if (likely(sent_pkts
)) {
2139 if (tcp_in_cwnd_reduction(sk
))
2140 tp
->prr_out
+= sent_pkts
;
2142 /* Send one loss probe per tail loss episode. */
2144 tcp_schedule_loss_probe(sk
);
2145 tcp_cwnd_validate(sk
, is_cwnd_limited
);
2148 return !tp
->packets_out
&& tcp_send_head(sk
);
2151 bool tcp_schedule_loss_probe(struct sock
*sk
)
2153 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2154 struct tcp_sock
*tp
= tcp_sk(sk
);
2155 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
2156 u32 rtt
= usecs_to_jiffies(tp
->srtt_us
>> 3);
2158 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
))
2160 /* No consecutive loss probes. */
2161 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
2165 /* Don't do any loss probe on a Fast Open connection before 3WHS
2168 if (sk
->sk_state
== TCP_SYN_RECV
)
2171 /* TLP is only scheduled when next timer event is RTO. */
2172 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
2175 /* Schedule a loss probe in 2*RTT for SACK capable connections
2176 * in Open state, that are either limited by cwnd or application.
2178 if (sysctl_tcp_early_retrans
< 3 || !tp
->srtt_us
|| !tp
->packets_out
||
2179 !tcp_is_sack(tp
) || inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
2182 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
2186 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2187 * for delayed ack when there's one outstanding packet.
2190 if (tp
->packets_out
== 1)
2191 timeout
= max_t(u32
, timeout
,
2192 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
2193 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
2195 /* If RTO is shorter, just schedule TLP in its place. */
2196 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
2197 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
2198 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
2199 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
2204 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2209 /* Thanks to skb fast clones, we can detect if a prior transmit of
2210 * a packet is still in a qdisc or driver queue.
2211 * In this case, there is very little point doing a retransmit !
2212 * Note: This is called from BH context only.
2214 static bool skb_still_in_host_queue(const struct sock
*sk
,
2215 const struct sk_buff
*skb
)
2217 if (unlikely(skb_fclone_busy(sk
, skb
))) {
2218 NET_INC_STATS_BH(sock_net(sk
),
2219 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
2225 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2226 * retransmit the last segment.
2228 void tcp_send_loss_probe(struct sock
*sk
)
2230 struct tcp_sock
*tp
= tcp_sk(sk
);
2231 struct sk_buff
*skb
;
2233 int mss
= tcp_current_mss(sk
);
2235 skb
= tcp_send_head(sk
);
2237 if (tcp_snd_wnd_test(tp
, skb
, mss
)) {
2238 pcount
= tp
->packets_out
;
2239 tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2240 if (tp
->packets_out
> pcount
)
2244 skb
= tcp_write_queue_prev(sk
, skb
);
2246 skb
= tcp_write_queue_tail(sk
);
2249 /* At most one outstanding TLP retransmission. */
2250 if (tp
->tlp_high_seq
)
2253 /* Retransmit last segment. */
2257 if (skb_still_in_host_queue(sk
, skb
))
2260 pcount
= tcp_skb_pcount(skb
);
2261 if (WARN_ON(!pcount
))
2264 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2265 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
,
2268 skb
= tcp_write_queue_next(sk
, skb
);
2271 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2274 if (__tcp_retransmit_skb(sk
, skb
))
2277 /* Record snd_nxt for loss detection. */
2278 tp
->tlp_high_seq
= tp
->snd_nxt
;
2281 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPLOSSPROBES
);
2282 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2283 inet_csk(sk
)->icsk_pending
= 0;
2288 /* Push out any pending frames which were held back due to
2289 * TCP_CORK or attempt at coalescing tiny packets.
2290 * The socket must be locked by the caller.
2292 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2295 /* If we are closed, the bytes will have to remain here.
2296 * In time closedown will finish, we empty the write queue and
2297 * all will be happy.
2299 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2302 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2303 sk_gfp_atomic(sk
, GFP_ATOMIC
)))
2304 tcp_check_probe_timer(sk
);
2307 /* Send _single_ skb sitting at the send head. This function requires
2308 * true push pending frames to setup probe timer etc.
2310 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2312 struct sk_buff
*skb
= tcp_send_head(sk
);
2314 BUG_ON(!skb
|| skb
->len
< mss_now
);
2316 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2319 /* This function returns the amount that we can raise the
2320 * usable window based on the following constraints
2322 * 1. The window can never be shrunk once it is offered (RFC 793)
2323 * 2. We limit memory per socket
2326 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2327 * RECV.NEXT + RCV.WIN fixed until:
2328 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2330 * i.e. don't raise the right edge of the window until you can raise
2331 * it at least MSS bytes.
2333 * Unfortunately, the recommended algorithm breaks header prediction,
2334 * since header prediction assumes th->window stays fixed.
2336 * Strictly speaking, keeping th->window fixed violates the receiver
2337 * side SWS prevention criteria. The problem is that under this rule
2338 * a stream of single byte packets will cause the right side of the
2339 * window to always advance by a single byte.
2341 * Of course, if the sender implements sender side SWS prevention
2342 * then this will not be a problem.
2344 * BSD seems to make the following compromise:
2346 * If the free space is less than the 1/4 of the maximum
2347 * space available and the free space is less than 1/2 mss,
2348 * then set the window to 0.
2349 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2350 * Otherwise, just prevent the window from shrinking
2351 * and from being larger than the largest representable value.
2353 * This prevents incremental opening of the window in the regime
2354 * where TCP is limited by the speed of the reader side taking
2355 * data out of the TCP receive queue. It does nothing about
2356 * those cases where the window is constrained on the sender side
2357 * because the pipeline is full.
2359 * BSD also seems to "accidentally" limit itself to windows that are a
2360 * multiple of MSS, at least until the free space gets quite small.
2361 * This would appear to be a side effect of the mbuf implementation.
2362 * Combining these two algorithms results in the observed behavior
2363 * of having a fixed window size at almost all times.
2365 * Below we obtain similar behavior by forcing the offered window to
2366 * a multiple of the mss when it is feasible to do so.
2368 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2369 * Regular options like TIMESTAMP are taken into account.
2371 u32
__tcp_select_window(struct sock
*sk
)
2373 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2374 struct tcp_sock
*tp
= tcp_sk(sk
);
2375 /* MSS for the peer's data. Previous versions used mss_clamp
2376 * here. I don't know if the value based on our guesses
2377 * of peer's MSS is better for the performance. It's more correct
2378 * but may be worse for the performance because of rcv_mss
2379 * fluctuations. --SAW 1998/11/1
2381 int mss
= icsk
->icsk_ack
.rcv_mss
;
2382 int free_space
= tcp_space(sk
);
2383 int allowed_space
= tcp_full_space(sk
);
2384 int full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2387 if (mss
> full_space
)
2390 if (free_space
< (full_space
>> 1)) {
2391 icsk
->icsk_ack
.quick
= 0;
2393 if (tcp_under_memory_pressure(sk
))
2394 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2397 /* free_space might become our new window, make sure we don't
2398 * increase it due to wscale.
2400 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2402 /* if free space is less than mss estimate, or is below 1/16th
2403 * of the maximum allowed, try to move to zero-window, else
2404 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2405 * new incoming data is dropped due to memory limits.
2406 * With large window, mss test triggers way too late in order
2407 * to announce zero window in time before rmem limit kicks in.
2409 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2413 if (free_space
> tp
->rcv_ssthresh
)
2414 free_space
= tp
->rcv_ssthresh
;
2416 /* Don't do rounding if we are using window scaling, since the
2417 * scaled window will not line up with the MSS boundary anyway.
2419 window
= tp
->rcv_wnd
;
2420 if (tp
->rx_opt
.rcv_wscale
) {
2421 window
= free_space
;
2423 /* Advertise enough space so that it won't get scaled away.
2424 * Import case: prevent zero window announcement if
2425 * 1<<rcv_wscale > mss.
2427 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2428 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2429 << tp
->rx_opt
.rcv_wscale
);
2431 /* Get the largest window that is a nice multiple of mss.
2432 * Window clamp already applied above.
2433 * If our current window offering is within 1 mss of the
2434 * free space we just keep it. This prevents the divide
2435 * and multiply from happening most of the time.
2436 * We also don't do any window rounding when the free space
2439 if (window
<= free_space
- mss
|| window
> free_space
)
2440 window
= (free_space
/ mss
) * mss
;
2441 else if (mss
== full_space
&&
2442 free_space
> window
+ (full_space
>> 1))
2443 window
= free_space
;
2449 /* Collapses two adjacent SKB's during retransmission. */
2450 static void tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2452 struct tcp_sock
*tp
= tcp_sk(sk
);
2453 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2454 int skb_size
, next_skb_size
;
2456 skb_size
= skb
->len
;
2457 next_skb_size
= next_skb
->len
;
2459 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2461 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2463 tcp_unlink_write_queue(next_skb
, sk
);
2465 skb_copy_from_linear_data(next_skb
, skb_put(skb
, next_skb_size
),
2468 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2469 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2471 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2472 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2474 /* Update sequence range on original skb. */
2475 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2477 /* Merge over control information. This moves PSH/FIN etc. over */
2478 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2480 /* All done, get rid of second SKB and account for it so
2481 * packet counting does not break.
2483 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2485 /* changed transmit queue under us so clear hints */
2486 tcp_clear_retrans_hints_partial(tp
);
2487 if (next_skb
== tp
->retransmit_skb_hint
)
2488 tp
->retransmit_skb_hint
= skb
;
2490 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2492 sk_wmem_free_skb(sk
, next_skb
);
2495 /* Check if coalescing SKBs is legal. */
2496 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2498 if (tcp_skb_pcount(skb
) > 1)
2500 /* TODO: SACK collapsing could be used to remove this condition */
2501 if (skb_shinfo(skb
)->nr_frags
!= 0)
2503 if (skb_cloned(skb
))
2505 if (skb
== tcp_send_head(sk
))
2507 /* Some heurestics for collapsing over SACK'd could be invented */
2508 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2514 /* Collapse packets in the retransmit queue to make to create
2515 * less packets on the wire. This is only done on retransmission.
2517 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2520 struct tcp_sock
*tp
= tcp_sk(sk
);
2521 struct sk_buff
*skb
= to
, *tmp
;
2524 if (!sysctl_tcp_retrans_collapse
)
2526 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2529 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2530 if (!tcp_can_collapse(sk
, skb
))
2542 /* Punt if not enough space exists in the first SKB for
2543 * the data in the second
2545 if (skb
->len
> skb_availroom(to
))
2548 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2551 tcp_collapse_retrans(sk
, to
);
2555 /* This retransmits one SKB. Policy decisions and retransmit queue
2556 * state updates are done by the caller. Returns non-zero if an
2557 * error occurred which prevented the send.
2559 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2561 struct tcp_sock
*tp
= tcp_sk(sk
);
2562 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2563 unsigned int cur_mss
;
2566 /* Inconslusive MTU probe */
2567 if (icsk
->icsk_mtup
.probe_size
) {
2568 icsk
->icsk_mtup
.probe_size
= 0;
2571 /* Do not sent more than we queued. 1/4 is reserved for possible
2572 * copying overhead: fragmentation, tunneling, mangling etc.
2574 if (atomic_read(&sk
->sk_wmem_alloc
) >
2575 min(sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2), sk
->sk_sndbuf
))
2578 if (skb_still_in_host_queue(sk
, skb
))
2581 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2582 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2584 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2588 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2589 return -EHOSTUNREACH
; /* Routing failure or similar. */
2591 cur_mss
= tcp_current_mss(sk
);
2593 /* If receiver has shrunk his window, and skb is out of
2594 * new window, do not retransmit it. The exception is the
2595 * case, when window is shrunk to zero. In this case
2596 * our retransmit serves as a zero window probe.
2598 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2599 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2602 if (skb
->len
> cur_mss
) {
2603 if (tcp_fragment(sk
, skb
, cur_mss
, cur_mss
, GFP_ATOMIC
))
2604 return -ENOMEM
; /* We'll try again later. */
2606 int oldpcount
= tcp_skb_pcount(skb
);
2608 if (unlikely(oldpcount
> 1)) {
2609 if (skb_unclone(skb
, GFP_ATOMIC
))
2611 tcp_init_tso_segs(skb
, cur_mss
);
2612 tcp_adjust_pcount(sk
, skb
, oldpcount
- tcp_skb_pcount(skb
));
2616 /* RFC3168, section 6.1.1.1. ECN fallback */
2617 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN_ECN
) == TCPHDR_SYN_ECN
)
2618 tcp_ecn_clear_syn(sk
, skb
);
2620 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2622 /* Make a copy, if the first transmission SKB clone we made
2623 * is still in somebody's hands, else make a clone.
2626 /* make sure skb->data is aligned on arches that require it
2627 * and check if ack-trimming & collapsing extended the headroom
2628 * beyond what csum_start can cover.
2630 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2631 skb_headroom(skb
) >= 0xFFFF)) {
2632 struct sk_buff
*nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
,
2634 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2637 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2641 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
2642 /* Update global TCP statistics. */
2643 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
2644 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2645 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
2646 tp
->total_retrans
++;
2651 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
)
2653 struct tcp_sock
*tp
= tcp_sk(sk
);
2654 int err
= __tcp_retransmit_skb(sk
, skb
);
2657 #if FASTRETRANS_DEBUG > 0
2658 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2659 net_dbg_ratelimited("retrans_out leaked\n");
2662 if (!tp
->retrans_out
)
2663 tp
->lost_retrans_low
= tp
->snd_nxt
;
2664 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2665 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2667 /* Save stamp of the first retransmit. */
2668 if (!tp
->retrans_stamp
)
2669 tp
->retrans_stamp
= tcp_skb_timestamp(skb
);
2671 /* snd_nxt is stored to detect loss of retransmitted segment,
2672 * see tcp_input.c tcp_sacktag_write_queue().
2674 TCP_SKB_CB(skb
)->ack_seq
= tp
->snd_nxt
;
2675 } else if (err
!= -EBUSY
) {
2676 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2679 if (tp
->undo_retrans
< 0)
2680 tp
->undo_retrans
= 0;
2681 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2685 /* Check if we forward retransmits are possible in the current
2686 * window/congestion state.
2688 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2690 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2691 const struct tcp_sock
*tp
= tcp_sk(sk
);
2693 /* Forward retransmissions are possible only during Recovery. */
2694 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2697 /* No forward retransmissions in Reno are possible. */
2698 if (tcp_is_reno(tp
))
2701 /* Yeah, we have to make difficult choice between forward transmission
2702 * and retransmission... Both ways have their merits...
2704 * For now we do not retransmit anything, while we have some new
2705 * segments to send. In the other cases, follow rule 3 for
2706 * NextSeg() specified in RFC3517.
2709 if (tcp_may_send_now(sk
))
2715 /* This gets called after a retransmit timeout, and the initially
2716 * retransmitted data is acknowledged. It tries to continue
2717 * resending the rest of the retransmit queue, until either
2718 * we've sent it all or the congestion window limit is reached.
2719 * If doing SACK, the first ACK which comes back for a timeout
2720 * based retransmit packet might feed us FACK information again.
2721 * If so, we use it to avoid unnecessarily retransmissions.
2723 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2725 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2726 struct tcp_sock
*tp
= tcp_sk(sk
);
2727 struct sk_buff
*skb
;
2728 struct sk_buff
*hole
= NULL
;
2731 int fwd_rexmitting
= 0;
2733 if (!tp
->packets_out
)
2737 tp
->retransmit_high
= tp
->snd_una
;
2739 if (tp
->retransmit_skb_hint
) {
2740 skb
= tp
->retransmit_skb_hint
;
2741 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2742 if (after(last_lost
, tp
->retransmit_high
))
2743 last_lost
= tp
->retransmit_high
;
2745 skb
= tcp_write_queue_head(sk
);
2746 last_lost
= tp
->snd_una
;
2749 tcp_for_write_queue_from(skb
, sk
) {
2750 __u8 sacked
= TCP_SKB_CB(skb
)->sacked
;
2752 if (skb
== tcp_send_head(sk
))
2754 /* we could do better than to assign each time */
2756 tp
->retransmit_skb_hint
= skb
;
2758 /* Assume this retransmit will generate
2759 * only one packet for congestion window
2760 * calculation purposes. This works because
2761 * tcp_retransmit_skb() will chop up the
2762 * packet to be MSS sized and all the
2763 * packet counting works out.
2765 if (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
)
2768 if (fwd_rexmitting
) {
2770 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2772 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2774 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2775 tp
->retransmit_high
= last_lost
;
2776 if (!tcp_can_forward_retransmit(sk
))
2778 /* Backtrack if necessary to non-L'ed skb */
2786 } else if (!(sacked
& TCPCB_LOST
)) {
2787 if (!hole
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2792 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2793 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2794 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2796 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2799 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2802 if (tcp_retransmit_skb(sk
, skb
))
2805 NET_INC_STATS_BH(sock_net(sk
), mib_idx
);
2807 if (tcp_in_cwnd_reduction(sk
))
2808 tp
->prr_out
+= tcp_skb_pcount(skb
);
2810 if (skb
== tcp_write_queue_head(sk
))
2811 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2812 inet_csk(sk
)->icsk_rto
,
2817 /* We allow to exceed memory limits for FIN packets to expedite
2818 * connection tear down and (memory) recovery.
2819 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2820 * or even be forced to close flow without any FIN.
2821 * In general, we want to allow one skb per socket to avoid hangs
2822 * with edge trigger epoll()
2824 void sk_forced_mem_schedule(struct sock
*sk
, int size
)
2828 if (size
<= sk
->sk_forward_alloc
)
2830 amt
= sk_mem_pages(size
);
2831 sk
->sk_forward_alloc
+= amt
* SK_MEM_QUANTUM
;
2832 sk_memory_allocated_add(sk
, amt
, &status
);
2835 /* Send a FIN. The caller locks the socket for us.
2836 * We should try to send a FIN packet really hard, but eventually give up.
2838 void tcp_send_fin(struct sock
*sk
)
2840 struct sk_buff
*skb
, *tskb
= tcp_write_queue_tail(sk
);
2841 struct tcp_sock
*tp
= tcp_sk(sk
);
2843 /* Optimization, tack on the FIN if we have one skb in write queue and
2844 * this skb was not yet sent, or we are under memory pressure.
2845 * Note: in the latter case, FIN packet will be sent after a timeout,
2846 * as TCP stack thinks it has already been transmitted.
2848 if (tskb
&& (tcp_send_head(sk
) || tcp_under_memory_pressure(sk
))) {
2850 TCP_SKB_CB(tskb
)->tcp_flags
|= TCPHDR_FIN
;
2851 TCP_SKB_CB(tskb
)->end_seq
++;
2853 if (!tcp_send_head(sk
)) {
2854 /* This means tskb was already sent.
2855 * Pretend we included the FIN on previous transmit.
2856 * We need to set tp->snd_nxt to the value it would have
2857 * if FIN had been sent. This is because retransmit path
2858 * does not change tp->snd_nxt.
2864 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, sk
->sk_allocation
);
2865 if (unlikely(!skb
)) {
2870 skb_reserve(skb
, MAX_TCP_HEADER
);
2871 sk_forced_mem_schedule(sk
, skb
->truesize
);
2872 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2873 tcp_init_nondata_skb(skb
, tp
->write_seq
,
2874 TCPHDR_ACK
| TCPHDR_FIN
);
2875 tcp_queue_skb(sk
, skb
);
2877 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
), TCP_NAGLE_OFF
);
2880 /* We get here when a process closes a file descriptor (either due to
2881 * an explicit close() or as a byproduct of exit()'ing) and there
2882 * was unread data in the receive queue. This behavior is recommended
2883 * by RFC 2525, section 2.17. -DaveM
2885 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
2887 struct sk_buff
*skb
;
2889 /* NOTE: No TCP options attached and we never retransmit this. */
2890 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
2892 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2896 /* Reserve space for headers and prepare control bits. */
2897 skb_reserve(skb
, MAX_TCP_HEADER
);
2898 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
2899 TCPHDR_ACK
| TCPHDR_RST
);
2901 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
2902 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
2904 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
2907 /* Send a crossed SYN-ACK during socket establishment.
2908 * WARNING: This routine must only be called when we have already sent
2909 * a SYN packet that crossed the incoming SYN that caused this routine
2910 * to get called. If this assumption fails then the initial rcv_wnd
2911 * and rcv_wscale values will not be correct.
2913 int tcp_send_synack(struct sock
*sk
)
2915 struct sk_buff
*skb
;
2917 skb
= tcp_write_queue_head(sk
);
2918 if (!skb
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
2919 pr_debug("%s: wrong queue state\n", __func__
);
2922 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
2923 if (skb_cloned(skb
)) {
2924 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
2927 tcp_unlink_write_queue(skb
, sk
);
2928 __skb_header_release(nskb
);
2929 __tcp_add_write_queue_head(sk
, nskb
);
2930 sk_wmem_free_skb(sk
, skb
);
2931 sk
->sk_wmem_queued
+= nskb
->truesize
;
2932 sk_mem_charge(sk
, nskb
->truesize
);
2936 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
2937 tcp_ecn_send_synack(sk
, skb
);
2939 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2943 * tcp_make_synack - Prepare a SYN-ACK.
2944 * sk: listener socket
2945 * dst: dst entry attached to the SYNACK
2946 * req: request_sock pointer
2948 * Allocate one skb and build a SYNACK packet.
2949 * @dst is consumed : Caller should not use it again.
2951 struct sk_buff
*tcp_make_synack(struct sock
*sk
, struct dst_entry
*dst
,
2952 struct request_sock
*req
,
2953 struct tcp_fastopen_cookie
*foc
)
2955 struct tcp_out_options opts
;
2956 struct inet_request_sock
*ireq
= inet_rsk(req
);
2957 struct tcp_sock
*tp
= tcp_sk(sk
);
2959 struct sk_buff
*skb
;
2960 struct tcp_md5sig_key
*md5
= NULL
;
2961 int tcp_header_size
;
2964 skb
= sock_wmalloc(sk
, MAX_TCP_HEADER
, 1, GFP_ATOMIC
);
2965 if (unlikely(!skb
)) {
2969 /* Reserve space for headers. */
2970 skb_reserve(skb
, MAX_TCP_HEADER
);
2972 skb_dst_set(skb
, dst
);
2974 mss
= dst_metric_advmss(dst
);
2975 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< mss
)
2976 mss
= tp
->rx_opt
.user_mss
;
2978 memset(&opts
, 0, sizeof(opts
));
2979 #ifdef CONFIG_SYN_COOKIES
2980 if (unlikely(req
->cookie_ts
))
2981 skb
->skb_mstamp
.stamp_jiffies
= cookie_init_timestamp(req
);
2984 skb_mstamp_get(&skb
->skb_mstamp
);
2986 #ifdef CONFIG_TCP_MD5SIG
2988 md5
= tcp_rsk(req
)->af_specific
->req_md5_lookup(sk
, req_to_sk(req
));
2990 tcp_header_size
= tcp_synack_options(sk
, req
, mss
, skb
, &opts
, md5
,
2993 skb_push(skb
, tcp_header_size
);
2994 skb_reset_transport_header(skb
);
2997 memset(th
, 0, sizeof(struct tcphdr
));
3000 tcp_ecn_make_synack(req
, th
, sk
);
3001 th
->source
= htons(ireq
->ir_num
);
3002 th
->dest
= ireq
->ir_rmt_port
;
3003 /* Setting of flags are superfluous here for callers (and ECE is
3004 * not even correctly set)
3006 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
3007 TCPHDR_SYN
| TCPHDR_ACK
);
3009 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
3010 /* XXX data is queued and acked as is. No buffer/window check */
3011 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
3013 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3014 th
->window
= htons(min(req
->rcv_wnd
, 65535U));
3015 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
3016 th
->doff
= (tcp_header_size
>> 2);
3017 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_OUTSEGS
);
3019 #ifdef CONFIG_TCP_MD5SIG
3020 /* Okay, we have all we need - do the md5 hash if needed */
3022 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
3023 md5
, req_to_sk(req
), skb
);
3027 /* Do not fool tcpdump (if any), clean our debris */
3028 skb
->tstamp
.tv64
= 0;
3031 EXPORT_SYMBOL(tcp_make_synack
);
3033 static void tcp_ca_dst_init(struct sock
*sk
, const struct dst_entry
*dst
)
3035 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3036 const struct tcp_congestion_ops
*ca
;
3037 u32 ca_key
= dst_metric(dst
, RTAX_CC_ALGO
);
3039 if (ca_key
== TCP_CA_UNSPEC
)
3043 ca
= tcp_ca_find_key(ca_key
);
3044 if (likely(ca
&& try_module_get(ca
->owner
))) {
3045 module_put(icsk
->icsk_ca_ops
->owner
);
3046 icsk
->icsk_ca_dst_locked
= tcp_ca_dst_locked(dst
);
3047 icsk
->icsk_ca_ops
= ca
;
3052 /* Do all connect socket setups that can be done AF independent. */
3053 static void tcp_connect_init(struct sock
*sk
)
3055 const struct dst_entry
*dst
= __sk_dst_get(sk
);
3056 struct tcp_sock
*tp
= tcp_sk(sk
);
3059 /* We'll fix this up when we get a response from the other end.
3060 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3062 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
3063 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
3065 #ifdef CONFIG_TCP_MD5SIG
3066 if (tp
->af_specific
->md5_lookup(sk
, sk
))
3067 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
3070 /* If user gave his TCP_MAXSEG, record it to clamp */
3071 if (tp
->rx_opt
.user_mss
)
3072 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3075 tcp_sync_mss(sk
, dst_mtu(dst
));
3077 tcp_ca_dst_init(sk
, dst
);
3079 if (!tp
->window_clamp
)
3080 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
3081 tp
->advmss
= dst_metric_advmss(dst
);
3082 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
3083 tp
->advmss
= tp
->rx_opt
.user_mss
;
3085 tcp_initialize_rcv_mss(sk
);
3087 /* limit the window selection if the user enforce a smaller rx buffer */
3088 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
3089 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
3090 tp
->window_clamp
= tcp_full_space(sk
);
3092 tcp_select_initial_window(tcp_full_space(sk
),
3093 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
3096 sysctl_tcp_window_scaling
,
3098 dst_metric(dst
, RTAX_INITRWND
));
3100 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
3101 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
3104 sock_reset_flag(sk
, SOCK_DONE
);
3107 tp
->snd_una
= tp
->write_seq
;
3108 tp
->snd_sml
= tp
->write_seq
;
3109 tp
->snd_up
= tp
->write_seq
;
3110 tp
->snd_nxt
= tp
->write_seq
;
3112 if (likely(!tp
->repair
))
3115 tp
->rcv_tstamp
= tcp_time_stamp
;
3116 tp
->rcv_wup
= tp
->rcv_nxt
;
3117 tp
->copied_seq
= tp
->rcv_nxt
;
3119 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
3120 inet_csk(sk
)->icsk_retransmits
= 0;
3121 tcp_clear_retrans(tp
);
3124 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
3126 struct tcp_sock
*tp
= tcp_sk(sk
);
3127 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
3129 tcb
->end_seq
+= skb
->len
;
3130 __skb_header_release(skb
);
3131 __tcp_add_write_queue_tail(sk
, skb
);
3132 sk
->sk_wmem_queued
+= skb
->truesize
;
3133 sk_mem_charge(sk
, skb
->truesize
);
3134 tp
->write_seq
= tcb
->end_seq
;
3135 tp
->packets_out
+= tcp_skb_pcount(skb
);
3138 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3139 * queue a data-only packet after the regular SYN, such that regular SYNs
3140 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3141 * only the SYN sequence, the data are retransmitted in the first ACK.
3142 * If cookie is not cached or other error occurs, falls back to send a
3143 * regular SYN with Fast Open cookie request option.
3145 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
3147 struct tcp_sock
*tp
= tcp_sk(sk
);
3148 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
3149 int syn_loss
= 0, space
, err
= 0, copied
;
3150 unsigned long last_syn_loss
= 0;
3151 struct sk_buff
*syn_data
;
3153 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
3154 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
3155 &syn_loss
, &last_syn_loss
);
3156 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3158 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
3159 fo
->cookie
.len
= -1;
3163 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
3164 fo
->cookie
.len
= -1;
3165 else if (fo
->cookie
.len
<= 0)
3168 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3169 * user-MSS. Reserve maximum option space for middleboxes that add
3170 * private TCP options. The cost is reduced data space in SYN :(
3172 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
3173 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3174 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
3175 MAX_TCP_OPTION_SPACE
;
3177 space
= min_t(size_t, space
, fo
->size
);
3179 /* limit to order-0 allocations */
3180 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
3182 syn_data
= sk_stream_alloc_skb(sk
, space
, sk
->sk_allocation
, false);
3185 syn_data
->ip_summed
= CHECKSUM_PARTIAL
;
3186 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
3187 copied
= copy_from_iter(skb_put(syn_data
, space
), space
,
3188 &fo
->data
->msg_iter
);
3189 if (unlikely(!copied
)) {
3190 kfree_skb(syn_data
);
3193 if (copied
!= space
) {
3194 skb_trim(syn_data
, copied
);
3198 /* No more data pending in inet_wait_for_connect() */
3199 if (space
== fo
->size
)
3203 tcp_connect_queue_skb(sk
, syn_data
);
3205 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
3207 syn
->skb_mstamp
= syn_data
->skb_mstamp
;
3209 /* Now full SYN+DATA was cloned and sent (or not),
3210 * remove the SYN from the original skb (syn_data)
3211 * we keep in write queue in case of a retransmit, as we
3212 * also have the SYN packet (with no data) in the same queue.
3214 TCP_SKB_CB(syn_data
)->seq
++;
3215 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
3217 tp
->syn_data
= (fo
->copied
> 0);
3218 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3223 /* Send a regular SYN with Fast Open cookie request option */
3224 if (fo
->cookie
.len
> 0)
3226 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3228 tp
->syn_fastopen
= 0;
3230 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3234 /* Build a SYN and send it off. */
3235 int tcp_connect(struct sock
*sk
)
3237 struct tcp_sock
*tp
= tcp_sk(sk
);
3238 struct sk_buff
*buff
;
3241 tcp_connect_init(sk
);
3243 if (unlikely(tp
->repair
)) {
3244 tcp_finish_connect(sk
, NULL
);
3248 buff
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
, true);
3249 if (unlikely(!buff
))
3252 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3253 tp
->retrans_stamp
= tcp_time_stamp
;
3254 tcp_connect_queue_skb(sk
, buff
);
3255 tcp_ecn_send_syn(sk
, buff
);
3257 /* Send off SYN; include data in Fast Open. */
3258 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3259 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3260 if (err
== -ECONNREFUSED
)
3263 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3264 * in order to make this packet get counted in tcpOutSegs.
3266 tp
->snd_nxt
= tp
->write_seq
;
3267 tp
->pushed_seq
= tp
->write_seq
;
3268 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3270 /* Timer for repeating the SYN until an answer. */
3271 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3272 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3275 EXPORT_SYMBOL(tcp_connect
);
3277 /* Send out a delayed ack, the caller does the policy checking
3278 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3281 void tcp_send_delayed_ack(struct sock
*sk
)
3283 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3284 int ato
= icsk
->icsk_ack
.ato
;
3285 unsigned long timeout
;
3287 tcp_ca_event(sk
, CA_EVENT_DELAYED_ACK
);
3289 if (ato
> TCP_DELACK_MIN
) {
3290 const struct tcp_sock
*tp
= tcp_sk(sk
);
3291 int max_ato
= HZ
/ 2;
3293 if (icsk
->icsk_ack
.pingpong
||
3294 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3295 max_ato
= TCP_DELACK_MAX
;
3297 /* Slow path, intersegment interval is "high". */
3299 /* If some rtt estimate is known, use it to bound delayed ack.
3300 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3304 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3311 ato
= min(ato
, max_ato
);
3314 /* Stay within the limit we were given */
3315 timeout
= jiffies
+ ato
;
3317 /* Use new timeout only if there wasn't a older one earlier. */
3318 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3319 /* If delack timer was blocked or is about to expire,
3322 if (icsk
->icsk_ack
.blocked
||
3323 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3328 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3329 timeout
= icsk
->icsk_ack
.timeout
;
3331 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3332 icsk
->icsk_ack
.timeout
= timeout
;
3333 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3336 /* This routine sends an ack and also updates the window. */
3337 void tcp_send_ack(struct sock
*sk
)
3339 struct sk_buff
*buff
;
3341 /* If we have been reset, we may not send again. */
3342 if (sk
->sk_state
== TCP_CLOSE
)
3345 tcp_ca_event(sk
, CA_EVENT_NON_DELAYED_ACK
);
3347 /* We are not putting this on the write queue, so
3348 * tcp_transmit_skb() will set the ownership to this
3351 buff
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3353 inet_csk_schedule_ack(sk
);
3354 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3355 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3356 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3360 /* Reserve space for headers and prepare control bits. */
3361 skb_reserve(buff
, MAX_TCP_HEADER
);
3362 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3364 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3366 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3367 * We also avoid tcp_wfree() overhead (cache line miss accessing
3368 * tp->tsq_flags) by using regular sock_wfree()
3370 skb_set_tcp_pure_ack(buff
);
3372 /* Send it off, this clears delayed acks for us. */
3373 skb_mstamp_get(&buff
->skb_mstamp
);
3374 tcp_transmit_skb(sk
, buff
, 0, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3376 EXPORT_SYMBOL_GPL(tcp_send_ack
);
3378 /* This routine sends a packet with an out of date sequence
3379 * number. It assumes the other end will try to ack it.
3381 * Question: what should we make while urgent mode?
3382 * 4.4BSD forces sending single byte of data. We cannot send
3383 * out of window data, because we have SND.NXT==SND.MAX...
3385 * Current solution: to send TWO zero-length segments in urgent mode:
3386 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3387 * out-of-date with SND.UNA-1 to probe window.
3389 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
, int mib
)
3391 struct tcp_sock
*tp
= tcp_sk(sk
);
3392 struct sk_buff
*skb
;
3394 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3395 skb
= alloc_skb(MAX_TCP_HEADER
, sk_gfp_atomic(sk
, GFP_ATOMIC
));
3399 /* Reserve space for headers and set control bits. */
3400 skb_reserve(skb
, MAX_TCP_HEADER
);
3401 /* Use a previous sequence. This should cause the other
3402 * end to send an ack. Don't queue or clone SKB, just
3405 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3406 skb_mstamp_get(&skb
->skb_mstamp
);
3407 NET_INC_STATS_BH(sock_net(sk
), mib
);
3408 return tcp_transmit_skb(sk
, skb
, 0, GFP_ATOMIC
);
3411 void tcp_send_window_probe(struct sock
*sk
)
3413 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3414 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3415 tcp_xmit_probe_skb(sk
, 0, LINUX_MIB_TCPWINPROBE
);
3419 /* Initiate keepalive or window probe from timer. */
3420 int tcp_write_wakeup(struct sock
*sk
, int mib
)
3422 struct tcp_sock
*tp
= tcp_sk(sk
);
3423 struct sk_buff
*skb
;
3425 if (sk
->sk_state
== TCP_CLOSE
)
3428 skb
= tcp_send_head(sk
);
3429 if (skb
&& before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3431 unsigned int mss
= tcp_current_mss(sk
);
3432 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3434 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3435 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3437 /* We are probing the opening of a window
3438 * but the window size is != 0
3439 * must have been a result SWS avoidance ( sender )
3441 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3443 seg_size
= min(seg_size
, mss
);
3444 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3445 if (tcp_fragment(sk
, skb
, seg_size
, mss
, GFP_ATOMIC
))
3447 } else if (!tcp_skb_pcount(skb
))
3448 tcp_set_skb_tso_segs(skb
, mss
);
3450 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3451 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3453 tcp_event_new_data_sent(sk
, skb
);
3456 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3457 tcp_xmit_probe_skb(sk
, 1, mib
);
3458 return tcp_xmit_probe_skb(sk
, 0, mib
);
3462 /* A window probe timeout has occurred. If window is not closed send
3463 * a partial packet else a zero probe.
3465 void tcp_send_probe0(struct sock
*sk
)
3467 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3468 struct tcp_sock
*tp
= tcp_sk(sk
);
3469 unsigned long probe_max
;
3472 err
= tcp_write_wakeup(sk
, LINUX_MIB_TCPWINPROBE
);
3474 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3475 /* Cancel probe timer, if it is not required. */
3476 icsk
->icsk_probes_out
= 0;
3477 icsk
->icsk_backoff
= 0;
3482 if (icsk
->icsk_backoff
< sysctl_tcp_retries2
)
3483 icsk
->icsk_backoff
++;
3484 icsk
->icsk_probes_out
++;
3485 probe_max
= TCP_RTO_MAX
;
3487 /* If packet was not sent due to local congestion,
3488 * do not backoff and do not remember icsk_probes_out.
3489 * Let local senders to fight for local resources.
3491 * Use accumulated backoff yet.
3493 if (!icsk
->icsk_probes_out
)
3494 icsk
->icsk_probes_out
= 1;
3495 probe_max
= TCP_RESOURCE_PROBE_INTERVAL
;
3497 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3498 tcp_probe0_when(sk
, probe_max
),
3502 int tcp_rtx_synack(struct sock
*sk
, struct request_sock
*req
)
3504 const struct tcp_request_sock_ops
*af_ops
= tcp_rsk(req
)->af_specific
;
3508 res
= af_ops
->send_synack(sk
, NULL
, &fl
, req
, 0, NULL
);
3510 TCP_INC_STATS_BH(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
3511 NET_INC_STATS_BH(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
3515 EXPORT_SYMBOL(tcp_rtx_synack
);