1 // SPDX-License-Identifier: GPL-2.0-only
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
38 #define pr_fmt(fmt) "TCP: " fmt
42 #include <linux/compiler.h>
43 #include <linux/gfp.h>
44 #include <linux/module.h>
45 #include <linux/static_key.h>
47 #include <trace/events/tcp.h>
49 /* Refresh clocks of a TCP socket,
50 * ensuring monotically increasing values.
52 void tcp_mstamp_refresh(struct tcp_sock
*tp
)
54 u64 val
= tcp_clock_ns();
56 tp
->tcp_clock_cache
= val
;
57 tp
->tcp_mstamp
= div_u64(val
, NSEC_PER_USEC
);
60 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
61 int push_one
, gfp_t gfp
);
63 /* Account for new data that has been sent to the network. */
64 static void tcp_event_new_data_sent(struct sock
*sk
, struct sk_buff
*skb
)
66 struct inet_connection_sock
*icsk
= inet_csk(sk
);
67 struct tcp_sock
*tp
= tcp_sk(sk
);
68 unsigned int prior_packets
= tp
->packets_out
;
70 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
72 __skb_unlink(skb
, &sk
->sk_write_queue
);
73 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, skb
);
75 tp
->packets_out
+= tcp_skb_pcount(skb
);
76 if (!prior_packets
|| icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
79 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
,
83 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
84 * window scaling factor due to loss of precision.
85 * If window has been shrunk, what should we make? It is not clear at all.
86 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
87 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
88 * invalid. OK, let's make this for now:
90 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
92 const struct tcp_sock
*tp
= tcp_sk(sk
);
94 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
) ||
95 (tp
->rx_opt
.wscale_ok
&&
96 ((tp
->snd_nxt
- tcp_wnd_end(tp
)) < (1 << tp
->rx_opt
.rcv_wscale
))))
99 return tcp_wnd_end(tp
);
102 /* Calculate mss to advertise in SYN segment.
103 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
105 * 1. It is independent of path mtu.
106 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
107 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
108 * attached devices, because some buggy hosts are confused by
110 * 4. We do not make 3, we advertise MSS, calculated from first
111 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
112 * This may be overridden via information stored in routing table.
113 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
114 * probably even Jumbo".
116 static __u16
tcp_advertise_mss(struct sock
*sk
)
118 struct tcp_sock
*tp
= tcp_sk(sk
);
119 const struct dst_entry
*dst
= __sk_dst_get(sk
);
120 int mss
= tp
->advmss
;
123 unsigned int metric
= dst_metric_advmss(dst
);
134 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
135 * This is the first part of cwnd validation mechanism.
137 void tcp_cwnd_restart(struct sock
*sk
, s32 delta
)
139 struct tcp_sock
*tp
= tcp_sk(sk
);
140 u32 restart_cwnd
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
141 u32 cwnd
= tp
->snd_cwnd
;
143 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
145 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
146 restart_cwnd
= min(restart_cwnd
, cwnd
);
148 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
150 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
151 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
152 tp
->snd_cwnd_used
= 0;
155 /* Congestion state accounting after a packet has been sent. */
156 static void tcp_event_data_sent(struct tcp_sock
*tp
,
159 struct inet_connection_sock
*icsk
= inet_csk(sk
);
160 const u32 now
= tcp_jiffies32
;
162 if (tcp_packets_in_flight(tp
) == 0)
163 tcp_ca_event(sk
, CA_EVENT_TX_START
);
165 /* If this is the first data packet sent in response to the
166 * previous received data,
167 * and it is a reply for ato after last received packet,
168 * increase pingpong count.
170 if (before(tp
->lsndtime
, icsk
->icsk_ack
.lrcvtime
) &&
171 (u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
172 inet_csk_inc_pingpong_cnt(sk
);
177 /* Account for an ACK we sent. */
178 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
,
181 struct tcp_sock
*tp
= tcp_sk(sk
);
183 if (unlikely(tp
->compressed_ack
> TCP_FASTRETRANS_THRESH
)) {
184 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPACKCOMPRESSED
,
185 tp
->compressed_ack
- TCP_FASTRETRANS_THRESH
);
186 tp
->compressed_ack
= TCP_FASTRETRANS_THRESH
;
187 if (hrtimer_try_to_cancel(&tp
->compressed_ack_timer
) == 1)
191 if (unlikely(rcv_nxt
!= tp
->rcv_nxt
))
192 return; /* Special ACK sent by DCTCP to reflect ECN */
193 tcp_dec_quickack_mode(sk
, pkts
);
194 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
197 /* Determine a window scaling and initial window to offer.
198 * Based on the assumption that the given amount of space
199 * will be offered. Store the results in the tp structure.
200 * NOTE: for smooth operation initial space offering should
201 * be a multiple of mss if possible. We assume here that mss >= 1.
202 * This MUST be enforced by all callers.
204 void tcp_select_initial_window(const struct sock
*sk
, int __space
, __u32 mss
,
205 __u32
*rcv_wnd
, __u32
*window_clamp
,
206 int wscale_ok
, __u8
*rcv_wscale
,
209 unsigned int space
= (__space
< 0 ? 0 : __space
);
211 /* If no clamp set the clamp to the max possible scaled window */
212 if (*window_clamp
== 0)
213 (*window_clamp
) = (U16_MAX
<< TCP_MAX_WSCALE
);
214 space
= min(*window_clamp
, space
);
216 /* Quantize space offering to a multiple of mss if possible. */
218 space
= rounddown(space
, mss
);
220 /* NOTE: offering an initial window larger than 32767
221 * will break some buggy TCP stacks. If the admin tells us
222 * it is likely we could be speaking with such a buggy stack
223 * we will truncate our initial window offering to 32K-1
224 * unless the remote has sent us a window scaling option,
225 * which we interpret as a sign the remote TCP is not
226 * misinterpreting the window field as a signed quantity.
228 if (sock_net(sk
)->ipv4
.sysctl_tcp_workaround_signed_windows
)
229 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
231 (*rcv_wnd
) = min_t(u32
, space
, U16_MAX
);
234 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
238 /* Set window scaling on max possible window */
239 space
= max_t(u32
, space
, sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[2]);
240 space
= max_t(u32
, space
, sysctl_rmem_max
);
241 space
= min_t(u32
, space
, *window_clamp
);
242 *rcv_wscale
= clamp_t(int, ilog2(space
) - 15,
245 /* Set the clamp no higher than max representable value */
246 (*window_clamp
) = min_t(__u32
, U16_MAX
<< (*rcv_wscale
), *window_clamp
);
248 EXPORT_SYMBOL(tcp_select_initial_window
);
250 /* Chose a new window to advertise, update state in tcp_sock for the
251 * socket, and return result with RFC1323 scaling applied. The return
252 * value can be stuffed directly into th->window for an outgoing
255 static u16
tcp_select_window(struct sock
*sk
)
257 struct tcp_sock
*tp
= tcp_sk(sk
);
258 u32 old_win
= tp
->rcv_wnd
;
259 u32 cur_win
= tcp_receive_window(tp
);
260 u32 new_win
= __tcp_select_window(sk
);
262 /* Never shrink the offered window */
263 if (new_win
< cur_win
) {
264 /* Danger Will Robinson!
265 * Don't update rcv_wup/rcv_wnd here or else
266 * we will not be able to advertise a zero
267 * window in time. --DaveM
269 * Relax Will Robinson.
272 NET_INC_STATS(sock_net(sk
),
273 LINUX_MIB_TCPWANTZEROWINDOWADV
);
274 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
276 tp
->rcv_wnd
= new_win
;
277 tp
->rcv_wup
= tp
->rcv_nxt
;
279 /* Make sure we do not exceed the maximum possible
282 if (!tp
->rx_opt
.rcv_wscale
&&
283 sock_net(sk
)->ipv4
.sysctl_tcp_workaround_signed_windows
)
284 new_win
= min(new_win
, MAX_TCP_WINDOW
);
286 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
288 /* RFC1323 scaling applied */
289 new_win
>>= tp
->rx_opt
.rcv_wscale
;
291 /* If we advertise zero window, disable fast path. */
295 NET_INC_STATS(sock_net(sk
),
296 LINUX_MIB_TCPTOZEROWINDOWADV
);
297 } else if (old_win
== 0) {
298 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFROMZEROWINDOWADV
);
304 /* Packet ECN state for a SYN-ACK */
305 static void tcp_ecn_send_synack(struct sock
*sk
, struct sk_buff
*skb
)
307 const struct tcp_sock
*tp
= tcp_sk(sk
);
309 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
310 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
311 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
312 else if (tcp_ca_needs_ecn(sk
) ||
313 tcp_bpf_ca_needs_ecn(sk
))
317 /* Packet ECN state for a SYN. */
318 static void tcp_ecn_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
320 struct tcp_sock
*tp
= tcp_sk(sk
);
321 bool bpf_needs_ecn
= tcp_bpf_ca_needs_ecn(sk
);
322 bool use_ecn
= sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1 ||
323 tcp_ca_needs_ecn(sk
) || bpf_needs_ecn
;
326 const struct dst_entry
*dst
= __sk_dst_get(sk
);
328 if (dst
&& dst_feature(dst
, RTAX_FEATURE_ECN
))
335 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
336 tp
->ecn_flags
= TCP_ECN_OK
;
337 if (tcp_ca_needs_ecn(sk
) || bpf_needs_ecn
)
342 static void tcp_ecn_clear_syn(struct sock
*sk
, struct sk_buff
*skb
)
344 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn_fallback
)
345 /* tp->ecn_flags are cleared at a later point in time when
346 * SYN ACK is ultimatively being received.
348 TCP_SKB_CB(skb
)->tcp_flags
&= ~(TCPHDR_ECE
| TCPHDR_CWR
);
352 tcp_ecn_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
354 if (inet_rsk(req
)->ecn_ok
)
358 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
361 static void tcp_ecn_send(struct sock
*sk
, struct sk_buff
*skb
,
362 struct tcphdr
*th
, int tcp_header_len
)
364 struct tcp_sock
*tp
= tcp_sk(sk
);
366 if (tp
->ecn_flags
& TCP_ECN_OK
) {
367 /* Not-retransmitted data segment: set ECT and inject CWR. */
368 if (skb
->len
!= tcp_header_len
&&
369 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
371 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
372 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
374 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
376 } else if (!tcp_ca_needs_ecn(sk
)) {
377 /* ACK or retransmitted segment: clear ECT|CE */
378 INET_ECN_dontxmit(sk
);
380 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
385 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
386 * auto increment end seqno.
388 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
390 skb
->ip_summed
= CHECKSUM_PARTIAL
;
392 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
393 TCP_SKB_CB(skb
)->sacked
= 0;
395 tcp_skb_pcount_set(skb
, 1);
397 TCP_SKB_CB(skb
)->seq
= seq
;
398 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
400 TCP_SKB_CB(skb
)->end_seq
= seq
;
403 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
405 return tp
->snd_una
!= tp
->snd_up
;
408 #define OPTION_SACK_ADVERTISE (1 << 0)
409 #define OPTION_TS (1 << 1)
410 #define OPTION_MD5 (1 << 2)
411 #define OPTION_WSCALE (1 << 3)
412 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
413 #define OPTION_SMC (1 << 9)
415 static void smc_options_write(__be32
*ptr
, u16
*options
)
417 #if IS_ENABLED(CONFIG_SMC)
418 if (static_branch_unlikely(&tcp_have_smc
)) {
419 if (unlikely(OPTION_SMC
& *options
)) {
420 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
423 (TCPOLEN_EXP_SMC_BASE
));
424 *ptr
++ = htonl(TCPOPT_SMC_MAGIC
);
430 struct tcp_out_options
{
431 u16 options
; /* bit field of OPTION_* */
432 u16 mss
; /* 0 to disable */
433 u8 ws
; /* window scale, 0 to disable */
434 u8 num_sack_blocks
; /* number of SACK blocks to include */
435 u8 hash_size
; /* bytes in hash_location */
436 __u8
*hash_location
; /* temporary pointer, overloaded */
437 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
438 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
441 /* Write previously computed TCP options to the packet.
443 * Beware: Something in the Internet is very sensitive to the ordering of
444 * TCP options, we learned this through the hard way, so be careful here.
445 * Luckily we can at least blame others for their non-compliance but from
446 * inter-operability perspective it seems that we're somewhat stuck with
447 * the ordering which we have been using if we want to keep working with
448 * those broken things (not that it currently hurts anybody as there isn't
449 * particular reason why the ordering would need to be changed).
451 * At least SACK_PERM as the first option is known to lead to a disaster
452 * (but it may well be that other scenarios fail similarly).
454 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
455 struct tcp_out_options
*opts
)
457 u16 options
= opts
->options
; /* mungable copy */
459 if (unlikely(OPTION_MD5
& options
)) {
460 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
461 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
462 /* overload cookie hash location */
463 opts
->hash_location
= (__u8
*)ptr
;
467 if (unlikely(opts
->mss
)) {
468 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
469 (TCPOLEN_MSS
<< 16) |
473 if (likely(OPTION_TS
& options
)) {
474 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
475 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
476 (TCPOLEN_SACK_PERM
<< 16) |
477 (TCPOPT_TIMESTAMP
<< 8) |
479 options
&= ~OPTION_SACK_ADVERTISE
;
481 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
483 (TCPOPT_TIMESTAMP
<< 8) |
486 *ptr
++ = htonl(opts
->tsval
);
487 *ptr
++ = htonl(opts
->tsecr
);
490 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
491 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
493 (TCPOPT_SACK_PERM
<< 8) |
497 if (unlikely(OPTION_WSCALE
& options
)) {
498 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
499 (TCPOPT_WINDOW
<< 16) |
500 (TCPOLEN_WINDOW
<< 8) |
504 if (unlikely(opts
->num_sack_blocks
)) {
505 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
506 tp
->duplicate_sack
: tp
->selective_acks
;
509 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
512 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
513 TCPOLEN_SACK_PERBLOCK
)));
515 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
517 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
518 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
521 tp
->rx_opt
.dsack
= 0;
524 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
525 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
527 u32 len
; /* Fast Open option length */
530 len
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
531 *ptr
= htonl((TCPOPT_EXP
<< 24) | (len
<< 16) |
532 TCPOPT_FASTOPEN_MAGIC
);
533 p
+= TCPOLEN_EXP_FASTOPEN_BASE
;
535 len
= TCPOLEN_FASTOPEN_BASE
+ foc
->len
;
536 *p
++ = TCPOPT_FASTOPEN
;
540 memcpy(p
, foc
->val
, foc
->len
);
541 if ((len
& 3) == 2) {
542 p
[foc
->len
] = TCPOPT_NOP
;
543 p
[foc
->len
+ 1] = TCPOPT_NOP
;
545 ptr
+= (len
+ 3) >> 2;
548 smc_options_write(ptr
, &options
);
551 static void smc_set_option(const struct tcp_sock
*tp
,
552 struct tcp_out_options
*opts
,
553 unsigned int *remaining
)
555 #if IS_ENABLED(CONFIG_SMC)
556 if (static_branch_unlikely(&tcp_have_smc
)) {
558 if (*remaining
>= TCPOLEN_EXP_SMC_BASE_ALIGNED
) {
559 opts
->options
|= OPTION_SMC
;
560 *remaining
-= TCPOLEN_EXP_SMC_BASE_ALIGNED
;
567 static void smc_set_option_cond(const struct tcp_sock
*tp
,
568 const struct inet_request_sock
*ireq
,
569 struct tcp_out_options
*opts
,
570 unsigned int *remaining
)
572 #if IS_ENABLED(CONFIG_SMC)
573 if (static_branch_unlikely(&tcp_have_smc
)) {
574 if (tp
->syn_smc
&& ireq
->smc_ok
) {
575 if (*remaining
>= TCPOLEN_EXP_SMC_BASE_ALIGNED
) {
576 opts
->options
|= OPTION_SMC
;
577 *remaining
-= TCPOLEN_EXP_SMC_BASE_ALIGNED
;
584 /* Compute TCP options for SYN packets. This is not the final
585 * network wire format yet.
587 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
588 struct tcp_out_options
*opts
,
589 struct tcp_md5sig_key
**md5
)
591 struct tcp_sock
*tp
= tcp_sk(sk
);
592 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
593 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
596 #ifdef CONFIG_TCP_MD5SIG
597 if (static_branch_unlikely(&tcp_md5_needed
) &&
598 rcu_access_pointer(tp
->md5sig_info
)) {
599 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
601 opts
->options
|= OPTION_MD5
;
602 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
607 /* We always get an MSS option. The option bytes which will be seen in
608 * normal data packets should timestamps be used, must be in the MSS
609 * advertised. But we subtract them from tp->mss_cache so that
610 * calculations in tcp_sendmsg are simpler etc. So account for this
611 * fact here if necessary. If we don't do this correctly, as a
612 * receiver we won't recognize data packets as being full sized when we
613 * should, and thus we won't abide by the delayed ACK rules correctly.
614 * SACKs don't matter, we never delay an ACK when we have any of those
616 opts
->mss
= tcp_advertise_mss(sk
);
617 remaining
-= TCPOLEN_MSS_ALIGNED
;
619 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_timestamps
&& !*md5
)) {
620 opts
->options
|= OPTION_TS
;
621 opts
->tsval
= tcp_skb_timestamp(skb
) + tp
->tsoffset
;
622 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
623 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
625 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_window_scaling
)) {
626 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
627 opts
->options
|= OPTION_WSCALE
;
628 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
630 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_sack
)) {
631 opts
->options
|= OPTION_SACK_ADVERTISE
;
632 if (unlikely(!(OPTION_TS
& opts
->options
)))
633 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
636 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
637 u32 need
= fastopen
->cookie
.len
;
639 need
+= fastopen
->cookie
.exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
640 TCPOLEN_FASTOPEN_BASE
;
641 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
642 if (remaining
>= need
) {
643 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
644 opts
->fastopen_cookie
= &fastopen
->cookie
;
646 tp
->syn_fastopen
= 1;
647 tp
->syn_fastopen_exp
= fastopen
->cookie
.exp
? 1 : 0;
651 smc_set_option(tp
, opts
, &remaining
);
653 return MAX_TCP_OPTION_SPACE
- remaining
;
656 /* Set up TCP options for SYN-ACKs. */
657 static unsigned int tcp_synack_options(const struct sock
*sk
,
658 struct request_sock
*req
,
659 unsigned int mss
, struct sk_buff
*skb
,
660 struct tcp_out_options
*opts
,
661 const struct tcp_md5sig_key
*md5
,
662 struct tcp_fastopen_cookie
*foc
)
664 struct inet_request_sock
*ireq
= inet_rsk(req
);
665 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
667 #ifdef CONFIG_TCP_MD5SIG
669 opts
->options
|= OPTION_MD5
;
670 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
672 /* We can't fit any SACK blocks in a packet with MD5 + TS
673 * options. There was discussion about disabling SACK
674 * rather than TS in order to fit in better with old,
675 * buggy kernels, but that was deemed to be unnecessary.
677 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
681 /* We always send an MSS option. */
683 remaining
-= TCPOLEN_MSS_ALIGNED
;
685 if (likely(ireq
->wscale_ok
)) {
686 opts
->ws
= ireq
->rcv_wscale
;
687 opts
->options
|= OPTION_WSCALE
;
688 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
690 if (likely(ireq
->tstamp_ok
)) {
691 opts
->options
|= OPTION_TS
;
692 opts
->tsval
= tcp_skb_timestamp(skb
) + tcp_rsk(req
)->ts_off
;
693 opts
->tsecr
= req
->ts_recent
;
694 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
696 if (likely(ireq
->sack_ok
)) {
697 opts
->options
|= OPTION_SACK_ADVERTISE
;
698 if (unlikely(!ireq
->tstamp_ok
))
699 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
701 if (foc
!= NULL
&& foc
->len
>= 0) {
704 need
+= foc
->exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
705 TCPOLEN_FASTOPEN_BASE
;
706 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
707 if (remaining
>= need
) {
708 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
709 opts
->fastopen_cookie
= foc
;
714 smc_set_option_cond(tcp_sk(sk
), ireq
, opts
, &remaining
);
716 return MAX_TCP_OPTION_SPACE
- remaining
;
719 /* Compute TCP options for ESTABLISHED sockets. This is not the
720 * final wire format yet.
722 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
723 struct tcp_out_options
*opts
,
724 struct tcp_md5sig_key
**md5
)
726 struct tcp_sock
*tp
= tcp_sk(sk
);
727 unsigned int size
= 0;
728 unsigned int eff_sacks
;
733 #ifdef CONFIG_TCP_MD5SIG
734 if (static_branch_unlikely(&tcp_md5_needed
) &&
735 rcu_access_pointer(tp
->md5sig_info
)) {
736 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
738 opts
->options
|= OPTION_MD5
;
739 size
+= TCPOLEN_MD5SIG_ALIGNED
;
744 if (likely(tp
->rx_opt
.tstamp_ok
)) {
745 opts
->options
|= OPTION_TS
;
746 opts
->tsval
= skb
? tcp_skb_timestamp(skb
) + tp
->tsoffset
: 0;
747 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
748 size
+= TCPOLEN_TSTAMP_ALIGNED
;
751 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
752 if (unlikely(eff_sacks
)) {
753 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
754 opts
->num_sack_blocks
=
755 min_t(unsigned int, eff_sacks
,
756 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
757 TCPOLEN_SACK_PERBLOCK
);
758 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
759 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
766 /* TCP SMALL QUEUES (TSQ)
768 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
769 * to reduce RTT and bufferbloat.
770 * We do this using a special skb destructor (tcp_wfree).
772 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
773 * needs to be reallocated in a driver.
774 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
776 * Since transmit from skb destructor is forbidden, we use a tasklet
777 * to process all sockets that eventually need to send more skbs.
778 * We use one tasklet per cpu, with its own queue of sockets.
781 struct tasklet_struct tasklet
;
782 struct list_head head
; /* queue of tcp sockets */
784 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
786 static void tcp_tsq_write(struct sock
*sk
)
788 if ((1 << sk
->sk_state
) &
789 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
790 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
)) {
791 struct tcp_sock
*tp
= tcp_sk(sk
);
793 if (tp
->lost_out
> tp
->retrans_out
&&
794 tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) {
795 tcp_mstamp_refresh(tp
);
796 tcp_xmit_retransmit_queue(sk
);
799 tcp_write_xmit(sk
, tcp_current_mss(sk
), tp
->nonagle
,
804 static void tcp_tsq_handler(struct sock
*sk
)
807 if (!sock_owned_by_user(sk
))
809 else if (!test_and_set_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
))
814 * One tasklet per cpu tries to send more skbs.
815 * We run in tasklet context but need to disable irqs when
816 * transferring tsq->head because tcp_wfree() might
817 * interrupt us (non NAPI drivers)
819 static void tcp_tasklet_func(unsigned long data
)
821 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
824 struct list_head
*q
, *n
;
828 local_irq_save(flags
);
829 list_splice_init(&tsq
->head
, &list
);
830 local_irq_restore(flags
);
832 list_for_each_safe(q
, n
, &list
) {
833 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
834 list_del(&tp
->tsq_node
);
836 sk
= (struct sock
*)tp
;
837 smp_mb__before_atomic();
838 clear_bit(TSQ_QUEUED
, &sk
->sk_tsq_flags
);
845 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
846 TCPF_WRITE_TIMER_DEFERRED | \
847 TCPF_DELACK_TIMER_DEFERRED | \
848 TCPF_MTU_REDUCED_DEFERRED)
850 * tcp_release_cb - tcp release_sock() callback
853 * called from release_sock() to perform protocol dependent
854 * actions before socket release.
856 void tcp_release_cb(struct sock
*sk
)
858 unsigned long flags
, nflags
;
860 /* perform an atomic operation only if at least one flag is set */
862 flags
= sk
->sk_tsq_flags
;
863 if (!(flags
& TCP_DEFERRED_ALL
))
865 nflags
= flags
& ~TCP_DEFERRED_ALL
;
866 } while (cmpxchg(&sk
->sk_tsq_flags
, flags
, nflags
) != flags
);
868 if (flags
& TCPF_TSQ_DEFERRED
) {
872 /* Here begins the tricky part :
873 * We are called from release_sock() with :
875 * 2) sk_lock.slock spinlock held
876 * 3) socket owned by us (sk->sk_lock.owned == 1)
878 * But following code is meant to be called from BH handlers,
879 * so we should keep BH disabled, but early release socket ownership
881 sock_release_ownership(sk
);
883 if (flags
& TCPF_WRITE_TIMER_DEFERRED
) {
884 tcp_write_timer_handler(sk
);
887 if (flags
& TCPF_DELACK_TIMER_DEFERRED
) {
888 tcp_delack_timer_handler(sk
);
891 if (flags
& TCPF_MTU_REDUCED_DEFERRED
) {
892 inet_csk(sk
)->icsk_af_ops
->mtu_reduced(sk
);
896 EXPORT_SYMBOL(tcp_release_cb
);
898 void __init
tcp_tasklet_init(void)
902 for_each_possible_cpu(i
) {
903 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
905 INIT_LIST_HEAD(&tsq
->head
);
906 tasklet_init(&tsq
->tasklet
,
913 * Write buffer destructor automatically called from kfree_skb.
914 * We can't xmit new skbs from this context, as we might already
917 void tcp_wfree(struct sk_buff
*skb
)
919 struct sock
*sk
= skb
->sk
;
920 struct tcp_sock
*tp
= tcp_sk(sk
);
921 unsigned long flags
, nval
, oval
;
923 /* Keep one reference on sk_wmem_alloc.
924 * Will be released by sk_free() from here or tcp_tasklet_func()
926 WARN_ON(refcount_sub_and_test(skb
->truesize
- 1, &sk
->sk_wmem_alloc
));
928 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
929 * Wait until our queues (qdisc + devices) are drained.
931 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
932 * - chance for incoming ACK (processed by another cpu maybe)
933 * to migrate this flow (skb->ooo_okay will be eventually set)
935 if (refcount_read(&sk
->sk_wmem_alloc
) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current
)
938 for (oval
= READ_ONCE(sk
->sk_tsq_flags
);; oval
= nval
) {
939 struct tsq_tasklet
*tsq
;
942 if (!(oval
& TSQF_THROTTLED
) || (oval
& TSQF_QUEUED
))
945 nval
= (oval
& ~TSQF_THROTTLED
) | TSQF_QUEUED
;
946 nval
= cmpxchg(&sk
->sk_tsq_flags
, oval
, nval
);
950 /* queue this socket to tasklet queue */
951 local_irq_save(flags
);
952 tsq
= this_cpu_ptr(&tsq_tasklet
);
953 empty
= list_empty(&tsq
->head
);
954 list_add(&tp
->tsq_node
, &tsq
->head
);
956 tasklet_schedule(&tsq
->tasklet
);
957 local_irq_restore(flags
);
964 /* Note: Called under soft irq.
965 * We can call TCP stack right away, unless socket is owned by user.
967 enum hrtimer_restart
tcp_pace_kick(struct hrtimer
*timer
)
969 struct tcp_sock
*tp
= container_of(timer
, struct tcp_sock
, pacing_timer
);
970 struct sock
*sk
= (struct sock
*)tp
;
975 return HRTIMER_NORESTART
;
978 static void tcp_update_skb_after_send(struct sock
*sk
, struct sk_buff
*skb
,
981 struct tcp_sock
*tp
= tcp_sk(sk
);
983 if (sk
->sk_pacing_status
!= SK_PACING_NONE
) {
984 unsigned long rate
= sk
->sk_pacing_rate
;
986 /* Original sch_fq does not pace first 10 MSS
987 * Note that tp->data_segs_out overflows after 2^32 packets,
988 * this is a minor annoyance.
990 if (rate
!= ~0UL && rate
&& tp
->data_segs_out
>= 10) {
991 u64 len_ns
= div64_ul((u64
)skb
->len
* NSEC_PER_SEC
, rate
);
992 u64 credit
= tp
->tcp_wstamp_ns
- prior_wstamp
;
994 /* take into account OS jitter */
995 len_ns
-= min_t(u64
, len_ns
/ 2, credit
);
996 tp
->tcp_wstamp_ns
+= len_ns
;
999 list_move_tail(&skb
->tcp_tsorted_anchor
, &tp
->tsorted_sent_queue
);
1002 /* This routine actually transmits TCP packets queued in by
1003 * tcp_do_sendmsg(). This is used by both the initial
1004 * transmission and possible later retransmissions.
1005 * All SKB's seen here are completely headerless. It is our
1006 * job to build the TCP header, and pass the packet down to
1007 * IP so it can do the same plus pass the packet off to the
1010 * We are working here with either a clone of the original
1011 * SKB, or a fresh unique copy made by the retransmit engine.
1013 static int __tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
,
1014 int clone_it
, gfp_t gfp_mask
, u32 rcv_nxt
)
1016 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1017 struct inet_sock
*inet
;
1018 struct tcp_sock
*tp
;
1019 struct tcp_skb_cb
*tcb
;
1020 struct tcp_out_options opts
;
1021 unsigned int tcp_options_size
, tcp_header_size
;
1022 struct sk_buff
*oskb
= NULL
;
1023 struct tcp_md5sig_key
*md5
;
1028 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
1030 prior_wstamp
= tp
->tcp_wstamp_ns
;
1031 tp
->tcp_wstamp_ns
= max(tp
->tcp_wstamp_ns
, tp
->tcp_clock_cache
);
1032 skb
->skb_mstamp_ns
= tp
->tcp_wstamp_ns
;
1034 TCP_SKB_CB(skb
)->tx
.in_flight
= TCP_SKB_CB(skb
)->end_seq
1038 tcp_skb_tsorted_save(oskb
) {
1039 if (unlikely(skb_cloned(oskb
)))
1040 skb
= pskb_copy(oskb
, gfp_mask
);
1042 skb
= skb_clone(oskb
, gfp_mask
);
1043 } tcp_skb_tsorted_restore(oskb
);
1050 tcb
= TCP_SKB_CB(skb
);
1051 memset(&opts
, 0, sizeof(opts
));
1053 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
1054 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
1056 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
1058 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
1060 /* if no packet is in qdisc/device queue, then allow XPS to select
1061 * another queue. We can be called from tcp_tsq_handler()
1062 * which holds one reference to sk.
1064 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1065 * One way to get this would be to set skb->truesize = 2 on them.
1067 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) < SKB_TRUESIZE(1);
1069 /* If we had to use memory reserve to allocate this skb,
1070 * this might cause drops if packet is looped back :
1071 * Other socket might not have SOCK_MEMALLOC.
1072 * Packets not looped back do not care about pfmemalloc.
1074 skb
->pfmemalloc
= 0;
1076 skb_push(skb
, tcp_header_size
);
1077 skb_reset_transport_header(skb
);
1081 skb
->destructor
= skb_is_tcp_pure_ack(skb
) ? __sock_wfree
: tcp_wfree
;
1082 skb_set_hash_from_sk(skb
, sk
);
1083 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
1085 skb_set_dst_pending_confirm(skb
, sk
->sk_dst_pending_confirm
);
1087 /* Build TCP header and checksum it. */
1088 th
= (struct tcphdr
*)skb
->data
;
1089 th
->source
= inet
->inet_sport
;
1090 th
->dest
= inet
->inet_dport
;
1091 th
->seq
= htonl(tcb
->seq
);
1092 th
->ack_seq
= htonl(rcv_nxt
);
1093 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
1099 /* The urg_mode check is necessary during a below snd_una win probe */
1100 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
1101 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
1102 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
1104 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
1105 th
->urg_ptr
= htons(0xFFFF);
1110 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
1111 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
1112 if (likely(!(tcb
->tcp_flags
& TCPHDR_SYN
))) {
1113 th
->window
= htons(tcp_select_window(sk
));
1114 tcp_ecn_send(sk
, skb
, th
, tcp_header_size
);
1116 /* RFC1323: The window in SYN & SYN/ACK segments
1119 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
1121 #ifdef CONFIG_TCP_MD5SIG
1122 /* Calculate the MD5 hash, as we have all we need now */
1124 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
1125 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
1130 icsk
->icsk_af_ops
->send_check(sk
, skb
);
1132 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
1133 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
), rcv_nxt
);
1135 if (skb
->len
!= tcp_header_size
) {
1136 tcp_event_data_sent(tp
, sk
);
1137 tp
->data_segs_out
+= tcp_skb_pcount(skb
);
1138 tp
->bytes_sent
+= skb
->len
- tcp_header_size
;
1141 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
1142 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
1143 tcp_skb_pcount(skb
));
1145 tp
->segs_out
+= tcp_skb_pcount(skb
);
1146 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1147 skb_shinfo(skb
)->gso_segs
= tcp_skb_pcount(skb
);
1148 skb_shinfo(skb
)->gso_size
= tcp_skb_mss(skb
);
1150 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1152 /* Cleanup our debris for IP stacks */
1153 memset(skb
->cb
, 0, max(sizeof(struct inet_skb_parm
),
1154 sizeof(struct inet6_skb_parm
)));
1156 err
= icsk
->icsk_af_ops
->queue_xmit(sk
, skb
, &inet
->cork
.fl
);
1158 if (unlikely(err
> 0)) {
1160 err
= net_xmit_eval(err
);
1163 tcp_update_skb_after_send(sk
, oskb
, prior_wstamp
);
1164 tcp_rate_skb_sent(sk
, oskb
);
1169 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
1172 return __tcp_transmit_skb(sk
, skb
, clone_it
, gfp_mask
,
1173 tcp_sk(sk
)->rcv_nxt
);
1176 /* This routine just queues the buffer for sending.
1178 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1179 * otherwise socket can stall.
1181 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1183 struct tcp_sock
*tp
= tcp_sk(sk
);
1185 /* Advance write_seq and place onto the write_queue. */
1186 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
1187 __skb_header_release(skb
);
1188 tcp_add_write_queue_tail(sk
, skb
);
1189 sk
->sk_wmem_queued
+= skb
->truesize
;
1190 sk_mem_charge(sk
, skb
->truesize
);
1193 /* Initialize TSO segments for a packet. */
1194 static void tcp_set_skb_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1196 if (skb
->len
<= mss_now
) {
1197 /* Avoid the costly divide in the normal
1200 tcp_skb_pcount_set(skb
, 1);
1201 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1203 tcp_skb_pcount_set(skb
, DIV_ROUND_UP(skb
->len
, mss_now
));
1204 TCP_SKB_CB(skb
)->tcp_gso_size
= mss_now
;
1208 /* Pcount in the middle of the write queue got changed, we need to do various
1209 * tweaks to fix counters
1211 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1213 struct tcp_sock
*tp
= tcp_sk(sk
);
1215 tp
->packets_out
-= decr
;
1217 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1218 tp
->sacked_out
-= decr
;
1219 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1220 tp
->retrans_out
-= decr
;
1221 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1222 tp
->lost_out
-= decr
;
1224 /* Reno case is special. Sigh... */
1225 if (tcp_is_reno(tp
) && decr
> 0)
1226 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1228 if (tp
->lost_skb_hint
&&
1229 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1230 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
1231 tp
->lost_cnt_hint
-= decr
;
1233 tcp_verify_left_out(tp
);
1236 static bool tcp_has_tx_tstamp(const struct sk_buff
*skb
)
1238 return TCP_SKB_CB(skb
)->txstamp_ack
||
1239 (skb_shinfo(skb
)->tx_flags
& SKBTX_ANY_TSTAMP
);
1242 static void tcp_fragment_tstamp(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1244 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1246 if (unlikely(tcp_has_tx_tstamp(skb
)) &&
1247 !before(shinfo
->tskey
, TCP_SKB_CB(skb2
)->seq
)) {
1248 struct skb_shared_info
*shinfo2
= skb_shinfo(skb2
);
1249 u8 tsflags
= shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
1251 shinfo
->tx_flags
&= ~tsflags
;
1252 shinfo2
->tx_flags
|= tsflags
;
1253 swap(shinfo
->tskey
, shinfo2
->tskey
);
1254 TCP_SKB_CB(skb2
)->txstamp_ack
= TCP_SKB_CB(skb
)->txstamp_ack
;
1255 TCP_SKB_CB(skb
)->txstamp_ack
= 0;
1259 static void tcp_skb_fragment_eor(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1261 TCP_SKB_CB(skb2
)->eor
= TCP_SKB_CB(skb
)->eor
;
1262 TCP_SKB_CB(skb
)->eor
= 0;
1265 /* Insert buff after skb on the write or rtx queue of sk. */
1266 static void tcp_insert_write_queue_after(struct sk_buff
*skb
,
1267 struct sk_buff
*buff
,
1269 enum tcp_queue tcp_queue
)
1271 if (tcp_queue
== TCP_FRAG_IN_WRITE_QUEUE
)
1272 __skb_queue_after(&sk
->sk_write_queue
, skb
, buff
);
1274 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, buff
);
1277 /* Function to create two new TCP segments. Shrinks the given segment
1278 * to the specified size and appends a new segment with the rest of the
1279 * packet to the list. This won't be called frequently, I hope.
1280 * Remember, these are still headerless SKBs at this point.
1282 int tcp_fragment(struct sock
*sk
, enum tcp_queue tcp_queue
,
1283 struct sk_buff
*skb
, u32 len
,
1284 unsigned int mss_now
, gfp_t gfp
)
1286 struct tcp_sock
*tp
= tcp_sk(sk
);
1287 struct sk_buff
*buff
;
1288 int nsize
, old_factor
;
1292 if (WARN_ON(len
> skb
->len
))
1295 nsize
= skb_headlen(skb
) - len
;
1299 if (skb_unclone(skb
, gfp
))
1302 /* Get a new skb... force flag on. */
1303 buff
= sk_stream_alloc_skb(sk
, nsize
, gfp
, true);
1305 return -ENOMEM
; /* We'll just try again later. */
1307 sk
->sk_wmem_queued
+= buff
->truesize
;
1308 sk_mem_charge(sk
, buff
->truesize
);
1309 nlen
= skb
->len
- len
- nsize
;
1310 buff
->truesize
+= nlen
;
1311 skb
->truesize
-= nlen
;
1313 /* Correct the sequence numbers. */
1314 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1315 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1316 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1318 /* PSH and FIN should only be set in the second packet. */
1319 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1320 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1321 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1322 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1323 tcp_skb_fragment_eor(skb
, buff
);
1325 skb_split(skb
, buff
, len
);
1327 buff
->ip_summed
= CHECKSUM_PARTIAL
;
1329 buff
->tstamp
= skb
->tstamp
;
1330 tcp_fragment_tstamp(skb
, buff
);
1332 old_factor
= tcp_skb_pcount(skb
);
1334 /* Fix up tso_factor for both original and new SKB. */
1335 tcp_set_skb_tso_segs(skb
, mss_now
);
1336 tcp_set_skb_tso_segs(buff
, mss_now
);
1338 /* Update delivered info for the new segment */
1339 TCP_SKB_CB(buff
)->tx
= TCP_SKB_CB(skb
)->tx
;
1341 /* If this packet has been sent out already, we must
1342 * adjust the various packet counters.
1344 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1345 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1346 tcp_skb_pcount(buff
);
1349 tcp_adjust_pcount(sk
, skb
, diff
);
1352 /* Link BUFF into the send queue. */
1353 __skb_header_release(buff
);
1354 tcp_insert_write_queue_after(skb
, buff
, sk
, tcp_queue
);
1355 if (tcp_queue
== TCP_FRAG_IN_RTX_QUEUE
)
1356 list_add(&buff
->tcp_tsorted_anchor
, &skb
->tcp_tsorted_anchor
);
1361 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1362 * data is not copied, but immediately discarded.
1364 static int __pskb_trim_head(struct sk_buff
*skb
, int len
)
1366 struct skb_shared_info
*shinfo
;
1369 eat
= min_t(int, len
, skb_headlen(skb
));
1371 __skb_pull(skb
, eat
);
1378 shinfo
= skb_shinfo(skb
);
1379 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1380 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1383 skb_frag_unref(skb
, i
);
1386 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1388 shinfo
->frags
[k
].page_offset
+= eat
;
1389 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1395 shinfo
->nr_frags
= k
;
1397 skb
->data_len
-= len
;
1398 skb
->len
= skb
->data_len
;
1402 /* Remove acked data from a packet in the transmit queue. */
1403 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1407 if (skb_unclone(skb
, GFP_ATOMIC
))
1410 delta_truesize
= __pskb_trim_head(skb
, len
);
1412 TCP_SKB_CB(skb
)->seq
+= len
;
1413 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1415 if (delta_truesize
) {
1416 skb
->truesize
-= delta_truesize
;
1417 sk
->sk_wmem_queued
-= delta_truesize
;
1418 sk_mem_uncharge(sk
, delta_truesize
);
1419 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1422 /* Any change of skb->len requires recalculation of tso factor. */
1423 if (tcp_skb_pcount(skb
) > 1)
1424 tcp_set_skb_tso_segs(skb
, tcp_skb_mss(skb
));
1429 /* Calculate MSS not accounting any TCP options. */
1430 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1432 const struct tcp_sock
*tp
= tcp_sk(sk
);
1433 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1436 /* Calculate base mss without TCP options:
1437 It is MMS_S - sizeof(tcphdr) of rfc1122
1439 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1441 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1442 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1443 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1445 if (dst
&& dst_allfrag(dst
))
1446 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1449 /* Clamp it (mss_clamp does not include tcp options) */
1450 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1451 mss_now
= tp
->rx_opt
.mss_clamp
;
1453 /* Now subtract optional transport overhead */
1454 mss_now
-= icsk
->icsk_ext_hdr_len
;
1456 /* Then reserve room for full set of TCP options and 8 bytes of data */
1462 /* Calculate MSS. Not accounting for SACKs here. */
1463 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1465 /* Subtract TCP options size, not including SACKs */
1466 return __tcp_mtu_to_mss(sk
, pmtu
) -
1467 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1470 /* Inverse of above */
1471 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1473 const struct tcp_sock
*tp
= tcp_sk(sk
);
1474 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1478 tp
->tcp_header_len
+
1479 icsk
->icsk_ext_hdr_len
+
1480 icsk
->icsk_af_ops
->net_header_len
;
1482 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1483 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1484 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1486 if (dst
&& dst_allfrag(dst
))
1487 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1491 EXPORT_SYMBOL(tcp_mss_to_mtu
);
1493 /* MTU probing init per socket */
1494 void tcp_mtup_init(struct sock
*sk
)
1496 struct tcp_sock
*tp
= tcp_sk(sk
);
1497 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1498 struct net
*net
= sock_net(sk
);
1500 icsk
->icsk_mtup
.enabled
= net
->ipv4
.sysctl_tcp_mtu_probing
> 1;
1501 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1502 icsk
->icsk_af_ops
->net_header_len
;
1503 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, net
->ipv4
.sysctl_tcp_base_mss
);
1504 icsk
->icsk_mtup
.probe_size
= 0;
1505 if (icsk
->icsk_mtup
.enabled
)
1506 icsk
->icsk_mtup
.probe_timestamp
= tcp_jiffies32
;
1508 EXPORT_SYMBOL(tcp_mtup_init
);
1510 /* This function synchronize snd mss to current pmtu/exthdr set.
1512 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1513 for TCP options, but includes only bare TCP header.
1515 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1516 It is minimum of user_mss and mss received with SYN.
1517 It also does not include TCP options.
1519 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1521 tp->mss_cache is current effective sending mss, including
1522 all tcp options except for SACKs. It is evaluated,
1523 taking into account current pmtu, but never exceeds
1524 tp->rx_opt.mss_clamp.
1526 NOTE1. rfc1122 clearly states that advertised MSS
1527 DOES NOT include either tcp or ip options.
1529 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1530 are READ ONLY outside this function. --ANK (980731)
1532 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1534 struct tcp_sock
*tp
= tcp_sk(sk
);
1535 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1538 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1539 icsk
->icsk_mtup
.search_high
= pmtu
;
1541 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1542 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1544 /* And store cached results */
1545 icsk
->icsk_pmtu_cookie
= pmtu
;
1546 if (icsk
->icsk_mtup
.enabled
)
1547 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1548 tp
->mss_cache
= mss_now
;
1552 EXPORT_SYMBOL(tcp_sync_mss
);
1554 /* Compute the current effective MSS, taking SACKs and IP options,
1555 * and even PMTU discovery events into account.
1557 unsigned int tcp_current_mss(struct sock
*sk
)
1559 const struct tcp_sock
*tp
= tcp_sk(sk
);
1560 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1562 unsigned int header_len
;
1563 struct tcp_out_options opts
;
1564 struct tcp_md5sig_key
*md5
;
1566 mss_now
= tp
->mss_cache
;
1569 u32 mtu
= dst_mtu(dst
);
1570 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1571 mss_now
= tcp_sync_mss(sk
, mtu
);
1574 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1575 sizeof(struct tcphdr
);
1576 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1577 * some common options. If this is an odd packet (because we have SACK
1578 * blocks etc) then our calculated header_len will be different, and
1579 * we have to adjust mss_now correspondingly */
1580 if (header_len
!= tp
->tcp_header_len
) {
1581 int delta
= (int) header_len
- tp
->tcp_header_len
;
1588 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1589 * As additional protections, we do not touch cwnd in retransmission phases,
1590 * and if application hit its sndbuf limit recently.
1592 static void tcp_cwnd_application_limited(struct sock
*sk
)
1594 struct tcp_sock
*tp
= tcp_sk(sk
);
1596 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1597 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1598 /* Limited by application or receiver window. */
1599 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1600 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1601 if (win_used
< tp
->snd_cwnd
) {
1602 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1603 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
1605 tp
->snd_cwnd_used
= 0;
1607 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1610 static void tcp_cwnd_validate(struct sock
*sk
, bool is_cwnd_limited
)
1612 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1613 struct tcp_sock
*tp
= tcp_sk(sk
);
1615 /* Track the maximum number of outstanding packets in each
1616 * window, and remember whether we were cwnd-limited then.
1618 if (!before(tp
->snd_una
, tp
->max_packets_seq
) ||
1619 tp
->packets_out
> tp
->max_packets_out
) {
1620 tp
->max_packets_out
= tp
->packets_out
;
1621 tp
->max_packets_seq
= tp
->snd_nxt
;
1622 tp
->is_cwnd_limited
= is_cwnd_limited
;
1625 if (tcp_is_cwnd_limited(sk
)) {
1626 /* Network is feed fully. */
1627 tp
->snd_cwnd_used
= 0;
1628 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1630 /* Network starves. */
1631 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1632 tp
->snd_cwnd_used
= tp
->packets_out
;
1634 if (sock_net(sk
)->ipv4
.sysctl_tcp_slow_start_after_idle
&&
1635 (s32
)(tcp_jiffies32
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
&&
1636 !ca_ops
->cong_control
)
1637 tcp_cwnd_application_limited(sk
);
1639 /* The following conditions together indicate the starvation
1640 * is caused by insufficient sender buffer:
1641 * 1) just sent some data (see tcp_write_xmit)
1642 * 2) not cwnd limited (this else condition)
1643 * 3) no more data to send (tcp_write_queue_empty())
1644 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1646 if (tcp_write_queue_empty(sk
) && sk
->sk_socket
&&
1647 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
) &&
1648 (1 << sk
->sk_state
) & (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
1649 tcp_chrono_start(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1653 /* Minshall's variant of the Nagle send check. */
1654 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1656 return after(tp
->snd_sml
, tp
->snd_una
) &&
1657 !after(tp
->snd_sml
, tp
->snd_nxt
);
1660 /* Update snd_sml if this skb is under mss
1661 * Note that a TSO packet might end with a sub-mss segment
1662 * The test is really :
1663 * if ((skb->len % mss) != 0)
1664 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1665 * But we can avoid doing the divide again given we already have
1666 * skb_pcount = skb->len / mss_now
1668 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1669 const struct sk_buff
*skb
)
1671 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1672 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1675 /* Return false, if packet can be sent now without violation Nagle's rules:
1676 * 1. It is full sized. (provided by caller in %partial bool)
1677 * 2. Or it contains FIN. (already checked by caller)
1678 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1679 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1680 * With Minshall's modification: all sent small packets are ACKed.
1682 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1686 ((nonagle
& TCP_NAGLE_CORK
) ||
1687 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1690 /* Return how many segs we'd like on a TSO packet,
1691 * to send one TSO packet per ms
1693 static u32
tcp_tso_autosize(const struct sock
*sk
, unsigned int mss_now
,
1698 bytes
= min_t(unsigned long,
1699 sk
->sk_pacing_rate
>> sk
->sk_pacing_shift
,
1700 sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
);
1702 /* Goal is to send at least one packet per ms,
1703 * not one big TSO packet every 100 ms.
1704 * This preserves ACK clocking and is consistent
1705 * with tcp_tso_should_defer() heuristic.
1707 segs
= max_t(u32
, bytes
/ mss_now
, min_tso_segs
);
1712 /* Return the number of segments we want in the skb we are transmitting.
1713 * See if congestion control module wants to decide; otherwise, autosize.
1715 static u32
tcp_tso_segs(struct sock
*sk
, unsigned int mss_now
)
1717 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1718 u32 min_tso
, tso_segs
;
1720 min_tso
= ca_ops
->min_tso_segs
?
1721 ca_ops
->min_tso_segs(sk
) :
1722 sock_net(sk
)->ipv4
.sysctl_tcp_min_tso_segs
;
1724 tso_segs
= tcp_tso_autosize(sk
, mss_now
, min_tso
);
1725 return min_t(u32
, tso_segs
, sk
->sk_gso_max_segs
);
1728 /* Returns the portion of skb which can be sent right away */
1729 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1730 const struct sk_buff
*skb
,
1731 unsigned int mss_now
,
1732 unsigned int max_segs
,
1735 const struct tcp_sock
*tp
= tcp_sk(sk
);
1736 u32 partial
, needed
, window
, max_len
;
1738 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1739 max_len
= mss_now
* max_segs
;
1741 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1744 needed
= min(skb
->len
, window
);
1746 if (max_len
<= needed
)
1749 partial
= needed
% mss_now
;
1750 /* If last segment is not a full MSS, check if Nagle rules allow us
1751 * to include this last segment in this skb.
1752 * Otherwise, we'll split the skb at last MSS boundary
1754 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
1755 return needed
- partial
;
1760 /* Can at least one segment of SKB be sent right now, according to the
1761 * congestion window rules? If so, return how many segments are allowed.
1763 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1764 const struct sk_buff
*skb
)
1766 u32 in_flight
, cwnd
, halfcwnd
;
1768 /* Don't be strict about the congestion window for the final FIN. */
1769 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1770 tcp_skb_pcount(skb
) == 1)
1773 in_flight
= tcp_packets_in_flight(tp
);
1774 cwnd
= tp
->snd_cwnd
;
1775 if (in_flight
>= cwnd
)
1778 /* For better scheduling, ensure we have at least
1779 * 2 GSO packets in flight.
1781 halfcwnd
= max(cwnd
>> 1, 1U);
1782 return min(halfcwnd
, cwnd
- in_flight
);
1785 /* Initialize TSO state of a skb.
1786 * This must be invoked the first time we consider transmitting
1787 * SKB onto the wire.
1789 static int tcp_init_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1791 int tso_segs
= tcp_skb_pcount(skb
);
1793 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1794 tcp_set_skb_tso_segs(skb
, mss_now
);
1795 tso_segs
= tcp_skb_pcount(skb
);
1801 /* Return true if the Nagle test allows this packet to be
1804 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1805 unsigned int cur_mss
, int nonagle
)
1807 /* Nagle rule does not apply to frames, which sit in the middle of the
1808 * write_queue (they have no chances to get new data).
1810 * This is implemented in the callers, where they modify the 'nonagle'
1811 * argument based upon the location of SKB in the send queue.
1813 if (nonagle
& TCP_NAGLE_PUSH
)
1816 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1817 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1820 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
1826 /* Does at least the first segment of SKB fit into the send window? */
1827 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1828 const struct sk_buff
*skb
,
1829 unsigned int cur_mss
)
1831 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1833 if (skb
->len
> cur_mss
)
1834 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1836 return !after(end_seq
, tcp_wnd_end(tp
));
1839 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1840 * which is put after SKB on the list. It is very much like
1841 * tcp_fragment() except that it may make several kinds of assumptions
1842 * in order to speed up the splitting operation. In particular, we
1843 * know that all the data is in scatter-gather pages, and that the
1844 * packet has never been sent out before (and thus is not cloned).
1846 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1847 unsigned int mss_now
, gfp_t gfp
)
1849 int nlen
= skb
->len
- len
;
1850 struct sk_buff
*buff
;
1853 /* All of a TSO frame must be composed of paged data. */
1854 if (skb
->len
!= skb
->data_len
)
1855 return tcp_fragment(sk
, TCP_FRAG_IN_WRITE_QUEUE
,
1856 skb
, len
, mss_now
, gfp
);
1858 buff
= sk_stream_alloc_skb(sk
, 0, gfp
, true);
1859 if (unlikely(!buff
))
1862 sk
->sk_wmem_queued
+= buff
->truesize
;
1863 sk_mem_charge(sk
, buff
->truesize
);
1864 buff
->truesize
+= nlen
;
1865 skb
->truesize
-= nlen
;
1867 /* Correct the sequence numbers. */
1868 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1869 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1870 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1872 /* PSH and FIN should only be set in the second packet. */
1873 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1874 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1875 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1877 /* This packet was never sent out yet, so no SACK bits. */
1878 TCP_SKB_CB(buff
)->sacked
= 0;
1880 tcp_skb_fragment_eor(skb
, buff
);
1882 buff
->ip_summed
= CHECKSUM_PARTIAL
;
1883 skb_split(skb
, buff
, len
);
1884 tcp_fragment_tstamp(skb
, buff
);
1886 /* Fix up tso_factor for both original and new SKB. */
1887 tcp_set_skb_tso_segs(skb
, mss_now
);
1888 tcp_set_skb_tso_segs(buff
, mss_now
);
1890 /* Link BUFF into the send queue. */
1891 __skb_header_release(buff
);
1892 tcp_insert_write_queue_after(skb
, buff
, sk
, TCP_FRAG_IN_WRITE_QUEUE
);
1897 /* Try to defer sending, if possible, in order to minimize the amount
1898 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1900 * This algorithm is from John Heffner.
1902 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
,
1903 bool *is_cwnd_limited
,
1904 bool *is_rwnd_limited
,
1907 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1908 u32 send_win
, cong_win
, limit
, in_flight
;
1909 struct tcp_sock
*tp
= tcp_sk(sk
);
1910 struct sk_buff
*head
;
1914 if (icsk
->icsk_ca_state
>= TCP_CA_Recovery
)
1917 /* Avoid bursty behavior by allowing defer
1918 * only if the last write was recent (1 ms).
1919 * Note that tp->tcp_wstamp_ns can be in the future if we have
1920 * packets waiting in a qdisc or device for EDT delivery.
1922 delta
= tp
->tcp_clock_cache
- tp
->tcp_wstamp_ns
- NSEC_PER_MSEC
;
1926 in_flight
= tcp_packets_in_flight(tp
);
1928 BUG_ON(tcp_skb_pcount(skb
) <= 1);
1929 BUG_ON(tp
->snd_cwnd
<= in_flight
);
1931 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1933 /* From in_flight test above, we know that cwnd > in_flight. */
1934 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1936 limit
= min(send_win
, cong_win
);
1938 /* If a full-sized TSO skb can be sent, do it. */
1939 if (limit
>= max_segs
* tp
->mss_cache
)
1942 /* Middle in queue won't get any more data, full sendable already? */
1943 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1946 win_divisor
= READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_tso_win_divisor
);
1948 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1950 /* If at least some fraction of a window is available,
1953 chunk
/= win_divisor
;
1957 /* Different approach, try not to defer past a single
1958 * ACK. Receiver should ACK every other full sized
1959 * frame, so if we have space for more than 3 frames
1962 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1966 /* TODO : use tsorted_sent_queue ? */
1967 head
= tcp_rtx_queue_head(sk
);
1970 delta
= tp
->tcp_clock_cache
- head
->tstamp
;
1971 /* If next ACK is likely to come too late (half srtt), do not defer */
1972 if ((s64
)(delta
- (u64
)NSEC_PER_USEC
* (tp
->srtt_us
>> 4)) < 0)
1975 /* Ok, it looks like it is advisable to defer.
1976 * Three cases are tracked :
1977 * 1) We are cwnd-limited
1978 * 2) We are rwnd-limited
1979 * 3) We are application limited.
1981 if (cong_win
< send_win
) {
1982 if (cong_win
<= skb
->len
) {
1983 *is_cwnd_limited
= true;
1987 if (send_win
<= skb
->len
) {
1988 *is_rwnd_limited
= true;
1993 /* If this packet won't get more data, do not wait. */
1994 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) ||
1995 TCP_SKB_CB(skb
)->eor
)
2004 static inline void tcp_mtu_check_reprobe(struct sock
*sk
)
2006 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2007 struct tcp_sock
*tp
= tcp_sk(sk
);
2008 struct net
*net
= sock_net(sk
);
2012 interval
= net
->ipv4
.sysctl_tcp_probe_interval
;
2013 delta
= tcp_jiffies32
- icsk
->icsk_mtup
.probe_timestamp
;
2014 if (unlikely(delta
>= interval
* HZ
)) {
2015 int mss
= tcp_current_mss(sk
);
2017 /* Update current search range */
2018 icsk
->icsk_mtup
.probe_size
= 0;
2019 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+
2020 sizeof(struct tcphdr
) +
2021 icsk
->icsk_af_ops
->net_header_len
;
2022 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
2024 /* Update probe time stamp */
2025 icsk
->icsk_mtup
.probe_timestamp
= tcp_jiffies32
;
2029 static bool tcp_can_coalesce_send_queue_head(struct sock
*sk
, int len
)
2031 struct sk_buff
*skb
, *next
;
2033 skb
= tcp_send_head(sk
);
2034 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
2035 if (len
<= skb
->len
)
2038 if (unlikely(TCP_SKB_CB(skb
)->eor
))
2047 /* Create a new MTU probe if we are ready.
2048 * MTU probe is regularly attempting to increase the path MTU by
2049 * deliberately sending larger packets. This discovers routing
2050 * changes resulting in larger path MTUs.
2052 * Returns 0 if we should wait to probe (no cwnd available),
2053 * 1 if a probe was sent,
2056 static int tcp_mtu_probe(struct sock
*sk
)
2058 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2059 struct tcp_sock
*tp
= tcp_sk(sk
);
2060 struct sk_buff
*skb
, *nskb
, *next
;
2061 struct net
*net
= sock_net(sk
);
2068 /* Not currently probing/verifying,
2070 * have enough cwnd, and
2071 * not SACKing (the variable headers throw things off)
2073 if (likely(!icsk
->icsk_mtup
.enabled
||
2074 icsk
->icsk_mtup
.probe_size
||
2075 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
2076 tp
->snd_cwnd
< 11 ||
2077 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
))
2080 /* Use binary search for probe_size between tcp_mss_base,
2081 * and current mss_clamp. if (search_high - search_low)
2082 * smaller than a threshold, backoff from probing.
2084 mss_now
= tcp_current_mss(sk
);
2085 probe_size
= tcp_mtu_to_mss(sk
, (icsk
->icsk_mtup
.search_high
+
2086 icsk
->icsk_mtup
.search_low
) >> 1);
2087 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
2088 interval
= icsk
->icsk_mtup
.search_high
- icsk
->icsk_mtup
.search_low
;
2089 /* When misfortune happens, we are reprobing actively,
2090 * and then reprobe timer has expired. We stick with current
2091 * probing process by not resetting search range to its orignal.
2093 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
) ||
2094 interval
< net
->ipv4
.sysctl_tcp_probe_threshold
) {
2095 /* Check whether enough time has elaplased for
2096 * another round of probing.
2098 tcp_mtu_check_reprobe(sk
);
2102 /* Have enough data in the send queue to probe? */
2103 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
2106 if (tp
->snd_wnd
< size_needed
)
2108 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
2111 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2112 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
2113 if (!tcp_packets_in_flight(tp
))
2119 if (!tcp_can_coalesce_send_queue_head(sk
, probe_size
))
2122 /* We're allowed to probe. Build it now. */
2123 nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
, false);
2126 sk
->sk_wmem_queued
+= nskb
->truesize
;
2127 sk_mem_charge(sk
, nskb
->truesize
);
2129 skb
= tcp_send_head(sk
);
2131 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
2132 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
2133 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
2134 TCP_SKB_CB(nskb
)->sacked
= 0;
2136 nskb
->ip_summed
= CHECKSUM_PARTIAL
;
2138 tcp_insert_write_queue_before(nskb
, skb
, sk
);
2139 tcp_highest_sack_replace(sk
, skb
, nskb
);
2142 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
2143 copy
= min_t(int, skb
->len
, probe_size
- len
);
2144 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
2146 if (skb
->len
<= copy
) {
2147 /* We've eaten all the data from this skb.
2149 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
2150 /* If this is the last SKB we copy and eor is set
2151 * we need to propagate it to the new skb.
2153 TCP_SKB_CB(nskb
)->eor
= TCP_SKB_CB(skb
)->eor
;
2154 tcp_unlink_write_queue(skb
, sk
);
2155 sk_wmem_free_skb(sk
, skb
);
2157 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
2158 ~(TCPHDR_FIN
|TCPHDR_PSH
);
2159 if (!skb_shinfo(skb
)->nr_frags
) {
2160 skb_pull(skb
, copy
);
2162 __pskb_trim_head(skb
, copy
);
2163 tcp_set_skb_tso_segs(skb
, mss_now
);
2165 TCP_SKB_CB(skb
)->seq
+= copy
;
2170 if (len
>= probe_size
)
2173 tcp_init_tso_segs(nskb
, nskb
->len
);
2175 /* We're ready to send. If this fails, the probe will
2176 * be resegmented into mss-sized pieces by tcp_write_xmit().
2178 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
2179 /* Decrement cwnd here because we are sending
2180 * effectively two packets. */
2182 tcp_event_new_data_sent(sk
, nskb
);
2184 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
2185 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
2186 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
2194 static bool tcp_pacing_check(struct sock
*sk
)
2196 struct tcp_sock
*tp
= tcp_sk(sk
);
2198 if (!tcp_needs_internal_pacing(sk
))
2201 if (tp
->tcp_wstamp_ns
<= tp
->tcp_clock_cache
)
2204 if (!hrtimer_is_queued(&tp
->pacing_timer
)) {
2205 hrtimer_start(&tp
->pacing_timer
,
2206 ns_to_ktime(tp
->tcp_wstamp_ns
),
2207 HRTIMER_MODE_ABS_PINNED_SOFT
);
2213 /* TCP Small Queues :
2214 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2215 * (These limits are doubled for retransmits)
2217 * - better RTT estimation and ACK scheduling
2220 * Alas, some drivers / subsystems require a fair amount
2221 * of queued bytes to ensure line rate.
2222 * One example is wifi aggregation (802.11 AMPDU)
2224 static bool tcp_small_queue_check(struct sock
*sk
, const struct sk_buff
*skb
,
2225 unsigned int factor
)
2227 unsigned long limit
;
2229 limit
= max_t(unsigned long,
2231 sk
->sk_pacing_rate
>> sk
->sk_pacing_shift
);
2232 if (sk
->sk_pacing_status
== SK_PACING_NONE
)
2233 limit
= min_t(unsigned long, limit
,
2234 sock_net(sk
)->ipv4
.sysctl_tcp_limit_output_bytes
);
2237 if (refcount_read(&sk
->sk_wmem_alloc
) > limit
) {
2238 /* Always send skb if rtx queue is empty.
2239 * No need to wait for TX completion to call us back,
2240 * after softirq/tasklet schedule.
2241 * This helps when TX completions are delayed too much.
2243 if (tcp_rtx_queue_empty(sk
))
2246 set_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
);
2247 /* It is possible TX completion already happened
2248 * before we set TSQ_THROTTLED, so we must
2249 * test again the condition.
2251 smp_mb__after_atomic();
2252 if (refcount_read(&sk
->sk_wmem_alloc
) > limit
)
2258 static void tcp_chrono_set(struct tcp_sock
*tp
, const enum tcp_chrono
new)
2260 const u32 now
= tcp_jiffies32
;
2261 enum tcp_chrono old
= tp
->chrono_type
;
2263 if (old
> TCP_CHRONO_UNSPEC
)
2264 tp
->chrono_stat
[old
- 1] += now
- tp
->chrono_start
;
2265 tp
->chrono_start
= now
;
2266 tp
->chrono_type
= new;
2269 void tcp_chrono_start(struct sock
*sk
, const enum tcp_chrono type
)
2271 struct tcp_sock
*tp
= tcp_sk(sk
);
2273 /* If there are multiple conditions worthy of tracking in a
2274 * chronograph then the highest priority enum takes precedence
2275 * over the other conditions. So that if something "more interesting"
2276 * starts happening, stop the previous chrono and start a new one.
2278 if (type
> tp
->chrono_type
)
2279 tcp_chrono_set(tp
, type
);
2282 void tcp_chrono_stop(struct sock
*sk
, const enum tcp_chrono type
)
2284 struct tcp_sock
*tp
= tcp_sk(sk
);
2287 /* There are multiple conditions worthy of tracking in a
2288 * chronograph, so that the highest priority enum takes
2289 * precedence over the other conditions (see tcp_chrono_start).
2290 * If a condition stops, we only stop chrono tracking if
2291 * it's the "most interesting" or current chrono we are
2292 * tracking and starts busy chrono if we have pending data.
2294 if (tcp_rtx_and_write_queues_empty(sk
))
2295 tcp_chrono_set(tp
, TCP_CHRONO_UNSPEC
);
2296 else if (type
== tp
->chrono_type
)
2297 tcp_chrono_set(tp
, TCP_CHRONO_BUSY
);
2300 /* This routine writes packets to the network. It advances the
2301 * send_head. This happens as incoming acks open up the remote
2304 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2305 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2306 * account rare use of URG, this is not a big flaw.
2308 * Send at most one packet when push_one > 0. Temporarily ignore
2309 * cwnd limit to force at most one packet out when push_one == 2.
2311 * Returns true, if no segments are in flight and we have queued segments,
2312 * but cannot send anything now because of SWS or another problem.
2314 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
2315 int push_one
, gfp_t gfp
)
2317 struct tcp_sock
*tp
= tcp_sk(sk
);
2318 struct sk_buff
*skb
;
2319 unsigned int tso_segs
, sent_pkts
;
2322 bool is_cwnd_limited
= false, is_rwnd_limited
= false;
2327 tcp_mstamp_refresh(tp
);
2329 /* Do MTU probing. */
2330 result
= tcp_mtu_probe(sk
);
2333 } else if (result
> 0) {
2338 max_segs
= tcp_tso_segs(sk
, mss_now
);
2339 while ((skb
= tcp_send_head(sk
))) {
2342 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
) {
2343 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2344 skb
->skb_mstamp_ns
= tp
->tcp_wstamp_ns
= tp
->tcp_clock_cache
;
2345 list_move_tail(&skb
->tcp_tsorted_anchor
, &tp
->tsorted_sent_queue
);
2346 tcp_init_tso_segs(skb
, mss_now
);
2347 goto repair
; /* Skip network transmission */
2350 if (tcp_pacing_check(sk
))
2353 tso_segs
= tcp_init_tso_segs(skb
, mss_now
);
2356 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2359 /* Force out a loss probe pkt. */
2365 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
))) {
2366 is_rwnd_limited
= true;
2370 if (tso_segs
== 1) {
2371 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2372 (tcp_skb_is_last(sk
, skb
) ?
2373 nonagle
: TCP_NAGLE_PUSH
))))
2377 tcp_tso_should_defer(sk
, skb
, &is_cwnd_limited
,
2378 &is_rwnd_limited
, max_segs
))
2383 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
2384 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2390 if (skb
->len
> limit
&&
2391 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
2394 if (tcp_small_queue_check(sk
, skb
, 0))
2397 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2401 /* Advance the send_head. This one is sent out.
2402 * This call will increment packets_out.
2404 tcp_event_new_data_sent(sk
, skb
);
2406 tcp_minshall_update(tp
, mss_now
, skb
);
2407 sent_pkts
+= tcp_skb_pcount(skb
);
2413 if (is_rwnd_limited
)
2414 tcp_chrono_start(sk
, TCP_CHRONO_RWND_LIMITED
);
2416 tcp_chrono_stop(sk
, TCP_CHRONO_RWND_LIMITED
);
2418 if (likely(sent_pkts
)) {
2419 if (tcp_in_cwnd_reduction(sk
))
2420 tp
->prr_out
+= sent_pkts
;
2422 /* Send one loss probe per tail loss episode. */
2424 tcp_schedule_loss_probe(sk
, false);
2425 is_cwnd_limited
|= (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
);
2426 tcp_cwnd_validate(sk
, is_cwnd_limited
);
2429 return !tp
->packets_out
&& !tcp_write_queue_empty(sk
);
2432 bool tcp_schedule_loss_probe(struct sock
*sk
, bool advancing_rto
)
2434 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2435 struct tcp_sock
*tp
= tcp_sk(sk
);
2436 u32 timeout
, rto_delta_us
;
2439 /* Don't do any loss probe on a Fast Open connection before 3WHS
2442 if (tp
->fastopen_rsk
)
2445 early_retrans
= sock_net(sk
)->ipv4
.sysctl_tcp_early_retrans
;
2446 /* Schedule a loss probe in 2*RTT for SACK capable connections
2447 * not in loss recovery, that are either limited by cwnd or application.
2449 if ((early_retrans
!= 3 && early_retrans
!= 4) ||
2450 !tp
->packets_out
|| !tcp_is_sack(tp
) ||
2451 (icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2452 icsk
->icsk_ca_state
!= TCP_CA_CWR
))
2455 /* Probe timeout is 2*rtt. Add minimum RTO to account
2456 * for delayed ack when there's one outstanding packet. If no RTT
2457 * sample is available then probe after TCP_TIMEOUT_INIT.
2460 timeout
= usecs_to_jiffies(tp
->srtt_us
>> 2);
2461 if (tp
->packets_out
== 1)
2462 timeout
+= TCP_RTO_MIN
;
2464 timeout
+= TCP_TIMEOUT_MIN
;
2466 timeout
= TCP_TIMEOUT_INIT
;
2469 /* If the RTO formula yields an earlier time, then use that time. */
2470 rto_delta_us
= advancing_rto
?
2471 jiffies_to_usecs(inet_csk(sk
)->icsk_rto
) :
2472 tcp_rto_delta_us(sk
); /* How far in future is RTO? */
2473 if (rto_delta_us
> 0)
2474 timeout
= min_t(u32
, timeout
, usecs_to_jiffies(rto_delta_us
));
2476 tcp_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2481 /* Thanks to skb fast clones, we can detect if a prior transmit of
2482 * a packet is still in a qdisc or driver queue.
2483 * In this case, there is very little point doing a retransmit !
2485 static bool skb_still_in_host_queue(const struct sock
*sk
,
2486 const struct sk_buff
*skb
)
2488 if (unlikely(skb_fclone_busy(sk
, skb
))) {
2489 NET_INC_STATS(sock_net(sk
),
2490 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
2496 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2497 * retransmit the last segment.
2499 void tcp_send_loss_probe(struct sock
*sk
)
2501 struct tcp_sock
*tp
= tcp_sk(sk
);
2502 struct sk_buff
*skb
;
2504 int mss
= tcp_current_mss(sk
);
2506 skb
= tcp_send_head(sk
);
2507 if (skb
&& tcp_snd_wnd_test(tp
, skb
, mss
)) {
2508 pcount
= tp
->packets_out
;
2509 tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2510 if (tp
->packets_out
> pcount
)
2514 skb
= skb_rb_last(&sk
->tcp_rtx_queue
);
2515 if (unlikely(!skb
)) {
2516 WARN_ONCE(tp
->packets_out
,
2517 "invalid inflight: %u state %u cwnd %u mss %d\n",
2518 tp
->packets_out
, sk
->sk_state
, tp
->snd_cwnd
, mss
);
2519 inet_csk(sk
)->icsk_pending
= 0;
2523 /* At most one outstanding TLP retransmission. */
2524 if (tp
->tlp_high_seq
)
2527 if (skb_still_in_host_queue(sk
, skb
))
2530 pcount
= tcp_skb_pcount(skb
);
2531 if (WARN_ON(!pcount
))
2534 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2535 if (unlikely(tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
,
2536 (pcount
- 1) * mss
, mss
,
2539 skb
= skb_rb_next(skb
);
2542 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2545 if (__tcp_retransmit_skb(sk
, skb
, 1))
2548 /* Record snd_nxt for loss detection. */
2549 tp
->tlp_high_seq
= tp
->snd_nxt
;
2552 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSPROBES
);
2553 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2554 inet_csk(sk
)->icsk_pending
= 0;
2559 /* Push out any pending frames which were held back due to
2560 * TCP_CORK or attempt at coalescing tiny packets.
2561 * The socket must be locked by the caller.
2563 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2566 /* If we are closed, the bytes will have to remain here.
2567 * In time closedown will finish, we empty the write queue and
2568 * all will be happy.
2570 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2573 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2574 sk_gfp_mask(sk
, GFP_ATOMIC
)))
2575 tcp_check_probe_timer(sk
);
2578 /* Send _single_ skb sitting at the send head. This function requires
2579 * true push pending frames to setup probe timer etc.
2581 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2583 struct sk_buff
*skb
= tcp_send_head(sk
);
2585 BUG_ON(!skb
|| skb
->len
< mss_now
);
2587 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2590 /* This function returns the amount that we can raise the
2591 * usable window based on the following constraints
2593 * 1. The window can never be shrunk once it is offered (RFC 793)
2594 * 2. We limit memory per socket
2597 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2598 * RECV.NEXT + RCV.WIN fixed until:
2599 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2601 * i.e. don't raise the right edge of the window until you can raise
2602 * it at least MSS bytes.
2604 * Unfortunately, the recommended algorithm breaks header prediction,
2605 * since header prediction assumes th->window stays fixed.
2607 * Strictly speaking, keeping th->window fixed violates the receiver
2608 * side SWS prevention criteria. The problem is that under this rule
2609 * a stream of single byte packets will cause the right side of the
2610 * window to always advance by a single byte.
2612 * Of course, if the sender implements sender side SWS prevention
2613 * then this will not be a problem.
2615 * BSD seems to make the following compromise:
2617 * If the free space is less than the 1/4 of the maximum
2618 * space available and the free space is less than 1/2 mss,
2619 * then set the window to 0.
2620 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2621 * Otherwise, just prevent the window from shrinking
2622 * and from being larger than the largest representable value.
2624 * This prevents incremental opening of the window in the regime
2625 * where TCP is limited by the speed of the reader side taking
2626 * data out of the TCP receive queue. It does nothing about
2627 * those cases where the window is constrained on the sender side
2628 * because the pipeline is full.
2630 * BSD also seems to "accidentally" limit itself to windows that are a
2631 * multiple of MSS, at least until the free space gets quite small.
2632 * This would appear to be a side effect of the mbuf implementation.
2633 * Combining these two algorithms results in the observed behavior
2634 * of having a fixed window size at almost all times.
2636 * Below we obtain similar behavior by forcing the offered window to
2637 * a multiple of the mss when it is feasible to do so.
2639 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2640 * Regular options like TIMESTAMP are taken into account.
2642 u32
__tcp_select_window(struct sock
*sk
)
2644 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2645 struct tcp_sock
*tp
= tcp_sk(sk
);
2646 /* MSS for the peer's data. Previous versions used mss_clamp
2647 * here. I don't know if the value based on our guesses
2648 * of peer's MSS is better for the performance. It's more correct
2649 * but may be worse for the performance because of rcv_mss
2650 * fluctuations. --SAW 1998/11/1
2652 int mss
= icsk
->icsk_ack
.rcv_mss
;
2653 int free_space
= tcp_space(sk
);
2654 int allowed_space
= tcp_full_space(sk
);
2655 int full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2658 if (unlikely(mss
> full_space
)) {
2663 if (free_space
< (full_space
>> 1)) {
2664 icsk
->icsk_ack
.quick
= 0;
2666 if (tcp_under_memory_pressure(sk
))
2667 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2670 /* free_space might become our new window, make sure we don't
2671 * increase it due to wscale.
2673 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2675 /* if free space is less than mss estimate, or is below 1/16th
2676 * of the maximum allowed, try to move to zero-window, else
2677 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2678 * new incoming data is dropped due to memory limits.
2679 * With large window, mss test triggers way too late in order
2680 * to announce zero window in time before rmem limit kicks in.
2682 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2686 if (free_space
> tp
->rcv_ssthresh
)
2687 free_space
= tp
->rcv_ssthresh
;
2689 /* Don't do rounding if we are using window scaling, since the
2690 * scaled window will not line up with the MSS boundary anyway.
2692 if (tp
->rx_opt
.rcv_wscale
) {
2693 window
= free_space
;
2695 /* Advertise enough space so that it won't get scaled away.
2696 * Import case: prevent zero window announcement if
2697 * 1<<rcv_wscale > mss.
2699 window
= ALIGN(window
, (1 << tp
->rx_opt
.rcv_wscale
));
2701 window
= tp
->rcv_wnd
;
2702 /* Get the largest window that is a nice multiple of mss.
2703 * Window clamp already applied above.
2704 * If our current window offering is within 1 mss of the
2705 * free space we just keep it. This prevents the divide
2706 * and multiply from happening most of the time.
2707 * We also don't do any window rounding when the free space
2710 if (window
<= free_space
- mss
|| window
> free_space
)
2711 window
= rounddown(free_space
, mss
);
2712 else if (mss
== full_space
&&
2713 free_space
> window
+ (full_space
>> 1))
2714 window
= free_space
;
2720 void tcp_skb_collapse_tstamp(struct sk_buff
*skb
,
2721 const struct sk_buff
*next_skb
)
2723 if (unlikely(tcp_has_tx_tstamp(next_skb
))) {
2724 const struct skb_shared_info
*next_shinfo
=
2725 skb_shinfo(next_skb
);
2726 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
2728 shinfo
->tx_flags
|= next_shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
2729 shinfo
->tskey
= next_shinfo
->tskey
;
2730 TCP_SKB_CB(skb
)->txstamp_ack
|=
2731 TCP_SKB_CB(next_skb
)->txstamp_ack
;
2735 /* Collapses two adjacent SKB's during retransmission. */
2736 static bool tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2738 struct tcp_sock
*tp
= tcp_sk(sk
);
2739 struct sk_buff
*next_skb
= skb_rb_next(skb
);
2742 next_skb_size
= next_skb
->len
;
2744 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2746 if (next_skb_size
) {
2747 if (next_skb_size
<= skb_availroom(skb
))
2748 skb_copy_bits(next_skb
, 0, skb_put(skb
, next_skb_size
),
2750 else if (!skb_shift(skb
, next_skb
, next_skb_size
))
2753 tcp_highest_sack_replace(sk
, next_skb
, skb
);
2755 /* Update sequence range on original skb. */
2756 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2758 /* Merge over control information. This moves PSH/FIN etc. over */
2759 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2761 /* All done, get rid of second SKB and account for it so
2762 * packet counting does not break.
2764 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2765 TCP_SKB_CB(skb
)->eor
= TCP_SKB_CB(next_skb
)->eor
;
2767 /* changed transmit queue under us so clear hints */
2768 tcp_clear_retrans_hints_partial(tp
);
2769 if (next_skb
== tp
->retransmit_skb_hint
)
2770 tp
->retransmit_skb_hint
= skb
;
2772 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2774 tcp_skb_collapse_tstamp(skb
, next_skb
);
2776 tcp_rtx_queue_unlink_and_free(next_skb
, sk
);
2780 /* Check if coalescing SKBs is legal. */
2781 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2783 if (tcp_skb_pcount(skb
) > 1)
2785 if (skb_cloned(skb
))
2787 /* Some heuristics for collapsing over SACK'd could be invented */
2788 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2794 /* Collapse packets in the retransmit queue to make to create
2795 * less packets on the wire. This is only done on retransmission.
2797 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2800 struct tcp_sock
*tp
= tcp_sk(sk
);
2801 struct sk_buff
*skb
= to
, *tmp
;
2804 if (!sock_net(sk
)->ipv4
.sysctl_tcp_retrans_collapse
)
2806 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2809 skb_rbtree_walk_from_safe(skb
, tmp
) {
2810 if (!tcp_can_collapse(sk
, skb
))
2813 if (!tcp_skb_can_collapse_to(to
))
2826 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2829 if (!tcp_collapse_retrans(sk
, to
))
2834 /* This retransmits one SKB. Policy decisions and retransmit queue
2835 * state updates are done by the caller. Returns non-zero if an
2836 * error occurred which prevented the send.
2838 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2840 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2841 struct tcp_sock
*tp
= tcp_sk(sk
);
2842 unsigned int cur_mss
;
2846 /* Inconclusive MTU probe */
2847 if (icsk
->icsk_mtup
.probe_size
)
2848 icsk
->icsk_mtup
.probe_size
= 0;
2850 /* Do not sent more than we queued. 1/4 is reserved for possible
2851 * copying overhead: fragmentation, tunneling, mangling etc.
2853 if (refcount_read(&sk
->sk_wmem_alloc
) >
2854 min_t(u32
, sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2),
2858 if (skb_still_in_host_queue(sk
, skb
))
2861 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2862 if (unlikely(before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))) {
2866 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2870 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2871 return -EHOSTUNREACH
; /* Routing failure or similar. */
2873 cur_mss
= tcp_current_mss(sk
);
2875 /* If receiver has shrunk his window, and skb is out of
2876 * new window, do not retransmit it. The exception is the
2877 * case, when window is shrunk to zero. In this case
2878 * our retransmit serves as a zero window probe.
2880 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2881 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2884 len
= cur_mss
* segs
;
2885 if (skb
->len
> len
) {
2886 if (tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
, len
,
2887 cur_mss
, GFP_ATOMIC
))
2888 return -ENOMEM
; /* We'll try again later. */
2890 if (skb_unclone(skb
, GFP_ATOMIC
))
2893 diff
= tcp_skb_pcount(skb
);
2894 tcp_set_skb_tso_segs(skb
, cur_mss
);
2895 diff
-= tcp_skb_pcount(skb
);
2897 tcp_adjust_pcount(sk
, skb
, diff
);
2898 if (skb
->len
< cur_mss
)
2899 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2902 /* RFC3168, section 6.1.1.1. ECN fallback */
2903 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN_ECN
) == TCPHDR_SYN_ECN
)
2904 tcp_ecn_clear_syn(sk
, skb
);
2906 /* Update global and local TCP statistics. */
2907 segs
= tcp_skb_pcount(skb
);
2908 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
, segs
);
2909 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2910 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
2911 tp
->total_retrans
+= segs
;
2912 tp
->bytes_retrans
+= skb
->len
;
2914 /* make sure skb->data is aligned on arches that require it
2915 * and check if ack-trimming & collapsing extended the headroom
2916 * beyond what csum_start can cover.
2918 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2919 skb_headroom(skb
) >= 0xFFFF)) {
2920 struct sk_buff
*nskb
;
2922 tcp_skb_tsorted_save(skb
) {
2923 nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
, GFP_ATOMIC
);
2924 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2926 } tcp_skb_tsorted_restore(skb
);
2929 tcp_update_skb_after_send(sk
, skb
, tp
->tcp_wstamp_ns
);
2930 tcp_rate_skb_sent(sk
, skb
);
2933 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2936 /* To avoid taking spuriously low RTT samples based on a timestamp
2937 * for a transmit that never happened, always mark EVER_RETRANS
2939 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
2941 if (BPF_SOCK_OPS_TEST_FLAG(tp
, BPF_SOCK_OPS_RETRANS_CB_FLAG
))
2942 tcp_call_bpf_3arg(sk
, BPF_SOCK_OPS_RETRANS_CB
,
2943 TCP_SKB_CB(skb
)->seq
, segs
, err
);
2946 trace_tcp_retransmit_skb(sk
, skb
);
2947 } else if (err
!= -EBUSY
) {
2948 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
, segs
);
2953 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2955 struct tcp_sock
*tp
= tcp_sk(sk
);
2956 int err
= __tcp_retransmit_skb(sk
, skb
, segs
);
2959 #if FASTRETRANS_DEBUG > 0
2960 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2961 net_dbg_ratelimited("retrans_out leaked\n");
2964 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2965 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2968 /* Save stamp of the first (attempted) retransmit. */
2969 if (!tp
->retrans_stamp
)
2970 tp
->retrans_stamp
= tcp_skb_timestamp(skb
);
2972 if (tp
->undo_retrans
< 0)
2973 tp
->undo_retrans
= 0;
2974 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2978 /* This gets called after a retransmit timeout, and the initially
2979 * retransmitted data is acknowledged. It tries to continue
2980 * resending the rest of the retransmit queue, until either
2981 * we've sent it all or the congestion window limit is reached.
2983 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2985 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2986 struct sk_buff
*skb
, *rtx_head
, *hole
= NULL
;
2987 struct tcp_sock
*tp
= tcp_sk(sk
);
2991 if (!tp
->packets_out
)
2994 rtx_head
= tcp_rtx_queue_head(sk
);
2995 skb
= tp
->retransmit_skb_hint
?: rtx_head
;
2996 max_segs
= tcp_tso_segs(sk
, tcp_current_mss(sk
));
2997 skb_rbtree_walk_from(skb
) {
3001 if (tcp_pacing_check(sk
))
3004 /* we could do better than to assign each time */
3006 tp
->retransmit_skb_hint
= skb
;
3008 segs
= tp
->snd_cwnd
- tcp_packets_in_flight(tp
);
3011 sacked
= TCP_SKB_CB(skb
)->sacked
;
3012 /* In case tcp_shift_skb_data() have aggregated large skbs,
3013 * we need to make sure not sending too bigs TSO packets
3015 segs
= min_t(int, segs
, max_segs
);
3017 if (tp
->retrans_out
>= tp
->lost_out
) {
3019 } else if (!(sacked
& TCPCB_LOST
)) {
3020 if (!hole
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
3025 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
3026 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
3028 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
3031 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
3034 if (tcp_small_queue_check(sk
, skb
, 1))
3037 if (tcp_retransmit_skb(sk
, skb
, segs
))
3040 NET_ADD_STATS(sock_net(sk
), mib_idx
, tcp_skb_pcount(skb
));
3042 if (tcp_in_cwnd_reduction(sk
))
3043 tp
->prr_out
+= tcp_skb_pcount(skb
);
3045 if (skb
== rtx_head
&&
3046 icsk
->icsk_pending
!= ICSK_TIME_REO_TIMEOUT
)
3047 tcp_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3048 inet_csk(sk
)->icsk_rto
,
3054 /* We allow to exceed memory limits for FIN packets to expedite
3055 * connection tear down and (memory) recovery.
3056 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3057 * or even be forced to close flow without any FIN.
3058 * In general, we want to allow one skb per socket to avoid hangs
3059 * with edge trigger epoll()
3061 void sk_forced_mem_schedule(struct sock
*sk
, int size
)
3065 if (size
<= sk
->sk_forward_alloc
)
3067 amt
= sk_mem_pages(size
);
3068 sk
->sk_forward_alloc
+= amt
* SK_MEM_QUANTUM
;
3069 sk_memory_allocated_add(sk
, amt
);
3071 if (mem_cgroup_sockets_enabled
&& sk
->sk_memcg
)
3072 mem_cgroup_charge_skmem(sk
->sk_memcg
, amt
);
3075 /* Send a FIN. The caller locks the socket for us.
3076 * We should try to send a FIN packet really hard, but eventually give up.
3078 void tcp_send_fin(struct sock
*sk
)
3080 struct sk_buff
*skb
, *tskb
= tcp_write_queue_tail(sk
);
3081 struct tcp_sock
*tp
= tcp_sk(sk
);
3083 /* Optimization, tack on the FIN if we have one skb in write queue and
3084 * this skb was not yet sent, or we are under memory pressure.
3085 * Note: in the latter case, FIN packet will be sent after a timeout,
3086 * as TCP stack thinks it has already been transmitted.
3088 if (!tskb
&& tcp_under_memory_pressure(sk
))
3089 tskb
= skb_rb_last(&sk
->tcp_rtx_queue
);
3092 TCP_SKB_CB(tskb
)->tcp_flags
|= TCPHDR_FIN
;
3093 TCP_SKB_CB(tskb
)->end_seq
++;
3095 if (tcp_write_queue_empty(sk
)) {
3096 /* This means tskb was already sent.
3097 * Pretend we included the FIN on previous transmit.
3098 * We need to set tp->snd_nxt to the value it would have
3099 * if FIN had been sent. This is because retransmit path
3100 * does not change tp->snd_nxt.
3106 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, sk
->sk_allocation
);
3110 INIT_LIST_HEAD(&skb
->tcp_tsorted_anchor
);
3111 skb_reserve(skb
, MAX_TCP_HEADER
);
3112 sk_forced_mem_schedule(sk
, skb
->truesize
);
3113 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3114 tcp_init_nondata_skb(skb
, tp
->write_seq
,
3115 TCPHDR_ACK
| TCPHDR_FIN
);
3116 tcp_queue_skb(sk
, skb
);
3118 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
), TCP_NAGLE_OFF
);
3121 /* We get here when a process closes a file descriptor (either due to
3122 * an explicit close() or as a byproduct of exit()'ing) and there
3123 * was unread data in the receive queue. This behavior is recommended
3124 * by RFC 2525, section 2.17. -DaveM
3126 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
3128 struct sk_buff
*skb
;
3130 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
3132 /* NOTE: No TCP options attached and we never retransmit this. */
3133 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
3135 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3139 /* Reserve space for headers and prepare control bits. */
3140 skb_reserve(skb
, MAX_TCP_HEADER
);
3141 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
3142 TCPHDR_ACK
| TCPHDR_RST
);
3143 tcp_mstamp_refresh(tcp_sk(sk
));
3145 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
3146 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3148 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3149 * skb here is different to the troublesome skb, so use NULL
3151 trace_tcp_send_reset(sk
, NULL
);
3154 /* Send a crossed SYN-ACK during socket establishment.
3155 * WARNING: This routine must only be called when we have already sent
3156 * a SYN packet that crossed the incoming SYN that caused this routine
3157 * to get called. If this assumption fails then the initial rcv_wnd
3158 * and rcv_wscale values will not be correct.
3160 int tcp_send_synack(struct sock
*sk
)
3162 struct sk_buff
*skb
;
3164 skb
= tcp_rtx_queue_head(sk
);
3165 if (!skb
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
3166 pr_err("%s: wrong queue state\n", __func__
);
3169 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
3170 if (skb_cloned(skb
)) {
3171 struct sk_buff
*nskb
;
3173 tcp_skb_tsorted_save(skb
) {
3174 nskb
= skb_copy(skb
, GFP_ATOMIC
);
3175 } tcp_skb_tsorted_restore(skb
);
3178 INIT_LIST_HEAD(&nskb
->tcp_tsorted_anchor
);
3179 tcp_rtx_queue_unlink_and_free(skb
, sk
);
3180 __skb_header_release(nskb
);
3181 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, nskb
);
3182 sk
->sk_wmem_queued
+= nskb
->truesize
;
3183 sk_mem_charge(sk
, nskb
->truesize
);
3187 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
3188 tcp_ecn_send_synack(sk
, skb
);
3190 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3194 * tcp_make_synack - Prepare a SYN-ACK.
3195 * sk: listener socket
3196 * dst: dst entry attached to the SYNACK
3197 * req: request_sock pointer
3199 * Allocate one skb and build a SYNACK packet.
3200 * @dst is consumed : Caller should not use it again.
3202 struct sk_buff
*tcp_make_synack(const struct sock
*sk
, struct dst_entry
*dst
,
3203 struct request_sock
*req
,
3204 struct tcp_fastopen_cookie
*foc
,
3205 enum tcp_synack_type synack_type
)
3207 struct inet_request_sock
*ireq
= inet_rsk(req
);
3208 const struct tcp_sock
*tp
= tcp_sk(sk
);
3209 struct tcp_md5sig_key
*md5
= NULL
;
3210 struct tcp_out_options opts
;
3211 struct sk_buff
*skb
;
3212 int tcp_header_size
;
3216 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
3217 if (unlikely(!skb
)) {
3221 /* Reserve space for headers. */
3222 skb_reserve(skb
, MAX_TCP_HEADER
);
3224 switch (synack_type
) {
3225 case TCP_SYNACK_NORMAL
:
3226 skb_set_owner_w(skb
, req_to_sk(req
));
3228 case TCP_SYNACK_COOKIE
:
3229 /* Under synflood, we do not attach skb to a socket,
3230 * to avoid false sharing.
3233 case TCP_SYNACK_FASTOPEN
:
3234 /* sk is a const pointer, because we want to express multiple
3235 * cpu might call us concurrently.
3236 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3238 skb_set_owner_w(skb
, (struct sock
*)sk
);
3241 skb_dst_set(skb
, dst
);
3243 mss
= tcp_mss_clamp(tp
, dst_metric_advmss(dst
));
3245 memset(&opts
, 0, sizeof(opts
));
3246 #ifdef CONFIG_SYN_COOKIES
3247 if (unlikely(req
->cookie_ts
))
3248 skb
->skb_mstamp_ns
= cookie_init_timestamp(req
);
3252 skb
->skb_mstamp_ns
= tcp_clock_ns();
3253 if (!tcp_rsk(req
)->snt_synack
) /* Timestamp first SYNACK */
3254 tcp_rsk(req
)->snt_synack
= tcp_skb_timestamp_us(skb
);
3257 #ifdef CONFIG_TCP_MD5SIG
3259 md5
= tcp_rsk(req
)->af_specific
->req_md5_lookup(sk
, req_to_sk(req
));
3261 skb_set_hash(skb
, tcp_rsk(req
)->txhash
, PKT_HASH_TYPE_L4
);
3262 tcp_header_size
= tcp_synack_options(sk
, req
, mss
, skb
, &opts
, md5
,
3265 skb_push(skb
, tcp_header_size
);
3266 skb_reset_transport_header(skb
);
3268 th
= (struct tcphdr
*)skb
->data
;
3269 memset(th
, 0, sizeof(struct tcphdr
));
3272 tcp_ecn_make_synack(req
, th
);
3273 th
->source
= htons(ireq
->ir_num
);
3274 th
->dest
= ireq
->ir_rmt_port
;
3275 skb
->mark
= ireq
->ir_mark
;
3276 skb
->ip_summed
= CHECKSUM_PARTIAL
;
3277 th
->seq
= htonl(tcp_rsk(req
)->snt_isn
);
3278 /* XXX data is queued and acked as is. No buffer/window check */
3279 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
3281 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3282 th
->window
= htons(min(req
->rsk_rcv_wnd
, 65535U));
3283 tcp_options_write((__be32
*)(th
+ 1), NULL
, &opts
);
3284 th
->doff
= (tcp_header_size
>> 2);
3285 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
3287 #ifdef CONFIG_TCP_MD5SIG
3288 /* Okay, we have all we need - do the md5 hash if needed */
3290 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
3291 md5
, req_to_sk(req
), skb
);
3295 /* Do not fool tcpdump (if any), clean our debris */
3299 EXPORT_SYMBOL(tcp_make_synack
);
3301 static void tcp_ca_dst_init(struct sock
*sk
, const struct dst_entry
*dst
)
3303 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3304 const struct tcp_congestion_ops
*ca
;
3305 u32 ca_key
= dst_metric(dst
, RTAX_CC_ALGO
);
3307 if (ca_key
== TCP_CA_UNSPEC
)
3311 ca
= tcp_ca_find_key(ca_key
);
3312 if (likely(ca
&& try_module_get(ca
->owner
))) {
3313 module_put(icsk
->icsk_ca_ops
->owner
);
3314 icsk
->icsk_ca_dst_locked
= tcp_ca_dst_locked(dst
);
3315 icsk
->icsk_ca_ops
= ca
;
3320 /* Do all connect socket setups that can be done AF independent. */
3321 static void tcp_connect_init(struct sock
*sk
)
3323 const struct dst_entry
*dst
= __sk_dst_get(sk
);
3324 struct tcp_sock
*tp
= tcp_sk(sk
);
3328 /* We'll fix this up when we get a response from the other end.
3329 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3331 tp
->tcp_header_len
= sizeof(struct tcphdr
);
3332 if (sock_net(sk
)->ipv4
.sysctl_tcp_timestamps
)
3333 tp
->tcp_header_len
+= TCPOLEN_TSTAMP_ALIGNED
;
3335 #ifdef CONFIG_TCP_MD5SIG
3336 if (tp
->af_specific
->md5_lookup(sk
, sk
))
3337 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
3340 /* If user gave his TCP_MAXSEG, record it to clamp */
3341 if (tp
->rx_opt
.user_mss
)
3342 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3345 tcp_sync_mss(sk
, dst_mtu(dst
));
3347 tcp_ca_dst_init(sk
, dst
);
3349 if (!tp
->window_clamp
)
3350 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
3351 tp
->advmss
= tcp_mss_clamp(tp
, dst_metric_advmss(dst
));
3353 tcp_initialize_rcv_mss(sk
);
3355 /* limit the window selection if the user enforce a smaller rx buffer */
3356 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
3357 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
3358 tp
->window_clamp
= tcp_full_space(sk
);
3360 rcv_wnd
= tcp_rwnd_init_bpf(sk
);
3362 rcv_wnd
= dst_metric(dst
, RTAX_INITRWND
);
3364 tcp_select_initial_window(sk
, tcp_full_space(sk
),
3365 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
3368 sock_net(sk
)->ipv4
.sysctl_tcp_window_scaling
,
3372 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
3373 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
3376 sock_reset_flag(sk
, SOCK_DONE
);
3379 tcp_write_queue_purge(sk
);
3380 tp
->snd_una
= tp
->write_seq
;
3381 tp
->snd_sml
= tp
->write_seq
;
3382 tp
->snd_up
= tp
->write_seq
;
3383 tp
->snd_nxt
= tp
->write_seq
;
3385 if (likely(!tp
->repair
))
3388 tp
->rcv_tstamp
= tcp_jiffies32
;
3389 tp
->rcv_wup
= tp
->rcv_nxt
;
3390 tp
->copied_seq
= tp
->rcv_nxt
;
3392 inet_csk(sk
)->icsk_rto
= tcp_timeout_init(sk
);
3393 inet_csk(sk
)->icsk_retransmits
= 0;
3394 tcp_clear_retrans(tp
);
3397 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
3399 struct tcp_sock
*tp
= tcp_sk(sk
);
3400 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
3402 tcb
->end_seq
+= skb
->len
;
3403 __skb_header_release(skb
);
3404 sk
->sk_wmem_queued
+= skb
->truesize
;
3405 sk_mem_charge(sk
, skb
->truesize
);
3406 tp
->write_seq
= tcb
->end_seq
;
3407 tp
->packets_out
+= tcp_skb_pcount(skb
);
3410 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3411 * queue a data-only packet after the regular SYN, such that regular SYNs
3412 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3413 * only the SYN sequence, the data are retransmitted in the first ACK.
3414 * If cookie is not cached or other error occurs, falls back to send a
3415 * regular SYN with Fast Open cookie request option.
3417 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
3419 struct tcp_sock
*tp
= tcp_sk(sk
);
3420 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
3422 struct sk_buff
*syn_data
;
3424 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
3425 if (!tcp_fastopen_cookie_check(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
))
3428 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3429 * user-MSS. Reserve maximum option space for middleboxes that add
3430 * private TCP options. The cost is reduced data space in SYN :(
3432 tp
->rx_opt
.mss_clamp
= tcp_mss_clamp(tp
, tp
->rx_opt
.mss_clamp
);
3434 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
3435 MAX_TCP_OPTION_SPACE
;
3437 space
= min_t(size_t, space
, fo
->size
);
3439 /* limit to order-0 allocations */
3440 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
3442 syn_data
= sk_stream_alloc_skb(sk
, space
, sk
->sk_allocation
, false);
3445 syn_data
->ip_summed
= CHECKSUM_PARTIAL
;
3446 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
3448 int copied
= copy_from_iter(skb_put(syn_data
, space
), space
,
3449 &fo
->data
->msg_iter
);
3450 if (unlikely(!copied
)) {
3451 tcp_skb_tsorted_anchor_cleanup(syn_data
);
3452 kfree_skb(syn_data
);
3455 if (copied
!= space
) {
3456 skb_trim(syn_data
, copied
);
3459 skb_zcopy_set(syn_data
, fo
->uarg
, NULL
);
3461 /* No more data pending in inet_wait_for_connect() */
3462 if (space
== fo
->size
)
3466 tcp_connect_queue_skb(sk
, syn_data
);
3468 tcp_chrono_start(sk
, TCP_CHRONO_BUSY
);
3470 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
3472 syn
->skb_mstamp_ns
= syn_data
->skb_mstamp_ns
;
3474 /* Now full SYN+DATA was cloned and sent (or not),
3475 * remove the SYN from the original skb (syn_data)
3476 * we keep in write queue in case of a retransmit, as we
3477 * also have the SYN packet (with no data) in the same queue.
3479 TCP_SKB_CB(syn_data
)->seq
++;
3480 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
3482 tp
->syn_data
= (fo
->copied
> 0);
3483 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, syn_data
);
3484 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3488 /* data was not sent, put it in write_queue */
3489 __skb_queue_tail(&sk
->sk_write_queue
, syn_data
);
3490 tp
->packets_out
-= tcp_skb_pcount(syn_data
);
3493 /* Send a regular SYN with Fast Open cookie request option */
3494 if (fo
->cookie
.len
> 0)
3496 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3498 tp
->syn_fastopen
= 0;
3500 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3504 /* Build a SYN and send it off. */
3505 int tcp_connect(struct sock
*sk
)
3507 struct tcp_sock
*tp
= tcp_sk(sk
);
3508 struct sk_buff
*buff
;
3511 tcp_call_bpf(sk
, BPF_SOCK_OPS_TCP_CONNECT_CB
, 0, NULL
);
3513 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
3514 return -EHOSTUNREACH
; /* Routing failure or similar. */
3516 tcp_connect_init(sk
);
3518 if (unlikely(tp
->repair
)) {
3519 tcp_finish_connect(sk
, NULL
);
3523 buff
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
, true);
3524 if (unlikely(!buff
))
3527 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3528 tcp_mstamp_refresh(tp
);
3529 tp
->retrans_stamp
= tcp_time_stamp(tp
);
3530 tcp_connect_queue_skb(sk
, buff
);
3531 tcp_ecn_send_syn(sk
, buff
);
3532 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, buff
);
3534 /* Send off SYN; include data in Fast Open. */
3535 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3536 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3537 if (err
== -ECONNREFUSED
)
3540 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3541 * in order to make this packet get counted in tcpOutSegs.
3543 tp
->snd_nxt
= tp
->write_seq
;
3544 tp
->pushed_seq
= tp
->write_seq
;
3545 buff
= tcp_send_head(sk
);
3546 if (unlikely(buff
)) {
3547 tp
->snd_nxt
= TCP_SKB_CB(buff
)->seq
;
3548 tp
->pushed_seq
= TCP_SKB_CB(buff
)->seq
;
3550 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3552 /* Timer for repeating the SYN until an answer. */
3553 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3554 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3557 EXPORT_SYMBOL(tcp_connect
);
3559 /* Send out a delayed ack, the caller does the policy checking
3560 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3563 void tcp_send_delayed_ack(struct sock
*sk
)
3565 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3566 int ato
= icsk
->icsk_ack
.ato
;
3567 unsigned long timeout
;
3569 if (ato
> TCP_DELACK_MIN
) {
3570 const struct tcp_sock
*tp
= tcp_sk(sk
);
3571 int max_ato
= HZ
/ 2;
3573 if (inet_csk_in_pingpong_mode(sk
) ||
3574 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3575 max_ato
= TCP_DELACK_MAX
;
3577 /* Slow path, intersegment interval is "high". */
3579 /* If some rtt estimate is known, use it to bound delayed ack.
3580 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3584 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3591 ato
= min(ato
, max_ato
);
3594 /* Stay within the limit we were given */
3595 timeout
= jiffies
+ ato
;
3597 /* Use new timeout only if there wasn't a older one earlier. */
3598 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3599 /* If delack timer was blocked or is about to expire,
3602 if (icsk
->icsk_ack
.blocked
||
3603 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3608 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3609 timeout
= icsk
->icsk_ack
.timeout
;
3611 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3612 icsk
->icsk_ack
.timeout
= timeout
;
3613 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3616 /* This routine sends an ack and also updates the window. */
3617 void __tcp_send_ack(struct sock
*sk
, u32 rcv_nxt
)
3619 struct sk_buff
*buff
;
3621 /* If we have been reset, we may not send again. */
3622 if (sk
->sk_state
== TCP_CLOSE
)
3625 /* We are not putting this on the write queue, so
3626 * tcp_transmit_skb() will set the ownership to this
3629 buff
= alloc_skb(MAX_TCP_HEADER
,
3630 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3631 if (unlikely(!buff
)) {
3632 inet_csk_schedule_ack(sk
);
3633 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3634 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3635 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3639 /* Reserve space for headers and prepare control bits. */
3640 skb_reserve(buff
, MAX_TCP_HEADER
);
3641 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3643 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3645 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3647 skb_set_tcp_pure_ack(buff
);
3649 /* Send it off, this clears delayed acks for us. */
3650 __tcp_transmit_skb(sk
, buff
, 0, (__force gfp_t
)0, rcv_nxt
);
3652 EXPORT_SYMBOL_GPL(__tcp_send_ack
);
3654 void tcp_send_ack(struct sock
*sk
)
3656 __tcp_send_ack(sk
, tcp_sk(sk
)->rcv_nxt
);
3659 /* This routine sends a packet with an out of date sequence
3660 * number. It assumes the other end will try to ack it.
3662 * Question: what should we make while urgent mode?
3663 * 4.4BSD forces sending single byte of data. We cannot send
3664 * out of window data, because we have SND.NXT==SND.MAX...
3666 * Current solution: to send TWO zero-length segments in urgent mode:
3667 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3668 * out-of-date with SND.UNA-1 to probe window.
3670 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
, int mib
)
3672 struct tcp_sock
*tp
= tcp_sk(sk
);
3673 struct sk_buff
*skb
;
3675 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3676 skb
= alloc_skb(MAX_TCP_HEADER
,
3677 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3681 /* Reserve space for headers and set control bits. */
3682 skb_reserve(skb
, MAX_TCP_HEADER
);
3683 /* Use a previous sequence. This should cause the other
3684 * end to send an ack. Don't queue or clone SKB, just
3687 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3688 NET_INC_STATS(sock_net(sk
), mib
);
3689 return tcp_transmit_skb(sk
, skb
, 0, (__force gfp_t
)0);
3692 /* Called from setsockopt( ... TCP_REPAIR ) */
3693 void tcp_send_window_probe(struct sock
*sk
)
3695 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3696 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3697 tcp_mstamp_refresh(tcp_sk(sk
));
3698 tcp_xmit_probe_skb(sk
, 0, LINUX_MIB_TCPWINPROBE
);
3702 /* Initiate keepalive or window probe from timer. */
3703 int tcp_write_wakeup(struct sock
*sk
, int mib
)
3705 struct tcp_sock
*tp
= tcp_sk(sk
);
3706 struct sk_buff
*skb
;
3708 if (sk
->sk_state
== TCP_CLOSE
)
3711 skb
= tcp_send_head(sk
);
3712 if (skb
&& before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3714 unsigned int mss
= tcp_current_mss(sk
);
3715 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3717 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3718 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3720 /* We are probing the opening of a window
3721 * but the window size is != 0
3722 * must have been a result SWS avoidance ( sender )
3724 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3726 seg_size
= min(seg_size
, mss
);
3727 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3728 if (tcp_fragment(sk
, TCP_FRAG_IN_WRITE_QUEUE
,
3729 skb
, seg_size
, mss
, GFP_ATOMIC
))
3731 } else if (!tcp_skb_pcount(skb
))
3732 tcp_set_skb_tso_segs(skb
, mss
);
3734 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3735 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3737 tcp_event_new_data_sent(sk
, skb
);
3740 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3741 tcp_xmit_probe_skb(sk
, 1, mib
);
3742 return tcp_xmit_probe_skb(sk
, 0, mib
);
3746 /* A window probe timeout has occurred. If window is not closed send
3747 * a partial packet else a zero probe.
3749 void tcp_send_probe0(struct sock
*sk
)
3751 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3752 struct tcp_sock
*tp
= tcp_sk(sk
);
3753 struct net
*net
= sock_net(sk
);
3754 unsigned long timeout
;
3757 err
= tcp_write_wakeup(sk
, LINUX_MIB_TCPWINPROBE
);
3759 if (tp
->packets_out
|| tcp_write_queue_empty(sk
)) {
3760 /* Cancel probe timer, if it is not required. */
3761 icsk
->icsk_probes_out
= 0;
3762 icsk
->icsk_backoff
= 0;
3766 icsk
->icsk_probes_out
++;
3768 if (icsk
->icsk_backoff
< net
->ipv4
.sysctl_tcp_retries2
)
3769 icsk
->icsk_backoff
++;
3770 timeout
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
3772 /* If packet was not sent due to local congestion,
3773 * Let senders fight for local resources conservatively.
3775 timeout
= TCP_RESOURCE_PROBE_INTERVAL
;
3777 tcp_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
, timeout
, TCP_RTO_MAX
, NULL
);
3780 int tcp_rtx_synack(const struct sock
*sk
, struct request_sock
*req
)
3782 const struct tcp_request_sock_ops
*af_ops
= tcp_rsk(req
)->af_specific
;
3786 tcp_rsk(req
)->txhash
= net_tx_rndhash();
3787 res
= af_ops
->send_synack(sk
, NULL
, &fl
, req
, NULL
, TCP_SYNACK_NORMAL
);
3789 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
3790 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
3791 if (unlikely(tcp_passive_fastopen(sk
)))
3792 tcp_sk(sk
)->total_retrans
++;
3793 trace_tcp_retransmit_synack(sk
, req
);
3797 EXPORT_SYMBOL(tcp_rtx_synack
);