2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
44 #include <linux/static_key.h>
46 #include <trace/events/tcp.h>
48 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
49 int push_one
, gfp_t gfp
);
51 /* Account for new data that has been sent to the network. */
52 static void tcp_event_new_data_sent(struct sock
*sk
, struct sk_buff
*skb
)
54 struct inet_connection_sock
*icsk
= inet_csk(sk
);
55 struct tcp_sock
*tp
= tcp_sk(sk
);
56 unsigned int prior_packets
= tp
->packets_out
;
58 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
60 __skb_unlink(skb
, &sk
->sk_write_queue
);
61 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, skb
);
63 if (tp
->highest_sack
== NULL
)
64 tp
->highest_sack
= skb
;
66 tp
->packets_out
+= tcp_skb_pcount(skb
);
67 if (!prior_packets
|| icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
70 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
,
74 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
75 * window scaling factor due to loss of precision.
76 * If window has been shrunk, what should we make? It is not clear at all.
77 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
78 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
79 * invalid. OK, let's make this for now:
81 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
83 const struct tcp_sock
*tp
= tcp_sk(sk
);
85 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
) ||
86 (tp
->rx_opt
.wscale_ok
&&
87 ((tp
->snd_nxt
- tcp_wnd_end(tp
)) < (1 << tp
->rx_opt
.rcv_wscale
))))
90 return tcp_wnd_end(tp
);
93 /* Calculate mss to advertise in SYN segment.
94 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
96 * 1. It is independent of path mtu.
97 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
98 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
99 * attached devices, because some buggy hosts are confused by
101 * 4. We do not make 3, we advertise MSS, calculated from first
102 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
103 * This may be overridden via information stored in routing table.
104 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
105 * probably even Jumbo".
107 static __u16
tcp_advertise_mss(struct sock
*sk
)
109 struct tcp_sock
*tp
= tcp_sk(sk
);
110 const struct dst_entry
*dst
= __sk_dst_get(sk
);
111 int mss
= tp
->advmss
;
114 unsigned int metric
= dst_metric_advmss(dst
);
125 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
126 * This is the first part of cwnd validation mechanism.
128 void tcp_cwnd_restart(struct sock
*sk
, s32 delta
)
130 struct tcp_sock
*tp
= tcp_sk(sk
);
131 u32 restart_cwnd
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
132 u32 cwnd
= tp
->snd_cwnd
;
134 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
136 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
137 restart_cwnd
= min(restart_cwnd
, cwnd
);
139 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
141 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
142 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
143 tp
->snd_cwnd_used
= 0;
146 /* Congestion state accounting after a packet has been sent. */
147 static void tcp_event_data_sent(struct tcp_sock
*tp
,
150 struct inet_connection_sock
*icsk
= inet_csk(sk
);
151 const u32 now
= tcp_jiffies32
;
153 if (tcp_packets_in_flight(tp
) == 0)
154 tcp_ca_event(sk
, CA_EVENT_TX_START
);
158 /* If it is a reply for ato after last received
159 * packet, enter pingpong mode.
161 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
162 icsk
->icsk_ack
.pingpong
= 1;
165 /* Account for an ACK we sent. */
166 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
,
169 struct tcp_sock
*tp
= tcp_sk(sk
);
171 if (unlikely(rcv_nxt
!= tp
->rcv_nxt
))
172 return; /* Special ACK sent by DCTCP to reflect ECN */
173 tcp_dec_quickack_mode(sk
, pkts
);
174 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
177 /* Determine a window scaling and initial window to offer.
178 * Based on the assumption that the given amount of space
179 * will be offered. Store the results in the tp structure.
180 * NOTE: for smooth operation initial space offering should
181 * be a multiple of mss if possible. We assume here that mss >= 1.
182 * This MUST be enforced by all callers.
184 void tcp_select_initial_window(const struct sock
*sk
, int __space
, __u32 mss
,
185 __u32
*rcv_wnd
, __u32
*window_clamp
,
186 int wscale_ok
, __u8
*rcv_wscale
,
189 unsigned int space
= (__space
< 0 ? 0 : __space
);
191 /* If no clamp set the clamp to the max possible scaled window */
192 if (*window_clamp
== 0)
193 (*window_clamp
) = (U16_MAX
<< TCP_MAX_WSCALE
);
194 space
= min(*window_clamp
, space
);
196 /* Quantize space offering to a multiple of mss if possible. */
198 space
= rounddown(space
, mss
);
200 /* NOTE: offering an initial window larger than 32767
201 * will break some buggy TCP stacks. If the admin tells us
202 * it is likely we could be speaking with such a buggy stack
203 * we will truncate our initial window offering to 32K-1
204 * unless the remote has sent us a window scaling option,
205 * which we interpret as a sign the remote TCP is not
206 * misinterpreting the window field as a signed quantity.
208 if (sock_net(sk
)->ipv4
.sysctl_tcp_workaround_signed_windows
)
209 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
211 (*rcv_wnd
) = min_t(u32
, space
, U16_MAX
);
214 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
218 /* Set window scaling on max possible window */
219 space
= max_t(u32
, space
, sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[2]);
220 space
= max_t(u32
, space
, sysctl_rmem_max
);
221 space
= min_t(u32
, space
, *window_clamp
);
222 while (space
> U16_MAX
&& (*rcv_wscale
) < TCP_MAX_WSCALE
) {
228 /* Set the clamp no higher than max representable value */
229 (*window_clamp
) = min_t(__u32
, U16_MAX
<< (*rcv_wscale
), *window_clamp
);
231 EXPORT_SYMBOL(tcp_select_initial_window
);
233 /* Chose a new window to advertise, update state in tcp_sock for the
234 * socket, and return result with RFC1323 scaling applied. The return
235 * value can be stuffed directly into th->window for an outgoing
238 static u16
tcp_select_window(struct sock
*sk
)
240 struct tcp_sock
*tp
= tcp_sk(sk
);
241 u32 old_win
= tp
->rcv_wnd
;
242 u32 cur_win
= tcp_receive_window(tp
);
243 u32 new_win
= __tcp_select_window(sk
);
245 /* Never shrink the offered window */
246 if (new_win
< cur_win
) {
247 /* Danger Will Robinson!
248 * Don't update rcv_wup/rcv_wnd here or else
249 * we will not be able to advertise a zero
250 * window in time. --DaveM
252 * Relax Will Robinson.
255 NET_INC_STATS(sock_net(sk
),
256 LINUX_MIB_TCPWANTZEROWINDOWADV
);
257 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
259 tp
->rcv_wnd
= new_win
;
260 tp
->rcv_wup
= tp
->rcv_nxt
;
262 /* Make sure we do not exceed the maximum possible
265 if (!tp
->rx_opt
.rcv_wscale
&&
266 sock_net(sk
)->ipv4
.sysctl_tcp_workaround_signed_windows
)
267 new_win
= min(new_win
, MAX_TCP_WINDOW
);
269 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
271 /* RFC1323 scaling applied */
272 new_win
>>= tp
->rx_opt
.rcv_wscale
;
274 /* If we advertise zero window, disable fast path. */
278 NET_INC_STATS(sock_net(sk
),
279 LINUX_MIB_TCPTOZEROWINDOWADV
);
280 } else if (old_win
== 0) {
281 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFROMZEROWINDOWADV
);
287 /* Packet ECN state for a SYN-ACK */
288 static void tcp_ecn_send_synack(struct sock
*sk
, struct sk_buff
*skb
)
290 const struct tcp_sock
*tp
= tcp_sk(sk
);
292 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
293 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
294 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
295 else if (tcp_ca_needs_ecn(sk
) ||
296 tcp_bpf_ca_needs_ecn(sk
))
300 /* Packet ECN state for a SYN. */
301 static void tcp_ecn_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
303 struct tcp_sock
*tp
= tcp_sk(sk
);
304 bool bpf_needs_ecn
= tcp_bpf_ca_needs_ecn(sk
);
305 bool use_ecn
= sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1 ||
306 tcp_ca_needs_ecn(sk
) || bpf_needs_ecn
;
309 const struct dst_entry
*dst
= __sk_dst_get(sk
);
311 if (dst
&& dst_feature(dst
, RTAX_FEATURE_ECN
))
318 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
319 tp
->ecn_flags
= TCP_ECN_OK
;
320 if (tcp_ca_needs_ecn(sk
) || bpf_needs_ecn
)
325 static void tcp_ecn_clear_syn(struct sock
*sk
, struct sk_buff
*skb
)
327 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn_fallback
)
328 /* tp->ecn_flags are cleared at a later point in time when
329 * SYN ACK is ultimatively being received.
331 TCP_SKB_CB(skb
)->tcp_flags
&= ~(TCPHDR_ECE
| TCPHDR_CWR
);
335 tcp_ecn_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
337 if (inet_rsk(req
)->ecn_ok
)
341 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
344 static void tcp_ecn_send(struct sock
*sk
, struct sk_buff
*skb
,
345 struct tcphdr
*th
, int tcp_header_len
)
347 struct tcp_sock
*tp
= tcp_sk(sk
);
349 if (tp
->ecn_flags
& TCP_ECN_OK
) {
350 /* Not-retransmitted data segment: set ECT and inject CWR. */
351 if (skb
->len
!= tcp_header_len
&&
352 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
354 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
355 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
357 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
359 } else if (!tcp_ca_needs_ecn(sk
)) {
360 /* ACK or retransmitted segment: clear ECT|CE */
361 INET_ECN_dontxmit(sk
);
363 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
368 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
369 * auto increment end seqno.
371 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
373 skb
->ip_summed
= CHECKSUM_PARTIAL
;
375 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
376 TCP_SKB_CB(skb
)->sacked
= 0;
378 tcp_skb_pcount_set(skb
, 1);
380 TCP_SKB_CB(skb
)->seq
= seq
;
381 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
383 TCP_SKB_CB(skb
)->end_seq
= seq
;
386 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
388 return tp
->snd_una
!= tp
->snd_up
;
391 #define OPTION_SACK_ADVERTISE (1 << 0)
392 #define OPTION_TS (1 << 1)
393 #define OPTION_MD5 (1 << 2)
394 #define OPTION_WSCALE (1 << 3)
395 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
396 #define OPTION_SMC (1 << 9)
398 static void smc_options_write(__be32
*ptr
, u16
*options
)
400 #if IS_ENABLED(CONFIG_SMC)
401 if (static_branch_unlikely(&tcp_have_smc
)) {
402 if (unlikely(OPTION_SMC
& *options
)) {
403 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
406 (TCPOLEN_EXP_SMC_BASE
));
407 *ptr
++ = htonl(TCPOPT_SMC_MAGIC
);
413 struct tcp_out_options
{
414 u16 options
; /* bit field of OPTION_* */
415 u16 mss
; /* 0 to disable */
416 u8 ws
; /* window scale, 0 to disable */
417 u8 num_sack_blocks
; /* number of SACK blocks to include */
418 u8 hash_size
; /* bytes in hash_location */
419 __u8
*hash_location
; /* temporary pointer, overloaded */
420 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
421 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
424 /* Write previously computed TCP options to the packet.
426 * Beware: Something in the Internet is very sensitive to the ordering of
427 * TCP options, we learned this through the hard way, so be careful here.
428 * Luckily we can at least blame others for their non-compliance but from
429 * inter-operability perspective it seems that we're somewhat stuck with
430 * the ordering which we have been using if we want to keep working with
431 * those broken things (not that it currently hurts anybody as there isn't
432 * particular reason why the ordering would need to be changed).
434 * At least SACK_PERM as the first option is known to lead to a disaster
435 * (but it may well be that other scenarios fail similarly).
437 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
438 struct tcp_out_options
*opts
)
440 u16 options
= opts
->options
; /* mungable copy */
442 if (unlikely(OPTION_MD5
& options
)) {
443 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
444 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
445 /* overload cookie hash location */
446 opts
->hash_location
= (__u8
*)ptr
;
450 if (unlikely(opts
->mss
)) {
451 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
452 (TCPOLEN_MSS
<< 16) |
456 if (likely(OPTION_TS
& options
)) {
457 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
458 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
459 (TCPOLEN_SACK_PERM
<< 16) |
460 (TCPOPT_TIMESTAMP
<< 8) |
462 options
&= ~OPTION_SACK_ADVERTISE
;
464 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
466 (TCPOPT_TIMESTAMP
<< 8) |
469 *ptr
++ = htonl(opts
->tsval
);
470 *ptr
++ = htonl(opts
->tsecr
);
473 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
474 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
476 (TCPOPT_SACK_PERM
<< 8) |
480 if (unlikely(OPTION_WSCALE
& options
)) {
481 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
482 (TCPOPT_WINDOW
<< 16) |
483 (TCPOLEN_WINDOW
<< 8) |
487 if (unlikely(opts
->num_sack_blocks
)) {
488 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
489 tp
->duplicate_sack
: tp
->selective_acks
;
492 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
495 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
496 TCPOLEN_SACK_PERBLOCK
)));
498 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
500 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
501 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
504 tp
->rx_opt
.dsack
= 0;
507 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
508 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
510 u32 len
; /* Fast Open option length */
513 len
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
514 *ptr
= htonl((TCPOPT_EXP
<< 24) | (len
<< 16) |
515 TCPOPT_FASTOPEN_MAGIC
);
516 p
+= TCPOLEN_EXP_FASTOPEN_BASE
;
518 len
= TCPOLEN_FASTOPEN_BASE
+ foc
->len
;
519 *p
++ = TCPOPT_FASTOPEN
;
523 memcpy(p
, foc
->val
, foc
->len
);
524 if ((len
& 3) == 2) {
525 p
[foc
->len
] = TCPOPT_NOP
;
526 p
[foc
->len
+ 1] = TCPOPT_NOP
;
528 ptr
+= (len
+ 3) >> 2;
531 smc_options_write(ptr
, &options
);
534 static void smc_set_option(const struct tcp_sock
*tp
,
535 struct tcp_out_options
*opts
,
536 unsigned int *remaining
)
538 #if IS_ENABLED(CONFIG_SMC)
539 if (static_branch_unlikely(&tcp_have_smc
)) {
541 if (*remaining
>= TCPOLEN_EXP_SMC_BASE_ALIGNED
) {
542 opts
->options
|= OPTION_SMC
;
543 *remaining
-= TCPOLEN_EXP_SMC_BASE_ALIGNED
;
550 static void smc_set_option_cond(const struct tcp_sock
*tp
,
551 const struct inet_request_sock
*ireq
,
552 struct tcp_out_options
*opts
,
553 unsigned int *remaining
)
555 #if IS_ENABLED(CONFIG_SMC)
556 if (static_branch_unlikely(&tcp_have_smc
)) {
557 if (tp
->syn_smc
&& ireq
->smc_ok
) {
558 if (*remaining
>= TCPOLEN_EXP_SMC_BASE_ALIGNED
) {
559 opts
->options
|= OPTION_SMC
;
560 *remaining
-= TCPOLEN_EXP_SMC_BASE_ALIGNED
;
567 /* Compute TCP options for SYN packets. This is not the final
568 * network wire format yet.
570 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
571 struct tcp_out_options
*opts
,
572 struct tcp_md5sig_key
**md5
)
574 struct tcp_sock
*tp
= tcp_sk(sk
);
575 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
576 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
578 #ifdef CONFIG_TCP_MD5SIG
579 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
581 opts
->options
|= OPTION_MD5
;
582 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
588 /* We always get an MSS option. The option bytes which will be seen in
589 * normal data packets should timestamps be used, must be in the MSS
590 * advertised. But we subtract them from tp->mss_cache so that
591 * calculations in tcp_sendmsg are simpler etc. So account for this
592 * fact here if necessary. If we don't do this correctly, as a
593 * receiver we won't recognize data packets as being full sized when we
594 * should, and thus we won't abide by the delayed ACK rules correctly.
595 * SACKs don't matter, we never delay an ACK when we have any of those
597 opts
->mss
= tcp_advertise_mss(sk
);
598 remaining
-= TCPOLEN_MSS_ALIGNED
;
600 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_timestamps
&& !*md5
)) {
601 opts
->options
|= OPTION_TS
;
602 opts
->tsval
= tcp_skb_timestamp(skb
) + tp
->tsoffset
;
603 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
604 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
606 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_window_scaling
)) {
607 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
608 opts
->options
|= OPTION_WSCALE
;
609 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
611 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_sack
)) {
612 opts
->options
|= OPTION_SACK_ADVERTISE
;
613 if (unlikely(!(OPTION_TS
& opts
->options
)))
614 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
617 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
618 u32 need
= fastopen
->cookie
.len
;
620 need
+= fastopen
->cookie
.exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
621 TCPOLEN_FASTOPEN_BASE
;
622 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
623 if (remaining
>= need
) {
624 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
625 opts
->fastopen_cookie
= &fastopen
->cookie
;
627 tp
->syn_fastopen
= 1;
628 tp
->syn_fastopen_exp
= fastopen
->cookie
.exp
? 1 : 0;
632 smc_set_option(tp
, opts
, &remaining
);
634 return MAX_TCP_OPTION_SPACE
- remaining
;
637 /* Set up TCP options for SYN-ACKs. */
638 static unsigned int tcp_synack_options(const struct sock
*sk
,
639 struct request_sock
*req
,
640 unsigned int mss
, struct sk_buff
*skb
,
641 struct tcp_out_options
*opts
,
642 const struct tcp_md5sig_key
*md5
,
643 struct tcp_fastopen_cookie
*foc
)
645 struct inet_request_sock
*ireq
= inet_rsk(req
);
646 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
648 #ifdef CONFIG_TCP_MD5SIG
650 opts
->options
|= OPTION_MD5
;
651 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
653 /* We can't fit any SACK blocks in a packet with MD5 + TS
654 * options. There was discussion about disabling SACK
655 * rather than TS in order to fit in better with old,
656 * buggy kernels, but that was deemed to be unnecessary.
658 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
662 /* We always send an MSS option. */
664 remaining
-= TCPOLEN_MSS_ALIGNED
;
666 if (likely(ireq
->wscale_ok
)) {
667 opts
->ws
= ireq
->rcv_wscale
;
668 opts
->options
|= OPTION_WSCALE
;
669 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
671 if (likely(ireq
->tstamp_ok
)) {
672 opts
->options
|= OPTION_TS
;
673 opts
->tsval
= tcp_skb_timestamp(skb
) + tcp_rsk(req
)->ts_off
;
674 opts
->tsecr
= req
->ts_recent
;
675 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
677 if (likely(ireq
->sack_ok
)) {
678 opts
->options
|= OPTION_SACK_ADVERTISE
;
679 if (unlikely(!ireq
->tstamp_ok
))
680 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
682 if (foc
!= NULL
&& foc
->len
>= 0) {
685 need
+= foc
->exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
686 TCPOLEN_FASTOPEN_BASE
;
687 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
688 if (remaining
>= need
) {
689 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
690 opts
->fastopen_cookie
= foc
;
695 smc_set_option_cond(tcp_sk(sk
), ireq
, opts
, &remaining
);
697 return MAX_TCP_OPTION_SPACE
- remaining
;
700 /* Compute TCP options for ESTABLISHED sockets. This is not the
701 * final wire format yet.
703 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
704 struct tcp_out_options
*opts
,
705 struct tcp_md5sig_key
**md5
)
707 struct tcp_sock
*tp
= tcp_sk(sk
);
708 unsigned int size
= 0;
709 unsigned int eff_sacks
;
713 #ifdef CONFIG_TCP_MD5SIG
714 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
715 if (unlikely(*md5
)) {
716 opts
->options
|= OPTION_MD5
;
717 size
+= TCPOLEN_MD5SIG_ALIGNED
;
723 if (likely(tp
->rx_opt
.tstamp_ok
)) {
724 opts
->options
|= OPTION_TS
;
725 opts
->tsval
= skb
? tcp_skb_timestamp(skb
) + tp
->tsoffset
: 0;
726 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
727 size
+= TCPOLEN_TSTAMP_ALIGNED
;
730 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
731 if (unlikely(eff_sacks
)) {
732 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
733 opts
->num_sack_blocks
=
734 min_t(unsigned int, eff_sacks
,
735 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
736 TCPOLEN_SACK_PERBLOCK
);
737 if (likely(opts
->num_sack_blocks
))
738 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
739 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
746 /* TCP SMALL QUEUES (TSQ)
748 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
749 * to reduce RTT and bufferbloat.
750 * We do this using a special skb destructor (tcp_wfree).
752 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
753 * needs to be reallocated in a driver.
754 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
756 * Since transmit from skb destructor is forbidden, we use a tasklet
757 * to process all sockets that eventually need to send more skbs.
758 * We use one tasklet per cpu, with its own queue of sockets.
761 struct tasklet_struct tasklet
;
762 struct list_head head
; /* queue of tcp sockets */
764 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
766 static void tcp_tsq_handler(struct sock
*sk
)
768 if ((1 << sk
->sk_state
) &
769 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
770 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
)) {
771 struct tcp_sock
*tp
= tcp_sk(sk
);
773 if (tp
->lost_out
> tp
->retrans_out
&&
774 tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) {
775 tcp_mstamp_refresh(tp
);
776 tcp_xmit_retransmit_queue(sk
);
779 tcp_write_xmit(sk
, tcp_current_mss(sk
), tp
->nonagle
,
784 * One tasklet per cpu tries to send more skbs.
785 * We run in tasklet context but need to disable irqs when
786 * transferring tsq->head because tcp_wfree() might
787 * interrupt us (non NAPI drivers)
789 static void tcp_tasklet_func(unsigned long data
)
791 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
794 struct list_head
*q
, *n
;
798 local_irq_save(flags
);
799 list_splice_init(&tsq
->head
, &list
);
800 local_irq_restore(flags
);
802 list_for_each_safe(q
, n
, &list
) {
803 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
804 list_del(&tp
->tsq_node
);
806 sk
= (struct sock
*)tp
;
807 smp_mb__before_atomic();
808 clear_bit(TSQ_QUEUED
, &sk
->sk_tsq_flags
);
810 if (!sk
->sk_lock
.owned
&&
811 test_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
)) {
813 if (!sock_owned_by_user(sk
)) {
814 clear_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
);
824 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
825 TCPF_WRITE_TIMER_DEFERRED | \
826 TCPF_DELACK_TIMER_DEFERRED | \
827 TCPF_MTU_REDUCED_DEFERRED)
829 * tcp_release_cb - tcp release_sock() callback
832 * called from release_sock() to perform protocol dependent
833 * actions before socket release.
835 void tcp_release_cb(struct sock
*sk
)
837 unsigned long flags
, nflags
;
839 /* perform an atomic operation only if at least one flag is set */
841 flags
= sk
->sk_tsq_flags
;
842 if (!(flags
& TCP_DEFERRED_ALL
))
844 nflags
= flags
& ~TCP_DEFERRED_ALL
;
845 } while (cmpxchg(&sk
->sk_tsq_flags
, flags
, nflags
) != flags
);
847 if (flags
& TCPF_TSQ_DEFERRED
)
850 /* Here begins the tricky part :
851 * We are called from release_sock() with :
853 * 2) sk_lock.slock spinlock held
854 * 3) socket owned by us (sk->sk_lock.owned == 1)
856 * But following code is meant to be called from BH handlers,
857 * so we should keep BH disabled, but early release socket ownership
859 sock_release_ownership(sk
);
861 if (flags
& TCPF_WRITE_TIMER_DEFERRED
) {
862 tcp_write_timer_handler(sk
);
865 if (flags
& TCPF_DELACK_TIMER_DEFERRED
) {
866 tcp_delack_timer_handler(sk
);
869 if (flags
& TCPF_MTU_REDUCED_DEFERRED
) {
870 inet_csk(sk
)->icsk_af_ops
->mtu_reduced(sk
);
874 EXPORT_SYMBOL(tcp_release_cb
);
876 void __init
tcp_tasklet_init(void)
880 for_each_possible_cpu(i
) {
881 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
883 INIT_LIST_HEAD(&tsq
->head
);
884 tasklet_init(&tsq
->tasklet
,
891 * Write buffer destructor automatically called from kfree_skb.
892 * We can't xmit new skbs from this context, as we might already
895 void tcp_wfree(struct sk_buff
*skb
)
897 struct sock
*sk
= skb
->sk
;
898 struct tcp_sock
*tp
= tcp_sk(sk
);
899 unsigned long flags
, nval
, oval
;
901 /* Keep one reference on sk_wmem_alloc.
902 * Will be released by sk_free() from here or tcp_tasklet_func()
904 WARN_ON(refcount_sub_and_test(skb
->truesize
- 1, &sk
->sk_wmem_alloc
));
906 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
907 * Wait until our queues (qdisc + devices) are drained.
909 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
910 * - chance for incoming ACK (processed by another cpu maybe)
911 * to migrate this flow (skb->ooo_okay will be eventually set)
913 if (refcount_read(&sk
->sk_wmem_alloc
) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current
)
916 for (oval
= READ_ONCE(sk
->sk_tsq_flags
);; oval
= nval
) {
917 struct tsq_tasklet
*tsq
;
920 if (!(oval
& TSQF_THROTTLED
) || (oval
& TSQF_QUEUED
))
923 nval
= (oval
& ~TSQF_THROTTLED
) | TSQF_QUEUED
| TCPF_TSQ_DEFERRED
;
924 nval
= cmpxchg(&sk
->sk_tsq_flags
, oval
, nval
);
928 /* queue this socket to tasklet queue */
929 local_irq_save(flags
);
930 tsq
= this_cpu_ptr(&tsq_tasklet
);
931 empty
= list_empty(&tsq
->head
);
932 list_add(&tp
->tsq_node
, &tsq
->head
);
934 tasklet_schedule(&tsq
->tasklet
);
935 local_irq_restore(flags
);
942 /* Note: Called under hard irq.
943 * We can not call TCP stack right away.
945 enum hrtimer_restart
tcp_pace_kick(struct hrtimer
*timer
)
947 struct tcp_sock
*tp
= container_of(timer
, struct tcp_sock
, pacing_timer
);
948 struct sock
*sk
= (struct sock
*)tp
;
949 unsigned long nval
, oval
;
951 for (oval
= READ_ONCE(sk
->sk_tsq_flags
);; oval
= nval
) {
952 struct tsq_tasklet
*tsq
;
955 if (oval
& TSQF_QUEUED
)
958 nval
= (oval
& ~TSQF_THROTTLED
) | TSQF_QUEUED
| TCPF_TSQ_DEFERRED
;
959 nval
= cmpxchg(&sk
->sk_tsq_flags
, oval
, nval
);
963 if (!refcount_inc_not_zero(&sk
->sk_wmem_alloc
))
965 /* queue this socket to tasklet queue */
966 tsq
= this_cpu_ptr(&tsq_tasklet
);
967 empty
= list_empty(&tsq
->head
);
968 list_add(&tp
->tsq_node
, &tsq
->head
);
970 tasklet_schedule(&tsq
->tasklet
);
973 return HRTIMER_NORESTART
;
976 /* BBR congestion control needs pacing.
977 * Same remark for SO_MAX_PACING_RATE.
978 * sch_fq packet scheduler is efficiently handling pacing,
979 * but is not always installed/used.
980 * Return true if TCP stack should pace packets itself.
982 static bool tcp_needs_internal_pacing(const struct sock
*sk
)
984 return smp_load_acquire(&sk
->sk_pacing_status
) == SK_PACING_NEEDED
;
987 static void tcp_internal_pacing(struct sock
*sk
, const struct sk_buff
*skb
)
992 if (!tcp_needs_internal_pacing(sk
))
994 rate
= sk
->sk_pacing_rate
;
995 if (!rate
|| rate
== ~0U)
998 /* Should account for header sizes as sch_fq does,
999 * but lets make things simple.
1001 len_ns
= (u64
)skb
->len
* NSEC_PER_SEC
;
1002 do_div(len_ns
, rate
);
1003 hrtimer_start(&tcp_sk(sk
)->pacing_timer
,
1004 ktime_add_ns(ktime_get(), len_ns
),
1005 HRTIMER_MODE_ABS_PINNED
);
1008 static void tcp_update_skb_after_send(struct tcp_sock
*tp
, struct sk_buff
*skb
)
1010 skb
->skb_mstamp
= tp
->tcp_mstamp
;
1011 list_move_tail(&skb
->tcp_tsorted_anchor
, &tp
->tsorted_sent_queue
);
1014 /* This routine actually transmits TCP packets queued in by
1015 * tcp_do_sendmsg(). This is used by both the initial
1016 * transmission and possible later retransmissions.
1017 * All SKB's seen here are completely headerless. It is our
1018 * job to build the TCP header, and pass the packet down to
1019 * IP so it can do the same plus pass the packet off to the
1022 * We are working here with either a clone of the original
1023 * SKB, or a fresh unique copy made by the retransmit engine.
1025 static int __tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
,
1026 int clone_it
, gfp_t gfp_mask
, u32 rcv_nxt
)
1028 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1029 struct inet_sock
*inet
;
1030 struct tcp_sock
*tp
;
1031 struct tcp_skb_cb
*tcb
;
1032 struct tcp_out_options opts
;
1033 unsigned int tcp_options_size
, tcp_header_size
;
1034 struct sk_buff
*oskb
= NULL
;
1035 struct tcp_md5sig_key
*md5
;
1039 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
1043 TCP_SKB_CB(skb
)->tx
.in_flight
= TCP_SKB_CB(skb
)->end_seq
1047 tcp_skb_tsorted_save(oskb
) {
1048 if (unlikely(skb_cloned(oskb
)))
1049 skb
= pskb_copy(oskb
, gfp_mask
);
1051 skb
= skb_clone(oskb
, gfp_mask
);
1052 } tcp_skb_tsorted_restore(oskb
);
1057 skb
->skb_mstamp
= tp
->tcp_mstamp
;
1060 tcb
= TCP_SKB_CB(skb
);
1061 memset(&opts
, 0, sizeof(opts
));
1063 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
1064 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
1066 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
1068 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
1070 /* if no packet is in qdisc/device queue, then allow XPS to select
1071 * another queue. We can be called from tcp_tsq_handler()
1072 * which holds one reference to sk_wmem_alloc.
1074 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1075 * One way to get this would be to set skb->truesize = 2 on them.
1077 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) < SKB_TRUESIZE(1);
1079 /* If we had to use memory reserve to allocate this skb,
1080 * this might cause drops if packet is looped back :
1081 * Other socket might not have SOCK_MEMALLOC.
1082 * Packets not looped back do not care about pfmemalloc.
1084 skb
->pfmemalloc
= 0;
1086 skb_push(skb
, tcp_header_size
);
1087 skb_reset_transport_header(skb
);
1091 skb
->destructor
= skb_is_tcp_pure_ack(skb
) ? __sock_wfree
: tcp_wfree
;
1092 skb_set_hash_from_sk(skb
, sk
);
1093 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
1095 skb_set_dst_pending_confirm(skb
, sk
->sk_dst_pending_confirm
);
1097 /* Build TCP header and checksum it. */
1098 th
= (struct tcphdr
*)skb
->data
;
1099 th
->source
= inet
->inet_sport
;
1100 th
->dest
= inet
->inet_dport
;
1101 th
->seq
= htonl(tcb
->seq
);
1102 th
->ack_seq
= htonl(rcv_nxt
);
1103 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
1109 /* The urg_mode check is necessary during a below snd_una win probe */
1110 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
1111 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
1112 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
1114 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
1115 th
->urg_ptr
= htons(0xFFFF);
1120 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
1121 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
1122 if (likely(!(tcb
->tcp_flags
& TCPHDR_SYN
))) {
1123 th
->window
= htons(tcp_select_window(sk
));
1124 tcp_ecn_send(sk
, skb
, th
, tcp_header_size
);
1126 /* RFC1323: The window in SYN & SYN/ACK segments
1129 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
1131 #ifdef CONFIG_TCP_MD5SIG
1132 /* Calculate the MD5 hash, as we have all we need now */
1134 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
1135 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
1140 icsk
->icsk_af_ops
->send_check(sk
, skb
);
1142 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
1143 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
), rcv_nxt
);
1145 if (skb
->len
!= tcp_header_size
) {
1146 tcp_event_data_sent(tp
, sk
);
1147 tp
->data_segs_out
+= tcp_skb_pcount(skb
);
1148 tcp_internal_pacing(sk
, skb
);
1151 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
1152 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
1153 tcp_skb_pcount(skb
));
1155 tp
->segs_out
+= tcp_skb_pcount(skb
);
1156 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1157 skb_shinfo(skb
)->gso_segs
= tcp_skb_pcount(skb
);
1158 skb_shinfo(skb
)->gso_size
= tcp_skb_mss(skb
);
1160 /* Our usage of tstamp should remain private */
1163 /* Cleanup our debris for IP stacks */
1164 memset(skb
->cb
, 0, max(sizeof(struct inet_skb_parm
),
1165 sizeof(struct inet6_skb_parm
)));
1167 err
= icsk
->icsk_af_ops
->queue_xmit(sk
, skb
, &inet
->cork
.fl
);
1169 if (unlikely(err
> 0)) {
1171 err
= net_xmit_eval(err
);
1174 tcp_update_skb_after_send(tp
, oskb
);
1175 tcp_rate_skb_sent(sk
, oskb
);
1180 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
1183 return __tcp_transmit_skb(sk
, skb
, clone_it
, gfp_mask
,
1184 tcp_sk(sk
)->rcv_nxt
);
1187 /* This routine just queues the buffer for sending.
1189 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1190 * otherwise socket can stall.
1192 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1194 struct tcp_sock
*tp
= tcp_sk(sk
);
1196 /* Advance write_seq and place onto the write_queue. */
1197 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
1198 __skb_header_release(skb
);
1199 tcp_add_write_queue_tail(sk
, skb
);
1200 sk
->sk_wmem_queued
+= skb
->truesize
;
1201 sk_mem_charge(sk
, skb
->truesize
);
1204 /* Initialize TSO segments for a packet. */
1205 static void tcp_set_skb_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1207 if (skb
->len
<= mss_now
|| skb
->ip_summed
== CHECKSUM_NONE
) {
1208 /* Avoid the costly divide in the normal
1211 tcp_skb_pcount_set(skb
, 1);
1212 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1214 tcp_skb_pcount_set(skb
, DIV_ROUND_UP(skb
->len
, mss_now
));
1215 TCP_SKB_CB(skb
)->tcp_gso_size
= mss_now
;
1219 /* Pcount in the middle of the write queue got changed, we need to do various
1220 * tweaks to fix counters
1222 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1224 struct tcp_sock
*tp
= tcp_sk(sk
);
1226 tp
->packets_out
-= decr
;
1228 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1229 tp
->sacked_out
-= decr
;
1230 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1231 tp
->retrans_out
-= decr
;
1232 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1233 tp
->lost_out
-= decr
;
1235 /* Reno case is special. Sigh... */
1236 if (tcp_is_reno(tp
) && decr
> 0)
1237 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1239 if (tp
->lost_skb_hint
&&
1240 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1241 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
1242 tp
->lost_cnt_hint
-= decr
;
1244 tcp_verify_left_out(tp
);
1247 static bool tcp_has_tx_tstamp(const struct sk_buff
*skb
)
1249 return TCP_SKB_CB(skb
)->txstamp_ack
||
1250 (skb_shinfo(skb
)->tx_flags
& SKBTX_ANY_TSTAMP
);
1253 static void tcp_fragment_tstamp(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1255 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1257 if (unlikely(tcp_has_tx_tstamp(skb
)) &&
1258 !before(shinfo
->tskey
, TCP_SKB_CB(skb2
)->seq
)) {
1259 struct skb_shared_info
*shinfo2
= skb_shinfo(skb2
);
1260 u8 tsflags
= shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
1262 shinfo
->tx_flags
&= ~tsflags
;
1263 shinfo2
->tx_flags
|= tsflags
;
1264 swap(shinfo
->tskey
, shinfo2
->tskey
);
1265 TCP_SKB_CB(skb2
)->txstamp_ack
= TCP_SKB_CB(skb
)->txstamp_ack
;
1266 TCP_SKB_CB(skb
)->txstamp_ack
= 0;
1270 static void tcp_skb_fragment_eor(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1272 TCP_SKB_CB(skb2
)->eor
= TCP_SKB_CB(skb
)->eor
;
1273 TCP_SKB_CB(skb
)->eor
= 0;
1276 /* Insert buff after skb on the write or rtx queue of sk. */
1277 static void tcp_insert_write_queue_after(struct sk_buff
*skb
,
1278 struct sk_buff
*buff
,
1280 enum tcp_queue tcp_queue
)
1282 if (tcp_queue
== TCP_FRAG_IN_WRITE_QUEUE
)
1283 __skb_queue_after(&sk
->sk_write_queue
, skb
, buff
);
1285 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, buff
);
1288 /* Function to create two new TCP segments. Shrinks the given segment
1289 * to the specified size and appends a new segment with the rest of the
1290 * packet to the list. This won't be called frequently, I hope.
1291 * Remember, these are still headerless SKBs at this point.
1293 int tcp_fragment(struct sock
*sk
, enum tcp_queue tcp_queue
,
1294 struct sk_buff
*skb
, u32 len
,
1295 unsigned int mss_now
, gfp_t gfp
)
1297 struct tcp_sock
*tp
= tcp_sk(sk
);
1298 struct sk_buff
*buff
;
1299 int nsize
, old_factor
;
1304 if (WARN_ON(len
> skb
->len
))
1307 nsize
= skb_headlen(skb
) - len
;
1311 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1312 * We need some allowance to not penalize applications setting small
1314 * Also allow first and last skb in retransmit queue to be split.
1316 limit
= sk
->sk_sndbuf
+ 2 * SKB_TRUESIZE(GSO_MAX_SIZE
);
1317 if (unlikely((sk
->sk_wmem_queued
>> 1) > limit
&&
1318 tcp_queue
!= TCP_FRAG_IN_WRITE_QUEUE
&&
1319 skb
!= tcp_rtx_queue_head(sk
) &&
1320 skb
!= tcp_rtx_queue_tail(sk
))) {
1321 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPWQUEUETOOBIG
);
1325 if (skb_unclone(skb
, gfp
))
1328 /* Get a new skb... force flag on. */
1329 buff
= sk_stream_alloc_skb(sk
, nsize
, gfp
, true);
1331 return -ENOMEM
; /* We'll just try again later. */
1333 sk
->sk_wmem_queued
+= buff
->truesize
;
1334 sk_mem_charge(sk
, buff
->truesize
);
1335 nlen
= skb
->len
- len
- nsize
;
1336 buff
->truesize
+= nlen
;
1337 skb
->truesize
-= nlen
;
1339 /* Correct the sequence numbers. */
1340 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1341 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1342 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1344 /* PSH and FIN should only be set in the second packet. */
1345 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1346 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1347 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1348 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1349 tcp_skb_fragment_eor(skb
, buff
);
1351 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1352 /* Copy and checksum data tail into the new buffer. */
1353 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1354 skb_put(buff
, nsize
),
1359 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1361 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1362 skb_split(skb
, buff
, len
);
1365 buff
->ip_summed
= skb
->ip_summed
;
1367 buff
->tstamp
= skb
->tstamp
;
1368 tcp_fragment_tstamp(skb
, buff
);
1370 old_factor
= tcp_skb_pcount(skb
);
1372 /* Fix up tso_factor for both original and new SKB. */
1373 tcp_set_skb_tso_segs(skb
, mss_now
);
1374 tcp_set_skb_tso_segs(buff
, mss_now
);
1376 /* Update delivered info for the new segment */
1377 TCP_SKB_CB(buff
)->tx
= TCP_SKB_CB(skb
)->tx
;
1379 /* If this packet has been sent out already, we must
1380 * adjust the various packet counters.
1382 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1383 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1384 tcp_skb_pcount(buff
);
1387 tcp_adjust_pcount(sk
, skb
, diff
);
1390 /* Link BUFF into the send queue. */
1391 __skb_header_release(buff
);
1392 tcp_insert_write_queue_after(skb
, buff
, sk
, tcp_queue
);
1393 if (tcp_queue
== TCP_FRAG_IN_RTX_QUEUE
)
1394 list_add(&buff
->tcp_tsorted_anchor
, &skb
->tcp_tsorted_anchor
);
1399 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1400 * data is not copied, but immediately discarded.
1402 static int __pskb_trim_head(struct sk_buff
*skb
, int len
)
1404 struct skb_shared_info
*shinfo
;
1407 eat
= min_t(int, len
, skb_headlen(skb
));
1409 __skb_pull(skb
, eat
);
1416 shinfo
= skb_shinfo(skb
);
1417 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1418 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1421 skb_frag_unref(skb
, i
);
1424 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1426 shinfo
->frags
[k
].page_offset
+= eat
;
1427 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1433 shinfo
->nr_frags
= k
;
1435 skb
->data_len
-= len
;
1436 skb
->len
= skb
->data_len
;
1440 /* Remove acked data from a packet in the transmit queue. */
1441 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1445 if (skb_unclone(skb
, GFP_ATOMIC
))
1448 delta_truesize
= __pskb_trim_head(skb
, len
);
1450 TCP_SKB_CB(skb
)->seq
+= len
;
1451 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1453 if (delta_truesize
) {
1454 skb
->truesize
-= delta_truesize
;
1455 sk
->sk_wmem_queued
-= delta_truesize
;
1456 sk_mem_uncharge(sk
, delta_truesize
);
1457 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1460 /* Any change of skb->len requires recalculation of tso factor. */
1461 if (tcp_skb_pcount(skb
) > 1)
1462 tcp_set_skb_tso_segs(skb
, tcp_skb_mss(skb
));
1467 /* Calculate MSS not accounting any TCP options. */
1468 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1470 const struct tcp_sock
*tp
= tcp_sk(sk
);
1471 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1474 /* Calculate base mss without TCP options:
1475 It is MMS_S - sizeof(tcphdr) of rfc1122
1477 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1479 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1480 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1481 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1483 if (dst
&& dst_allfrag(dst
))
1484 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1487 /* Clamp it (mss_clamp does not include tcp options) */
1488 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1489 mss_now
= tp
->rx_opt
.mss_clamp
;
1491 /* Now subtract optional transport overhead */
1492 mss_now
-= icsk
->icsk_ext_hdr_len
;
1494 /* Then reserve room for full set of TCP options and 8 bytes of data */
1495 mss_now
= max(mss_now
, sock_net(sk
)->ipv4
.sysctl_tcp_min_snd_mss
);
1499 /* Calculate MSS. Not accounting for SACKs here. */
1500 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1502 /* Subtract TCP options size, not including SACKs */
1503 return __tcp_mtu_to_mss(sk
, pmtu
) -
1504 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1507 /* Inverse of above */
1508 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1510 const struct tcp_sock
*tp
= tcp_sk(sk
);
1511 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1515 tp
->tcp_header_len
+
1516 icsk
->icsk_ext_hdr_len
+
1517 icsk
->icsk_af_ops
->net_header_len
;
1519 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1520 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1521 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1523 if (dst
&& dst_allfrag(dst
))
1524 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1528 EXPORT_SYMBOL(tcp_mss_to_mtu
);
1530 /* MTU probing init per socket */
1531 void tcp_mtup_init(struct sock
*sk
)
1533 struct tcp_sock
*tp
= tcp_sk(sk
);
1534 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1535 struct net
*net
= sock_net(sk
);
1537 icsk
->icsk_mtup
.enabled
= net
->ipv4
.sysctl_tcp_mtu_probing
> 1;
1538 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1539 icsk
->icsk_af_ops
->net_header_len
;
1540 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, net
->ipv4
.sysctl_tcp_base_mss
);
1541 icsk
->icsk_mtup
.probe_size
= 0;
1542 if (icsk
->icsk_mtup
.enabled
)
1543 icsk
->icsk_mtup
.probe_timestamp
= tcp_jiffies32
;
1545 EXPORT_SYMBOL(tcp_mtup_init
);
1547 /* This function synchronize snd mss to current pmtu/exthdr set.
1549 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1550 for TCP options, but includes only bare TCP header.
1552 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1553 It is minimum of user_mss and mss received with SYN.
1554 It also does not include TCP options.
1556 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1558 tp->mss_cache is current effective sending mss, including
1559 all tcp options except for SACKs. It is evaluated,
1560 taking into account current pmtu, but never exceeds
1561 tp->rx_opt.mss_clamp.
1563 NOTE1. rfc1122 clearly states that advertised MSS
1564 DOES NOT include either tcp or ip options.
1566 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1567 are READ ONLY outside this function. --ANK (980731)
1569 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1571 struct tcp_sock
*tp
= tcp_sk(sk
);
1572 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1575 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1576 icsk
->icsk_mtup
.search_high
= pmtu
;
1578 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1579 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1581 /* And store cached results */
1582 icsk
->icsk_pmtu_cookie
= pmtu
;
1583 if (icsk
->icsk_mtup
.enabled
)
1584 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1585 tp
->mss_cache
= mss_now
;
1589 EXPORT_SYMBOL(tcp_sync_mss
);
1591 /* Compute the current effective MSS, taking SACKs and IP options,
1592 * and even PMTU discovery events into account.
1594 unsigned int tcp_current_mss(struct sock
*sk
)
1596 const struct tcp_sock
*tp
= tcp_sk(sk
);
1597 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1599 unsigned int header_len
;
1600 struct tcp_out_options opts
;
1601 struct tcp_md5sig_key
*md5
;
1603 mss_now
= tp
->mss_cache
;
1606 u32 mtu
= dst_mtu(dst
);
1607 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1608 mss_now
= tcp_sync_mss(sk
, mtu
);
1611 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1612 sizeof(struct tcphdr
);
1613 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1614 * some common options. If this is an odd packet (because we have SACK
1615 * blocks etc) then our calculated header_len will be different, and
1616 * we have to adjust mss_now correspondingly */
1617 if (header_len
!= tp
->tcp_header_len
) {
1618 int delta
= (int) header_len
- tp
->tcp_header_len
;
1625 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1626 * As additional protections, we do not touch cwnd in retransmission phases,
1627 * and if application hit its sndbuf limit recently.
1629 static void tcp_cwnd_application_limited(struct sock
*sk
)
1631 struct tcp_sock
*tp
= tcp_sk(sk
);
1633 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1634 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1635 /* Limited by application or receiver window. */
1636 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1637 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1638 if (win_used
< tp
->snd_cwnd
) {
1639 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1640 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
1642 tp
->snd_cwnd_used
= 0;
1644 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1647 static void tcp_cwnd_validate(struct sock
*sk
, bool is_cwnd_limited
)
1649 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1650 struct tcp_sock
*tp
= tcp_sk(sk
);
1652 /* Track the maximum number of outstanding packets in each
1653 * window, and remember whether we were cwnd-limited then.
1655 if (!before(tp
->snd_una
, tp
->max_packets_seq
) ||
1656 tp
->packets_out
> tp
->max_packets_out
) {
1657 tp
->max_packets_out
= tp
->packets_out
;
1658 tp
->max_packets_seq
= tp
->snd_nxt
;
1659 tp
->is_cwnd_limited
= is_cwnd_limited
;
1662 if (tcp_is_cwnd_limited(sk
)) {
1663 /* Network is feed fully. */
1664 tp
->snd_cwnd_used
= 0;
1665 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1667 /* Network starves. */
1668 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1669 tp
->snd_cwnd_used
= tp
->packets_out
;
1671 if (sock_net(sk
)->ipv4
.sysctl_tcp_slow_start_after_idle
&&
1672 (s32
)(tcp_jiffies32
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
&&
1673 !ca_ops
->cong_control
)
1674 tcp_cwnd_application_limited(sk
);
1676 /* The following conditions together indicate the starvation
1677 * is caused by insufficient sender buffer:
1678 * 1) just sent some data (see tcp_write_xmit)
1679 * 2) not cwnd limited (this else condition)
1680 * 3) no more data to send (tcp_write_queue_empty())
1681 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1683 if (tcp_write_queue_empty(sk
) && sk
->sk_socket
&&
1684 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
) &&
1685 (1 << sk
->sk_state
) & (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
1686 tcp_chrono_start(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1690 /* Minshall's variant of the Nagle send check. */
1691 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1693 return after(tp
->snd_sml
, tp
->snd_una
) &&
1694 !after(tp
->snd_sml
, tp
->snd_nxt
);
1697 /* Update snd_sml if this skb is under mss
1698 * Note that a TSO packet might end with a sub-mss segment
1699 * The test is really :
1700 * if ((skb->len % mss) != 0)
1701 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1702 * But we can avoid doing the divide again given we already have
1703 * skb_pcount = skb->len / mss_now
1705 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1706 const struct sk_buff
*skb
)
1708 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1709 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1712 /* Return false, if packet can be sent now without violation Nagle's rules:
1713 * 1. It is full sized. (provided by caller in %partial bool)
1714 * 2. Or it contains FIN. (already checked by caller)
1715 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1716 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1717 * With Minshall's modification: all sent small packets are ACKed.
1719 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1723 ((nonagle
& TCP_NAGLE_CORK
) ||
1724 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1727 /* Return how many segs we'd like on a TSO packet,
1728 * to send one TSO packet per ms
1730 u32
tcp_tso_autosize(const struct sock
*sk
, unsigned int mss_now
,
1735 bytes
= min(sk
->sk_pacing_rate
>> sk
->sk_pacing_shift
,
1736 sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
);
1738 /* Goal is to send at least one packet per ms,
1739 * not one big TSO packet every 100 ms.
1740 * This preserves ACK clocking and is consistent
1741 * with tcp_tso_should_defer() heuristic.
1743 segs
= max_t(u32
, bytes
/ mss_now
, min_tso_segs
);
1747 EXPORT_SYMBOL(tcp_tso_autosize
);
1749 /* Return the number of segments we want in the skb we are transmitting.
1750 * See if congestion control module wants to decide; otherwise, autosize.
1752 static u32
tcp_tso_segs(struct sock
*sk
, unsigned int mss_now
)
1754 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1755 u32 tso_segs
= ca_ops
->tso_segs_goal
? ca_ops
->tso_segs_goal(sk
) : 0;
1758 tso_segs
= tcp_tso_autosize(sk
, mss_now
,
1759 sock_net(sk
)->ipv4
.sysctl_tcp_min_tso_segs
);
1760 return min_t(u32
, tso_segs
, sk
->sk_gso_max_segs
);
1763 /* Returns the portion of skb which can be sent right away */
1764 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1765 const struct sk_buff
*skb
,
1766 unsigned int mss_now
,
1767 unsigned int max_segs
,
1770 const struct tcp_sock
*tp
= tcp_sk(sk
);
1771 u32 partial
, needed
, window
, max_len
;
1773 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1774 max_len
= mss_now
* max_segs
;
1776 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1779 needed
= min(skb
->len
, window
);
1781 if (max_len
<= needed
)
1784 partial
= needed
% mss_now
;
1785 /* If last segment is not a full MSS, check if Nagle rules allow us
1786 * to include this last segment in this skb.
1787 * Otherwise, we'll split the skb at last MSS boundary
1789 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
1790 return needed
- partial
;
1795 /* Can at least one segment of SKB be sent right now, according to the
1796 * congestion window rules? If so, return how many segments are allowed.
1798 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1799 const struct sk_buff
*skb
)
1801 u32 in_flight
, cwnd
, halfcwnd
;
1803 /* Don't be strict about the congestion window for the final FIN. */
1804 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1805 tcp_skb_pcount(skb
) == 1)
1808 in_flight
= tcp_packets_in_flight(tp
);
1809 cwnd
= tp
->snd_cwnd
;
1810 if (in_flight
>= cwnd
)
1813 /* For better scheduling, ensure we have at least
1814 * 2 GSO packets in flight.
1816 halfcwnd
= max(cwnd
>> 1, 1U);
1817 return min(halfcwnd
, cwnd
- in_flight
);
1820 /* Initialize TSO state of a skb.
1821 * This must be invoked the first time we consider transmitting
1822 * SKB onto the wire.
1824 static int tcp_init_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1826 int tso_segs
= tcp_skb_pcount(skb
);
1828 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1829 tcp_set_skb_tso_segs(skb
, mss_now
);
1830 tso_segs
= tcp_skb_pcount(skb
);
1836 /* Return true if the Nagle test allows this packet to be
1839 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1840 unsigned int cur_mss
, int nonagle
)
1842 /* Nagle rule does not apply to frames, which sit in the middle of the
1843 * write_queue (they have no chances to get new data).
1845 * This is implemented in the callers, where they modify the 'nonagle'
1846 * argument based upon the location of SKB in the send queue.
1848 if (nonagle
& TCP_NAGLE_PUSH
)
1851 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1852 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1855 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
1861 /* Does at least the first segment of SKB fit into the send window? */
1862 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1863 const struct sk_buff
*skb
,
1864 unsigned int cur_mss
)
1866 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1868 if (skb
->len
> cur_mss
)
1869 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1871 return !after(end_seq
, tcp_wnd_end(tp
));
1874 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1875 * which is put after SKB on the list. It is very much like
1876 * tcp_fragment() except that it may make several kinds of assumptions
1877 * in order to speed up the splitting operation. In particular, we
1878 * know that all the data is in scatter-gather pages, and that the
1879 * packet has never been sent out before (and thus is not cloned).
1881 static int tso_fragment(struct sock
*sk
, enum tcp_queue tcp_queue
,
1882 struct sk_buff
*skb
, unsigned int len
,
1883 unsigned int mss_now
, gfp_t gfp
)
1885 struct sk_buff
*buff
;
1886 int nlen
= skb
->len
- len
;
1889 /* All of a TSO frame must be composed of paged data. */
1890 if (skb
->len
!= skb
->data_len
)
1891 return tcp_fragment(sk
, tcp_queue
, skb
, len
, mss_now
, gfp
);
1893 buff
= sk_stream_alloc_skb(sk
, 0, gfp
, true);
1894 if (unlikely(!buff
))
1897 sk
->sk_wmem_queued
+= buff
->truesize
;
1898 sk_mem_charge(sk
, buff
->truesize
);
1899 buff
->truesize
+= nlen
;
1900 skb
->truesize
-= nlen
;
1902 /* Correct the sequence numbers. */
1903 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1904 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1905 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1907 /* PSH and FIN should only be set in the second packet. */
1908 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1909 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1910 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1912 /* This packet was never sent out yet, so no SACK bits. */
1913 TCP_SKB_CB(buff
)->sacked
= 0;
1915 tcp_skb_fragment_eor(skb
, buff
);
1917 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1918 skb_split(skb
, buff
, len
);
1919 tcp_fragment_tstamp(skb
, buff
);
1921 /* Fix up tso_factor for both original and new SKB. */
1922 tcp_set_skb_tso_segs(skb
, mss_now
);
1923 tcp_set_skb_tso_segs(buff
, mss_now
);
1925 /* Link BUFF into the send queue. */
1926 __skb_header_release(buff
);
1927 tcp_insert_write_queue_after(skb
, buff
, sk
, tcp_queue
);
1932 /* Try to defer sending, if possible, in order to minimize the amount
1933 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1935 * This algorithm is from John Heffner.
1937 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
,
1938 bool *is_cwnd_limited
,
1939 bool *is_rwnd_limited
,
1942 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1943 u32 age
, send_win
, cong_win
, limit
, in_flight
;
1944 struct tcp_sock
*tp
= tcp_sk(sk
);
1945 struct sk_buff
*head
;
1948 if (icsk
->icsk_ca_state
>= TCP_CA_Recovery
)
1951 /* Avoid bursty behavior by allowing defer
1952 * only if the last write was recent.
1954 if ((s32
)(tcp_jiffies32
- tp
->lsndtime
) > 0)
1957 in_flight
= tcp_packets_in_flight(tp
);
1959 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1961 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1963 /* From in_flight test above, we know that cwnd > in_flight. */
1964 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1966 limit
= min(send_win
, cong_win
);
1968 /* If a full-sized TSO skb can be sent, do it. */
1969 if (limit
>= max_segs
* tp
->mss_cache
)
1972 /* Middle in queue won't get any more data, full sendable already? */
1973 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1976 win_divisor
= READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_tso_win_divisor
);
1978 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1980 /* If at least some fraction of a window is available,
1983 chunk
/= win_divisor
;
1987 /* Different approach, try not to defer past a single
1988 * ACK. Receiver should ACK every other full sized
1989 * frame, so if we have space for more than 3 frames
1992 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1996 /* TODO : use tsorted_sent_queue ? */
1997 head
= tcp_rtx_queue_head(sk
);
2000 age
= tcp_stamp_us_delta(tp
->tcp_mstamp
, head
->skb_mstamp
);
2001 /* If next ACK is likely to come too late (half srtt), do not defer */
2002 if (age
< (tp
->srtt_us
>> 4))
2005 /* Ok, it looks like it is advisable to defer.
2006 * Three cases are tracked :
2007 * 1) We are cwnd-limited
2008 * 2) We are rwnd-limited
2009 * 3) We are application limited.
2011 if (cong_win
< send_win
) {
2012 if (cong_win
<= skb
->len
) {
2013 *is_cwnd_limited
= true;
2017 if (send_win
<= skb
->len
) {
2018 *is_rwnd_limited
= true;
2023 /* If this packet won't get more data, do not wait. */
2024 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
2033 static inline void tcp_mtu_check_reprobe(struct sock
*sk
)
2035 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2036 struct tcp_sock
*tp
= tcp_sk(sk
);
2037 struct net
*net
= sock_net(sk
);
2041 interval
= net
->ipv4
.sysctl_tcp_probe_interval
;
2042 delta
= tcp_jiffies32
- icsk
->icsk_mtup
.probe_timestamp
;
2043 if (unlikely(delta
>= interval
* HZ
)) {
2044 int mss
= tcp_current_mss(sk
);
2046 /* Update current search range */
2047 icsk
->icsk_mtup
.probe_size
= 0;
2048 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+
2049 sizeof(struct tcphdr
) +
2050 icsk
->icsk_af_ops
->net_header_len
;
2051 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
2053 /* Update probe time stamp */
2054 icsk
->icsk_mtup
.probe_timestamp
= tcp_jiffies32
;
2058 static bool tcp_can_coalesce_send_queue_head(struct sock
*sk
, int len
)
2060 struct sk_buff
*skb
, *next
;
2062 skb
= tcp_send_head(sk
);
2063 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
2064 if (len
<= skb
->len
)
2067 if (unlikely(TCP_SKB_CB(skb
)->eor
) || tcp_has_tx_tstamp(skb
))
2076 /* Create a new MTU probe if we are ready.
2077 * MTU probe is regularly attempting to increase the path MTU by
2078 * deliberately sending larger packets. This discovers routing
2079 * changes resulting in larger path MTUs.
2081 * Returns 0 if we should wait to probe (no cwnd available),
2082 * 1 if a probe was sent,
2085 static int tcp_mtu_probe(struct sock
*sk
)
2087 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2088 struct tcp_sock
*tp
= tcp_sk(sk
);
2089 struct sk_buff
*skb
, *nskb
, *next
;
2090 struct net
*net
= sock_net(sk
);
2097 /* Not currently probing/verifying,
2099 * have enough cwnd, and
2100 * not SACKing (the variable headers throw things off)
2102 if (likely(!icsk
->icsk_mtup
.enabled
||
2103 icsk
->icsk_mtup
.probe_size
||
2104 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
2105 tp
->snd_cwnd
< 11 ||
2106 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
))
2109 /* Use binary search for probe_size between tcp_mss_base,
2110 * and current mss_clamp. if (search_high - search_low)
2111 * smaller than a threshold, backoff from probing.
2113 mss_now
= tcp_current_mss(sk
);
2114 probe_size
= tcp_mtu_to_mss(sk
, (icsk
->icsk_mtup
.search_high
+
2115 icsk
->icsk_mtup
.search_low
) >> 1);
2116 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
2117 interval
= icsk
->icsk_mtup
.search_high
- icsk
->icsk_mtup
.search_low
;
2118 /* When misfortune happens, we are reprobing actively,
2119 * and then reprobe timer has expired. We stick with current
2120 * probing process by not resetting search range to its orignal.
2122 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
) ||
2123 interval
< net
->ipv4
.sysctl_tcp_probe_threshold
) {
2124 /* Check whether enough time has elaplased for
2125 * another round of probing.
2127 tcp_mtu_check_reprobe(sk
);
2131 /* Have enough data in the send queue to probe? */
2132 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
2135 if (tp
->snd_wnd
< size_needed
)
2137 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
2140 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2141 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
2142 if (!tcp_packets_in_flight(tp
))
2148 if (!tcp_can_coalesce_send_queue_head(sk
, probe_size
))
2151 /* We're allowed to probe. Build it now. */
2152 nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
, false);
2155 sk
->sk_wmem_queued
+= nskb
->truesize
;
2156 sk_mem_charge(sk
, nskb
->truesize
);
2158 skb
= tcp_send_head(sk
);
2160 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
2161 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
2162 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
2163 TCP_SKB_CB(nskb
)->sacked
= 0;
2165 nskb
->ip_summed
= skb
->ip_summed
;
2167 tcp_insert_write_queue_before(nskb
, skb
, sk
);
2168 tcp_highest_sack_replace(sk
, skb
, nskb
);
2171 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
2172 copy
= min_t(int, skb
->len
, probe_size
- len
);
2173 if (nskb
->ip_summed
) {
2174 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
2176 __wsum csum
= skb_copy_and_csum_bits(skb
, 0,
2177 skb_put(nskb
, copy
),
2179 nskb
->csum
= csum_block_add(nskb
->csum
, csum
, len
);
2182 if (skb
->len
<= copy
) {
2183 /* We've eaten all the data from this skb.
2185 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
2186 /* If this is the last SKB we copy and eor is set
2187 * we need to propagate it to the new skb.
2189 TCP_SKB_CB(nskb
)->eor
= TCP_SKB_CB(skb
)->eor
;
2190 tcp_skb_collapse_tstamp(nskb
, skb
);
2191 tcp_unlink_write_queue(skb
, sk
);
2192 sk_wmem_free_skb(sk
, skb
);
2194 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
2195 ~(TCPHDR_FIN
|TCPHDR_PSH
);
2196 if (!skb_shinfo(skb
)->nr_frags
) {
2197 skb_pull(skb
, copy
);
2198 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2199 skb
->csum
= csum_partial(skb
->data
,
2202 __pskb_trim_head(skb
, copy
);
2203 tcp_set_skb_tso_segs(skb
, mss_now
);
2205 TCP_SKB_CB(skb
)->seq
+= copy
;
2210 if (len
>= probe_size
)
2213 tcp_init_tso_segs(nskb
, nskb
->len
);
2215 /* We're ready to send. If this fails, the probe will
2216 * be resegmented into mss-sized pieces by tcp_write_xmit().
2218 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
2219 /* Decrement cwnd here because we are sending
2220 * effectively two packets. */
2222 tcp_event_new_data_sent(sk
, nskb
);
2224 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
2225 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
2226 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
2234 static bool tcp_pacing_check(const struct sock
*sk
)
2236 return tcp_needs_internal_pacing(sk
) &&
2237 hrtimer_active(&tcp_sk(sk
)->pacing_timer
);
2240 /* TCP Small Queues :
2241 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2242 * (These limits are doubled for retransmits)
2244 * - better RTT estimation and ACK scheduling
2247 * Alas, some drivers / subsystems require a fair amount
2248 * of queued bytes to ensure line rate.
2249 * One example is wifi aggregation (802.11 AMPDU)
2251 static bool tcp_small_queue_check(struct sock
*sk
, const struct sk_buff
*skb
,
2252 unsigned int factor
)
2256 limit
= max(2 * skb
->truesize
, sk
->sk_pacing_rate
>> sk
->sk_pacing_shift
);
2257 limit
= min_t(u32
, limit
,
2258 sock_net(sk
)->ipv4
.sysctl_tcp_limit_output_bytes
);
2261 if (refcount_read(&sk
->sk_wmem_alloc
) > limit
) {
2262 /* Always send skb if rtx queue is empty.
2263 * No need to wait for TX completion to call us back,
2264 * after softirq/tasklet schedule.
2265 * This helps when TX completions are delayed too much.
2267 if (tcp_rtx_queue_empty(sk
))
2270 set_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
);
2271 /* It is possible TX completion already happened
2272 * before we set TSQ_THROTTLED, so we must
2273 * test again the condition.
2275 smp_mb__after_atomic();
2276 if (refcount_read(&sk
->sk_wmem_alloc
) > limit
)
2282 static void tcp_chrono_set(struct tcp_sock
*tp
, const enum tcp_chrono
new)
2284 const u32 now
= tcp_jiffies32
;
2285 enum tcp_chrono old
= tp
->chrono_type
;
2287 if (old
> TCP_CHRONO_UNSPEC
)
2288 tp
->chrono_stat
[old
- 1] += now
- tp
->chrono_start
;
2289 tp
->chrono_start
= now
;
2290 tp
->chrono_type
= new;
2293 void tcp_chrono_start(struct sock
*sk
, const enum tcp_chrono type
)
2295 struct tcp_sock
*tp
= tcp_sk(sk
);
2297 /* If there are multiple conditions worthy of tracking in a
2298 * chronograph then the highest priority enum takes precedence
2299 * over the other conditions. So that if something "more interesting"
2300 * starts happening, stop the previous chrono and start a new one.
2302 if (type
> tp
->chrono_type
)
2303 tcp_chrono_set(tp
, type
);
2306 void tcp_chrono_stop(struct sock
*sk
, const enum tcp_chrono type
)
2308 struct tcp_sock
*tp
= tcp_sk(sk
);
2311 /* There are multiple conditions worthy of tracking in a
2312 * chronograph, so that the highest priority enum takes
2313 * precedence over the other conditions (see tcp_chrono_start).
2314 * If a condition stops, we only stop chrono tracking if
2315 * it's the "most interesting" or current chrono we are
2316 * tracking and starts busy chrono if we have pending data.
2318 if (tcp_rtx_and_write_queues_empty(sk
))
2319 tcp_chrono_set(tp
, TCP_CHRONO_UNSPEC
);
2320 else if (type
== tp
->chrono_type
)
2321 tcp_chrono_set(tp
, TCP_CHRONO_BUSY
);
2324 /* This routine writes packets to the network. It advances the
2325 * send_head. This happens as incoming acks open up the remote
2328 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2329 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2330 * account rare use of URG, this is not a big flaw.
2332 * Send at most one packet when push_one > 0. Temporarily ignore
2333 * cwnd limit to force at most one packet out when push_one == 2.
2335 * Returns true, if no segments are in flight and we have queued segments,
2336 * but cannot send anything now because of SWS or another problem.
2338 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
2339 int push_one
, gfp_t gfp
)
2341 struct tcp_sock
*tp
= tcp_sk(sk
);
2342 struct sk_buff
*skb
;
2343 unsigned int tso_segs
, sent_pkts
;
2346 bool is_cwnd_limited
= false, is_rwnd_limited
= false;
2351 tcp_mstamp_refresh(tp
);
2353 /* Do MTU probing. */
2354 result
= tcp_mtu_probe(sk
);
2357 } else if (result
> 0) {
2362 max_segs
= tcp_tso_segs(sk
, mss_now
);
2363 while ((skb
= tcp_send_head(sk
))) {
2366 if (tcp_pacing_check(sk
))
2369 tso_segs
= tcp_init_tso_segs(skb
, mss_now
);
2372 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
) {
2373 /* "skb_mstamp" is used as a start point for the retransmit timer */
2374 tcp_update_skb_after_send(tp
, skb
);
2375 goto repair
; /* Skip network transmission */
2378 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2381 /* Force out a loss probe pkt. */
2387 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
))) {
2388 is_rwnd_limited
= true;
2392 if (tso_segs
== 1) {
2393 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2394 (tcp_skb_is_last(sk
, skb
) ?
2395 nonagle
: TCP_NAGLE_PUSH
))))
2399 tcp_tso_should_defer(sk
, skb
, &is_cwnd_limited
,
2400 &is_rwnd_limited
, max_segs
))
2405 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
2406 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2412 if (skb
->len
> limit
&&
2413 unlikely(tso_fragment(sk
, TCP_FRAG_IN_WRITE_QUEUE
,
2414 skb
, limit
, mss_now
, gfp
)))
2417 if (test_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
))
2418 clear_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
);
2419 if (tcp_small_queue_check(sk
, skb
, 0))
2422 /* Argh, we hit an empty skb(), presumably a thread
2423 * is sleeping in sendmsg()/sk_stream_wait_memory().
2424 * We do not want to send a pure-ack packet and have
2425 * a strange looking rtx queue with empty packet(s).
2427 if (TCP_SKB_CB(skb
)->end_seq
== TCP_SKB_CB(skb
)->seq
)
2430 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2434 /* Advance the send_head. This one is sent out.
2435 * This call will increment packets_out.
2437 tcp_event_new_data_sent(sk
, skb
);
2439 tcp_minshall_update(tp
, mss_now
, skb
);
2440 sent_pkts
+= tcp_skb_pcount(skb
);
2446 if (is_rwnd_limited
)
2447 tcp_chrono_start(sk
, TCP_CHRONO_RWND_LIMITED
);
2449 tcp_chrono_stop(sk
, TCP_CHRONO_RWND_LIMITED
);
2451 if (likely(sent_pkts
)) {
2452 if (tcp_in_cwnd_reduction(sk
))
2453 tp
->prr_out
+= sent_pkts
;
2455 /* Send one loss probe per tail loss episode. */
2457 tcp_schedule_loss_probe(sk
, false);
2458 is_cwnd_limited
|= (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
);
2459 tcp_cwnd_validate(sk
, is_cwnd_limited
);
2462 return !tp
->packets_out
&& !tcp_write_queue_empty(sk
);
2465 bool tcp_schedule_loss_probe(struct sock
*sk
, bool advancing_rto
)
2467 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2468 struct tcp_sock
*tp
= tcp_sk(sk
);
2469 u32 timeout
, rto_delta_us
;
2472 /* Don't do any loss probe on a Fast Open connection before 3WHS
2475 if (tp
->fastopen_rsk
)
2478 early_retrans
= sock_net(sk
)->ipv4
.sysctl_tcp_early_retrans
;
2479 /* Schedule a loss probe in 2*RTT for SACK capable connections
2480 * not in loss recovery, that are either limited by cwnd or application.
2482 if ((early_retrans
!= 3 && early_retrans
!= 4) ||
2483 !tp
->packets_out
|| !tcp_is_sack(tp
) ||
2484 (icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2485 icsk
->icsk_ca_state
!= TCP_CA_CWR
))
2488 /* Probe timeout is 2*rtt. Add minimum RTO to account
2489 * for delayed ack when there's one outstanding packet. If no RTT
2490 * sample is available then probe after TCP_TIMEOUT_INIT.
2493 timeout
= usecs_to_jiffies(tp
->srtt_us
>> 2);
2494 if (tp
->packets_out
== 1)
2495 timeout
+= TCP_RTO_MIN
;
2497 timeout
+= TCP_TIMEOUT_MIN
;
2499 timeout
= TCP_TIMEOUT_INIT
;
2502 /* If the RTO formula yields an earlier time, then use that time. */
2503 rto_delta_us
= advancing_rto
?
2504 jiffies_to_usecs(inet_csk(sk
)->icsk_rto
) :
2505 tcp_rto_delta_us(sk
); /* How far in future is RTO? */
2506 if (rto_delta_us
> 0)
2507 timeout
= min_t(u32
, timeout
, usecs_to_jiffies(rto_delta_us
));
2509 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2514 /* Thanks to skb fast clones, we can detect if a prior transmit of
2515 * a packet is still in a qdisc or driver queue.
2516 * In this case, there is very little point doing a retransmit !
2518 static bool skb_still_in_host_queue(const struct sock
*sk
,
2519 const struct sk_buff
*skb
)
2521 if (unlikely(skb_fclone_busy(sk
, skb
))) {
2522 NET_INC_STATS(sock_net(sk
),
2523 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
2529 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2530 * retransmit the last segment.
2532 void tcp_send_loss_probe(struct sock
*sk
)
2534 struct tcp_sock
*tp
= tcp_sk(sk
);
2535 struct sk_buff
*skb
;
2537 int mss
= tcp_current_mss(sk
);
2539 skb
= tcp_send_head(sk
);
2540 if (skb
&& tcp_snd_wnd_test(tp
, skb
, mss
)) {
2541 pcount
= tp
->packets_out
;
2542 tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2543 if (tp
->packets_out
> pcount
)
2547 skb
= skb_rb_last(&sk
->tcp_rtx_queue
);
2549 if (unlikely(!skb
)) {
2550 WARN_ONCE(tp
->packets_out
,
2551 "invalid inflight: %u state %u cwnd %u mss %d\n",
2552 tp
->packets_out
, sk
->sk_state
, tp
->snd_cwnd
, mss
);
2553 inet_csk(sk
)->icsk_pending
= 0;
2557 /* At most one outstanding TLP retransmission. */
2558 if (tp
->tlp_high_seq
)
2561 if (skb_still_in_host_queue(sk
, skb
))
2564 pcount
= tcp_skb_pcount(skb
);
2565 if (WARN_ON(!pcount
))
2568 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2569 if (unlikely(tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
,
2570 (pcount
- 1) * mss
, mss
,
2573 skb
= skb_rb_next(skb
);
2576 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2579 if (__tcp_retransmit_skb(sk
, skb
, 1))
2582 /* Record snd_nxt for loss detection. */
2583 tp
->tlp_high_seq
= tp
->snd_nxt
;
2586 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSPROBES
);
2587 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2588 inet_csk(sk
)->icsk_pending
= 0;
2593 /* Push out any pending frames which were held back due to
2594 * TCP_CORK or attempt at coalescing tiny packets.
2595 * The socket must be locked by the caller.
2597 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2600 /* If we are closed, the bytes will have to remain here.
2601 * In time closedown will finish, we empty the write queue and
2602 * all will be happy.
2604 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2607 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2608 sk_gfp_mask(sk
, GFP_ATOMIC
)))
2609 tcp_check_probe_timer(sk
);
2612 /* Send _single_ skb sitting at the send head. This function requires
2613 * true push pending frames to setup probe timer etc.
2615 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2617 struct sk_buff
*skb
= tcp_send_head(sk
);
2619 BUG_ON(!skb
|| skb
->len
< mss_now
);
2621 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2624 /* This function returns the amount that we can raise the
2625 * usable window based on the following constraints
2627 * 1. The window can never be shrunk once it is offered (RFC 793)
2628 * 2. We limit memory per socket
2631 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2632 * RECV.NEXT + RCV.WIN fixed until:
2633 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2635 * i.e. don't raise the right edge of the window until you can raise
2636 * it at least MSS bytes.
2638 * Unfortunately, the recommended algorithm breaks header prediction,
2639 * since header prediction assumes th->window stays fixed.
2641 * Strictly speaking, keeping th->window fixed violates the receiver
2642 * side SWS prevention criteria. The problem is that under this rule
2643 * a stream of single byte packets will cause the right side of the
2644 * window to always advance by a single byte.
2646 * Of course, if the sender implements sender side SWS prevention
2647 * then this will not be a problem.
2649 * BSD seems to make the following compromise:
2651 * If the free space is less than the 1/4 of the maximum
2652 * space available and the free space is less than 1/2 mss,
2653 * then set the window to 0.
2654 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2655 * Otherwise, just prevent the window from shrinking
2656 * and from being larger than the largest representable value.
2658 * This prevents incremental opening of the window in the regime
2659 * where TCP is limited by the speed of the reader side taking
2660 * data out of the TCP receive queue. It does nothing about
2661 * those cases where the window is constrained on the sender side
2662 * because the pipeline is full.
2664 * BSD also seems to "accidentally" limit itself to windows that are a
2665 * multiple of MSS, at least until the free space gets quite small.
2666 * This would appear to be a side effect of the mbuf implementation.
2667 * Combining these two algorithms results in the observed behavior
2668 * of having a fixed window size at almost all times.
2670 * Below we obtain similar behavior by forcing the offered window to
2671 * a multiple of the mss when it is feasible to do so.
2673 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2674 * Regular options like TIMESTAMP are taken into account.
2676 u32
__tcp_select_window(struct sock
*sk
)
2678 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2679 struct tcp_sock
*tp
= tcp_sk(sk
);
2680 /* MSS for the peer's data. Previous versions used mss_clamp
2681 * here. I don't know if the value based on our guesses
2682 * of peer's MSS is better for the performance. It's more correct
2683 * but may be worse for the performance because of rcv_mss
2684 * fluctuations. --SAW 1998/11/1
2686 int mss
= icsk
->icsk_ack
.rcv_mss
;
2687 int free_space
= tcp_space(sk
);
2688 int allowed_space
= tcp_full_space(sk
);
2689 int full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2692 if (unlikely(mss
> full_space
)) {
2697 if (free_space
< (full_space
>> 1)) {
2698 icsk
->icsk_ack
.quick
= 0;
2700 if (tcp_under_memory_pressure(sk
))
2701 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2704 /* free_space might become our new window, make sure we don't
2705 * increase it due to wscale.
2707 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2709 /* if free space is less than mss estimate, or is below 1/16th
2710 * of the maximum allowed, try to move to zero-window, else
2711 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2712 * new incoming data is dropped due to memory limits.
2713 * With large window, mss test triggers way too late in order
2714 * to announce zero window in time before rmem limit kicks in.
2716 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2720 if (free_space
> tp
->rcv_ssthresh
)
2721 free_space
= tp
->rcv_ssthresh
;
2723 /* Don't do rounding if we are using window scaling, since the
2724 * scaled window will not line up with the MSS boundary anyway.
2726 if (tp
->rx_opt
.rcv_wscale
) {
2727 window
= free_space
;
2729 /* Advertise enough space so that it won't get scaled away.
2730 * Import case: prevent zero window announcement if
2731 * 1<<rcv_wscale > mss.
2733 window
= ALIGN(window
, (1 << tp
->rx_opt
.rcv_wscale
));
2735 window
= tp
->rcv_wnd
;
2736 /* Get the largest window that is a nice multiple of mss.
2737 * Window clamp already applied above.
2738 * If our current window offering is within 1 mss of the
2739 * free space we just keep it. This prevents the divide
2740 * and multiply from happening most of the time.
2741 * We also don't do any window rounding when the free space
2744 if (window
<= free_space
- mss
|| window
> free_space
)
2745 window
= rounddown(free_space
, mss
);
2746 else if (mss
== full_space
&&
2747 free_space
> window
+ (full_space
>> 1))
2748 window
= free_space
;
2754 void tcp_skb_collapse_tstamp(struct sk_buff
*skb
,
2755 const struct sk_buff
*next_skb
)
2757 if (unlikely(tcp_has_tx_tstamp(next_skb
))) {
2758 const struct skb_shared_info
*next_shinfo
=
2759 skb_shinfo(next_skb
);
2760 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
2762 shinfo
->tx_flags
|= next_shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
2763 shinfo
->tskey
= next_shinfo
->tskey
;
2764 TCP_SKB_CB(skb
)->txstamp_ack
|=
2765 TCP_SKB_CB(next_skb
)->txstamp_ack
;
2769 /* Collapses two adjacent SKB's during retransmission. */
2770 static bool tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2772 struct tcp_sock
*tp
= tcp_sk(sk
);
2773 struct sk_buff
*next_skb
= skb_rb_next(skb
);
2774 int skb_size
, next_skb_size
;
2776 skb_size
= skb
->len
;
2777 next_skb_size
= next_skb
->len
;
2779 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2781 if (next_skb_size
) {
2782 if (next_skb_size
<= skb_availroom(skb
))
2783 skb_copy_bits(next_skb
, 0, skb_put(skb
, next_skb_size
),
2785 else if (!tcp_skb_shift(skb
, next_skb
, 1, next_skb_size
))
2788 tcp_highest_sack_replace(sk
, next_skb
, skb
);
2790 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2791 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2793 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2794 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2796 /* Update sequence range on original skb. */
2797 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2799 /* Merge over control information. This moves PSH/FIN etc. over */
2800 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2802 /* All done, get rid of second SKB and account for it so
2803 * packet counting does not break.
2805 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2806 TCP_SKB_CB(skb
)->eor
= TCP_SKB_CB(next_skb
)->eor
;
2808 /* changed transmit queue under us so clear hints */
2809 tcp_clear_retrans_hints_partial(tp
);
2810 if (next_skb
== tp
->retransmit_skb_hint
)
2811 tp
->retransmit_skb_hint
= skb
;
2813 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2815 tcp_skb_collapse_tstamp(skb
, next_skb
);
2817 tcp_rtx_queue_unlink_and_free(next_skb
, sk
);
2821 /* Check if coalescing SKBs is legal. */
2822 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2824 if (tcp_skb_pcount(skb
) > 1)
2826 if (skb_cloned(skb
))
2828 /* Some heuristics for collapsing over SACK'd could be invented */
2829 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2835 /* Collapse packets in the retransmit queue to make to create
2836 * less packets on the wire. This is only done on retransmission.
2838 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2841 struct tcp_sock
*tp
= tcp_sk(sk
);
2842 struct sk_buff
*skb
= to
, *tmp
;
2845 if (!sock_net(sk
)->ipv4
.sysctl_tcp_retrans_collapse
)
2847 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2850 skb_rbtree_walk_from_safe(skb
, tmp
) {
2851 if (!tcp_can_collapse(sk
, skb
))
2854 if (!tcp_skb_can_collapse_to(to
))
2867 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2870 if (!tcp_collapse_retrans(sk
, to
))
2875 /* This retransmits one SKB. Policy decisions and retransmit queue
2876 * state updates are done by the caller. Returns non-zero if an
2877 * error occurred which prevented the send.
2879 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2881 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2882 struct tcp_sock
*tp
= tcp_sk(sk
);
2883 unsigned int cur_mss
;
2887 /* Inconclusive MTU probe */
2888 if (icsk
->icsk_mtup
.probe_size
)
2889 icsk
->icsk_mtup
.probe_size
= 0;
2891 /* Do not sent more than we queued. 1/4 is reserved for possible
2892 * copying overhead: fragmentation, tunneling, mangling etc.
2894 if (refcount_read(&sk
->sk_wmem_alloc
) >
2895 min_t(u32
, sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2),
2899 if (skb_still_in_host_queue(sk
, skb
))
2902 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2903 if (unlikely(before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))) {
2907 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2911 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2912 return -EHOSTUNREACH
; /* Routing failure or similar. */
2914 cur_mss
= tcp_current_mss(sk
);
2916 /* If receiver has shrunk his window, and skb is out of
2917 * new window, do not retransmit it. The exception is the
2918 * case, when window is shrunk to zero. In this case
2919 * our retransmit serves as a zero window probe.
2921 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2922 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2925 len
= cur_mss
* segs
;
2926 if (skb
->len
> len
) {
2927 if (tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
, len
,
2928 cur_mss
, GFP_ATOMIC
))
2929 return -ENOMEM
; /* We'll try again later. */
2931 if (skb_unclone(skb
, GFP_ATOMIC
))
2934 diff
= tcp_skb_pcount(skb
);
2935 tcp_set_skb_tso_segs(skb
, cur_mss
);
2936 diff
-= tcp_skb_pcount(skb
);
2938 tcp_adjust_pcount(sk
, skb
, diff
);
2939 if (skb
->len
< cur_mss
)
2940 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2943 /* RFC3168, section 6.1.1.1. ECN fallback */
2944 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN_ECN
) == TCPHDR_SYN_ECN
)
2945 tcp_ecn_clear_syn(sk
, skb
);
2947 /* Update global and local TCP statistics. */
2948 segs
= tcp_skb_pcount(skb
);
2949 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
, segs
);
2950 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2951 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
2952 tp
->total_retrans
+= segs
;
2954 /* make sure skb->data is aligned on arches that require it
2955 * and check if ack-trimming & collapsing extended the headroom
2956 * beyond what csum_start can cover.
2958 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2959 skb_headroom(skb
) >= 0xFFFF)) {
2960 struct sk_buff
*nskb
;
2962 tcp_skb_tsorted_save(skb
) {
2963 nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
, GFP_ATOMIC
);
2964 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2966 } tcp_skb_tsorted_restore(skb
);
2969 tcp_update_skb_after_send(tp
, skb
);
2970 tcp_rate_skb_sent(sk
, skb
);
2973 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2977 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
2978 trace_tcp_retransmit_skb(sk
, skb
);
2979 } else if (err
!= -EBUSY
) {
2980 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
, segs
);
2985 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2987 struct tcp_sock
*tp
= tcp_sk(sk
);
2988 int err
= __tcp_retransmit_skb(sk
, skb
, segs
);
2991 #if FASTRETRANS_DEBUG > 0
2992 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2993 net_dbg_ratelimited("retrans_out leaked\n");
2996 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2997 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2999 /* Save stamp of the first retransmit. */
3000 if (!tp
->retrans_stamp
)
3001 tp
->retrans_stamp
= tcp_skb_timestamp(skb
);
3005 if (tp
->undo_retrans
< 0)
3006 tp
->undo_retrans
= 0;
3007 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
3011 /* This gets called after a retransmit timeout, and the initially
3012 * retransmitted data is acknowledged. It tries to continue
3013 * resending the rest of the retransmit queue, until either
3014 * we've sent it all or the congestion window limit is reached.
3016 void tcp_xmit_retransmit_queue(struct sock
*sk
)
3018 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3019 struct sk_buff
*skb
, *rtx_head
, *hole
= NULL
;
3020 struct tcp_sock
*tp
= tcp_sk(sk
);
3024 if (!tp
->packets_out
)
3027 rtx_head
= tcp_rtx_queue_head(sk
);
3028 skb
= tp
->retransmit_skb_hint
?: rtx_head
;
3029 max_segs
= tcp_tso_segs(sk
, tcp_current_mss(sk
));
3030 skb_rbtree_walk_from(skb
) {
3034 if (tcp_pacing_check(sk
))
3037 /* we could do better than to assign each time */
3039 tp
->retransmit_skb_hint
= skb
;
3041 segs
= tp
->snd_cwnd
- tcp_packets_in_flight(tp
);
3044 sacked
= TCP_SKB_CB(skb
)->sacked
;
3045 /* In case tcp_shift_skb_data() have aggregated large skbs,
3046 * we need to make sure not sending too bigs TSO packets
3048 segs
= min_t(int, segs
, max_segs
);
3050 if (tp
->retrans_out
>= tp
->lost_out
) {
3052 } else if (!(sacked
& TCPCB_LOST
)) {
3053 if (!hole
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
3058 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
3059 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
3061 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
3064 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
3067 if (tcp_small_queue_check(sk
, skb
, 1))
3070 if (tcp_retransmit_skb(sk
, skb
, segs
))
3073 NET_ADD_STATS(sock_net(sk
), mib_idx
, tcp_skb_pcount(skb
));
3075 if (tcp_in_cwnd_reduction(sk
))
3076 tp
->prr_out
+= tcp_skb_pcount(skb
);
3078 if (skb
== rtx_head
&&
3079 icsk
->icsk_pending
!= ICSK_TIME_REO_TIMEOUT
)
3080 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3081 inet_csk(sk
)->icsk_rto
,
3086 /* We allow to exceed memory limits for FIN packets to expedite
3087 * connection tear down and (memory) recovery.
3088 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3089 * or even be forced to close flow without any FIN.
3090 * In general, we want to allow one skb per socket to avoid hangs
3091 * with edge trigger epoll()
3093 void sk_forced_mem_schedule(struct sock
*sk
, int size
)
3097 if (size
<= sk
->sk_forward_alloc
)
3099 amt
= sk_mem_pages(size
);
3100 sk
->sk_forward_alloc
+= amt
* SK_MEM_QUANTUM
;
3101 sk_memory_allocated_add(sk
, amt
);
3103 if (mem_cgroup_sockets_enabled
&& sk
->sk_memcg
)
3104 mem_cgroup_charge_skmem(sk
->sk_memcg
, amt
);
3107 /* Send a FIN. The caller locks the socket for us.
3108 * We should try to send a FIN packet really hard, but eventually give up.
3110 void tcp_send_fin(struct sock
*sk
)
3112 struct sk_buff
*skb
, *tskb
= tcp_write_queue_tail(sk
);
3113 struct tcp_sock
*tp
= tcp_sk(sk
);
3115 /* Optimization, tack on the FIN if we have one skb in write queue and
3116 * this skb was not yet sent, or we are under memory pressure.
3117 * Note: in the latter case, FIN packet will be sent after a timeout,
3118 * as TCP stack thinks it has already been transmitted.
3120 if (!tskb
&& tcp_under_memory_pressure(sk
))
3121 tskb
= skb_rb_last(&sk
->tcp_rtx_queue
);
3125 TCP_SKB_CB(tskb
)->tcp_flags
|= TCPHDR_FIN
;
3126 TCP_SKB_CB(tskb
)->end_seq
++;
3128 if (tcp_write_queue_empty(sk
)) {
3129 /* This means tskb was already sent.
3130 * Pretend we included the FIN on previous transmit.
3131 * We need to set tp->snd_nxt to the value it would have
3132 * if FIN had been sent. This is because retransmit path
3133 * does not change tp->snd_nxt.
3139 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, sk
->sk_allocation
);
3140 if (unlikely(!skb
)) {
3145 INIT_LIST_HEAD(&skb
->tcp_tsorted_anchor
);
3146 skb_reserve(skb
, MAX_TCP_HEADER
);
3147 sk_forced_mem_schedule(sk
, skb
->truesize
);
3148 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3149 tcp_init_nondata_skb(skb
, tp
->write_seq
,
3150 TCPHDR_ACK
| TCPHDR_FIN
);
3151 tcp_queue_skb(sk
, skb
);
3153 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
), TCP_NAGLE_OFF
);
3156 /* We get here when a process closes a file descriptor (either due to
3157 * an explicit close() or as a byproduct of exit()'ing) and there
3158 * was unread data in the receive queue. This behavior is recommended
3159 * by RFC 2525, section 2.17. -DaveM
3161 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
3163 struct sk_buff
*skb
;
3165 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
3167 /* NOTE: No TCP options attached and we never retransmit this. */
3168 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
3170 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3174 /* Reserve space for headers and prepare control bits. */
3175 skb_reserve(skb
, MAX_TCP_HEADER
);
3176 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
3177 TCPHDR_ACK
| TCPHDR_RST
);
3178 tcp_mstamp_refresh(tcp_sk(sk
));
3180 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
3181 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3183 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3184 * skb here is different to the troublesome skb, so use NULL
3186 trace_tcp_send_reset(sk
, NULL
);
3189 /* Send a crossed SYN-ACK during socket establishment.
3190 * WARNING: This routine must only be called when we have already sent
3191 * a SYN packet that crossed the incoming SYN that caused this routine
3192 * to get called. If this assumption fails then the initial rcv_wnd
3193 * and rcv_wscale values will not be correct.
3195 int tcp_send_synack(struct sock
*sk
)
3197 struct sk_buff
*skb
;
3199 skb
= tcp_rtx_queue_head(sk
);
3200 if (!skb
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
3201 pr_err("%s: wrong queue state\n", __func__
);
3204 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
3205 if (skb_cloned(skb
)) {
3206 struct sk_buff
*nskb
;
3208 tcp_skb_tsorted_save(skb
) {
3209 nskb
= skb_copy(skb
, GFP_ATOMIC
);
3210 } tcp_skb_tsorted_restore(skb
);
3213 INIT_LIST_HEAD(&nskb
->tcp_tsorted_anchor
);
3214 tcp_highest_sack_replace(sk
, skb
, nskb
);
3215 tcp_rtx_queue_unlink_and_free(skb
, sk
);
3216 __skb_header_release(nskb
);
3217 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, nskb
);
3218 sk
->sk_wmem_queued
+= nskb
->truesize
;
3219 sk_mem_charge(sk
, nskb
->truesize
);
3223 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
3224 tcp_ecn_send_synack(sk
, skb
);
3226 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3230 * tcp_make_synack - Prepare a SYN-ACK.
3231 * sk: listener socket
3232 * dst: dst entry attached to the SYNACK
3233 * req: request_sock pointer
3235 * Allocate one skb and build a SYNACK packet.
3236 * @dst is consumed : Caller should not use it again.
3238 struct sk_buff
*tcp_make_synack(const struct sock
*sk
, struct dst_entry
*dst
,
3239 struct request_sock
*req
,
3240 struct tcp_fastopen_cookie
*foc
,
3241 enum tcp_synack_type synack_type
)
3243 struct inet_request_sock
*ireq
= inet_rsk(req
);
3244 const struct tcp_sock
*tp
= tcp_sk(sk
);
3245 struct tcp_md5sig_key
*md5
= NULL
;
3246 struct tcp_out_options opts
;
3247 struct sk_buff
*skb
;
3248 int tcp_header_size
;
3252 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
3253 if (unlikely(!skb
)) {
3257 /* Reserve space for headers. */
3258 skb_reserve(skb
, MAX_TCP_HEADER
);
3260 switch (synack_type
) {
3261 case TCP_SYNACK_NORMAL
:
3262 skb_set_owner_w(skb
, req_to_sk(req
));
3264 case TCP_SYNACK_COOKIE
:
3265 /* Under synflood, we do not attach skb to a socket,
3266 * to avoid false sharing.
3269 case TCP_SYNACK_FASTOPEN
:
3270 /* sk is a const pointer, because we want to express multiple
3271 * cpu might call us concurrently.
3272 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3274 skb_set_owner_w(skb
, (struct sock
*)sk
);
3277 skb_dst_set(skb
, dst
);
3279 mss
= tcp_mss_clamp(tp
, dst_metric_advmss(dst
));
3281 memset(&opts
, 0, sizeof(opts
));
3282 #ifdef CONFIG_SYN_COOKIES
3283 if (unlikely(req
->cookie_ts
))
3284 skb
->skb_mstamp
= cookie_init_timestamp(req
);
3287 skb
->skb_mstamp
= tcp_clock_us();
3289 #ifdef CONFIG_TCP_MD5SIG
3291 md5
= tcp_rsk(req
)->af_specific
->req_md5_lookup(sk
, req_to_sk(req
));
3293 skb_set_hash(skb
, tcp_rsk(req
)->txhash
, PKT_HASH_TYPE_L4
);
3294 tcp_header_size
= tcp_synack_options(sk
, req
, mss
, skb
, &opts
, md5
,
3297 skb_push(skb
, tcp_header_size
);
3298 skb_reset_transport_header(skb
);
3300 th
= (struct tcphdr
*)skb
->data
;
3301 memset(th
, 0, sizeof(struct tcphdr
));
3304 tcp_ecn_make_synack(req
, th
);
3305 th
->source
= htons(ireq
->ir_num
);
3306 th
->dest
= ireq
->ir_rmt_port
;
3307 skb
->mark
= ireq
->ir_mark
;
3308 skb
->ip_summed
= CHECKSUM_PARTIAL
;
3309 th
->seq
= htonl(tcp_rsk(req
)->snt_isn
);
3310 /* XXX data is queued and acked as is. No buffer/window check */
3311 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
3313 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3314 th
->window
= htons(min(req
->rsk_rcv_wnd
, 65535U));
3315 tcp_options_write((__be32
*)(th
+ 1), NULL
, &opts
);
3316 th
->doff
= (tcp_header_size
>> 2);
3317 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
3319 #ifdef CONFIG_TCP_MD5SIG
3320 /* Okay, we have all we need - do the md5 hash if needed */
3322 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
3323 md5
, req_to_sk(req
), skb
);
3327 /* Do not fool tcpdump (if any), clean our debris */
3331 EXPORT_SYMBOL(tcp_make_synack
);
3333 static void tcp_ca_dst_init(struct sock
*sk
, const struct dst_entry
*dst
)
3335 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3336 const struct tcp_congestion_ops
*ca
;
3337 u32 ca_key
= dst_metric(dst
, RTAX_CC_ALGO
);
3339 if (ca_key
== TCP_CA_UNSPEC
)
3343 ca
= tcp_ca_find_key(ca_key
);
3344 if (likely(ca
&& try_module_get(ca
->owner
))) {
3345 module_put(icsk
->icsk_ca_ops
->owner
);
3346 icsk
->icsk_ca_dst_locked
= tcp_ca_dst_locked(dst
);
3347 icsk
->icsk_ca_ops
= ca
;
3352 /* Do all connect socket setups that can be done AF independent. */
3353 static void tcp_connect_init(struct sock
*sk
)
3355 const struct dst_entry
*dst
= __sk_dst_get(sk
);
3356 struct tcp_sock
*tp
= tcp_sk(sk
);
3360 /* We'll fix this up when we get a response from the other end.
3361 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3363 tp
->tcp_header_len
= sizeof(struct tcphdr
);
3364 if (sock_net(sk
)->ipv4
.sysctl_tcp_timestamps
)
3365 tp
->tcp_header_len
+= TCPOLEN_TSTAMP_ALIGNED
;
3367 #ifdef CONFIG_TCP_MD5SIG
3368 if (tp
->af_specific
->md5_lookup(sk
, sk
))
3369 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
3372 /* If user gave his TCP_MAXSEG, record it to clamp */
3373 if (tp
->rx_opt
.user_mss
)
3374 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3377 tcp_sync_mss(sk
, dst_mtu(dst
));
3379 tcp_ca_dst_init(sk
, dst
);
3381 if (!tp
->window_clamp
)
3382 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
3383 tp
->advmss
= tcp_mss_clamp(tp
, dst_metric_advmss(dst
));
3385 tcp_initialize_rcv_mss(sk
);
3387 /* limit the window selection if the user enforce a smaller rx buffer */
3388 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
3389 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
3390 tp
->window_clamp
= tcp_full_space(sk
);
3392 rcv_wnd
= tcp_rwnd_init_bpf(sk
);
3394 rcv_wnd
= dst_metric(dst
, RTAX_INITRWND
);
3396 tcp_select_initial_window(sk
, tcp_full_space(sk
),
3397 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
3400 sock_net(sk
)->ipv4
.sysctl_tcp_window_scaling
,
3404 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
3405 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
3408 sock_reset_flag(sk
, SOCK_DONE
);
3411 tcp_write_queue_purge(sk
);
3412 tp
->snd_una
= tp
->write_seq
;
3413 tp
->snd_sml
= tp
->write_seq
;
3414 tp
->snd_up
= tp
->write_seq
;
3415 tp
->snd_nxt
= tp
->write_seq
;
3417 if (likely(!tp
->repair
))
3420 tp
->rcv_tstamp
= tcp_jiffies32
;
3421 tp
->rcv_wup
= tp
->rcv_nxt
;
3422 tp
->copied_seq
= tp
->rcv_nxt
;
3424 inet_csk(sk
)->icsk_rto
= tcp_timeout_init(sk
);
3425 inet_csk(sk
)->icsk_retransmits
= 0;
3426 tcp_clear_retrans(tp
);
3429 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
3431 struct tcp_sock
*tp
= tcp_sk(sk
);
3432 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
3434 tcb
->end_seq
+= skb
->len
;
3435 __skb_header_release(skb
);
3436 sk
->sk_wmem_queued
+= skb
->truesize
;
3437 sk_mem_charge(sk
, skb
->truesize
);
3438 tp
->write_seq
= tcb
->end_seq
;
3439 tp
->packets_out
+= tcp_skb_pcount(skb
);
3442 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3443 * queue a data-only packet after the regular SYN, such that regular SYNs
3444 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3445 * only the SYN sequence, the data are retransmitted in the first ACK.
3446 * If cookie is not cached or other error occurs, falls back to send a
3447 * regular SYN with Fast Open cookie request option.
3449 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
3451 struct tcp_sock
*tp
= tcp_sk(sk
);
3452 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
3454 struct sk_buff
*syn_data
;
3456 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
3457 if (!tcp_fastopen_cookie_check(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
))
3460 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3461 * user-MSS. Reserve maximum option space for middleboxes that add
3462 * private TCP options. The cost is reduced data space in SYN :(
3464 tp
->rx_opt
.mss_clamp
= tcp_mss_clamp(tp
, tp
->rx_opt
.mss_clamp
);
3466 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
3467 MAX_TCP_OPTION_SPACE
;
3469 space
= min_t(size_t, space
, fo
->size
);
3471 /* limit to order-0 allocations */
3472 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
3474 syn_data
= sk_stream_alloc_skb(sk
, space
, sk
->sk_allocation
, false);
3477 syn_data
->ip_summed
= CHECKSUM_PARTIAL
;
3478 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
3480 int copied
= copy_from_iter(skb_put(syn_data
, space
), space
,
3481 &fo
->data
->msg_iter
);
3482 if (unlikely(!copied
)) {
3483 tcp_skb_tsorted_anchor_cleanup(syn_data
);
3484 kfree_skb(syn_data
);
3487 if (copied
!= space
) {
3488 skb_trim(syn_data
, copied
);
3492 /* No more data pending in inet_wait_for_connect() */
3493 if (space
== fo
->size
)
3497 tcp_connect_queue_skb(sk
, syn_data
);
3499 tcp_chrono_start(sk
, TCP_CHRONO_BUSY
);
3501 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
3503 syn
->skb_mstamp
= syn_data
->skb_mstamp
;
3505 /* Now full SYN+DATA was cloned and sent (or not),
3506 * remove the SYN from the original skb (syn_data)
3507 * we keep in write queue in case of a retransmit, as we
3508 * also have the SYN packet (with no data) in the same queue.
3510 TCP_SKB_CB(syn_data
)->seq
++;
3511 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
3513 tp
->syn_data
= (fo
->copied
> 0);
3514 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, syn_data
);
3515 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3519 /* data was not sent, put it in write_queue */
3520 __skb_queue_tail(&sk
->sk_write_queue
, syn_data
);
3521 tp
->packets_out
-= tcp_skb_pcount(syn_data
);
3524 /* Send a regular SYN with Fast Open cookie request option */
3525 if (fo
->cookie
.len
> 0)
3527 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3529 tp
->syn_fastopen
= 0;
3531 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3535 /* Build a SYN and send it off. */
3536 int tcp_connect(struct sock
*sk
)
3538 struct tcp_sock
*tp
= tcp_sk(sk
);
3539 struct sk_buff
*buff
;
3542 tcp_call_bpf(sk
, BPF_SOCK_OPS_TCP_CONNECT_CB
);
3544 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
3545 return -EHOSTUNREACH
; /* Routing failure or similar. */
3547 tcp_connect_init(sk
);
3549 if (unlikely(tp
->repair
)) {
3550 tcp_finish_connect(sk
, NULL
);
3554 buff
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
, true);
3555 if (unlikely(!buff
))
3558 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3559 tcp_mstamp_refresh(tp
);
3560 tp
->retrans_stamp
= tcp_time_stamp(tp
);
3561 tcp_connect_queue_skb(sk
, buff
);
3562 tcp_ecn_send_syn(sk
, buff
);
3563 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, buff
);
3565 /* Send off SYN; include data in Fast Open. */
3566 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3567 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3568 if (err
== -ECONNREFUSED
)
3571 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3572 * in order to make this packet get counted in tcpOutSegs.
3574 tp
->snd_nxt
= tp
->write_seq
;
3575 tp
->pushed_seq
= tp
->write_seq
;
3576 buff
= tcp_send_head(sk
);
3577 if (unlikely(buff
)) {
3578 tp
->snd_nxt
= TCP_SKB_CB(buff
)->seq
;
3579 tp
->pushed_seq
= TCP_SKB_CB(buff
)->seq
;
3581 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3583 /* Timer for repeating the SYN until an answer. */
3584 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3585 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3588 EXPORT_SYMBOL(tcp_connect
);
3590 /* Send out a delayed ack, the caller does the policy checking
3591 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3594 void tcp_send_delayed_ack(struct sock
*sk
)
3596 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3597 int ato
= icsk
->icsk_ack
.ato
;
3598 unsigned long timeout
;
3600 if (ato
> TCP_DELACK_MIN
) {
3601 const struct tcp_sock
*tp
= tcp_sk(sk
);
3602 int max_ato
= HZ
/ 2;
3604 if (icsk
->icsk_ack
.pingpong
||
3605 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3606 max_ato
= TCP_DELACK_MAX
;
3608 /* Slow path, intersegment interval is "high". */
3610 /* If some rtt estimate is known, use it to bound delayed ack.
3611 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3615 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3622 ato
= min(ato
, max_ato
);
3625 /* Stay within the limit we were given */
3626 timeout
= jiffies
+ ato
;
3628 /* Use new timeout only if there wasn't a older one earlier. */
3629 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3630 /* If delack timer was blocked or is about to expire,
3633 if (icsk
->icsk_ack
.blocked
||
3634 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3639 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3640 timeout
= icsk
->icsk_ack
.timeout
;
3642 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3643 icsk
->icsk_ack
.timeout
= timeout
;
3644 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3647 /* This routine sends an ack and also updates the window. */
3648 void __tcp_send_ack(struct sock
*sk
, u32 rcv_nxt
)
3650 struct sk_buff
*buff
;
3652 /* If we have been reset, we may not send again. */
3653 if (sk
->sk_state
== TCP_CLOSE
)
3656 /* We are not putting this on the write queue, so
3657 * tcp_transmit_skb() will set the ownership to this
3660 buff
= alloc_skb(MAX_TCP_HEADER
,
3661 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3662 if (unlikely(!buff
)) {
3663 inet_csk_schedule_ack(sk
);
3664 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3665 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3666 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3670 /* Reserve space for headers and prepare control bits. */
3671 skb_reserve(buff
, MAX_TCP_HEADER
);
3672 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3674 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3676 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3678 skb_set_tcp_pure_ack(buff
);
3680 /* Send it off, this clears delayed acks for us. */
3681 __tcp_transmit_skb(sk
, buff
, 0, (__force gfp_t
)0, rcv_nxt
);
3683 EXPORT_SYMBOL_GPL(__tcp_send_ack
);
3685 void tcp_send_ack(struct sock
*sk
)
3687 __tcp_send_ack(sk
, tcp_sk(sk
)->rcv_nxt
);
3690 /* This routine sends a packet with an out of date sequence
3691 * number. It assumes the other end will try to ack it.
3693 * Question: what should we make while urgent mode?
3694 * 4.4BSD forces sending single byte of data. We cannot send
3695 * out of window data, because we have SND.NXT==SND.MAX...
3697 * Current solution: to send TWO zero-length segments in urgent mode:
3698 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3699 * out-of-date with SND.UNA-1 to probe window.
3701 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
, int mib
)
3703 struct tcp_sock
*tp
= tcp_sk(sk
);
3704 struct sk_buff
*skb
;
3706 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3707 skb
= alloc_skb(MAX_TCP_HEADER
,
3708 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3712 /* Reserve space for headers and set control bits. */
3713 skb_reserve(skb
, MAX_TCP_HEADER
);
3714 /* Use a previous sequence. This should cause the other
3715 * end to send an ack. Don't queue or clone SKB, just
3718 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3719 NET_INC_STATS(sock_net(sk
), mib
);
3720 return tcp_transmit_skb(sk
, skb
, 0, (__force gfp_t
)0);
3723 /* Called from setsockopt( ... TCP_REPAIR ) */
3724 void tcp_send_window_probe(struct sock
*sk
)
3726 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3727 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3728 tcp_mstamp_refresh(tcp_sk(sk
));
3729 tcp_xmit_probe_skb(sk
, 0, LINUX_MIB_TCPWINPROBE
);
3733 /* Initiate keepalive or window probe from timer. */
3734 int tcp_write_wakeup(struct sock
*sk
, int mib
)
3736 struct tcp_sock
*tp
= tcp_sk(sk
);
3737 struct sk_buff
*skb
;
3739 if (sk
->sk_state
== TCP_CLOSE
)
3742 skb
= tcp_send_head(sk
);
3743 if (skb
&& before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3745 unsigned int mss
= tcp_current_mss(sk
);
3746 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3748 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3749 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3751 /* We are probing the opening of a window
3752 * but the window size is != 0
3753 * must have been a result SWS avoidance ( sender )
3755 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3757 seg_size
= min(seg_size
, mss
);
3758 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3759 if (tcp_fragment(sk
, TCP_FRAG_IN_WRITE_QUEUE
,
3760 skb
, seg_size
, mss
, GFP_ATOMIC
))
3762 } else if (!tcp_skb_pcount(skb
))
3763 tcp_set_skb_tso_segs(skb
, mss
);
3765 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3766 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3768 tcp_event_new_data_sent(sk
, skb
);
3771 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3772 tcp_xmit_probe_skb(sk
, 1, mib
);
3773 return tcp_xmit_probe_skb(sk
, 0, mib
);
3777 /* A window probe timeout has occurred. If window is not closed send
3778 * a partial packet else a zero probe.
3780 void tcp_send_probe0(struct sock
*sk
)
3782 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3783 struct tcp_sock
*tp
= tcp_sk(sk
);
3784 struct net
*net
= sock_net(sk
);
3785 unsigned long probe_max
;
3788 err
= tcp_write_wakeup(sk
, LINUX_MIB_TCPWINPROBE
);
3790 if (tp
->packets_out
|| tcp_write_queue_empty(sk
)) {
3791 /* Cancel probe timer, if it is not required. */
3792 icsk
->icsk_probes_out
= 0;
3793 icsk
->icsk_backoff
= 0;
3798 if (icsk
->icsk_backoff
< net
->ipv4
.sysctl_tcp_retries2
)
3799 icsk
->icsk_backoff
++;
3800 icsk
->icsk_probes_out
++;
3801 probe_max
= TCP_RTO_MAX
;
3803 /* If packet was not sent due to local congestion,
3804 * do not backoff and do not remember icsk_probes_out.
3805 * Let local senders to fight for local resources.
3807 * Use accumulated backoff yet.
3809 if (!icsk
->icsk_probes_out
)
3810 icsk
->icsk_probes_out
= 1;
3811 probe_max
= TCP_RESOURCE_PROBE_INTERVAL
;
3813 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3814 tcp_probe0_when(sk
, probe_max
),
3818 int tcp_rtx_synack(const struct sock
*sk
, struct request_sock
*req
)
3820 const struct tcp_request_sock_ops
*af_ops
= tcp_rsk(req
)->af_specific
;
3824 tcp_rsk(req
)->txhash
= net_tx_rndhash();
3825 res
= af_ops
->send_synack(sk
, NULL
, &fl
, req
, NULL
, TCP_SYNACK_NORMAL
);
3827 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
3828 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
3829 if (unlikely(tcp_passive_fastopen(sk
)))
3830 tcp_sk(sk
)->total_retrans
++;
3831 trace_tcp_retransmit_synack(sk
, req
);
3835 EXPORT_SYMBOL(tcp_rtx_synack
);