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 tp
->packets_out
+= tcp_skb_pcount(skb
);
64 if (!prior_packets
|| icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
67 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
,
71 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
72 * window scaling factor due to loss of precision.
73 * If window has been shrunk, what should we make? It is not clear at all.
74 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
75 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
76 * invalid. OK, let's make this for now:
78 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
80 const struct tcp_sock
*tp
= tcp_sk(sk
);
82 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
) ||
83 (tp
->rx_opt
.wscale_ok
&&
84 ((tp
->snd_nxt
- tcp_wnd_end(tp
)) < (1 << tp
->rx_opt
.rcv_wscale
))))
87 return tcp_wnd_end(tp
);
90 /* Calculate mss to advertise in SYN segment.
91 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
93 * 1. It is independent of path mtu.
94 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
95 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
96 * attached devices, because some buggy hosts are confused by
98 * 4. We do not make 3, we advertise MSS, calculated from first
99 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
100 * This may be overridden via information stored in routing table.
101 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
102 * probably even Jumbo".
104 static __u16
tcp_advertise_mss(struct sock
*sk
)
106 struct tcp_sock
*tp
= tcp_sk(sk
);
107 const struct dst_entry
*dst
= __sk_dst_get(sk
);
108 int mss
= tp
->advmss
;
111 unsigned int metric
= dst_metric_advmss(dst
);
122 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
123 * This is the first part of cwnd validation mechanism.
125 void tcp_cwnd_restart(struct sock
*sk
, s32 delta
)
127 struct tcp_sock
*tp
= tcp_sk(sk
);
128 u32 restart_cwnd
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
129 u32 cwnd
= tp
->snd_cwnd
;
131 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
133 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
134 restart_cwnd
= min(restart_cwnd
, cwnd
);
136 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
138 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
139 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
140 tp
->snd_cwnd_used
= 0;
143 /* Congestion state accounting after a packet has been sent. */
144 static void tcp_event_data_sent(struct tcp_sock
*tp
,
147 struct inet_connection_sock
*icsk
= inet_csk(sk
);
148 const u32 now
= tcp_jiffies32
;
150 if (tcp_packets_in_flight(tp
) == 0)
151 tcp_ca_event(sk
, CA_EVENT_TX_START
);
155 /* If it is a reply for ato after last received
156 * packet, enter pingpong mode.
158 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
159 icsk
->icsk_ack
.pingpong
= 1;
162 /* Account for an ACK we sent. */
163 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
,
166 struct tcp_sock
*tp
= tcp_sk(sk
);
168 if (unlikely(rcv_nxt
!= tp
->rcv_nxt
))
169 return; /* Special ACK sent by DCTCP to reflect ECN */
170 tcp_dec_quickack_mode(sk
, pkts
);
171 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
174 /* Determine a window scaling and initial window to offer.
175 * Based on the assumption that the given amount of space
176 * will be offered. Store the results in the tp structure.
177 * NOTE: for smooth operation initial space offering should
178 * be a multiple of mss if possible. We assume here that mss >= 1.
179 * This MUST be enforced by all callers.
181 void tcp_select_initial_window(const struct sock
*sk
, int __space
, __u32 mss
,
182 __u32
*rcv_wnd
, __u32
*window_clamp
,
183 int wscale_ok
, __u8
*rcv_wscale
,
186 unsigned int space
= (__space
< 0 ? 0 : __space
);
188 /* If no clamp set the clamp to the max possible scaled window */
189 if (*window_clamp
== 0)
190 (*window_clamp
) = (U16_MAX
<< TCP_MAX_WSCALE
);
191 space
= min(*window_clamp
, space
);
193 /* Quantize space offering to a multiple of mss if possible. */
195 space
= rounddown(space
, mss
);
197 /* NOTE: offering an initial window larger than 32767
198 * will break some buggy TCP stacks. If the admin tells us
199 * it is likely we could be speaking with such a buggy stack
200 * we will truncate our initial window offering to 32K-1
201 * unless the remote has sent us a window scaling option,
202 * which we interpret as a sign the remote TCP is not
203 * misinterpreting the window field as a signed quantity.
205 if (sock_net(sk
)->ipv4
.sysctl_tcp_workaround_signed_windows
)
206 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
208 (*rcv_wnd
) = min_t(u32
, space
, U16_MAX
);
211 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
215 /* Set window scaling on max possible window */
216 space
= max_t(u32
, space
, sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[2]);
217 space
= max_t(u32
, space
, sysctl_rmem_max
);
218 space
= min_t(u32
, space
, *window_clamp
);
219 while (space
> U16_MAX
&& (*rcv_wscale
) < TCP_MAX_WSCALE
) {
225 /* Set the clamp no higher than max representable value */
226 (*window_clamp
) = min_t(__u32
, U16_MAX
<< (*rcv_wscale
), *window_clamp
);
228 EXPORT_SYMBOL(tcp_select_initial_window
);
230 /* Chose a new window to advertise, update state in tcp_sock for the
231 * socket, and return result with RFC1323 scaling applied. The return
232 * value can be stuffed directly into th->window for an outgoing
235 static u16
tcp_select_window(struct sock
*sk
)
237 struct tcp_sock
*tp
= tcp_sk(sk
);
238 u32 old_win
= tp
->rcv_wnd
;
239 u32 cur_win
= tcp_receive_window(tp
);
240 u32 new_win
= __tcp_select_window(sk
);
242 /* Never shrink the offered window */
243 if (new_win
< cur_win
) {
244 /* Danger Will Robinson!
245 * Don't update rcv_wup/rcv_wnd here or else
246 * we will not be able to advertise a zero
247 * window in time. --DaveM
249 * Relax Will Robinson.
252 NET_INC_STATS(sock_net(sk
),
253 LINUX_MIB_TCPWANTZEROWINDOWADV
);
254 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
256 tp
->rcv_wnd
= new_win
;
257 tp
->rcv_wup
= tp
->rcv_nxt
;
259 /* Make sure we do not exceed the maximum possible
262 if (!tp
->rx_opt
.rcv_wscale
&&
263 sock_net(sk
)->ipv4
.sysctl_tcp_workaround_signed_windows
)
264 new_win
= min(new_win
, MAX_TCP_WINDOW
);
266 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
268 /* RFC1323 scaling applied */
269 new_win
>>= tp
->rx_opt
.rcv_wscale
;
271 /* If we advertise zero window, disable fast path. */
275 NET_INC_STATS(sock_net(sk
),
276 LINUX_MIB_TCPTOZEROWINDOWADV
);
277 } else if (old_win
== 0) {
278 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFROMZEROWINDOWADV
);
284 /* Packet ECN state for a SYN-ACK */
285 static void tcp_ecn_send_synack(struct sock
*sk
, struct sk_buff
*skb
)
287 const struct tcp_sock
*tp
= tcp_sk(sk
);
289 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
290 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
291 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
292 else if (tcp_ca_needs_ecn(sk
) ||
293 tcp_bpf_ca_needs_ecn(sk
))
297 /* Packet ECN state for a SYN. */
298 static void tcp_ecn_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
300 struct tcp_sock
*tp
= tcp_sk(sk
);
301 bool bpf_needs_ecn
= tcp_bpf_ca_needs_ecn(sk
);
302 bool use_ecn
= sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1 ||
303 tcp_ca_needs_ecn(sk
) || bpf_needs_ecn
;
306 const struct dst_entry
*dst
= __sk_dst_get(sk
);
308 if (dst
&& dst_feature(dst
, RTAX_FEATURE_ECN
))
315 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
316 tp
->ecn_flags
= TCP_ECN_OK
;
317 if (tcp_ca_needs_ecn(sk
) || bpf_needs_ecn
)
322 static void tcp_ecn_clear_syn(struct sock
*sk
, struct sk_buff
*skb
)
324 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn_fallback
)
325 /* tp->ecn_flags are cleared at a later point in time when
326 * SYN ACK is ultimatively being received.
328 TCP_SKB_CB(skb
)->tcp_flags
&= ~(TCPHDR_ECE
| TCPHDR_CWR
);
332 tcp_ecn_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
334 if (inet_rsk(req
)->ecn_ok
)
338 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
341 static void tcp_ecn_send(struct sock
*sk
, struct sk_buff
*skb
,
342 struct tcphdr
*th
, int tcp_header_len
)
344 struct tcp_sock
*tp
= tcp_sk(sk
);
346 if (tp
->ecn_flags
& TCP_ECN_OK
) {
347 /* Not-retransmitted data segment: set ECT and inject CWR. */
348 if (skb
->len
!= tcp_header_len
&&
349 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
351 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
352 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
354 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
356 } else if (!tcp_ca_needs_ecn(sk
)) {
357 /* ACK or retransmitted segment: clear ECT|CE */
358 INET_ECN_dontxmit(sk
);
360 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
365 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
366 * auto increment end seqno.
368 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
370 skb
->ip_summed
= CHECKSUM_PARTIAL
;
372 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
373 TCP_SKB_CB(skb
)->sacked
= 0;
375 tcp_skb_pcount_set(skb
, 1);
377 TCP_SKB_CB(skb
)->seq
= seq
;
378 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
380 TCP_SKB_CB(skb
)->end_seq
= seq
;
383 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
385 return tp
->snd_una
!= tp
->snd_up
;
388 #define OPTION_SACK_ADVERTISE (1 << 0)
389 #define OPTION_TS (1 << 1)
390 #define OPTION_MD5 (1 << 2)
391 #define OPTION_WSCALE (1 << 3)
392 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
393 #define OPTION_SMC (1 << 9)
395 static void smc_options_write(__be32
*ptr
, u16
*options
)
397 #if IS_ENABLED(CONFIG_SMC)
398 if (static_branch_unlikely(&tcp_have_smc
)) {
399 if (unlikely(OPTION_SMC
& *options
)) {
400 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
403 (TCPOLEN_EXP_SMC_BASE
));
404 *ptr
++ = htonl(TCPOPT_SMC_MAGIC
);
410 struct tcp_out_options
{
411 u16 options
; /* bit field of OPTION_* */
412 u16 mss
; /* 0 to disable */
413 u8 ws
; /* window scale, 0 to disable */
414 u8 num_sack_blocks
; /* number of SACK blocks to include */
415 u8 hash_size
; /* bytes in hash_location */
416 __u8
*hash_location
; /* temporary pointer, overloaded */
417 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
418 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
421 /* Write previously computed TCP options to the packet.
423 * Beware: Something in the Internet is very sensitive to the ordering of
424 * TCP options, we learned this through the hard way, so be careful here.
425 * Luckily we can at least blame others for their non-compliance but from
426 * inter-operability perspective it seems that we're somewhat stuck with
427 * the ordering which we have been using if we want to keep working with
428 * those broken things (not that it currently hurts anybody as there isn't
429 * particular reason why the ordering would need to be changed).
431 * At least SACK_PERM as the first option is known to lead to a disaster
432 * (but it may well be that other scenarios fail similarly).
434 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
435 struct tcp_out_options
*opts
)
437 u16 options
= opts
->options
; /* mungable copy */
439 if (unlikely(OPTION_MD5
& options
)) {
440 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
441 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
442 /* overload cookie hash location */
443 opts
->hash_location
= (__u8
*)ptr
;
447 if (unlikely(opts
->mss
)) {
448 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
449 (TCPOLEN_MSS
<< 16) |
453 if (likely(OPTION_TS
& options
)) {
454 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
455 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
456 (TCPOLEN_SACK_PERM
<< 16) |
457 (TCPOPT_TIMESTAMP
<< 8) |
459 options
&= ~OPTION_SACK_ADVERTISE
;
461 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
463 (TCPOPT_TIMESTAMP
<< 8) |
466 *ptr
++ = htonl(opts
->tsval
);
467 *ptr
++ = htonl(opts
->tsecr
);
470 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
471 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
473 (TCPOPT_SACK_PERM
<< 8) |
477 if (unlikely(OPTION_WSCALE
& options
)) {
478 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
479 (TCPOPT_WINDOW
<< 16) |
480 (TCPOLEN_WINDOW
<< 8) |
484 if (unlikely(opts
->num_sack_blocks
)) {
485 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
486 tp
->duplicate_sack
: tp
->selective_acks
;
489 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
492 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
493 TCPOLEN_SACK_PERBLOCK
)));
495 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
497 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
498 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
501 tp
->rx_opt
.dsack
= 0;
504 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
505 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
507 u32 len
; /* Fast Open option length */
510 len
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
511 *ptr
= htonl((TCPOPT_EXP
<< 24) | (len
<< 16) |
512 TCPOPT_FASTOPEN_MAGIC
);
513 p
+= TCPOLEN_EXP_FASTOPEN_BASE
;
515 len
= TCPOLEN_FASTOPEN_BASE
+ foc
->len
;
516 *p
++ = TCPOPT_FASTOPEN
;
520 memcpy(p
, foc
->val
, foc
->len
);
521 if ((len
& 3) == 2) {
522 p
[foc
->len
] = TCPOPT_NOP
;
523 p
[foc
->len
+ 1] = TCPOPT_NOP
;
525 ptr
+= (len
+ 3) >> 2;
528 smc_options_write(ptr
, &options
);
531 static void smc_set_option(const struct tcp_sock
*tp
,
532 struct tcp_out_options
*opts
,
533 unsigned int *remaining
)
535 #if IS_ENABLED(CONFIG_SMC)
536 if (static_branch_unlikely(&tcp_have_smc
)) {
538 if (*remaining
>= TCPOLEN_EXP_SMC_BASE_ALIGNED
) {
539 opts
->options
|= OPTION_SMC
;
540 *remaining
-= TCPOLEN_EXP_SMC_BASE_ALIGNED
;
547 static void smc_set_option_cond(const struct tcp_sock
*tp
,
548 const struct inet_request_sock
*ireq
,
549 struct tcp_out_options
*opts
,
550 unsigned int *remaining
)
552 #if IS_ENABLED(CONFIG_SMC)
553 if (static_branch_unlikely(&tcp_have_smc
)) {
554 if (tp
->syn_smc
&& ireq
->smc_ok
) {
555 if (*remaining
>= TCPOLEN_EXP_SMC_BASE_ALIGNED
) {
556 opts
->options
|= OPTION_SMC
;
557 *remaining
-= TCPOLEN_EXP_SMC_BASE_ALIGNED
;
564 /* Compute TCP options for SYN packets. This is not the final
565 * network wire format yet.
567 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
568 struct tcp_out_options
*opts
,
569 struct tcp_md5sig_key
**md5
)
571 struct tcp_sock
*tp
= tcp_sk(sk
);
572 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
573 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
575 #ifdef CONFIG_TCP_MD5SIG
576 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
578 opts
->options
|= OPTION_MD5
;
579 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
585 /* We always get an MSS option. The option bytes which will be seen in
586 * normal data packets should timestamps be used, must be in the MSS
587 * advertised. But we subtract them from tp->mss_cache so that
588 * calculations in tcp_sendmsg are simpler etc. So account for this
589 * fact here if necessary. If we don't do this correctly, as a
590 * receiver we won't recognize data packets as being full sized when we
591 * should, and thus we won't abide by the delayed ACK rules correctly.
592 * SACKs don't matter, we never delay an ACK when we have any of those
594 opts
->mss
= tcp_advertise_mss(sk
);
595 remaining
-= TCPOLEN_MSS_ALIGNED
;
597 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_timestamps
&& !*md5
)) {
598 opts
->options
|= OPTION_TS
;
599 opts
->tsval
= tcp_skb_timestamp(skb
) + tp
->tsoffset
;
600 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
601 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
603 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_window_scaling
)) {
604 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
605 opts
->options
|= OPTION_WSCALE
;
606 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
608 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_sack
)) {
609 opts
->options
|= OPTION_SACK_ADVERTISE
;
610 if (unlikely(!(OPTION_TS
& opts
->options
)))
611 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
614 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
615 u32 need
= fastopen
->cookie
.len
;
617 need
+= fastopen
->cookie
.exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
618 TCPOLEN_FASTOPEN_BASE
;
619 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
620 if (remaining
>= need
) {
621 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
622 opts
->fastopen_cookie
= &fastopen
->cookie
;
624 tp
->syn_fastopen
= 1;
625 tp
->syn_fastopen_exp
= fastopen
->cookie
.exp
? 1 : 0;
629 smc_set_option(tp
, opts
, &remaining
);
631 return MAX_TCP_OPTION_SPACE
- remaining
;
634 /* Set up TCP options for SYN-ACKs. */
635 static unsigned int tcp_synack_options(const struct sock
*sk
,
636 struct request_sock
*req
,
637 unsigned int mss
, struct sk_buff
*skb
,
638 struct tcp_out_options
*opts
,
639 const struct tcp_md5sig_key
*md5
,
640 struct tcp_fastopen_cookie
*foc
)
642 struct inet_request_sock
*ireq
= inet_rsk(req
);
643 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
645 #ifdef CONFIG_TCP_MD5SIG
647 opts
->options
|= OPTION_MD5
;
648 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
650 /* We can't fit any SACK blocks in a packet with MD5 + TS
651 * options. There was discussion about disabling SACK
652 * rather than TS in order to fit in better with old,
653 * buggy kernels, but that was deemed to be unnecessary.
655 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
659 /* We always send an MSS option. */
661 remaining
-= TCPOLEN_MSS_ALIGNED
;
663 if (likely(ireq
->wscale_ok
)) {
664 opts
->ws
= ireq
->rcv_wscale
;
665 opts
->options
|= OPTION_WSCALE
;
666 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
668 if (likely(ireq
->tstamp_ok
)) {
669 opts
->options
|= OPTION_TS
;
670 opts
->tsval
= tcp_skb_timestamp(skb
) + tcp_rsk(req
)->ts_off
;
671 opts
->tsecr
= req
->ts_recent
;
672 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
674 if (likely(ireq
->sack_ok
)) {
675 opts
->options
|= OPTION_SACK_ADVERTISE
;
676 if (unlikely(!ireq
->tstamp_ok
))
677 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
679 if (foc
!= NULL
&& foc
->len
>= 0) {
682 need
+= foc
->exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
683 TCPOLEN_FASTOPEN_BASE
;
684 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
685 if (remaining
>= need
) {
686 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
687 opts
->fastopen_cookie
= foc
;
692 smc_set_option_cond(tcp_sk(sk
), ireq
, opts
, &remaining
);
694 return MAX_TCP_OPTION_SPACE
- remaining
;
697 /* Compute TCP options for ESTABLISHED sockets. This is not the
698 * final wire format yet.
700 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
701 struct tcp_out_options
*opts
,
702 struct tcp_md5sig_key
**md5
)
704 struct tcp_sock
*tp
= tcp_sk(sk
);
705 unsigned int size
= 0;
706 unsigned int eff_sacks
;
710 #ifdef CONFIG_TCP_MD5SIG
711 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
712 if (unlikely(*md5
)) {
713 opts
->options
|= OPTION_MD5
;
714 size
+= TCPOLEN_MD5SIG_ALIGNED
;
720 if (likely(tp
->rx_opt
.tstamp_ok
)) {
721 opts
->options
|= OPTION_TS
;
722 opts
->tsval
= skb
? tcp_skb_timestamp(skb
) + tp
->tsoffset
: 0;
723 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
724 size
+= TCPOLEN_TSTAMP_ALIGNED
;
727 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
728 if (unlikely(eff_sacks
)) {
729 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
730 opts
->num_sack_blocks
=
731 min_t(unsigned int, eff_sacks
,
732 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
733 TCPOLEN_SACK_PERBLOCK
);
734 if (likely(opts
->num_sack_blocks
))
735 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
736 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
743 /* TCP SMALL QUEUES (TSQ)
745 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
746 * to reduce RTT and bufferbloat.
747 * We do this using a special skb destructor (tcp_wfree).
749 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
750 * needs to be reallocated in a driver.
751 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
753 * Since transmit from skb destructor is forbidden, we use a tasklet
754 * to process all sockets that eventually need to send more skbs.
755 * We use one tasklet per cpu, with its own queue of sockets.
758 struct tasklet_struct tasklet
;
759 struct list_head head
; /* queue of tcp sockets */
761 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
763 static void tcp_tsq_handler(struct sock
*sk
)
765 if ((1 << sk
->sk_state
) &
766 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
767 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
)) {
768 struct tcp_sock
*tp
= tcp_sk(sk
);
770 if (tp
->lost_out
> tp
->retrans_out
&&
771 tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) {
772 tcp_mstamp_refresh(tp
);
773 tcp_xmit_retransmit_queue(sk
);
776 tcp_write_xmit(sk
, tcp_current_mss(sk
), tp
->nonagle
,
781 * One tasklet per cpu tries to send more skbs.
782 * We run in tasklet context but need to disable irqs when
783 * transferring tsq->head because tcp_wfree() might
784 * interrupt us (non NAPI drivers)
786 static void tcp_tasklet_func(unsigned long data
)
788 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
791 struct list_head
*q
, *n
;
795 local_irq_save(flags
);
796 list_splice_init(&tsq
->head
, &list
);
797 local_irq_restore(flags
);
799 list_for_each_safe(q
, n
, &list
) {
800 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
801 list_del(&tp
->tsq_node
);
803 sk
= (struct sock
*)tp
;
804 smp_mb__before_atomic();
805 clear_bit(TSQ_QUEUED
, &sk
->sk_tsq_flags
);
807 if (!sk
->sk_lock
.owned
&&
808 test_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
)) {
810 if (!sock_owned_by_user(sk
)) {
811 clear_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
);
821 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
822 TCPF_WRITE_TIMER_DEFERRED | \
823 TCPF_DELACK_TIMER_DEFERRED | \
824 TCPF_MTU_REDUCED_DEFERRED)
826 * tcp_release_cb - tcp release_sock() callback
829 * called from release_sock() to perform protocol dependent
830 * actions before socket release.
832 void tcp_release_cb(struct sock
*sk
)
834 unsigned long flags
, nflags
;
836 /* perform an atomic operation only if at least one flag is set */
838 flags
= sk
->sk_tsq_flags
;
839 if (!(flags
& TCP_DEFERRED_ALL
))
841 nflags
= flags
& ~TCP_DEFERRED_ALL
;
842 } while (cmpxchg(&sk
->sk_tsq_flags
, flags
, nflags
) != flags
);
844 if (flags
& TCPF_TSQ_DEFERRED
)
847 /* Here begins the tricky part :
848 * We are called from release_sock() with :
850 * 2) sk_lock.slock spinlock held
851 * 3) socket owned by us (sk->sk_lock.owned == 1)
853 * But following code is meant to be called from BH handlers,
854 * so we should keep BH disabled, but early release socket ownership
856 sock_release_ownership(sk
);
858 if (flags
& TCPF_WRITE_TIMER_DEFERRED
) {
859 tcp_write_timer_handler(sk
);
862 if (flags
& TCPF_DELACK_TIMER_DEFERRED
) {
863 tcp_delack_timer_handler(sk
);
866 if (flags
& TCPF_MTU_REDUCED_DEFERRED
) {
867 inet_csk(sk
)->icsk_af_ops
->mtu_reduced(sk
);
871 EXPORT_SYMBOL(tcp_release_cb
);
873 void __init
tcp_tasklet_init(void)
877 for_each_possible_cpu(i
) {
878 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
880 INIT_LIST_HEAD(&tsq
->head
);
881 tasklet_init(&tsq
->tasklet
,
888 * Write buffer destructor automatically called from kfree_skb.
889 * We can't xmit new skbs from this context, as we might already
892 void tcp_wfree(struct sk_buff
*skb
)
894 struct sock
*sk
= skb
->sk
;
895 struct tcp_sock
*tp
= tcp_sk(sk
);
896 unsigned long flags
, nval
, oval
;
898 /* Keep one reference on sk_wmem_alloc.
899 * Will be released by sk_free() from here or tcp_tasklet_func()
901 WARN_ON(refcount_sub_and_test(skb
->truesize
- 1, &sk
->sk_wmem_alloc
));
903 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
904 * Wait until our queues (qdisc + devices) are drained.
906 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
907 * - chance for incoming ACK (processed by another cpu maybe)
908 * to migrate this flow (skb->ooo_okay will be eventually set)
910 if (refcount_read(&sk
->sk_wmem_alloc
) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current
)
913 for (oval
= READ_ONCE(sk
->sk_tsq_flags
);; oval
= nval
) {
914 struct tsq_tasklet
*tsq
;
917 if (!(oval
& TSQF_THROTTLED
) || (oval
& TSQF_QUEUED
))
920 nval
= (oval
& ~TSQF_THROTTLED
) | TSQF_QUEUED
| TCPF_TSQ_DEFERRED
;
921 nval
= cmpxchg(&sk
->sk_tsq_flags
, oval
, nval
);
925 /* queue this socket to tasklet queue */
926 local_irq_save(flags
);
927 tsq
= this_cpu_ptr(&tsq_tasklet
);
928 empty
= list_empty(&tsq
->head
);
929 list_add(&tp
->tsq_node
, &tsq
->head
);
931 tasklet_schedule(&tsq
->tasklet
);
932 local_irq_restore(flags
);
939 /* Note: Called under hard irq.
940 * We can not call TCP stack right away.
942 enum hrtimer_restart
tcp_pace_kick(struct hrtimer
*timer
)
944 struct tcp_sock
*tp
= container_of(timer
, struct tcp_sock
, pacing_timer
);
945 struct sock
*sk
= (struct sock
*)tp
;
946 unsigned long nval
, oval
;
948 for (oval
= READ_ONCE(sk
->sk_tsq_flags
);; oval
= nval
) {
949 struct tsq_tasklet
*tsq
;
952 if (oval
& TSQF_QUEUED
)
955 nval
= (oval
& ~TSQF_THROTTLED
) | TSQF_QUEUED
| TCPF_TSQ_DEFERRED
;
956 nval
= cmpxchg(&sk
->sk_tsq_flags
, oval
, nval
);
960 if (!refcount_inc_not_zero(&sk
->sk_wmem_alloc
))
962 /* queue this socket to tasklet queue */
963 tsq
= this_cpu_ptr(&tsq_tasklet
);
964 empty
= list_empty(&tsq
->head
);
965 list_add(&tp
->tsq_node
, &tsq
->head
);
967 tasklet_schedule(&tsq
->tasklet
);
970 return HRTIMER_NORESTART
;
973 /* BBR congestion control needs pacing.
974 * Same remark for SO_MAX_PACING_RATE.
975 * sch_fq packet scheduler is efficiently handling pacing,
976 * but is not always installed/used.
977 * Return true if TCP stack should pace packets itself.
979 static bool tcp_needs_internal_pacing(const struct sock
*sk
)
981 return smp_load_acquire(&sk
->sk_pacing_status
) == SK_PACING_NEEDED
;
984 static void tcp_internal_pacing(struct sock
*sk
, const struct sk_buff
*skb
)
989 if (!tcp_needs_internal_pacing(sk
))
991 rate
= sk
->sk_pacing_rate
;
992 if (!rate
|| rate
== ~0U)
995 /* Should account for header sizes as sch_fq does,
996 * but lets make things simple.
998 len_ns
= (u64
)skb
->len
* NSEC_PER_SEC
;
999 do_div(len_ns
, rate
);
1000 hrtimer_start(&tcp_sk(sk
)->pacing_timer
,
1001 ktime_add_ns(ktime_get(), len_ns
),
1002 HRTIMER_MODE_ABS_PINNED
);
1005 static void tcp_update_skb_after_send(struct tcp_sock
*tp
, struct sk_buff
*skb
)
1007 skb
->skb_mstamp
= tp
->tcp_mstamp
;
1008 list_move_tail(&skb
->tcp_tsorted_anchor
, &tp
->tsorted_sent_queue
);
1011 /* This routine actually transmits TCP packets queued in by
1012 * tcp_do_sendmsg(). This is used by both the initial
1013 * transmission and possible later retransmissions.
1014 * All SKB's seen here are completely headerless. It is our
1015 * job to build the TCP header, and pass the packet down to
1016 * IP so it can do the same plus pass the packet off to the
1019 * We are working here with either a clone of the original
1020 * SKB, or a fresh unique copy made by the retransmit engine.
1022 static int __tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
,
1023 int clone_it
, gfp_t gfp_mask
, u32 rcv_nxt
)
1025 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1026 struct inet_sock
*inet
;
1027 struct tcp_sock
*tp
;
1028 struct tcp_skb_cb
*tcb
;
1029 struct tcp_out_options opts
;
1030 unsigned int tcp_options_size
, tcp_header_size
;
1031 struct sk_buff
*oskb
= NULL
;
1032 struct tcp_md5sig_key
*md5
;
1036 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
1040 TCP_SKB_CB(skb
)->tx
.in_flight
= TCP_SKB_CB(skb
)->end_seq
1044 tcp_skb_tsorted_save(oskb
) {
1045 if (unlikely(skb_cloned(oskb
)))
1046 skb
= pskb_copy(oskb
, gfp_mask
);
1048 skb
= skb_clone(oskb
, gfp_mask
);
1049 } tcp_skb_tsorted_restore(oskb
);
1054 skb
->skb_mstamp
= tp
->tcp_mstamp
;
1057 tcb
= TCP_SKB_CB(skb
);
1058 memset(&opts
, 0, sizeof(opts
));
1060 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
1061 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
1063 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
1065 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
1067 /* if no packet is in qdisc/device queue, then allow XPS to select
1068 * another queue. We can be called from tcp_tsq_handler()
1069 * which holds one reference to sk_wmem_alloc.
1071 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1072 * One way to get this would be to set skb->truesize = 2 on them.
1074 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) < SKB_TRUESIZE(1);
1076 /* If we had to use memory reserve to allocate this skb,
1077 * this might cause drops if packet is looped back :
1078 * Other socket might not have SOCK_MEMALLOC.
1079 * Packets not looped back do not care about pfmemalloc.
1081 skb
->pfmemalloc
= 0;
1083 skb_push(skb
, tcp_header_size
);
1084 skb_reset_transport_header(skb
);
1088 skb
->destructor
= skb_is_tcp_pure_ack(skb
) ? __sock_wfree
: tcp_wfree
;
1089 skb_set_hash_from_sk(skb
, sk
);
1090 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
1092 skb_set_dst_pending_confirm(skb
, sk
->sk_dst_pending_confirm
);
1094 /* Build TCP header and checksum it. */
1095 th
= (struct tcphdr
*)skb
->data
;
1096 th
->source
= inet
->inet_sport
;
1097 th
->dest
= inet
->inet_dport
;
1098 th
->seq
= htonl(tcb
->seq
);
1099 th
->ack_seq
= htonl(rcv_nxt
);
1100 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
1106 /* The urg_mode check is necessary during a below snd_una win probe */
1107 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
1108 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
1109 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
1111 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
1112 th
->urg_ptr
= htons(0xFFFF);
1117 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
1118 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
1119 if (likely(!(tcb
->tcp_flags
& TCPHDR_SYN
))) {
1120 th
->window
= htons(tcp_select_window(sk
));
1121 tcp_ecn_send(sk
, skb
, th
, tcp_header_size
);
1123 /* RFC1323: The window in SYN & SYN/ACK segments
1126 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
1128 #ifdef CONFIG_TCP_MD5SIG
1129 /* Calculate the MD5 hash, as we have all we need now */
1131 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
1132 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
1137 icsk
->icsk_af_ops
->send_check(sk
, skb
);
1139 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
1140 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
), rcv_nxt
);
1142 if (skb
->len
!= tcp_header_size
) {
1143 tcp_event_data_sent(tp
, sk
);
1144 tp
->data_segs_out
+= tcp_skb_pcount(skb
);
1145 tcp_internal_pacing(sk
, skb
);
1148 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
1149 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
1150 tcp_skb_pcount(skb
));
1152 tp
->segs_out
+= tcp_skb_pcount(skb
);
1153 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1154 skb_shinfo(skb
)->gso_segs
= tcp_skb_pcount(skb
);
1155 skb_shinfo(skb
)->gso_size
= tcp_skb_mss(skb
);
1157 /* Our usage of tstamp should remain private */
1160 /* Cleanup our debris for IP stacks */
1161 memset(skb
->cb
, 0, max(sizeof(struct inet_skb_parm
),
1162 sizeof(struct inet6_skb_parm
)));
1164 err
= icsk
->icsk_af_ops
->queue_xmit(sk
, skb
, &inet
->cork
.fl
);
1166 if (unlikely(err
> 0)) {
1168 err
= net_xmit_eval(err
);
1171 tcp_update_skb_after_send(tp
, oskb
);
1172 tcp_rate_skb_sent(sk
, oskb
);
1177 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
1180 return __tcp_transmit_skb(sk
, skb
, clone_it
, gfp_mask
,
1181 tcp_sk(sk
)->rcv_nxt
);
1184 /* This routine just queues the buffer for sending.
1186 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1187 * otherwise socket can stall.
1189 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1191 struct tcp_sock
*tp
= tcp_sk(sk
);
1193 /* Advance write_seq and place onto the write_queue. */
1194 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
1195 __skb_header_release(skb
);
1196 tcp_add_write_queue_tail(sk
, skb
);
1197 sk
->sk_wmem_queued
+= skb
->truesize
;
1198 sk_mem_charge(sk
, skb
->truesize
);
1201 /* Initialize TSO segments for a packet. */
1202 static void tcp_set_skb_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1204 if (skb
->len
<= mss_now
|| skb
->ip_summed
== CHECKSUM_NONE
) {
1205 /* Avoid the costly divide in the normal
1208 tcp_skb_pcount_set(skb
, 1);
1209 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1211 tcp_skb_pcount_set(skb
, DIV_ROUND_UP(skb
->len
, mss_now
));
1212 TCP_SKB_CB(skb
)->tcp_gso_size
= mss_now
;
1216 /* Pcount in the middle of the write queue got changed, we need to do various
1217 * tweaks to fix counters
1219 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1221 struct tcp_sock
*tp
= tcp_sk(sk
);
1223 tp
->packets_out
-= decr
;
1225 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1226 tp
->sacked_out
-= decr
;
1227 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1228 tp
->retrans_out
-= decr
;
1229 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1230 tp
->lost_out
-= decr
;
1232 /* Reno case is special. Sigh... */
1233 if (tcp_is_reno(tp
) && decr
> 0)
1234 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1236 if (tp
->lost_skb_hint
&&
1237 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1238 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
1239 tp
->lost_cnt_hint
-= decr
;
1241 tcp_verify_left_out(tp
);
1244 static bool tcp_has_tx_tstamp(const struct sk_buff
*skb
)
1246 return TCP_SKB_CB(skb
)->txstamp_ack
||
1247 (skb_shinfo(skb
)->tx_flags
& SKBTX_ANY_TSTAMP
);
1250 static void tcp_fragment_tstamp(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1252 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1254 if (unlikely(tcp_has_tx_tstamp(skb
)) &&
1255 !before(shinfo
->tskey
, TCP_SKB_CB(skb2
)->seq
)) {
1256 struct skb_shared_info
*shinfo2
= skb_shinfo(skb2
);
1257 u8 tsflags
= shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
1259 shinfo
->tx_flags
&= ~tsflags
;
1260 shinfo2
->tx_flags
|= tsflags
;
1261 swap(shinfo
->tskey
, shinfo2
->tskey
);
1262 TCP_SKB_CB(skb2
)->txstamp_ack
= TCP_SKB_CB(skb
)->txstamp_ack
;
1263 TCP_SKB_CB(skb
)->txstamp_ack
= 0;
1267 static void tcp_skb_fragment_eor(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1269 TCP_SKB_CB(skb2
)->eor
= TCP_SKB_CB(skb
)->eor
;
1270 TCP_SKB_CB(skb
)->eor
= 0;
1273 /* Insert buff after skb on the write or rtx queue of sk. */
1274 static void tcp_insert_write_queue_after(struct sk_buff
*skb
,
1275 struct sk_buff
*buff
,
1277 enum tcp_queue tcp_queue
)
1279 if (tcp_queue
== TCP_FRAG_IN_WRITE_QUEUE
)
1280 __skb_queue_after(&sk
->sk_write_queue
, skb
, buff
);
1282 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, buff
);
1285 /* Function to create two new TCP segments. Shrinks the given segment
1286 * to the specified size and appends a new segment with the rest of the
1287 * packet to the list. This won't be called frequently, I hope.
1288 * Remember, these are still headerless SKBs at this point.
1290 int tcp_fragment(struct sock
*sk
, enum tcp_queue tcp_queue
,
1291 struct sk_buff
*skb
, u32 len
,
1292 unsigned int mss_now
, gfp_t gfp
)
1294 struct tcp_sock
*tp
= tcp_sk(sk
);
1295 struct sk_buff
*buff
;
1296 int nsize
, old_factor
;
1301 if (WARN_ON(len
> skb
->len
))
1304 nsize
= skb_headlen(skb
) - len
;
1308 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1309 * We need some allowance to not penalize applications setting small
1311 * Also allow first and last skb in retransmit queue to be split.
1313 limit
= sk
->sk_sndbuf
+ 2 * SKB_TRUESIZE(GSO_MAX_SIZE
);
1314 if (unlikely((sk
->sk_wmem_queued
>> 1) > limit
&&
1315 tcp_queue
!= TCP_FRAG_IN_WRITE_QUEUE
&&
1316 skb
!= tcp_rtx_queue_head(sk
) &&
1317 skb
!= tcp_rtx_queue_tail(sk
))) {
1318 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPWQUEUETOOBIG
);
1322 if (skb_unclone(skb
, gfp
))
1325 /* Get a new skb... force flag on. */
1326 buff
= sk_stream_alloc_skb(sk
, nsize
, gfp
, true);
1328 return -ENOMEM
; /* We'll just try again later. */
1330 sk
->sk_wmem_queued
+= buff
->truesize
;
1331 sk_mem_charge(sk
, buff
->truesize
);
1332 nlen
= skb
->len
- len
- nsize
;
1333 buff
->truesize
+= nlen
;
1334 skb
->truesize
-= nlen
;
1336 /* Correct the sequence numbers. */
1337 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1338 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1339 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1341 /* PSH and FIN should only be set in the second packet. */
1342 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1343 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1344 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1345 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1346 tcp_skb_fragment_eor(skb
, buff
);
1348 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1349 /* Copy and checksum data tail into the new buffer. */
1350 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1351 skb_put(buff
, nsize
),
1356 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1358 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1359 skb_split(skb
, buff
, len
);
1362 buff
->ip_summed
= skb
->ip_summed
;
1364 buff
->tstamp
= skb
->tstamp
;
1365 tcp_fragment_tstamp(skb
, buff
);
1367 old_factor
= tcp_skb_pcount(skb
);
1369 /* Fix up tso_factor for both original and new SKB. */
1370 tcp_set_skb_tso_segs(skb
, mss_now
);
1371 tcp_set_skb_tso_segs(buff
, mss_now
);
1373 /* Update delivered info for the new segment */
1374 TCP_SKB_CB(buff
)->tx
= TCP_SKB_CB(skb
)->tx
;
1376 /* If this packet has been sent out already, we must
1377 * adjust the various packet counters.
1379 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1380 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1381 tcp_skb_pcount(buff
);
1384 tcp_adjust_pcount(sk
, skb
, diff
);
1387 /* Link BUFF into the send queue. */
1388 __skb_header_release(buff
);
1389 tcp_insert_write_queue_after(skb
, buff
, sk
, tcp_queue
);
1390 if (tcp_queue
== TCP_FRAG_IN_RTX_QUEUE
)
1391 list_add(&buff
->tcp_tsorted_anchor
, &skb
->tcp_tsorted_anchor
);
1396 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1397 * data is not copied, but immediately discarded.
1399 static int __pskb_trim_head(struct sk_buff
*skb
, int len
)
1401 struct skb_shared_info
*shinfo
;
1404 eat
= min_t(int, len
, skb_headlen(skb
));
1406 __skb_pull(skb
, eat
);
1413 shinfo
= skb_shinfo(skb
);
1414 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1415 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1418 skb_frag_unref(skb
, i
);
1421 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1423 shinfo
->frags
[k
].page_offset
+= eat
;
1424 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1430 shinfo
->nr_frags
= k
;
1432 skb
->data_len
-= len
;
1433 skb
->len
= skb
->data_len
;
1437 /* Remove acked data from a packet in the transmit queue. */
1438 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1442 if (skb_unclone(skb
, GFP_ATOMIC
))
1445 delta_truesize
= __pskb_trim_head(skb
, len
);
1447 TCP_SKB_CB(skb
)->seq
+= len
;
1448 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1450 if (delta_truesize
) {
1451 skb
->truesize
-= delta_truesize
;
1452 sk
->sk_wmem_queued
-= delta_truesize
;
1453 sk_mem_uncharge(sk
, delta_truesize
);
1454 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1457 /* Any change of skb->len requires recalculation of tso factor. */
1458 if (tcp_skb_pcount(skb
) > 1)
1459 tcp_set_skb_tso_segs(skb
, tcp_skb_mss(skb
));
1464 /* Calculate MSS not accounting any TCP options. */
1465 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1467 const struct tcp_sock
*tp
= tcp_sk(sk
);
1468 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1471 /* Calculate base mss without TCP options:
1472 It is MMS_S - sizeof(tcphdr) of rfc1122
1474 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1476 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1477 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1478 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1480 if (dst
&& dst_allfrag(dst
))
1481 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1484 /* Clamp it (mss_clamp does not include tcp options) */
1485 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1486 mss_now
= tp
->rx_opt
.mss_clamp
;
1488 /* Now subtract optional transport overhead */
1489 mss_now
-= icsk
->icsk_ext_hdr_len
;
1491 /* Then reserve room for full set of TCP options and 8 bytes of data */
1492 mss_now
= max(mss_now
, sock_net(sk
)->ipv4
.sysctl_tcp_min_snd_mss
);
1496 /* Calculate MSS. Not accounting for SACKs here. */
1497 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1499 /* Subtract TCP options size, not including SACKs */
1500 return __tcp_mtu_to_mss(sk
, pmtu
) -
1501 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1504 /* Inverse of above */
1505 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1507 const struct tcp_sock
*tp
= tcp_sk(sk
);
1508 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1512 tp
->tcp_header_len
+
1513 icsk
->icsk_ext_hdr_len
+
1514 icsk
->icsk_af_ops
->net_header_len
;
1516 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1517 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1518 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1520 if (dst
&& dst_allfrag(dst
))
1521 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1525 EXPORT_SYMBOL(tcp_mss_to_mtu
);
1527 /* MTU probing init per socket */
1528 void tcp_mtup_init(struct sock
*sk
)
1530 struct tcp_sock
*tp
= tcp_sk(sk
);
1531 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1532 struct net
*net
= sock_net(sk
);
1534 icsk
->icsk_mtup
.enabled
= net
->ipv4
.sysctl_tcp_mtu_probing
> 1;
1535 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1536 icsk
->icsk_af_ops
->net_header_len
;
1537 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, net
->ipv4
.sysctl_tcp_base_mss
);
1538 icsk
->icsk_mtup
.probe_size
= 0;
1539 if (icsk
->icsk_mtup
.enabled
)
1540 icsk
->icsk_mtup
.probe_timestamp
= tcp_jiffies32
;
1542 EXPORT_SYMBOL(tcp_mtup_init
);
1544 /* This function synchronize snd mss to current pmtu/exthdr set.
1546 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1547 for TCP options, but includes only bare TCP header.
1549 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1550 It is minimum of user_mss and mss received with SYN.
1551 It also does not include TCP options.
1553 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1555 tp->mss_cache is current effective sending mss, including
1556 all tcp options except for SACKs. It is evaluated,
1557 taking into account current pmtu, but never exceeds
1558 tp->rx_opt.mss_clamp.
1560 NOTE1. rfc1122 clearly states that advertised MSS
1561 DOES NOT include either tcp or ip options.
1563 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1564 are READ ONLY outside this function. --ANK (980731)
1566 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1568 struct tcp_sock
*tp
= tcp_sk(sk
);
1569 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1572 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1573 icsk
->icsk_mtup
.search_high
= pmtu
;
1575 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1576 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1578 /* And store cached results */
1579 icsk
->icsk_pmtu_cookie
= pmtu
;
1580 if (icsk
->icsk_mtup
.enabled
)
1581 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1582 tp
->mss_cache
= mss_now
;
1586 EXPORT_SYMBOL(tcp_sync_mss
);
1588 /* Compute the current effective MSS, taking SACKs and IP options,
1589 * and even PMTU discovery events into account.
1591 unsigned int tcp_current_mss(struct sock
*sk
)
1593 const struct tcp_sock
*tp
= tcp_sk(sk
);
1594 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1596 unsigned int header_len
;
1597 struct tcp_out_options opts
;
1598 struct tcp_md5sig_key
*md5
;
1600 mss_now
= tp
->mss_cache
;
1603 u32 mtu
= dst_mtu(dst
);
1604 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1605 mss_now
= tcp_sync_mss(sk
, mtu
);
1608 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1609 sizeof(struct tcphdr
);
1610 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1611 * some common options. If this is an odd packet (because we have SACK
1612 * blocks etc) then our calculated header_len will be different, and
1613 * we have to adjust mss_now correspondingly */
1614 if (header_len
!= tp
->tcp_header_len
) {
1615 int delta
= (int) header_len
- tp
->tcp_header_len
;
1622 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1623 * As additional protections, we do not touch cwnd in retransmission phases,
1624 * and if application hit its sndbuf limit recently.
1626 static void tcp_cwnd_application_limited(struct sock
*sk
)
1628 struct tcp_sock
*tp
= tcp_sk(sk
);
1630 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1631 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1632 /* Limited by application or receiver window. */
1633 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1634 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1635 if (win_used
< tp
->snd_cwnd
) {
1636 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1637 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
1639 tp
->snd_cwnd_used
= 0;
1641 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1644 static void tcp_cwnd_validate(struct sock
*sk
, bool is_cwnd_limited
)
1646 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1647 struct tcp_sock
*tp
= tcp_sk(sk
);
1649 /* Track the maximum number of outstanding packets in each
1650 * window, and remember whether we were cwnd-limited then.
1652 if (!before(tp
->snd_una
, tp
->max_packets_seq
) ||
1653 tp
->packets_out
> tp
->max_packets_out
) {
1654 tp
->max_packets_out
= tp
->packets_out
;
1655 tp
->max_packets_seq
= tp
->snd_nxt
;
1656 tp
->is_cwnd_limited
= is_cwnd_limited
;
1659 if (tcp_is_cwnd_limited(sk
)) {
1660 /* Network is feed fully. */
1661 tp
->snd_cwnd_used
= 0;
1662 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1664 /* Network starves. */
1665 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1666 tp
->snd_cwnd_used
= tp
->packets_out
;
1668 if (sock_net(sk
)->ipv4
.sysctl_tcp_slow_start_after_idle
&&
1669 (s32
)(tcp_jiffies32
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
&&
1670 !ca_ops
->cong_control
)
1671 tcp_cwnd_application_limited(sk
);
1673 /* The following conditions together indicate the starvation
1674 * is caused by insufficient sender buffer:
1675 * 1) just sent some data (see tcp_write_xmit)
1676 * 2) not cwnd limited (this else condition)
1677 * 3) no more data to send (tcp_write_queue_empty())
1678 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1680 if (tcp_write_queue_empty(sk
) && sk
->sk_socket
&&
1681 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
) &&
1682 (1 << sk
->sk_state
) & (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
1683 tcp_chrono_start(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1687 /* Minshall's variant of the Nagle send check. */
1688 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1690 return after(tp
->snd_sml
, tp
->snd_una
) &&
1691 !after(tp
->snd_sml
, tp
->snd_nxt
);
1694 /* Update snd_sml if this skb is under mss
1695 * Note that a TSO packet might end with a sub-mss segment
1696 * The test is really :
1697 * if ((skb->len % mss) != 0)
1698 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1699 * But we can avoid doing the divide again given we already have
1700 * skb_pcount = skb->len / mss_now
1702 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1703 const struct sk_buff
*skb
)
1705 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1706 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1709 /* Return false, if packet can be sent now without violation Nagle's rules:
1710 * 1. It is full sized. (provided by caller in %partial bool)
1711 * 2. Or it contains FIN. (already checked by caller)
1712 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1713 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1714 * With Minshall's modification: all sent small packets are ACKed.
1716 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1720 ((nonagle
& TCP_NAGLE_CORK
) ||
1721 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1724 /* Return how many segs we'd like on a TSO packet,
1725 * to send one TSO packet per ms
1727 u32
tcp_tso_autosize(const struct sock
*sk
, unsigned int mss_now
,
1732 bytes
= min(sk
->sk_pacing_rate
>> sk
->sk_pacing_shift
,
1733 sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
);
1735 /* Goal is to send at least one packet per ms,
1736 * not one big TSO packet every 100 ms.
1737 * This preserves ACK clocking and is consistent
1738 * with tcp_tso_should_defer() heuristic.
1740 segs
= max_t(u32
, bytes
/ mss_now
, min_tso_segs
);
1744 EXPORT_SYMBOL(tcp_tso_autosize
);
1746 /* Return the number of segments we want in the skb we are transmitting.
1747 * See if congestion control module wants to decide; otherwise, autosize.
1749 static u32
tcp_tso_segs(struct sock
*sk
, unsigned int mss_now
)
1751 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1752 u32 tso_segs
= ca_ops
->tso_segs_goal
? ca_ops
->tso_segs_goal(sk
) : 0;
1755 tso_segs
= tcp_tso_autosize(sk
, mss_now
,
1756 sock_net(sk
)->ipv4
.sysctl_tcp_min_tso_segs
);
1757 return min_t(u32
, tso_segs
, sk
->sk_gso_max_segs
);
1760 /* Returns the portion of skb which can be sent right away */
1761 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1762 const struct sk_buff
*skb
,
1763 unsigned int mss_now
,
1764 unsigned int max_segs
,
1767 const struct tcp_sock
*tp
= tcp_sk(sk
);
1768 u32 partial
, needed
, window
, max_len
;
1770 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1771 max_len
= mss_now
* max_segs
;
1773 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1776 needed
= min(skb
->len
, window
);
1778 if (max_len
<= needed
)
1781 partial
= needed
% mss_now
;
1782 /* If last segment is not a full MSS, check if Nagle rules allow us
1783 * to include this last segment in this skb.
1784 * Otherwise, we'll split the skb at last MSS boundary
1786 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
1787 return needed
- partial
;
1792 /* Can at least one segment of SKB be sent right now, according to the
1793 * congestion window rules? If so, return how many segments are allowed.
1795 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1796 const struct sk_buff
*skb
)
1798 u32 in_flight
, cwnd
, halfcwnd
;
1800 /* Don't be strict about the congestion window for the final FIN. */
1801 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1802 tcp_skb_pcount(skb
) == 1)
1805 in_flight
= tcp_packets_in_flight(tp
);
1806 cwnd
= tp
->snd_cwnd
;
1807 if (in_flight
>= cwnd
)
1810 /* For better scheduling, ensure we have at least
1811 * 2 GSO packets in flight.
1813 halfcwnd
= max(cwnd
>> 1, 1U);
1814 return min(halfcwnd
, cwnd
- in_flight
);
1817 /* Initialize TSO state of a skb.
1818 * This must be invoked the first time we consider transmitting
1819 * SKB onto the wire.
1821 static int tcp_init_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1823 int tso_segs
= tcp_skb_pcount(skb
);
1825 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1826 tcp_set_skb_tso_segs(skb
, mss_now
);
1827 tso_segs
= tcp_skb_pcount(skb
);
1833 /* Return true if the Nagle test allows this packet to be
1836 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1837 unsigned int cur_mss
, int nonagle
)
1839 /* Nagle rule does not apply to frames, which sit in the middle of the
1840 * write_queue (they have no chances to get new data).
1842 * This is implemented in the callers, where they modify the 'nonagle'
1843 * argument based upon the location of SKB in the send queue.
1845 if (nonagle
& TCP_NAGLE_PUSH
)
1848 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1849 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1852 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
1858 /* Does at least the first segment of SKB fit into the send window? */
1859 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1860 const struct sk_buff
*skb
,
1861 unsigned int cur_mss
)
1863 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1865 if (skb
->len
> cur_mss
)
1866 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1868 return !after(end_seq
, tcp_wnd_end(tp
));
1871 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1872 * which is put after SKB on the list. It is very much like
1873 * tcp_fragment() except that it may make several kinds of assumptions
1874 * in order to speed up the splitting operation. In particular, we
1875 * know that all the data is in scatter-gather pages, and that the
1876 * packet has never been sent out before (and thus is not cloned).
1878 static int tso_fragment(struct sock
*sk
, enum tcp_queue tcp_queue
,
1879 struct sk_buff
*skb
, unsigned int len
,
1880 unsigned int mss_now
, gfp_t gfp
)
1882 struct sk_buff
*buff
;
1883 int nlen
= skb
->len
- len
;
1886 /* All of a TSO frame must be composed of paged data. */
1887 if (skb
->len
!= skb
->data_len
)
1888 return tcp_fragment(sk
, tcp_queue
, skb
, len
, mss_now
, gfp
);
1890 buff
= sk_stream_alloc_skb(sk
, 0, gfp
, true);
1891 if (unlikely(!buff
))
1894 sk
->sk_wmem_queued
+= buff
->truesize
;
1895 sk_mem_charge(sk
, buff
->truesize
);
1896 buff
->truesize
+= nlen
;
1897 skb
->truesize
-= nlen
;
1899 /* Correct the sequence numbers. */
1900 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1901 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1902 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1904 /* PSH and FIN should only be set in the second packet. */
1905 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1906 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1907 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1909 /* This packet was never sent out yet, so no SACK bits. */
1910 TCP_SKB_CB(buff
)->sacked
= 0;
1912 tcp_skb_fragment_eor(skb
, buff
);
1914 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1915 skb_split(skb
, buff
, len
);
1916 tcp_fragment_tstamp(skb
, buff
);
1918 /* Fix up tso_factor for both original and new SKB. */
1919 tcp_set_skb_tso_segs(skb
, mss_now
);
1920 tcp_set_skb_tso_segs(buff
, mss_now
);
1922 /* Link BUFF into the send queue. */
1923 __skb_header_release(buff
);
1924 tcp_insert_write_queue_after(skb
, buff
, sk
, tcp_queue
);
1929 /* Try to defer sending, if possible, in order to minimize the amount
1930 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1932 * This algorithm is from John Heffner.
1934 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
,
1935 bool *is_cwnd_limited
,
1936 bool *is_rwnd_limited
,
1939 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1940 u32 age
, send_win
, cong_win
, limit
, in_flight
;
1941 struct tcp_sock
*tp
= tcp_sk(sk
);
1942 struct sk_buff
*head
;
1945 if (icsk
->icsk_ca_state
>= TCP_CA_Recovery
)
1948 /* Avoid bursty behavior by allowing defer
1949 * only if the last write was recent.
1951 if ((s32
)(tcp_jiffies32
- tp
->lsndtime
) > 0)
1954 in_flight
= tcp_packets_in_flight(tp
);
1956 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1958 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1960 /* From in_flight test above, we know that cwnd > in_flight. */
1961 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1963 limit
= min(send_win
, cong_win
);
1965 /* If a full-sized TSO skb can be sent, do it. */
1966 if (limit
>= max_segs
* tp
->mss_cache
)
1969 /* Middle in queue won't get any more data, full sendable already? */
1970 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1973 win_divisor
= READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_tso_win_divisor
);
1975 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1977 /* If at least some fraction of a window is available,
1980 chunk
/= win_divisor
;
1984 /* Different approach, try not to defer past a single
1985 * ACK. Receiver should ACK every other full sized
1986 * frame, so if we have space for more than 3 frames
1989 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1993 /* TODO : use tsorted_sent_queue ? */
1994 head
= tcp_rtx_queue_head(sk
);
1997 age
= tcp_stamp_us_delta(tp
->tcp_mstamp
, head
->skb_mstamp
);
1998 /* If next ACK is likely to come too late (half srtt), do not defer */
1999 if (age
< (tp
->srtt_us
>> 4))
2002 /* Ok, it looks like it is advisable to defer.
2003 * Three cases are tracked :
2004 * 1) We are cwnd-limited
2005 * 2) We are rwnd-limited
2006 * 3) We are application limited.
2008 if (cong_win
< send_win
) {
2009 if (cong_win
<= skb
->len
) {
2010 *is_cwnd_limited
= true;
2014 if (send_win
<= skb
->len
) {
2015 *is_rwnd_limited
= true;
2020 /* If this packet won't get more data, do not wait. */
2021 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
2030 static inline void tcp_mtu_check_reprobe(struct sock
*sk
)
2032 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2033 struct tcp_sock
*tp
= tcp_sk(sk
);
2034 struct net
*net
= sock_net(sk
);
2038 interval
= net
->ipv4
.sysctl_tcp_probe_interval
;
2039 delta
= tcp_jiffies32
- icsk
->icsk_mtup
.probe_timestamp
;
2040 if (unlikely(delta
>= interval
* HZ
)) {
2041 int mss
= tcp_current_mss(sk
);
2043 /* Update current search range */
2044 icsk
->icsk_mtup
.probe_size
= 0;
2045 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+
2046 sizeof(struct tcphdr
) +
2047 icsk
->icsk_af_ops
->net_header_len
;
2048 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
2050 /* Update probe time stamp */
2051 icsk
->icsk_mtup
.probe_timestamp
= tcp_jiffies32
;
2055 static bool tcp_can_coalesce_send_queue_head(struct sock
*sk
, int len
)
2057 struct sk_buff
*skb
, *next
;
2059 skb
= tcp_send_head(sk
);
2060 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
2061 if (len
<= skb
->len
)
2064 if (unlikely(TCP_SKB_CB(skb
)->eor
) || tcp_has_tx_tstamp(skb
))
2073 /* Create a new MTU probe if we are ready.
2074 * MTU probe is regularly attempting to increase the path MTU by
2075 * deliberately sending larger packets. This discovers routing
2076 * changes resulting in larger path MTUs.
2078 * Returns 0 if we should wait to probe (no cwnd available),
2079 * 1 if a probe was sent,
2082 static int tcp_mtu_probe(struct sock
*sk
)
2084 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2085 struct tcp_sock
*tp
= tcp_sk(sk
);
2086 struct sk_buff
*skb
, *nskb
, *next
;
2087 struct net
*net
= sock_net(sk
);
2094 /* Not currently probing/verifying,
2096 * have enough cwnd, and
2097 * not SACKing (the variable headers throw things off)
2099 if (likely(!icsk
->icsk_mtup
.enabled
||
2100 icsk
->icsk_mtup
.probe_size
||
2101 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
2102 tp
->snd_cwnd
< 11 ||
2103 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
))
2106 /* Use binary search for probe_size between tcp_mss_base,
2107 * and current mss_clamp. if (search_high - search_low)
2108 * smaller than a threshold, backoff from probing.
2110 mss_now
= tcp_current_mss(sk
);
2111 probe_size
= tcp_mtu_to_mss(sk
, (icsk
->icsk_mtup
.search_high
+
2112 icsk
->icsk_mtup
.search_low
) >> 1);
2113 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
2114 interval
= icsk
->icsk_mtup
.search_high
- icsk
->icsk_mtup
.search_low
;
2115 /* When misfortune happens, we are reprobing actively,
2116 * and then reprobe timer has expired. We stick with current
2117 * probing process by not resetting search range to its orignal.
2119 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
) ||
2120 interval
< net
->ipv4
.sysctl_tcp_probe_threshold
) {
2121 /* Check whether enough time has elaplased for
2122 * another round of probing.
2124 tcp_mtu_check_reprobe(sk
);
2128 /* Have enough data in the send queue to probe? */
2129 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
2132 if (tp
->snd_wnd
< size_needed
)
2134 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
2137 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2138 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
2139 if (!tcp_packets_in_flight(tp
))
2145 if (!tcp_can_coalesce_send_queue_head(sk
, probe_size
))
2148 /* We're allowed to probe. Build it now. */
2149 nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
, false);
2152 sk
->sk_wmem_queued
+= nskb
->truesize
;
2153 sk_mem_charge(sk
, nskb
->truesize
);
2155 skb
= tcp_send_head(sk
);
2157 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
2158 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
2159 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
2160 TCP_SKB_CB(nskb
)->sacked
= 0;
2162 nskb
->ip_summed
= skb
->ip_summed
;
2164 tcp_insert_write_queue_before(nskb
, skb
, sk
);
2165 tcp_highest_sack_replace(sk
, skb
, nskb
);
2168 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
2169 copy
= min_t(int, skb
->len
, probe_size
- len
);
2170 if (nskb
->ip_summed
) {
2171 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
2173 __wsum csum
= skb_copy_and_csum_bits(skb
, 0,
2174 skb_put(nskb
, copy
),
2176 nskb
->csum
= csum_block_add(nskb
->csum
, csum
, len
);
2179 if (skb
->len
<= copy
) {
2180 /* We've eaten all the data from this skb.
2182 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
2183 /* If this is the last SKB we copy and eor is set
2184 * we need to propagate it to the new skb.
2186 TCP_SKB_CB(nskb
)->eor
= TCP_SKB_CB(skb
)->eor
;
2187 tcp_skb_collapse_tstamp(nskb
, skb
);
2188 tcp_unlink_write_queue(skb
, sk
);
2189 sk_wmem_free_skb(sk
, skb
);
2191 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
2192 ~(TCPHDR_FIN
|TCPHDR_PSH
);
2193 if (!skb_shinfo(skb
)->nr_frags
) {
2194 skb_pull(skb
, copy
);
2195 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2196 skb
->csum
= csum_partial(skb
->data
,
2199 __pskb_trim_head(skb
, copy
);
2200 tcp_set_skb_tso_segs(skb
, mss_now
);
2202 TCP_SKB_CB(skb
)->seq
+= copy
;
2207 if (len
>= probe_size
)
2210 tcp_init_tso_segs(nskb
, nskb
->len
);
2212 /* We're ready to send. If this fails, the probe will
2213 * be resegmented into mss-sized pieces by tcp_write_xmit().
2215 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
2216 /* Decrement cwnd here because we are sending
2217 * effectively two packets. */
2219 tcp_event_new_data_sent(sk
, nskb
);
2221 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
2222 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
2223 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
2231 static bool tcp_pacing_check(const struct sock
*sk
)
2233 return tcp_needs_internal_pacing(sk
) &&
2234 hrtimer_active(&tcp_sk(sk
)->pacing_timer
);
2237 /* TCP Small Queues :
2238 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2239 * (These limits are doubled for retransmits)
2241 * - better RTT estimation and ACK scheduling
2244 * Alas, some drivers / subsystems require a fair amount
2245 * of queued bytes to ensure line rate.
2246 * One example is wifi aggregation (802.11 AMPDU)
2248 static bool tcp_small_queue_check(struct sock
*sk
, const struct sk_buff
*skb
,
2249 unsigned int factor
)
2253 limit
= max(2 * skb
->truesize
, sk
->sk_pacing_rate
>> sk
->sk_pacing_shift
);
2254 limit
= min_t(u32
, limit
,
2255 sock_net(sk
)->ipv4
.sysctl_tcp_limit_output_bytes
);
2258 if (refcount_read(&sk
->sk_wmem_alloc
) > limit
) {
2259 /* Always send skb if rtx queue is empty.
2260 * No need to wait for TX completion to call us back,
2261 * after softirq/tasklet schedule.
2262 * This helps when TX completions are delayed too much.
2264 if (tcp_rtx_queue_empty(sk
))
2267 set_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
);
2268 /* It is possible TX completion already happened
2269 * before we set TSQ_THROTTLED, so we must
2270 * test again the condition.
2272 smp_mb__after_atomic();
2273 if (refcount_read(&sk
->sk_wmem_alloc
) > limit
)
2279 static void tcp_chrono_set(struct tcp_sock
*tp
, const enum tcp_chrono
new)
2281 const u32 now
= tcp_jiffies32
;
2282 enum tcp_chrono old
= tp
->chrono_type
;
2284 if (old
> TCP_CHRONO_UNSPEC
)
2285 tp
->chrono_stat
[old
- 1] += now
- tp
->chrono_start
;
2286 tp
->chrono_start
= now
;
2287 tp
->chrono_type
= new;
2290 void tcp_chrono_start(struct sock
*sk
, const enum tcp_chrono type
)
2292 struct tcp_sock
*tp
= tcp_sk(sk
);
2294 /* If there are multiple conditions worthy of tracking in a
2295 * chronograph then the highest priority enum takes precedence
2296 * over the other conditions. So that if something "more interesting"
2297 * starts happening, stop the previous chrono and start a new one.
2299 if (type
> tp
->chrono_type
)
2300 tcp_chrono_set(tp
, type
);
2303 void tcp_chrono_stop(struct sock
*sk
, const enum tcp_chrono type
)
2305 struct tcp_sock
*tp
= tcp_sk(sk
);
2308 /* There are multiple conditions worthy of tracking in a
2309 * chronograph, so that the highest priority enum takes
2310 * precedence over the other conditions (see tcp_chrono_start).
2311 * If a condition stops, we only stop chrono tracking if
2312 * it's the "most interesting" or current chrono we are
2313 * tracking and starts busy chrono if we have pending data.
2315 if (tcp_rtx_and_write_queues_empty(sk
))
2316 tcp_chrono_set(tp
, TCP_CHRONO_UNSPEC
);
2317 else if (type
== tp
->chrono_type
)
2318 tcp_chrono_set(tp
, TCP_CHRONO_BUSY
);
2321 /* This routine writes packets to the network. It advances the
2322 * send_head. This happens as incoming acks open up the remote
2325 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2326 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2327 * account rare use of URG, this is not a big flaw.
2329 * Send at most one packet when push_one > 0. Temporarily ignore
2330 * cwnd limit to force at most one packet out when push_one == 2.
2332 * Returns true, if no segments are in flight and we have queued segments,
2333 * but cannot send anything now because of SWS or another problem.
2335 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
2336 int push_one
, gfp_t gfp
)
2338 struct tcp_sock
*tp
= tcp_sk(sk
);
2339 struct sk_buff
*skb
;
2340 unsigned int tso_segs
, sent_pkts
;
2343 bool is_cwnd_limited
= false, is_rwnd_limited
= false;
2348 tcp_mstamp_refresh(tp
);
2350 /* Do MTU probing. */
2351 result
= tcp_mtu_probe(sk
);
2354 } else if (result
> 0) {
2359 max_segs
= tcp_tso_segs(sk
, mss_now
);
2360 while ((skb
= tcp_send_head(sk
))) {
2363 if (tcp_pacing_check(sk
))
2366 tso_segs
= tcp_init_tso_segs(skb
, mss_now
);
2369 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
) {
2370 /* "skb_mstamp" is used as a start point for the retransmit timer */
2371 tcp_update_skb_after_send(tp
, skb
);
2372 goto repair
; /* Skip network transmission */
2375 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2378 /* Force out a loss probe pkt. */
2384 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
))) {
2385 is_rwnd_limited
= true;
2389 if (tso_segs
== 1) {
2390 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2391 (tcp_skb_is_last(sk
, skb
) ?
2392 nonagle
: TCP_NAGLE_PUSH
))))
2396 tcp_tso_should_defer(sk
, skb
, &is_cwnd_limited
,
2397 &is_rwnd_limited
, max_segs
))
2402 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
2403 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2409 if (skb
->len
> limit
&&
2410 unlikely(tso_fragment(sk
, TCP_FRAG_IN_WRITE_QUEUE
,
2411 skb
, limit
, mss_now
, gfp
)))
2414 if (test_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
))
2415 clear_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
);
2416 if (tcp_small_queue_check(sk
, skb
, 0))
2419 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2423 /* Advance the send_head. This one is sent out.
2424 * This call will increment packets_out.
2426 tcp_event_new_data_sent(sk
, skb
);
2428 tcp_minshall_update(tp
, mss_now
, skb
);
2429 sent_pkts
+= tcp_skb_pcount(skb
);
2435 if (is_rwnd_limited
)
2436 tcp_chrono_start(sk
, TCP_CHRONO_RWND_LIMITED
);
2438 tcp_chrono_stop(sk
, TCP_CHRONO_RWND_LIMITED
);
2440 if (likely(sent_pkts
)) {
2441 if (tcp_in_cwnd_reduction(sk
))
2442 tp
->prr_out
+= sent_pkts
;
2444 /* Send one loss probe per tail loss episode. */
2446 tcp_schedule_loss_probe(sk
, false);
2447 is_cwnd_limited
|= (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
);
2448 tcp_cwnd_validate(sk
, is_cwnd_limited
);
2451 return !tp
->packets_out
&& !tcp_write_queue_empty(sk
);
2454 bool tcp_schedule_loss_probe(struct sock
*sk
, bool advancing_rto
)
2456 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2457 struct tcp_sock
*tp
= tcp_sk(sk
);
2458 u32 timeout
, rto_delta_us
;
2461 /* Don't do any loss probe on a Fast Open connection before 3WHS
2464 if (tp
->fastopen_rsk
)
2467 early_retrans
= sock_net(sk
)->ipv4
.sysctl_tcp_early_retrans
;
2468 /* Schedule a loss probe in 2*RTT for SACK capable connections
2469 * not in loss recovery, that are either limited by cwnd or application.
2471 if ((early_retrans
!= 3 && early_retrans
!= 4) ||
2472 !tp
->packets_out
|| !tcp_is_sack(tp
) ||
2473 (icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2474 icsk
->icsk_ca_state
!= TCP_CA_CWR
))
2477 /* Probe timeout is 2*rtt. Add minimum RTO to account
2478 * for delayed ack when there's one outstanding packet. If no RTT
2479 * sample is available then probe after TCP_TIMEOUT_INIT.
2482 timeout
= usecs_to_jiffies(tp
->srtt_us
>> 2);
2483 if (tp
->packets_out
== 1)
2484 timeout
+= TCP_RTO_MIN
;
2486 timeout
+= TCP_TIMEOUT_MIN
;
2488 timeout
= TCP_TIMEOUT_INIT
;
2491 /* If the RTO formula yields an earlier time, then use that time. */
2492 rto_delta_us
= advancing_rto
?
2493 jiffies_to_usecs(inet_csk(sk
)->icsk_rto
) :
2494 tcp_rto_delta_us(sk
); /* How far in future is RTO? */
2495 if (rto_delta_us
> 0)
2496 timeout
= min_t(u32
, timeout
, usecs_to_jiffies(rto_delta_us
));
2498 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2503 /* Thanks to skb fast clones, we can detect if a prior transmit of
2504 * a packet is still in a qdisc or driver queue.
2505 * In this case, there is very little point doing a retransmit !
2507 static bool skb_still_in_host_queue(const struct sock
*sk
,
2508 const struct sk_buff
*skb
)
2510 if (unlikely(skb_fclone_busy(sk
, skb
))) {
2511 NET_INC_STATS(sock_net(sk
),
2512 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
2518 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2519 * retransmit the last segment.
2521 void tcp_send_loss_probe(struct sock
*sk
)
2523 struct tcp_sock
*tp
= tcp_sk(sk
);
2524 struct sk_buff
*skb
;
2526 int mss
= tcp_current_mss(sk
);
2528 skb
= tcp_send_head(sk
);
2529 if (skb
&& tcp_snd_wnd_test(tp
, skb
, mss
)) {
2530 pcount
= tp
->packets_out
;
2531 tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2532 if (tp
->packets_out
> pcount
)
2536 skb
= skb_rb_last(&sk
->tcp_rtx_queue
);
2538 if (unlikely(!skb
)) {
2539 WARN_ONCE(tp
->packets_out
,
2540 "invalid inflight: %u state %u cwnd %u mss %d\n",
2541 tp
->packets_out
, sk
->sk_state
, tp
->snd_cwnd
, mss
);
2542 inet_csk(sk
)->icsk_pending
= 0;
2546 /* At most one outstanding TLP retransmission. */
2547 if (tp
->tlp_high_seq
)
2550 if (skb_still_in_host_queue(sk
, skb
))
2553 pcount
= tcp_skb_pcount(skb
);
2554 if (WARN_ON(!pcount
))
2557 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2558 if (unlikely(tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
,
2559 (pcount
- 1) * mss
, mss
,
2562 skb
= skb_rb_next(skb
);
2565 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2568 if (__tcp_retransmit_skb(sk
, skb
, 1))
2571 /* Record snd_nxt for loss detection. */
2572 tp
->tlp_high_seq
= tp
->snd_nxt
;
2575 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSPROBES
);
2576 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2577 inet_csk(sk
)->icsk_pending
= 0;
2582 /* Push out any pending frames which were held back due to
2583 * TCP_CORK or attempt at coalescing tiny packets.
2584 * The socket must be locked by the caller.
2586 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2589 /* If we are closed, the bytes will have to remain here.
2590 * In time closedown will finish, we empty the write queue and
2591 * all will be happy.
2593 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2596 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2597 sk_gfp_mask(sk
, GFP_ATOMIC
)))
2598 tcp_check_probe_timer(sk
);
2601 /* Send _single_ skb sitting at the send head. This function requires
2602 * true push pending frames to setup probe timer etc.
2604 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2606 struct sk_buff
*skb
= tcp_send_head(sk
);
2608 BUG_ON(!skb
|| skb
->len
< mss_now
);
2610 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2613 /* This function returns the amount that we can raise the
2614 * usable window based on the following constraints
2616 * 1. The window can never be shrunk once it is offered (RFC 793)
2617 * 2. We limit memory per socket
2620 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2621 * RECV.NEXT + RCV.WIN fixed until:
2622 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2624 * i.e. don't raise the right edge of the window until you can raise
2625 * it at least MSS bytes.
2627 * Unfortunately, the recommended algorithm breaks header prediction,
2628 * since header prediction assumes th->window stays fixed.
2630 * Strictly speaking, keeping th->window fixed violates the receiver
2631 * side SWS prevention criteria. The problem is that under this rule
2632 * a stream of single byte packets will cause the right side of the
2633 * window to always advance by a single byte.
2635 * Of course, if the sender implements sender side SWS prevention
2636 * then this will not be a problem.
2638 * BSD seems to make the following compromise:
2640 * If the free space is less than the 1/4 of the maximum
2641 * space available and the free space is less than 1/2 mss,
2642 * then set the window to 0.
2643 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2644 * Otherwise, just prevent the window from shrinking
2645 * and from being larger than the largest representable value.
2647 * This prevents incremental opening of the window in the regime
2648 * where TCP is limited by the speed of the reader side taking
2649 * data out of the TCP receive queue. It does nothing about
2650 * those cases where the window is constrained on the sender side
2651 * because the pipeline is full.
2653 * BSD also seems to "accidentally" limit itself to windows that are a
2654 * multiple of MSS, at least until the free space gets quite small.
2655 * This would appear to be a side effect of the mbuf implementation.
2656 * Combining these two algorithms results in the observed behavior
2657 * of having a fixed window size at almost all times.
2659 * Below we obtain similar behavior by forcing the offered window to
2660 * a multiple of the mss when it is feasible to do so.
2662 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2663 * Regular options like TIMESTAMP are taken into account.
2665 u32
__tcp_select_window(struct sock
*sk
)
2667 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2668 struct tcp_sock
*tp
= tcp_sk(sk
);
2669 /* MSS for the peer's data. Previous versions used mss_clamp
2670 * here. I don't know if the value based on our guesses
2671 * of peer's MSS is better for the performance. It's more correct
2672 * but may be worse for the performance because of rcv_mss
2673 * fluctuations. --SAW 1998/11/1
2675 int mss
= icsk
->icsk_ack
.rcv_mss
;
2676 int free_space
= tcp_space(sk
);
2677 int allowed_space
= tcp_full_space(sk
);
2678 int full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2681 if (unlikely(mss
> full_space
)) {
2686 if (free_space
< (full_space
>> 1)) {
2687 icsk
->icsk_ack
.quick
= 0;
2689 if (tcp_under_memory_pressure(sk
))
2690 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2693 /* free_space might become our new window, make sure we don't
2694 * increase it due to wscale.
2696 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2698 /* if free space is less than mss estimate, or is below 1/16th
2699 * of the maximum allowed, try to move to zero-window, else
2700 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2701 * new incoming data is dropped due to memory limits.
2702 * With large window, mss test triggers way too late in order
2703 * to announce zero window in time before rmem limit kicks in.
2705 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2709 if (free_space
> tp
->rcv_ssthresh
)
2710 free_space
= tp
->rcv_ssthresh
;
2712 /* Don't do rounding if we are using window scaling, since the
2713 * scaled window will not line up with the MSS boundary anyway.
2715 if (tp
->rx_opt
.rcv_wscale
) {
2716 window
= free_space
;
2718 /* Advertise enough space so that it won't get scaled away.
2719 * Import case: prevent zero window announcement if
2720 * 1<<rcv_wscale > mss.
2722 window
= ALIGN(window
, (1 << tp
->rx_opt
.rcv_wscale
));
2724 window
= tp
->rcv_wnd
;
2725 /* Get the largest window that is a nice multiple of mss.
2726 * Window clamp already applied above.
2727 * If our current window offering is within 1 mss of the
2728 * free space we just keep it. This prevents the divide
2729 * and multiply from happening most of the time.
2730 * We also don't do any window rounding when the free space
2733 if (window
<= free_space
- mss
|| window
> free_space
)
2734 window
= rounddown(free_space
, mss
);
2735 else if (mss
== full_space
&&
2736 free_space
> window
+ (full_space
>> 1))
2737 window
= free_space
;
2743 void tcp_skb_collapse_tstamp(struct sk_buff
*skb
,
2744 const struct sk_buff
*next_skb
)
2746 if (unlikely(tcp_has_tx_tstamp(next_skb
))) {
2747 const struct skb_shared_info
*next_shinfo
=
2748 skb_shinfo(next_skb
);
2749 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
2751 shinfo
->tx_flags
|= next_shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
2752 shinfo
->tskey
= next_shinfo
->tskey
;
2753 TCP_SKB_CB(skb
)->txstamp_ack
|=
2754 TCP_SKB_CB(next_skb
)->txstamp_ack
;
2758 /* Collapses two adjacent SKB's during retransmission. */
2759 static bool tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2761 struct tcp_sock
*tp
= tcp_sk(sk
);
2762 struct sk_buff
*next_skb
= skb_rb_next(skb
);
2763 int skb_size
, next_skb_size
;
2765 skb_size
= skb
->len
;
2766 next_skb_size
= next_skb
->len
;
2768 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2770 if (next_skb_size
) {
2771 if (next_skb_size
<= skb_availroom(skb
))
2772 skb_copy_bits(next_skb
, 0, skb_put(skb
, next_skb_size
),
2774 else if (!tcp_skb_shift(skb
, next_skb
, 1, next_skb_size
))
2777 tcp_highest_sack_replace(sk
, next_skb
, skb
);
2779 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2780 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2782 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2783 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2785 /* Update sequence range on original skb. */
2786 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2788 /* Merge over control information. This moves PSH/FIN etc. over */
2789 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2791 /* All done, get rid of second SKB and account for it so
2792 * packet counting does not break.
2794 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2795 TCP_SKB_CB(skb
)->eor
= TCP_SKB_CB(next_skb
)->eor
;
2797 /* changed transmit queue under us so clear hints */
2798 tcp_clear_retrans_hints_partial(tp
);
2799 if (next_skb
== tp
->retransmit_skb_hint
)
2800 tp
->retransmit_skb_hint
= skb
;
2802 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2804 tcp_skb_collapse_tstamp(skb
, next_skb
);
2806 tcp_rtx_queue_unlink_and_free(next_skb
, sk
);
2810 /* Check if coalescing SKBs is legal. */
2811 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2813 if (tcp_skb_pcount(skb
) > 1)
2815 if (skb_cloned(skb
))
2817 /* Some heuristics for collapsing over SACK'd could be invented */
2818 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2824 /* Collapse packets in the retransmit queue to make to create
2825 * less packets on the wire. This is only done on retransmission.
2827 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2830 struct tcp_sock
*tp
= tcp_sk(sk
);
2831 struct sk_buff
*skb
= to
, *tmp
;
2834 if (!sock_net(sk
)->ipv4
.sysctl_tcp_retrans_collapse
)
2836 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2839 skb_rbtree_walk_from_safe(skb
, tmp
) {
2840 if (!tcp_can_collapse(sk
, skb
))
2843 if (!tcp_skb_can_collapse_to(to
))
2856 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2859 if (!tcp_collapse_retrans(sk
, to
))
2864 /* This retransmits one SKB. Policy decisions and retransmit queue
2865 * state updates are done by the caller. Returns non-zero if an
2866 * error occurred which prevented the send.
2868 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2870 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2871 struct tcp_sock
*tp
= tcp_sk(sk
);
2872 unsigned int cur_mss
;
2876 /* Inconclusive MTU probe */
2877 if (icsk
->icsk_mtup
.probe_size
)
2878 icsk
->icsk_mtup
.probe_size
= 0;
2880 /* Do not sent more than we queued. 1/4 is reserved for possible
2881 * copying overhead: fragmentation, tunneling, mangling etc.
2883 if (refcount_read(&sk
->sk_wmem_alloc
) >
2884 min_t(u32
, sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2),
2888 if (skb_still_in_host_queue(sk
, skb
))
2891 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2892 if (unlikely(before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))) {
2896 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2900 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2901 return -EHOSTUNREACH
; /* Routing failure or similar. */
2903 cur_mss
= tcp_current_mss(sk
);
2905 /* If receiver has shrunk his window, and skb is out of
2906 * new window, do not retransmit it. The exception is the
2907 * case, when window is shrunk to zero. In this case
2908 * our retransmit serves as a zero window probe.
2910 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2911 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2914 len
= cur_mss
* segs
;
2915 if (skb
->len
> len
) {
2916 if (tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
, len
,
2917 cur_mss
, GFP_ATOMIC
))
2918 return -ENOMEM
; /* We'll try again later. */
2920 if (skb_unclone(skb
, GFP_ATOMIC
))
2923 diff
= tcp_skb_pcount(skb
);
2924 tcp_set_skb_tso_segs(skb
, cur_mss
);
2925 diff
-= tcp_skb_pcount(skb
);
2927 tcp_adjust_pcount(sk
, skb
, diff
);
2928 if (skb
->len
< cur_mss
)
2929 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2932 /* RFC3168, section 6.1.1.1. ECN fallback */
2933 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN_ECN
) == TCPHDR_SYN_ECN
)
2934 tcp_ecn_clear_syn(sk
, skb
);
2936 /* Update global and local TCP statistics. */
2937 segs
= tcp_skb_pcount(skb
);
2938 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
, segs
);
2939 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2940 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
2941 tp
->total_retrans
+= segs
;
2943 /* make sure skb->data is aligned on arches that require it
2944 * and check if ack-trimming & collapsing extended the headroom
2945 * beyond what csum_start can cover.
2947 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2948 skb_headroom(skb
) >= 0xFFFF)) {
2949 struct sk_buff
*nskb
;
2951 tcp_skb_tsorted_save(skb
) {
2952 nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
, GFP_ATOMIC
);
2953 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2955 } tcp_skb_tsorted_restore(skb
);
2958 tcp_update_skb_after_send(tp
, skb
);
2959 tcp_rate_skb_sent(sk
, skb
);
2962 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2966 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
2967 trace_tcp_retransmit_skb(sk
, skb
);
2968 } else if (err
!= -EBUSY
) {
2969 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
, segs
);
2974 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2976 struct tcp_sock
*tp
= tcp_sk(sk
);
2977 int err
= __tcp_retransmit_skb(sk
, skb
, segs
);
2980 #if FASTRETRANS_DEBUG > 0
2981 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2982 net_dbg_ratelimited("retrans_out leaked\n");
2985 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2986 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2988 /* Save stamp of the first retransmit. */
2989 if (!tp
->retrans_stamp
)
2990 tp
->retrans_stamp
= tcp_skb_timestamp(skb
);
2994 if (tp
->undo_retrans
< 0)
2995 tp
->undo_retrans
= 0;
2996 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
3000 /* This gets called after a retransmit timeout, and the initially
3001 * retransmitted data is acknowledged. It tries to continue
3002 * resending the rest of the retransmit queue, until either
3003 * we've sent it all or the congestion window limit is reached.
3005 void tcp_xmit_retransmit_queue(struct sock
*sk
)
3007 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3008 struct sk_buff
*skb
, *rtx_head
, *hole
= NULL
;
3009 struct tcp_sock
*tp
= tcp_sk(sk
);
3013 if (!tp
->packets_out
)
3016 rtx_head
= tcp_rtx_queue_head(sk
);
3017 skb
= tp
->retransmit_skb_hint
?: rtx_head
;
3018 max_segs
= tcp_tso_segs(sk
, tcp_current_mss(sk
));
3019 skb_rbtree_walk_from(skb
) {
3023 if (tcp_pacing_check(sk
))
3026 /* we could do better than to assign each time */
3028 tp
->retransmit_skb_hint
= skb
;
3030 segs
= tp
->snd_cwnd
- tcp_packets_in_flight(tp
);
3033 sacked
= TCP_SKB_CB(skb
)->sacked
;
3034 /* In case tcp_shift_skb_data() have aggregated large skbs,
3035 * we need to make sure not sending too bigs TSO packets
3037 segs
= min_t(int, segs
, max_segs
);
3039 if (tp
->retrans_out
>= tp
->lost_out
) {
3041 } else if (!(sacked
& TCPCB_LOST
)) {
3042 if (!hole
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
3047 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
3048 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
3050 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
3053 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
3056 if (tcp_small_queue_check(sk
, skb
, 1))
3059 if (tcp_retransmit_skb(sk
, skb
, segs
))
3062 NET_ADD_STATS(sock_net(sk
), mib_idx
, tcp_skb_pcount(skb
));
3064 if (tcp_in_cwnd_reduction(sk
))
3065 tp
->prr_out
+= tcp_skb_pcount(skb
);
3067 if (skb
== rtx_head
&&
3068 icsk
->icsk_pending
!= ICSK_TIME_REO_TIMEOUT
)
3069 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3070 inet_csk(sk
)->icsk_rto
,
3075 /* We allow to exceed memory limits for FIN packets to expedite
3076 * connection tear down and (memory) recovery.
3077 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3078 * or even be forced to close flow without any FIN.
3079 * In general, we want to allow one skb per socket to avoid hangs
3080 * with edge trigger epoll()
3082 void sk_forced_mem_schedule(struct sock
*sk
, int size
)
3086 if (size
<= sk
->sk_forward_alloc
)
3088 amt
= sk_mem_pages(size
);
3089 sk
->sk_forward_alloc
+= amt
* SK_MEM_QUANTUM
;
3090 sk_memory_allocated_add(sk
, amt
);
3092 if (mem_cgroup_sockets_enabled
&& sk
->sk_memcg
)
3093 mem_cgroup_charge_skmem(sk
->sk_memcg
, amt
);
3096 /* Send a FIN. The caller locks the socket for us.
3097 * We should try to send a FIN packet really hard, but eventually give up.
3099 void tcp_send_fin(struct sock
*sk
)
3101 struct sk_buff
*skb
, *tskb
= tcp_write_queue_tail(sk
);
3102 struct tcp_sock
*tp
= tcp_sk(sk
);
3104 /* Optimization, tack on the FIN if we have one skb in write queue and
3105 * this skb was not yet sent, or we are under memory pressure.
3106 * Note: in the latter case, FIN packet will be sent after a timeout,
3107 * as TCP stack thinks it has already been transmitted.
3109 if (!tskb
&& tcp_under_memory_pressure(sk
))
3110 tskb
= skb_rb_last(&sk
->tcp_rtx_queue
);
3114 TCP_SKB_CB(tskb
)->tcp_flags
|= TCPHDR_FIN
;
3115 TCP_SKB_CB(tskb
)->end_seq
++;
3117 if (tcp_write_queue_empty(sk
)) {
3118 /* This means tskb was already sent.
3119 * Pretend we included the FIN on previous transmit.
3120 * We need to set tp->snd_nxt to the value it would have
3121 * if FIN had been sent. This is because retransmit path
3122 * does not change tp->snd_nxt.
3128 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, sk
->sk_allocation
);
3129 if (unlikely(!skb
)) {
3134 INIT_LIST_HEAD(&skb
->tcp_tsorted_anchor
);
3135 skb_reserve(skb
, MAX_TCP_HEADER
);
3136 sk_forced_mem_schedule(sk
, skb
->truesize
);
3137 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3138 tcp_init_nondata_skb(skb
, tp
->write_seq
,
3139 TCPHDR_ACK
| TCPHDR_FIN
);
3140 tcp_queue_skb(sk
, skb
);
3142 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
), TCP_NAGLE_OFF
);
3145 /* We get here when a process closes a file descriptor (either due to
3146 * an explicit close() or as a byproduct of exit()'ing) and there
3147 * was unread data in the receive queue. This behavior is recommended
3148 * by RFC 2525, section 2.17. -DaveM
3150 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
3152 struct sk_buff
*skb
;
3154 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
3156 /* NOTE: No TCP options attached and we never retransmit this. */
3157 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
3159 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3163 /* Reserve space for headers and prepare control bits. */
3164 skb_reserve(skb
, MAX_TCP_HEADER
);
3165 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
3166 TCPHDR_ACK
| TCPHDR_RST
);
3167 tcp_mstamp_refresh(tcp_sk(sk
));
3169 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
3170 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3172 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3173 * skb here is different to the troublesome skb, so use NULL
3175 trace_tcp_send_reset(sk
, NULL
);
3178 /* Send a crossed SYN-ACK during socket establishment.
3179 * WARNING: This routine must only be called when we have already sent
3180 * a SYN packet that crossed the incoming SYN that caused this routine
3181 * to get called. If this assumption fails then the initial rcv_wnd
3182 * and rcv_wscale values will not be correct.
3184 int tcp_send_synack(struct sock
*sk
)
3186 struct sk_buff
*skb
;
3188 skb
= tcp_rtx_queue_head(sk
);
3189 if (!skb
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
3190 pr_err("%s: wrong queue state\n", __func__
);
3193 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
3194 if (skb_cloned(skb
)) {
3195 struct sk_buff
*nskb
;
3197 tcp_skb_tsorted_save(skb
) {
3198 nskb
= skb_copy(skb
, GFP_ATOMIC
);
3199 } tcp_skb_tsorted_restore(skb
);
3202 INIT_LIST_HEAD(&nskb
->tcp_tsorted_anchor
);
3203 tcp_rtx_queue_unlink_and_free(skb
, sk
);
3204 __skb_header_release(nskb
);
3205 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, nskb
);
3206 sk
->sk_wmem_queued
+= nskb
->truesize
;
3207 sk_mem_charge(sk
, nskb
->truesize
);
3211 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
3212 tcp_ecn_send_synack(sk
, skb
);
3214 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3218 * tcp_make_synack - Prepare a SYN-ACK.
3219 * sk: listener socket
3220 * dst: dst entry attached to the SYNACK
3221 * req: request_sock pointer
3223 * Allocate one skb and build a SYNACK packet.
3224 * @dst is consumed : Caller should not use it again.
3226 struct sk_buff
*tcp_make_synack(const struct sock
*sk
, struct dst_entry
*dst
,
3227 struct request_sock
*req
,
3228 struct tcp_fastopen_cookie
*foc
,
3229 enum tcp_synack_type synack_type
)
3231 struct inet_request_sock
*ireq
= inet_rsk(req
);
3232 const struct tcp_sock
*tp
= tcp_sk(sk
);
3233 struct tcp_md5sig_key
*md5
= NULL
;
3234 struct tcp_out_options opts
;
3235 struct sk_buff
*skb
;
3236 int tcp_header_size
;
3240 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
3241 if (unlikely(!skb
)) {
3245 /* Reserve space for headers. */
3246 skb_reserve(skb
, MAX_TCP_HEADER
);
3248 switch (synack_type
) {
3249 case TCP_SYNACK_NORMAL
:
3250 skb_set_owner_w(skb
, req_to_sk(req
));
3252 case TCP_SYNACK_COOKIE
:
3253 /* Under synflood, we do not attach skb to a socket,
3254 * to avoid false sharing.
3257 case TCP_SYNACK_FASTOPEN
:
3258 /* sk is a const pointer, because we want to express multiple
3259 * cpu might call us concurrently.
3260 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3262 skb_set_owner_w(skb
, (struct sock
*)sk
);
3265 skb_dst_set(skb
, dst
);
3267 mss
= tcp_mss_clamp(tp
, dst_metric_advmss(dst
));
3269 memset(&opts
, 0, sizeof(opts
));
3270 #ifdef CONFIG_SYN_COOKIES
3271 if (unlikely(req
->cookie_ts
))
3272 skb
->skb_mstamp
= cookie_init_timestamp(req
);
3275 skb
->skb_mstamp
= tcp_clock_us();
3277 #ifdef CONFIG_TCP_MD5SIG
3279 md5
= tcp_rsk(req
)->af_specific
->req_md5_lookup(sk
, req_to_sk(req
));
3281 skb_set_hash(skb
, tcp_rsk(req
)->txhash
, PKT_HASH_TYPE_L4
);
3282 tcp_header_size
= tcp_synack_options(sk
, req
, mss
, skb
, &opts
, md5
,
3285 skb_push(skb
, tcp_header_size
);
3286 skb_reset_transport_header(skb
);
3288 th
= (struct tcphdr
*)skb
->data
;
3289 memset(th
, 0, sizeof(struct tcphdr
));
3292 tcp_ecn_make_synack(req
, th
);
3293 th
->source
= htons(ireq
->ir_num
);
3294 th
->dest
= ireq
->ir_rmt_port
;
3295 skb
->mark
= ireq
->ir_mark
;
3296 skb
->ip_summed
= CHECKSUM_PARTIAL
;
3297 th
->seq
= htonl(tcp_rsk(req
)->snt_isn
);
3298 /* XXX data is queued and acked as is. No buffer/window check */
3299 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
3301 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3302 th
->window
= htons(min(req
->rsk_rcv_wnd
, 65535U));
3303 tcp_options_write((__be32
*)(th
+ 1), NULL
, &opts
);
3304 th
->doff
= (tcp_header_size
>> 2);
3305 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
3307 #ifdef CONFIG_TCP_MD5SIG
3308 /* Okay, we have all we need - do the md5 hash if needed */
3310 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
3311 md5
, req_to_sk(req
), skb
);
3315 /* Do not fool tcpdump (if any), clean our debris */
3319 EXPORT_SYMBOL(tcp_make_synack
);
3321 static void tcp_ca_dst_init(struct sock
*sk
, const struct dst_entry
*dst
)
3323 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3324 const struct tcp_congestion_ops
*ca
;
3325 u32 ca_key
= dst_metric(dst
, RTAX_CC_ALGO
);
3327 if (ca_key
== TCP_CA_UNSPEC
)
3331 ca
= tcp_ca_find_key(ca_key
);
3332 if (likely(ca
&& try_module_get(ca
->owner
))) {
3333 module_put(icsk
->icsk_ca_ops
->owner
);
3334 icsk
->icsk_ca_dst_locked
= tcp_ca_dst_locked(dst
);
3335 icsk
->icsk_ca_ops
= ca
;
3340 /* Do all connect socket setups that can be done AF independent. */
3341 static void tcp_connect_init(struct sock
*sk
)
3343 const struct dst_entry
*dst
= __sk_dst_get(sk
);
3344 struct tcp_sock
*tp
= tcp_sk(sk
);
3348 /* We'll fix this up when we get a response from the other end.
3349 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3351 tp
->tcp_header_len
= sizeof(struct tcphdr
);
3352 if (sock_net(sk
)->ipv4
.sysctl_tcp_timestamps
)
3353 tp
->tcp_header_len
+= TCPOLEN_TSTAMP_ALIGNED
;
3355 #ifdef CONFIG_TCP_MD5SIG
3356 if (tp
->af_specific
->md5_lookup(sk
, sk
))
3357 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
3360 /* If user gave his TCP_MAXSEG, record it to clamp */
3361 if (tp
->rx_opt
.user_mss
)
3362 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3365 tcp_sync_mss(sk
, dst_mtu(dst
));
3367 tcp_ca_dst_init(sk
, dst
);
3369 if (!tp
->window_clamp
)
3370 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
3371 tp
->advmss
= tcp_mss_clamp(tp
, dst_metric_advmss(dst
));
3373 tcp_initialize_rcv_mss(sk
);
3375 /* limit the window selection if the user enforce a smaller rx buffer */
3376 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
3377 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
3378 tp
->window_clamp
= tcp_full_space(sk
);
3380 rcv_wnd
= tcp_rwnd_init_bpf(sk
);
3382 rcv_wnd
= dst_metric(dst
, RTAX_INITRWND
);
3384 tcp_select_initial_window(sk
, tcp_full_space(sk
),
3385 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
3388 sock_net(sk
)->ipv4
.sysctl_tcp_window_scaling
,
3392 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
3393 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
3396 sock_reset_flag(sk
, SOCK_DONE
);
3399 tcp_write_queue_purge(sk
);
3400 tp
->snd_una
= tp
->write_seq
;
3401 tp
->snd_sml
= tp
->write_seq
;
3402 tp
->snd_up
= tp
->write_seq
;
3403 tp
->snd_nxt
= tp
->write_seq
;
3405 if (likely(!tp
->repair
))
3408 tp
->rcv_tstamp
= tcp_jiffies32
;
3409 tp
->rcv_wup
= tp
->rcv_nxt
;
3410 tp
->copied_seq
= tp
->rcv_nxt
;
3412 inet_csk(sk
)->icsk_rto
= tcp_timeout_init(sk
);
3413 inet_csk(sk
)->icsk_retransmits
= 0;
3414 tcp_clear_retrans(tp
);
3417 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
3419 struct tcp_sock
*tp
= tcp_sk(sk
);
3420 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
3422 tcb
->end_seq
+= skb
->len
;
3423 __skb_header_release(skb
);
3424 sk
->sk_wmem_queued
+= skb
->truesize
;
3425 sk_mem_charge(sk
, skb
->truesize
);
3426 tp
->write_seq
= tcb
->end_seq
;
3427 tp
->packets_out
+= tcp_skb_pcount(skb
);
3430 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3431 * queue a data-only packet after the regular SYN, such that regular SYNs
3432 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3433 * only the SYN sequence, the data are retransmitted in the first ACK.
3434 * If cookie is not cached or other error occurs, falls back to send a
3435 * regular SYN with Fast Open cookie request option.
3437 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
3439 struct tcp_sock
*tp
= tcp_sk(sk
);
3440 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
3442 struct sk_buff
*syn_data
;
3444 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
3445 if (!tcp_fastopen_cookie_check(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
))
3448 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3449 * user-MSS. Reserve maximum option space for middleboxes that add
3450 * private TCP options. The cost is reduced data space in SYN :(
3452 tp
->rx_opt
.mss_clamp
= tcp_mss_clamp(tp
, tp
->rx_opt
.mss_clamp
);
3454 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
3455 MAX_TCP_OPTION_SPACE
;
3457 space
= min_t(size_t, space
, fo
->size
);
3459 /* limit to order-0 allocations */
3460 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
3462 syn_data
= sk_stream_alloc_skb(sk
, space
, sk
->sk_allocation
, false);
3465 syn_data
->ip_summed
= CHECKSUM_PARTIAL
;
3466 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
3468 int copied
= copy_from_iter(skb_put(syn_data
, space
), space
,
3469 &fo
->data
->msg_iter
);
3470 if (unlikely(!copied
)) {
3471 tcp_skb_tsorted_anchor_cleanup(syn_data
);
3472 kfree_skb(syn_data
);
3475 if (copied
!= space
) {
3476 skb_trim(syn_data
, copied
);
3480 /* No more data pending in inet_wait_for_connect() */
3481 if (space
== fo
->size
)
3485 tcp_connect_queue_skb(sk
, syn_data
);
3487 tcp_chrono_start(sk
, TCP_CHRONO_BUSY
);
3489 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
3491 syn
->skb_mstamp
= syn_data
->skb_mstamp
;
3493 /* Now full SYN+DATA was cloned and sent (or not),
3494 * remove the SYN from the original skb (syn_data)
3495 * we keep in write queue in case of a retransmit, as we
3496 * also have the SYN packet (with no data) in the same queue.
3498 TCP_SKB_CB(syn_data
)->seq
++;
3499 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
3501 tp
->syn_data
= (fo
->copied
> 0);
3502 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, syn_data
);
3503 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3507 /* data was not sent, put it in write_queue */
3508 __skb_queue_tail(&sk
->sk_write_queue
, syn_data
);
3509 tp
->packets_out
-= tcp_skb_pcount(syn_data
);
3512 /* Send a regular SYN with Fast Open cookie request option */
3513 if (fo
->cookie
.len
> 0)
3515 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3517 tp
->syn_fastopen
= 0;
3519 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3523 /* Build a SYN and send it off. */
3524 int tcp_connect(struct sock
*sk
)
3526 struct tcp_sock
*tp
= tcp_sk(sk
);
3527 struct sk_buff
*buff
;
3530 tcp_call_bpf(sk
, BPF_SOCK_OPS_TCP_CONNECT_CB
);
3532 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
3533 return -EHOSTUNREACH
; /* Routing failure or similar. */
3535 tcp_connect_init(sk
);
3537 if (unlikely(tp
->repair
)) {
3538 tcp_finish_connect(sk
, NULL
);
3542 buff
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
, true);
3543 if (unlikely(!buff
))
3546 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3547 tcp_mstamp_refresh(tp
);
3548 tp
->retrans_stamp
= tcp_time_stamp(tp
);
3549 tcp_connect_queue_skb(sk
, buff
);
3550 tcp_ecn_send_syn(sk
, buff
);
3551 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, buff
);
3553 /* Send off SYN; include data in Fast Open. */
3554 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3555 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3556 if (err
== -ECONNREFUSED
)
3559 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3560 * in order to make this packet get counted in tcpOutSegs.
3562 tp
->snd_nxt
= tp
->write_seq
;
3563 tp
->pushed_seq
= tp
->write_seq
;
3564 buff
= tcp_send_head(sk
);
3565 if (unlikely(buff
)) {
3566 tp
->snd_nxt
= TCP_SKB_CB(buff
)->seq
;
3567 tp
->pushed_seq
= TCP_SKB_CB(buff
)->seq
;
3569 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3571 /* Timer for repeating the SYN until an answer. */
3572 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3573 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3576 EXPORT_SYMBOL(tcp_connect
);
3578 /* Send out a delayed ack, the caller does the policy checking
3579 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3582 void tcp_send_delayed_ack(struct sock
*sk
)
3584 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3585 int ato
= icsk
->icsk_ack
.ato
;
3586 unsigned long timeout
;
3588 if (ato
> TCP_DELACK_MIN
) {
3589 const struct tcp_sock
*tp
= tcp_sk(sk
);
3590 int max_ato
= HZ
/ 2;
3592 if (icsk
->icsk_ack
.pingpong
||
3593 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3594 max_ato
= TCP_DELACK_MAX
;
3596 /* Slow path, intersegment interval is "high". */
3598 /* If some rtt estimate is known, use it to bound delayed ack.
3599 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3603 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3610 ato
= min(ato
, max_ato
);
3613 /* Stay within the limit we were given */
3614 timeout
= jiffies
+ ato
;
3616 /* Use new timeout only if there wasn't a older one earlier. */
3617 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3618 /* If delack timer was blocked or is about to expire,
3621 if (icsk
->icsk_ack
.blocked
||
3622 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3627 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3628 timeout
= icsk
->icsk_ack
.timeout
;
3630 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3631 icsk
->icsk_ack
.timeout
= timeout
;
3632 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3635 /* This routine sends an ack and also updates the window. */
3636 void __tcp_send_ack(struct sock
*sk
, u32 rcv_nxt
)
3638 struct sk_buff
*buff
;
3640 /* If we have been reset, we may not send again. */
3641 if (sk
->sk_state
== TCP_CLOSE
)
3644 /* We are not putting this on the write queue, so
3645 * tcp_transmit_skb() will set the ownership to this
3648 buff
= alloc_skb(MAX_TCP_HEADER
,
3649 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3650 if (unlikely(!buff
)) {
3651 inet_csk_schedule_ack(sk
);
3652 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3653 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3654 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3658 /* Reserve space for headers and prepare control bits. */
3659 skb_reserve(buff
, MAX_TCP_HEADER
);
3660 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3662 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3664 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3666 skb_set_tcp_pure_ack(buff
);
3668 /* Send it off, this clears delayed acks for us. */
3669 __tcp_transmit_skb(sk
, buff
, 0, (__force gfp_t
)0, rcv_nxt
);
3671 EXPORT_SYMBOL_GPL(__tcp_send_ack
);
3673 void tcp_send_ack(struct sock
*sk
)
3675 __tcp_send_ack(sk
, tcp_sk(sk
)->rcv_nxt
);
3678 /* This routine sends a packet with an out of date sequence
3679 * number. It assumes the other end will try to ack it.
3681 * Question: what should we make while urgent mode?
3682 * 4.4BSD forces sending single byte of data. We cannot send
3683 * out of window data, because we have SND.NXT==SND.MAX...
3685 * Current solution: to send TWO zero-length segments in urgent mode:
3686 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3687 * out-of-date with SND.UNA-1 to probe window.
3689 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
, int mib
)
3691 struct tcp_sock
*tp
= tcp_sk(sk
);
3692 struct sk_buff
*skb
;
3694 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3695 skb
= alloc_skb(MAX_TCP_HEADER
,
3696 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3700 /* Reserve space for headers and set control bits. */
3701 skb_reserve(skb
, MAX_TCP_HEADER
);
3702 /* Use a previous sequence. This should cause the other
3703 * end to send an ack. Don't queue or clone SKB, just
3706 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3707 NET_INC_STATS(sock_net(sk
), mib
);
3708 return tcp_transmit_skb(sk
, skb
, 0, (__force gfp_t
)0);
3711 /* Called from setsockopt( ... TCP_REPAIR ) */
3712 void tcp_send_window_probe(struct sock
*sk
)
3714 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3715 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3716 tcp_mstamp_refresh(tcp_sk(sk
));
3717 tcp_xmit_probe_skb(sk
, 0, LINUX_MIB_TCPWINPROBE
);
3721 /* Initiate keepalive or window probe from timer. */
3722 int tcp_write_wakeup(struct sock
*sk
, int mib
)
3724 struct tcp_sock
*tp
= tcp_sk(sk
);
3725 struct sk_buff
*skb
;
3727 if (sk
->sk_state
== TCP_CLOSE
)
3730 skb
= tcp_send_head(sk
);
3731 if (skb
&& before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3733 unsigned int mss
= tcp_current_mss(sk
);
3734 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3736 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3737 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3739 /* We are probing the opening of a window
3740 * but the window size is != 0
3741 * must have been a result SWS avoidance ( sender )
3743 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3745 seg_size
= min(seg_size
, mss
);
3746 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3747 if (tcp_fragment(sk
, TCP_FRAG_IN_WRITE_QUEUE
,
3748 skb
, seg_size
, mss
, GFP_ATOMIC
))
3750 } else if (!tcp_skb_pcount(skb
))
3751 tcp_set_skb_tso_segs(skb
, mss
);
3753 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3754 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3756 tcp_event_new_data_sent(sk
, skb
);
3759 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3760 tcp_xmit_probe_skb(sk
, 1, mib
);
3761 return tcp_xmit_probe_skb(sk
, 0, mib
);
3765 /* A window probe timeout has occurred. If window is not closed send
3766 * a partial packet else a zero probe.
3768 void tcp_send_probe0(struct sock
*sk
)
3770 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3771 struct tcp_sock
*tp
= tcp_sk(sk
);
3772 struct net
*net
= sock_net(sk
);
3773 unsigned long probe_max
;
3776 err
= tcp_write_wakeup(sk
, LINUX_MIB_TCPWINPROBE
);
3778 if (tp
->packets_out
|| tcp_write_queue_empty(sk
)) {
3779 /* Cancel probe timer, if it is not required. */
3780 icsk
->icsk_probes_out
= 0;
3781 icsk
->icsk_backoff
= 0;
3786 if (icsk
->icsk_backoff
< net
->ipv4
.sysctl_tcp_retries2
)
3787 icsk
->icsk_backoff
++;
3788 icsk
->icsk_probes_out
++;
3789 probe_max
= TCP_RTO_MAX
;
3791 /* If packet was not sent due to local congestion,
3792 * do not backoff and do not remember icsk_probes_out.
3793 * Let local senders to fight for local resources.
3795 * Use accumulated backoff yet.
3797 if (!icsk
->icsk_probes_out
)
3798 icsk
->icsk_probes_out
= 1;
3799 probe_max
= TCP_RESOURCE_PROBE_INTERVAL
;
3801 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3802 tcp_probe0_when(sk
, probe_max
),
3806 int tcp_rtx_synack(const struct sock
*sk
, struct request_sock
*req
)
3808 const struct tcp_request_sock_ops
*af_ops
= tcp_rsk(req
)->af_specific
;
3812 tcp_rsk(req
)->txhash
= net_tx_rndhash();
3813 res
= af_ops
->send_synack(sk
, NULL
, &fl
, req
, NULL
, TCP_SYNACK_NORMAL
);
3815 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
3816 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
3817 if (unlikely(tcp_passive_fastopen(sk
)))
3818 tcp_sk(sk
)->total_retrans
++;
3819 trace_tcp_retransmit_synack(sk
, req
);
3823 EXPORT_SYMBOL(tcp_rtx_synack
);