2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly
= 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly
= 0;
53 /* Default TSQ limit of four TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly
= 262144;
56 /* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
60 int sysctl_tcp_tso_win_divisor __read_mostly
= 3;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly
= 1;
65 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
66 int push_one
, gfp_t gfp
);
68 /* Account for new data that has been sent to the network. */
69 static void tcp_event_new_data_sent(struct sock
*sk
, const struct sk_buff
*skb
)
71 struct inet_connection_sock
*icsk
= inet_csk(sk
);
72 struct tcp_sock
*tp
= tcp_sk(sk
);
73 unsigned int prior_packets
= tp
->packets_out
;
75 tcp_advance_send_head(sk
, skb
);
76 tp
->snd_nxt
= TCP_SKB_CB(skb
)->end_seq
;
78 tp
->packets_out
+= tcp_skb_pcount(skb
);
79 if (!prior_packets
|| icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
||
80 icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
) {
84 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
,
88 /* SND.NXT, if window was not shrunk.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
94 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
96 const struct tcp_sock
*tp
= tcp_sk(sk
);
98 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
))
101 return tcp_wnd_end(tp
);
104 /* Calculate mss to advertise in SYN segment.
105 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
107 * 1. It is independent of path mtu.
108 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
109 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
110 * attached devices, because some buggy hosts are confused by
112 * 4. We do not make 3, we advertise MSS, calculated from first
113 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
114 * This may be overridden via information stored in routing table.
115 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
116 * probably even Jumbo".
118 static __u16
tcp_advertise_mss(struct sock
*sk
)
120 struct tcp_sock
*tp
= tcp_sk(sk
);
121 const struct dst_entry
*dst
= __sk_dst_get(sk
);
122 int mss
= tp
->advmss
;
125 unsigned int metric
= dst_metric_advmss(dst
);
136 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
137 * This is the first part of cwnd validation mechanism.
139 void tcp_cwnd_restart(struct sock
*sk
, s32 delta
)
141 struct tcp_sock
*tp
= tcp_sk(sk
);
142 u32 restart_cwnd
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
143 u32 cwnd
= tp
->snd_cwnd
;
145 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
147 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
148 restart_cwnd
= min(restart_cwnd
, cwnd
);
150 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
152 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
153 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
154 tp
->snd_cwnd_used
= 0;
157 /* Congestion state accounting after a packet has been sent. */
158 static void tcp_event_data_sent(struct tcp_sock
*tp
,
161 struct inet_connection_sock
*icsk
= inet_csk(sk
);
162 const u32 now
= tcp_time_stamp
;
164 if (tcp_packets_in_flight(tp
) == 0)
165 tcp_ca_event(sk
, CA_EVENT_TX_START
);
169 /* If it is a reply for ato after last received
170 * packet, enter pingpong mode.
172 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
173 icsk
->icsk_ack
.pingpong
= 1;
176 /* Account for an ACK we sent. */
177 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
179 tcp_dec_quickack_mode(sk
, pkts
);
180 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
184 u32
tcp_default_init_rwnd(u32 mss
)
186 /* Initial receive window should be twice of TCP_INIT_CWND to
187 * enable proper sending of new unsent data during fast recovery
188 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
189 * limit when mss is larger than 1460.
191 u32 init_rwnd
= TCP_INIT_CWND
* 2;
194 init_rwnd
= max((1460 * init_rwnd
) / mss
, 2U);
198 /* Determine a window scaling and initial window to offer.
199 * Based on the assumption that the given amount of space
200 * will be offered. Store the results in the tp structure.
201 * NOTE: for smooth operation initial space offering should
202 * be a multiple of mss if possible. We assume here that mss >= 1.
203 * This MUST be enforced by all callers.
205 void tcp_select_initial_window(int __space
, __u32 mss
,
206 __u32
*rcv_wnd
, __u32
*window_clamp
,
207 int wscale_ok
, __u8
*rcv_wscale
,
210 unsigned int space
= (__space
< 0 ? 0 : __space
);
212 /* If no clamp set the clamp to the max possible scaled window */
213 if (*window_clamp
== 0)
214 (*window_clamp
) = (65535 << 14);
215 space
= min(*window_clamp
, space
);
217 /* Quantize space offering to a multiple of mss if possible. */
219 space
= (space
/ mss
) * mss
;
221 /* NOTE: offering an initial window larger than 32767
222 * will break some buggy TCP stacks. If the admin tells us
223 * it is likely we could be speaking with such a buggy stack
224 * we will truncate our initial window offering to 32K-1
225 * unless the remote has sent us a window scaling option,
226 * which we interpret as a sign the remote TCP is not
227 * misinterpreting the window field as a signed quantity.
229 if (sysctl_tcp_workaround_signed_windows
)
230 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
236 /* Set window scaling on max possible window
237 * See RFC1323 for an explanation of the limit to 14
239 space
= max_t(u32
, space
, sysctl_tcp_rmem
[2]);
240 space
= max_t(u32
, space
, sysctl_rmem_max
);
241 space
= min_t(u32
, space
, *window_clamp
);
242 while (space
> 65535 && (*rcv_wscale
) < 14) {
248 if (mss
> (1 << *rcv_wscale
)) {
249 if (!init_rcv_wnd
) /* Use default unless specified otherwise */
250 init_rcv_wnd
= tcp_default_init_rwnd(mss
);
251 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
254 /* Set the clamp no higher than max representable value */
255 (*window_clamp
) = min(65535U << (*rcv_wscale
), *window_clamp
);
257 EXPORT_SYMBOL(tcp_select_initial_window
);
259 /* Chose a new window to advertise, update state in tcp_sock for the
260 * socket, and return result with RFC1323 scaling applied. The return
261 * value can be stuffed directly into th->window for an outgoing
264 static u16
tcp_select_window(struct sock
*sk
)
266 struct tcp_sock
*tp
= tcp_sk(sk
);
267 u32 old_win
= tp
->rcv_wnd
;
268 u32 cur_win
= tcp_receive_window(tp
);
269 u32 new_win
= __tcp_select_window(sk
);
271 /* Never shrink the offered window */
272 if (new_win
< cur_win
) {
273 /* Danger Will Robinson!
274 * Don't update rcv_wup/rcv_wnd here or else
275 * we will not be able to advertise a zero
276 * window in time. --DaveM
278 * Relax Will Robinson.
281 NET_INC_STATS(sock_net(sk
),
282 LINUX_MIB_TCPWANTZEROWINDOWADV
);
283 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
285 tp
->rcv_wnd
= new_win
;
286 tp
->rcv_wup
= tp
->rcv_nxt
;
288 /* Make sure we do not exceed the maximum possible
291 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
292 new_win
= min(new_win
, MAX_TCP_WINDOW
);
294 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
296 /* RFC1323 scaling applied */
297 new_win
>>= tp
->rx_opt
.rcv_wscale
;
299 /* If we advertise zero window, disable fast path. */
303 NET_INC_STATS(sock_net(sk
),
304 LINUX_MIB_TCPTOZEROWINDOWADV
);
305 } else if (old_win
== 0) {
306 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFROMZEROWINDOWADV
);
312 /* Packet ECN state for a SYN-ACK */
313 static void tcp_ecn_send_synack(struct sock
*sk
, struct sk_buff
*skb
)
315 const struct tcp_sock
*tp
= tcp_sk(sk
);
317 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
318 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
319 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
320 else if (tcp_ca_needs_ecn(sk
))
324 /* Packet ECN state for a SYN. */
325 static void tcp_ecn_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
327 struct tcp_sock
*tp
= tcp_sk(sk
);
328 bool use_ecn
= sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1 ||
329 tcp_ca_needs_ecn(sk
);
332 const struct dst_entry
*dst
= __sk_dst_get(sk
);
334 if (dst
&& dst_feature(dst
, RTAX_FEATURE_ECN
))
341 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
342 tp
->ecn_flags
= TCP_ECN_OK
;
343 if (tcp_ca_needs_ecn(sk
))
348 static void tcp_ecn_clear_syn(struct sock
*sk
, struct sk_buff
*skb
)
350 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn_fallback
)
351 /* tp->ecn_flags are cleared at a later point in time when
352 * SYN ACK is ultimatively being received.
354 TCP_SKB_CB(skb
)->tcp_flags
&= ~(TCPHDR_ECE
| TCPHDR_CWR
);
358 tcp_ecn_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
360 if (inet_rsk(req
)->ecn_ok
)
364 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
367 static void tcp_ecn_send(struct sock
*sk
, struct sk_buff
*skb
,
368 struct tcphdr
*th
, int tcp_header_len
)
370 struct tcp_sock
*tp
= tcp_sk(sk
);
372 if (tp
->ecn_flags
& TCP_ECN_OK
) {
373 /* Not-retransmitted data segment: set ECT and inject CWR. */
374 if (skb
->len
!= tcp_header_len
&&
375 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
377 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
378 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
380 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
382 } else if (!tcp_ca_needs_ecn(sk
)) {
383 /* ACK or retransmitted segment: clear ECT|CE */
384 INET_ECN_dontxmit(sk
);
386 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
391 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
392 * auto increment end seqno.
394 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
396 skb
->ip_summed
= CHECKSUM_PARTIAL
;
399 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
400 TCP_SKB_CB(skb
)->sacked
= 0;
402 tcp_skb_pcount_set(skb
, 1);
404 TCP_SKB_CB(skb
)->seq
= seq
;
405 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
407 TCP_SKB_CB(skb
)->end_seq
= seq
;
410 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
412 return tp
->snd_una
!= tp
->snd_up
;
415 #define OPTION_SACK_ADVERTISE (1 << 0)
416 #define OPTION_TS (1 << 1)
417 #define OPTION_MD5 (1 << 2)
418 #define OPTION_WSCALE (1 << 3)
419 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
421 struct tcp_out_options
{
422 u16 options
; /* bit field of OPTION_* */
423 u16 mss
; /* 0 to disable */
424 u8 ws
; /* window scale, 0 to disable */
425 u8 num_sack_blocks
; /* number of SACK blocks to include */
426 u8 hash_size
; /* bytes in hash_location */
427 __u8
*hash_location
; /* temporary pointer, overloaded */
428 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
429 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
432 /* Write previously computed TCP options to the packet.
434 * Beware: Something in the Internet is very sensitive to the ordering of
435 * TCP options, we learned this through the hard way, so be careful here.
436 * Luckily we can at least blame others for their non-compliance but from
437 * inter-operability perspective it seems that we're somewhat stuck with
438 * the ordering which we have been using if we want to keep working with
439 * those broken things (not that it currently hurts anybody as there isn't
440 * particular reason why the ordering would need to be changed).
442 * At least SACK_PERM as the first option is known to lead to a disaster
443 * (but it may well be that other scenarios fail similarly).
445 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
446 struct tcp_out_options
*opts
)
448 u16 options
= opts
->options
; /* mungable copy */
450 if (unlikely(OPTION_MD5
& options
)) {
451 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
452 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
453 /* overload cookie hash location */
454 opts
->hash_location
= (__u8
*)ptr
;
458 if (unlikely(opts
->mss
)) {
459 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
460 (TCPOLEN_MSS
<< 16) |
464 if (likely(OPTION_TS
& options
)) {
465 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
466 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
467 (TCPOLEN_SACK_PERM
<< 16) |
468 (TCPOPT_TIMESTAMP
<< 8) |
470 options
&= ~OPTION_SACK_ADVERTISE
;
472 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
474 (TCPOPT_TIMESTAMP
<< 8) |
477 *ptr
++ = htonl(opts
->tsval
);
478 *ptr
++ = htonl(opts
->tsecr
);
481 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
482 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
484 (TCPOPT_SACK_PERM
<< 8) |
488 if (unlikely(OPTION_WSCALE
& options
)) {
489 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
490 (TCPOPT_WINDOW
<< 16) |
491 (TCPOLEN_WINDOW
<< 8) |
495 if (unlikely(opts
->num_sack_blocks
)) {
496 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
497 tp
->duplicate_sack
: tp
->selective_acks
;
500 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
503 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
504 TCPOLEN_SACK_PERBLOCK
)));
506 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
508 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
509 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
512 tp
->rx_opt
.dsack
= 0;
515 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
516 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
518 u32 len
; /* Fast Open option length */
521 len
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
522 *ptr
= htonl((TCPOPT_EXP
<< 24) | (len
<< 16) |
523 TCPOPT_FASTOPEN_MAGIC
);
524 p
+= TCPOLEN_EXP_FASTOPEN_BASE
;
526 len
= TCPOLEN_FASTOPEN_BASE
+ foc
->len
;
527 *p
++ = TCPOPT_FASTOPEN
;
531 memcpy(p
, foc
->val
, foc
->len
);
532 if ((len
& 3) == 2) {
533 p
[foc
->len
] = TCPOPT_NOP
;
534 p
[foc
->len
+ 1] = TCPOPT_NOP
;
536 ptr
+= (len
+ 3) >> 2;
540 /* Compute TCP options for SYN packets. This is not the final
541 * network wire format yet.
543 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
544 struct tcp_out_options
*opts
,
545 struct tcp_md5sig_key
**md5
)
547 struct tcp_sock
*tp
= tcp_sk(sk
);
548 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
549 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
551 #ifdef CONFIG_TCP_MD5SIG
552 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
554 opts
->options
|= OPTION_MD5
;
555 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
561 /* We always get an MSS option. The option bytes which will be seen in
562 * normal data packets should timestamps be used, must be in the MSS
563 * advertised. But we subtract them from tp->mss_cache so that
564 * calculations in tcp_sendmsg are simpler etc. So account for this
565 * fact here if necessary. If we don't do this correctly, as a
566 * receiver we won't recognize data packets as being full sized when we
567 * should, and thus we won't abide by the delayed ACK rules correctly.
568 * SACKs don't matter, we never delay an ACK when we have any of those
570 opts
->mss
= tcp_advertise_mss(sk
);
571 remaining
-= TCPOLEN_MSS_ALIGNED
;
573 if (likely(sysctl_tcp_timestamps
&& !*md5
)) {
574 opts
->options
|= OPTION_TS
;
575 opts
->tsval
= tcp_skb_timestamp(skb
) + tp
->tsoffset
;
576 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
577 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
579 if (likely(sysctl_tcp_window_scaling
)) {
580 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
581 opts
->options
|= OPTION_WSCALE
;
582 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
584 if (likely(sysctl_tcp_sack
)) {
585 opts
->options
|= OPTION_SACK_ADVERTISE
;
586 if (unlikely(!(OPTION_TS
& opts
->options
)))
587 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
590 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
591 u32 need
= fastopen
->cookie
.len
;
593 need
+= fastopen
->cookie
.exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
594 TCPOLEN_FASTOPEN_BASE
;
595 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
596 if (remaining
>= need
) {
597 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
598 opts
->fastopen_cookie
= &fastopen
->cookie
;
600 tp
->syn_fastopen
= 1;
601 tp
->syn_fastopen_exp
= fastopen
->cookie
.exp
? 1 : 0;
605 return MAX_TCP_OPTION_SPACE
- remaining
;
608 /* Set up TCP options for SYN-ACKs. */
609 static unsigned int tcp_synack_options(struct request_sock
*req
,
610 unsigned int mss
, struct sk_buff
*skb
,
611 struct tcp_out_options
*opts
,
612 const struct tcp_md5sig_key
*md5
,
613 struct tcp_fastopen_cookie
*foc
)
615 struct inet_request_sock
*ireq
= inet_rsk(req
);
616 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
618 #ifdef CONFIG_TCP_MD5SIG
620 opts
->options
|= OPTION_MD5
;
621 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
623 /* We can't fit any SACK blocks in a packet with MD5 + TS
624 * options. There was discussion about disabling SACK
625 * rather than TS in order to fit in better with old,
626 * buggy kernels, but that was deemed to be unnecessary.
628 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
632 /* We always send an MSS option. */
634 remaining
-= TCPOLEN_MSS_ALIGNED
;
636 if (likely(ireq
->wscale_ok
)) {
637 opts
->ws
= ireq
->rcv_wscale
;
638 opts
->options
|= OPTION_WSCALE
;
639 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
641 if (likely(ireq
->tstamp_ok
)) {
642 opts
->options
|= OPTION_TS
;
643 opts
->tsval
= tcp_skb_timestamp(skb
) + tcp_rsk(req
)->ts_off
;
644 opts
->tsecr
= req
->ts_recent
;
645 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
647 if (likely(ireq
->sack_ok
)) {
648 opts
->options
|= OPTION_SACK_ADVERTISE
;
649 if (unlikely(!ireq
->tstamp_ok
))
650 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
652 if (foc
!= NULL
&& foc
->len
>= 0) {
655 need
+= foc
->exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
656 TCPOLEN_FASTOPEN_BASE
;
657 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
658 if (remaining
>= need
) {
659 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
660 opts
->fastopen_cookie
= foc
;
665 return MAX_TCP_OPTION_SPACE
- remaining
;
668 /* Compute TCP options for ESTABLISHED sockets. This is not the
669 * final wire format yet.
671 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
672 struct tcp_out_options
*opts
,
673 struct tcp_md5sig_key
**md5
)
675 struct tcp_sock
*tp
= tcp_sk(sk
);
676 unsigned int size
= 0;
677 unsigned int eff_sacks
;
681 #ifdef CONFIG_TCP_MD5SIG
682 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
683 if (unlikely(*md5
)) {
684 opts
->options
|= OPTION_MD5
;
685 size
+= TCPOLEN_MD5SIG_ALIGNED
;
691 if (likely(tp
->rx_opt
.tstamp_ok
)) {
692 opts
->options
|= OPTION_TS
;
693 opts
->tsval
= skb
? tcp_skb_timestamp(skb
) + tp
->tsoffset
: 0;
694 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
695 size
+= TCPOLEN_TSTAMP_ALIGNED
;
698 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
699 if (unlikely(eff_sacks
)) {
700 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
701 opts
->num_sack_blocks
=
702 min_t(unsigned int, eff_sacks
,
703 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
704 TCPOLEN_SACK_PERBLOCK
);
705 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
706 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
713 /* TCP SMALL QUEUES (TSQ)
715 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
716 * to reduce RTT and bufferbloat.
717 * We do this using a special skb destructor (tcp_wfree).
719 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
720 * needs to be reallocated in a driver.
721 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
723 * Since transmit from skb destructor is forbidden, we use a tasklet
724 * to process all sockets that eventually need to send more skbs.
725 * We use one tasklet per cpu, with its own queue of sockets.
728 struct tasklet_struct tasklet
;
729 struct list_head head
; /* queue of tcp sockets */
731 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
733 static void tcp_tsq_handler(struct sock
*sk
)
735 if ((1 << sk
->sk_state
) &
736 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
737 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
)) {
738 struct tcp_sock
*tp
= tcp_sk(sk
);
740 if (tp
->lost_out
> tp
->retrans_out
&&
741 tp
->snd_cwnd
> tcp_packets_in_flight(tp
))
742 tcp_xmit_retransmit_queue(sk
);
744 tcp_write_xmit(sk
, tcp_current_mss(sk
), tp
->nonagle
,
749 * One tasklet per cpu tries to send more skbs.
750 * We run in tasklet context but need to disable irqs when
751 * transferring tsq->head because tcp_wfree() might
752 * interrupt us (non NAPI drivers)
754 static void tcp_tasklet_func(unsigned long data
)
756 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
759 struct list_head
*q
, *n
;
763 local_irq_save(flags
);
764 list_splice_init(&tsq
->head
, &list
);
765 local_irq_restore(flags
);
767 list_for_each_safe(q
, n
, &list
) {
768 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
769 list_del(&tp
->tsq_node
);
771 sk
= (struct sock
*)tp
;
772 smp_mb__before_atomic();
773 clear_bit(TSQ_QUEUED
, &sk
->sk_tsq_flags
);
775 if (!sk
->sk_lock
.owned
&&
776 test_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
)) {
778 if (!sock_owned_by_user(sk
)) {
779 clear_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
);
789 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
790 TCPF_WRITE_TIMER_DEFERRED | \
791 TCPF_DELACK_TIMER_DEFERRED | \
792 TCPF_MTU_REDUCED_DEFERRED)
794 * tcp_release_cb - tcp release_sock() callback
797 * called from release_sock() to perform protocol dependent
798 * actions before socket release.
800 void tcp_release_cb(struct sock
*sk
)
802 unsigned long flags
, nflags
;
804 /* perform an atomic operation only if at least one flag is set */
806 flags
= sk
->sk_tsq_flags
;
807 if (!(flags
& TCP_DEFERRED_ALL
))
809 nflags
= flags
& ~TCP_DEFERRED_ALL
;
810 } while (cmpxchg(&sk
->sk_tsq_flags
, flags
, nflags
) != flags
);
812 if (flags
& TCPF_TSQ_DEFERRED
)
815 /* Here begins the tricky part :
816 * We are called from release_sock() with :
818 * 2) sk_lock.slock spinlock held
819 * 3) socket owned by us (sk->sk_lock.owned == 1)
821 * But following code is meant to be called from BH handlers,
822 * so we should keep BH disabled, but early release socket ownership
824 sock_release_ownership(sk
);
826 if (flags
& TCPF_WRITE_TIMER_DEFERRED
) {
827 tcp_write_timer_handler(sk
);
830 if (flags
& TCPF_DELACK_TIMER_DEFERRED
) {
831 tcp_delack_timer_handler(sk
);
834 if (flags
& TCPF_MTU_REDUCED_DEFERRED
) {
835 inet_csk(sk
)->icsk_af_ops
->mtu_reduced(sk
);
839 EXPORT_SYMBOL(tcp_release_cb
);
841 void __init
tcp_tasklet_init(void)
845 for_each_possible_cpu(i
) {
846 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
848 INIT_LIST_HEAD(&tsq
->head
);
849 tasklet_init(&tsq
->tasklet
,
856 * Write buffer destructor automatically called from kfree_skb.
857 * We can't xmit new skbs from this context, as we might already
860 void tcp_wfree(struct sk_buff
*skb
)
862 struct sock
*sk
= skb
->sk
;
863 struct tcp_sock
*tp
= tcp_sk(sk
);
864 unsigned long flags
, nval
, oval
;
867 /* Keep one reference on sk_wmem_alloc.
868 * Will be released by sk_free() from here or tcp_tasklet_func()
870 wmem
= atomic_sub_return(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
872 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
873 * Wait until our queues (qdisc + devices) are drained.
875 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
876 * - chance for incoming ACK (processed by another cpu maybe)
877 * to migrate this flow (skb->ooo_okay will be eventually set)
879 if (wmem
>= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current
)
882 for (oval
= READ_ONCE(sk
->sk_tsq_flags
);; oval
= nval
) {
883 struct tsq_tasklet
*tsq
;
886 if (!(oval
& TSQF_THROTTLED
) || (oval
& TSQF_QUEUED
))
889 nval
= (oval
& ~TSQF_THROTTLED
) | TSQF_QUEUED
| TCPF_TSQ_DEFERRED
;
890 nval
= cmpxchg(&sk
->sk_tsq_flags
, oval
, nval
);
894 /* queue this socket to tasklet queue */
895 local_irq_save(flags
);
896 tsq
= this_cpu_ptr(&tsq_tasklet
);
897 empty
= list_empty(&tsq
->head
);
898 list_add(&tp
->tsq_node
, &tsq
->head
);
900 tasklet_schedule(&tsq
->tasklet
);
901 local_irq_restore(flags
);
908 /* This routine actually transmits TCP packets queued in by
909 * tcp_do_sendmsg(). This is used by both the initial
910 * transmission and possible later retransmissions.
911 * All SKB's seen here are completely headerless. It is our
912 * job to build the TCP header, and pass the packet down to
913 * IP so it can do the same plus pass the packet off to the
916 * We are working here with either a clone of the original
917 * SKB, or a fresh unique copy made by the retransmit engine.
919 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
922 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
923 struct inet_sock
*inet
;
925 struct tcp_skb_cb
*tcb
;
926 struct tcp_out_options opts
;
927 unsigned int tcp_options_size
, tcp_header_size
;
928 struct tcp_md5sig_key
*md5
;
932 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
936 skb_mstamp_get(&skb
->skb_mstamp
);
937 TCP_SKB_CB(skb
)->tx
.in_flight
= TCP_SKB_CB(skb
)->end_seq
939 tcp_rate_skb_sent(sk
, skb
);
941 if (unlikely(skb_cloned(skb
)))
942 skb
= pskb_copy(skb
, gfp_mask
);
944 skb
= skb_clone(skb
, gfp_mask
);
950 tcb
= TCP_SKB_CB(skb
);
951 memset(&opts
, 0, sizeof(opts
));
953 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
954 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
956 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
958 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
960 /* if no packet is in qdisc/device queue, then allow XPS to select
961 * another queue. We can be called from tcp_tsq_handler()
962 * which holds one reference to sk_wmem_alloc.
964 * TODO: Ideally, in-flight pure ACK packets should not matter here.
965 * One way to get this would be to set skb->truesize = 2 on them.
967 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) < SKB_TRUESIZE(1);
969 skb_push(skb
, tcp_header_size
);
970 skb_reset_transport_header(skb
);
974 skb
->destructor
= skb_is_tcp_pure_ack(skb
) ? __sock_wfree
: tcp_wfree
;
975 skb_set_hash_from_sk(skb
, sk
);
976 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
978 /* Build TCP header and checksum it. */
979 th
= (struct tcphdr
*)skb
->data
;
980 th
->source
= inet
->inet_sport
;
981 th
->dest
= inet
->inet_dport
;
982 th
->seq
= htonl(tcb
->seq
);
983 th
->ack_seq
= htonl(tp
->rcv_nxt
);
984 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
990 /* The urg_mode check is necessary during a below snd_una win probe */
991 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
992 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
993 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
995 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
996 th
->urg_ptr
= htons(0xFFFF);
1001 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
1002 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
1003 if (likely(!(tcb
->tcp_flags
& TCPHDR_SYN
))) {
1004 th
->window
= htons(tcp_select_window(sk
));
1005 tcp_ecn_send(sk
, skb
, th
, tcp_header_size
);
1007 /* RFC1323: The window in SYN & SYN/ACK segments
1010 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
1012 #ifdef CONFIG_TCP_MD5SIG
1013 /* Calculate the MD5 hash, as we have all we need now */
1015 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
1016 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
1021 icsk
->icsk_af_ops
->send_check(sk
, skb
);
1023 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
1024 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
1026 if (skb
->len
!= tcp_header_size
) {
1027 tcp_event_data_sent(tp
, sk
);
1028 tp
->data_segs_out
+= tcp_skb_pcount(skb
);
1031 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
1032 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
1033 tcp_skb_pcount(skb
));
1035 tp
->segs_out
+= tcp_skb_pcount(skb
);
1036 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1037 skb_shinfo(skb
)->gso_segs
= tcp_skb_pcount(skb
);
1038 skb_shinfo(skb
)->gso_size
= tcp_skb_mss(skb
);
1040 /* Our usage of tstamp should remain private */
1043 /* Cleanup our debris for IP stacks */
1044 memset(skb
->cb
, 0, max(sizeof(struct inet_skb_parm
),
1045 sizeof(struct inet6_skb_parm
)));
1047 err
= icsk
->icsk_af_ops
->queue_xmit(sk
, skb
, &inet
->cork
.fl
);
1049 if (likely(err
<= 0))
1054 return net_xmit_eval(err
);
1057 /* This routine just queues the buffer for sending.
1059 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1060 * otherwise socket can stall.
1062 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1064 struct tcp_sock
*tp
= tcp_sk(sk
);
1066 /* Advance write_seq and place onto the write_queue. */
1067 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
1068 __skb_header_release(skb
);
1069 tcp_add_write_queue_tail(sk
, skb
);
1070 sk
->sk_wmem_queued
+= skb
->truesize
;
1071 sk_mem_charge(sk
, skb
->truesize
);
1074 /* Initialize TSO segments for a packet. */
1075 static void tcp_set_skb_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1077 if (skb
->len
<= mss_now
|| skb
->ip_summed
== CHECKSUM_NONE
) {
1078 /* Avoid the costly divide in the normal
1081 tcp_skb_pcount_set(skb
, 1);
1082 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1084 tcp_skb_pcount_set(skb
, DIV_ROUND_UP(skb
->len
, mss_now
));
1085 TCP_SKB_CB(skb
)->tcp_gso_size
= mss_now
;
1089 /* When a modification to fackets out becomes necessary, we need to check
1090 * skb is counted to fackets_out or not.
1092 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1095 struct tcp_sock
*tp
= tcp_sk(sk
);
1097 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1100 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1101 tp
->fackets_out
-= decr
;
1104 /* Pcount in the middle of the write queue got changed, we need to do various
1105 * tweaks to fix counters
1107 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1109 struct tcp_sock
*tp
= tcp_sk(sk
);
1111 tp
->packets_out
-= decr
;
1113 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1114 tp
->sacked_out
-= decr
;
1115 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1116 tp
->retrans_out
-= decr
;
1117 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1118 tp
->lost_out
-= decr
;
1120 /* Reno case is special. Sigh... */
1121 if (tcp_is_reno(tp
) && decr
> 0)
1122 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1124 tcp_adjust_fackets_out(sk
, skb
, decr
);
1126 if (tp
->lost_skb_hint
&&
1127 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1128 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1129 tp
->lost_cnt_hint
-= decr
;
1131 tcp_verify_left_out(tp
);
1134 static bool tcp_has_tx_tstamp(const struct sk_buff
*skb
)
1136 return TCP_SKB_CB(skb
)->txstamp_ack
||
1137 (skb_shinfo(skb
)->tx_flags
& SKBTX_ANY_TSTAMP
);
1140 static void tcp_fragment_tstamp(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1142 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1144 if (unlikely(tcp_has_tx_tstamp(skb
)) &&
1145 !before(shinfo
->tskey
, TCP_SKB_CB(skb2
)->seq
)) {
1146 struct skb_shared_info
*shinfo2
= skb_shinfo(skb2
);
1147 u8 tsflags
= shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
1149 shinfo
->tx_flags
&= ~tsflags
;
1150 shinfo2
->tx_flags
|= tsflags
;
1151 swap(shinfo
->tskey
, shinfo2
->tskey
);
1152 TCP_SKB_CB(skb2
)->txstamp_ack
= TCP_SKB_CB(skb
)->txstamp_ack
;
1153 TCP_SKB_CB(skb
)->txstamp_ack
= 0;
1157 static void tcp_skb_fragment_eor(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1159 TCP_SKB_CB(skb2
)->eor
= TCP_SKB_CB(skb
)->eor
;
1160 TCP_SKB_CB(skb
)->eor
= 0;
1163 /* Function to create two new TCP segments. Shrinks the given segment
1164 * to the specified size and appends a new segment with the rest of the
1165 * packet to the list. This won't be called frequently, I hope.
1166 * Remember, these are still headerless SKBs at this point.
1168 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1169 unsigned int mss_now
, gfp_t gfp
)
1171 struct tcp_sock
*tp
= tcp_sk(sk
);
1172 struct sk_buff
*buff
;
1173 int nsize
, old_factor
;
1177 if (WARN_ON(len
> skb
->len
))
1180 nsize
= skb_headlen(skb
) - len
;
1184 if (skb_unclone(skb
, gfp
))
1187 /* Get a new skb... force flag on. */
1188 buff
= sk_stream_alloc_skb(sk
, nsize
, gfp
, true);
1190 return -ENOMEM
; /* We'll just try again later. */
1192 sk
->sk_wmem_queued
+= buff
->truesize
;
1193 sk_mem_charge(sk
, buff
->truesize
);
1194 nlen
= skb
->len
- len
- nsize
;
1195 buff
->truesize
+= nlen
;
1196 skb
->truesize
-= nlen
;
1198 /* Correct the sequence numbers. */
1199 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1200 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1201 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1203 /* PSH and FIN should only be set in the second packet. */
1204 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1205 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1206 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1207 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1208 tcp_skb_fragment_eor(skb
, buff
);
1210 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1211 /* Copy and checksum data tail into the new buffer. */
1212 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1213 skb_put(buff
, nsize
),
1218 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1220 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1221 skb_split(skb
, buff
, len
);
1224 buff
->ip_summed
= skb
->ip_summed
;
1226 buff
->tstamp
= skb
->tstamp
;
1227 tcp_fragment_tstamp(skb
, buff
);
1229 old_factor
= tcp_skb_pcount(skb
);
1231 /* Fix up tso_factor for both original and new SKB. */
1232 tcp_set_skb_tso_segs(skb
, mss_now
);
1233 tcp_set_skb_tso_segs(buff
, mss_now
);
1235 /* Update delivered info for the new segment */
1236 TCP_SKB_CB(buff
)->tx
= TCP_SKB_CB(skb
)->tx
;
1238 /* If this packet has been sent out already, we must
1239 * adjust the various packet counters.
1241 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1242 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1243 tcp_skb_pcount(buff
);
1246 tcp_adjust_pcount(sk
, skb
, diff
);
1249 /* Link BUFF into the send queue. */
1250 __skb_header_release(buff
);
1251 tcp_insert_write_queue_after(skb
, buff
, sk
);
1256 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1257 * eventually). The difference is that pulled data not copied, but
1258 * immediately discarded.
1260 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1262 struct skb_shared_info
*shinfo
;
1265 eat
= min_t(int, len
, skb_headlen(skb
));
1267 __skb_pull(skb
, eat
);
1274 shinfo
= skb_shinfo(skb
);
1275 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1276 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1279 skb_frag_unref(skb
, i
);
1282 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1284 shinfo
->frags
[k
].page_offset
+= eat
;
1285 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1291 shinfo
->nr_frags
= k
;
1293 skb_reset_tail_pointer(skb
);
1294 skb
->data_len
-= len
;
1295 skb
->len
= skb
->data_len
;
1298 /* Remove acked data from a packet in the transmit queue. */
1299 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1301 if (skb_unclone(skb
, GFP_ATOMIC
))
1304 __pskb_trim_head(skb
, len
);
1306 TCP_SKB_CB(skb
)->seq
+= len
;
1307 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1309 skb
->truesize
-= len
;
1310 sk
->sk_wmem_queued
-= len
;
1311 sk_mem_uncharge(sk
, len
);
1312 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1314 /* Any change of skb->len requires recalculation of tso factor. */
1315 if (tcp_skb_pcount(skb
) > 1)
1316 tcp_set_skb_tso_segs(skb
, tcp_skb_mss(skb
));
1321 /* Calculate MSS not accounting any TCP options. */
1322 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1324 const struct tcp_sock
*tp
= tcp_sk(sk
);
1325 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1328 /* Calculate base mss without TCP options:
1329 It is MMS_S - sizeof(tcphdr) of rfc1122
1331 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1333 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1334 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1335 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1337 if (dst
&& dst_allfrag(dst
))
1338 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1341 /* Clamp it (mss_clamp does not include tcp options) */
1342 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1343 mss_now
= tp
->rx_opt
.mss_clamp
;
1345 /* Now subtract optional transport overhead */
1346 mss_now
-= icsk
->icsk_ext_hdr_len
;
1348 /* Then reserve room for full set of TCP options and 8 bytes of data */
1354 /* Calculate MSS. Not accounting for SACKs here. */
1355 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1357 /* Subtract TCP options size, not including SACKs */
1358 return __tcp_mtu_to_mss(sk
, pmtu
) -
1359 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1362 /* Inverse of above */
1363 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1365 const struct tcp_sock
*tp
= tcp_sk(sk
);
1366 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1370 tp
->tcp_header_len
+
1371 icsk
->icsk_ext_hdr_len
+
1372 icsk
->icsk_af_ops
->net_header_len
;
1374 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1375 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1376 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1378 if (dst
&& dst_allfrag(dst
))
1379 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1383 EXPORT_SYMBOL(tcp_mss_to_mtu
);
1385 /* MTU probing init per socket */
1386 void tcp_mtup_init(struct sock
*sk
)
1388 struct tcp_sock
*tp
= tcp_sk(sk
);
1389 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1390 struct net
*net
= sock_net(sk
);
1392 icsk
->icsk_mtup
.enabled
= net
->ipv4
.sysctl_tcp_mtu_probing
> 1;
1393 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1394 icsk
->icsk_af_ops
->net_header_len
;
1395 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, net
->ipv4
.sysctl_tcp_base_mss
);
1396 icsk
->icsk_mtup
.probe_size
= 0;
1397 if (icsk
->icsk_mtup
.enabled
)
1398 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1400 EXPORT_SYMBOL(tcp_mtup_init
);
1402 /* This function synchronize snd mss to current pmtu/exthdr set.
1404 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1405 for TCP options, but includes only bare TCP header.
1407 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1408 It is minimum of user_mss and mss received with SYN.
1409 It also does not include TCP options.
1411 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1413 tp->mss_cache is current effective sending mss, including
1414 all tcp options except for SACKs. It is evaluated,
1415 taking into account current pmtu, but never exceeds
1416 tp->rx_opt.mss_clamp.
1418 NOTE1. rfc1122 clearly states that advertised MSS
1419 DOES NOT include either tcp or ip options.
1421 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1422 are READ ONLY outside this function. --ANK (980731)
1424 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1426 struct tcp_sock
*tp
= tcp_sk(sk
);
1427 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1430 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1431 icsk
->icsk_mtup
.search_high
= pmtu
;
1433 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1434 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1436 /* And store cached results */
1437 icsk
->icsk_pmtu_cookie
= pmtu
;
1438 if (icsk
->icsk_mtup
.enabled
)
1439 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1440 tp
->mss_cache
= mss_now
;
1444 EXPORT_SYMBOL(tcp_sync_mss
);
1446 /* Compute the current effective MSS, taking SACKs and IP options,
1447 * and even PMTU discovery events into account.
1449 unsigned int tcp_current_mss(struct sock
*sk
)
1451 const struct tcp_sock
*tp
= tcp_sk(sk
);
1452 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1454 unsigned int header_len
;
1455 struct tcp_out_options opts
;
1456 struct tcp_md5sig_key
*md5
;
1458 mss_now
= tp
->mss_cache
;
1461 u32 mtu
= dst_mtu(dst
);
1462 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1463 mss_now
= tcp_sync_mss(sk
, mtu
);
1466 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1467 sizeof(struct tcphdr
);
1468 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1469 * some common options. If this is an odd packet (because we have SACK
1470 * blocks etc) then our calculated header_len will be different, and
1471 * we have to adjust mss_now correspondingly */
1472 if (header_len
!= tp
->tcp_header_len
) {
1473 int delta
= (int) header_len
- tp
->tcp_header_len
;
1480 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1481 * As additional protections, we do not touch cwnd in retransmission phases,
1482 * and if application hit its sndbuf limit recently.
1484 static void tcp_cwnd_application_limited(struct sock
*sk
)
1486 struct tcp_sock
*tp
= tcp_sk(sk
);
1488 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1489 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1490 /* Limited by application or receiver window. */
1491 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1492 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1493 if (win_used
< tp
->snd_cwnd
) {
1494 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1495 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
1497 tp
->snd_cwnd_used
= 0;
1499 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1502 static void tcp_cwnd_validate(struct sock
*sk
, bool is_cwnd_limited
)
1504 struct tcp_sock
*tp
= tcp_sk(sk
);
1506 /* Track the maximum number of outstanding packets in each
1507 * window, and remember whether we were cwnd-limited then.
1509 if (!before(tp
->snd_una
, tp
->max_packets_seq
) ||
1510 tp
->packets_out
> tp
->max_packets_out
) {
1511 tp
->max_packets_out
= tp
->packets_out
;
1512 tp
->max_packets_seq
= tp
->snd_nxt
;
1513 tp
->is_cwnd_limited
= is_cwnd_limited
;
1516 if (tcp_is_cwnd_limited(sk
)) {
1517 /* Network is feed fully. */
1518 tp
->snd_cwnd_used
= 0;
1519 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1521 /* Network starves. */
1522 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1523 tp
->snd_cwnd_used
= tp
->packets_out
;
1525 if (sysctl_tcp_slow_start_after_idle
&&
1526 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1527 tcp_cwnd_application_limited(sk
);
1529 /* The following conditions together indicate the starvation
1530 * is caused by insufficient sender buffer:
1531 * 1) just sent some data (see tcp_write_xmit)
1532 * 2) not cwnd limited (this else condition)
1533 * 3) no more data to send (null tcp_send_head )
1534 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1536 if (!tcp_send_head(sk
) && sk
->sk_socket
&&
1537 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
) &&
1538 (1 << sk
->sk_state
) & (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
1539 tcp_chrono_start(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1543 /* Minshall's variant of the Nagle send check. */
1544 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1546 return after(tp
->snd_sml
, tp
->snd_una
) &&
1547 !after(tp
->snd_sml
, tp
->snd_nxt
);
1550 /* Update snd_sml if this skb is under mss
1551 * Note that a TSO packet might end with a sub-mss segment
1552 * The test is really :
1553 * if ((skb->len % mss) != 0)
1554 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1555 * But we can avoid doing the divide again given we already have
1556 * skb_pcount = skb->len / mss_now
1558 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1559 const struct sk_buff
*skb
)
1561 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1562 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1565 /* Return false, if packet can be sent now without violation Nagle's rules:
1566 * 1. It is full sized. (provided by caller in %partial bool)
1567 * 2. Or it contains FIN. (already checked by caller)
1568 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1569 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1570 * With Minshall's modification: all sent small packets are ACKed.
1572 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1576 ((nonagle
& TCP_NAGLE_CORK
) ||
1577 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1580 /* Return how many segs we'd like on a TSO packet,
1581 * to send one TSO packet per ms
1583 u32
tcp_tso_autosize(const struct sock
*sk
, unsigned int mss_now
,
1588 bytes
= min(sk
->sk_pacing_rate
>> 10,
1589 sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
);
1591 /* Goal is to send at least one packet per ms,
1592 * not one big TSO packet every 100 ms.
1593 * This preserves ACK clocking and is consistent
1594 * with tcp_tso_should_defer() heuristic.
1596 segs
= max_t(u32
, bytes
/ mss_now
, min_tso_segs
);
1598 return min_t(u32
, segs
, sk
->sk_gso_max_segs
);
1600 EXPORT_SYMBOL(tcp_tso_autosize
);
1602 /* Return the number of segments we want in the skb we are transmitting.
1603 * See if congestion control module wants to decide; otherwise, autosize.
1605 static u32
tcp_tso_segs(struct sock
*sk
, unsigned int mss_now
)
1607 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1608 u32 tso_segs
= ca_ops
->tso_segs_goal
? ca_ops
->tso_segs_goal(sk
) : 0;
1611 tcp_tso_autosize(sk
, mss_now
, sysctl_tcp_min_tso_segs
);
1614 /* Returns the portion of skb which can be sent right away */
1615 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1616 const struct sk_buff
*skb
,
1617 unsigned int mss_now
,
1618 unsigned int max_segs
,
1621 const struct tcp_sock
*tp
= tcp_sk(sk
);
1622 u32 partial
, needed
, window
, max_len
;
1624 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1625 max_len
= mss_now
* max_segs
;
1627 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1630 needed
= min(skb
->len
, window
);
1632 if (max_len
<= needed
)
1635 partial
= needed
% mss_now
;
1636 /* If last segment is not a full MSS, check if Nagle rules allow us
1637 * to include this last segment in this skb.
1638 * Otherwise, we'll split the skb at last MSS boundary
1640 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
1641 return needed
- partial
;
1646 /* Can at least one segment of SKB be sent right now, according to the
1647 * congestion window rules? If so, return how many segments are allowed.
1649 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1650 const struct sk_buff
*skb
)
1652 u32 in_flight
, cwnd
, halfcwnd
;
1654 /* Don't be strict about the congestion window for the final FIN. */
1655 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1656 tcp_skb_pcount(skb
) == 1)
1659 in_flight
= tcp_packets_in_flight(tp
);
1660 cwnd
= tp
->snd_cwnd
;
1661 if (in_flight
>= cwnd
)
1664 /* For better scheduling, ensure we have at least
1665 * 2 GSO packets in flight.
1667 halfcwnd
= max(cwnd
>> 1, 1U);
1668 return min(halfcwnd
, cwnd
- in_flight
);
1671 /* Initialize TSO state of a skb.
1672 * This must be invoked the first time we consider transmitting
1673 * SKB onto the wire.
1675 static int tcp_init_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1677 int tso_segs
= tcp_skb_pcount(skb
);
1679 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1680 tcp_set_skb_tso_segs(skb
, mss_now
);
1681 tso_segs
= tcp_skb_pcount(skb
);
1687 /* Return true if the Nagle test allows this packet to be
1690 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1691 unsigned int cur_mss
, int nonagle
)
1693 /* Nagle rule does not apply to frames, which sit in the middle of the
1694 * write_queue (they have no chances to get new data).
1696 * This is implemented in the callers, where they modify the 'nonagle'
1697 * argument based upon the location of SKB in the send queue.
1699 if (nonagle
& TCP_NAGLE_PUSH
)
1702 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1703 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1706 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
1712 /* Does at least the first segment of SKB fit into the send window? */
1713 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1714 const struct sk_buff
*skb
,
1715 unsigned int cur_mss
)
1717 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1719 if (skb
->len
> cur_mss
)
1720 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1722 return !after(end_seq
, tcp_wnd_end(tp
));
1725 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1726 * should be put on the wire right now. If so, it returns the number of
1727 * packets allowed by the congestion window.
1729 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1730 unsigned int cur_mss
, int nonagle
)
1732 const struct tcp_sock
*tp
= tcp_sk(sk
);
1733 unsigned int cwnd_quota
;
1735 tcp_init_tso_segs(skb
, cur_mss
);
1737 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1740 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1741 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1747 /* Test if sending is allowed right now. */
1748 bool tcp_may_send_now(struct sock
*sk
)
1750 const struct tcp_sock
*tp
= tcp_sk(sk
);
1751 struct sk_buff
*skb
= tcp_send_head(sk
);
1754 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1755 (tcp_skb_is_last(sk
, skb
) ?
1756 tp
->nonagle
: TCP_NAGLE_PUSH
));
1759 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1760 * which is put after SKB on the list. It is very much like
1761 * tcp_fragment() except that it may make several kinds of assumptions
1762 * in order to speed up the splitting operation. In particular, we
1763 * know that all the data is in scatter-gather pages, and that the
1764 * packet has never been sent out before (and thus is not cloned).
1766 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1767 unsigned int mss_now
, gfp_t gfp
)
1769 struct sk_buff
*buff
;
1770 int nlen
= skb
->len
- len
;
1773 /* All of a TSO frame must be composed of paged data. */
1774 if (skb
->len
!= skb
->data_len
)
1775 return tcp_fragment(sk
, skb
, len
, mss_now
, gfp
);
1777 buff
= sk_stream_alloc_skb(sk
, 0, gfp
, true);
1778 if (unlikely(!buff
))
1781 sk
->sk_wmem_queued
+= buff
->truesize
;
1782 sk_mem_charge(sk
, buff
->truesize
);
1783 buff
->truesize
+= nlen
;
1784 skb
->truesize
-= nlen
;
1786 /* Correct the sequence numbers. */
1787 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1788 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1789 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1791 /* PSH and FIN should only be set in the second packet. */
1792 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1793 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1794 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1796 /* This packet was never sent out yet, so no SACK bits. */
1797 TCP_SKB_CB(buff
)->sacked
= 0;
1799 tcp_skb_fragment_eor(skb
, buff
);
1801 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1802 skb_split(skb
, buff
, len
);
1803 tcp_fragment_tstamp(skb
, buff
);
1805 /* Fix up tso_factor for both original and new SKB. */
1806 tcp_set_skb_tso_segs(skb
, mss_now
);
1807 tcp_set_skb_tso_segs(buff
, mss_now
);
1809 /* Link BUFF into the send queue. */
1810 __skb_header_release(buff
);
1811 tcp_insert_write_queue_after(skb
, buff
, sk
);
1816 /* Try to defer sending, if possible, in order to minimize the amount
1817 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1819 * This algorithm is from John Heffner.
1821 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
,
1822 bool *is_cwnd_limited
, u32 max_segs
)
1824 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1825 u32 age
, send_win
, cong_win
, limit
, in_flight
;
1826 struct tcp_sock
*tp
= tcp_sk(sk
);
1827 struct skb_mstamp now
;
1828 struct sk_buff
*head
;
1831 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1834 if (icsk
->icsk_ca_state
>= TCP_CA_Recovery
)
1837 /* Avoid bursty behavior by allowing defer
1838 * only if the last write was recent.
1840 if ((s32
)(tcp_time_stamp
- tp
->lsndtime
) > 0)
1843 in_flight
= tcp_packets_in_flight(tp
);
1845 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1847 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1849 /* From in_flight test above, we know that cwnd > in_flight. */
1850 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1852 limit
= min(send_win
, cong_win
);
1854 /* If a full-sized TSO skb can be sent, do it. */
1855 if (limit
>= max_segs
* tp
->mss_cache
)
1858 /* Middle in queue won't get any more data, full sendable already? */
1859 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1862 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1864 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1866 /* If at least some fraction of a window is available,
1869 chunk
/= win_divisor
;
1873 /* Different approach, try not to defer past a single
1874 * ACK. Receiver should ACK every other full sized
1875 * frame, so if we have space for more than 3 frames
1878 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1882 head
= tcp_write_queue_head(sk
);
1883 skb_mstamp_get(&now
);
1884 age
= skb_mstamp_us_delta(&now
, &head
->skb_mstamp
);
1885 /* If next ACK is likely to come too late (half srtt), do not defer */
1886 if (age
< (tp
->srtt_us
>> 4))
1889 /* Ok, it looks like it is advisable to defer. */
1891 if (cong_win
< send_win
&& cong_win
<= skb
->len
)
1892 *is_cwnd_limited
= true;
1900 static inline void tcp_mtu_check_reprobe(struct sock
*sk
)
1902 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1903 struct tcp_sock
*tp
= tcp_sk(sk
);
1904 struct net
*net
= sock_net(sk
);
1908 interval
= net
->ipv4
.sysctl_tcp_probe_interval
;
1909 delta
= tcp_time_stamp
- icsk
->icsk_mtup
.probe_timestamp
;
1910 if (unlikely(delta
>= interval
* HZ
)) {
1911 int mss
= tcp_current_mss(sk
);
1913 /* Update current search range */
1914 icsk
->icsk_mtup
.probe_size
= 0;
1915 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+
1916 sizeof(struct tcphdr
) +
1917 icsk
->icsk_af_ops
->net_header_len
;
1918 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
1920 /* Update probe time stamp */
1921 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1925 /* Create a new MTU probe if we are ready.
1926 * MTU probe is regularly attempting to increase the path MTU by
1927 * deliberately sending larger packets. This discovers routing
1928 * changes resulting in larger path MTUs.
1930 * Returns 0 if we should wait to probe (no cwnd available),
1931 * 1 if a probe was sent,
1934 static int tcp_mtu_probe(struct sock
*sk
)
1936 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1937 struct tcp_sock
*tp
= tcp_sk(sk
);
1938 struct sk_buff
*skb
, *nskb
, *next
;
1939 struct net
*net
= sock_net(sk
);
1946 /* Not currently probing/verifying,
1948 * have enough cwnd, and
1949 * not SACKing (the variable headers throw things off)
1951 if (likely(!icsk
->icsk_mtup
.enabled
||
1952 icsk
->icsk_mtup
.probe_size
||
1953 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1954 tp
->snd_cwnd
< 11 ||
1955 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
))
1958 /* Use binary search for probe_size between tcp_mss_base,
1959 * and current mss_clamp. if (search_high - search_low)
1960 * smaller than a threshold, backoff from probing.
1962 mss_now
= tcp_current_mss(sk
);
1963 probe_size
= tcp_mtu_to_mss(sk
, (icsk
->icsk_mtup
.search_high
+
1964 icsk
->icsk_mtup
.search_low
) >> 1);
1965 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1966 interval
= icsk
->icsk_mtup
.search_high
- icsk
->icsk_mtup
.search_low
;
1967 /* When misfortune happens, we are reprobing actively,
1968 * and then reprobe timer has expired. We stick with current
1969 * probing process by not resetting search range to its orignal.
1971 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
) ||
1972 interval
< net
->ipv4
.sysctl_tcp_probe_threshold
) {
1973 /* Check whether enough time has elaplased for
1974 * another round of probing.
1976 tcp_mtu_check_reprobe(sk
);
1980 /* Have enough data in the send queue to probe? */
1981 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1984 if (tp
->snd_wnd
< size_needed
)
1986 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1989 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1990 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1991 if (!tcp_packets_in_flight(tp
))
1997 /* We're allowed to probe. Build it now. */
1998 nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
, false);
2001 sk
->sk_wmem_queued
+= nskb
->truesize
;
2002 sk_mem_charge(sk
, nskb
->truesize
);
2004 skb
= tcp_send_head(sk
);
2006 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
2007 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
2008 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
2009 TCP_SKB_CB(nskb
)->sacked
= 0;
2011 nskb
->ip_summed
= skb
->ip_summed
;
2013 tcp_insert_write_queue_before(nskb
, skb
, sk
);
2016 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
2017 copy
= min_t(int, skb
->len
, probe_size
- len
);
2018 if (nskb
->ip_summed
) {
2019 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
2021 __wsum csum
= skb_copy_and_csum_bits(skb
, 0,
2022 skb_put(nskb
, copy
),
2024 nskb
->csum
= csum_block_add(nskb
->csum
, csum
, len
);
2027 if (skb
->len
<= copy
) {
2028 /* We've eaten all the data from this skb.
2030 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
2031 tcp_unlink_write_queue(skb
, sk
);
2032 sk_wmem_free_skb(sk
, skb
);
2034 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
2035 ~(TCPHDR_FIN
|TCPHDR_PSH
);
2036 if (!skb_shinfo(skb
)->nr_frags
) {
2037 skb_pull(skb
, copy
);
2038 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2039 skb
->csum
= csum_partial(skb
->data
,
2042 __pskb_trim_head(skb
, copy
);
2043 tcp_set_skb_tso_segs(skb
, mss_now
);
2045 TCP_SKB_CB(skb
)->seq
+= copy
;
2050 if (len
>= probe_size
)
2053 tcp_init_tso_segs(nskb
, nskb
->len
);
2055 /* We're ready to send. If this fails, the probe will
2056 * be resegmented into mss-sized pieces by tcp_write_xmit().
2058 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
2059 /* Decrement cwnd here because we are sending
2060 * effectively two packets. */
2062 tcp_event_new_data_sent(sk
, nskb
);
2064 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
2065 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
2066 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
2074 /* TCP Small Queues :
2075 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2076 * (These limits are doubled for retransmits)
2078 * - better RTT estimation and ACK scheduling
2081 * Alas, some drivers / subsystems require a fair amount
2082 * of queued bytes to ensure line rate.
2083 * One example is wifi aggregation (802.11 AMPDU)
2085 static bool tcp_small_queue_check(struct sock
*sk
, const struct sk_buff
*skb
,
2086 unsigned int factor
)
2090 limit
= max(2 * skb
->truesize
, sk
->sk_pacing_rate
>> 10);
2091 limit
= min_t(u32
, limit
, sysctl_tcp_limit_output_bytes
);
2094 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
) {
2095 /* Always send the 1st or 2nd skb in write queue.
2096 * No need to wait for TX completion to call us back,
2097 * after softirq/tasklet schedule.
2098 * This helps when TX completions are delayed too much.
2100 if (skb
== sk
->sk_write_queue
.next
||
2101 skb
->prev
== sk
->sk_write_queue
.next
)
2104 set_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
);
2105 /* It is possible TX completion already happened
2106 * before we set TSQ_THROTTLED, so we must
2107 * test again the condition.
2109 smp_mb__after_atomic();
2110 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
)
2116 static void tcp_chrono_set(struct tcp_sock
*tp
, const enum tcp_chrono
new)
2118 const u32 now
= tcp_time_stamp
;
2120 if (tp
->chrono_type
> TCP_CHRONO_UNSPEC
)
2121 tp
->chrono_stat
[tp
->chrono_type
- 1] += now
- tp
->chrono_start
;
2122 tp
->chrono_start
= now
;
2123 tp
->chrono_type
= new;
2126 void tcp_chrono_start(struct sock
*sk
, const enum tcp_chrono type
)
2128 struct tcp_sock
*tp
= tcp_sk(sk
);
2130 /* If there are multiple conditions worthy of tracking in a
2131 * chronograph then the highest priority enum takes precedence
2132 * over the other conditions. So that if something "more interesting"
2133 * starts happening, stop the previous chrono and start a new one.
2135 if (type
> tp
->chrono_type
)
2136 tcp_chrono_set(tp
, type
);
2139 void tcp_chrono_stop(struct sock
*sk
, const enum tcp_chrono type
)
2141 struct tcp_sock
*tp
= tcp_sk(sk
);
2144 /* There are multiple conditions worthy of tracking in a
2145 * chronograph, so that the highest priority enum takes
2146 * precedence over the other conditions (see tcp_chrono_start).
2147 * If a condition stops, we only stop chrono tracking if
2148 * it's the "most interesting" or current chrono we are
2149 * tracking and starts busy chrono if we have pending data.
2151 if (tcp_write_queue_empty(sk
))
2152 tcp_chrono_set(tp
, TCP_CHRONO_UNSPEC
);
2153 else if (type
== tp
->chrono_type
)
2154 tcp_chrono_set(tp
, TCP_CHRONO_BUSY
);
2157 /* This routine writes packets to the network. It advances the
2158 * send_head. This happens as incoming acks open up the remote
2161 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2162 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2163 * account rare use of URG, this is not a big flaw.
2165 * Send at most one packet when push_one > 0. Temporarily ignore
2166 * cwnd limit to force at most one packet out when push_one == 2.
2168 * Returns true, if no segments are in flight and we have queued segments,
2169 * but cannot send anything now because of SWS or another problem.
2171 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
2172 int push_one
, gfp_t gfp
)
2174 struct tcp_sock
*tp
= tcp_sk(sk
);
2175 struct sk_buff
*skb
;
2176 unsigned int tso_segs
, sent_pkts
;
2179 bool is_cwnd_limited
= false, is_rwnd_limited
= false;
2185 /* Do MTU probing. */
2186 result
= tcp_mtu_probe(sk
);
2189 } else if (result
> 0) {
2194 max_segs
= tcp_tso_segs(sk
, mss_now
);
2195 while ((skb
= tcp_send_head(sk
))) {
2198 tso_segs
= tcp_init_tso_segs(skb
, mss_now
);
2201 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
) {
2202 /* "skb_mstamp" is used as a start point for the retransmit timer */
2203 skb_mstamp_get(&skb
->skb_mstamp
);
2204 goto repair
; /* Skip network transmission */
2207 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2210 /* Force out a loss probe pkt. */
2216 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
))) {
2217 is_rwnd_limited
= true;
2221 if (tso_segs
== 1) {
2222 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2223 (tcp_skb_is_last(sk
, skb
) ?
2224 nonagle
: TCP_NAGLE_PUSH
))))
2228 tcp_tso_should_defer(sk
, skb
, &is_cwnd_limited
,
2234 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
2235 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2241 if (skb
->len
> limit
&&
2242 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
2245 if (test_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
))
2246 clear_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
);
2247 if (tcp_small_queue_check(sk
, skb
, 0))
2250 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2254 /* Advance the send_head. This one is sent out.
2255 * This call will increment packets_out.
2257 tcp_event_new_data_sent(sk
, skb
);
2259 tcp_minshall_update(tp
, mss_now
, skb
);
2260 sent_pkts
+= tcp_skb_pcount(skb
);
2266 if (is_rwnd_limited
)
2267 tcp_chrono_start(sk
, TCP_CHRONO_RWND_LIMITED
);
2269 tcp_chrono_stop(sk
, TCP_CHRONO_RWND_LIMITED
);
2271 if (likely(sent_pkts
)) {
2272 if (tcp_in_cwnd_reduction(sk
))
2273 tp
->prr_out
+= sent_pkts
;
2275 /* Send one loss probe per tail loss episode. */
2277 tcp_schedule_loss_probe(sk
);
2278 is_cwnd_limited
|= (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
);
2279 tcp_cwnd_validate(sk
, is_cwnd_limited
);
2282 return !tp
->packets_out
&& tcp_send_head(sk
);
2285 bool tcp_schedule_loss_probe(struct sock
*sk
)
2287 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2288 struct tcp_sock
*tp
= tcp_sk(sk
);
2289 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
2290 u32 rtt
= usecs_to_jiffies(tp
->srtt_us
>> 3);
2292 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
))
2294 /* No consecutive loss probes. */
2295 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
2299 /* Don't do any loss probe on a Fast Open connection before 3WHS
2302 if (tp
->fastopen_rsk
)
2305 /* TLP is only scheduled when next timer event is RTO. */
2306 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
2309 /* Schedule a loss probe in 2*RTT for SACK capable connections
2310 * in Open state, that are either limited by cwnd or application.
2312 if (sysctl_tcp_early_retrans
< 3 || !tp
->packets_out
||
2313 !tcp_is_sack(tp
) || inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
2316 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
2320 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2321 * for delayed ack when there's one outstanding packet. If no RTT
2322 * sample is available then probe after TCP_TIMEOUT_INIT.
2324 timeout
= rtt
<< 1 ? : TCP_TIMEOUT_INIT
;
2325 if (tp
->packets_out
== 1)
2326 timeout
= max_t(u32
, timeout
,
2327 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
2328 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
2330 /* If RTO is shorter, just schedule TLP in its place. */
2331 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
2332 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
2333 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
2334 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
2339 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2344 /* Thanks to skb fast clones, we can detect if a prior transmit of
2345 * a packet is still in a qdisc or driver queue.
2346 * In this case, there is very little point doing a retransmit !
2348 static bool skb_still_in_host_queue(const struct sock
*sk
,
2349 const struct sk_buff
*skb
)
2351 if (unlikely(skb_fclone_busy(sk
, skb
))) {
2352 NET_INC_STATS(sock_net(sk
),
2353 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
2359 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2360 * retransmit the last segment.
2362 void tcp_send_loss_probe(struct sock
*sk
)
2364 struct tcp_sock
*tp
= tcp_sk(sk
);
2365 struct sk_buff
*skb
;
2367 int mss
= tcp_current_mss(sk
);
2369 skb
= tcp_send_head(sk
);
2371 if (tcp_snd_wnd_test(tp
, skb
, mss
)) {
2372 pcount
= tp
->packets_out
;
2373 tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2374 if (tp
->packets_out
> pcount
)
2378 skb
= tcp_write_queue_prev(sk
, skb
);
2380 skb
= tcp_write_queue_tail(sk
);
2383 /* At most one outstanding TLP retransmission. */
2384 if (tp
->tlp_high_seq
)
2387 /* Retransmit last segment. */
2391 if (skb_still_in_host_queue(sk
, skb
))
2394 pcount
= tcp_skb_pcount(skb
);
2395 if (WARN_ON(!pcount
))
2398 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2399 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
,
2402 skb
= tcp_write_queue_next(sk
, skb
);
2405 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2408 if (__tcp_retransmit_skb(sk
, skb
, 1))
2411 /* Record snd_nxt for loss detection. */
2412 tp
->tlp_high_seq
= tp
->snd_nxt
;
2415 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSPROBES
);
2416 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2417 inet_csk(sk
)->icsk_pending
= 0;
2422 /* Push out any pending frames which were held back due to
2423 * TCP_CORK or attempt at coalescing tiny packets.
2424 * The socket must be locked by the caller.
2426 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2429 /* If we are closed, the bytes will have to remain here.
2430 * In time closedown will finish, we empty the write queue and
2431 * all will be happy.
2433 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2436 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2437 sk_gfp_mask(sk
, GFP_ATOMIC
)))
2438 tcp_check_probe_timer(sk
);
2441 /* Send _single_ skb sitting at the send head. This function requires
2442 * true push pending frames to setup probe timer etc.
2444 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2446 struct sk_buff
*skb
= tcp_send_head(sk
);
2448 BUG_ON(!skb
|| skb
->len
< mss_now
);
2450 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2453 /* This function returns the amount that we can raise the
2454 * usable window based on the following constraints
2456 * 1. The window can never be shrunk once it is offered (RFC 793)
2457 * 2. We limit memory per socket
2460 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2461 * RECV.NEXT + RCV.WIN fixed until:
2462 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2464 * i.e. don't raise the right edge of the window until you can raise
2465 * it at least MSS bytes.
2467 * Unfortunately, the recommended algorithm breaks header prediction,
2468 * since header prediction assumes th->window stays fixed.
2470 * Strictly speaking, keeping th->window fixed violates the receiver
2471 * side SWS prevention criteria. The problem is that under this rule
2472 * a stream of single byte packets will cause the right side of the
2473 * window to always advance by a single byte.
2475 * Of course, if the sender implements sender side SWS prevention
2476 * then this will not be a problem.
2478 * BSD seems to make the following compromise:
2480 * If the free space is less than the 1/4 of the maximum
2481 * space available and the free space is less than 1/2 mss,
2482 * then set the window to 0.
2483 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2484 * Otherwise, just prevent the window from shrinking
2485 * and from being larger than the largest representable value.
2487 * This prevents incremental opening of the window in the regime
2488 * where TCP is limited by the speed of the reader side taking
2489 * data out of the TCP receive queue. It does nothing about
2490 * those cases where the window is constrained on the sender side
2491 * because the pipeline is full.
2493 * BSD also seems to "accidentally" limit itself to windows that are a
2494 * multiple of MSS, at least until the free space gets quite small.
2495 * This would appear to be a side effect of the mbuf implementation.
2496 * Combining these two algorithms results in the observed behavior
2497 * of having a fixed window size at almost all times.
2499 * Below we obtain similar behavior by forcing the offered window to
2500 * a multiple of the mss when it is feasible to do so.
2502 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2503 * Regular options like TIMESTAMP are taken into account.
2505 u32
__tcp_select_window(struct sock
*sk
)
2507 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2508 struct tcp_sock
*tp
= tcp_sk(sk
);
2509 /* MSS for the peer's data. Previous versions used mss_clamp
2510 * here. I don't know if the value based on our guesses
2511 * of peer's MSS is better for the performance. It's more correct
2512 * but may be worse for the performance because of rcv_mss
2513 * fluctuations. --SAW 1998/11/1
2515 int mss
= icsk
->icsk_ack
.rcv_mss
;
2516 int free_space
= tcp_space(sk
);
2517 int allowed_space
= tcp_full_space(sk
);
2518 int full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2521 if (mss
> full_space
)
2524 if (free_space
< (full_space
>> 1)) {
2525 icsk
->icsk_ack
.quick
= 0;
2527 if (tcp_under_memory_pressure(sk
))
2528 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2531 /* free_space might become our new window, make sure we don't
2532 * increase it due to wscale.
2534 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2536 /* if free space is less than mss estimate, or is below 1/16th
2537 * of the maximum allowed, try to move to zero-window, else
2538 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2539 * new incoming data is dropped due to memory limits.
2540 * With large window, mss test triggers way too late in order
2541 * to announce zero window in time before rmem limit kicks in.
2543 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2547 if (free_space
> tp
->rcv_ssthresh
)
2548 free_space
= tp
->rcv_ssthresh
;
2550 /* Don't do rounding if we are using window scaling, since the
2551 * scaled window will not line up with the MSS boundary anyway.
2553 window
= tp
->rcv_wnd
;
2554 if (tp
->rx_opt
.rcv_wscale
) {
2555 window
= free_space
;
2557 /* Advertise enough space so that it won't get scaled away.
2558 * Import case: prevent zero window announcement if
2559 * 1<<rcv_wscale > mss.
2561 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2562 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2563 << tp
->rx_opt
.rcv_wscale
);
2565 /* Get the largest window that is a nice multiple of mss.
2566 * Window clamp already applied above.
2567 * If our current window offering is within 1 mss of the
2568 * free space we just keep it. This prevents the divide
2569 * and multiply from happening most of the time.
2570 * We also don't do any window rounding when the free space
2573 if (window
<= free_space
- mss
|| window
> free_space
)
2574 window
= (free_space
/ mss
) * mss
;
2575 else if (mss
== full_space
&&
2576 free_space
> window
+ (full_space
>> 1))
2577 window
= free_space
;
2583 void tcp_skb_collapse_tstamp(struct sk_buff
*skb
,
2584 const struct sk_buff
*next_skb
)
2586 if (unlikely(tcp_has_tx_tstamp(next_skb
))) {
2587 const struct skb_shared_info
*next_shinfo
=
2588 skb_shinfo(next_skb
);
2589 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
2591 shinfo
->tx_flags
|= next_shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
2592 shinfo
->tskey
= next_shinfo
->tskey
;
2593 TCP_SKB_CB(skb
)->txstamp_ack
|=
2594 TCP_SKB_CB(next_skb
)->txstamp_ack
;
2598 /* Collapses two adjacent SKB's during retransmission. */
2599 static bool tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2601 struct tcp_sock
*tp
= tcp_sk(sk
);
2602 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2603 int skb_size
, next_skb_size
;
2605 skb_size
= skb
->len
;
2606 next_skb_size
= next_skb
->len
;
2608 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2610 if (next_skb_size
) {
2611 if (next_skb_size
<= skb_availroom(skb
))
2612 skb_copy_bits(next_skb
, 0, skb_put(skb
, next_skb_size
),
2614 else if (!skb_shift(skb
, next_skb
, next_skb_size
))
2617 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2619 tcp_unlink_write_queue(next_skb
, sk
);
2621 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2622 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2624 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2625 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2627 /* Update sequence range on original skb. */
2628 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2630 /* Merge over control information. This moves PSH/FIN etc. over */
2631 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2633 /* All done, get rid of second SKB and account for it so
2634 * packet counting does not break.
2636 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2637 TCP_SKB_CB(skb
)->eor
= TCP_SKB_CB(next_skb
)->eor
;
2639 /* changed transmit queue under us so clear hints */
2640 tcp_clear_retrans_hints_partial(tp
);
2641 if (next_skb
== tp
->retransmit_skb_hint
)
2642 tp
->retransmit_skb_hint
= skb
;
2644 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2646 tcp_skb_collapse_tstamp(skb
, next_skb
);
2648 sk_wmem_free_skb(sk
, next_skb
);
2652 /* Check if coalescing SKBs is legal. */
2653 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2655 if (tcp_skb_pcount(skb
) > 1)
2657 if (skb_cloned(skb
))
2659 if (skb
== tcp_send_head(sk
))
2661 /* Some heuristics for collapsing over SACK'd could be invented */
2662 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2668 /* Collapse packets in the retransmit queue to make to create
2669 * less packets on the wire. This is only done on retransmission.
2671 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2674 struct tcp_sock
*tp
= tcp_sk(sk
);
2675 struct sk_buff
*skb
= to
, *tmp
;
2678 if (!sysctl_tcp_retrans_collapse
)
2680 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2683 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2684 if (!tcp_can_collapse(sk
, skb
))
2687 if (!tcp_skb_can_collapse_to(to
))
2700 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2703 if (!tcp_collapse_retrans(sk
, to
))
2708 /* This retransmits one SKB. Policy decisions and retransmit queue
2709 * state updates are done by the caller. Returns non-zero if an
2710 * error occurred which prevented the send.
2712 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2714 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2715 struct tcp_sock
*tp
= tcp_sk(sk
);
2716 unsigned int cur_mss
;
2720 /* Inconclusive MTU probe */
2721 if (icsk
->icsk_mtup
.probe_size
)
2722 icsk
->icsk_mtup
.probe_size
= 0;
2724 /* Do not sent more than we queued. 1/4 is reserved for possible
2725 * copying overhead: fragmentation, tunneling, mangling etc.
2727 if (atomic_read(&sk
->sk_wmem_alloc
) >
2728 min_t(u32
, sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2),
2732 if (skb_still_in_host_queue(sk
, skb
))
2735 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2736 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2738 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2742 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2743 return -EHOSTUNREACH
; /* Routing failure or similar. */
2745 cur_mss
= tcp_current_mss(sk
);
2747 /* If receiver has shrunk his window, and skb is out of
2748 * new window, do not retransmit it. The exception is the
2749 * case, when window is shrunk to zero. In this case
2750 * our retransmit serves as a zero window probe.
2752 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2753 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2756 len
= cur_mss
* segs
;
2757 if (skb
->len
> len
) {
2758 if (tcp_fragment(sk
, skb
, len
, cur_mss
, GFP_ATOMIC
))
2759 return -ENOMEM
; /* We'll try again later. */
2761 if (skb_unclone(skb
, GFP_ATOMIC
))
2764 diff
= tcp_skb_pcount(skb
);
2765 tcp_set_skb_tso_segs(skb
, cur_mss
);
2766 diff
-= tcp_skb_pcount(skb
);
2768 tcp_adjust_pcount(sk
, skb
, diff
);
2769 if (skb
->len
< cur_mss
)
2770 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2773 /* RFC3168, section 6.1.1.1. ECN fallback */
2774 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN_ECN
) == TCPHDR_SYN_ECN
)
2775 tcp_ecn_clear_syn(sk
, skb
);
2777 /* make sure skb->data is aligned on arches that require it
2778 * and check if ack-trimming & collapsing extended the headroom
2779 * beyond what csum_start can cover.
2781 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2782 skb_headroom(skb
) >= 0xFFFF)) {
2783 struct sk_buff
*nskb
;
2785 skb_mstamp_get(&skb
->skb_mstamp
);
2786 nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
, GFP_ATOMIC
);
2787 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2790 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2794 segs
= tcp_skb_pcount(skb
);
2796 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
2797 /* Update global TCP statistics. */
2798 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
, segs
);
2799 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2800 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
2801 tp
->total_retrans
+= segs
;
2806 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2808 struct tcp_sock
*tp
= tcp_sk(sk
);
2809 int err
= __tcp_retransmit_skb(sk
, skb
, segs
);
2812 #if FASTRETRANS_DEBUG > 0
2813 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2814 net_dbg_ratelimited("retrans_out leaked\n");
2817 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2818 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2820 /* Save stamp of the first retransmit. */
2821 if (!tp
->retrans_stamp
)
2822 tp
->retrans_stamp
= tcp_skb_timestamp(skb
);
2824 } else if (err
!= -EBUSY
) {
2825 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2828 if (tp
->undo_retrans
< 0)
2829 tp
->undo_retrans
= 0;
2830 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2834 /* Check if we forward retransmits are possible in the current
2835 * window/congestion state.
2837 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2839 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2840 const struct tcp_sock
*tp
= tcp_sk(sk
);
2842 /* Forward retransmissions are possible only during Recovery. */
2843 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2846 /* No forward retransmissions in Reno are possible. */
2847 if (tcp_is_reno(tp
))
2850 /* Yeah, we have to make difficult choice between forward transmission
2851 * and retransmission... Both ways have their merits...
2853 * For now we do not retransmit anything, while we have some new
2854 * segments to send. In the other cases, follow rule 3 for
2855 * NextSeg() specified in RFC3517.
2858 if (tcp_may_send_now(sk
))
2864 /* This gets called after a retransmit timeout, and the initially
2865 * retransmitted data is acknowledged. It tries to continue
2866 * resending the rest of the retransmit queue, until either
2867 * we've sent it all or the congestion window limit is reached.
2868 * If doing SACK, the first ACK which comes back for a timeout
2869 * based retransmit packet might feed us FACK information again.
2870 * If so, we use it to avoid unnecessarily retransmissions.
2872 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2874 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2875 struct tcp_sock
*tp
= tcp_sk(sk
);
2876 struct sk_buff
*skb
;
2877 struct sk_buff
*hole
= NULL
;
2878 u32 max_segs
, last_lost
;
2880 int fwd_rexmitting
= 0;
2882 if (!tp
->packets_out
)
2886 tp
->retransmit_high
= tp
->snd_una
;
2888 if (tp
->retransmit_skb_hint
) {
2889 skb
= tp
->retransmit_skb_hint
;
2890 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2891 if (after(last_lost
, tp
->retransmit_high
))
2892 last_lost
= tp
->retransmit_high
;
2894 skb
= tcp_write_queue_head(sk
);
2895 last_lost
= tp
->snd_una
;
2898 max_segs
= tcp_tso_segs(sk
, tcp_current_mss(sk
));
2899 tcp_for_write_queue_from(skb
, sk
) {
2903 if (skb
== tcp_send_head(sk
))
2905 /* we could do better than to assign each time */
2907 tp
->retransmit_skb_hint
= skb
;
2909 segs
= tp
->snd_cwnd
- tcp_packets_in_flight(tp
);
2912 sacked
= TCP_SKB_CB(skb
)->sacked
;
2913 /* In case tcp_shift_skb_data() have aggregated large skbs,
2914 * we need to make sure not sending too bigs TSO packets
2916 segs
= min_t(int, segs
, max_segs
);
2918 if (fwd_rexmitting
) {
2920 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2922 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2924 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2925 tp
->retransmit_high
= last_lost
;
2926 if (!tcp_can_forward_retransmit(sk
))
2928 /* Backtrack if necessary to non-L'ed skb */
2936 } else if (!(sacked
& TCPCB_LOST
)) {
2937 if (!hole
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2942 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2943 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2944 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2946 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2949 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2952 if (tcp_small_queue_check(sk
, skb
, 1))
2955 if (tcp_retransmit_skb(sk
, skb
, segs
))
2958 NET_ADD_STATS(sock_net(sk
), mib_idx
, tcp_skb_pcount(skb
));
2960 if (tcp_in_cwnd_reduction(sk
))
2961 tp
->prr_out
+= tcp_skb_pcount(skb
);
2963 if (skb
== tcp_write_queue_head(sk
))
2964 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2965 inet_csk(sk
)->icsk_rto
,
2970 /* We allow to exceed memory limits for FIN packets to expedite
2971 * connection tear down and (memory) recovery.
2972 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2973 * or even be forced to close flow without any FIN.
2974 * In general, we want to allow one skb per socket to avoid hangs
2975 * with edge trigger epoll()
2977 void sk_forced_mem_schedule(struct sock
*sk
, int size
)
2981 if (size
<= sk
->sk_forward_alloc
)
2983 amt
= sk_mem_pages(size
);
2984 sk
->sk_forward_alloc
+= amt
* SK_MEM_QUANTUM
;
2985 sk_memory_allocated_add(sk
, amt
);
2987 if (mem_cgroup_sockets_enabled
&& sk
->sk_memcg
)
2988 mem_cgroup_charge_skmem(sk
->sk_memcg
, amt
);
2991 /* Send a FIN. The caller locks the socket for us.
2992 * We should try to send a FIN packet really hard, but eventually give up.
2994 void tcp_send_fin(struct sock
*sk
)
2996 struct sk_buff
*skb
, *tskb
= tcp_write_queue_tail(sk
);
2997 struct tcp_sock
*tp
= tcp_sk(sk
);
2999 /* Optimization, tack on the FIN if we have one skb in write queue and
3000 * this skb was not yet sent, or we are under memory pressure.
3001 * Note: in the latter case, FIN packet will be sent after a timeout,
3002 * as TCP stack thinks it has already been transmitted.
3004 if (tskb
&& (tcp_send_head(sk
) || tcp_under_memory_pressure(sk
))) {
3006 TCP_SKB_CB(tskb
)->tcp_flags
|= TCPHDR_FIN
;
3007 TCP_SKB_CB(tskb
)->end_seq
++;
3009 if (!tcp_send_head(sk
)) {
3010 /* This means tskb was already sent.
3011 * Pretend we included the FIN on previous transmit.
3012 * We need to set tp->snd_nxt to the value it would have
3013 * if FIN had been sent. This is because retransmit path
3014 * does not change tp->snd_nxt.
3020 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, sk
->sk_allocation
);
3021 if (unlikely(!skb
)) {
3026 skb_reserve(skb
, MAX_TCP_HEADER
);
3027 sk_forced_mem_schedule(sk
, skb
->truesize
);
3028 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3029 tcp_init_nondata_skb(skb
, tp
->write_seq
,
3030 TCPHDR_ACK
| TCPHDR_FIN
);
3031 tcp_queue_skb(sk
, skb
);
3033 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
), TCP_NAGLE_OFF
);
3036 /* We get here when a process closes a file descriptor (either due to
3037 * an explicit close() or as a byproduct of exit()'ing) and there
3038 * was unread data in the receive queue. This behavior is recommended
3039 * by RFC 2525, section 2.17. -DaveM
3041 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
3043 struct sk_buff
*skb
;
3045 /* NOTE: No TCP options attached and we never retransmit this. */
3046 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
3048 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3052 /* Reserve space for headers and prepare control bits. */
3053 skb_reserve(skb
, MAX_TCP_HEADER
);
3054 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
3055 TCPHDR_ACK
| TCPHDR_RST
);
3056 skb_mstamp_get(&skb
->skb_mstamp
);
3058 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
3059 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3061 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
3064 /* Send a crossed SYN-ACK during socket establishment.
3065 * WARNING: This routine must only be called when we have already sent
3066 * a SYN packet that crossed the incoming SYN that caused this routine
3067 * to get called. If this assumption fails then the initial rcv_wnd
3068 * and rcv_wscale values will not be correct.
3070 int tcp_send_synack(struct sock
*sk
)
3072 struct sk_buff
*skb
;
3074 skb
= tcp_write_queue_head(sk
);
3075 if (!skb
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
3076 pr_debug("%s: wrong queue state\n", __func__
);
3079 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
3080 if (skb_cloned(skb
)) {
3081 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
3084 tcp_unlink_write_queue(skb
, sk
);
3085 __skb_header_release(nskb
);
3086 __tcp_add_write_queue_head(sk
, nskb
);
3087 sk_wmem_free_skb(sk
, skb
);
3088 sk
->sk_wmem_queued
+= nskb
->truesize
;
3089 sk_mem_charge(sk
, nskb
->truesize
);
3093 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
3094 tcp_ecn_send_synack(sk
, skb
);
3096 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3100 * tcp_make_synack - Prepare a SYN-ACK.
3101 * sk: listener socket
3102 * dst: dst entry attached to the SYNACK
3103 * req: request_sock pointer
3105 * Allocate one skb and build a SYNACK packet.
3106 * @dst is consumed : Caller should not use it again.
3108 struct sk_buff
*tcp_make_synack(const struct sock
*sk
, struct dst_entry
*dst
,
3109 struct request_sock
*req
,
3110 struct tcp_fastopen_cookie
*foc
,
3111 enum tcp_synack_type synack_type
)
3113 struct inet_request_sock
*ireq
= inet_rsk(req
);
3114 const struct tcp_sock
*tp
= tcp_sk(sk
);
3115 struct tcp_md5sig_key
*md5
= NULL
;
3116 struct tcp_out_options opts
;
3117 struct sk_buff
*skb
;
3118 int tcp_header_size
;
3123 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
3124 if (unlikely(!skb
)) {
3128 /* Reserve space for headers. */
3129 skb_reserve(skb
, MAX_TCP_HEADER
);
3131 switch (synack_type
) {
3132 case TCP_SYNACK_NORMAL
:
3133 skb_set_owner_w(skb
, req_to_sk(req
));
3135 case TCP_SYNACK_COOKIE
:
3136 /* Under synflood, we do not attach skb to a socket,
3137 * to avoid false sharing.
3140 case TCP_SYNACK_FASTOPEN
:
3141 /* sk is a const pointer, because we want to express multiple
3142 * cpu might call us concurrently.
3143 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3145 skb_set_owner_w(skb
, (struct sock
*)sk
);
3148 skb_dst_set(skb
, dst
);
3150 mss
= dst_metric_advmss(dst
);
3151 user_mss
= READ_ONCE(tp
->rx_opt
.user_mss
);
3152 if (user_mss
&& user_mss
< mss
)
3155 memset(&opts
, 0, sizeof(opts
));
3156 #ifdef CONFIG_SYN_COOKIES
3157 if (unlikely(req
->cookie_ts
))
3158 skb
->skb_mstamp
.stamp_jiffies
= cookie_init_timestamp(req
);
3161 skb_mstamp_get(&skb
->skb_mstamp
);
3163 #ifdef CONFIG_TCP_MD5SIG
3165 md5
= tcp_rsk(req
)->af_specific
->req_md5_lookup(sk
, req_to_sk(req
));
3167 skb_set_hash(skb
, tcp_rsk(req
)->txhash
, PKT_HASH_TYPE_L4
);
3168 tcp_header_size
= tcp_synack_options(req
, mss
, skb
, &opts
, md5
, foc
) +
3171 skb_push(skb
, tcp_header_size
);
3172 skb_reset_transport_header(skb
);
3174 th
= (struct tcphdr
*)skb
->data
;
3175 memset(th
, 0, sizeof(struct tcphdr
));
3178 tcp_ecn_make_synack(req
, th
);
3179 th
->source
= htons(ireq
->ir_num
);
3180 th
->dest
= ireq
->ir_rmt_port
;
3181 /* Setting of flags are superfluous here for callers (and ECE is
3182 * not even correctly set)
3184 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
3185 TCPHDR_SYN
| TCPHDR_ACK
);
3187 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
3188 /* XXX data is queued and acked as is. No buffer/window check */
3189 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
3191 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3192 th
->window
= htons(min(req
->rsk_rcv_wnd
, 65535U));
3193 tcp_options_write((__be32
*)(th
+ 1), NULL
, &opts
);
3194 th
->doff
= (tcp_header_size
>> 2);
3195 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
3197 #ifdef CONFIG_TCP_MD5SIG
3198 /* Okay, we have all we need - do the md5 hash if needed */
3200 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
3201 md5
, req_to_sk(req
), skb
);
3205 /* Do not fool tcpdump (if any), clean our debris */
3209 EXPORT_SYMBOL(tcp_make_synack
);
3211 static void tcp_ca_dst_init(struct sock
*sk
, const struct dst_entry
*dst
)
3213 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3214 const struct tcp_congestion_ops
*ca
;
3215 u32 ca_key
= dst_metric(dst
, RTAX_CC_ALGO
);
3217 if (ca_key
== TCP_CA_UNSPEC
)
3221 ca
= tcp_ca_find_key(ca_key
);
3222 if (likely(ca
&& try_module_get(ca
->owner
))) {
3223 module_put(icsk
->icsk_ca_ops
->owner
);
3224 icsk
->icsk_ca_dst_locked
= tcp_ca_dst_locked(dst
);
3225 icsk
->icsk_ca_ops
= ca
;
3230 /* Do all connect socket setups that can be done AF independent. */
3231 static void tcp_connect_init(struct sock
*sk
)
3233 const struct dst_entry
*dst
= __sk_dst_get(sk
);
3234 struct tcp_sock
*tp
= tcp_sk(sk
);
3237 /* We'll fix this up when we get a response from the other end.
3238 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3240 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
3241 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
3243 #ifdef CONFIG_TCP_MD5SIG
3244 if (tp
->af_specific
->md5_lookup(sk
, sk
))
3245 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
3248 /* If user gave his TCP_MAXSEG, record it to clamp */
3249 if (tp
->rx_opt
.user_mss
)
3250 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3253 tcp_sync_mss(sk
, dst_mtu(dst
));
3255 tcp_ca_dst_init(sk
, dst
);
3257 if (!tp
->window_clamp
)
3258 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
3259 tp
->advmss
= dst_metric_advmss(dst
);
3260 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
3261 tp
->advmss
= tp
->rx_opt
.user_mss
;
3263 tcp_initialize_rcv_mss(sk
);
3265 /* limit the window selection if the user enforce a smaller rx buffer */
3266 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
3267 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
3268 tp
->window_clamp
= tcp_full_space(sk
);
3270 tcp_select_initial_window(tcp_full_space(sk
),
3271 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
3274 sysctl_tcp_window_scaling
,
3276 dst_metric(dst
, RTAX_INITRWND
));
3278 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
3279 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
3282 sock_reset_flag(sk
, SOCK_DONE
);
3285 tp
->snd_una
= tp
->write_seq
;
3286 tp
->snd_sml
= tp
->write_seq
;
3287 tp
->snd_up
= tp
->write_seq
;
3288 tp
->snd_nxt
= tp
->write_seq
;
3290 if (likely(!tp
->repair
))
3293 tp
->rcv_tstamp
= tcp_time_stamp
;
3294 tp
->rcv_wup
= tp
->rcv_nxt
;
3295 tp
->copied_seq
= tp
->rcv_nxt
;
3297 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
3298 inet_csk(sk
)->icsk_retransmits
= 0;
3299 tcp_clear_retrans(tp
);
3302 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
3304 struct tcp_sock
*tp
= tcp_sk(sk
);
3305 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
3307 tcb
->end_seq
+= skb
->len
;
3308 __skb_header_release(skb
);
3309 __tcp_add_write_queue_tail(sk
, skb
);
3310 sk
->sk_wmem_queued
+= skb
->truesize
;
3311 sk_mem_charge(sk
, skb
->truesize
);
3312 tp
->write_seq
= tcb
->end_seq
;
3313 tp
->packets_out
+= tcp_skb_pcount(skb
);
3316 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3317 * queue a data-only packet after the regular SYN, such that regular SYNs
3318 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3319 * only the SYN sequence, the data are retransmitted in the first ACK.
3320 * If cookie is not cached or other error occurs, falls back to send a
3321 * regular SYN with Fast Open cookie request option.
3323 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
3325 struct tcp_sock
*tp
= tcp_sk(sk
);
3326 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
3327 int syn_loss
= 0, space
, err
= 0;
3328 unsigned long last_syn_loss
= 0;
3329 struct sk_buff
*syn_data
;
3331 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
3332 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
3333 &syn_loss
, &last_syn_loss
);
3334 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3336 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
3337 fo
->cookie
.len
= -1;
3341 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
3342 fo
->cookie
.len
= -1;
3343 else if (fo
->cookie
.len
<= 0)
3346 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3347 * user-MSS. Reserve maximum option space for middleboxes that add
3348 * private TCP options. The cost is reduced data space in SYN :(
3350 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
3351 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3352 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
3353 MAX_TCP_OPTION_SPACE
;
3355 space
= min_t(size_t, space
, fo
->size
);
3357 /* limit to order-0 allocations */
3358 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
3360 syn_data
= sk_stream_alloc_skb(sk
, space
, sk
->sk_allocation
, false);
3363 syn_data
->ip_summed
= CHECKSUM_PARTIAL
;
3364 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
3366 int copied
= copy_from_iter(skb_put(syn_data
, space
), space
,
3367 &fo
->data
->msg_iter
);
3368 if (unlikely(!copied
)) {
3369 kfree_skb(syn_data
);
3372 if (copied
!= space
) {
3373 skb_trim(syn_data
, copied
);
3377 /* No more data pending in inet_wait_for_connect() */
3378 if (space
== fo
->size
)
3382 tcp_connect_queue_skb(sk
, syn_data
);
3384 tcp_chrono_start(sk
, TCP_CHRONO_BUSY
);
3386 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
3388 syn
->skb_mstamp
= syn_data
->skb_mstamp
;
3390 /* Now full SYN+DATA was cloned and sent (or not),
3391 * remove the SYN from the original skb (syn_data)
3392 * we keep in write queue in case of a retransmit, as we
3393 * also have the SYN packet (with no data) in the same queue.
3395 TCP_SKB_CB(syn_data
)->seq
++;
3396 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
3398 tp
->syn_data
= (fo
->copied
> 0);
3399 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3404 /* Send a regular SYN with Fast Open cookie request option */
3405 if (fo
->cookie
.len
> 0)
3407 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3409 tp
->syn_fastopen
= 0;
3411 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3415 /* Build a SYN and send it off. */
3416 int tcp_connect(struct sock
*sk
)
3418 struct tcp_sock
*tp
= tcp_sk(sk
);
3419 struct sk_buff
*buff
;
3422 tcp_connect_init(sk
);
3424 if (unlikely(tp
->repair
)) {
3425 tcp_finish_connect(sk
, NULL
);
3429 buff
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
, true);
3430 if (unlikely(!buff
))
3433 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3434 tp
->retrans_stamp
= tcp_time_stamp
;
3435 tcp_connect_queue_skb(sk
, buff
);
3436 tcp_ecn_send_syn(sk
, buff
);
3438 /* Send off SYN; include data in Fast Open. */
3439 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3440 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3441 if (err
== -ECONNREFUSED
)
3444 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3445 * in order to make this packet get counted in tcpOutSegs.
3447 tp
->snd_nxt
= tp
->write_seq
;
3448 tp
->pushed_seq
= tp
->write_seq
;
3449 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3451 /* Timer for repeating the SYN until an answer. */
3452 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3453 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3456 EXPORT_SYMBOL(tcp_connect
);
3458 /* Send out a delayed ack, the caller does the policy checking
3459 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3462 void tcp_send_delayed_ack(struct sock
*sk
)
3464 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3465 int ato
= icsk
->icsk_ack
.ato
;
3466 unsigned long timeout
;
3468 tcp_ca_event(sk
, CA_EVENT_DELAYED_ACK
);
3470 if (ato
> TCP_DELACK_MIN
) {
3471 const struct tcp_sock
*tp
= tcp_sk(sk
);
3472 int max_ato
= HZ
/ 2;
3474 if (icsk
->icsk_ack
.pingpong
||
3475 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3476 max_ato
= TCP_DELACK_MAX
;
3478 /* Slow path, intersegment interval is "high". */
3480 /* If some rtt estimate is known, use it to bound delayed ack.
3481 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3485 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3492 ato
= min(ato
, max_ato
);
3495 /* Stay within the limit we were given */
3496 timeout
= jiffies
+ ato
;
3498 /* Use new timeout only if there wasn't a older one earlier. */
3499 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3500 /* If delack timer was blocked or is about to expire,
3503 if (icsk
->icsk_ack
.blocked
||
3504 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3509 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3510 timeout
= icsk
->icsk_ack
.timeout
;
3512 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3513 icsk
->icsk_ack
.timeout
= timeout
;
3514 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3517 /* This routine sends an ack and also updates the window. */
3518 void tcp_send_ack(struct sock
*sk
)
3520 struct sk_buff
*buff
;
3522 /* If we have been reset, we may not send again. */
3523 if (sk
->sk_state
== TCP_CLOSE
)
3526 tcp_ca_event(sk
, CA_EVENT_NON_DELAYED_ACK
);
3528 /* We are not putting this on the write queue, so
3529 * tcp_transmit_skb() will set the ownership to this
3532 buff
= alloc_skb(MAX_TCP_HEADER
,
3533 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3534 if (unlikely(!buff
)) {
3535 inet_csk_schedule_ack(sk
);
3536 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3537 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3538 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3542 /* Reserve space for headers and prepare control bits. */
3543 skb_reserve(buff
, MAX_TCP_HEADER
);
3544 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3546 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3548 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3550 skb_set_tcp_pure_ack(buff
);
3552 /* Send it off, this clears delayed acks for us. */
3553 skb_mstamp_get(&buff
->skb_mstamp
);
3554 tcp_transmit_skb(sk
, buff
, 0, (__force gfp_t
)0);
3556 EXPORT_SYMBOL_GPL(tcp_send_ack
);
3558 /* This routine sends a packet with an out of date sequence
3559 * number. It assumes the other end will try to ack it.
3561 * Question: what should we make while urgent mode?
3562 * 4.4BSD forces sending single byte of data. We cannot send
3563 * out of window data, because we have SND.NXT==SND.MAX...
3565 * Current solution: to send TWO zero-length segments in urgent mode:
3566 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3567 * out-of-date with SND.UNA-1 to probe window.
3569 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
, int mib
)
3571 struct tcp_sock
*tp
= tcp_sk(sk
);
3572 struct sk_buff
*skb
;
3574 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3575 skb
= alloc_skb(MAX_TCP_HEADER
,
3576 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3580 /* Reserve space for headers and set control bits. */
3581 skb_reserve(skb
, MAX_TCP_HEADER
);
3582 /* Use a previous sequence. This should cause the other
3583 * end to send an ack. Don't queue or clone SKB, just
3586 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3587 skb_mstamp_get(&skb
->skb_mstamp
);
3588 NET_INC_STATS(sock_net(sk
), mib
);
3589 return tcp_transmit_skb(sk
, skb
, 0, (__force gfp_t
)0);
3592 void tcp_send_window_probe(struct sock
*sk
)
3594 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3595 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3596 tcp_xmit_probe_skb(sk
, 0, LINUX_MIB_TCPWINPROBE
);
3600 /* Initiate keepalive or window probe from timer. */
3601 int tcp_write_wakeup(struct sock
*sk
, int mib
)
3603 struct tcp_sock
*tp
= tcp_sk(sk
);
3604 struct sk_buff
*skb
;
3606 if (sk
->sk_state
== TCP_CLOSE
)
3609 skb
= tcp_send_head(sk
);
3610 if (skb
&& before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3612 unsigned int mss
= tcp_current_mss(sk
);
3613 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3615 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3616 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3618 /* We are probing the opening of a window
3619 * but the window size is != 0
3620 * must have been a result SWS avoidance ( sender )
3622 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3624 seg_size
= min(seg_size
, mss
);
3625 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3626 if (tcp_fragment(sk
, skb
, seg_size
, mss
, GFP_ATOMIC
))
3628 } else if (!tcp_skb_pcount(skb
))
3629 tcp_set_skb_tso_segs(skb
, mss
);
3631 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3632 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3634 tcp_event_new_data_sent(sk
, skb
);
3637 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3638 tcp_xmit_probe_skb(sk
, 1, mib
);
3639 return tcp_xmit_probe_skb(sk
, 0, mib
);
3643 /* A window probe timeout has occurred. If window is not closed send
3644 * a partial packet else a zero probe.
3646 void tcp_send_probe0(struct sock
*sk
)
3648 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3649 struct tcp_sock
*tp
= tcp_sk(sk
);
3650 struct net
*net
= sock_net(sk
);
3651 unsigned long probe_max
;
3654 err
= tcp_write_wakeup(sk
, LINUX_MIB_TCPWINPROBE
);
3656 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3657 /* Cancel probe timer, if it is not required. */
3658 icsk
->icsk_probes_out
= 0;
3659 icsk
->icsk_backoff
= 0;
3664 if (icsk
->icsk_backoff
< net
->ipv4
.sysctl_tcp_retries2
)
3665 icsk
->icsk_backoff
++;
3666 icsk
->icsk_probes_out
++;
3667 probe_max
= TCP_RTO_MAX
;
3669 /* If packet was not sent due to local congestion,
3670 * do not backoff and do not remember icsk_probes_out.
3671 * Let local senders to fight for local resources.
3673 * Use accumulated backoff yet.
3675 if (!icsk
->icsk_probes_out
)
3676 icsk
->icsk_probes_out
= 1;
3677 probe_max
= TCP_RESOURCE_PROBE_INTERVAL
;
3679 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3680 tcp_probe0_when(sk
, probe_max
),
3684 int tcp_rtx_synack(const struct sock
*sk
, struct request_sock
*req
)
3686 const struct tcp_request_sock_ops
*af_ops
= tcp_rsk(req
)->af_specific
;
3690 tcp_rsk(req
)->txhash
= net_tx_rndhash();
3691 res
= af_ops
->send_synack(sk
, NULL
, &fl
, req
, NULL
, TCP_SYNACK_NORMAL
);
3693 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
3694 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
3695 if (unlikely(tcp_passive_fastopen(sk
)))
3696 tcp_sk(sk
)->total_retrans
++;
3700 EXPORT_SYMBOL(tcp_rtx_synack
);