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 clear_bit(TSQ_QUEUED
, &sk
->sk_tsq_flags
);
774 if (!sk
->sk_lock
.owned
&&
775 test_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
)) {
777 if (!sock_owned_by_user(sk
)) {
778 clear_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
);
788 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
789 TCPF_WRITE_TIMER_DEFERRED | \
790 TCPF_DELACK_TIMER_DEFERRED | \
791 TCPF_MTU_REDUCED_DEFERRED)
793 * tcp_release_cb - tcp release_sock() callback
796 * called from release_sock() to perform protocol dependent
797 * actions before socket release.
799 void tcp_release_cb(struct sock
*sk
)
801 unsigned long flags
, nflags
;
803 /* perform an atomic operation only if at least one flag is set */
805 flags
= sk
->sk_tsq_flags
;
806 if (!(flags
& TCP_DEFERRED_ALL
))
808 nflags
= flags
& ~TCP_DEFERRED_ALL
;
809 } while (cmpxchg(&sk
->sk_tsq_flags
, flags
, nflags
) != flags
);
811 if (flags
& TCPF_TSQ_DEFERRED
)
814 /* Here begins the tricky part :
815 * We are called from release_sock() with :
817 * 2) sk_lock.slock spinlock held
818 * 3) socket owned by us (sk->sk_lock.owned == 1)
820 * But following code is meant to be called from BH handlers,
821 * so we should keep BH disabled, but early release socket ownership
823 sock_release_ownership(sk
);
825 if (flags
& TCPF_WRITE_TIMER_DEFERRED
) {
826 tcp_write_timer_handler(sk
);
829 if (flags
& TCPF_DELACK_TIMER_DEFERRED
) {
830 tcp_delack_timer_handler(sk
);
833 if (flags
& TCPF_MTU_REDUCED_DEFERRED
) {
834 inet_csk(sk
)->icsk_af_ops
->mtu_reduced(sk
);
838 EXPORT_SYMBOL(tcp_release_cb
);
840 void __init
tcp_tasklet_init(void)
844 for_each_possible_cpu(i
) {
845 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
847 INIT_LIST_HEAD(&tsq
->head
);
848 tasklet_init(&tsq
->tasklet
,
855 * Write buffer destructor automatically called from kfree_skb.
856 * We can't xmit new skbs from this context, as we might already
859 void tcp_wfree(struct sk_buff
*skb
)
861 struct sock
*sk
= skb
->sk
;
862 struct tcp_sock
*tp
= tcp_sk(sk
);
863 unsigned long flags
, nval
, oval
;
866 /* Keep one reference on sk_wmem_alloc.
867 * Will be released by sk_free() from here or tcp_tasklet_func()
869 wmem
= atomic_sub_return(skb
->truesize
- 1, &sk
->sk_wmem_alloc
);
871 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
872 * Wait until our queues (qdisc + devices) are drained.
874 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
875 * - chance for incoming ACK (processed by another cpu maybe)
876 * to migrate this flow (skb->ooo_okay will be eventually set)
878 if (wmem
>= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current
)
881 for (oval
= READ_ONCE(sk
->sk_tsq_flags
);; oval
= nval
) {
882 struct tsq_tasklet
*tsq
;
885 if (!(oval
& TSQF_THROTTLED
) || (oval
& TSQF_QUEUED
))
888 nval
= (oval
& ~TSQF_THROTTLED
) | TSQF_QUEUED
| TCPF_TSQ_DEFERRED
;
889 nval
= cmpxchg(&sk
->sk_tsq_flags
, oval
, nval
);
893 /* queue this socket to tasklet queue */
894 local_irq_save(flags
);
895 tsq
= this_cpu_ptr(&tsq_tasklet
);
896 empty
= list_empty(&tsq
->head
);
897 list_add(&tp
->tsq_node
, &tsq
->head
);
899 tasklet_schedule(&tsq
->tasklet
);
900 local_irq_restore(flags
);
907 /* This routine actually transmits TCP packets queued in by
908 * tcp_do_sendmsg(). This is used by both the initial
909 * transmission and possible later retransmissions.
910 * All SKB's seen here are completely headerless. It is our
911 * job to build the TCP header, and pass the packet down to
912 * IP so it can do the same plus pass the packet off to the
915 * We are working here with either a clone of the original
916 * SKB, or a fresh unique copy made by the retransmit engine.
918 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
921 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
922 struct inet_sock
*inet
;
924 struct tcp_skb_cb
*tcb
;
925 struct tcp_out_options opts
;
926 unsigned int tcp_options_size
, tcp_header_size
;
927 struct tcp_md5sig_key
*md5
;
931 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
935 skb_mstamp_get(&skb
->skb_mstamp
);
936 TCP_SKB_CB(skb
)->tx
.in_flight
= TCP_SKB_CB(skb
)->end_seq
938 tcp_rate_skb_sent(sk
, skb
);
940 if (unlikely(skb_cloned(skb
)))
941 skb
= pskb_copy(skb
, gfp_mask
);
943 skb
= skb_clone(skb
, gfp_mask
);
949 tcb
= TCP_SKB_CB(skb
);
950 memset(&opts
, 0, sizeof(opts
));
952 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
953 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
955 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
957 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
959 /* if no packet is in qdisc/device queue, then allow XPS to select
960 * another queue. We can be called from tcp_tsq_handler()
961 * which holds one reference to sk_wmem_alloc.
963 * TODO: Ideally, in-flight pure ACK packets should not matter here.
964 * One way to get this would be to set skb->truesize = 2 on them.
966 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) < SKB_TRUESIZE(1);
968 skb_push(skb
, tcp_header_size
);
969 skb_reset_transport_header(skb
);
973 skb
->destructor
= skb_is_tcp_pure_ack(skb
) ? __sock_wfree
: tcp_wfree
;
974 skb_set_hash_from_sk(skb
, sk
);
975 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
977 /* Build TCP header and checksum it. */
978 th
= (struct tcphdr
*)skb
->data
;
979 th
->source
= inet
->inet_sport
;
980 th
->dest
= inet
->inet_dport
;
981 th
->seq
= htonl(tcb
->seq
);
982 th
->ack_seq
= htonl(tp
->rcv_nxt
);
983 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
989 /* The urg_mode check is necessary during a below snd_una win probe */
990 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
991 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
992 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
994 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
995 th
->urg_ptr
= htons(0xFFFF);
1000 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
1001 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
1002 if (likely(!(tcb
->tcp_flags
& TCPHDR_SYN
))) {
1003 th
->window
= htons(tcp_select_window(sk
));
1004 tcp_ecn_send(sk
, skb
, th
, tcp_header_size
);
1006 /* RFC1323: The window in SYN & SYN/ACK segments
1009 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
1011 #ifdef CONFIG_TCP_MD5SIG
1012 /* Calculate the MD5 hash, as we have all we need now */
1014 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
1015 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
1020 icsk
->icsk_af_ops
->send_check(sk
, skb
);
1022 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
1023 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
1025 if (skb
->len
!= tcp_header_size
) {
1026 tcp_event_data_sent(tp
, sk
);
1027 tp
->data_segs_out
+= tcp_skb_pcount(skb
);
1030 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
1031 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
1032 tcp_skb_pcount(skb
));
1034 tp
->segs_out
+= tcp_skb_pcount(skb
);
1035 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1036 skb_shinfo(skb
)->gso_segs
= tcp_skb_pcount(skb
);
1037 skb_shinfo(skb
)->gso_size
= tcp_skb_mss(skb
);
1039 /* Our usage of tstamp should remain private */
1040 skb
->tstamp
.tv64
= 0;
1042 /* Cleanup our debris for IP stacks */
1043 memset(skb
->cb
, 0, max(sizeof(struct inet_skb_parm
),
1044 sizeof(struct inet6_skb_parm
)));
1046 err
= icsk
->icsk_af_ops
->queue_xmit(sk
, skb
, &inet
->cork
.fl
);
1048 if (likely(err
<= 0))
1053 return net_xmit_eval(err
);
1056 /* This routine just queues the buffer for sending.
1058 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1059 * otherwise socket can stall.
1061 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1063 struct tcp_sock
*tp
= tcp_sk(sk
);
1065 /* Advance write_seq and place onto the write_queue. */
1066 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
1067 __skb_header_release(skb
);
1068 tcp_add_write_queue_tail(sk
, skb
);
1069 sk
->sk_wmem_queued
+= skb
->truesize
;
1070 sk_mem_charge(sk
, skb
->truesize
);
1073 /* Initialize TSO segments for a packet. */
1074 static void tcp_set_skb_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1076 if (skb
->len
<= mss_now
|| skb
->ip_summed
== CHECKSUM_NONE
) {
1077 /* Avoid the costly divide in the normal
1080 tcp_skb_pcount_set(skb
, 1);
1081 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1083 tcp_skb_pcount_set(skb
, DIV_ROUND_UP(skb
->len
, mss_now
));
1084 TCP_SKB_CB(skb
)->tcp_gso_size
= mss_now
;
1088 /* When a modification to fackets out becomes necessary, we need to check
1089 * skb is counted to fackets_out or not.
1091 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1094 struct tcp_sock
*tp
= tcp_sk(sk
);
1096 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1099 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1100 tp
->fackets_out
-= decr
;
1103 /* Pcount in the middle of the write queue got changed, we need to do various
1104 * tweaks to fix counters
1106 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1108 struct tcp_sock
*tp
= tcp_sk(sk
);
1110 tp
->packets_out
-= decr
;
1112 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1113 tp
->sacked_out
-= decr
;
1114 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1115 tp
->retrans_out
-= decr
;
1116 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1117 tp
->lost_out
-= decr
;
1119 /* Reno case is special. Sigh... */
1120 if (tcp_is_reno(tp
) && decr
> 0)
1121 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1123 tcp_adjust_fackets_out(sk
, skb
, decr
);
1125 if (tp
->lost_skb_hint
&&
1126 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1127 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1128 tp
->lost_cnt_hint
-= decr
;
1130 tcp_verify_left_out(tp
);
1133 static bool tcp_has_tx_tstamp(const struct sk_buff
*skb
)
1135 return TCP_SKB_CB(skb
)->txstamp_ack
||
1136 (skb_shinfo(skb
)->tx_flags
& SKBTX_ANY_TSTAMP
);
1139 static void tcp_fragment_tstamp(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1141 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1143 if (unlikely(tcp_has_tx_tstamp(skb
)) &&
1144 !before(shinfo
->tskey
, TCP_SKB_CB(skb2
)->seq
)) {
1145 struct skb_shared_info
*shinfo2
= skb_shinfo(skb2
);
1146 u8 tsflags
= shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
1148 shinfo
->tx_flags
&= ~tsflags
;
1149 shinfo2
->tx_flags
|= tsflags
;
1150 swap(shinfo
->tskey
, shinfo2
->tskey
);
1151 TCP_SKB_CB(skb2
)->txstamp_ack
= TCP_SKB_CB(skb
)->txstamp_ack
;
1152 TCP_SKB_CB(skb
)->txstamp_ack
= 0;
1156 static void tcp_skb_fragment_eor(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1158 TCP_SKB_CB(skb2
)->eor
= TCP_SKB_CB(skb
)->eor
;
1159 TCP_SKB_CB(skb
)->eor
= 0;
1162 /* Function to create two new TCP segments. Shrinks the given segment
1163 * to the specified size and appends a new segment with the rest of the
1164 * packet to the list. This won't be called frequently, I hope.
1165 * Remember, these are still headerless SKBs at this point.
1167 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1168 unsigned int mss_now
, gfp_t gfp
)
1170 struct tcp_sock
*tp
= tcp_sk(sk
);
1171 struct sk_buff
*buff
;
1172 int nsize
, old_factor
;
1176 if (WARN_ON(len
> skb
->len
))
1179 nsize
= skb_headlen(skb
) - len
;
1183 if (skb_unclone(skb
, gfp
))
1186 /* Get a new skb... force flag on. */
1187 buff
= sk_stream_alloc_skb(sk
, nsize
, gfp
, true);
1189 return -ENOMEM
; /* We'll just try again later. */
1191 sk
->sk_wmem_queued
+= buff
->truesize
;
1192 sk_mem_charge(sk
, buff
->truesize
);
1193 nlen
= skb
->len
- len
- nsize
;
1194 buff
->truesize
+= nlen
;
1195 skb
->truesize
-= nlen
;
1197 /* Correct the sequence numbers. */
1198 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1199 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1200 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1202 /* PSH and FIN should only be set in the second packet. */
1203 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1204 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1205 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1206 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1207 tcp_skb_fragment_eor(skb
, buff
);
1209 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1210 /* Copy and checksum data tail into the new buffer. */
1211 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1212 skb_put(buff
, nsize
),
1217 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1219 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1220 skb_split(skb
, buff
, len
);
1223 buff
->ip_summed
= skb
->ip_summed
;
1225 buff
->tstamp
= skb
->tstamp
;
1226 tcp_fragment_tstamp(skb
, buff
);
1228 old_factor
= tcp_skb_pcount(skb
);
1230 /* Fix up tso_factor for both original and new SKB. */
1231 tcp_set_skb_tso_segs(skb
, mss_now
);
1232 tcp_set_skb_tso_segs(buff
, mss_now
);
1234 /* Update delivered info for the new segment */
1235 TCP_SKB_CB(buff
)->tx
= TCP_SKB_CB(skb
)->tx
;
1237 /* If this packet has been sent out already, we must
1238 * adjust the various packet counters.
1240 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1241 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1242 tcp_skb_pcount(buff
);
1245 tcp_adjust_pcount(sk
, skb
, diff
);
1248 /* Link BUFF into the send queue. */
1249 __skb_header_release(buff
);
1250 tcp_insert_write_queue_after(skb
, buff
, sk
);
1255 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1256 * eventually). The difference is that pulled data not copied, but
1257 * immediately discarded.
1259 static void __pskb_trim_head(struct sk_buff
*skb
, int len
)
1261 struct skb_shared_info
*shinfo
;
1264 eat
= min_t(int, len
, skb_headlen(skb
));
1266 __skb_pull(skb
, eat
);
1273 shinfo
= skb_shinfo(skb
);
1274 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1275 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1278 skb_frag_unref(skb
, i
);
1281 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1283 shinfo
->frags
[k
].page_offset
+= eat
;
1284 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1290 shinfo
->nr_frags
= k
;
1292 skb_reset_tail_pointer(skb
);
1293 skb
->data_len
-= len
;
1294 skb
->len
= skb
->data_len
;
1297 /* Remove acked data from a packet in the transmit queue. */
1298 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1300 if (skb_unclone(skb
, GFP_ATOMIC
))
1303 __pskb_trim_head(skb
, len
);
1305 TCP_SKB_CB(skb
)->seq
+= len
;
1306 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1308 skb
->truesize
-= len
;
1309 sk
->sk_wmem_queued
-= len
;
1310 sk_mem_uncharge(sk
, len
);
1311 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1313 /* Any change of skb->len requires recalculation of tso factor. */
1314 if (tcp_skb_pcount(skb
) > 1)
1315 tcp_set_skb_tso_segs(skb
, tcp_skb_mss(skb
));
1320 /* Calculate MSS not accounting any TCP options. */
1321 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1323 const struct tcp_sock
*tp
= tcp_sk(sk
);
1324 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1327 /* Calculate base mss without TCP options:
1328 It is MMS_S - sizeof(tcphdr) of rfc1122
1330 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1332 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1333 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1334 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1336 if (dst
&& dst_allfrag(dst
))
1337 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1340 /* Clamp it (mss_clamp does not include tcp options) */
1341 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1342 mss_now
= tp
->rx_opt
.mss_clamp
;
1344 /* Now subtract optional transport overhead */
1345 mss_now
-= icsk
->icsk_ext_hdr_len
;
1347 /* Then reserve room for full set of TCP options and 8 bytes of data */
1353 /* Calculate MSS. Not accounting for SACKs here. */
1354 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1356 /* Subtract TCP options size, not including SACKs */
1357 return __tcp_mtu_to_mss(sk
, pmtu
) -
1358 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1361 /* Inverse of above */
1362 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1364 const struct tcp_sock
*tp
= tcp_sk(sk
);
1365 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1369 tp
->tcp_header_len
+
1370 icsk
->icsk_ext_hdr_len
+
1371 icsk
->icsk_af_ops
->net_header_len
;
1373 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1374 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1375 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1377 if (dst
&& dst_allfrag(dst
))
1378 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1382 EXPORT_SYMBOL(tcp_mss_to_mtu
);
1384 /* MTU probing init per socket */
1385 void tcp_mtup_init(struct sock
*sk
)
1387 struct tcp_sock
*tp
= tcp_sk(sk
);
1388 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1389 struct net
*net
= sock_net(sk
);
1391 icsk
->icsk_mtup
.enabled
= net
->ipv4
.sysctl_tcp_mtu_probing
> 1;
1392 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1393 icsk
->icsk_af_ops
->net_header_len
;
1394 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, net
->ipv4
.sysctl_tcp_base_mss
);
1395 icsk
->icsk_mtup
.probe_size
= 0;
1396 if (icsk
->icsk_mtup
.enabled
)
1397 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1399 EXPORT_SYMBOL(tcp_mtup_init
);
1401 /* This function synchronize snd mss to current pmtu/exthdr set.
1403 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1404 for TCP options, but includes only bare TCP header.
1406 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1407 It is minimum of user_mss and mss received with SYN.
1408 It also does not include TCP options.
1410 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1412 tp->mss_cache is current effective sending mss, including
1413 all tcp options except for SACKs. It is evaluated,
1414 taking into account current pmtu, but never exceeds
1415 tp->rx_opt.mss_clamp.
1417 NOTE1. rfc1122 clearly states that advertised MSS
1418 DOES NOT include either tcp or ip options.
1420 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1421 are READ ONLY outside this function. --ANK (980731)
1423 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1425 struct tcp_sock
*tp
= tcp_sk(sk
);
1426 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1429 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1430 icsk
->icsk_mtup
.search_high
= pmtu
;
1432 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1433 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1435 /* And store cached results */
1436 icsk
->icsk_pmtu_cookie
= pmtu
;
1437 if (icsk
->icsk_mtup
.enabled
)
1438 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1439 tp
->mss_cache
= mss_now
;
1443 EXPORT_SYMBOL(tcp_sync_mss
);
1445 /* Compute the current effective MSS, taking SACKs and IP options,
1446 * and even PMTU discovery events into account.
1448 unsigned int tcp_current_mss(struct sock
*sk
)
1450 const struct tcp_sock
*tp
= tcp_sk(sk
);
1451 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1453 unsigned int header_len
;
1454 struct tcp_out_options opts
;
1455 struct tcp_md5sig_key
*md5
;
1457 mss_now
= tp
->mss_cache
;
1460 u32 mtu
= dst_mtu(dst
);
1461 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1462 mss_now
= tcp_sync_mss(sk
, mtu
);
1465 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1466 sizeof(struct tcphdr
);
1467 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1468 * some common options. If this is an odd packet (because we have SACK
1469 * blocks etc) then our calculated header_len will be different, and
1470 * we have to adjust mss_now correspondingly */
1471 if (header_len
!= tp
->tcp_header_len
) {
1472 int delta
= (int) header_len
- tp
->tcp_header_len
;
1479 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1480 * As additional protections, we do not touch cwnd in retransmission phases,
1481 * and if application hit its sndbuf limit recently.
1483 static void tcp_cwnd_application_limited(struct sock
*sk
)
1485 struct tcp_sock
*tp
= tcp_sk(sk
);
1487 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1488 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1489 /* Limited by application or receiver window. */
1490 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1491 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1492 if (win_used
< tp
->snd_cwnd
) {
1493 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1494 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
1496 tp
->snd_cwnd_used
= 0;
1498 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1501 static void tcp_cwnd_validate(struct sock
*sk
, bool is_cwnd_limited
)
1503 struct tcp_sock
*tp
= tcp_sk(sk
);
1505 /* Track the maximum number of outstanding packets in each
1506 * window, and remember whether we were cwnd-limited then.
1508 if (!before(tp
->snd_una
, tp
->max_packets_seq
) ||
1509 tp
->packets_out
> tp
->max_packets_out
) {
1510 tp
->max_packets_out
= tp
->packets_out
;
1511 tp
->max_packets_seq
= tp
->snd_nxt
;
1512 tp
->is_cwnd_limited
= is_cwnd_limited
;
1515 if (tcp_is_cwnd_limited(sk
)) {
1516 /* Network is feed fully. */
1517 tp
->snd_cwnd_used
= 0;
1518 tp
->snd_cwnd_stamp
= tcp_time_stamp
;
1520 /* Network starves. */
1521 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1522 tp
->snd_cwnd_used
= tp
->packets_out
;
1524 if (sysctl_tcp_slow_start_after_idle
&&
1525 (s32
)(tcp_time_stamp
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
)
1526 tcp_cwnd_application_limited(sk
);
1528 /* The following conditions together indicate the starvation
1529 * is caused by insufficient sender buffer:
1530 * 1) just sent some data (see tcp_write_xmit)
1531 * 2) not cwnd limited (this else condition)
1532 * 3) no more data to send (null tcp_send_head )
1533 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1535 if (!tcp_send_head(sk
) && sk
->sk_socket
&&
1536 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
) &&
1537 (1 << sk
->sk_state
) & (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
1538 tcp_chrono_start(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1542 /* Minshall's variant of the Nagle send check. */
1543 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1545 return after(tp
->snd_sml
, tp
->snd_una
) &&
1546 !after(tp
->snd_sml
, tp
->snd_nxt
);
1549 /* Update snd_sml if this skb is under mss
1550 * Note that a TSO packet might end with a sub-mss segment
1551 * The test is really :
1552 * if ((skb->len % mss) != 0)
1553 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1554 * But we can avoid doing the divide again given we already have
1555 * skb_pcount = skb->len / mss_now
1557 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1558 const struct sk_buff
*skb
)
1560 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1561 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1564 /* Return false, if packet can be sent now without violation Nagle's rules:
1565 * 1. It is full sized. (provided by caller in %partial bool)
1566 * 2. Or it contains FIN. (already checked by caller)
1567 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1568 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1569 * With Minshall's modification: all sent small packets are ACKed.
1571 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1575 ((nonagle
& TCP_NAGLE_CORK
) ||
1576 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1579 /* Return how many segs we'd like on a TSO packet,
1580 * to send one TSO packet per ms
1582 u32
tcp_tso_autosize(const struct sock
*sk
, unsigned int mss_now
,
1587 bytes
= min(sk
->sk_pacing_rate
>> 10,
1588 sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
);
1590 /* Goal is to send at least one packet per ms,
1591 * not one big TSO packet every 100 ms.
1592 * This preserves ACK clocking and is consistent
1593 * with tcp_tso_should_defer() heuristic.
1595 segs
= max_t(u32
, bytes
/ mss_now
, min_tso_segs
);
1597 return min_t(u32
, segs
, sk
->sk_gso_max_segs
);
1599 EXPORT_SYMBOL(tcp_tso_autosize
);
1601 /* Return the number of segments we want in the skb we are transmitting.
1602 * See if congestion control module wants to decide; otherwise, autosize.
1604 static u32
tcp_tso_segs(struct sock
*sk
, unsigned int mss_now
)
1606 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1607 u32 tso_segs
= ca_ops
->tso_segs_goal
? ca_ops
->tso_segs_goal(sk
) : 0;
1610 tcp_tso_autosize(sk
, mss_now
, sysctl_tcp_min_tso_segs
);
1613 /* Returns the portion of skb which can be sent right away */
1614 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1615 const struct sk_buff
*skb
,
1616 unsigned int mss_now
,
1617 unsigned int max_segs
,
1620 const struct tcp_sock
*tp
= tcp_sk(sk
);
1621 u32 partial
, needed
, window
, max_len
;
1623 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1624 max_len
= mss_now
* max_segs
;
1626 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1629 needed
= min(skb
->len
, window
);
1631 if (max_len
<= needed
)
1634 partial
= needed
% mss_now
;
1635 /* If last segment is not a full MSS, check if Nagle rules allow us
1636 * to include this last segment in this skb.
1637 * Otherwise, we'll split the skb at last MSS boundary
1639 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
1640 return needed
- partial
;
1645 /* Can at least one segment of SKB be sent right now, according to the
1646 * congestion window rules? If so, return how many segments are allowed.
1648 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1649 const struct sk_buff
*skb
)
1651 u32 in_flight
, cwnd
, halfcwnd
;
1653 /* Don't be strict about the congestion window for the final FIN. */
1654 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1655 tcp_skb_pcount(skb
) == 1)
1658 in_flight
= tcp_packets_in_flight(tp
);
1659 cwnd
= tp
->snd_cwnd
;
1660 if (in_flight
>= cwnd
)
1663 /* For better scheduling, ensure we have at least
1664 * 2 GSO packets in flight.
1666 halfcwnd
= max(cwnd
>> 1, 1U);
1667 return min(halfcwnd
, cwnd
- in_flight
);
1670 /* Initialize TSO state of a skb.
1671 * This must be invoked the first time we consider transmitting
1672 * SKB onto the wire.
1674 static int tcp_init_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1676 int tso_segs
= tcp_skb_pcount(skb
);
1678 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1679 tcp_set_skb_tso_segs(skb
, mss_now
);
1680 tso_segs
= tcp_skb_pcount(skb
);
1686 /* Return true if the Nagle test allows this packet to be
1689 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1690 unsigned int cur_mss
, int nonagle
)
1692 /* Nagle rule does not apply to frames, which sit in the middle of the
1693 * write_queue (they have no chances to get new data).
1695 * This is implemented in the callers, where they modify the 'nonagle'
1696 * argument based upon the location of SKB in the send queue.
1698 if (nonagle
& TCP_NAGLE_PUSH
)
1701 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1702 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1705 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
1711 /* Does at least the first segment of SKB fit into the send window? */
1712 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1713 const struct sk_buff
*skb
,
1714 unsigned int cur_mss
)
1716 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1718 if (skb
->len
> cur_mss
)
1719 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1721 return !after(end_seq
, tcp_wnd_end(tp
));
1724 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1725 * should be put on the wire right now. If so, it returns the number of
1726 * packets allowed by the congestion window.
1728 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1729 unsigned int cur_mss
, int nonagle
)
1731 const struct tcp_sock
*tp
= tcp_sk(sk
);
1732 unsigned int cwnd_quota
;
1734 tcp_init_tso_segs(skb
, cur_mss
);
1736 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1739 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1740 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1746 /* Test if sending is allowed right now. */
1747 bool tcp_may_send_now(struct sock
*sk
)
1749 const struct tcp_sock
*tp
= tcp_sk(sk
);
1750 struct sk_buff
*skb
= tcp_send_head(sk
);
1753 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1754 (tcp_skb_is_last(sk
, skb
) ?
1755 tp
->nonagle
: TCP_NAGLE_PUSH
));
1758 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1759 * which is put after SKB on the list. It is very much like
1760 * tcp_fragment() except that it may make several kinds of assumptions
1761 * in order to speed up the splitting operation. In particular, we
1762 * know that all the data is in scatter-gather pages, and that the
1763 * packet has never been sent out before (and thus is not cloned).
1765 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1766 unsigned int mss_now
, gfp_t gfp
)
1768 struct sk_buff
*buff
;
1769 int nlen
= skb
->len
- len
;
1772 /* All of a TSO frame must be composed of paged data. */
1773 if (skb
->len
!= skb
->data_len
)
1774 return tcp_fragment(sk
, skb
, len
, mss_now
, gfp
);
1776 buff
= sk_stream_alloc_skb(sk
, 0, gfp
, true);
1777 if (unlikely(!buff
))
1780 sk
->sk_wmem_queued
+= buff
->truesize
;
1781 sk_mem_charge(sk
, buff
->truesize
);
1782 buff
->truesize
+= nlen
;
1783 skb
->truesize
-= nlen
;
1785 /* Correct the sequence numbers. */
1786 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1787 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1788 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1790 /* PSH and FIN should only be set in the second packet. */
1791 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1792 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1793 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1795 /* This packet was never sent out yet, so no SACK bits. */
1796 TCP_SKB_CB(buff
)->sacked
= 0;
1798 tcp_skb_fragment_eor(skb
, buff
);
1800 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1801 skb_split(skb
, buff
, len
);
1802 tcp_fragment_tstamp(skb
, buff
);
1804 /* Fix up tso_factor for both original and new SKB. */
1805 tcp_set_skb_tso_segs(skb
, mss_now
);
1806 tcp_set_skb_tso_segs(buff
, mss_now
);
1808 /* Link BUFF into the send queue. */
1809 __skb_header_release(buff
);
1810 tcp_insert_write_queue_after(skb
, buff
, sk
);
1815 /* Try to defer sending, if possible, in order to minimize the amount
1816 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1818 * This algorithm is from John Heffner.
1820 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
,
1821 bool *is_cwnd_limited
, u32 max_segs
)
1823 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1824 u32 age
, send_win
, cong_win
, limit
, in_flight
;
1825 struct tcp_sock
*tp
= tcp_sk(sk
);
1826 struct skb_mstamp now
;
1827 struct sk_buff
*head
;
1830 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1833 if (icsk
->icsk_ca_state
>= TCP_CA_Recovery
)
1836 /* Avoid bursty behavior by allowing defer
1837 * only if the last write was recent.
1839 if ((s32
)(tcp_time_stamp
- tp
->lsndtime
) > 0)
1842 in_flight
= tcp_packets_in_flight(tp
);
1844 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1846 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1848 /* From in_flight test above, we know that cwnd > in_flight. */
1849 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1851 limit
= min(send_win
, cong_win
);
1853 /* If a full-sized TSO skb can be sent, do it. */
1854 if (limit
>= max_segs
* tp
->mss_cache
)
1857 /* Middle in queue won't get any more data, full sendable already? */
1858 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1861 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1863 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1865 /* If at least some fraction of a window is available,
1868 chunk
/= win_divisor
;
1872 /* Different approach, try not to defer past a single
1873 * ACK. Receiver should ACK every other full sized
1874 * frame, so if we have space for more than 3 frames
1877 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1881 head
= tcp_write_queue_head(sk
);
1882 skb_mstamp_get(&now
);
1883 age
= skb_mstamp_us_delta(&now
, &head
->skb_mstamp
);
1884 /* If next ACK is likely to come too late (half srtt), do not defer */
1885 if (age
< (tp
->srtt_us
>> 4))
1888 /* Ok, it looks like it is advisable to defer. */
1890 if (cong_win
< send_win
&& cong_win
<= skb
->len
)
1891 *is_cwnd_limited
= true;
1899 static inline void tcp_mtu_check_reprobe(struct sock
*sk
)
1901 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1902 struct tcp_sock
*tp
= tcp_sk(sk
);
1903 struct net
*net
= sock_net(sk
);
1907 interval
= net
->ipv4
.sysctl_tcp_probe_interval
;
1908 delta
= tcp_time_stamp
- icsk
->icsk_mtup
.probe_timestamp
;
1909 if (unlikely(delta
>= interval
* HZ
)) {
1910 int mss
= tcp_current_mss(sk
);
1912 /* Update current search range */
1913 icsk
->icsk_mtup
.probe_size
= 0;
1914 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+
1915 sizeof(struct tcphdr
) +
1916 icsk
->icsk_af_ops
->net_header_len
;
1917 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
1919 /* Update probe time stamp */
1920 icsk
->icsk_mtup
.probe_timestamp
= tcp_time_stamp
;
1924 /* Create a new MTU probe if we are ready.
1925 * MTU probe is regularly attempting to increase the path MTU by
1926 * deliberately sending larger packets. This discovers routing
1927 * changes resulting in larger path MTUs.
1929 * Returns 0 if we should wait to probe (no cwnd available),
1930 * 1 if a probe was sent,
1933 static int tcp_mtu_probe(struct sock
*sk
)
1935 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1936 struct tcp_sock
*tp
= tcp_sk(sk
);
1937 struct sk_buff
*skb
, *nskb
, *next
;
1938 struct net
*net
= sock_net(sk
);
1945 /* Not currently probing/verifying,
1947 * have enough cwnd, and
1948 * not SACKing (the variable headers throw things off)
1950 if (likely(!icsk
->icsk_mtup
.enabled
||
1951 icsk
->icsk_mtup
.probe_size
||
1952 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
1953 tp
->snd_cwnd
< 11 ||
1954 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
))
1957 /* Use binary search for probe_size between tcp_mss_base,
1958 * and current mss_clamp. if (search_high - search_low)
1959 * smaller than a threshold, backoff from probing.
1961 mss_now
= tcp_current_mss(sk
);
1962 probe_size
= tcp_mtu_to_mss(sk
, (icsk
->icsk_mtup
.search_high
+
1963 icsk
->icsk_mtup
.search_low
) >> 1);
1964 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
1965 interval
= icsk
->icsk_mtup
.search_high
- icsk
->icsk_mtup
.search_low
;
1966 /* When misfortune happens, we are reprobing actively,
1967 * and then reprobe timer has expired. We stick with current
1968 * probing process by not resetting search range to its orignal.
1970 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
) ||
1971 interval
< net
->ipv4
.sysctl_tcp_probe_threshold
) {
1972 /* Check whether enough time has elaplased for
1973 * another round of probing.
1975 tcp_mtu_check_reprobe(sk
);
1979 /* Have enough data in the send queue to probe? */
1980 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
1983 if (tp
->snd_wnd
< size_needed
)
1985 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
1988 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1989 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
1990 if (!tcp_packets_in_flight(tp
))
1996 /* We're allowed to probe. Build it now. */
1997 nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
, false);
2000 sk
->sk_wmem_queued
+= nskb
->truesize
;
2001 sk_mem_charge(sk
, nskb
->truesize
);
2003 skb
= tcp_send_head(sk
);
2005 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
2006 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
2007 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
2008 TCP_SKB_CB(nskb
)->sacked
= 0;
2010 nskb
->ip_summed
= skb
->ip_summed
;
2012 tcp_insert_write_queue_before(nskb
, skb
, sk
);
2015 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
2016 copy
= min_t(int, skb
->len
, probe_size
- len
);
2017 if (nskb
->ip_summed
) {
2018 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
2020 __wsum csum
= skb_copy_and_csum_bits(skb
, 0,
2021 skb_put(nskb
, copy
),
2023 nskb
->csum
= csum_block_add(nskb
->csum
, csum
, len
);
2026 if (skb
->len
<= copy
) {
2027 /* We've eaten all the data from this skb.
2029 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
2030 tcp_unlink_write_queue(skb
, sk
);
2031 sk_wmem_free_skb(sk
, skb
);
2033 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
2034 ~(TCPHDR_FIN
|TCPHDR_PSH
);
2035 if (!skb_shinfo(skb
)->nr_frags
) {
2036 skb_pull(skb
, copy
);
2037 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2038 skb
->csum
= csum_partial(skb
->data
,
2041 __pskb_trim_head(skb
, copy
);
2042 tcp_set_skb_tso_segs(skb
, mss_now
);
2044 TCP_SKB_CB(skb
)->seq
+= copy
;
2049 if (len
>= probe_size
)
2052 tcp_init_tso_segs(nskb
, nskb
->len
);
2054 /* We're ready to send. If this fails, the probe will
2055 * be resegmented into mss-sized pieces by tcp_write_xmit().
2057 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
2058 /* Decrement cwnd here because we are sending
2059 * effectively two packets. */
2061 tcp_event_new_data_sent(sk
, nskb
);
2063 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
2064 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
2065 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
2073 /* TCP Small Queues :
2074 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2075 * (These limits are doubled for retransmits)
2077 * - better RTT estimation and ACK scheduling
2080 * Alas, some drivers / subsystems require a fair amount
2081 * of queued bytes to ensure line rate.
2082 * One example is wifi aggregation (802.11 AMPDU)
2084 static bool tcp_small_queue_check(struct sock
*sk
, const struct sk_buff
*skb
,
2085 unsigned int factor
)
2089 limit
= max(2 * skb
->truesize
, sk
->sk_pacing_rate
>> 10);
2090 limit
= min_t(u32
, limit
, sysctl_tcp_limit_output_bytes
);
2093 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
) {
2094 /* Always send the 1st or 2nd skb in write queue.
2095 * No need to wait for TX completion to call us back,
2096 * after softirq/tasklet schedule.
2097 * This helps when TX completions are delayed too much.
2099 if (skb
== sk
->sk_write_queue
.next
||
2100 skb
->prev
== sk
->sk_write_queue
.next
)
2103 set_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
);
2104 /* It is possible TX completion already happened
2105 * before we set TSQ_THROTTLED, so we must
2106 * test again the condition.
2108 smp_mb__after_atomic();
2109 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
)
2115 static void tcp_chrono_set(struct tcp_sock
*tp
, const enum tcp_chrono
new)
2117 const u32 now
= tcp_time_stamp
;
2119 if (tp
->chrono_type
> TCP_CHRONO_UNSPEC
)
2120 tp
->chrono_stat
[tp
->chrono_type
- 1] += now
- tp
->chrono_start
;
2121 tp
->chrono_start
= now
;
2122 tp
->chrono_type
= new;
2125 void tcp_chrono_start(struct sock
*sk
, const enum tcp_chrono type
)
2127 struct tcp_sock
*tp
= tcp_sk(sk
);
2129 /* If there are multiple conditions worthy of tracking in a
2130 * chronograph then the highest priority enum takes precedence
2131 * over the other conditions. So that if something "more interesting"
2132 * starts happening, stop the previous chrono and start a new one.
2134 if (type
> tp
->chrono_type
)
2135 tcp_chrono_set(tp
, type
);
2138 void tcp_chrono_stop(struct sock
*sk
, const enum tcp_chrono type
)
2140 struct tcp_sock
*tp
= tcp_sk(sk
);
2143 /* There are multiple conditions worthy of tracking in a
2144 * chronograph, so that the highest priority enum takes
2145 * precedence over the other conditions (see tcp_chrono_start).
2146 * If a condition stops, we only stop chrono tracking if
2147 * it's the "most interesting" or current chrono we are
2148 * tracking and starts busy chrono if we have pending data.
2150 if (tcp_write_queue_empty(sk
))
2151 tcp_chrono_set(tp
, TCP_CHRONO_UNSPEC
);
2152 else if (type
== tp
->chrono_type
)
2153 tcp_chrono_set(tp
, TCP_CHRONO_BUSY
);
2156 /* This routine writes packets to the network. It advances the
2157 * send_head. This happens as incoming acks open up the remote
2160 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2161 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2162 * account rare use of URG, this is not a big flaw.
2164 * Send at most one packet when push_one > 0. Temporarily ignore
2165 * cwnd limit to force at most one packet out when push_one == 2.
2167 * Returns true, if no segments are in flight and we have queued segments,
2168 * but cannot send anything now because of SWS or another problem.
2170 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
2171 int push_one
, gfp_t gfp
)
2173 struct tcp_sock
*tp
= tcp_sk(sk
);
2174 struct sk_buff
*skb
;
2175 unsigned int tso_segs
, sent_pkts
;
2178 bool is_cwnd_limited
= false, is_rwnd_limited
= false;
2184 /* Do MTU probing. */
2185 result
= tcp_mtu_probe(sk
);
2188 } else if (result
> 0) {
2193 max_segs
= tcp_tso_segs(sk
, mss_now
);
2194 while ((skb
= tcp_send_head(sk
))) {
2197 tso_segs
= tcp_init_tso_segs(skb
, mss_now
);
2200 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
) {
2201 /* "skb_mstamp" is used as a start point for the retransmit timer */
2202 skb_mstamp_get(&skb
->skb_mstamp
);
2203 goto repair
; /* Skip network transmission */
2206 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2209 /* Force out a loss probe pkt. */
2215 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
))) {
2216 is_rwnd_limited
= true;
2220 if (tso_segs
== 1) {
2221 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2222 (tcp_skb_is_last(sk
, skb
) ?
2223 nonagle
: TCP_NAGLE_PUSH
))))
2227 tcp_tso_should_defer(sk
, skb
, &is_cwnd_limited
,
2233 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
2234 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2240 if (skb
->len
> limit
&&
2241 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
2244 if (test_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
))
2245 clear_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
);
2246 if (tcp_small_queue_check(sk
, skb
, 0))
2249 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2253 /* Advance the send_head. This one is sent out.
2254 * This call will increment packets_out.
2256 tcp_event_new_data_sent(sk
, skb
);
2258 tcp_minshall_update(tp
, mss_now
, skb
);
2259 sent_pkts
+= tcp_skb_pcount(skb
);
2265 if (is_rwnd_limited
)
2266 tcp_chrono_start(sk
, TCP_CHRONO_RWND_LIMITED
);
2268 tcp_chrono_stop(sk
, TCP_CHRONO_RWND_LIMITED
);
2270 if (likely(sent_pkts
)) {
2271 if (tcp_in_cwnd_reduction(sk
))
2272 tp
->prr_out
+= sent_pkts
;
2274 /* Send one loss probe per tail loss episode. */
2276 tcp_schedule_loss_probe(sk
);
2277 is_cwnd_limited
|= (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
);
2278 tcp_cwnd_validate(sk
, is_cwnd_limited
);
2281 return !tp
->packets_out
&& tcp_send_head(sk
);
2284 bool tcp_schedule_loss_probe(struct sock
*sk
)
2286 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2287 struct tcp_sock
*tp
= tcp_sk(sk
);
2288 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
2289 u32 rtt
= usecs_to_jiffies(tp
->srtt_us
>> 3);
2291 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_EARLY_RETRANS
))
2293 /* No consecutive loss probes. */
2294 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
2298 /* Don't do any loss probe on a Fast Open connection before 3WHS
2301 if (tp
->fastopen_rsk
)
2304 /* TLP is only scheduled when next timer event is RTO. */
2305 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
2308 /* Schedule a loss probe in 2*RTT for SACK capable connections
2309 * in Open state, that are either limited by cwnd or application.
2311 if (sysctl_tcp_early_retrans
< 3 || !tp
->packets_out
||
2312 !tcp_is_sack(tp
) || inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
)
2315 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
2319 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2320 * for delayed ack when there's one outstanding packet. If no RTT
2321 * sample is available then probe after TCP_TIMEOUT_INIT.
2323 timeout
= rtt
<< 1 ? : TCP_TIMEOUT_INIT
;
2324 if (tp
->packets_out
== 1)
2325 timeout
= max_t(u32
, timeout
,
2326 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
2327 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
2329 /* If RTO is shorter, just schedule TLP in its place. */
2330 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
2331 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
2332 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
2333 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
2338 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2343 /* Thanks to skb fast clones, we can detect if a prior transmit of
2344 * a packet is still in a qdisc or driver queue.
2345 * In this case, there is very little point doing a retransmit !
2347 static bool skb_still_in_host_queue(const struct sock
*sk
,
2348 const struct sk_buff
*skb
)
2350 if (unlikely(skb_fclone_busy(sk
, skb
))) {
2351 NET_INC_STATS(sock_net(sk
),
2352 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
2358 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2359 * retransmit the last segment.
2361 void tcp_send_loss_probe(struct sock
*sk
)
2363 struct tcp_sock
*tp
= tcp_sk(sk
);
2364 struct sk_buff
*skb
;
2366 int mss
= tcp_current_mss(sk
);
2368 skb
= tcp_send_head(sk
);
2370 if (tcp_snd_wnd_test(tp
, skb
, mss
)) {
2371 pcount
= tp
->packets_out
;
2372 tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2373 if (tp
->packets_out
> pcount
)
2377 skb
= tcp_write_queue_prev(sk
, skb
);
2379 skb
= tcp_write_queue_tail(sk
);
2382 /* At most one outstanding TLP retransmission. */
2383 if (tp
->tlp_high_seq
)
2386 /* Retransmit last segment. */
2390 if (skb_still_in_host_queue(sk
, skb
))
2393 pcount
= tcp_skb_pcount(skb
);
2394 if (WARN_ON(!pcount
))
2397 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2398 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
,
2401 skb
= tcp_write_queue_next(sk
, skb
);
2404 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2407 if (__tcp_retransmit_skb(sk
, skb
, 1))
2410 /* Record snd_nxt for loss detection. */
2411 tp
->tlp_high_seq
= tp
->snd_nxt
;
2414 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSPROBES
);
2415 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2416 inet_csk(sk
)->icsk_pending
= 0;
2421 /* Push out any pending frames which were held back due to
2422 * TCP_CORK or attempt at coalescing tiny packets.
2423 * The socket must be locked by the caller.
2425 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2428 /* If we are closed, the bytes will have to remain here.
2429 * In time closedown will finish, we empty the write queue and
2430 * all will be happy.
2432 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2435 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2436 sk_gfp_mask(sk
, GFP_ATOMIC
)))
2437 tcp_check_probe_timer(sk
);
2440 /* Send _single_ skb sitting at the send head. This function requires
2441 * true push pending frames to setup probe timer etc.
2443 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2445 struct sk_buff
*skb
= tcp_send_head(sk
);
2447 BUG_ON(!skb
|| skb
->len
< mss_now
);
2449 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2452 /* This function returns the amount that we can raise the
2453 * usable window based on the following constraints
2455 * 1. The window can never be shrunk once it is offered (RFC 793)
2456 * 2. We limit memory per socket
2459 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2460 * RECV.NEXT + RCV.WIN fixed until:
2461 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2463 * i.e. don't raise the right edge of the window until you can raise
2464 * it at least MSS bytes.
2466 * Unfortunately, the recommended algorithm breaks header prediction,
2467 * since header prediction assumes th->window stays fixed.
2469 * Strictly speaking, keeping th->window fixed violates the receiver
2470 * side SWS prevention criteria. The problem is that under this rule
2471 * a stream of single byte packets will cause the right side of the
2472 * window to always advance by a single byte.
2474 * Of course, if the sender implements sender side SWS prevention
2475 * then this will not be a problem.
2477 * BSD seems to make the following compromise:
2479 * If the free space is less than the 1/4 of the maximum
2480 * space available and the free space is less than 1/2 mss,
2481 * then set the window to 0.
2482 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2483 * Otherwise, just prevent the window from shrinking
2484 * and from being larger than the largest representable value.
2486 * This prevents incremental opening of the window in the regime
2487 * where TCP is limited by the speed of the reader side taking
2488 * data out of the TCP receive queue. It does nothing about
2489 * those cases where the window is constrained on the sender side
2490 * because the pipeline is full.
2492 * BSD also seems to "accidentally" limit itself to windows that are a
2493 * multiple of MSS, at least until the free space gets quite small.
2494 * This would appear to be a side effect of the mbuf implementation.
2495 * Combining these two algorithms results in the observed behavior
2496 * of having a fixed window size at almost all times.
2498 * Below we obtain similar behavior by forcing the offered window to
2499 * a multiple of the mss when it is feasible to do so.
2501 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2502 * Regular options like TIMESTAMP are taken into account.
2504 u32
__tcp_select_window(struct sock
*sk
)
2506 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2507 struct tcp_sock
*tp
= tcp_sk(sk
);
2508 /* MSS for the peer's data. Previous versions used mss_clamp
2509 * here. I don't know if the value based on our guesses
2510 * of peer's MSS is better for the performance. It's more correct
2511 * but may be worse for the performance because of rcv_mss
2512 * fluctuations. --SAW 1998/11/1
2514 int mss
= icsk
->icsk_ack
.rcv_mss
;
2515 int free_space
= tcp_space(sk
);
2516 int allowed_space
= tcp_full_space(sk
);
2517 int full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2520 if (mss
> full_space
)
2523 if (free_space
< (full_space
>> 1)) {
2524 icsk
->icsk_ack
.quick
= 0;
2526 if (tcp_under_memory_pressure(sk
))
2527 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2530 /* free_space might become our new window, make sure we don't
2531 * increase it due to wscale.
2533 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2535 /* if free space is less than mss estimate, or is below 1/16th
2536 * of the maximum allowed, try to move to zero-window, else
2537 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2538 * new incoming data is dropped due to memory limits.
2539 * With large window, mss test triggers way too late in order
2540 * to announce zero window in time before rmem limit kicks in.
2542 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2546 if (free_space
> tp
->rcv_ssthresh
)
2547 free_space
= tp
->rcv_ssthresh
;
2549 /* Don't do rounding if we are using window scaling, since the
2550 * scaled window will not line up with the MSS boundary anyway.
2552 window
= tp
->rcv_wnd
;
2553 if (tp
->rx_opt
.rcv_wscale
) {
2554 window
= free_space
;
2556 /* Advertise enough space so that it won't get scaled away.
2557 * Import case: prevent zero window announcement if
2558 * 1<<rcv_wscale > mss.
2560 if (((window
>> tp
->rx_opt
.rcv_wscale
) << tp
->rx_opt
.rcv_wscale
) != window
)
2561 window
= (((window
>> tp
->rx_opt
.rcv_wscale
) + 1)
2562 << tp
->rx_opt
.rcv_wscale
);
2564 /* Get the largest window that is a nice multiple of mss.
2565 * Window clamp already applied above.
2566 * If our current window offering is within 1 mss of the
2567 * free space we just keep it. This prevents the divide
2568 * and multiply from happening most of the time.
2569 * We also don't do any window rounding when the free space
2572 if (window
<= free_space
- mss
|| window
> free_space
)
2573 window
= (free_space
/ mss
) * mss
;
2574 else if (mss
== full_space
&&
2575 free_space
> window
+ (full_space
>> 1))
2576 window
= free_space
;
2582 void tcp_skb_collapse_tstamp(struct sk_buff
*skb
,
2583 const struct sk_buff
*next_skb
)
2585 if (unlikely(tcp_has_tx_tstamp(next_skb
))) {
2586 const struct skb_shared_info
*next_shinfo
=
2587 skb_shinfo(next_skb
);
2588 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
2590 shinfo
->tx_flags
|= next_shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
2591 shinfo
->tskey
= next_shinfo
->tskey
;
2592 TCP_SKB_CB(skb
)->txstamp_ack
|=
2593 TCP_SKB_CB(next_skb
)->txstamp_ack
;
2597 /* Collapses two adjacent SKB's during retransmission. */
2598 static bool tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2600 struct tcp_sock
*tp
= tcp_sk(sk
);
2601 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2602 int skb_size
, next_skb_size
;
2604 skb_size
= skb
->len
;
2605 next_skb_size
= next_skb
->len
;
2607 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2609 if (next_skb_size
) {
2610 if (next_skb_size
<= skb_availroom(skb
))
2611 skb_copy_bits(next_skb
, 0, skb_put(skb
, next_skb_size
),
2613 else if (!skb_shift(skb
, next_skb
, next_skb_size
))
2616 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2618 tcp_unlink_write_queue(next_skb
, sk
);
2620 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2621 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2623 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2624 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2626 /* Update sequence range on original skb. */
2627 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2629 /* Merge over control information. This moves PSH/FIN etc. over */
2630 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2632 /* All done, get rid of second SKB and account for it so
2633 * packet counting does not break.
2635 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2636 TCP_SKB_CB(skb
)->eor
= TCP_SKB_CB(next_skb
)->eor
;
2638 /* changed transmit queue under us so clear hints */
2639 tcp_clear_retrans_hints_partial(tp
);
2640 if (next_skb
== tp
->retransmit_skb_hint
)
2641 tp
->retransmit_skb_hint
= skb
;
2643 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2645 tcp_skb_collapse_tstamp(skb
, next_skb
);
2647 sk_wmem_free_skb(sk
, next_skb
);
2651 /* Check if coalescing SKBs is legal. */
2652 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2654 if (tcp_skb_pcount(skb
) > 1)
2656 if (skb_cloned(skb
))
2658 if (skb
== tcp_send_head(sk
))
2660 /* Some heuristics for collapsing over SACK'd could be invented */
2661 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2667 /* Collapse packets in the retransmit queue to make to create
2668 * less packets on the wire. This is only done on retransmission.
2670 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2673 struct tcp_sock
*tp
= tcp_sk(sk
);
2674 struct sk_buff
*skb
= to
, *tmp
;
2677 if (!sysctl_tcp_retrans_collapse
)
2679 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2682 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2683 if (!tcp_can_collapse(sk
, skb
))
2686 if (!tcp_skb_can_collapse_to(to
))
2699 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2702 if (!tcp_collapse_retrans(sk
, to
))
2707 /* This retransmits one SKB. Policy decisions and retransmit queue
2708 * state updates are done by the caller. Returns non-zero if an
2709 * error occurred which prevented the send.
2711 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2713 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2714 struct tcp_sock
*tp
= tcp_sk(sk
);
2715 unsigned int cur_mss
;
2719 /* Inconclusive MTU probe */
2720 if (icsk
->icsk_mtup
.probe_size
)
2721 icsk
->icsk_mtup
.probe_size
= 0;
2723 /* Do not sent more than we queued. 1/4 is reserved for possible
2724 * copying overhead: fragmentation, tunneling, mangling etc.
2726 if (atomic_read(&sk
->sk_wmem_alloc
) >
2727 min_t(u32
, sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2),
2731 if (skb_still_in_host_queue(sk
, skb
))
2734 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2735 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2737 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2741 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2742 return -EHOSTUNREACH
; /* Routing failure or similar. */
2744 cur_mss
= tcp_current_mss(sk
);
2746 /* If receiver has shrunk his window, and skb is out of
2747 * new window, do not retransmit it. The exception is the
2748 * case, when window is shrunk to zero. In this case
2749 * our retransmit serves as a zero window probe.
2751 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2752 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2755 len
= cur_mss
* segs
;
2756 if (skb
->len
> len
) {
2757 if (tcp_fragment(sk
, skb
, len
, cur_mss
, GFP_ATOMIC
))
2758 return -ENOMEM
; /* We'll try again later. */
2760 if (skb_unclone(skb
, GFP_ATOMIC
))
2763 diff
= tcp_skb_pcount(skb
);
2764 tcp_set_skb_tso_segs(skb
, cur_mss
);
2765 diff
-= tcp_skb_pcount(skb
);
2767 tcp_adjust_pcount(sk
, skb
, diff
);
2768 if (skb
->len
< cur_mss
)
2769 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2772 /* RFC3168, section 6.1.1.1. ECN fallback */
2773 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN_ECN
) == TCPHDR_SYN_ECN
)
2774 tcp_ecn_clear_syn(sk
, skb
);
2776 /* make sure skb->data is aligned on arches that require it
2777 * and check if ack-trimming & collapsing extended the headroom
2778 * beyond what csum_start can cover.
2780 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2781 skb_headroom(skb
) >= 0xFFFF)) {
2782 struct sk_buff
*nskb
;
2784 skb_mstamp_get(&skb
->skb_mstamp
);
2785 nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
, GFP_ATOMIC
);
2786 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2789 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2793 segs
= tcp_skb_pcount(skb
);
2795 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
2796 /* Update global TCP statistics. */
2797 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
, segs
);
2798 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2799 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
2800 tp
->total_retrans
+= segs
;
2805 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2807 struct tcp_sock
*tp
= tcp_sk(sk
);
2808 int err
= __tcp_retransmit_skb(sk
, skb
, segs
);
2811 #if FASTRETRANS_DEBUG > 0
2812 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2813 net_dbg_ratelimited("retrans_out leaked\n");
2816 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2817 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2819 /* Save stamp of the first retransmit. */
2820 if (!tp
->retrans_stamp
)
2821 tp
->retrans_stamp
= tcp_skb_timestamp(skb
);
2823 } else if (err
!= -EBUSY
) {
2824 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2827 if (tp
->undo_retrans
< 0)
2828 tp
->undo_retrans
= 0;
2829 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2833 /* Check if we forward retransmits are possible in the current
2834 * window/congestion state.
2836 static bool tcp_can_forward_retransmit(struct sock
*sk
)
2838 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2839 const struct tcp_sock
*tp
= tcp_sk(sk
);
2841 /* Forward retransmissions are possible only during Recovery. */
2842 if (icsk
->icsk_ca_state
!= TCP_CA_Recovery
)
2845 /* No forward retransmissions in Reno are possible. */
2846 if (tcp_is_reno(tp
))
2849 /* Yeah, we have to make difficult choice between forward transmission
2850 * and retransmission... Both ways have their merits...
2852 * For now we do not retransmit anything, while we have some new
2853 * segments to send. In the other cases, follow rule 3 for
2854 * NextSeg() specified in RFC3517.
2857 if (tcp_may_send_now(sk
))
2863 /* This gets called after a retransmit timeout, and the initially
2864 * retransmitted data is acknowledged. It tries to continue
2865 * resending the rest of the retransmit queue, until either
2866 * we've sent it all or the congestion window limit is reached.
2867 * If doing SACK, the first ACK which comes back for a timeout
2868 * based retransmit packet might feed us FACK information again.
2869 * If so, we use it to avoid unnecessarily retransmissions.
2871 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2873 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2874 struct tcp_sock
*tp
= tcp_sk(sk
);
2875 struct sk_buff
*skb
;
2876 struct sk_buff
*hole
= NULL
;
2877 u32 max_segs
, last_lost
;
2879 int fwd_rexmitting
= 0;
2881 if (!tp
->packets_out
)
2885 tp
->retransmit_high
= tp
->snd_una
;
2887 if (tp
->retransmit_skb_hint
) {
2888 skb
= tp
->retransmit_skb_hint
;
2889 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2890 if (after(last_lost
, tp
->retransmit_high
))
2891 last_lost
= tp
->retransmit_high
;
2893 skb
= tcp_write_queue_head(sk
);
2894 last_lost
= tp
->snd_una
;
2897 max_segs
= tcp_tso_segs(sk
, tcp_current_mss(sk
));
2898 tcp_for_write_queue_from(skb
, sk
) {
2902 if (skb
== tcp_send_head(sk
))
2904 /* we could do better than to assign each time */
2906 tp
->retransmit_skb_hint
= skb
;
2908 segs
= tp
->snd_cwnd
- tcp_packets_in_flight(tp
);
2911 sacked
= TCP_SKB_CB(skb
)->sacked
;
2912 /* In case tcp_shift_skb_data() have aggregated large skbs,
2913 * we need to make sure not sending too bigs TSO packets
2915 segs
= min_t(int, segs
, max_segs
);
2917 if (fwd_rexmitting
) {
2919 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_highest_sack_seq(tp
)))
2921 mib_idx
= LINUX_MIB_TCPFORWARDRETRANS
;
2923 } else if (!before(TCP_SKB_CB(skb
)->seq
, tp
->retransmit_high
)) {
2924 tp
->retransmit_high
= last_lost
;
2925 if (!tcp_can_forward_retransmit(sk
))
2927 /* Backtrack if necessary to non-L'ed skb */
2935 } else if (!(sacked
& TCPCB_LOST
)) {
2936 if (!hole
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2941 last_lost
= TCP_SKB_CB(skb
)->end_seq
;
2942 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2943 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2945 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2948 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2951 if (tcp_small_queue_check(sk
, skb
, 1))
2954 if (tcp_retransmit_skb(sk
, skb
, segs
))
2957 NET_ADD_STATS(sock_net(sk
), mib_idx
, tcp_skb_pcount(skb
));
2959 if (tcp_in_cwnd_reduction(sk
))
2960 tp
->prr_out
+= tcp_skb_pcount(skb
);
2962 if (skb
== tcp_write_queue_head(sk
))
2963 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2964 inet_csk(sk
)->icsk_rto
,
2969 /* We allow to exceed memory limits for FIN packets to expedite
2970 * connection tear down and (memory) recovery.
2971 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2972 * or even be forced to close flow without any FIN.
2973 * In general, we want to allow one skb per socket to avoid hangs
2974 * with edge trigger epoll()
2976 void sk_forced_mem_schedule(struct sock
*sk
, int size
)
2980 if (size
<= sk
->sk_forward_alloc
)
2982 amt
= sk_mem_pages(size
);
2983 sk
->sk_forward_alloc
+= amt
* SK_MEM_QUANTUM
;
2984 sk_memory_allocated_add(sk
, amt
);
2986 if (mem_cgroup_sockets_enabled
&& sk
->sk_memcg
)
2987 mem_cgroup_charge_skmem(sk
->sk_memcg
, amt
);
2990 /* Send a FIN. The caller locks the socket for us.
2991 * We should try to send a FIN packet really hard, but eventually give up.
2993 void tcp_send_fin(struct sock
*sk
)
2995 struct sk_buff
*skb
, *tskb
= tcp_write_queue_tail(sk
);
2996 struct tcp_sock
*tp
= tcp_sk(sk
);
2998 /* Optimization, tack on the FIN if we have one skb in write queue and
2999 * this skb was not yet sent, or we are under memory pressure.
3000 * Note: in the latter case, FIN packet will be sent after a timeout,
3001 * as TCP stack thinks it has already been transmitted.
3003 if (tskb
&& (tcp_send_head(sk
) || tcp_under_memory_pressure(sk
))) {
3005 TCP_SKB_CB(tskb
)->tcp_flags
|= TCPHDR_FIN
;
3006 TCP_SKB_CB(tskb
)->end_seq
++;
3008 if (!tcp_send_head(sk
)) {
3009 /* This means tskb was already sent.
3010 * Pretend we included the FIN on previous transmit.
3011 * We need to set tp->snd_nxt to the value it would have
3012 * if FIN had been sent. This is because retransmit path
3013 * does not change tp->snd_nxt.
3019 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, sk
->sk_allocation
);
3020 if (unlikely(!skb
)) {
3025 skb_reserve(skb
, MAX_TCP_HEADER
);
3026 sk_forced_mem_schedule(sk
, skb
->truesize
);
3027 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3028 tcp_init_nondata_skb(skb
, tp
->write_seq
,
3029 TCPHDR_ACK
| TCPHDR_FIN
);
3030 tcp_queue_skb(sk
, skb
);
3032 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
), TCP_NAGLE_OFF
);
3035 /* We get here when a process closes a file descriptor (either due to
3036 * an explicit close() or as a byproduct of exit()'ing) and there
3037 * was unread data in the receive queue. This behavior is recommended
3038 * by RFC 2525, section 2.17. -DaveM
3040 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
3042 struct sk_buff
*skb
;
3044 /* NOTE: No TCP options attached and we never retransmit this. */
3045 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
3047 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3051 /* Reserve space for headers and prepare control bits. */
3052 skb_reserve(skb
, MAX_TCP_HEADER
);
3053 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
3054 TCPHDR_ACK
| TCPHDR_RST
);
3055 skb_mstamp_get(&skb
->skb_mstamp
);
3057 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
3058 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3060 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
3063 /* Send a crossed SYN-ACK during socket establishment.
3064 * WARNING: This routine must only be called when we have already sent
3065 * a SYN packet that crossed the incoming SYN that caused this routine
3066 * to get called. If this assumption fails then the initial rcv_wnd
3067 * and rcv_wscale values will not be correct.
3069 int tcp_send_synack(struct sock
*sk
)
3071 struct sk_buff
*skb
;
3073 skb
= tcp_write_queue_head(sk
);
3074 if (!skb
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
3075 pr_debug("%s: wrong queue state\n", __func__
);
3078 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
3079 if (skb_cloned(skb
)) {
3080 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
3083 tcp_unlink_write_queue(skb
, sk
);
3084 __skb_header_release(nskb
);
3085 __tcp_add_write_queue_head(sk
, nskb
);
3086 sk_wmem_free_skb(sk
, skb
);
3087 sk
->sk_wmem_queued
+= nskb
->truesize
;
3088 sk_mem_charge(sk
, nskb
->truesize
);
3092 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
3093 tcp_ecn_send_synack(sk
, skb
);
3095 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3099 * tcp_make_synack - Prepare a SYN-ACK.
3100 * sk: listener socket
3101 * dst: dst entry attached to the SYNACK
3102 * req: request_sock pointer
3104 * Allocate one skb and build a SYNACK packet.
3105 * @dst is consumed : Caller should not use it again.
3107 struct sk_buff
*tcp_make_synack(const struct sock
*sk
, struct dst_entry
*dst
,
3108 struct request_sock
*req
,
3109 struct tcp_fastopen_cookie
*foc
,
3110 enum tcp_synack_type synack_type
)
3112 struct inet_request_sock
*ireq
= inet_rsk(req
);
3113 const struct tcp_sock
*tp
= tcp_sk(sk
);
3114 struct tcp_md5sig_key
*md5
= NULL
;
3115 struct tcp_out_options opts
;
3116 struct sk_buff
*skb
;
3117 int tcp_header_size
;
3122 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
3123 if (unlikely(!skb
)) {
3127 /* Reserve space for headers. */
3128 skb_reserve(skb
, MAX_TCP_HEADER
);
3130 switch (synack_type
) {
3131 case TCP_SYNACK_NORMAL
:
3132 skb_set_owner_w(skb
, req_to_sk(req
));
3134 case TCP_SYNACK_COOKIE
:
3135 /* Under synflood, we do not attach skb to a socket,
3136 * to avoid false sharing.
3139 case TCP_SYNACK_FASTOPEN
:
3140 /* sk is a const pointer, because we want to express multiple
3141 * cpu might call us concurrently.
3142 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3144 skb_set_owner_w(skb
, (struct sock
*)sk
);
3147 skb_dst_set(skb
, dst
);
3149 mss
= dst_metric_advmss(dst
);
3150 user_mss
= READ_ONCE(tp
->rx_opt
.user_mss
);
3151 if (user_mss
&& user_mss
< mss
)
3154 memset(&opts
, 0, sizeof(opts
));
3155 #ifdef CONFIG_SYN_COOKIES
3156 if (unlikely(req
->cookie_ts
))
3157 skb
->skb_mstamp
.stamp_jiffies
= cookie_init_timestamp(req
);
3160 skb_mstamp_get(&skb
->skb_mstamp
);
3162 #ifdef CONFIG_TCP_MD5SIG
3164 md5
= tcp_rsk(req
)->af_specific
->req_md5_lookup(sk
, req_to_sk(req
));
3166 skb_set_hash(skb
, tcp_rsk(req
)->txhash
, PKT_HASH_TYPE_L4
);
3167 tcp_header_size
= tcp_synack_options(req
, mss
, skb
, &opts
, md5
, foc
) +
3170 skb_push(skb
, tcp_header_size
);
3171 skb_reset_transport_header(skb
);
3173 th
= (struct tcphdr
*)skb
->data
;
3174 memset(th
, 0, sizeof(struct tcphdr
));
3177 tcp_ecn_make_synack(req
, th
);
3178 th
->source
= htons(ireq
->ir_num
);
3179 th
->dest
= ireq
->ir_rmt_port
;
3180 /* Setting of flags are superfluous here for callers (and ECE is
3181 * not even correctly set)
3183 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
3184 TCPHDR_SYN
| TCPHDR_ACK
);
3186 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
3187 /* XXX data is queued and acked as is. No buffer/window check */
3188 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
3190 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3191 th
->window
= htons(min(req
->rsk_rcv_wnd
, 65535U));
3192 tcp_options_write((__be32
*)(th
+ 1), NULL
, &opts
);
3193 th
->doff
= (tcp_header_size
>> 2);
3194 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
3196 #ifdef CONFIG_TCP_MD5SIG
3197 /* Okay, we have all we need - do the md5 hash if needed */
3199 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
3200 md5
, req_to_sk(req
), skb
);
3204 /* Do not fool tcpdump (if any), clean our debris */
3205 skb
->tstamp
.tv64
= 0;
3208 EXPORT_SYMBOL(tcp_make_synack
);
3210 static void tcp_ca_dst_init(struct sock
*sk
, const struct dst_entry
*dst
)
3212 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3213 const struct tcp_congestion_ops
*ca
;
3214 u32 ca_key
= dst_metric(dst
, RTAX_CC_ALGO
);
3216 if (ca_key
== TCP_CA_UNSPEC
)
3220 ca
= tcp_ca_find_key(ca_key
);
3221 if (likely(ca
&& try_module_get(ca
->owner
))) {
3222 module_put(icsk
->icsk_ca_ops
->owner
);
3223 icsk
->icsk_ca_dst_locked
= tcp_ca_dst_locked(dst
);
3224 icsk
->icsk_ca_ops
= ca
;
3229 /* Do all connect socket setups that can be done AF independent. */
3230 static void tcp_connect_init(struct sock
*sk
)
3232 const struct dst_entry
*dst
= __sk_dst_get(sk
);
3233 struct tcp_sock
*tp
= tcp_sk(sk
);
3236 /* We'll fix this up when we get a response from the other end.
3237 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3239 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
3240 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
3242 #ifdef CONFIG_TCP_MD5SIG
3243 if (tp
->af_specific
->md5_lookup(sk
, sk
))
3244 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
3247 /* If user gave his TCP_MAXSEG, record it to clamp */
3248 if (tp
->rx_opt
.user_mss
)
3249 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3252 tcp_sync_mss(sk
, dst_mtu(dst
));
3254 tcp_ca_dst_init(sk
, dst
);
3256 if (!tp
->window_clamp
)
3257 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
3258 tp
->advmss
= dst_metric_advmss(dst
);
3259 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->advmss
)
3260 tp
->advmss
= tp
->rx_opt
.user_mss
;
3262 tcp_initialize_rcv_mss(sk
);
3264 /* limit the window selection if the user enforce a smaller rx buffer */
3265 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
3266 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
3267 tp
->window_clamp
= tcp_full_space(sk
);
3269 tcp_select_initial_window(tcp_full_space(sk
),
3270 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
3273 sysctl_tcp_window_scaling
,
3275 dst_metric(dst
, RTAX_INITRWND
));
3277 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
3278 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
3281 sock_reset_flag(sk
, SOCK_DONE
);
3284 tp
->snd_una
= tp
->write_seq
;
3285 tp
->snd_sml
= tp
->write_seq
;
3286 tp
->snd_up
= tp
->write_seq
;
3287 tp
->snd_nxt
= tp
->write_seq
;
3289 if (likely(!tp
->repair
))
3292 tp
->rcv_tstamp
= tcp_time_stamp
;
3293 tp
->rcv_wup
= tp
->rcv_nxt
;
3294 tp
->copied_seq
= tp
->rcv_nxt
;
3296 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
3297 inet_csk(sk
)->icsk_retransmits
= 0;
3298 tcp_clear_retrans(tp
);
3301 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
3303 struct tcp_sock
*tp
= tcp_sk(sk
);
3304 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
3306 tcb
->end_seq
+= skb
->len
;
3307 __skb_header_release(skb
);
3308 __tcp_add_write_queue_tail(sk
, skb
);
3309 sk
->sk_wmem_queued
+= skb
->truesize
;
3310 sk_mem_charge(sk
, skb
->truesize
);
3311 tp
->write_seq
= tcb
->end_seq
;
3312 tp
->packets_out
+= tcp_skb_pcount(skb
);
3315 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3316 * queue a data-only packet after the regular SYN, such that regular SYNs
3317 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3318 * only the SYN sequence, the data are retransmitted in the first ACK.
3319 * If cookie is not cached or other error occurs, falls back to send a
3320 * regular SYN with Fast Open cookie request option.
3322 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
3324 struct tcp_sock
*tp
= tcp_sk(sk
);
3325 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
3326 int syn_loss
= 0, space
, err
= 0;
3327 unsigned long last_syn_loss
= 0;
3328 struct sk_buff
*syn_data
;
3330 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
3331 tcp_fastopen_cache_get(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
,
3332 &syn_loss
, &last_syn_loss
);
3333 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3335 time_before(jiffies
, last_syn_loss
+ (60*HZ
<< syn_loss
))) {
3336 fo
->cookie
.len
= -1;
3340 if (sysctl_tcp_fastopen
& TFO_CLIENT_NO_COOKIE
)
3341 fo
->cookie
.len
= -1;
3342 else if (fo
->cookie
.len
<= 0)
3345 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3346 * user-MSS. Reserve maximum option space for middleboxes that add
3347 * private TCP options. The cost is reduced data space in SYN :(
3349 if (tp
->rx_opt
.user_mss
&& tp
->rx_opt
.user_mss
< tp
->rx_opt
.mss_clamp
)
3350 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3351 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
3352 MAX_TCP_OPTION_SPACE
;
3354 space
= min_t(size_t, space
, fo
->size
);
3356 /* limit to order-0 allocations */
3357 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
3359 syn_data
= sk_stream_alloc_skb(sk
, space
, sk
->sk_allocation
, false);
3362 syn_data
->ip_summed
= CHECKSUM_PARTIAL
;
3363 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
3365 int copied
= copy_from_iter(skb_put(syn_data
, space
), space
,
3366 &fo
->data
->msg_iter
);
3367 if (unlikely(!copied
)) {
3368 kfree_skb(syn_data
);
3371 if (copied
!= space
) {
3372 skb_trim(syn_data
, copied
);
3376 /* No more data pending in inet_wait_for_connect() */
3377 if (space
== fo
->size
)
3381 tcp_connect_queue_skb(sk
, syn_data
);
3383 tcp_chrono_start(sk
, TCP_CHRONO_BUSY
);
3385 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
3387 syn
->skb_mstamp
= syn_data
->skb_mstamp
;
3389 /* Now full SYN+DATA was cloned and sent (or not),
3390 * remove the SYN from the original skb (syn_data)
3391 * we keep in write queue in case of a retransmit, as we
3392 * also have the SYN packet (with no data) in the same queue.
3394 TCP_SKB_CB(syn_data
)->seq
++;
3395 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
3397 tp
->syn_data
= (fo
->copied
> 0);
3398 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3403 /* Send a regular SYN with Fast Open cookie request option */
3404 if (fo
->cookie
.len
> 0)
3406 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3408 tp
->syn_fastopen
= 0;
3410 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3414 /* Build a SYN and send it off. */
3415 int tcp_connect(struct sock
*sk
)
3417 struct tcp_sock
*tp
= tcp_sk(sk
);
3418 struct sk_buff
*buff
;
3421 tcp_connect_init(sk
);
3423 if (unlikely(tp
->repair
)) {
3424 tcp_finish_connect(sk
, NULL
);
3428 buff
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
, true);
3429 if (unlikely(!buff
))
3432 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3433 tp
->retrans_stamp
= tcp_time_stamp
;
3434 tcp_connect_queue_skb(sk
, buff
);
3435 tcp_ecn_send_syn(sk
, buff
);
3437 /* Send off SYN; include data in Fast Open. */
3438 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3439 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3440 if (err
== -ECONNREFUSED
)
3443 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3444 * in order to make this packet get counted in tcpOutSegs.
3446 tp
->snd_nxt
= tp
->write_seq
;
3447 tp
->pushed_seq
= tp
->write_seq
;
3448 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3450 /* Timer for repeating the SYN until an answer. */
3451 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3452 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3455 EXPORT_SYMBOL(tcp_connect
);
3457 /* Send out a delayed ack, the caller does the policy checking
3458 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3461 void tcp_send_delayed_ack(struct sock
*sk
)
3463 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3464 int ato
= icsk
->icsk_ack
.ato
;
3465 unsigned long timeout
;
3467 tcp_ca_event(sk
, CA_EVENT_DELAYED_ACK
);
3469 if (ato
> TCP_DELACK_MIN
) {
3470 const struct tcp_sock
*tp
= tcp_sk(sk
);
3471 int max_ato
= HZ
/ 2;
3473 if (icsk
->icsk_ack
.pingpong
||
3474 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3475 max_ato
= TCP_DELACK_MAX
;
3477 /* Slow path, intersegment interval is "high". */
3479 /* If some rtt estimate is known, use it to bound delayed ack.
3480 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3484 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3491 ato
= min(ato
, max_ato
);
3494 /* Stay within the limit we were given */
3495 timeout
= jiffies
+ ato
;
3497 /* Use new timeout only if there wasn't a older one earlier. */
3498 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3499 /* If delack timer was blocked or is about to expire,
3502 if (icsk
->icsk_ack
.blocked
||
3503 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3508 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3509 timeout
= icsk
->icsk_ack
.timeout
;
3511 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3512 icsk
->icsk_ack
.timeout
= timeout
;
3513 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3516 /* This routine sends an ack and also updates the window. */
3517 void tcp_send_ack(struct sock
*sk
)
3519 struct sk_buff
*buff
;
3521 /* If we have been reset, we may not send again. */
3522 if (sk
->sk_state
== TCP_CLOSE
)
3525 tcp_ca_event(sk
, CA_EVENT_NON_DELAYED_ACK
);
3527 /* We are not putting this on the write queue, so
3528 * tcp_transmit_skb() will set the ownership to this
3531 buff
= alloc_skb(MAX_TCP_HEADER
,
3532 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3533 if (unlikely(!buff
)) {
3534 inet_csk_schedule_ack(sk
);
3535 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3536 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3537 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3541 /* Reserve space for headers and prepare control bits. */
3542 skb_reserve(buff
, MAX_TCP_HEADER
);
3543 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3545 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3547 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3549 skb_set_tcp_pure_ack(buff
);
3551 /* Send it off, this clears delayed acks for us. */
3552 skb_mstamp_get(&buff
->skb_mstamp
);
3553 tcp_transmit_skb(sk
, buff
, 0, (__force gfp_t
)0);
3555 EXPORT_SYMBOL_GPL(tcp_send_ack
);
3557 /* This routine sends a packet with an out of date sequence
3558 * number. It assumes the other end will try to ack it.
3560 * Question: what should we make while urgent mode?
3561 * 4.4BSD forces sending single byte of data. We cannot send
3562 * out of window data, because we have SND.NXT==SND.MAX...
3564 * Current solution: to send TWO zero-length segments in urgent mode:
3565 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3566 * out-of-date with SND.UNA-1 to probe window.
3568 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
, int mib
)
3570 struct tcp_sock
*tp
= tcp_sk(sk
);
3571 struct sk_buff
*skb
;
3573 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3574 skb
= alloc_skb(MAX_TCP_HEADER
,
3575 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3579 /* Reserve space for headers and set control bits. */
3580 skb_reserve(skb
, MAX_TCP_HEADER
);
3581 /* Use a previous sequence. This should cause the other
3582 * end to send an ack. Don't queue or clone SKB, just
3585 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3586 skb_mstamp_get(&skb
->skb_mstamp
);
3587 NET_INC_STATS(sock_net(sk
), mib
);
3588 return tcp_transmit_skb(sk
, skb
, 0, (__force gfp_t
)0);
3591 void tcp_send_window_probe(struct sock
*sk
)
3593 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3594 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3595 tcp_xmit_probe_skb(sk
, 0, LINUX_MIB_TCPWINPROBE
);
3599 /* Initiate keepalive or window probe from timer. */
3600 int tcp_write_wakeup(struct sock
*sk
, int mib
)
3602 struct tcp_sock
*tp
= tcp_sk(sk
);
3603 struct sk_buff
*skb
;
3605 if (sk
->sk_state
== TCP_CLOSE
)
3608 skb
= tcp_send_head(sk
);
3609 if (skb
&& before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3611 unsigned int mss
= tcp_current_mss(sk
);
3612 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3614 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3615 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3617 /* We are probing the opening of a window
3618 * but the window size is != 0
3619 * must have been a result SWS avoidance ( sender )
3621 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3623 seg_size
= min(seg_size
, mss
);
3624 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3625 if (tcp_fragment(sk
, skb
, seg_size
, mss
, GFP_ATOMIC
))
3627 } else if (!tcp_skb_pcount(skb
))
3628 tcp_set_skb_tso_segs(skb
, mss
);
3630 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3631 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3633 tcp_event_new_data_sent(sk
, skb
);
3636 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3637 tcp_xmit_probe_skb(sk
, 1, mib
);
3638 return tcp_xmit_probe_skb(sk
, 0, mib
);
3642 /* A window probe timeout has occurred. If window is not closed send
3643 * a partial packet else a zero probe.
3645 void tcp_send_probe0(struct sock
*sk
)
3647 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3648 struct tcp_sock
*tp
= tcp_sk(sk
);
3649 struct net
*net
= sock_net(sk
);
3650 unsigned long probe_max
;
3653 err
= tcp_write_wakeup(sk
, LINUX_MIB_TCPWINPROBE
);
3655 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3656 /* Cancel probe timer, if it is not required. */
3657 icsk
->icsk_probes_out
= 0;
3658 icsk
->icsk_backoff
= 0;
3663 if (icsk
->icsk_backoff
< net
->ipv4
.sysctl_tcp_retries2
)
3664 icsk
->icsk_backoff
++;
3665 icsk
->icsk_probes_out
++;
3666 probe_max
= TCP_RTO_MAX
;
3668 /* If packet was not sent due to local congestion,
3669 * do not backoff and do not remember icsk_probes_out.
3670 * Let local senders to fight for local resources.
3672 * Use accumulated backoff yet.
3674 if (!icsk
->icsk_probes_out
)
3675 icsk
->icsk_probes_out
= 1;
3676 probe_max
= TCP_RESOURCE_PROBE_INTERVAL
;
3678 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3679 tcp_probe0_when(sk
, probe_max
),
3683 int tcp_rtx_synack(const struct sock
*sk
, struct request_sock
*req
)
3685 const struct tcp_request_sock_ops
*af_ops
= tcp_rsk(req
)->af_specific
;
3689 tcp_rsk(req
)->txhash
= net_tx_rndhash();
3690 res
= af_ops
->send_synack(sk
, NULL
, &fl
, req
, NULL
, TCP_SYNACK_NORMAL
);
3692 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
3693 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
3694 if (unlikely(tcp_passive_fastopen(sk
)))
3695 tcp_sk(sk
)->total_retrans
++;
3699 EXPORT_SYMBOL(tcp_rtx_synack
);