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_LOSS_PROBE
)
82 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
,
86 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
87 * window scaling factor due to loss of precision.
88 * If window has been shrunk, what should we make? It is not clear at all.
89 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
90 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
91 * invalid. OK, let's make this for now:
93 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
95 const struct tcp_sock
*tp
= tcp_sk(sk
);
97 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
) ||
98 (tp
->rx_opt
.wscale_ok
&&
99 ((tp
->snd_nxt
- tcp_wnd_end(tp
)) < (1 << tp
->rx_opt
.rcv_wscale
))))
102 return tcp_wnd_end(tp
);
105 /* Calculate mss to advertise in SYN segment.
106 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
108 * 1. It is independent of path mtu.
109 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
110 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
111 * attached devices, because some buggy hosts are confused by
113 * 4. We do not make 3, we advertise MSS, calculated from first
114 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
115 * This may be overridden via information stored in routing table.
116 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
117 * probably even Jumbo".
119 static __u16
tcp_advertise_mss(struct sock
*sk
)
121 struct tcp_sock
*tp
= tcp_sk(sk
);
122 const struct dst_entry
*dst
= __sk_dst_get(sk
);
123 int mss
= tp
->advmss
;
126 unsigned int metric
= dst_metric_advmss(dst
);
137 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
138 * This is the first part of cwnd validation mechanism.
140 void tcp_cwnd_restart(struct sock
*sk
, s32 delta
)
142 struct tcp_sock
*tp
= tcp_sk(sk
);
143 u32 restart_cwnd
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
144 u32 cwnd
= tp
->snd_cwnd
;
146 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
148 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
149 restart_cwnd
= min(restart_cwnd
, cwnd
);
151 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
153 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
154 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
155 tp
->snd_cwnd_used
= 0;
158 /* Congestion state accounting after a packet has been sent. */
159 static void tcp_event_data_sent(struct tcp_sock
*tp
,
162 struct inet_connection_sock
*icsk
= inet_csk(sk
);
163 const u32 now
= tcp_jiffies32
;
165 if (tcp_packets_in_flight(tp
) == 0)
166 tcp_ca_event(sk
, CA_EVENT_TX_START
);
170 /* If it is a reply for ato after last received
171 * packet, enter pingpong mode.
173 if ((u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
174 icsk
->icsk_ack
.pingpong
= 1;
177 /* Account for an ACK we sent. */
178 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
)
180 tcp_dec_quickack_mode(sk
, pkts
);
181 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
185 u32
tcp_default_init_rwnd(u32 mss
)
187 /* Initial receive window should be twice of TCP_INIT_CWND to
188 * enable proper sending of new unsent data during fast recovery
189 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
190 * limit when mss is larger than 1460.
192 u32 init_rwnd
= TCP_INIT_CWND
* 2;
195 init_rwnd
= max((1460 * init_rwnd
) / mss
, 2U);
199 /* Determine a window scaling and initial window to offer.
200 * Based on the assumption that the given amount of space
201 * will be offered. Store the results in the tp structure.
202 * NOTE: for smooth operation initial space offering should
203 * be a multiple of mss if possible. We assume here that mss >= 1.
204 * This MUST be enforced by all callers.
206 void tcp_select_initial_window(int __space
, __u32 mss
,
207 __u32
*rcv_wnd
, __u32
*window_clamp
,
208 int wscale_ok
, __u8
*rcv_wscale
,
211 unsigned int space
= (__space
< 0 ? 0 : __space
);
213 /* If no clamp set the clamp to the max possible scaled window */
214 if (*window_clamp
== 0)
215 (*window_clamp
) = (U16_MAX
<< TCP_MAX_WSCALE
);
216 space
= min(*window_clamp
, space
);
218 /* Quantize space offering to a multiple of mss if possible. */
220 space
= rounddown(space
, mss
);
222 /* NOTE: offering an initial window larger than 32767
223 * will break some buggy TCP stacks. If the admin tells us
224 * it is likely we could be speaking with such a buggy stack
225 * we will truncate our initial window offering to 32K-1
226 * unless the remote has sent us a window scaling option,
227 * which we interpret as a sign the remote TCP is not
228 * misinterpreting the window field as a signed quantity.
230 if (sysctl_tcp_workaround_signed_windows
)
231 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
237 /* Set window scaling on max possible window */
238 space
= max_t(u32
, space
, sysctl_tcp_rmem
[2]);
239 space
= max_t(u32
, space
, sysctl_rmem_max
);
240 space
= min_t(u32
, space
, *window_clamp
);
241 while (space
> U16_MAX
&& (*rcv_wscale
) < TCP_MAX_WSCALE
) {
247 if (mss
> (1 << *rcv_wscale
)) {
248 if (!init_rcv_wnd
) /* Use default unless specified otherwise */
249 init_rcv_wnd
= tcp_default_init_rwnd(mss
);
250 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
253 /* Set the clamp no higher than max representable value */
254 (*window_clamp
) = min_t(__u32
, U16_MAX
<< (*rcv_wscale
), *window_clamp
);
256 EXPORT_SYMBOL(tcp_select_initial_window
);
258 /* Chose a new window to advertise, update state in tcp_sock for the
259 * socket, and return result with RFC1323 scaling applied. The return
260 * value can be stuffed directly into th->window for an outgoing
263 static u16
tcp_select_window(struct sock
*sk
)
265 struct tcp_sock
*tp
= tcp_sk(sk
);
266 u32 old_win
= tp
->rcv_wnd
;
267 u32 cur_win
= tcp_receive_window(tp
);
268 u32 new_win
= __tcp_select_window(sk
);
270 /* Never shrink the offered window */
271 if (new_win
< cur_win
) {
272 /* Danger Will Robinson!
273 * Don't update rcv_wup/rcv_wnd here or else
274 * we will not be able to advertise a zero
275 * window in time. --DaveM
277 * Relax Will Robinson.
280 NET_INC_STATS(sock_net(sk
),
281 LINUX_MIB_TCPWANTZEROWINDOWADV
);
282 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
284 tp
->rcv_wnd
= new_win
;
285 tp
->rcv_wup
= tp
->rcv_nxt
;
287 /* Make sure we do not exceed the maximum possible
290 if (!tp
->rx_opt
.rcv_wscale
&& sysctl_tcp_workaround_signed_windows
)
291 new_win
= min(new_win
, MAX_TCP_WINDOW
);
293 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
295 /* RFC1323 scaling applied */
296 new_win
>>= tp
->rx_opt
.rcv_wscale
;
298 /* If we advertise zero window, disable fast path. */
302 NET_INC_STATS(sock_net(sk
),
303 LINUX_MIB_TCPTOZEROWINDOWADV
);
304 } else if (old_win
== 0) {
305 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFROMZEROWINDOWADV
);
311 /* Packet ECN state for a SYN-ACK */
312 static void tcp_ecn_send_synack(struct sock
*sk
, struct sk_buff
*skb
)
314 const struct tcp_sock
*tp
= tcp_sk(sk
);
316 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
317 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
318 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
319 else if (tcp_ca_needs_ecn(sk
))
323 /* Packet ECN state for a SYN. */
324 static void tcp_ecn_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
326 struct tcp_sock
*tp
= tcp_sk(sk
);
327 bool use_ecn
= sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1 ||
328 tcp_ca_needs_ecn(sk
);
331 const struct dst_entry
*dst
= __sk_dst_get(sk
);
333 if (dst
&& dst_feature(dst
, RTAX_FEATURE_ECN
))
340 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
341 tp
->ecn_flags
= TCP_ECN_OK
;
342 if (tcp_ca_needs_ecn(sk
))
347 static void tcp_ecn_clear_syn(struct sock
*sk
, struct sk_buff
*skb
)
349 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn_fallback
)
350 /* tp->ecn_flags are cleared at a later point in time when
351 * SYN ACK is ultimatively being received.
353 TCP_SKB_CB(skb
)->tcp_flags
&= ~(TCPHDR_ECE
| TCPHDR_CWR
);
357 tcp_ecn_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
359 if (inet_rsk(req
)->ecn_ok
)
363 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
366 static void tcp_ecn_send(struct sock
*sk
, struct sk_buff
*skb
,
367 struct tcphdr
*th
, int tcp_header_len
)
369 struct tcp_sock
*tp
= tcp_sk(sk
);
371 if (tp
->ecn_flags
& TCP_ECN_OK
) {
372 /* Not-retransmitted data segment: set ECT and inject CWR. */
373 if (skb
->len
!= tcp_header_len
&&
374 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
376 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
377 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
379 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
381 } else if (!tcp_ca_needs_ecn(sk
)) {
382 /* ACK or retransmitted segment: clear ECT|CE */
383 INET_ECN_dontxmit(sk
);
385 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
390 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
391 * auto increment end seqno.
393 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
395 skb
->ip_summed
= CHECKSUM_PARTIAL
;
398 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
399 TCP_SKB_CB(skb
)->sacked
= 0;
401 tcp_skb_pcount_set(skb
, 1);
403 TCP_SKB_CB(skb
)->seq
= seq
;
404 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
406 TCP_SKB_CB(skb
)->end_seq
= seq
;
409 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
411 return tp
->snd_una
!= tp
->snd_up
;
414 #define OPTION_SACK_ADVERTISE (1 << 0)
415 #define OPTION_TS (1 << 1)
416 #define OPTION_MD5 (1 << 2)
417 #define OPTION_WSCALE (1 << 3)
418 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
420 struct tcp_out_options
{
421 u16 options
; /* bit field of OPTION_* */
422 u16 mss
; /* 0 to disable */
423 u8 ws
; /* window scale, 0 to disable */
424 u8 num_sack_blocks
; /* number of SACK blocks to include */
425 u8 hash_size
; /* bytes in hash_location */
426 __u8
*hash_location
; /* temporary pointer, overloaded */
427 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
428 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
431 /* Write previously computed TCP options to the packet.
433 * Beware: Something in the Internet is very sensitive to the ordering of
434 * TCP options, we learned this through the hard way, so be careful here.
435 * Luckily we can at least blame others for their non-compliance but from
436 * inter-operability perspective it seems that we're somewhat stuck with
437 * the ordering which we have been using if we want to keep working with
438 * those broken things (not that it currently hurts anybody as there isn't
439 * particular reason why the ordering would need to be changed).
441 * At least SACK_PERM as the first option is known to lead to a disaster
442 * (but it may well be that other scenarios fail similarly).
444 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
445 struct tcp_out_options
*opts
)
447 u16 options
= opts
->options
; /* mungable copy */
449 if (unlikely(OPTION_MD5
& options
)) {
450 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
451 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
452 /* overload cookie hash location */
453 opts
->hash_location
= (__u8
*)ptr
;
457 if (unlikely(opts
->mss
)) {
458 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
459 (TCPOLEN_MSS
<< 16) |
463 if (likely(OPTION_TS
& options
)) {
464 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
465 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
466 (TCPOLEN_SACK_PERM
<< 16) |
467 (TCPOPT_TIMESTAMP
<< 8) |
469 options
&= ~OPTION_SACK_ADVERTISE
;
471 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
473 (TCPOPT_TIMESTAMP
<< 8) |
476 *ptr
++ = htonl(opts
->tsval
);
477 *ptr
++ = htonl(opts
->tsecr
);
480 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
481 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
483 (TCPOPT_SACK_PERM
<< 8) |
487 if (unlikely(OPTION_WSCALE
& options
)) {
488 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
489 (TCPOPT_WINDOW
<< 16) |
490 (TCPOLEN_WINDOW
<< 8) |
494 if (unlikely(opts
->num_sack_blocks
)) {
495 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
496 tp
->duplicate_sack
: tp
->selective_acks
;
499 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
502 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
503 TCPOLEN_SACK_PERBLOCK
)));
505 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
507 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
508 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
511 tp
->rx_opt
.dsack
= 0;
514 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
515 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
517 u32 len
; /* Fast Open option length */
520 len
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
521 *ptr
= htonl((TCPOPT_EXP
<< 24) | (len
<< 16) |
522 TCPOPT_FASTOPEN_MAGIC
);
523 p
+= TCPOLEN_EXP_FASTOPEN_BASE
;
525 len
= TCPOLEN_FASTOPEN_BASE
+ foc
->len
;
526 *p
++ = TCPOPT_FASTOPEN
;
530 memcpy(p
, foc
->val
, foc
->len
);
531 if ((len
& 3) == 2) {
532 p
[foc
->len
] = TCPOPT_NOP
;
533 p
[foc
->len
+ 1] = TCPOPT_NOP
;
535 ptr
+= (len
+ 3) >> 2;
539 /* Compute TCP options for SYN packets. This is not the final
540 * network wire format yet.
542 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
543 struct tcp_out_options
*opts
,
544 struct tcp_md5sig_key
**md5
)
546 struct tcp_sock
*tp
= tcp_sk(sk
);
547 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
548 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
550 #ifdef CONFIG_TCP_MD5SIG
551 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
553 opts
->options
|= OPTION_MD5
;
554 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
560 /* We always get an MSS option. The option bytes which will be seen in
561 * normal data packets should timestamps be used, must be in the MSS
562 * advertised. But we subtract them from tp->mss_cache so that
563 * calculations in tcp_sendmsg are simpler etc. So account for this
564 * fact here if necessary. If we don't do this correctly, as a
565 * receiver we won't recognize data packets as being full sized when we
566 * should, and thus we won't abide by the delayed ACK rules correctly.
567 * SACKs don't matter, we never delay an ACK when we have any of those
569 opts
->mss
= tcp_advertise_mss(sk
);
570 remaining
-= TCPOLEN_MSS_ALIGNED
;
572 if (likely(sysctl_tcp_timestamps
&& !*md5
)) {
573 opts
->options
|= OPTION_TS
;
574 opts
->tsval
= tcp_skb_timestamp(skb
) + tp
->tsoffset
;
575 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
576 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
578 if (likely(sysctl_tcp_window_scaling
)) {
579 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
580 opts
->options
|= OPTION_WSCALE
;
581 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
583 if (likely(sysctl_tcp_sack
)) {
584 opts
->options
|= OPTION_SACK_ADVERTISE
;
585 if (unlikely(!(OPTION_TS
& opts
->options
)))
586 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
589 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
590 u32 need
= fastopen
->cookie
.len
;
592 need
+= fastopen
->cookie
.exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
593 TCPOLEN_FASTOPEN_BASE
;
594 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
595 if (remaining
>= need
) {
596 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
597 opts
->fastopen_cookie
= &fastopen
->cookie
;
599 tp
->syn_fastopen
= 1;
600 tp
->syn_fastopen_exp
= fastopen
->cookie
.exp
? 1 : 0;
604 return MAX_TCP_OPTION_SPACE
- remaining
;
607 /* Set up TCP options for SYN-ACKs. */
608 static unsigned int tcp_synack_options(struct request_sock
*req
,
609 unsigned int mss
, struct sk_buff
*skb
,
610 struct tcp_out_options
*opts
,
611 const struct tcp_md5sig_key
*md5
,
612 struct tcp_fastopen_cookie
*foc
)
614 struct inet_request_sock
*ireq
= inet_rsk(req
);
615 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
617 #ifdef CONFIG_TCP_MD5SIG
619 opts
->options
|= OPTION_MD5
;
620 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
622 /* We can't fit any SACK blocks in a packet with MD5 + TS
623 * options. There was discussion about disabling SACK
624 * rather than TS in order to fit in better with old,
625 * buggy kernels, but that was deemed to be unnecessary.
627 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
631 /* We always send an MSS option. */
633 remaining
-= TCPOLEN_MSS_ALIGNED
;
635 if (likely(ireq
->wscale_ok
)) {
636 opts
->ws
= ireq
->rcv_wscale
;
637 opts
->options
|= OPTION_WSCALE
;
638 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
640 if (likely(ireq
->tstamp_ok
)) {
641 opts
->options
|= OPTION_TS
;
642 opts
->tsval
= tcp_skb_timestamp(skb
) + tcp_rsk(req
)->ts_off
;
643 opts
->tsecr
= req
->ts_recent
;
644 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
646 if (likely(ireq
->sack_ok
)) {
647 opts
->options
|= OPTION_SACK_ADVERTISE
;
648 if (unlikely(!ireq
->tstamp_ok
))
649 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
651 if (foc
!= NULL
&& foc
->len
>= 0) {
654 need
+= foc
->exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
655 TCPOLEN_FASTOPEN_BASE
;
656 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
657 if (remaining
>= need
) {
658 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
659 opts
->fastopen_cookie
= foc
;
664 return MAX_TCP_OPTION_SPACE
- remaining
;
667 /* Compute TCP options for ESTABLISHED sockets. This is not the
668 * final wire format yet.
670 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
671 struct tcp_out_options
*opts
,
672 struct tcp_md5sig_key
**md5
)
674 struct tcp_sock
*tp
= tcp_sk(sk
);
675 unsigned int size
= 0;
676 unsigned int eff_sacks
;
680 #ifdef CONFIG_TCP_MD5SIG
681 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
682 if (unlikely(*md5
)) {
683 opts
->options
|= OPTION_MD5
;
684 size
+= TCPOLEN_MD5SIG_ALIGNED
;
690 if (likely(tp
->rx_opt
.tstamp_ok
)) {
691 opts
->options
|= OPTION_TS
;
692 opts
->tsval
= skb
? tcp_skb_timestamp(skb
) + tp
->tsoffset
: 0;
693 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
694 size
+= TCPOLEN_TSTAMP_ALIGNED
;
697 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
698 if (unlikely(eff_sacks
)) {
699 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
700 opts
->num_sack_blocks
=
701 min_t(unsigned int, eff_sacks
,
702 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
703 TCPOLEN_SACK_PERBLOCK
);
704 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
705 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
712 /* TCP SMALL QUEUES (TSQ)
714 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
715 * to reduce RTT and bufferbloat.
716 * We do this using a special skb destructor (tcp_wfree).
718 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
719 * needs to be reallocated in a driver.
720 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
722 * Since transmit from skb destructor is forbidden, we use a tasklet
723 * to process all sockets that eventually need to send more skbs.
724 * We use one tasklet per cpu, with its own queue of sockets.
727 struct tasklet_struct tasklet
;
728 struct list_head head
; /* queue of tcp sockets */
730 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
732 static void tcp_tsq_handler(struct sock
*sk
)
734 if ((1 << sk
->sk_state
) &
735 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
736 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
)) {
737 struct tcp_sock
*tp
= tcp_sk(sk
);
739 if (tp
->lost_out
> tp
->retrans_out
&&
740 tp
->snd_cwnd
> tcp_packets_in_flight(tp
))
741 tcp_xmit_retransmit_queue(sk
);
743 tcp_write_xmit(sk
, tcp_current_mss(sk
), tp
->nonagle
,
748 * One tasklet per cpu tries to send more skbs.
749 * We run in tasklet context but need to disable irqs when
750 * transferring tsq->head because tcp_wfree() might
751 * interrupt us (non NAPI drivers)
753 static void tcp_tasklet_func(unsigned long data
)
755 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
758 struct list_head
*q
, *n
;
762 local_irq_save(flags
);
763 list_splice_init(&tsq
->head
, &list
);
764 local_irq_restore(flags
);
766 list_for_each_safe(q
, n
, &list
) {
767 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
768 list_del(&tp
->tsq_node
);
770 sk
= (struct sock
*)tp
;
771 smp_mb__before_atomic();
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 /* Note: Called under hard irq.
908 * We can not call TCP stack right away.
910 enum hrtimer_restart
tcp_pace_kick(struct hrtimer
*timer
)
912 struct tcp_sock
*tp
= container_of(timer
, struct tcp_sock
, pacing_timer
);
913 struct sock
*sk
= (struct sock
*)tp
;
914 unsigned long nval
, oval
;
916 for (oval
= READ_ONCE(sk
->sk_tsq_flags
);; oval
= nval
) {
917 struct tsq_tasklet
*tsq
;
920 if (oval
& TSQF_QUEUED
)
923 nval
= (oval
& ~TSQF_THROTTLED
) | TSQF_QUEUED
| TCPF_TSQ_DEFERRED
;
924 nval
= cmpxchg(&sk
->sk_tsq_flags
, oval
, nval
);
928 if (!atomic_inc_not_zero(&sk
->sk_wmem_alloc
))
930 /* queue this socket to tasklet queue */
931 tsq
= this_cpu_ptr(&tsq_tasklet
);
932 empty
= list_empty(&tsq
->head
);
933 list_add(&tp
->tsq_node
, &tsq
->head
);
935 tasklet_schedule(&tsq
->tasklet
);
938 return HRTIMER_NORESTART
;
941 /* BBR congestion control needs pacing.
942 * Same remark for SO_MAX_PACING_RATE.
943 * sch_fq packet scheduler is efficiently handling pacing,
944 * but is not always installed/used.
945 * Return true if TCP stack should pace packets itself.
947 static bool tcp_needs_internal_pacing(const struct sock
*sk
)
949 return smp_load_acquire(&sk
->sk_pacing_status
) == SK_PACING_NEEDED
;
952 static void tcp_internal_pacing(struct sock
*sk
, const struct sk_buff
*skb
)
957 if (!tcp_needs_internal_pacing(sk
))
959 rate
= sk
->sk_pacing_rate
;
960 if (!rate
|| rate
== ~0U)
963 /* Should account for header sizes as sch_fq does,
964 * but lets make things simple.
966 len_ns
= (u64
)skb
->len
* NSEC_PER_SEC
;
967 do_div(len_ns
, rate
);
968 hrtimer_start(&tcp_sk(sk
)->pacing_timer
,
969 ktime_add_ns(ktime_get(), len_ns
),
970 HRTIMER_MODE_ABS_PINNED
);
973 /* This routine actually transmits TCP packets queued in by
974 * tcp_do_sendmsg(). This is used by both the initial
975 * transmission and possible later retransmissions.
976 * All SKB's seen here are completely headerless. It is our
977 * job to build the TCP header, and pass the packet down to
978 * IP so it can do the same plus pass the packet off to the
981 * We are working here with either a clone of the original
982 * SKB, or a fresh unique copy made by the retransmit engine.
984 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
987 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
988 struct inet_sock
*inet
;
990 struct tcp_skb_cb
*tcb
;
991 struct tcp_out_options opts
;
992 unsigned int tcp_options_size
, tcp_header_size
;
993 struct tcp_md5sig_key
*md5
;
997 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
1000 skb
->skb_mstamp
= tp
->tcp_mstamp
;
1002 TCP_SKB_CB(skb
)->tx
.in_flight
= TCP_SKB_CB(skb
)->end_seq
1004 tcp_rate_skb_sent(sk
, skb
);
1006 if (unlikely(skb_cloned(skb
)))
1007 skb
= pskb_copy(skb
, gfp_mask
);
1009 skb
= skb_clone(skb
, gfp_mask
);
1015 tcb
= TCP_SKB_CB(skb
);
1016 memset(&opts
, 0, sizeof(opts
));
1018 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
1019 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
1021 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
1023 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
1025 /* if no packet is in qdisc/device queue, then allow XPS to select
1026 * another queue. We can be called from tcp_tsq_handler()
1027 * which holds one reference to sk_wmem_alloc.
1029 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1030 * One way to get this would be to set skb->truesize = 2 on them.
1032 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) < SKB_TRUESIZE(1);
1034 /* If we had to use memory reserve to allocate this skb,
1035 * this might cause drops if packet is looped back :
1036 * Other socket might not have SOCK_MEMALLOC.
1037 * Packets not looped back do not care about pfmemalloc.
1039 skb
->pfmemalloc
= 0;
1041 skb_push(skb
, tcp_header_size
);
1042 skb_reset_transport_header(skb
);
1046 skb
->destructor
= skb_is_tcp_pure_ack(skb
) ? __sock_wfree
: tcp_wfree
;
1047 skb_set_hash_from_sk(skb
, sk
);
1048 atomic_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
1050 skb_set_dst_pending_confirm(skb
, sk
->sk_dst_pending_confirm
);
1052 /* Build TCP header and checksum it. */
1053 th
= (struct tcphdr
*)skb
->data
;
1054 th
->source
= inet
->inet_sport
;
1055 th
->dest
= inet
->inet_dport
;
1056 th
->seq
= htonl(tcb
->seq
);
1057 th
->ack_seq
= htonl(tp
->rcv_nxt
);
1058 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
1064 /* The urg_mode check is necessary during a below snd_una win probe */
1065 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
1066 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
1067 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
1069 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
1070 th
->urg_ptr
= htons(0xFFFF);
1075 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
1076 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
1077 if (likely(!(tcb
->tcp_flags
& TCPHDR_SYN
))) {
1078 th
->window
= htons(tcp_select_window(sk
));
1079 tcp_ecn_send(sk
, skb
, th
, tcp_header_size
);
1081 /* RFC1323: The window in SYN & SYN/ACK segments
1084 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
1086 #ifdef CONFIG_TCP_MD5SIG
1087 /* Calculate the MD5 hash, as we have all we need now */
1089 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
1090 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
1095 icsk
->icsk_af_ops
->send_check(sk
, skb
);
1097 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
1098 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
1100 if (skb
->len
!= tcp_header_size
) {
1101 tcp_event_data_sent(tp
, sk
);
1102 tp
->data_segs_out
+= tcp_skb_pcount(skb
);
1103 tcp_internal_pacing(sk
, skb
);
1106 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
1107 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
1108 tcp_skb_pcount(skb
));
1110 tp
->segs_out
+= tcp_skb_pcount(skb
);
1111 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1112 skb_shinfo(skb
)->gso_segs
= tcp_skb_pcount(skb
);
1113 skb_shinfo(skb
)->gso_size
= tcp_skb_mss(skb
);
1115 /* Our usage of tstamp should remain private */
1118 /* Cleanup our debris for IP stacks */
1119 memset(skb
->cb
, 0, max(sizeof(struct inet_skb_parm
),
1120 sizeof(struct inet6_skb_parm
)));
1122 err
= icsk
->icsk_af_ops
->queue_xmit(sk
, skb
, &inet
->cork
.fl
);
1124 if (likely(err
<= 0))
1129 return net_xmit_eval(err
);
1132 /* This routine just queues the buffer for sending.
1134 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1135 * otherwise socket can stall.
1137 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1139 struct tcp_sock
*tp
= tcp_sk(sk
);
1141 /* Advance write_seq and place onto the write_queue. */
1142 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
1143 __skb_header_release(skb
);
1144 tcp_add_write_queue_tail(sk
, skb
);
1145 sk
->sk_wmem_queued
+= skb
->truesize
;
1146 sk_mem_charge(sk
, skb
->truesize
);
1149 /* Initialize TSO segments for a packet. */
1150 static void tcp_set_skb_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1152 if (skb
->len
<= mss_now
|| skb
->ip_summed
== CHECKSUM_NONE
) {
1153 /* Avoid the costly divide in the normal
1156 tcp_skb_pcount_set(skb
, 1);
1157 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1159 tcp_skb_pcount_set(skb
, DIV_ROUND_UP(skb
->len
, mss_now
));
1160 TCP_SKB_CB(skb
)->tcp_gso_size
= mss_now
;
1164 /* When a modification to fackets out becomes necessary, we need to check
1165 * skb is counted to fackets_out or not.
1167 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1170 struct tcp_sock
*tp
= tcp_sk(sk
);
1172 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1175 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1176 tp
->fackets_out
-= decr
;
1179 /* Pcount in the middle of the write queue got changed, we need to do various
1180 * tweaks to fix counters
1182 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1184 struct tcp_sock
*tp
= tcp_sk(sk
);
1186 tp
->packets_out
-= decr
;
1188 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1189 tp
->sacked_out
-= decr
;
1190 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1191 tp
->retrans_out
-= decr
;
1192 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1193 tp
->lost_out
-= decr
;
1195 /* Reno case is special. Sigh... */
1196 if (tcp_is_reno(tp
) && decr
> 0)
1197 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1199 tcp_adjust_fackets_out(sk
, skb
, decr
);
1201 if (tp
->lost_skb_hint
&&
1202 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1203 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1204 tp
->lost_cnt_hint
-= decr
;
1206 tcp_verify_left_out(tp
);
1209 static bool tcp_has_tx_tstamp(const struct sk_buff
*skb
)
1211 return TCP_SKB_CB(skb
)->txstamp_ack
||
1212 (skb_shinfo(skb
)->tx_flags
& SKBTX_ANY_TSTAMP
);
1215 static void tcp_fragment_tstamp(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1217 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1219 if (unlikely(tcp_has_tx_tstamp(skb
)) &&
1220 !before(shinfo
->tskey
, TCP_SKB_CB(skb2
)->seq
)) {
1221 struct skb_shared_info
*shinfo2
= skb_shinfo(skb2
);
1222 u8 tsflags
= shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
1224 shinfo
->tx_flags
&= ~tsflags
;
1225 shinfo2
->tx_flags
|= tsflags
;
1226 swap(shinfo
->tskey
, shinfo2
->tskey
);
1227 TCP_SKB_CB(skb2
)->txstamp_ack
= TCP_SKB_CB(skb
)->txstamp_ack
;
1228 TCP_SKB_CB(skb
)->txstamp_ack
= 0;
1232 static void tcp_skb_fragment_eor(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1234 TCP_SKB_CB(skb2
)->eor
= TCP_SKB_CB(skb
)->eor
;
1235 TCP_SKB_CB(skb
)->eor
= 0;
1238 /* Function to create two new TCP segments. Shrinks the given segment
1239 * to the specified size and appends a new segment with the rest of the
1240 * packet to the list. This won't be called frequently, I hope.
1241 * Remember, these are still headerless SKBs at this point.
1243 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1244 unsigned int mss_now
, gfp_t gfp
)
1246 struct tcp_sock
*tp
= tcp_sk(sk
);
1247 struct sk_buff
*buff
;
1248 int nsize
, old_factor
;
1252 if (WARN_ON(len
> skb
->len
))
1255 nsize
= skb_headlen(skb
) - len
;
1259 if (skb_unclone(skb
, gfp
))
1262 /* Get a new skb... force flag on. */
1263 buff
= sk_stream_alloc_skb(sk
, nsize
, gfp
, true);
1265 return -ENOMEM
; /* We'll just try again later. */
1267 sk
->sk_wmem_queued
+= buff
->truesize
;
1268 sk_mem_charge(sk
, buff
->truesize
);
1269 nlen
= skb
->len
- len
- nsize
;
1270 buff
->truesize
+= nlen
;
1271 skb
->truesize
-= nlen
;
1273 /* Correct the sequence numbers. */
1274 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1275 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1276 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1278 /* PSH and FIN should only be set in the second packet. */
1279 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1280 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1281 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1282 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1283 tcp_skb_fragment_eor(skb
, buff
);
1285 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1286 /* Copy and checksum data tail into the new buffer. */
1287 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1288 skb_put(buff
, nsize
),
1293 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1295 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1296 skb_split(skb
, buff
, len
);
1299 buff
->ip_summed
= skb
->ip_summed
;
1301 buff
->tstamp
= skb
->tstamp
;
1302 tcp_fragment_tstamp(skb
, buff
);
1304 old_factor
= tcp_skb_pcount(skb
);
1306 /* Fix up tso_factor for both original and new SKB. */
1307 tcp_set_skb_tso_segs(skb
, mss_now
);
1308 tcp_set_skb_tso_segs(buff
, mss_now
);
1310 /* Update delivered info for the new segment */
1311 TCP_SKB_CB(buff
)->tx
= TCP_SKB_CB(skb
)->tx
;
1313 /* If this packet has been sent out already, we must
1314 * adjust the various packet counters.
1316 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1317 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1318 tcp_skb_pcount(buff
);
1321 tcp_adjust_pcount(sk
, skb
, diff
);
1324 /* Link BUFF into the send queue. */
1325 __skb_header_release(buff
);
1326 tcp_insert_write_queue_after(skb
, buff
, sk
);
1331 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1332 * eventually). The difference is that pulled data not copied, but
1333 * immediately discarded.
1335 static int __pskb_trim_head(struct sk_buff
*skb
, int len
)
1337 struct skb_shared_info
*shinfo
;
1340 eat
= min_t(int, len
, skb_headlen(skb
));
1342 __skb_pull(skb
, eat
);
1349 shinfo
= skb_shinfo(skb
);
1350 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1351 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1354 skb_frag_unref(skb
, i
);
1357 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1359 shinfo
->frags
[k
].page_offset
+= eat
;
1360 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1366 shinfo
->nr_frags
= k
;
1368 skb_reset_tail_pointer(skb
);
1369 skb
->data_len
-= len
;
1370 skb
->len
= skb
->data_len
;
1374 /* Remove acked data from a packet in the transmit queue. */
1375 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1379 if (skb_unclone(skb
, GFP_ATOMIC
))
1382 delta_truesize
= __pskb_trim_head(skb
, len
);
1384 TCP_SKB_CB(skb
)->seq
+= len
;
1385 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1387 if (delta_truesize
) {
1388 skb
->truesize
-= delta_truesize
;
1389 sk
->sk_wmem_queued
-= delta_truesize
;
1390 sk_mem_uncharge(sk
, delta_truesize
);
1391 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1394 /* Any change of skb->len requires recalculation of tso factor. */
1395 if (tcp_skb_pcount(skb
) > 1)
1396 tcp_set_skb_tso_segs(skb
, tcp_skb_mss(skb
));
1401 /* Calculate MSS not accounting any TCP options. */
1402 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1404 const struct tcp_sock
*tp
= tcp_sk(sk
);
1405 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1408 /* Calculate base mss without TCP options:
1409 It is MMS_S - sizeof(tcphdr) of rfc1122
1411 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1413 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1414 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1415 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1417 if (dst
&& dst_allfrag(dst
))
1418 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1421 /* Clamp it (mss_clamp does not include tcp options) */
1422 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1423 mss_now
= tp
->rx_opt
.mss_clamp
;
1425 /* Now subtract optional transport overhead */
1426 mss_now
-= icsk
->icsk_ext_hdr_len
;
1428 /* Then reserve room for full set of TCP options and 8 bytes of data */
1434 /* Calculate MSS. Not accounting for SACKs here. */
1435 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1437 /* Subtract TCP options size, not including SACKs */
1438 return __tcp_mtu_to_mss(sk
, pmtu
) -
1439 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1442 /* Inverse of above */
1443 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1445 const struct tcp_sock
*tp
= tcp_sk(sk
);
1446 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1450 tp
->tcp_header_len
+
1451 icsk
->icsk_ext_hdr_len
+
1452 icsk
->icsk_af_ops
->net_header_len
;
1454 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1455 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1456 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1458 if (dst
&& dst_allfrag(dst
))
1459 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1463 EXPORT_SYMBOL(tcp_mss_to_mtu
);
1465 /* MTU probing init per socket */
1466 void tcp_mtup_init(struct sock
*sk
)
1468 struct tcp_sock
*tp
= tcp_sk(sk
);
1469 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1470 struct net
*net
= sock_net(sk
);
1472 icsk
->icsk_mtup
.enabled
= net
->ipv4
.sysctl_tcp_mtu_probing
> 1;
1473 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1474 icsk
->icsk_af_ops
->net_header_len
;
1475 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, net
->ipv4
.sysctl_tcp_base_mss
);
1476 icsk
->icsk_mtup
.probe_size
= 0;
1477 if (icsk
->icsk_mtup
.enabled
)
1478 icsk
->icsk_mtup
.probe_timestamp
= tcp_jiffies32
;
1480 EXPORT_SYMBOL(tcp_mtup_init
);
1482 /* This function synchronize snd mss to current pmtu/exthdr set.
1484 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1485 for TCP options, but includes only bare TCP header.
1487 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1488 It is minimum of user_mss and mss received with SYN.
1489 It also does not include TCP options.
1491 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1493 tp->mss_cache is current effective sending mss, including
1494 all tcp options except for SACKs. It is evaluated,
1495 taking into account current pmtu, but never exceeds
1496 tp->rx_opt.mss_clamp.
1498 NOTE1. rfc1122 clearly states that advertised MSS
1499 DOES NOT include either tcp or ip options.
1501 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1502 are READ ONLY outside this function. --ANK (980731)
1504 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1506 struct tcp_sock
*tp
= tcp_sk(sk
);
1507 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1510 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1511 icsk
->icsk_mtup
.search_high
= pmtu
;
1513 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1514 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1516 /* And store cached results */
1517 icsk
->icsk_pmtu_cookie
= pmtu
;
1518 if (icsk
->icsk_mtup
.enabled
)
1519 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1520 tp
->mss_cache
= mss_now
;
1524 EXPORT_SYMBOL(tcp_sync_mss
);
1526 /* Compute the current effective MSS, taking SACKs and IP options,
1527 * and even PMTU discovery events into account.
1529 unsigned int tcp_current_mss(struct sock
*sk
)
1531 const struct tcp_sock
*tp
= tcp_sk(sk
);
1532 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1534 unsigned int header_len
;
1535 struct tcp_out_options opts
;
1536 struct tcp_md5sig_key
*md5
;
1538 mss_now
= tp
->mss_cache
;
1541 u32 mtu
= dst_mtu(dst
);
1542 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1543 mss_now
= tcp_sync_mss(sk
, mtu
);
1546 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1547 sizeof(struct tcphdr
);
1548 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1549 * some common options. If this is an odd packet (because we have SACK
1550 * blocks etc) then our calculated header_len will be different, and
1551 * we have to adjust mss_now correspondingly */
1552 if (header_len
!= tp
->tcp_header_len
) {
1553 int delta
= (int) header_len
- tp
->tcp_header_len
;
1560 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1561 * As additional protections, we do not touch cwnd in retransmission phases,
1562 * and if application hit its sndbuf limit recently.
1564 static void tcp_cwnd_application_limited(struct sock
*sk
)
1566 struct tcp_sock
*tp
= tcp_sk(sk
);
1568 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1569 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1570 /* Limited by application or receiver window. */
1571 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1572 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1573 if (win_used
< tp
->snd_cwnd
) {
1574 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1575 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
1577 tp
->snd_cwnd_used
= 0;
1579 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1582 static void tcp_cwnd_validate(struct sock
*sk
, bool is_cwnd_limited
)
1584 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1585 struct tcp_sock
*tp
= tcp_sk(sk
);
1587 /* Track the maximum number of outstanding packets in each
1588 * window, and remember whether we were cwnd-limited then.
1590 if (!before(tp
->snd_una
, tp
->max_packets_seq
) ||
1591 tp
->packets_out
> tp
->max_packets_out
) {
1592 tp
->max_packets_out
= tp
->packets_out
;
1593 tp
->max_packets_seq
= tp
->snd_nxt
;
1594 tp
->is_cwnd_limited
= is_cwnd_limited
;
1597 if (tcp_is_cwnd_limited(sk
)) {
1598 /* Network is feed fully. */
1599 tp
->snd_cwnd_used
= 0;
1600 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1602 /* Network starves. */
1603 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1604 tp
->snd_cwnd_used
= tp
->packets_out
;
1606 if (sysctl_tcp_slow_start_after_idle
&&
1607 (s32
)(tcp_jiffies32
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
&&
1608 !ca_ops
->cong_control
)
1609 tcp_cwnd_application_limited(sk
);
1611 /* The following conditions together indicate the starvation
1612 * is caused by insufficient sender buffer:
1613 * 1) just sent some data (see tcp_write_xmit)
1614 * 2) not cwnd limited (this else condition)
1615 * 3) no more data to send (null tcp_send_head )
1616 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1618 if (!tcp_send_head(sk
) && sk
->sk_socket
&&
1619 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
) &&
1620 (1 << sk
->sk_state
) & (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
1621 tcp_chrono_start(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1625 /* Minshall's variant of the Nagle send check. */
1626 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1628 return after(tp
->snd_sml
, tp
->snd_una
) &&
1629 !after(tp
->snd_sml
, tp
->snd_nxt
);
1632 /* Update snd_sml if this skb is under mss
1633 * Note that a TSO packet might end with a sub-mss segment
1634 * The test is really :
1635 * if ((skb->len % mss) != 0)
1636 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1637 * But we can avoid doing the divide again given we already have
1638 * skb_pcount = skb->len / mss_now
1640 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1641 const struct sk_buff
*skb
)
1643 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1644 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1647 /* Return false, if packet can be sent now without violation Nagle's rules:
1648 * 1. It is full sized. (provided by caller in %partial bool)
1649 * 2. Or it contains FIN. (already checked by caller)
1650 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1651 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1652 * With Minshall's modification: all sent small packets are ACKed.
1654 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1658 ((nonagle
& TCP_NAGLE_CORK
) ||
1659 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1662 /* Return how many segs we'd like on a TSO packet,
1663 * to send one TSO packet per ms
1665 u32
tcp_tso_autosize(const struct sock
*sk
, unsigned int mss_now
,
1670 bytes
= min(sk
->sk_pacing_rate
>> 10,
1671 sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
);
1673 /* Goal is to send at least one packet per ms,
1674 * not one big TSO packet every 100 ms.
1675 * This preserves ACK clocking and is consistent
1676 * with tcp_tso_should_defer() heuristic.
1678 segs
= max_t(u32
, bytes
/ mss_now
, min_tso_segs
);
1680 return min_t(u32
, segs
, sk
->sk_gso_max_segs
);
1682 EXPORT_SYMBOL(tcp_tso_autosize
);
1684 /* Return the number of segments we want in the skb we are transmitting.
1685 * See if congestion control module wants to decide; otherwise, autosize.
1687 static u32
tcp_tso_segs(struct sock
*sk
, unsigned int mss_now
)
1689 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1690 u32 tso_segs
= ca_ops
->tso_segs_goal
? ca_ops
->tso_segs_goal(sk
) : 0;
1693 tcp_tso_autosize(sk
, mss_now
, sysctl_tcp_min_tso_segs
);
1696 /* Returns the portion of skb which can be sent right away */
1697 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1698 const struct sk_buff
*skb
,
1699 unsigned int mss_now
,
1700 unsigned int max_segs
,
1703 const struct tcp_sock
*tp
= tcp_sk(sk
);
1704 u32 partial
, needed
, window
, max_len
;
1706 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1707 max_len
= mss_now
* max_segs
;
1709 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1712 needed
= min(skb
->len
, window
);
1714 if (max_len
<= needed
)
1717 partial
= needed
% mss_now
;
1718 /* If last segment is not a full MSS, check if Nagle rules allow us
1719 * to include this last segment in this skb.
1720 * Otherwise, we'll split the skb at last MSS boundary
1722 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
1723 return needed
- partial
;
1728 /* Can at least one segment of SKB be sent right now, according to the
1729 * congestion window rules? If so, return how many segments are allowed.
1731 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1732 const struct sk_buff
*skb
)
1734 u32 in_flight
, cwnd
, halfcwnd
;
1736 /* Don't be strict about the congestion window for the final FIN. */
1737 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1738 tcp_skb_pcount(skb
) == 1)
1741 in_flight
= tcp_packets_in_flight(tp
);
1742 cwnd
= tp
->snd_cwnd
;
1743 if (in_flight
>= cwnd
)
1746 /* For better scheduling, ensure we have at least
1747 * 2 GSO packets in flight.
1749 halfcwnd
= max(cwnd
>> 1, 1U);
1750 return min(halfcwnd
, cwnd
- in_flight
);
1753 /* Initialize TSO state of a skb.
1754 * This must be invoked the first time we consider transmitting
1755 * SKB onto the wire.
1757 static int tcp_init_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1759 int tso_segs
= tcp_skb_pcount(skb
);
1761 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1762 tcp_set_skb_tso_segs(skb
, mss_now
);
1763 tso_segs
= tcp_skb_pcount(skb
);
1769 /* Return true if the Nagle test allows this packet to be
1772 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1773 unsigned int cur_mss
, int nonagle
)
1775 /* Nagle rule does not apply to frames, which sit in the middle of the
1776 * write_queue (they have no chances to get new data).
1778 * This is implemented in the callers, where they modify the 'nonagle'
1779 * argument based upon the location of SKB in the send queue.
1781 if (nonagle
& TCP_NAGLE_PUSH
)
1784 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1785 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1788 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
1794 /* Does at least the first segment of SKB fit into the send window? */
1795 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1796 const struct sk_buff
*skb
,
1797 unsigned int cur_mss
)
1799 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1801 if (skb
->len
> cur_mss
)
1802 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1804 return !after(end_seq
, tcp_wnd_end(tp
));
1807 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1808 * should be put on the wire right now. If so, it returns the number of
1809 * packets allowed by the congestion window.
1811 static unsigned int tcp_snd_test(const struct sock
*sk
, struct sk_buff
*skb
,
1812 unsigned int cur_mss
, int nonagle
)
1814 const struct tcp_sock
*tp
= tcp_sk(sk
);
1815 unsigned int cwnd_quota
;
1817 tcp_init_tso_segs(skb
, cur_mss
);
1819 if (!tcp_nagle_test(tp
, skb
, cur_mss
, nonagle
))
1822 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
1823 if (cwnd_quota
&& !tcp_snd_wnd_test(tp
, skb
, cur_mss
))
1829 /* Test if sending is allowed right now. */
1830 bool tcp_may_send_now(struct sock
*sk
)
1832 const struct tcp_sock
*tp
= tcp_sk(sk
);
1833 struct sk_buff
*skb
= tcp_send_head(sk
);
1836 tcp_snd_test(sk
, skb
, tcp_current_mss(sk
),
1837 (tcp_skb_is_last(sk
, skb
) ?
1838 tp
->nonagle
: TCP_NAGLE_PUSH
));
1841 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1842 * which is put after SKB on the list. It is very much like
1843 * tcp_fragment() except that it may make several kinds of assumptions
1844 * in order to speed up the splitting operation. In particular, we
1845 * know that all the data is in scatter-gather pages, and that the
1846 * packet has never been sent out before (and thus is not cloned).
1848 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1849 unsigned int mss_now
, gfp_t gfp
)
1851 struct sk_buff
*buff
;
1852 int nlen
= skb
->len
- len
;
1855 /* All of a TSO frame must be composed of paged data. */
1856 if (skb
->len
!= skb
->data_len
)
1857 return tcp_fragment(sk
, skb
, len
, mss_now
, gfp
);
1859 buff
= sk_stream_alloc_skb(sk
, 0, gfp
, true);
1860 if (unlikely(!buff
))
1863 sk
->sk_wmem_queued
+= buff
->truesize
;
1864 sk_mem_charge(sk
, buff
->truesize
);
1865 buff
->truesize
+= nlen
;
1866 skb
->truesize
-= nlen
;
1868 /* Correct the sequence numbers. */
1869 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1870 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1871 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1873 /* PSH and FIN should only be set in the second packet. */
1874 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1875 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1876 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1878 /* This packet was never sent out yet, so no SACK bits. */
1879 TCP_SKB_CB(buff
)->sacked
= 0;
1881 tcp_skb_fragment_eor(skb
, buff
);
1883 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1884 skb_split(skb
, buff
, len
);
1885 tcp_fragment_tstamp(skb
, buff
);
1887 /* Fix up tso_factor for both original and new SKB. */
1888 tcp_set_skb_tso_segs(skb
, mss_now
);
1889 tcp_set_skb_tso_segs(buff
, mss_now
);
1891 /* Link BUFF into the send queue. */
1892 __skb_header_release(buff
);
1893 tcp_insert_write_queue_after(skb
, buff
, sk
);
1898 /* Try to defer sending, if possible, in order to minimize the amount
1899 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1901 * This algorithm is from John Heffner.
1903 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
,
1904 bool *is_cwnd_limited
, u32 max_segs
)
1906 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1907 u32 age
, send_win
, cong_win
, limit
, in_flight
;
1908 struct tcp_sock
*tp
= tcp_sk(sk
);
1909 struct sk_buff
*head
;
1912 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1915 if (icsk
->icsk_ca_state
>= TCP_CA_Recovery
)
1918 /* Avoid bursty behavior by allowing defer
1919 * only if the last write was recent.
1921 if ((s32
)(tcp_jiffies32
- tp
->lsndtime
) > 0)
1924 in_flight
= tcp_packets_in_flight(tp
);
1926 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1928 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1930 /* From in_flight test above, we know that cwnd > in_flight. */
1931 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1933 limit
= min(send_win
, cong_win
);
1935 /* If a full-sized TSO skb can be sent, do it. */
1936 if (limit
>= max_segs
* tp
->mss_cache
)
1939 /* Middle in queue won't get any more data, full sendable already? */
1940 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1943 win_divisor
= ACCESS_ONCE(sysctl_tcp_tso_win_divisor
);
1945 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1947 /* If at least some fraction of a window is available,
1950 chunk
/= win_divisor
;
1954 /* Different approach, try not to defer past a single
1955 * ACK. Receiver should ACK every other full sized
1956 * frame, so if we have space for more than 3 frames
1959 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1963 head
= tcp_write_queue_head(sk
);
1965 age
= skb_mstamp_us_delta(&tp
->tcp_mstamp
, &head
->skb_mstamp
);
1966 /* If next ACK is likely to come too late (half srtt), do not defer */
1967 if (age
< (tp
->srtt_us
>> 4))
1970 /* Ok, it looks like it is advisable to defer. */
1972 if (cong_win
< send_win
&& cong_win
<= skb
->len
)
1973 *is_cwnd_limited
= true;
1981 static inline void tcp_mtu_check_reprobe(struct sock
*sk
)
1983 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1984 struct tcp_sock
*tp
= tcp_sk(sk
);
1985 struct net
*net
= sock_net(sk
);
1989 interval
= net
->ipv4
.sysctl_tcp_probe_interval
;
1990 delta
= tcp_jiffies32
- icsk
->icsk_mtup
.probe_timestamp
;
1991 if (unlikely(delta
>= interval
* HZ
)) {
1992 int mss
= tcp_current_mss(sk
);
1994 /* Update current search range */
1995 icsk
->icsk_mtup
.probe_size
= 0;
1996 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+
1997 sizeof(struct tcphdr
) +
1998 icsk
->icsk_af_ops
->net_header_len
;
1999 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
2001 /* Update probe time stamp */
2002 icsk
->icsk_mtup
.probe_timestamp
= tcp_jiffies32
;
2006 /* Create a new MTU probe if we are ready.
2007 * MTU probe is regularly attempting to increase the path MTU by
2008 * deliberately sending larger packets. This discovers routing
2009 * changes resulting in larger path MTUs.
2011 * Returns 0 if we should wait to probe (no cwnd available),
2012 * 1 if a probe was sent,
2015 static int tcp_mtu_probe(struct sock
*sk
)
2017 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2018 struct tcp_sock
*tp
= tcp_sk(sk
);
2019 struct sk_buff
*skb
, *nskb
, *next
;
2020 struct net
*net
= sock_net(sk
);
2027 /* Not currently probing/verifying,
2029 * have enough cwnd, and
2030 * not SACKing (the variable headers throw things off)
2032 if (likely(!icsk
->icsk_mtup
.enabled
||
2033 icsk
->icsk_mtup
.probe_size
||
2034 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
2035 tp
->snd_cwnd
< 11 ||
2036 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
))
2039 /* Use binary search for probe_size between tcp_mss_base,
2040 * and current mss_clamp. if (search_high - search_low)
2041 * smaller than a threshold, backoff from probing.
2043 mss_now
= tcp_current_mss(sk
);
2044 probe_size
= tcp_mtu_to_mss(sk
, (icsk
->icsk_mtup
.search_high
+
2045 icsk
->icsk_mtup
.search_low
) >> 1);
2046 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
2047 interval
= icsk
->icsk_mtup
.search_high
- icsk
->icsk_mtup
.search_low
;
2048 /* When misfortune happens, we are reprobing actively,
2049 * and then reprobe timer has expired. We stick with current
2050 * probing process by not resetting search range to its orignal.
2052 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
) ||
2053 interval
< net
->ipv4
.sysctl_tcp_probe_threshold
) {
2054 /* Check whether enough time has elaplased for
2055 * another round of probing.
2057 tcp_mtu_check_reprobe(sk
);
2061 /* Have enough data in the send queue to probe? */
2062 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
2065 if (tp
->snd_wnd
< size_needed
)
2067 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
2070 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2071 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
2072 if (!tcp_packets_in_flight(tp
))
2078 /* We're allowed to probe. Build it now. */
2079 nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
, false);
2082 sk
->sk_wmem_queued
+= nskb
->truesize
;
2083 sk_mem_charge(sk
, nskb
->truesize
);
2085 skb
= tcp_send_head(sk
);
2087 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
2088 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
2089 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
2090 TCP_SKB_CB(nskb
)->sacked
= 0;
2092 nskb
->ip_summed
= skb
->ip_summed
;
2094 tcp_insert_write_queue_before(nskb
, skb
, sk
);
2097 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
2098 copy
= min_t(int, skb
->len
, probe_size
- len
);
2099 if (nskb
->ip_summed
) {
2100 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
2102 __wsum csum
= skb_copy_and_csum_bits(skb
, 0,
2103 skb_put(nskb
, copy
),
2105 nskb
->csum
= csum_block_add(nskb
->csum
, csum
, len
);
2108 if (skb
->len
<= copy
) {
2109 /* We've eaten all the data from this skb.
2111 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
2112 tcp_unlink_write_queue(skb
, sk
);
2113 sk_wmem_free_skb(sk
, skb
);
2115 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
2116 ~(TCPHDR_FIN
|TCPHDR_PSH
);
2117 if (!skb_shinfo(skb
)->nr_frags
) {
2118 skb_pull(skb
, copy
);
2119 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2120 skb
->csum
= csum_partial(skb
->data
,
2123 __pskb_trim_head(skb
, copy
);
2124 tcp_set_skb_tso_segs(skb
, mss_now
);
2126 TCP_SKB_CB(skb
)->seq
+= copy
;
2131 if (len
>= probe_size
)
2134 tcp_init_tso_segs(nskb
, nskb
->len
);
2136 /* We're ready to send. If this fails, the probe will
2137 * be resegmented into mss-sized pieces by tcp_write_xmit().
2139 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
2140 /* Decrement cwnd here because we are sending
2141 * effectively two packets. */
2143 tcp_event_new_data_sent(sk
, nskb
);
2145 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
2146 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
2147 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
2155 static bool tcp_pacing_check(const struct sock
*sk
)
2157 return tcp_needs_internal_pacing(sk
) &&
2158 hrtimer_active(&tcp_sk(sk
)->pacing_timer
);
2161 /* TCP Small Queues :
2162 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2163 * (These limits are doubled for retransmits)
2165 * - better RTT estimation and ACK scheduling
2168 * Alas, some drivers / subsystems require a fair amount
2169 * of queued bytes to ensure line rate.
2170 * One example is wifi aggregation (802.11 AMPDU)
2172 static bool tcp_small_queue_check(struct sock
*sk
, const struct sk_buff
*skb
,
2173 unsigned int factor
)
2177 limit
= max(2 * skb
->truesize
, sk
->sk_pacing_rate
>> 10);
2178 limit
= min_t(u32
, limit
, sysctl_tcp_limit_output_bytes
);
2181 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
) {
2182 /* Always send the 1st or 2nd skb in write queue.
2183 * No need to wait for TX completion to call us back,
2184 * after softirq/tasklet schedule.
2185 * This helps when TX completions are delayed too much.
2187 if (skb
== sk
->sk_write_queue
.next
||
2188 skb
->prev
== sk
->sk_write_queue
.next
)
2191 set_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
);
2192 /* It is possible TX completion already happened
2193 * before we set TSQ_THROTTLED, so we must
2194 * test again the condition.
2196 smp_mb__after_atomic();
2197 if (atomic_read(&sk
->sk_wmem_alloc
) > limit
)
2203 static void tcp_chrono_set(struct tcp_sock
*tp
, const enum tcp_chrono
new)
2205 const u32 now
= tcp_jiffies32
;
2207 if (tp
->chrono_type
> TCP_CHRONO_UNSPEC
)
2208 tp
->chrono_stat
[tp
->chrono_type
- 1] += now
- tp
->chrono_start
;
2209 tp
->chrono_start
= now
;
2210 tp
->chrono_type
= new;
2213 void tcp_chrono_start(struct sock
*sk
, const enum tcp_chrono type
)
2215 struct tcp_sock
*tp
= tcp_sk(sk
);
2217 /* If there are multiple conditions worthy of tracking in a
2218 * chronograph then the highest priority enum takes precedence
2219 * over the other conditions. So that if something "more interesting"
2220 * starts happening, stop the previous chrono and start a new one.
2222 if (type
> tp
->chrono_type
)
2223 tcp_chrono_set(tp
, type
);
2226 void tcp_chrono_stop(struct sock
*sk
, const enum tcp_chrono type
)
2228 struct tcp_sock
*tp
= tcp_sk(sk
);
2231 /* There are multiple conditions worthy of tracking in a
2232 * chronograph, so that the highest priority enum takes
2233 * precedence over the other conditions (see tcp_chrono_start).
2234 * If a condition stops, we only stop chrono tracking if
2235 * it's the "most interesting" or current chrono we are
2236 * tracking and starts busy chrono if we have pending data.
2238 if (tcp_write_queue_empty(sk
))
2239 tcp_chrono_set(tp
, TCP_CHRONO_UNSPEC
);
2240 else if (type
== tp
->chrono_type
)
2241 tcp_chrono_set(tp
, TCP_CHRONO_BUSY
);
2244 /* This routine writes packets to the network. It advances the
2245 * send_head. This happens as incoming acks open up the remote
2248 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2249 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2250 * account rare use of URG, this is not a big flaw.
2252 * Send at most one packet when push_one > 0. Temporarily ignore
2253 * cwnd limit to force at most one packet out when push_one == 2.
2255 * Returns true, if no segments are in flight and we have queued segments,
2256 * but cannot send anything now because of SWS or another problem.
2258 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
2259 int push_one
, gfp_t gfp
)
2261 struct tcp_sock
*tp
= tcp_sk(sk
);
2262 struct sk_buff
*skb
;
2263 unsigned int tso_segs
, sent_pkts
;
2266 bool is_cwnd_limited
= false, is_rwnd_limited
= false;
2272 /* Do MTU probing. */
2273 result
= tcp_mtu_probe(sk
);
2276 } else if (result
> 0) {
2281 max_segs
= tcp_tso_segs(sk
, mss_now
);
2282 skb_mstamp_get(&tp
->tcp_mstamp
);
2283 while ((skb
= tcp_send_head(sk
))) {
2286 if (tcp_pacing_check(sk
))
2289 tso_segs
= tcp_init_tso_segs(skb
, mss_now
);
2292 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
) {
2293 /* "skb_mstamp" is used as a start point for the retransmit timer */
2294 skb
->skb_mstamp
= tp
->tcp_mstamp
;
2295 goto repair
; /* Skip network transmission */
2298 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2301 /* Force out a loss probe pkt. */
2307 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
))) {
2308 is_rwnd_limited
= true;
2312 if (tso_segs
== 1) {
2313 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2314 (tcp_skb_is_last(sk
, skb
) ?
2315 nonagle
: TCP_NAGLE_PUSH
))))
2319 tcp_tso_should_defer(sk
, skb
, &is_cwnd_limited
,
2325 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
2326 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2332 if (skb
->len
> limit
&&
2333 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
2336 if (test_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
))
2337 clear_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
);
2338 if (tcp_small_queue_check(sk
, skb
, 0))
2341 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2345 /* Advance the send_head. This one is sent out.
2346 * This call will increment packets_out.
2348 tcp_event_new_data_sent(sk
, skb
);
2350 tcp_minshall_update(tp
, mss_now
, skb
);
2351 sent_pkts
+= tcp_skb_pcount(skb
);
2357 if (is_rwnd_limited
)
2358 tcp_chrono_start(sk
, TCP_CHRONO_RWND_LIMITED
);
2360 tcp_chrono_stop(sk
, TCP_CHRONO_RWND_LIMITED
);
2362 if (likely(sent_pkts
)) {
2363 if (tcp_in_cwnd_reduction(sk
))
2364 tp
->prr_out
+= sent_pkts
;
2366 /* Send one loss probe per tail loss episode. */
2368 tcp_schedule_loss_probe(sk
);
2369 is_cwnd_limited
|= (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
);
2370 tcp_cwnd_validate(sk
, is_cwnd_limited
);
2373 return !tp
->packets_out
&& tcp_send_head(sk
);
2376 bool tcp_schedule_loss_probe(struct sock
*sk
)
2378 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2379 struct tcp_sock
*tp
= tcp_sk(sk
);
2380 u32 timeout
, tlp_time_stamp
, rto_time_stamp
;
2381 u32 rtt
= usecs_to_jiffies(tp
->srtt_us
>> 3);
2383 /* No consecutive loss probes. */
2384 if (WARN_ON(icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)) {
2388 /* Don't do any loss probe on a Fast Open connection before 3WHS
2391 if (tp
->fastopen_rsk
)
2394 /* TLP is only scheduled when next timer event is RTO. */
2395 if (icsk
->icsk_pending
!= ICSK_TIME_RETRANS
)
2398 /* Schedule a loss probe in 2*RTT for SACK capable connections
2399 * in Open state, that are either limited by cwnd or application.
2401 if ((sysctl_tcp_early_retrans
!= 3 && sysctl_tcp_early_retrans
!= 4) ||
2402 !tp
->packets_out
|| !tcp_is_sack(tp
) ||
2403 icsk
->icsk_ca_state
!= TCP_CA_Open
)
2406 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
2410 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2411 * for delayed ack when there's one outstanding packet. If no RTT
2412 * sample is available then probe after TCP_TIMEOUT_INIT.
2414 timeout
= rtt
<< 1 ? : TCP_TIMEOUT_INIT
;
2415 if (tp
->packets_out
== 1)
2416 timeout
= max_t(u32
, timeout
,
2417 (rtt
+ (rtt
>> 1) + TCP_DELACK_MAX
));
2418 timeout
= max_t(u32
, timeout
, msecs_to_jiffies(10));
2420 /* If RTO is shorter, just schedule TLP in its place. */
2421 tlp_time_stamp
= tcp_time_stamp
+ timeout
;
2422 rto_time_stamp
= (u32
)inet_csk(sk
)->icsk_timeout
;
2423 if ((s32
)(tlp_time_stamp
- rto_time_stamp
) > 0) {
2424 s32 delta
= rto_time_stamp
- tcp_time_stamp
;
2429 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2434 /* Thanks to skb fast clones, we can detect if a prior transmit of
2435 * a packet is still in a qdisc or driver queue.
2436 * In this case, there is very little point doing a retransmit !
2438 static bool skb_still_in_host_queue(const struct sock
*sk
,
2439 const struct sk_buff
*skb
)
2441 if (unlikely(skb_fclone_busy(sk
, skb
))) {
2442 NET_INC_STATS(sock_net(sk
),
2443 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
2449 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2450 * retransmit the last segment.
2452 void tcp_send_loss_probe(struct sock
*sk
)
2454 struct tcp_sock
*tp
= tcp_sk(sk
);
2455 struct sk_buff
*skb
;
2457 int mss
= tcp_current_mss(sk
);
2459 skb
= tcp_send_head(sk
);
2461 if (tcp_snd_wnd_test(tp
, skb
, mss
)) {
2462 pcount
= tp
->packets_out
;
2463 tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2464 if (tp
->packets_out
> pcount
)
2468 skb
= tcp_write_queue_prev(sk
, skb
);
2470 skb
= tcp_write_queue_tail(sk
);
2473 /* At most one outstanding TLP retransmission. */
2474 if (tp
->tlp_high_seq
)
2477 /* Retransmit last segment. */
2481 if (skb_still_in_host_queue(sk
, skb
))
2484 pcount
= tcp_skb_pcount(skb
);
2485 if (WARN_ON(!pcount
))
2488 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2489 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
,
2492 skb
= tcp_write_queue_next(sk
, skb
);
2495 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2498 if (__tcp_retransmit_skb(sk
, skb
, 1))
2501 /* Record snd_nxt for loss detection. */
2502 tp
->tlp_high_seq
= tp
->snd_nxt
;
2505 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSPROBES
);
2506 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2507 inet_csk(sk
)->icsk_pending
= 0;
2512 /* Push out any pending frames which were held back due to
2513 * TCP_CORK or attempt at coalescing tiny packets.
2514 * The socket must be locked by the caller.
2516 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2519 /* If we are closed, the bytes will have to remain here.
2520 * In time closedown will finish, we empty the write queue and
2521 * all will be happy.
2523 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2526 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2527 sk_gfp_mask(sk
, GFP_ATOMIC
)))
2528 tcp_check_probe_timer(sk
);
2531 /* Send _single_ skb sitting at the send head. This function requires
2532 * true push pending frames to setup probe timer etc.
2534 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2536 struct sk_buff
*skb
= tcp_send_head(sk
);
2538 BUG_ON(!skb
|| skb
->len
< mss_now
);
2540 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2543 /* This function returns the amount that we can raise the
2544 * usable window based on the following constraints
2546 * 1. The window can never be shrunk once it is offered (RFC 793)
2547 * 2. We limit memory per socket
2550 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2551 * RECV.NEXT + RCV.WIN fixed until:
2552 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2554 * i.e. don't raise the right edge of the window until you can raise
2555 * it at least MSS bytes.
2557 * Unfortunately, the recommended algorithm breaks header prediction,
2558 * since header prediction assumes th->window stays fixed.
2560 * Strictly speaking, keeping th->window fixed violates the receiver
2561 * side SWS prevention criteria. The problem is that under this rule
2562 * a stream of single byte packets will cause the right side of the
2563 * window to always advance by a single byte.
2565 * Of course, if the sender implements sender side SWS prevention
2566 * then this will not be a problem.
2568 * BSD seems to make the following compromise:
2570 * If the free space is less than the 1/4 of the maximum
2571 * space available and the free space is less than 1/2 mss,
2572 * then set the window to 0.
2573 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2574 * Otherwise, just prevent the window from shrinking
2575 * and from being larger than the largest representable value.
2577 * This prevents incremental opening of the window in the regime
2578 * where TCP is limited by the speed of the reader side taking
2579 * data out of the TCP receive queue. It does nothing about
2580 * those cases where the window is constrained on the sender side
2581 * because the pipeline is full.
2583 * BSD also seems to "accidentally" limit itself to windows that are a
2584 * multiple of MSS, at least until the free space gets quite small.
2585 * This would appear to be a side effect of the mbuf implementation.
2586 * Combining these two algorithms results in the observed behavior
2587 * of having a fixed window size at almost all times.
2589 * Below we obtain similar behavior by forcing the offered window to
2590 * a multiple of the mss when it is feasible to do so.
2592 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2593 * Regular options like TIMESTAMP are taken into account.
2595 u32
__tcp_select_window(struct sock
*sk
)
2597 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2598 struct tcp_sock
*tp
= tcp_sk(sk
);
2599 /* MSS for the peer's data. Previous versions used mss_clamp
2600 * here. I don't know if the value based on our guesses
2601 * of peer's MSS is better for the performance. It's more correct
2602 * but may be worse for the performance because of rcv_mss
2603 * fluctuations. --SAW 1998/11/1
2605 int mss
= icsk
->icsk_ack
.rcv_mss
;
2606 int free_space
= tcp_space(sk
);
2607 int allowed_space
= tcp_full_space(sk
);
2608 int full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2611 if (unlikely(mss
> full_space
)) {
2616 if (free_space
< (full_space
>> 1)) {
2617 icsk
->icsk_ack
.quick
= 0;
2619 if (tcp_under_memory_pressure(sk
))
2620 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2623 /* free_space might become our new window, make sure we don't
2624 * increase it due to wscale.
2626 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2628 /* if free space is less than mss estimate, or is below 1/16th
2629 * of the maximum allowed, try to move to zero-window, else
2630 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2631 * new incoming data is dropped due to memory limits.
2632 * With large window, mss test triggers way too late in order
2633 * to announce zero window in time before rmem limit kicks in.
2635 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2639 if (free_space
> tp
->rcv_ssthresh
)
2640 free_space
= tp
->rcv_ssthresh
;
2642 /* Don't do rounding if we are using window scaling, since the
2643 * scaled window will not line up with the MSS boundary anyway.
2645 if (tp
->rx_opt
.rcv_wscale
) {
2646 window
= free_space
;
2648 /* Advertise enough space so that it won't get scaled away.
2649 * Import case: prevent zero window announcement if
2650 * 1<<rcv_wscale > mss.
2652 window
= ALIGN(window
, (1 << tp
->rx_opt
.rcv_wscale
));
2654 window
= tp
->rcv_wnd
;
2655 /* Get the largest window that is a nice multiple of mss.
2656 * Window clamp already applied above.
2657 * If our current window offering is within 1 mss of the
2658 * free space we just keep it. This prevents the divide
2659 * and multiply from happening most of the time.
2660 * We also don't do any window rounding when the free space
2663 if (window
<= free_space
- mss
|| window
> free_space
)
2664 window
= rounddown(free_space
, mss
);
2665 else if (mss
== full_space
&&
2666 free_space
> window
+ (full_space
>> 1))
2667 window
= free_space
;
2673 void tcp_skb_collapse_tstamp(struct sk_buff
*skb
,
2674 const struct sk_buff
*next_skb
)
2676 if (unlikely(tcp_has_tx_tstamp(next_skb
))) {
2677 const struct skb_shared_info
*next_shinfo
=
2678 skb_shinfo(next_skb
);
2679 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
2681 shinfo
->tx_flags
|= next_shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
2682 shinfo
->tskey
= next_shinfo
->tskey
;
2683 TCP_SKB_CB(skb
)->txstamp_ack
|=
2684 TCP_SKB_CB(next_skb
)->txstamp_ack
;
2688 /* Collapses two adjacent SKB's during retransmission. */
2689 static bool tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2691 struct tcp_sock
*tp
= tcp_sk(sk
);
2692 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2693 int skb_size
, next_skb_size
;
2695 skb_size
= skb
->len
;
2696 next_skb_size
= next_skb
->len
;
2698 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2700 if (next_skb_size
) {
2701 if (next_skb_size
<= skb_availroom(skb
))
2702 skb_copy_bits(next_skb
, 0, skb_put(skb
, next_skb_size
),
2704 else if (!skb_shift(skb
, next_skb
, next_skb_size
))
2707 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2709 tcp_unlink_write_queue(next_skb
, sk
);
2711 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2712 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2714 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2715 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2717 /* Update sequence range on original skb. */
2718 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2720 /* Merge over control information. This moves PSH/FIN etc. over */
2721 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2723 /* All done, get rid of second SKB and account for it so
2724 * packet counting does not break.
2726 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2727 TCP_SKB_CB(skb
)->eor
= TCP_SKB_CB(next_skb
)->eor
;
2729 /* changed transmit queue under us so clear hints */
2730 tcp_clear_retrans_hints_partial(tp
);
2731 if (next_skb
== tp
->retransmit_skb_hint
)
2732 tp
->retransmit_skb_hint
= skb
;
2734 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2736 tcp_skb_collapse_tstamp(skb
, next_skb
);
2738 sk_wmem_free_skb(sk
, next_skb
);
2742 /* Check if coalescing SKBs is legal. */
2743 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2745 if (tcp_skb_pcount(skb
) > 1)
2747 if (skb_cloned(skb
))
2749 if (skb
== tcp_send_head(sk
))
2751 /* Some heuristics for collapsing over SACK'd could be invented */
2752 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2758 /* Collapse packets in the retransmit queue to make to create
2759 * less packets on the wire. This is only done on retransmission.
2761 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2764 struct tcp_sock
*tp
= tcp_sk(sk
);
2765 struct sk_buff
*skb
= to
, *tmp
;
2768 if (!sysctl_tcp_retrans_collapse
)
2770 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2773 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2774 if (!tcp_can_collapse(sk
, skb
))
2777 if (!tcp_skb_can_collapse_to(to
))
2790 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2793 if (!tcp_collapse_retrans(sk
, to
))
2798 /* This retransmits one SKB. Policy decisions and retransmit queue
2799 * state updates are done by the caller. Returns non-zero if an
2800 * error occurred which prevented the send.
2802 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2804 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2805 struct tcp_sock
*tp
= tcp_sk(sk
);
2806 unsigned int cur_mss
;
2810 /* Inconclusive MTU probe */
2811 if (icsk
->icsk_mtup
.probe_size
)
2812 icsk
->icsk_mtup
.probe_size
= 0;
2814 /* Do not sent more than we queued. 1/4 is reserved for possible
2815 * copying overhead: fragmentation, tunneling, mangling etc.
2817 if (atomic_read(&sk
->sk_wmem_alloc
) >
2818 min_t(u32
, sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2),
2822 if (skb_still_in_host_queue(sk
, skb
))
2825 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2826 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2828 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2832 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2833 return -EHOSTUNREACH
; /* Routing failure or similar. */
2835 cur_mss
= tcp_current_mss(sk
);
2837 /* If receiver has shrunk his window, and skb is out of
2838 * new window, do not retransmit it. The exception is the
2839 * case, when window is shrunk to zero. In this case
2840 * our retransmit serves as a zero window probe.
2842 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2843 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2846 len
= cur_mss
* segs
;
2847 if (skb
->len
> len
) {
2848 if (tcp_fragment(sk
, skb
, len
, cur_mss
, GFP_ATOMIC
))
2849 return -ENOMEM
; /* We'll try again later. */
2851 if (skb_unclone(skb
, GFP_ATOMIC
))
2854 diff
= tcp_skb_pcount(skb
);
2855 tcp_set_skb_tso_segs(skb
, cur_mss
);
2856 diff
-= tcp_skb_pcount(skb
);
2858 tcp_adjust_pcount(sk
, skb
, diff
);
2859 if (skb
->len
< cur_mss
)
2860 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2863 /* RFC3168, section 6.1.1.1. ECN fallback */
2864 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN_ECN
) == TCPHDR_SYN_ECN
)
2865 tcp_ecn_clear_syn(sk
, skb
);
2867 /* Update global and local TCP statistics. */
2868 segs
= tcp_skb_pcount(skb
);
2869 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
, segs
);
2870 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2871 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
2872 tp
->total_retrans
+= segs
;
2874 /* make sure skb->data is aligned on arches that require it
2875 * and check if ack-trimming & collapsing extended the headroom
2876 * beyond what csum_start can cover.
2878 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2879 skb_headroom(skb
) >= 0xFFFF)) {
2880 struct sk_buff
*nskb
;
2882 skb
->skb_mstamp
= tp
->tcp_mstamp
;
2883 nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
, GFP_ATOMIC
);
2884 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2887 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2891 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
2892 } else if (err
!= -EBUSY
) {
2893 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2898 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2900 struct tcp_sock
*tp
= tcp_sk(sk
);
2901 int err
= __tcp_retransmit_skb(sk
, skb
, segs
);
2904 #if FASTRETRANS_DEBUG > 0
2905 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2906 net_dbg_ratelimited("retrans_out leaked\n");
2909 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2910 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2912 /* Save stamp of the first retransmit. */
2913 if (!tp
->retrans_stamp
)
2914 tp
->retrans_stamp
= tcp_skb_timestamp(skb
);
2918 if (tp
->undo_retrans
< 0)
2919 tp
->undo_retrans
= 0;
2920 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2924 /* This gets called after a retransmit timeout, and the initially
2925 * retransmitted data is acknowledged. It tries to continue
2926 * resending the rest of the retransmit queue, until either
2927 * we've sent it all or the congestion window limit is reached.
2928 * If doing SACK, the first ACK which comes back for a timeout
2929 * based retransmit packet might feed us FACK information again.
2930 * If so, we use it to avoid unnecessarily retransmissions.
2932 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2934 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2935 struct tcp_sock
*tp
= tcp_sk(sk
);
2936 struct sk_buff
*skb
;
2937 struct sk_buff
*hole
= NULL
;
2941 if (!tp
->packets_out
)
2944 if (tp
->retransmit_skb_hint
) {
2945 skb
= tp
->retransmit_skb_hint
;
2947 skb
= tcp_write_queue_head(sk
);
2950 max_segs
= tcp_tso_segs(sk
, tcp_current_mss(sk
));
2951 tcp_for_write_queue_from(skb
, sk
) {
2955 if (skb
== tcp_send_head(sk
))
2958 if (tcp_pacing_check(sk
))
2961 /* we could do better than to assign each time */
2963 tp
->retransmit_skb_hint
= skb
;
2965 segs
= tp
->snd_cwnd
- tcp_packets_in_flight(tp
);
2968 sacked
= TCP_SKB_CB(skb
)->sacked
;
2969 /* In case tcp_shift_skb_data() have aggregated large skbs,
2970 * we need to make sure not sending too bigs TSO packets
2972 segs
= min_t(int, segs
, max_segs
);
2974 if (tp
->retrans_out
>= tp
->lost_out
) {
2976 } else if (!(sacked
& TCPCB_LOST
)) {
2977 if (!hole
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2982 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2983 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2985 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2988 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2991 if (tcp_small_queue_check(sk
, skb
, 1))
2994 if (tcp_retransmit_skb(sk
, skb
, segs
))
2997 NET_ADD_STATS(sock_net(sk
), mib_idx
, tcp_skb_pcount(skb
));
2999 if (tcp_in_cwnd_reduction(sk
))
3000 tp
->prr_out
+= tcp_skb_pcount(skb
);
3002 if (skb
== tcp_write_queue_head(sk
) &&
3003 icsk
->icsk_pending
!= ICSK_TIME_REO_TIMEOUT
)
3004 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3005 inet_csk(sk
)->icsk_rto
,
3010 /* We allow to exceed memory limits for FIN packets to expedite
3011 * connection tear down and (memory) recovery.
3012 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3013 * or even be forced to close flow without any FIN.
3014 * In general, we want to allow one skb per socket to avoid hangs
3015 * with edge trigger epoll()
3017 void sk_forced_mem_schedule(struct sock
*sk
, int size
)
3021 if (size
<= sk
->sk_forward_alloc
)
3023 amt
= sk_mem_pages(size
);
3024 sk
->sk_forward_alloc
+= amt
* SK_MEM_QUANTUM
;
3025 sk_memory_allocated_add(sk
, amt
);
3027 if (mem_cgroup_sockets_enabled
&& sk
->sk_memcg
)
3028 mem_cgroup_charge_skmem(sk
->sk_memcg
, amt
);
3031 /* Send a FIN. The caller locks the socket for us.
3032 * We should try to send a FIN packet really hard, but eventually give up.
3034 void tcp_send_fin(struct sock
*sk
)
3036 struct sk_buff
*skb
, *tskb
= tcp_write_queue_tail(sk
);
3037 struct tcp_sock
*tp
= tcp_sk(sk
);
3039 /* Optimization, tack on the FIN if we have one skb in write queue and
3040 * this skb was not yet sent, or we are under memory pressure.
3041 * Note: in the latter case, FIN packet will be sent after a timeout,
3042 * as TCP stack thinks it has already been transmitted.
3044 if (tskb
&& (tcp_send_head(sk
) || tcp_under_memory_pressure(sk
))) {
3046 TCP_SKB_CB(tskb
)->tcp_flags
|= TCPHDR_FIN
;
3047 TCP_SKB_CB(tskb
)->end_seq
++;
3049 if (!tcp_send_head(sk
)) {
3050 /* This means tskb was already sent.
3051 * Pretend we included the FIN on previous transmit.
3052 * We need to set tp->snd_nxt to the value it would have
3053 * if FIN had been sent. This is because retransmit path
3054 * does not change tp->snd_nxt.
3060 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, sk
->sk_allocation
);
3061 if (unlikely(!skb
)) {
3066 skb_reserve(skb
, MAX_TCP_HEADER
);
3067 sk_forced_mem_schedule(sk
, skb
->truesize
);
3068 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3069 tcp_init_nondata_skb(skb
, tp
->write_seq
,
3070 TCPHDR_ACK
| TCPHDR_FIN
);
3071 tcp_queue_skb(sk
, skb
);
3073 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
), TCP_NAGLE_OFF
);
3076 /* We get here when a process closes a file descriptor (either due to
3077 * an explicit close() or as a byproduct of exit()'ing) and there
3078 * was unread data in the receive queue. This behavior is recommended
3079 * by RFC 2525, section 2.17. -DaveM
3081 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
3083 struct sk_buff
*skb
;
3085 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
3087 /* NOTE: No TCP options attached and we never retransmit this. */
3088 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
3090 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3094 /* Reserve space for headers and prepare control bits. */
3095 skb_reserve(skb
, MAX_TCP_HEADER
);
3096 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
3097 TCPHDR_ACK
| TCPHDR_RST
);
3098 skb_mstamp_get(&tcp_sk(sk
)->tcp_mstamp
);
3100 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
3101 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3104 /* Send a crossed SYN-ACK during socket establishment.
3105 * WARNING: This routine must only be called when we have already sent
3106 * a SYN packet that crossed the incoming SYN that caused this routine
3107 * to get called. If this assumption fails then the initial rcv_wnd
3108 * and rcv_wscale values will not be correct.
3110 int tcp_send_synack(struct sock
*sk
)
3112 struct sk_buff
*skb
;
3114 skb
= tcp_write_queue_head(sk
);
3115 if (!skb
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
3116 pr_debug("%s: wrong queue state\n", __func__
);
3119 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
3120 if (skb_cloned(skb
)) {
3121 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
3124 tcp_unlink_write_queue(skb
, sk
);
3125 __skb_header_release(nskb
);
3126 __tcp_add_write_queue_head(sk
, nskb
);
3127 sk_wmem_free_skb(sk
, skb
);
3128 sk
->sk_wmem_queued
+= nskb
->truesize
;
3129 sk_mem_charge(sk
, nskb
->truesize
);
3133 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
3134 tcp_ecn_send_synack(sk
, skb
);
3136 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3140 * tcp_make_synack - Prepare a SYN-ACK.
3141 * sk: listener socket
3142 * dst: dst entry attached to the SYNACK
3143 * req: request_sock pointer
3145 * Allocate one skb and build a SYNACK packet.
3146 * @dst is consumed : Caller should not use it again.
3148 struct sk_buff
*tcp_make_synack(const struct sock
*sk
, struct dst_entry
*dst
,
3149 struct request_sock
*req
,
3150 struct tcp_fastopen_cookie
*foc
,
3151 enum tcp_synack_type synack_type
)
3153 struct inet_request_sock
*ireq
= inet_rsk(req
);
3154 const struct tcp_sock
*tp
= tcp_sk(sk
);
3155 struct tcp_md5sig_key
*md5
= NULL
;
3156 struct tcp_out_options opts
;
3157 struct sk_buff
*skb
;
3158 int tcp_header_size
;
3162 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
3163 if (unlikely(!skb
)) {
3167 /* Reserve space for headers. */
3168 skb_reserve(skb
, MAX_TCP_HEADER
);
3170 switch (synack_type
) {
3171 case TCP_SYNACK_NORMAL
:
3172 skb_set_owner_w(skb
, req_to_sk(req
));
3174 case TCP_SYNACK_COOKIE
:
3175 /* Under synflood, we do not attach skb to a socket,
3176 * to avoid false sharing.
3179 case TCP_SYNACK_FASTOPEN
:
3180 /* sk is a const pointer, because we want to express multiple
3181 * cpu might call us concurrently.
3182 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3184 skb_set_owner_w(skb
, (struct sock
*)sk
);
3187 skb_dst_set(skb
, dst
);
3189 mss
= tcp_mss_clamp(tp
, dst_metric_advmss(dst
));
3191 memset(&opts
, 0, sizeof(opts
));
3192 #ifdef CONFIG_SYN_COOKIES
3193 if (unlikely(req
->cookie_ts
))
3194 skb
->skb_mstamp
.stamp_jiffies
= cookie_init_timestamp(req
);
3197 skb_mstamp_get(&skb
->skb_mstamp
);
3199 #ifdef CONFIG_TCP_MD5SIG
3201 md5
= tcp_rsk(req
)->af_specific
->req_md5_lookup(sk
, req_to_sk(req
));
3203 skb_set_hash(skb
, tcp_rsk(req
)->txhash
, PKT_HASH_TYPE_L4
);
3204 tcp_header_size
= tcp_synack_options(req
, mss
, skb
, &opts
, md5
, foc
) +
3207 skb_push(skb
, tcp_header_size
);
3208 skb_reset_transport_header(skb
);
3210 th
= (struct tcphdr
*)skb
->data
;
3211 memset(th
, 0, sizeof(struct tcphdr
));
3214 tcp_ecn_make_synack(req
, th
);
3215 th
->source
= htons(ireq
->ir_num
);
3216 th
->dest
= ireq
->ir_rmt_port
;
3217 /* Setting of flags are superfluous here for callers (and ECE is
3218 * not even correctly set)
3220 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
3221 TCPHDR_SYN
| TCPHDR_ACK
);
3223 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
3224 /* XXX data is queued and acked as is. No buffer/window check */
3225 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
3227 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3228 th
->window
= htons(min(req
->rsk_rcv_wnd
, 65535U));
3229 tcp_options_write((__be32
*)(th
+ 1), NULL
, &opts
);
3230 th
->doff
= (tcp_header_size
>> 2);
3231 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
3233 #ifdef CONFIG_TCP_MD5SIG
3234 /* Okay, we have all we need - do the md5 hash if needed */
3236 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
3237 md5
, req_to_sk(req
), skb
);
3241 /* Do not fool tcpdump (if any), clean our debris */
3245 EXPORT_SYMBOL(tcp_make_synack
);
3247 static void tcp_ca_dst_init(struct sock
*sk
, const struct dst_entry
*dst
)
3249 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3250 const struct tcp_congestion_ops
*ca
;
3251 u32 ca_key
= dst_metric(dst
, RTAX_CC_ALGO
);
3253 if (ca_key
== TCP_CA_UNSPEC
)
3257 ca
= tcp_ca_find_key(ca_key
);
3258 if (likely(ca
&& try_module_get(ca
->owner
))) {
3259 module_put(icsk
->icsk_ca_ops
->owner
);
3260 icsk
->icsk_ca_dst_locked
= tcp_ca_dst_locked(dst
);
3261 icsk
->icsk_ca_ops
= ca
;
3266 /* Do all connect socket setups that can be done AF independent. */
3267 static void tcp_connect_init(struct sock
*sk
)
3269 const struct dst_entry
*dst
= __sk_dst_get(sk
);
3270 struct tcp_sock
*tp
= tcp_sk(sk
);
3273 /* We'll fix this up when we get a response from the other end.
3274 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3276 tp
->tcp_header_len
= sizeof(struct tcphdr
) +
3277 (sysctl_tcp_timestamps
? TCPOLEN_TSTAMP_ALIGNED
: 0);
3279 #ifdef CONFIG_TCP_MD5SIG
3280 if (tp
->af_specific
->md5_lookup(sk
, sk
))
3281 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
3284 /* If user gave his TCP_MAXSEG, record it to clamp */
3285 if (tp
->rx_opt
.user_mss
)
3286 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3289 tcp_sync_mss(sk
, dst_mtu(dst
));
3291 tcp_ca_dst_init(sk
, dst
);
3293 if (!tp
->window_clamp
)
3294 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
3295 tp
->advmss
= tcp_mss_clamp(tp
, dst_metric_advmss(dst
));
3297 tcp_initialize_rcv_mss(sk
);
3299 /* limit the window selection if the user enforce a smaller rx buffer */
3300 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
3301 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
3302 tp
->window_clamp
= tcp_full_space(sk
);
3304 tcp_select_initial_window(tcp_full_space(sk
),
3305 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
3308 sysctl_tcp_window_scaling
,
3310 dst_metric(dst
, RTAX_INITRWND
));
3312 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
3313 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
3316 sock_reset_flag(sk
, SOCK_DONE
);
3319 tp
->snd_una
= tp
->write_seq
;
3320 tp
->snd_sml
= tp
->write_seq
;
3321 tp
->snd_up
= tp
->write_seq
;
3322 tp
->snd_nxt
= tp
->write_seq
;
3324 if (likely(!tp
->repair
))
3327 tp
->rcv_tstamp
= tcp_jiffies32
;
3328 tp
->rcv_wup
= tp
->rcv_nxt
;
3329 tp
->copied_seq
= tp
->rcv_nxt
;
3331 inet_csk(sk
)->icsk_rto
= TCP_TIMEOUT_INIT
;
3332 inet_csk(sk
)->icsk_retransmits
= 0;
3333 tcp_clear_retrans(tp
);
3336 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
3338 struct tcp_sock
*tp
= tcp_sk(sk
);
3339 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
3341 tcb
->end_seq
+= skb
->len
;
3342 __skb_header_release(skb
);
3343 __tcp_add_write_queue_tail(sk
, skb
);
3344 sk
->sk_wmem_queued
+= skb
->truesize
;
3345 sk_mem_charge(sk
, skb
->truesize
);
3346 tp
->write_seq
= tcb
->end_seq
;
3347 tp
->packets_out
+= tcp_skb_pcount(skb
);
3350 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3351 * queue a data-only packet after the regular SYN, such that regular SYNs
3352 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3353 * only the SYN sequence, the data are retransmitted in the first ACK.
3354 * If cookie is not cached or other error occurs, falls back to send a
3355 * regular SYN with Fast Open cookie request option.
3357 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
3359 struct tcp_sock
*tp
= tcp_sk(sk
);
3360 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
3362 struct sk_buff
*syn_data
;
3364 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
3365 if (!tcp_fastopen_cookie_check(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
))
3368 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3369 * user-MSS. Reserve maximum option space for middleboxes that add
3370 * private TCP options. The cost is reduced data space in SYN :(
3372 tp
->rx_opt
.mss_clamp
= tcp_mss_clamp(tp
, tp
->rx_opt
.mss_clamp
);
3374 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
3375 MAX_TCP_OPTION_SPACE
;
3377 space
= min_t(size_t, space
, fo
->size
);
3379 /* limit to order-0 allocations */
3380 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
3382 syn_data
= sk_stream_alloc_skb(sk
, space
, sk
->sk_allocation
, false);
3385 syn_data
->ip_summed
= CHECKSUM_PARTIAL
;
3386 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
3388 int copied
= copy_from_iter(skb_put(syn_data
, space
), space
,
3389 &fo
->data
->msg_iter
);
3390 if (unlikely(!copied
)) {
3391 kfree_skb(syn_data
);
3394 if (copied
!= space
) {
3395 skb_trim(syn_data
, copied
);
3399 /* No more data pending in inet_wait_for_connect() */
3400 if (space
== fo
->size
)
3404 tcp_connect_queue_skb(sk
, syn_data
);
3406 tcp_chrono_start(sk
, TCP_CHRONO_BUSY
);
3408 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
3410 syn
->skb_mstamp
= syn_data
->skb_mstamp
;
3412 /* Now full SYN+DATA was cloned and sent (or not),
3413 * remove the SYN from the original skb (syn_data)
3414 * we keep in write queue in case of a retransmit, as we
3415 * also have the SYN packet (with no data) in the same queue.
3417 TCP_SKB_CB(syn_data
)->seq
++;
3418 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
3420 tp
->syn_data
= (fo
->copied
> 0);
3421 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3426 /* Send a regular SYN with Fast Open cookie request option */
3427 if (fo
->cookie
.len
> 0)
3429 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3431 tp
->syn_fastopen
= 0;
3433 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3437 /* Build a SYN and send it off. */
3438 int tcp_connect(struct sock
*sk
)
3440 struct tcp_sock
*tp
= tcp_sk(sk
);
3441 struct sk_buff
*buff
;
3444 tcp_connect_init(sk
);
3446 if (unlikely(tp
->repair
)) {
3447 tcp_finish_connect(sk
, NULL
);
3451 buff
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
, true);
3452 if (unlikely(!buff
))
3455 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3456 skb_mstamp_get(&tp
->tcp_mstamp
);
3457 tp
->retrans_stamp
= tp
->tcp_mstamp
.stamp_jiffies
;
3458 tcp_connect_queue_skb(sk
, buff
);
3459 tcp_ecn_send_syn(sk
, buff
);
3461 /* Send off SYN; include data in Fast Open. */
3462 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3463 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3464 if (err
== -ECONNREFUSED
)
3467 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3468 * in order to make this packet get counted in tcpOutSegs.
3470 tp
->snd_nxt
= tp
->write_seq
;
3471 tp
->pushed_seq
= tp
->write_seq
;
3472 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3474 /* Timer for repeating the SYN until an answer. */
3475 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3476 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3479 EXPORT_SYMBOL(tcp_connect
);
3481 /* Send out a delayed ack, the caller does the policy checking
3482 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3485 void tcp_send_delayed_ack(struct sock
*sk
)
3487 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3488 int ato
= icsk
->icsk_ack
.ato
;
3489 unsigned long timeout
;
3491 tcp_ca_event(sk
, CA_EVENT_DELAYED_ACK
);
3493 if (ato
> TCP_DELACK_MIN
) {
3494 const struct tcp_sock
*tp
= tcp_sk(sk
);
3495 int max_ato
= HZ
/ 2;
3497 if (icsk
->icsk_ack
.pingpong
||
3498 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3499 max_ato
= TCP_DELACK_MAX
;
3501 /* Slow path, intersegment interval is "high". */
3503 /* If some rtt estimate is known, use it to bound delayed ack.
3504 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3508 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3515 ato
= min(ato
, max_ato
);
3518 /* Stay within the limit we were given */
3519 timeout
= jiffies
+ ato
;
3521 /* Use new timeout only if there wasn't a older one earlier. */
3522 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3523 /* If delack timer was blocked or is about to expire,
3526 if (icsk
->icsk_ack
.blocked
||
3527 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3532 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3533 timeout
= icsk
->icsk_ack
.timeout
;
3535 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3536 icsk
->icsk_ack
.timeout
= timeout
;
3537 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3540 /* This routine sends an ack and also updates the window. */
3541 void tcp_send_ack(struct sock
*sk
)
3543 struct sk_buff
*buff
;
3545 /* If we have been reset, we may not send again. */
3546 if (sk
->sk_state
== TCP_CLOSE
)
3549 tcp_ca_event(sk
, CA_EVENT_NON_DELAYED_ACK
);
3551 /* We are not putting this on the write queue, so
3552 * tcp_transmit_skb() will set the ownership to this
3555 buff
= alloc_skb(MAX_TCP_HEADER
,
3556 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3557 if (unlikely(!buff
)) {
3558 inet_csk_schedule_ack(sk
);
3559 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3560 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3561 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3565 /* Reserve space for headers and prepare control bits. */
3566 skb_reserve(buff
, MAX_TCP_HEADER
);
3567 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3569 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3571 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3573 skb_set_tcp_pure_ack(buff
);
3575 /* Send it off, this clears delayed acks for us. */
3576 tcp_transmit_skb(sk
, buff
, 0, (__force gfp_t
)0);
3578 EXPORT_SYMBOL_GPL(tcp_send_ack
);
3580 /* This routine sends a packet with an out of date sequence
3581 * number. It assumes the other end will try to ack it.
3583 * Question: what should we make while urgent mode?
3584 * 4.4BSD forces sending single byte of data. We cannot send
3585 * out of window data, because we have SND.NXT==SND.MAX...
3587 * Current solution: to send TWO zero-length segments in urgent mode:
3588 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3589 * out-of-date with SND.UNA-1 to probe window.
3591 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
, int mib
)
3593 struct tcp_sock
*tp
= tcp_sk(sk
);
3594 struct sk_buff
*skb
;
3596 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3597 skb
= alloc_skb(MAX_TCP_HEADER
,
3598 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3602 /* Reserve space for headers and set control bits. */
3603 skb_reserve(skb
, MAX_TCP_HEADER
);
3604 /* Use a previous sequence. This should cause the other
3605 * end to send an ack. Don't queue or clone SKB, just
3608 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3609 NET_INC_STATS(sock_net(sk
), mib
);
3610 return tcp_transmit_skb(sk
, skb
, 0, (__force gfp_t
)0);
3613 /* Called from setsockopt( ... TCP_REPAIR ) */
3614 void tcp_send_window_probe(struct sock
*sk
)
3616 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3617 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3618 skb_mstamp_get(&tcp_sk(sk
)->tcp_mstamp
);
3619 tcp_xmit_probe_skb(sk
, 0, LINUX_MIB_TCPWINPROBE
);
3623 /* Initiate keepalive or window probe from timer. */
3624 int tcp_write_wakeup(struct sock
*sk
, int mib
)
3626 struct tcp_sock
*tp
= tcp_sk(sk
);
3627 struct sk_buff
*skb
;
3629 if (sk
->sk_state
== TCP_CLOSE
)
3632 skb
= tcp_send_head(sk
);
3633 if (skb
&& before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3635 unsigned int mss
= tcp_current_mss(sk
);
3636 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3638 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3639 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3641 /* We are probing the opening of a window
3642 * but the window size is != 0
3643 * must have been a result SWS avoidance ( sender )
3645 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3647 seg_size
= min(seg_size
, mss
);
3648 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3649 if (tcp_fragment(sk
, skb
, seg_size
, mss
, GFP_ATOMIC
))
3651 } else if (!tcp_skb_pcount(skb
))
3652 tcp_set_skb_tso_segs(skb
, mss
);
3654 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3655 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3657 tcp_event_new_data_sent(sk
, skb
);
3660 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3661 tcp_xmit_probe_skb(sk
, 1, mib
);
3662 return tcp_xmit_probe_skb(sk
, 0, mib
);
3666 /* A window probe timeout has occurred. If window is not closed send
3667 * a partial packet else a zero probe.
3669 void tcp_send_probe0(struct sock
*sk
)
3671 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3672 struct tcp_sock
*tp
= tcp_sk(sk
);
3673 struct net
*net
= sock_net(sk
);
3674 unsigned long probe_max
;
3677 err
= tcp_write_wakeup(sk
, LINUX_MIB_TCPWINPROBE
);
3679 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3680 /* Cancel probe timer, if it is not required. */
3681 icsk
->icsk_probes_out
= 0;
3682 icsk
->icsk_backoff
= 0;
3687 if (icsk
->icsk_backoff
< net
->ipv4
.sysctl_tcp_retries2
)
3688 icsk
->icsk_backoff
++;
3689 icsk
->icsk_probes_out
++;
3690 probe_max
= TCP_RTO_MAX
;
3692 /* If packet was not sent due to local congestion,
3693 * do not backoff and do not remember icsk_probes_out.
3694 * Let local senders to fight for local resources.
3696 * Use accumulated backoff yet.
3698 if (!icsk
->icsk_probes_out
)
3699 icsk
->icsk_probes_out
= 1;
3700 probe_max
= TCP_RESOURCE_PROBE_INTERVAL
;
3702 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3703 tcp_probe0_when(sk
, probe_max
),
3707 int tcp_rtx_synack(const struct sock
*sk
, struct request_sock
*req
)
3709 const struct tcp_request_sock_ops
*af_ops
= tcp_rsk(req
)->af_specific
;
3713 tcp_rsk(req
)->txhash
= net_tx_rndhash();
3714 res
= af_ops
->send_synack(sk
, NULL
, &fl
, req
, NULL
, TCP_SYNACK_NORMAL
);
3716 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
3717 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
3718 if (unlikely(tcp_passive_fastopen(sk
)))
3719 tcp_sk(sk
)->total_retrans
++;
3723 EXPORT_SYMBOL(tcp_rtx_synack
);