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
) ||
320 tcp_bpf_ca_needs_ecn(sk
))
324 /* Packet ECN state for a SYN. */
325 static void tcp_ecn_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
327 struct tcp_sock
*tp
= tcp_sk(sk
);
328 bool bpf_needs_ecn
= tcp_bpf_ca_needs_ecn(sk
);
329 bool use_ecn
= sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1 ||
330 tcp_ca_needs_ecn(sk
) || bpf_needs_ecn
;
333 const struct dst_entry
*dst
= __sk_dst_get(sk
);
335 if (dst
&& dst_feature(dst
, RTAX_FEATURE_ECN
))
342 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
343 tp
->ecn_flags
= TCP_ECN_OK
;
344 if (tcp_ca_needs_ecn(sk
) || bpf_needs_ecn
)
349 static void tcp_ecn_clear_syn(struct sock
*sk
, struct sk_buff
*skb
)
351 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn_fallback
)
352 /* tp->ecn_flags are cleared at a later point in time when
353 * SYN ACK is ultimatively being received.
355 TCP_SKB_CB(skb
)->tcp_flags
&= ~(TCPHDR_ECE
| TCPHDR_CWR
);
359 tcp_ecn_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
361 if (inet_rsk(req
)->ecn_ok
)
365 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
368 static void tcp_ecn_send(struct sock
*sk
, struct sk_buff
*skb
,
369 struct tcphdr
*th
, int tcp_header_len
)
371 struct tcp_sock
*tp
= tcp_sk(sk
);
373 if (tp
->ecn_flags
& TCP_ECN_OK
) {
374 /* Not-retransmitted data segment: set ECT and inject CWR. */
375 if (skb
->len
!= tcp_header_len
&&
376 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
378 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
379 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
381 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
383 } else if (!tcp_ca_needs_ecn(sk
)) {
384 /* ACK or retransmitted segment: clear ECT|CE */
385 INET_ECN_dontxmit(sk
);
387 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
392 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
393 * auto increment end seqno.
395 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
397 skb
->ip_summed
= CHECKSUM_PARTIAL
;
400 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
401 TCP_SKB_CB(skb
)->sacked
= 0;
403 tcp_skb_pcount_set(skb
, 1);
405 TCP_SKB_CB(skb
)->seq
= seq
;
406 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
408 TCP_SKB_CB(skb
)->end_seq
= seq
;
411 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
413 return tp
->snd_una
!= tp
->snd_up
;
416 #define OPTION_SACK_ADVERTISE (1 << 0)
417 #define OPTION_TS (1 << 1)
418 #define OPTION_MD5 (1 << 2)
419 #define OPTION_WSCALE (1 << 3)
420 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
422 struct tcp_out_options
{
423 u16 options
; /* bit field of OPTION_* */
424 u16 mss
; /* 0 to disable */
425 u8 ws
; /* window scale, 0 to disable */
426 u8 num_sack_blocks
; /* number of SACK blocks to include */
427 u8 hash_size
; /* bytes in hash_location */
428 __u8
*hash_location
; /* temporary pointer, overloaded */
429 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
430 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
433 /* Write previously computed TCP options to the packet.
435 * Beware: Something in the Internet is very sensitive to the ordering of
436 * TCP options, we learned this through the hard way, so be careful here.
437 * Luckily we can at least blame others for their non-compliance but from
438 * inter-operability perspective it seems that we're somewhat stuck with
439 * the ordering which we have been using if we want to keep working with
440 * those broken things (not that it currently hurts anybody as there isn't
441 * particular reason why the ordering would need to be changed).
443 * At least SACK_PERM as the first option is known to lead to a disaster
444 * (but it may well be that other scenarios fail similarly).
446 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
447 struct tcp_out_options
*opts
)
449 u16 options
= opts
->options
; /* mungable copy */
451 if (unlikely(OPTION_MD5
& options
)) {
452 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
453 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
454 /* overload cookie hash location */
455 opts
->hash_location
= (__u8
*)ptr
;
459 if (unlikely(opts
->mss
)) {
460 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
461 (TCPOLEN_MSS
<< 16) |
465 if (likely(OPTION_TS
& options
)) {
466 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
467 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
468 (TCPOLEN_SACK_PERM
<< 16) |
469 (TCPOPT_TIMESTAMP
<< 8) |
471 options
&= ~OPTION_SACK_ADVERTISE
;
473 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
475 (TCPOPT_TIMESTAMP
<< 8) |
478 *ptr
++ = htonl(opts
->tsval
);
479 *ptr
++ = htonl(opts
->tsecr
);
482 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
483 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
485 (TCPOPT_SACK_PERM
<< 8) |
489 if (unlikely(OPTION_WSCALE
& options
)) {
490 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
491 (TCPOPT_WINDOW
<< 16) |
492 (TCPOLEN_WINDOW
<< 8) |
496 if (unlikely(opts
->num_sack_blocks
)) {
497 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
498 tp
->duplicate_sack
: tp
->selective_acks
;
501 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
504 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
505 TCPOLEN_SACK_PERBLOCK
)));
507 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
509 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
510 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
513 tp
->rx_opt
.dsack
= 0;
516 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
517 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
519 u32 len
; /* Fast Open option length */
522 len
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
523 *ptr
= htonl((TCPOPT_EXP
<< 24) | (len
<< 16) |
524 TCPOPT_FASTOPEN_MAGIC
);
525 p
+= TCPOLEN_EXP_FASTOPEN_BASE
;
527 len
= TCPOLEN_FASTOPEN_BASE
+ foc
->len
;
528 *p
++ = TCPOPT_FASTOPEN
;
532 memcpy(p
, foc
->val
, foc
->len
);
533 if ((len
& 3) == 2) {
534 p
[foc
->len
] = TCPOPT_NOP
;
535 p
[foc
->len
+ 1] = TCPOPT_NOP
;
537 ptr
+= (len
+ 3) >> 2;
541 /* Compute TCP options for SYN packets. This is not the final
542 * network wire format yet.
544 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
545 struct tcp_out_options
*opts
,
546 struct tcp_md5sig_key
**md5
)
548 struct tcp_sock
*tp
= tcp_sk(sk
);
549 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
550 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
552 #ifdef CONFIG_TCP_MD5SIG
553 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
555 opts
->options
|= OPTION_MD5
;
556 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
562 /* We always get an MSS option. The option bytes which will be seen in
563 * normal data packets should timestamps be used, must be in the MSS
564 * advertised. But we subtract them from tp->mss_cache so that
565 * calculations in tcp_sendmsg are simpler etc. So account for this
566 * fact here if necessary. If we don't do this correctly, as a
567 * receiver we won't recognize data packets as being full sized when we
568 * should, and thus we won't abide by the delayed ACK rules correctly.
569 * SACKs don't matter, we never delay an ACK when we have any of those
571 opts
->mss
= tcp_advertise_mss(sk
);
572 remaining
-= TCPOLEN_MSS_ALIGNED
;
574 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_timestamps
&& !*md5
)) {
575 opts
->options
|= OPTION_TS
;
576 opts
->tsval
= tcp_skb_timestamp(skb
) + tp
->tsoffset
;
577 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
578 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
580 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_window_scaling
)) {
581 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
582 opts
->options
|= OPTION_WSCALE
;
583 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
585 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_sack
)) {
586 opts
->options
|= OPTION_SACK_ADVERTISE
;
587 if (unlikely(!(OPTION_TS
& opts
->options
)))
588 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
591 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
592 u32 need
= fastopen
->cookie
.len
;
594 need
+= fastopen
->cookie
.exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
595 TCPOLEN_FASTOPEN_BASE
;
596 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
597 if (remaining
>= need
) {
598 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
599 opts
->fastopen_cookie
= &fastopen
->cookie
;
601 tp
->syn_fastopen
= 1;
602 tp
->syn_fastopen_exp
= fastopen
->cookie
.exp
? 1 : 0;
606 return MAX_TCP_OPTION_SPACE
- remaining
;
609 /* Set up TCP options for SYN-ACKs. */
610 static unsigned int tcp_synack_options(struct request_sock
*req
,
611 unsigned int mss
, struct sk_buff
*skb
,
612 struct tcp_out_options
*opts
,
613 const struct tcp_md5sig_key
*md5
,
614 struct tcp_fastopen_cookie
*foc
)
616 struct inet_request_sock
*ireq
= inet_rsk(req
);
617 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
619 #ifdef CONFIG_TCP_MD5SIG
621 opts
->options
|= OPTION_MD5
;
622 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
624 /* We can't fit any SACK blocks in a packet with MD5 + TS
625 * options. There was discussion about disabling SACK
626 * rather than TS in order to fit in better with old,
627 * buggy kernels, but that was deemed to be unnecessary.
629 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
633 /* We always send an MSS option. */
635 remaining
-= TCPOLEN_MSS_ALIGNED
;
637 if (likely(ireq
->wscale_ok
)) {
638 opts
->ws
= ireq
->rcv_wscale
;
639 opts
->options
|= OPTION_WSCALE
;
640 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
642 if (likely(ireq
->tstamp_ok
)) {
643 opts
->options
|= OPTION_TS
;
644 opts
->tsval
= tcp_skb_timestamp(skb
) + tcp_rsk(req
)->ts_off
;
645 opts
->tsecr
= req
->ts_recent
;
646 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
648 if (likely(ireq
->sack_ok
)) {
649 opts
->options
|= OPTION_SACK_ADVERTISE
;
650 if (unlikely(!ireq
->tstamp_ok
))
651 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
653 if (foc
!= NULL
&& foc
->len
>= 0) {
656 need
+= foc
->exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
657 TCPOLEN_FASTOPEN_BASE
;
658 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
659 if (remaining
>= need
) {
660 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
661 opts
->fastopen_cookie
= foc
;
666 return MAX_TCP_OPTION_SPACE
- remaining
;
669 /* Compute TCP options for ESTABLISHED sockets. This is not the
670 * final wire format yet.
672 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
673 struct tcp_out_options
*opts
,
674 struct tcp_md5sig_key
**md5
)
676 struct tcp_sock
*tp
= tcp_sk(sk
);
677 unsigned int size
= 0;
678 unsigned int eff_sacks
;
682 #ifdef CONFIG_TCP_MD5SIG
683 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
684 if (unlikely(*md5
)) {
685 opts
->options
|= OPTION_MD5
;
686 size
+= TCPOLEN_MD5SIG_ALIGNED
;
692 if (likely(tp
->rx_opt
.tstamp_ok
)) {
693 opts
->options
|= OPTION_TS
;
694 opts
->tsval
= skb
? tcp_skb_timestamp(skb
) + tp
->tsoffset
: 0;
695 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
696 size
+= TCPOLEN_TSTAMP_ALIGNED
;
699 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
700 if (unlikely(eff_sacks
)) {
701 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
702 opts
->num_sack_blocks
=
703 min_t(unsigned int, eff_sacks
,
704 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
705 TCPOLEN_SACK_PERBLOCK
);
706 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
707 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
714 /* TCP SMALL QUEUES (TSQ)
716 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
717 * to reduce RTT and bufferbloat.
718 * We do this using a special skb destructor (tcp_wfree).
720 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
721 * needs to be reallocated in a driver.
722 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
724 * Since transmit from skb destructor is forbidden, we use a tasklet
725 * to process all sockets that eventually need to send more skbs.
726 * We use one tasklet per cpu, with its own queue of sockets.
729 struct tasklet_struct tasklet
;
730 struct list_head head
; /* queue of tcp sockets */
732 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
734 static void tcp_tsq_handler(struct sock
*sk
)
736 if ((1 << sk
->sk_state
) &
737 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
738 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
)) {
739 struct tcp_sock
*tp
= tcp_sk(sk
);
741 if (tp
->lost_out
> tp
->retrans_out
&&
742 tp
->snd_cwnd
> tcp_packets_in_flight(tp
))
743 tcp_xmit_retransmit_queue(sk
);
745 tcp_write_xmit(sk
, tcp_current_mss(sk
), tp
->nonagle
,
750 * One tasklet per cpu tries to send more skbs.
751 * We run in tasklet context but need to disable irqs when
752 * transferring tsq->head because tcp_wfree() might
753 * interrupt us (non NAPI drivers)
755 static void tcp_tasklet_func(unsigned long data
)
757 struct tsq_tasklet
*tsq
= (struct tsq_tasklet
*)data
;
760 struct list_head
*q
, *n
;
764 local_irq_save(flags
);
765 list_splice_init(&tsq
->head
, &list
);
766 local_irq_restore(flags
);
768 list_for_each_safe(q
, n
, &list
) {
769 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
770 list_del(&tp
->tsq_node
);
772 sk
= (struct sock
*)tp
;
773 smp_mb__before_atomic();
774 clear_bit(TSQ_QUEUED
, &sk
->sk_tsq_flags
);
776 if (!sk
->sk_lock
.owned
&&
777 test_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
)) {
779 if (!sock_owned_by_user(sk
)) {
780 clear_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
);
790 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
791 TCPF_WRITE_TIMER_DEFERRED | \
792 TCPF_DELACK_TIMER_DEFERRED | \
793 TCPF_MTU_REDUCED_DEFERRED)
795 * tcp_release_cb - tcp release_sock() callback
798 * called from release_sock() to perform protocol dependent
799 * actions before socket release.
801 void tcp_release_cb(struct sock
*sk
)
803 unsigned long flags
, nflags
;
805 /* perform an atomic operation only if at least one flag is set */
807 flags
= sk
->sk_tsq_flags
;
808 if (!(flags
& TCP_DEFERRED_ALL
))
810 nflags
= flags
& ~TCP_DEFERRED_ALL
;
811 } while (cmpxchg(&sk
->sk_tsq_flags
, flags
, nflags
) != flags
);
813 if (flags
& TCPF_TSQ_DEFERRED
)
816 /* Here begins the tricky part :
817 * We are called from release_sock() with :
819 * 2) sk_lock.slock spinlock held
820 * 3) socket owned by us (sk->sk_lock.owned == 1)
822 * But following code is meant to be called from BH handlers,
823 * so we should keep BH disabled, but early release socket ownership
825 sock_release_ownership(sk
);
827 if (flags
& TCPF_WRITE_TIMER_DEFERRED
) {
828 tcp_write_timer_handler(sk
);
831 if (flags
& TCPF_DELACK_TIMER_DEFERRED
) {
832 tcp_delack_timer_handler(sk
);
835 if (flags
& TCPF_MTU_REDUCED_DEFERRED
) {
836 inet_csk(sk
)->icsk_af_ops
->mtu_reduced(sk
);
840 EXPORT_SYMBOL(tcp_release_cb
);
842 void __init
tcp_tasklet_init(void)
846 for_each_possible_cpu(i
) {
847 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
849 INIT_LIST_HEAD(&tsq
->head
);
850 tasklet_init(&tsq
->tasklet
,
857 * Write buffer destructor automatically called from kfree_skb.
858 * We can't xmit new skbs from this context, as we might already
861 void tcp_wfree(struct sk_buff
*skb
)
863 struct sock
*sk
= skb
->sk
;
864 struct tcp_sock
*tp
= tcp_sk(sk
);
865 unsigned long flags
, nval
, oval
;
867 /* Keep one reference on sk_wmem_alloc.
868 * Will be released by sk_free() from here or tcp_tasklet_func()
870 WARN_ON(refcount_sub_and_test(skb
->truesize
- 1, &sk
->sk_wmem_alloc
));
872 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
873 * Wait until our queues (qdisc + devices) are drained.
875 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
876 * - chance for incoming ACK (processed by another cpu maybe)
877 * to migrate this flow (skb->ooo_okay will be eventually set)
879 if (refcount_read(&sk
->sk_wmem_alloc
) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current
)
882 for (oval
= READ_ONCE(sk
->sk_tsq_flags
);; oval
= nval
) {
883 struct tsq_tasklet
*tsq
;
886 if (!(oval
& TSQF_THROTTLED
) || (oval
& TSQF_QUEUED
))
889 nval
= (oval
& ~TSQF_THROTTLED
) | TSQF_QUEUED
| TCPF_TSQ_DEFERRED
;
890 nval
= cmpxchg(&sk
->sk_tsq_flags
, oval
, nval
);
894 /* queue this socket to tasklet queue */
895 local_irq_save(flags
);
896 tsq
= this_cpu_ptr(&tsq_tasklet
);
897 empty
= list_empty(&tsq
->head
);
898 list_add(&tp
->tsq_node
, &tsq
->head
);
900 tasklet_schedule(&tsq
->tasklet
);
901 local_irq_restore(flags
);
908 /* Note: Called under hard irq.
909 * We can not call TCP stack right away.
911 enum hrtimer_restart
tcp_pace_kick(struct hrtimer
*timer
)
913 struct tcp_sock
*tp
= container_of(timer
, struct tcp_sock
, pacing_timer
);
914 struct sock
*sk
= (struct sock
*)tp
;
915 unsigned long nval
, oval
;
917 for (oval
= READ_ONCE(sk
->sk_tsq_flags
);; oval
= nval
) {
918 struct tsq_tasklet
*tsq
;
921 if (oval
& TSQF_QUEUED
)
924 nval
= (oval
& ~TSQF_THROTTLED
) | TSQF_QUEUED
| TCPF_TSQ_DEFERRED
;
925 nval
= cmpxchg(&sk
->sk_tsq_flags
, oval
, nval
);
929 if (!refcount_inc_not_zero(&sk
->sk_wmem_alloc
))
931 /* queue this socket to tasklet queue */
932 tsq
= this_cpu_ptr(&tsq_tasklet
);
933 empty
= list_empty(&tsq
->head
);
934 list_add(&tp
->tsq_node
, &tsq
->head
);
936 tasklet_schedule(&tsq
->tasklet
);
939 return HRTIMER_NORESTART
;
942 /* BBR congestion control needs pacing.
943 * Same remark for SO_MAX_PACING_RATE.
944 * sch_fq packet scheduler is efficiently handling pacing,
945 * but is not always installed/used.
946 * Return true if TCP stack should pace packets itself.
948 static bool tcp_needs_internal_pacing(const struct sock
*sk
)
950 return smp_load_acquire(&sk
->sk_pacing_status
) == SK_PACING_NEEDED
;
953 static void tcp_internal_pacing(struct sock
*sk
, const struct sk_buff
*skb
)
958 if (!tcp_needs_internal_pacing(sk
))
960 rate
= sk
->sk_pacing_rate
;
961 if (!rate
|| rate
== ~0U)
964 /* Should account for header sizes as sch_fq does,
965 * but lets make things simple.
967 len_ns
= (u64
)skb
->len
* NSEC_PER_SEC
;
968 do_div(len_ns
, rate
);
969 hrtimer_start(&tcp_sk(sk
)->pacing_timer
,
970 ktime_add_ns(ktime_get(), len_ns
),
971 HRTIMER_MODE_ABS_PINNED
);
974 /* This routine actually transmits TCP packets queued in by
975 * tcp_do_sendmsg(). This is used by both the initial
976 * transmission and possible later retransmissions.
977 * All SKB's seen here are completely headerless. It is our
978 * job to build the TCP header, and pass the packet down to
979 * IP so it can do the same plus pass the packet off to the
982 * We are working here with either a clone of the original
983 * SKB, or a fresh unique copy made by the retransmit engine.
985 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
988 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
989 struct inet_sock
*inet
;
991 struct tcp_skb_cb
*tcb
;
992 struct tcp_out_options opts
;
993 unsigned int tcp_options_size
, tcp_header_size
;
994 struct sk_buff
*oskb
= NULL
;
995 struct tcp_md5sig_key
*md5
;
999 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
1003 TCP_SKB_CB(skb
)->tx
.in_flight
= TCP_SKB_CB(skb
)->end_seq
1006 if (unlikely(skb_cloned(skb
)))
1007 skb
= pskb_copy(skb
, gfp_mask
);
1009 skb
= skb_clone(skb
, gfp_mask
);
1013 skb
->skb_mstamp
= tp
->tcp_mstamp
;
1016 tcb
= TCP_SKB_CB(skb
);
1017 memset(&opts
, 0, sizeof(opts
));
1019 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
))
1020 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
1022 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
1024 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
1026 /* if no packet is in qdisc/device queue, then allow XPS to select
1027 * another queue. We can be called from tcp_tsq_handler()
1028 * which holds one reference to sk_wmem_alloc.
1030 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1031 * One way to get this would be to set skb->truesize = 2 on them.
1033 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) < SKB_TRUESIZE(1);
1035 /* If we had to use memory reserve to allocate this skb,
1036 * this might cause drops if packet is looped back :
1037 * Other socket might not have SOCK_MEMALLOC.
1038 * Packets not looped back do not care about pfmemalloc.
1040 skb
->pfmemalloc
= 0;
1042 skb_push(skb
, tcp_header_size
);
1043 skb_reset_transport_header(skb
);
1047 skb
->destructor
= skb_is_tcp_pure_ack(skb
) ? __sock_wfree
: tcp_wfree
;
1048 skb_set_hash_from_sk(skb
, sk
);
1049 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
1051 skb_set_dst_pending_confirm(skb
, sk
->sk_dst_pending_confirm
);
1053 /* Build TCP header and checksum it. */
1054 th
= (struct tcphdr
*)skb
->data
;
1055 th
->source
= inet
->inet_sport
;
1056 th
->dest
= inet
->inet_dport
;
1057 th
->seq
= htonl(tcb
->seq
);
1058 th
->ack_seq
= htonl(tp
->rcv_nxt
);
1059 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
1065 /* The urg_mode check is necessary during a below snd_una win probe */
1066 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
1067 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
1068 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
1070 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
1071 th
->urg_ptr
= htons(0xFFFF);
1076 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
1077 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
1078 if (likely(!(tcb
->tcp_flags
& TCPHDR_SYN
))) {
1079 th
->window
= htons(tcp_select_window(sk
));
1080 tcp_ecn_send(sk
, skb
, th
, tcp_header_size
);
1082 /* RFC1323: The window in SYN & SYN/ACK segments
1085 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
1087 #ifdef CONFIG_TCP_MD5SIG
1088 /* Calculate the MD5 hash, as we have all we need now */
1090 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
1091 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
1096 icsk
->icsk_af_ops
->send_check(sk
, skb
);
1098 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
1099 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
));
1101 if (skb
->len
!= tcp_header_size
) {
1102 tcp_event_data_sent(tp
, sk
);
1103 tp
->data_segs_out
+= tcp_skb_pcount(skb
);
1104 tcp_internal_pacing(sk
, skb
);
1107 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
1108 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
1109 tcp_skb_pcount(skb
));
1111 tp
->segs_out
+= tcp_skb_pcount(skb
);
1112 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1113 skb_shinfo(skb
)->gso_segs
= tcp_skb_pcount(skb
);
1114 skb_shinfo(skb
)->gso_size
= tcp_skb_mss(skb
);
1116 /* Our usage of tstamp should remain private */
1119 /* Cleanup our debris for IP stacks */
1120 memset(skb
->cb
, 0, max(sizeof(struct inet_skb_parm
),
1121 sizeof(struct inet6_skb_parm
)));
1123 err
= icsk
->icsk_af_ops
->queue_xmit(sk
, skb
, &inet
->cork
.fl
);
1125 if (unlikely(err
> 0)) {
1127 err
= net_xmit_eval(err
);
1130 oskb
->skb_mstamp
= tp
->tcp_mstamp
;
1131 tcp_rate_skb_sent(sk
, oskb
);
1136 /* This routine just queues the buffer for sending.
1138 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1139 * otherwise socket can stall.
1141 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1143 struct tcp_sock
*tp
= tcp_sk(sk
);
1145 /* Advance write_seq and place onto the write_queue. */
1146 tp
->write_seq
= TCP_SKB_CB(skb
)->end_seq
;
1147 __skb_header_release(skb
);
1148 tcp_add_write_queue_tail(sk
, skb
);
1149 sk
->sk_wmem_queued
+= skb
->truesize
;
1150 sk_mem_charge(sk
, skb
->truesize
);
1153 /* Initialize TSO segments for a packet. */
1154 static void tcp_set_skb_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1156 if (skb
->len
<= mss_now
|| skb
->ip_summed
== CHECKSUM_NONE
) {
1157 /* Avoid the costly divide in the normal
1160 tcp_skb_pcount_set(skb
, 1);
1161 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1163 tcp_skb_pcount_set(skb
, DIV_ROUND_UP(skb
->len
, mss_now
));
1164 TCP_SKB_CB(skb
)->tcp_gso_size
= mss_now
;
1168 /* When a modification to fackets out becomes necessary, we need to check
1169 * skb is counted to fackets_out or not.
1171 static void tcp_adjust_fackets_out(struct sock
*sk
, const struct sk_buff
*skb
,
1174 struct tcp_sock
*tp
= tcp_sk(sk
);
1176 if (!tp
->sacked_out
|| tcp_is_reno(tp
))
1179 if (after(tcp_highest_sack_seq(tp
), TCP_SKB_CB(skb
)->seq
))
1180 tp
->fackets_out
-= decr
;
1183 /* Pcount in the middle of the write queue got changed, we need to do various
1184 * tweaks to fix counters
1186 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1188 struct tcp_sock
*tp
= tcp_sk(sk
);
1190 tp
->packets_out
-= decr
;
1192 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1193 tp
->sacked_out
-= decr
;
1194 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1195 tp
->retrans_out
-= decr
;
1196 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1197 tp
->lost_out
-= decr
;
1199 /* Reno case is special. Sigh... */
1200 if (tcp_is_reno(tp
) && decr
> 0)
1201 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1203 tcp_adjust_fackets_out(sk
, skb
, decr
);
1205 if (tp
->lost_skb_hint
&&
1206 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1207 (tcp_is_fack(tp
) || (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)))
1208 tp
->lost_cnt_hint
-= decr
;
1210 tcp_verify_left_out(tp
);
1213 static bool tcp_has_tx_tstamp(const struct sk_buff
*skb
)
1215 return TCP_SKB_CB(skb
)->txstamp_ack
||
1216 (skb_shinfo(skb
)->tx_flags
& SKBTX_ANY_TSTAMP
);
1219 static void tcp_fragment_tstamp(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1221 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1223 if (unlikely(tcp_has_tx_tstamp(skb
)) &&
1224 !before(shinfo
->tskey
, TCP_SKB_CB(skb2
)->seq
)) {
1225 struct skb_shared_info
*shinfo2
= skb_shinfo(skb2
);
1226 u8 tsflags
= shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
1228 shinfo
->tx_flags
&= ~tsflags
;
1229 shinfo2
->tx_flags
|= tsflags
;
1230 swap(shinfo
->tskey
, shinfo2
->tskey
);
1231 TCP_SKB_CB(skb2
)->txstamp_ack
= TCP_SKB_CB(skb
)->txstamp_ack
;
1232 TCP_SKB_CB(skb
)->txstamp_ack
= 0;
1236 static void tcp_skb_fragment_eor(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1238 TCP_SKB_CB(skb2
)->eor
= TCP_SKB_CB(skb
)->eor
;
1239 TCP_SKB_CB(skb
)->eor
= 0;
1242 /* Function to create two new TCP segments. Shrinks the given segment
1243 * to the specified size and appends a new segment with the rest of the
1244 * packet to the list. This won't be called frequently, I hope.
1245 * Remember, these are still headerless SKBs at this point.
1247 int tcp_fragment(struct sock
*sk
, struct sk_buff
*skb
, u32 len
,
1248 unsigned int mss_now
, gfp_t gfp
)
1250 struct tcp_sock
*tp
= tcp_sk(sk
);
1251 struct sk_buff
*buff
;
1252 int nsize
, old_factor
;
1256 if (WARN_ON(len
> skb
->len
))
1259 nsize
= skb_headlen(skb
) - len
;
1263 if (skb_unclone(skb
, gfp
))
1266 /* Get a new skb... force flag on. */
1267 buff
= sk_stream_alloc_skb(sk
, nsize
, gfp
, true);
1269 return -ENOMEM
; /* We'll just try again later. */
1271 sk
->sk_wmem_queued
+= buff
->truesize
;
1272 sk_mem_charge(sk
, buff
->truesize
);
1273 nlen
= skb
->len
- len
- nsize
;
1274 buff
->truesize
+= nlen
;
1275 skb
->truesize
-= nlen
;
1277 /* Correct the sequence numbers. */
1278 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1279 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1280 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1282 /* PSH and FIN should only be set in the second packet. */
1283 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1284 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1285 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1286 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1287 tcp_skb_fragment_eor(skb
, buff
);
1289 if (!skb_shinfo(skb
)->nr_frags
&& skb
->ip_summed
!= CHECKSUM_PARTIAL
) {
1290 /* Copy and checksum data tail into the new buffer. */
1291 buff
->csum
= csum_partial_copy_nocheck(skb
->data
+ len
,
1292 skb_put(buff
, nsize
),
1297 skb
->csum
= csum_block_sub(skb
->csum
, buff
->csum
, len
);
1299 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1300 skb_split(skb
, buff
, len
);
1303 buff
->ip_summed
= skb
->ip_summed
;
1305 buff
->tstamp
= skb
->tstamp
;
1306 tcp_fragment_tstamp(skb
, buff
);
1308 old_factor
= tcp_skb_pcount(skb
);
1310 /* Fix up tso_factor for both original and new SKB. */
1311 tcp_set_skb_tso_segs(skb
, mss_now
);
1312 tcp_set_skb_tso_segs(buff
, mss_now
);
1314 /* Update delivered info for the new segment */
1315 TCP_SKB_CB(buff
)->tx
= TCP_SKB_CB(skb
)->tx
;
1317 /* If this packet has been sent out already, we must
1318 * adjust the various packet counters.
1320 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1321 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1322 tcp_skb_pcount(buff
);
1325 tcp_adjust_pcount(sk
, skb
, diff
);
1328 /* Link BUFF into the send queue. */
1329 __skb_header_release(buff
);
1330 tcp_insert_write_queue_after(skb
, buff
, sk
);
1335 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1336 * data is not copied, but immediately discarded.
1338 static int __pskb_trim_head(struct sk_buff
*skb
, int len
)
1340 struct skb_shared_info
*shinfo
;
1343 eat
= min_t(int, len
, skb_headlen(skb
));
1345 __skb_pull(skb
, eat
);
1352 shinfo
= skb_shinfo(skb
);
1353 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1354 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1357 skb_frag_unref(skb
, i
);
1360 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1362 shinfo
->frags
[k
].page_offset
+= eat
;
1363 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1369 shinfo
->nr_frags
= k
;
1371 skb
->data_len
-= len
;
1372 skb
->len
= skb
->data_len
;
1376 /* Remove acked data from a packet in the transmit queue. */
1377 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1381 if (skb_unclone(skb
, GFP_ATOMIC
))
1384 delta_truesize
= __pskb_trim_head(skb
, len
);
1386 TCP_SKB_CB(skb
)->seq
+= len
;
1387 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1389 if (delta_truesize
) {
1390 skb
->truesize
-= delta_truesize
;
1391 sk
->sk_wmem_queued
-= delta_truesize
;
1392 sk_mem_uncharge(sk
, delta_truesize
);
1393 sock_set_flag(sk
, SOCK_QUEUE_SHRUNK
);
1396 /* Any change of skb->len requires recalculation of tso factor. */
1397 if (tcp_skb_pcount(skb
) > 1)
1398 tcp_set_skb_tso_segs(skb
, tcp_skb_mss(skb
));
1403 /* Calculate MSS not accounting any TCP options. */
1404 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1406 const struct tcp_sock
*tp
= tcp_sk(sk
);
1407 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1410 /* Calculate base mss without TCP options:
1411 It is MMS_S - sizeof(tcphdr) of rfc1122
1413 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1415 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1416 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1417 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1419 if (dst
&& dst_allfrag(dst
))
1420 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1423 /* Clamp it (mss_clamp does not include tcp options) */
1424 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1425 mss_now
= tp
->rx_opt
.mss_clamp
;
1427 /* Now subtract optional transport overhead */
1428 mss_now
-= icsk
->icsk_ext_hdr_len
;
1430 /* Then reserve room for full set of TCP options and 8 bytes of data */
1436 /* Calculate MSS. Not accounting for SACKs here. */
1437 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1439 /* Subtract TCP options size, not including SACKs */
1440 return __tcp_mtu_to_mss(sk
, pmtu
) -
1441 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1444 /* Inverse of above */
1445 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1447 const struct tcp_sock
*tp
= tcp_sk(sk
);
1448 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1452 tp
->tcp_header_len
+
1453 icsk
->icsk_ext_hdr_len
+
1454 icsk
->icsk_af_ops
->net_header_len
;
1456 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1457 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1458 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1460 if (dst
&& dst_allfrag(dst
))
1461 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1465 EXPORT_SYMBOL(tcp_mss_to_mtu
);
1467 /* MTU probing init per socket */
1468 void tcp_mtup_init(struct sock
*sk
)
1470 struct tcp_sock
*tp
= tcp_sk(sk
);
1471 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1472 struct net
*net
= sock_net(sk
);
1474 icsk
->icsk_mtup
.enabled
= net
->ipv4
.sysctl_tcp_mtu_probing
> 1;
1475 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1476 icsk
->icsk_af_ops
->net_header_len
;
1477 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, net
->ipv4
.sysctl_tcp_base_mss
);
1478 icsk
->icsk_mtup
.probe_size
= 0;
1479 if (icsk
->icsk_mtup
.enabled
)
1480 icsk
->icsk_mtup
.probe_timestamp
= tcp_jiffies32
;
1482 EXPORT_SYMBOL(tcp_mtup_init
);
1484 /* This function synchronize snd mss to current pmtu/exthdr set.
1486 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1487 for TCP options, but includes only bare TCP header.
1489 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1490 It is minimum of user_mss and mss received with SYN.
1491 It also does not include TCP options.
1493 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1495 tp->mss_cache is current effective sending mss, including
1496 all tcp options except for SACKs. It is evaluated,
1497 taking into account current pmtu, but never exceeds
1498 tp->rx_opt.mss_clamp.
1500 NOTE1. rfc1122 clearly states that advertised MSS
1501 DOES NOT include either tcp or ip options.
1503 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1504 are READ ONLY outside this function. --ANK (980731)
1506 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1508 struct tcp_sock
*tp
= tcp_sk(sk
);
1509 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1512 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1513 icsk
->icsk_mtup
.search_high
= pmtu
;
1515 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1516 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1518 /* And store cached results */
1519 icsk
->icsk_pmtu_cookie
= pmtu
;
1520 if (icsk
->icsk_mtup
.enabled
)
1521 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1522 tp
->mss_cache
= mss_now
;
1526 EXPORT_SYMBOL(tcp_sync_mss
);
1528 /* Compute the current effective MSS, taking SACKs and IP options,
1529 * and even PMTU discovery events into account.
1531 unsigned int tcp_current_mss(struct sock
*sk
)
1533 const struct tcp_sock
*tp
= tcp_sk(sk
);
1534 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1536 unsigned int header_len
;
1537 struct tcp_out_options opts
;
1538 struct tcp_md5sig_key
*md5
;
1540 mss_now
= tp
->mss_cache
;
1543 u32 mtu
= dst_mtu(dst
);
1544 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1545 mss_now
= tcp_sync_mss(sk
, mtu
);
1548 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1549 sizeof(struct tcphdr
);
1550 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1551 * some common options. If this is an odd packet (because we have SACK
1552 * blocks etc) then our calculated header_len will be different, and
1553 * we have to adjust mss_now correspondingly */
1554 if (header_len
!= tp
->tcp_header_len
) {
1555 int delta
= (int) header_len
- tp
->tcp_header_len
;
1562 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1563 * As additional protections, we do not touch cwnd in retransmission phases,
1564 * and if application hit its sndbuf limit recently.
1566 static void tcp_cwnd_application_limited(struct sock
*sk
)
1568 struct tcp_sock
*tp
= tcp_sk(sk
);
1570 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1571 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1572 /* Limited by application or receiver window. */
1573 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1574 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1575 if (win_used
< tp
->snd_cwnd
) {
1576 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1577 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
1579 tp
->snd_cwnd_used
= 0;
1581 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1584 static void tcp_cwnd_validate(struct sock
*sk
, bool is_cwnd_limited
)
1586 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1587 struct tcp_sock
*tp
= tcp_sk(sk
);
1589 /* Track the maximum number of outstanding packets in each
1590 * window, and remember whether we were cwnd-limited then.
1592 if (!before(tp
->snd_una
, tp
->max_packets_seq
) ||
1593 tp
->packets_out
> tp
->max_packets_out
) {
1594 tp
->max_packets_out
= tp
->packets_out
;
1595 tp
->max_packets_seq
= tp
->snd_nxt
;
1596 tp
->is_cwnd_limited
= is_cwnd_limited
;
1599 if (tcp_is_cwnd_limited(sk
)) {
1600 /* Network is feed fully. */
1601 tp
->snd_cwnd_used
= 0;
1602 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1604 /* Network starves. */
1605 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1606 tp
->snd_cwnd_used
= tp
->packets_out
;
1608 if (sysctl_tcp_slow_start_after_idle
&&
1609 (s32
)(tcp_jiffies32
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
&&
1610 !ca_ops
->cong_control
)
1611 tcp_cwnd_application_limited(sk
);
1613 /* The following conditions together indicate the starvation
1614 * is caused by insufficient sender buffer:
1615 * 1) just sent some data (see tcp_write_xmit)
1616 * 2) not cwnd limited (this else condition)
1617 * 3) no more data to send (null tcp_send_head )
1618 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1620 if (!tcp_send_head(sk
) && sk
->sk_socket
&&
1621 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
) &&
1622 (1 << sk
->sk_state
) & (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
1623 tcp_chrono_start(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1627 /* Minshall's variant of the Nagle send check. */
1628 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1630 return after(tp
->snd_sml
, tp
->snd_una
) &&
1631 !after(tp
->snd_sml
, tp
->snd_nxt
);
1634 /* Update snd_sml if this skb is under mss
1635 * Note that a TSO packet might end with a sub-mss segment
1636 * The test is really :
1637 * if ((skb->len % mss) != 0)
1638 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1639 * But we can avoid doing the divide again given we already have
1640 * skb_pcount = skb->len / mss_now
1642 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1643 const struct sk_buff
*skb
)
1645 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1646 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1649 /* Return false, if packet can be sent now without violation Nagle's rules:
1650 * 1. It is full sized. (provided by caller in %partial bool)
1651 * 2. Or it contains FIN. (already checked by caller)
1652 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1653 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1654 * With Minshall's modification: all sent small packets are ACKed.
1656 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1660 ((nonagle
& TCP_NAGLE_CORK
) ||
1661 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1664 /* Return how many segs we'd like on a TSO packet,
1665 * to send one TSO packet per ms
1667 u32
tcp_tso_autosize(const struct sock
*sk
, unsigned int mss_now
,
1672 bytes
= min(sk
->sk_pacing_rate
>> 10,
1673 sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
);
1675 /* Goal is to send at least one packet per ms,
1676 * not one big TSO packet every 100 ms.
1677 * This preserves ACK clocking and is consistent
1678 * with tcp_tso_should_defer() heuristic.
1680 segs
= max_t(u32
, bytes
/ mss_now
, min_tso_segs
);
1682 return min_t(u32
, segs
, sk
->sk_gso_max_segs
);
1684 EXPORT_SYMBOL(tcp_tso_autosize
);
1686 /* Return the number of segments we want in the skb we are transmitting.
1687 * See if congestion control module wants to decide; otherwise, autosize.
1689 static u32
tcp_tso_segs(struct sock
*sk
, unsigned int mss_now
)
1691 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1692 u32 tso_segs
= ca_ops
->tso_segs_goal
? ca_ops
->tso_segs_goal(sk
) : 0;
1695 tcp_tso_autosize(sk
, mss_now
, sysctl_tcp_min_tso_segs
);
1698 /* Returns the portion of skb which can be sent right away */
1699 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1700 const struct sk_buff
*skb
,
1701 unsigned int mss_now
,
1702 unsigned int max_segs
,
1705 const struct tcp_sock
*tp
= tcp_sk(sk
);
1706 u32 partial
, needed
, window
, max_len
;
1708 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1709 max_len
= mss_now
* max_segs
;
1711 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
1714 needed
= min(skb
->len
, window
);
1716 if (max_len
<= needed
)
1719 partial
= needed
% mss_now
;
1720 /* If last segment is not a full MSS, check if Nagle rules allow us
1721 * to include this last segment in this skb.
1722 * Otherwise, we'll split the skb at last MSS boundary
1724 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
1725 return needed
- partial
;
1730 /* Can at least one segment of SKB be sent right now, according to the
1731 * congestion window rules? If so, return how many segments are allowed.
1733 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
1734 const struct sk_buff
*skb
)
1736 u32 in_flight
, cwnd
, halfcwnd
;
1738 /* Don't be strict about the congestion window for the final FIN. */
1739 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
1740 tcp_skb_pcount(skb
) == 1)
1743 in_flight
= tcp_packets_in_flight(tp
);
1744 cwnd
= tp
->snd_cwnd
;
1745 if (in_flight
>= cwnd
)
1748 /* For better scheduling, ensure we have at least
1749 * 2 GSO packets in flight.
1751 halfcwnd
= max(cwnd
>> 1, 1U);
1752 return min(halfcwnd
, cwnd
- in_flight
);
1755 /* Initialize TSO state of a skb.
1756 * This must be invoked the first time we consider transmitting
1757 * SKB onto the wire.
1759 static int tcp_init_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1761 int tso_segs
= tcp_skb_pcount(skb
);
1763 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
1764 tcp_set_skb_tso_segs(skb
, mss_now
);
1765 tso_segs
= tcp_skb_pcount(skb
);
1771 /* Return true if the Nagle test allows this packet to be
1774 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
1775 unsigned int cur_mss
, int nonagle
)
1777 /* Nagle rule does not apply to frames, which sit in the middle of the
1778 * write_queue (they have no chances to get new data).
1780 * This is implemented in the callers, where they modify the 'nonagle'
1781 * argument based upon the location of SKB in the send queue.
1783 if (nonagle
& TCP_NAGLE_PUSH
)
1786 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1787 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
1790 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
1796 /* Does at least the first segment of SKB fit into the send window? */
1797 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
1798 const struct sk_buff
*skb
,
1799 unsigned int cur_mss
)
1801 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1803 if (skb
->len
> cur_mss
)
1804 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
1806 return !after(end_seq
, tcp_wnd_end(tp
));
1809 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1810 * which is put after SKB on the list. It is very much like
1811 * tcp_fragment() except that it may make several kinds of assumptions
1812 * in order to speed up the splitting operation. In particular, we
1813 * know that all the data is in scatter-gather pages, and that the
1814 * packet has never been sent out before (and thus is not cloned).
1816 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
1817 unsigned int mss_now
, gfp_t gfp
)
1819 struct sk_buff
*buff
;
1820 int nlen
= skb
->len
- len
;
1823 /* All of a TSO frame must be composed of paged data. */
1824 if (skb
->len
!= skb
->data_len
)
1825 return tcp_fragment(sk
, skb
, len
, mss_now
, gfp
);
1827 buff
= sk_stream_alloc_skb(sk
, 0, gfp
, true);
1828 if (unlikely(!buff
))
1831 sk
->sk_wmem_queued
+= buff
->truesize
;
1832 sk_mem_charge(sk
, buff
->truesize
);
1833 buff
->truesize
+= nlen
;
1834 skb
->truesize
-= nlen
;
1836 /* Correct the sequence numbers. */
1837 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1838 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1839 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1841 /* PSH and FIN should only be set in the second packet. */
1842 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1843 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1844 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1846 /* This packet was never sent out yet, so no SACK bits. */
1847 TCP_SKB_CB(buff
)->sacked
= 0;
1849 tcp_skb_fragment_eor(skb
, buff
);
1851 buff
->ip_summed
= skb
->ip_summed
= CHECKSUM_PARTIAL
;
1852 skb_split(skb
, buff
, len
);
1853 tcp_fragment_tstamp(skb
, buff
);
1855 /* Fix up tso_factor for both original and new SKB. */
1856 tcp_set_skb_tso_segs(skb
, mss_now
);
1857 tcp_set_skb_tso_segs(buff
, mss_now
);
1859 /* Link BUFF into the send queue. */
1860 __skb_header_release(buff
);
1861 tcp_insert_write_queue_after(skb
, buff
, sk
);
1866 /* Try to defer sending, if possible, in order to minimize the amount
1867 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1869 * This algorithm is from John Heffner.
1871 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
,
1872 bool *is_cwnd_limited
, u32 max_segs
)
1874 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1875 u32 age
, send_win
, cong_win
, limit
, in_flight
;
1876 struct tcp_sock
*tp
= tcp_sk(sk
);
1877 struct sk_buff
*head
;
1880 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
)
1883 if (icsk
->icsk_ca_state
>= TCP_CA_Recovery
)
1886 /* Avoid bursty behavior by allowing defer
1887 * only if the last write was recent.
1889 if ((s32
)(tcp_jiffies32
- tp
->lsndtime
) > 0)
1892 in_flight
= tcp_packets_in_flight(tp
);
1894 BUG_ON(tcp_skb_pcount(skb
) <= 1 || (tp
->snd_cwnd
<= in_flight
));
1896 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
1898 /* From in_flight test above, we know that cwnd > in_flight. */
1899 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
1901 limit
= min(send_win
, cong_win
);
1903 /* If a full-sized TSO skb can be sent, do it. */
1904 if (limit
>= max_segs
* tp
->mss_cache
)
1907 /* Middle in queue won't get any more data, full sendable already? */
1908 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
1911 win_divisor
= READ_ONCE(sysctl_tcp_tso_win_divisor
);
1913 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
1915 /* If at least some fraction of a window is available,
1918 chunk
/= win_divisor
;
1922 /* Different approach, try not to defer past a single
1923 * ACK. Receiver should ACK every other full sized
1924 * frame, so if we have space for more than 3 frames
1927 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
1931 head
= tcp_write_queue_head(sk
);
1933 age
= tcp_stamp_us_delta(tp
->tcp_mstamp
, head
->skb_mstamp
);
1934 /* If next ACK is likely to come too late (half srtt), do not defer */
1935 if (age
< (tp
->srtt_us
>> 4))
1938 /* Ok, it looks like it is advisable to defer. */
1940 if (cong_win
< send_win
&& cong_win
<= skb
->len
)
1941 *is_cwnd_limited
= true;
1949 static inline void tcp_mtu_check_reprobe(struct sock
*sk
)
1951 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1952 struct tcp_sock
*tp
= tcp_sk(sk
);
1953 struct net
*net
= sock_net(sk
);
1957 interval
= net
->ipv4
.sysctl_tcp_probe_interval
;
1958 delta
= tcp_jiffies32
- icsk
->icsk_mtup
.probe_timestamp
;
1959 if (unlikely(delta
>= interval
* HZ
)) {
1960 int mss
= tcp_current_mss(sk
);
1962 /* Update current search range */
1963 icsk
->icsk_mtup
.probe_size
= 0;
1964 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+
1965 sizeof(struct tcphdr
) +
1966 icsk
->icsk_af_ops
->net_header_len
;
1967 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
1969 /* Update probe time stamp */
1970 icsk
->icsk_mtup
.probe_timestamp
= tcp_jiffies32
;
1974 /* Create a new MTU probe if we are ready.
1975 * MTU probe is regularly attempting to increase the path MTU by
1976 * deliberately sending larger packets. This discovers routing
1977 * changes resulting in larger path MTUs.
1979 * Returns 0 if we should wait to probe (no cwnd available),
1980 * 1 if a probe was sent,
1983 static int tcp_mtu_probe(struct sock
*sk
)
1985 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1986 struct tcp_sock
*tp
= tcp_sk(sk
);
1987 struct sk_buff
*skb
, *nskb
, *next
;
1988 struct net
*net
= sock_net(sk
);
1995 /* Not currently probing/verifying,
1997 * have enough cwnd, and
1998 * not SACKing (the variable headers throw things off)
2000 if (likely(!icsk
->icsk_mtup
.enabled
||
2001 icsk
->icsk_mtup
.probe_size
||
2002 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
2003 tp
->snd_cwnd
< 11 ||
2004 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
))
2007 /* Use binary search for probe_size between tcp_mss_base,
2008 * and current mss_clamp. if (search_high - search_low)
2009 * smaller than a threshold, backoff from probing.
2011 mss_now
= tcp_current_mss(sk
);
2012 probe_size
= tcp_mtu_to_mss(sk
, (icsk
->icsk_mtup
.search_high
+
2013 icsk
->icsk_mtup
.search_low
) >> 1);
2014 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
2015 interval
= icsk
->icsk_mtup
.search_high
- icsk
->icsk_mtup
.search_low
;
2016 /* When misfortune happens, we are reprobing actively,
2017 * and then reprobe timer has expired. We stick with current
2018 * probing process by not resetting search range to its orignal.
2020 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
) ||
2021 interval
< net
->ipv4
.sysctl_tcp_probe_threshold
) {
2022 /* Check whether enough time has elaplased for
2023 * another round of probing.
2025 tcp_mtu_check_reprobe(sk
);
2029 /* Have enough data in the send queue to probe? */
2030 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
2033 if (tp
->snd_wnd
< size_needed
)
2035 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
2038 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2039 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
2040 if (!tcp_packets_in_flight(tp
))
2046 /* We're allowed to probe. Build it now. */
2047 nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
, false);
2050 sk
->sk_wmem_queued
+= nskb
->truesize
;
2051 sk_mem_charge(sk
, nskb
->truesize
);
2053 skb
= tcp_send_head(sk
);
2055 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
2056 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
2057 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
2058 TCP_SKB_CB(nskb
)->sacked
= 0;
2060 nskb
->ip_summed
= skb
->ip_summed
;
2062 tcp_insert_write_queue_before(nskb
, skb
, sk
);
2065 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
2066 copy
= min_t(int, skb
->len
, probe_size
- len
);
2067 if (nskb
->ip_summed
) {
2068 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
2070 __wsum csum
= skb_copy_and_csum_bits(skb
, 0,
2071 skb_put(nskb
, copy
),
2073 nskb
->csum
= csum_block_add(nskb
->csum
, csum
, len
);
2076 if (skb
->len
<= copy
) {
2077 /* We've eaten all the data from this skb.
2079 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
2080 tcp_unlink_write_queue(skb
, sk
);
2081 sk_wmem_free_skb(sk
, skb
);
2083 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
2084 ~(TCPHDR_FIN
|TCPHDR_PSH
);
2085 if (!skb_shinfo(skb
)->nr_frags
) {
2086 skb_pull(skb
, copy
);
2087 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2088 skb
->csum
= csum_partial(skb
->data
,
2091 __pskb_trim_head(skb
, copy
);
2092 tcp_set_skb_tso_segs(skb
, mss_now
);
2094 TCP_SKB_CB(skb
)->seq
+= copy
;
2099 if (len
>= probe_size
)
2102 tcp_init_tso_segs(nskb
, nskb
->len
);
2104 /* We're ready to send. If this fails, the probe will
2105 * be resegmented into mss-sized pieces by tcp_write_xmit().
2107 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
2108 /* Decrement cwnd here because we are sending
2109 * effectively two packets. */
2111 tcp_event_new_data_sent(sk
, nskb
);
2113 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
2114 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
2115 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
2123 static bool tcp_pacing_check(const struct sock
*sk
)
2125 return tcp_needs_internal_pacing(sk
) &&
2126 hrtimer_active(&tcp_sk(sk
)->pacing_timer
);
2129 /* TCP Small Queues :
2130 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2131 * (These limits are doubled for retransmits)
2133 * - better RTT estimation and ACK scheduling
2136 * Alas, some drivers / subsystems require a fair amount
2137 * of queued bytes to ensure line rate.
2138 * One example is wifi aggregation (802.11 AMPDU)
2140 static bool tcp_small_queue_check(struct sock
*sk
, const struct sk_buff
*skb
,
2141 unsigned int factor
)
2145 limit
= max(2 * skb
->truesize
, sk
->sk_pacing_rate
>> 10);
2146 limit
= min_t(u32
, limit
, sysctl_tcp_limit_output_bytes
);
2149 if (refcount_read(&sk
->sk_wmem_alloc
) > limit
) {
2150 /* Always send the 1st or 2nd skb in write queue.
2151 * No need to wait for TX completion to call us back,
2152 * after softirq/tasklet schedule.
2153 * This helps when TX completions are delayed too much.
2155 if (skb
== sk
->sk_write_queue
.next
||
2156 skb
->prev
== sk
->sk_write_queue
.next
)
2159 set_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
);
2160 /* It is possible TX completion already happened
2161 * before we set TSQ_THROTTLED, so we must
2162 * test again the condition.
2164 smp_mb__after_atomic();
2165 if (refcount_read(&sk
->sk_wmem_alloc
) > limit
)
2171 static void tcp_chrono_set(struct tcp_sock
*tp
, const enum tcp_chrono
new)
2173 const u32 now
= tcp_jiffies32
;
2174 enum tcp_chrono old
= tp
->chrono_type
;
2176 if (old
> TCP_CHRONO_UNSPEC
)
2177 tp
->chrono_stat
[old
- 1] += now
- tp
->chrono_start
;
2178 tp
->chrono_start
= now
;
2179 tp
->chrono_type
= new;
2182 void tcp_chrono_start(struct sock
*sk
, const enum tcp_chrono type
)
2184 struct tcp_sock
*tp
= tcp_sk(sk
);
2186 /* If there are multiple conditions worthy of tracking in a
2187 * chronograph then the highest priority enum takes precedence
2188 * over the other conditions. So that if something "more interesting"
2189 * starts happening, stop the previous chrono and start a new one.
2191 if (type
> tp
->chrono_type
)
2192 tcp_chrono_set(tp
, type
);
2195 void tcp_chrono_stop(struct sock
*sk
, const enum tcp_chrono type
)
2197 struct tcp_sock
*tp
= tcp_sk(sk
);
2200 /* There are multiple conditions worthy of tracking in a
2201 * chronograph, so that the highest priority enum takes
2202 * precedence over the other conditions (see tcp_chrono_start).
2203 * If a condition stops, we only stop chrono tracking if
2204 * it's the "most interesting" or current chrono we are
2205 * tracking and starts busy chrono if we have pending data.
2207 if (tcp_write_queue_empty(sk
))
2208 tcp_chrono_set(tp
, TCP_CHRONO_UNSPEC
);
2209 else if (type
== tp
->chrono_type
)
2210 tcp_chrono_set(tp
, TCP_CHRONO_BUSY
);
2213 /* This routine writes packets to the network. It advances the
2214 * send_head. This happens as incoming acks open up the remote
2217 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2218 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2219 * account rare use of URG, this is not a big flaw.
2221 * Send at most one packet when push_one > 0. Temporarily ignore
2222 * cwnd limit to force at most one packet out when push_one == 2.
2224 * Returns true, if no segments are in flight and we have queued segments,
2225 * but cannot send anything now because of SWS or another problem.
2227 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
2228 int push_one
, gfp_t gfp
)
2230 struct tcp_sock
*tp
= tcp_sk(sk
);
2231 struct sk_buff
*skb
;
2232 unsigned int tso_segs
, sent_pkts
;
2235 bool is_cwnd_limited
= false, is_rwnd_limited
= false;
2241 /* Do MTU probing. */
2242 result
= tcp_mtu_probe(sk
);
2245 } else if (result
> 0) {
2250 max_segs
= tcp_tso_segs(sk
, mss_now
);
2251 tcp_mstamp_refresh(tp
);
2252 while ((skb
= tcp_send_head(sk
))) {
2255 if (tcp_pacing_check(sk
))
2258 tso_segs
= tcp_init_tso_segs(skb
, mss_now
);
2261 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
) {
2262 /* "skb_mstamp" is used as a start point for the retransmit timer */
2263 skb
->skb_mstamp
= tp
->tcp_mstamp
;
2264 goto repair
; /* Skip network transmission */
2267 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2270 /* Force out a loss probe pkt. */
2276 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
))) {
2277 is_rwnd_limited
= true;
2281 if (tso_segs
== 1) {
2282 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2283 (tcp_skb_is_last(sk
, skb
) ?
2284 nonagle
: TCP_NAGLE_PUSH
))))
2288 tcp_tso_should_defer(sk
, skb
, &is_cwnd_limited
,
2294 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
2295 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2301 if (skb
->len
> limit
&&
2302 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
2305 if (test_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
))
2306 clear_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
);
2307 if (tcp_small_queue_check(sk
, skb
, 0))
2310 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2314 /* Advance the send_head. This one is sent out.
2315 * This call will increment packets_out.
2317 tcp_event_new_data_sent(sk
, skb
);
2319 tcp_minshall_update(tp
, mss_now
, skb
);
2320 sent_pkts
+= tcp_skb_pcount(skb
);
2326 if (is_rwnd_limited
)
2327 tcp_chrono_start(sk
, TCP_CHRONO_RWND_LIMITED
);
2329 tcp_chrono_stop(sk
, TCP_CHRONO_RWND_LIMITED
);
2331 if (likely(sent_pkts
)) {
2332 if (tcp_in_cwnd_reduction(sk
))
2333 tp
->prr_out
+= sent_pkts
;
2335 /* Send one loss probe per tail loss episode. */
2337 tcp_schedule_loss_probe(sk
);
2338 is_cwnd_limited
|= (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
);
2339 tcp_cwnd_validate(sk
, is_cwnd_limited
);
2342 return !tp
->packets_out
&& tcp_send_head(sk
);
2345 bool tcp_schedule_loss_probe(struct sock
*sk
)
2347 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2348 struct tcp_sock
*tp
= tcp_sk(sk
);
2349 u32 timeout
, rto_delta_us
;
2351 /* Don't do any loss probe on a Fast Open connection before 3WHS
2354 if (tp
->fastopen_rsk
)
2357 /* Schedule a loss probe in 2*RTT for SACK capable connections
2358 * in Open state, that are either limited by cwnd or application.
2360 if ((sysctl_tcp_early_retrans
!= 3 && sysctl_tcp_early_retrans
!= 4) ||
2361 !tp
->packets_out
|| !tcp_is_sack(tp
) ||
2362 icsk
->icsk_ca_state
!= TCP_CA_Open
)
2365 if ((tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) &&
2369 /* Probe timeout is 2*rtt. Add minimum RTO to account
2370 * for delayed ack when there's one outstanding packet. If no RTT
2371 * sample is available then probe after TCP_TIMEOUT_INIT.
2374 timeout
= usecs_to_jiffies(tp
->srtt_us
>> 2);
2375 if (tp
->packets_out
== 1)
2376 timeout
+= TCP_RTO_MIN
;
2378 timeout
+= TCP_TIMEOUT_MIN
;
2380 timeout
= TCP_TIMEOUT_INIT
;
2383 /* If the RTO formula yields an earlier time, then use that time. */
2384 rto_delta_us
= tcp_rto_delta_us(sk
); /* How far in future is RTO? */
2385 if (rto_delta_us
> 0)
2386 timeout
= min_t(u32
, timeout
, usecs_to_jiffies(rto_delta_us
));
2388 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
,
2393 /* Thanks to skb fast clones, we can detect if a prior transmit of
2394 * a packet is still in a qdisc or driver queue.
2395 * In this case, there is very little point doing a retransmit !
2397 static bool skb_still_in_host_queue(const struct sock
*sk
,
2398 const struct sk_buff
*skb
)
2400 if (unlikely(skb_fclone_busy(sk
, skb
))) {
2401 NET_INC_STATS(sock_net(sk
),
2402 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
2408 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2409 * retransmit the last segment.
2411 void tcp_send_loss_probe(struct sock
*sk
)
2413 struct tcp_sock
*tp
= tcp_sk(sk
);
2414 struct sk_buff
*skb
;
2416 int mss
= tcp_current_mss(sk
);
2418 skb
= tcp_send_head(sk
);
2420 if (tcp_snd_wnd_test(tp
, skb
, mss
)) {
2421 pcount
= tp
->packets_out
;
2422 tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2423 if (tp
->packets_out
> pcount
)
2427 skb
= tcp_write_queue_prev(sk
, skb
);
2429 skb
= tcp_write_queue_tail(sk
);
2432 /* At most one outstanding TLP retransmission. */
2433 if (tp
->tlp_high_seq
)
2436 /* Retransmit last segment. */
2440 if (skb_still_in_host_queue(sk
, skb
))
2443 pcount
= tcp_skb_pcount(skb
);
2444 if (WARN_ON(!pcount
))
2447 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2448 if (unlikely(tcp_fragment(sk
, skb
, (pcount
- 1) * mss
, mss
,
2451 skb
= tcp_write_queue_next(sk
, skb
);
2454 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2457 if (__tcp_retransmit_skb(sk
, skb
, 1))
2460 /* Record snd_nxt for loss detection. */
2461 tp
->tlp_high_seq
= tp
->snd_nxt
;
2464 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSPROBES
);
2465 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2466 inet_csk(sk
)->icsk_pending
= 0;
2471 /* Push out any pending frames which were held back due to
2472 * TCP_CORK or attempt at coalescing tiny packets.
2473 * The socket must be locked by the caller.
2475 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2478 /* If we are closed, the bytes will have to remain here.
2479 * In time closedown will finish, we empty the write queue and
2480 * all will be happy.
2482 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2485 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2486 sk_gfp_mask(sk
, GFP_ATOMIC
)))
2487 tcp_check_probe_timer(sk
);
2490 /* Send _single_ skb sitting at the send head. This function requires
2491 * true push pending frames to setup probe timer etc.
2493 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2495 struct sk_buff
*skb
= tcp_send_head(sk
);
2497 BUG_ON(!skb
|| skb
->len
< mss_now
);
2499 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2502 /* This function returns the amount that we can raise the
2503 * usable window based on the following constraints
2505 * 1. The window can never be shrunk once it is offered (RFC 793)
2506 * 2. We limit memory per socket
2509 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2510 * RECV.NEXT + RCV.WIN fixed until:
2511 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2513 * i.e. don't raise the right edge of the window until you can raise
2514 * it at least MSS bytes.
2516 * Unfortunately, the recommended algorithm breaks header prediction,
2517 * since header prediction assumes th->window stays fixed.
2519 * Strictly speaking, keeping th->window fixed violates the receiver
2520 * side SWS prevention criteria. The problem is that under this rule
2521 * a stream of single byte packets will cause the right side of the
2522 * window to always advance by a single byte.
2524 * Of course, if the sender implements sender side SWS prevention
2525 * then this will not be a problem.
2527 * BSD seems to make the following compromise:
2529 * If the free space is less than the 1/4 of the maximum
2530 * space available and the free space is less than 1/2 mss,
2531 * then set the window to 0.
2532 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2533 * Otherwise, just prevent the window from shrinking
2534 * and from being larger than the largest representable value.
2536 * This prevents incremental opening of the window in the regime
2537 * where TCP is limited by the speed of the reader side taking
2538 * data out of the TCP receive queue. It does nothing about
2539 * those cases where the window is constrained on the sender side
2540 * because the pipeline is full.
2542 * BSD also seems to "accidentally" limit itself to windows that are a
2543 * multiple of MSS, at least until the free space gets quite small.
2544 * This would appear to be a side effect of the mbuf implementation.
2545 * Combining these two algorithms results in the observed behavior
2546 * of having a fixed window size at almost all times.
2548 * Below we obtain similar behavior by forcing the offered window to
2549 * a multiple of the mss when it is feasible to do so.
2551 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2552 * Regular options like TIMESTAMP are taken into account.
2554 u32
__tcp_select_window(struct sock
*sk
)
2556 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2557 struct tcp_sock
*tp
= tcp_sk(sk
);
2558 /* MSS for the peer's data. Previous versions used mss_clamp
2559 * here. I don't know if the value based on our guesses
2560 * of peer's MSS is better for the performance. It's more correct
2561 * but may be worse for the performance because of rcv_mss
2562 * fluctuations. --SAW 1998/11/1
2564 int mss
= icsk
->icsk_ack
.rcv_mss
;
2565 int free_space
= tcp_space(sk
);
2566 int allowed_space
= tcp_full_space(sk
);
2567 int full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2570 if (unlikely(mss
> full_space
)) {
2575 if (free_space
< (full_space
>> 1)) {
2576 icsk
->icsk_ack
.quick
= 0;
2578 if (tcp_under_memory_pressure(sk
))
2579 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2582 /* free_space might become our new window, make sure we don't
2583 * increase it due to wscale.
2585 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2587 /* if free space is less than mss estimate, or is below 1/16th
2588 * of the maximum allowed, try to move to zero-window, else
2589 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2590 * new incoming data is dropped due to memory limits.
2591 * With large window, mss test triggers way too late in order
2592 * to announce zero window in time before rmem limit kicks in.
2594 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2598 if (free_space
> tp
->rcv_ssthresh
)
2599 free_space
= tp
->rcv_ssthresh
;
2601 /* Don't do rounding if we are using window scaling, since the
2602 * scaled window will not line up with the MSS boundary anyway.
2604 if (tp
->rx_opt
.rcv_wscale
) {
2605 window
= free_space
;
2607 /* Advertise enough space so that it won't get scaled away.
2608 * Import case: prevent zero window announcement if
2609 * 1<<rcv_wscale > mss.
2611 window
= ALIGN(window
, (1 << tp
->rx_opt
.rcv_wscale
));
2613 window
= tp
->rcv_wnd
;
2614 /* Get the largest window that is a nice multiple of mss.
2615 * Window clamp already applied above.
2616 * If our current window offering is within 1 mss of the
2617 * free space we just keep it. This prevents the divide
2618 * and multiply from happening most of the time.
2619 * We also don't do any window rounding when the free space
2622 if (window
<= free_space
- mss
|| window
> free_space
)
2623 window
= rounddown(free_space
, mss
);
2624 else if (mss
== full_space
&&
2625 free_space
> window
+ (full_space
>> 1))
2626 window
= free_space
;
2632 void tcp_skb_collapse_tstamp(struct sk_buff
*skb
,
2633 const struct sk_buff
*next_skb
)
2635 if (unlikely(tcp_has_tx_tstamp(next_skb
))) {
2636 const struct skb_shared_info
*next_shinfo
=
2637 skb_shinfo(next_skb
);
2638 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
2640 shinfo
->tx_flags
|= next_shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
2641 shinfo
->tskey
= next_shinfo
->tskey
;
2642 TCP_SKB_CB(skb
)->txstamp_ack
|=
2643 TCP_SKB_CB(next_skb
)->txstamp_ack
;
2647 /* Collapses two adjacent SKB's during retransmission. */
2648 static bool tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
2650 struct tcp_sock
*tp
= tcp_sk(sk
);
2651 struct sk_buff
*next_skb
= tcp_write_queue_next(sk
, skb
);
2652 int skb_size
, next_skb_size
;
2654 skb_size
= skb
->len
;
2655 next_skb_size
= next_skb
->len
;
2657 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
2659 if (next_skb_size
) {
2660 if (next_skb_size
<= skb_availroom(skb
))
2661 skb_copy_bits(next_skb
, 0, skb_put(skb
, next_skb_size
),
2663 else if (!skb_shift(skb
, next_skb
, next_skb_size
))
2666 tcp_highest_sack_combine(sk
, next_skb
, skb
);
2668 tcp_unlink_write_queue(next_skb
, sk
);
2670 if (next_skb
->ip_summed
== CHECKSUM_PARTIAL
)
2671 skb
->ip_summed
= CHECKSUM_PARTIAL
;
2673 if (skb
->ip_summed
!= CHECKSUM_PARTIAL
)
2674 skb
->csum
= csum_block_add(skb
->csum
, next_skb
->csum
, skb_size
);
2676 /* Update sequence range on original skb. */
2677 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
2679 /* Merge over control information. This moves PSH/FIN etc. over */
2680 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
2682 /* All done, get rid of second SKB and account for it so
2683 * packet counting does not break.
2685 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
2686 TCP_SKB_CB(skb
)->eor
= TCP_SKB_CB(next_skb
)->eor
;
2688 /* changed transmit queue under us so clear hints */
2689 tcp_clear_retrans_hints_partial(tp
);
2690 if (next_skb
== tp
->retransmit_skb_hint
)
2691 tp
->retransmit_skb_hint
= skb
;
2693 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
2695 tcp_skb_collapse_tstamp(skb
, next_skb
);
2697 sk_wmem_free_skb(sk
, next_skb
);
2701 /* Check if coalescing SKBs is legal. */
2702 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
2704 if (tcp_skb_pcount(skb
) > 1)
2706 if (skb_cloned(skb
))
2708 if (skb
== tcp_send_head(sk
))
2710 /* Some heuristics for collapsing over SACK'd could be invented */
2711 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
2717 /* Collapse packets in the retransmit queue to make to create
2718 * less packets on the wire. This is only done on retransmission.
2720 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
2723 struct tcp_sock
*tp
= tcp_sk(sk
);
2724 struct sk_buff
*skb
= to
, *tmp
;
2727 if (!sysctl_tcp_retrans_collapse
)
2729 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2732 tcp_for_write_queue_from_safe(skb
, tmp
, sk
) {
2733 if (!tcp_can_collapse(sk
, skb
))
2736 if (!tcp_skb_can_collapse_to(to
))
2749 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
2752 if (!tcp_collapse_retrans(sk
, to
))
2757 /* This retransmits one SKB. Policy decisions and retransmit queue
2758 * state updates are done by the caller. Returns non-zero if an
2759 * error occurred which prevented the send.
2761 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2763 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2764 struct tcp_sock
*tp
= tcp_sk(sk
);
2765 unsigned int cur_mss
;
2769 /* Inconclusive MTU probe */
2770 if (icsk
->icsk_mtup
.probe_size
)
2771 icsk
->icsk_mtup
.probe_size
= 0;
2773 /* Do not sent more than we queued. 1/4 is reserved for possible
2774 * copying overhead: fragmentation, tunneling, mangling etc.
2776 if (refcount_read(&sk
->sk_wmem_alloc
) >
2777 min_t(u32
, sk
->sk_wmem_queued
+ (sk
->sk_wmem_queued
>> 2),
2781 if (skb_still_in_host_queue(sk
, skb
))
2784 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
2785 if (before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))
2787 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
2791 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
2792 return -EHOSTUNREACH
; /* Routing failure or similar. */
2794 cur_mss
= tcp_current_mss(sk
);
2796 /* If receiver has shrunk his window, and skb is out of
2797 * new window, do not retransmit it. The exception is the
2798 * case, when window is shrunk to zero. In this case
2799 * our retransmit serves as a zero window probe.
2801 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
2802 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
2805 len
= cur_mss
* segs
;
2806 if (skb
->len
> len
) {
2807 if (tcp_fragment(sk
, skb
, len
, cur_mss
, GFP_ATOMIC
))
2808 return -ENOMEM
; /* We'll try again later. */
2810 if (skb_unclone(skb
, GFP_ATOMIC
))
2813 diff
= tcp_skb_pcount(skb
);
2814 tcp_set_skb_tso_segs(skb
, cur_mss
);
2815 diff
-= tcp_skb_pcount(skb
);
2817 tcp_adjust_pcount(sk
, skb
, diff
);
2818 if (skb
->len
< cur_mss
)
2819 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
2822 /* RFC3168, section 6.1.1.1. ECN fallback */
2823 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN_ECN
) == TCPHDR_SYN_ECN
)
2824 tcp_ecn_clear_syn(sk
, skb
);
2826 /* Update global and local TCP statistics. */
2827 segs
= tcp_skb_pcount(skb
);
2828 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
, segs
);
2829 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
2830 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
2831 tp
->total_retrans
+= segs
;
2833 /* make sure skb->data is aligned on arches that require it
2834 * and check if ack-trimming & collapsing extended the headroom
2835 * beyond what csum_start can cover.
2837 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
2838 skb_headroom(skb
) >= 0xFFFF)) {
2839 struct sk_buff
*nskb
;
2841 nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
, GFP_ATOMIC
);
2842 err
= nskb
? tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
) :
2845 skb
->skb_mstamp
= tp
->tcp_mstamp
;
2847 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
2851 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
2852 } else if (err
!= -EBUSY
) {
2853 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
);
2858 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
2860 struct tcp_sock
*tp
= tcp_sk(sk
);
2861 int err
= __tcp_retransmit_skb(sk
, skb
, segs
);
2864 #if FASTRETRANS_DEBUG > 0
2865 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
2866 net_dbg_ratelimited("retrans_out leaked\n");
2869 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
2870 tp
->retrans_out
+= tcp_skb_pcount(skb
);
2872 /* Save stamp of the first retransmit. */
2873 if (!tp
->retrans_stamp
)
2874 tp
->retrans_stamp
= tcp_skb_timestamp(skb
);
2878 if (tp
->undo_retrans
< 0)
2879 tp
->undo_retrans
= 0;
2880 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
2884 /* This gets called after a retransmit timeout, and the initially
2885 * retransmitted data is acknowledged. It tries to continue
2886 * resending the rest of the retransmit queue, until either
2887 * we've sent it all or the congestion window limit is reached.
2888 * If doing SACK, the first ACK which comes back for a timeout
2889 * based retransmit packet might feed us FACK information again.
2890 * If so, we use it to avoid unnecessarily retransmissions.
2892 void tcp_xmit_retransmit_queue(struct sock
*sk
)
2894 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2895 struct tcp_sock
*tp
= tcp_sk(sk
);
2896 struct sk_buff
*skb
;
2897 struct sk_buff
*hole
= NULL
;
2901 if (!tp
->packets_out
)
2904 if (tp
->retransmit_skb_hint
) {
2905 skb
= tp
->retransmit_skb_hint
;
2907 skb
= tcp_write_queue_head(sk
);
2910 max_segs
= tcp_tso_segs(sk
, tcp_current_mss(sk
));
2911 tcp_for_write_queue_from(skb
, sk
) {
2915 if (skb
== tcp_send_head(sk
))
2918 if (tcp_pacing_check(sk
))
2921 /* we could do better than to assign each time */
2923 tp
->retransmit_skb_hint
= skb
;
2925 segs
= tp
->snd_cwnd
- tcp_packets_in_flight(tp
);
2928 sacked
= TCP_SKB_CB(skb
)->sacked
;
2929 /* In case tcp_shift_skb_data() have aggregated large skbs,
2930 * we need to make sure not sending too bigs TSO packets
2932 segs
= min_t(int, segs
, max_segs
);
2934 if (tp
->retrans_out
>= tp
->lost_out
) {
2936 } else if (!(sacked
& TCPCB_LOST
)) {
2937 if (!hole
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
2942 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
2943 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
2945 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
2948 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
2951 if (tcp_small_queue_check(sk
, skb
, 1))
2954 if (tcp_retransmit_skb(sk
, skb
, segs
))
2957 NET_ADD_STATS(sock_net(sk
), mib_idx
, tcp_skb_pcount(skb
));
2959 if (tcp_in_cwnd_reduction(sk
))
2960 tp
->prr_out
+= tcp_skb_pcount(skb
);
2962 if (skb
== tcp_write_queue_head(sk
) &&
2963 icsk
->icsk_pending
!= ICSK_TIME_REO_TIMEOUT
)
2964 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
2965 inet_csk(sk
)->icsk_rto
,
2970 /* We allow to exceed memory limits for FIN packets to expedite
2971 * connection tear down and (memory) recovery.
2972 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2973 * or even be forced to close flow without any FIN.
2974 * In general, we want to allow one skb per socket to avoid hangs
2975 * with edge trigger epoll()
2977 void sk_forced_mem_schedule(struct sock
*sk
, int size
)
2981 if (size
<= sk
->sk_forward_alloc
)
2983 amt
= sk_mem_pages(size
);
2984 sk
->sk_forward_alloc
+= amt
* SK_MEM_QUANTUM
;
2985 sk_memory_allocated_add(sk
, amt
);
2987 if (mem_cgroup_sockets_enabled
&& sk
->sk_memcg
)
2988 mem_cgroup_charge_skmem(sk
->sk_memcg
, amt
);
2991 /* Send a FIN. The caller locks the socket for us.
2992 * We should try to send a FIN packet really hard, but eventually give up.
2994 void tcp_send_fin(struct sock
*sk
)
2996 struct sk_buff
*skb
, *tskb
= tcp_write_queue_tail(sk
);
2997 struct tcp_sock
*tp
= tcp_sk(sk
);
2999 /* Optimization, tack on the FIN if we have one skb in write queue and
3000 * this skb was not yet sent, or we are under memory pressure.
3001 * Note: in the latter case, FIN packet will be sent after a timeout,
3002 * as TCP stack thinks it has already been transmitted.
3004 if (tskb
&& (tcp_send_head(sk
) || tcp_under_memory_pressure(sk
))) {
3006 TCP_SKB_CB(tskb
)->tcp_flags
|= TCPHDR_FIN
;
3007 TCP_SKB_CB(tskb
)->end_seq
++;
3009 if (!tcp_send_head(sk
)) {
3010 /* This means tskb was already sent.
3011 * Pretend we included the FIN on previous transmit.
3012 * We need to set tp->snd_nxt to the value it would have
3013 * if FIN had been sent. This is because retransmit path
3014 * does not change tp->snd_nxt.
3020 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, sk
->sk_allocation
);
3021 if (unlikely(!skb
)) {
3026 skb_reserve(skb
, MAX_TCP_HEADER
);
3027 sk_forced_mem_schedule(sk
, skb
->truesize
);
3028 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3029 tcp_init_nondata_skb(skb
, tp
->write_seq
,
3030 TCPHDR_ACK
| TCPHDR_FIN
);
3031 tcp_queue_skb(sk
, skb
);
3033 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
), TCP_NAGLE_OFF
);
3036 /* We get here when a process closes a file descriptor (either due to
3037 * an explicit close() or as a byproduct of exit()'ing) and there
3038 * was unread data in the receive queue. This behavior is recommended
3039 * by RFC 2525, section 2.17. -DaveM
3041 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
3043 struct sk_buff
*skb
;
3045 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
3047 /* NOTE: No TCP options attached and we never retransmit this. */
3048 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
3050 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3054 /* Reserve space for headers and prepare control bits. */
3055 skb_reserve(skb
, MAX_TCP_HEADER
);
3056 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
3057 TCPHDR_ACK
| TCPHDR_RST
);
3058 tcp_mstamp_refresh(tcp_sk(sk
));
3060 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
3061 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3064 /* Send a crossed SYN-ACK during socket establishment.
3065 * WARNING: This routine must only be called when we have already sent
3066 * a SYN packet that crossed the incoming SYN that caused this routine
3067 * to get called. If this assumption fails then the initial rcv_wnd
3068 * and rcv_wscale values will not be correct.
3070 int tcp_send_synack(struct sock
*sk
)
3072 struct sk_buff
*skb
;
3074 skb
= tcp_write_queue_head(sk
);
3075 if (!skb
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
3076 pr_debug("%s: wrong queue state\n", __func__
);
3079 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
3080 if (skb_cloned(skb
)) {
3081 struct sk_buff
*nskb
= skb_copy(skb
, GFP_ATOMIC
);
3084 tcp_unlink_write_queue(skb
, sk
);
3085 __skb_header_release(nskb
);
3086 __tcp_add_write_queue_head(sk
, nskb
);
3087 sk_wmem_free_skb(sk
, skb
);
3088 sk
->sk_wmem_queued
+= nskb
->truesize
;
3089 sk_mem_charge(sk
, nskb
->truesize
);
3093 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
3094 tcp_ecn_send_synack(sk
, skb
);
3096 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3100 * tcp_make_synack - Prepare a SYN-ACK.
3101 * sk: listener socket
3102 * dst: dst entry attached to the SYNACK
3103 * req: request_sock pointer
3105 * Allocate one skb and build a SYNACK packet.
3106 * @dst is consumed : Caller should not use it again.
3108 struct sk_buff
*tcp_make_synack(const struct sock
*sk
, struct dst_entry
*dst
,
3109 struct request_sock
*req
,
3110 struct tcp_fastopen_cookie
*foc
,
3111 enum tcp_synack_type synack_type
)
3113 struct inet_request_sock
*ireq
= inet_rsk(req
);
3114 const struct tcp_sock
*tp
= tcp_sk(sk
);
3115 struct tcp_md5sig_key
*md5
= NULL
;
3116 struct tcp_out_options opts
;
3117 struct sk_buff
*skb
;
3118 int tcp_header_size
;
3122 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
3123 if (unlikely(!skb
)) {
3127 /* Reserve space for headers. */
3128 skb_reserve(skb
, MAX_TCP_HEADER
);
3130 switch (synack_type
) {
3131 case TCP_SYNACK_NORMAL
:
3132 skb_set_owner_w(skb
, req_to_sk(req
));
3134 case TCP_SYNACK_COOKIE
:
3135 /* Under synflood, we do not attach skb to a socket,
3136 * to avoid false sharing.
3139 case TCP_SYNACK_FASTOPEN
:
3140 /* sk is a const pointer, because we want to express multiple
3141 * cpu might call us concurrently.
3142 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3144 skb_set_owner_w(skb
, (struct sock
*)sk
);
3147 skb_dst_set(skb
, dst
);
3149 mss
= tcp_mss_clamp(tp
, dst_metric_advmss(dst
));
3151 memset(&opts
, 0, sizeof(opts
));
3152 #ifdef CONFIG_SYN_COOKIES
3153 if (unlikely(req
->cookie_ts
))
3154 skb
->skb_mstamp
= cookie_init_timestamp(req
);
3157 skb
->skb_mstamp
= tcp_clock_us();
3159 #ifdef CONFIG_TCP_MD5SIG
3161 md5
= tcp_rsk(req
)->af_specific
->req_md5_lookup(sk
, req_to_sk(req
));
3163 skb_set_hash(skb
, tcp_rsk(req
)->txhash
, PKT_HASH_TYPE_L4
);
3164 tcp_header_size
= tcp_synack_options(req
, mss
, skb
, &opts
, md5
, foc
) +
3167 skb_push(skb
, tcp_header_size
);
3168 skb_reset_transport_header(skb
);
3170 th
= (struct tcphdr
*)skb
->data
;
3171 memset(th
, 0, sizeof(struct tcphdr
));
3174 tcp_ecn_make_synack(req
, th
);
3175 th
->source
= htons(ireq
->ir_num
);
3176 th
->dest
= ireq
->ir_rmt_port
;
3177 skb
->mark
= ireq
->ir_mark
;
3178 /* Setting of flags are superfluous here for callers (and ECE is
3179 * not even correctly set)
3181 tcp_init_nondata_skb(skb
, tcp_rsk(req
)->snt_isn
,
3182 TCPHDR_SYN
| TCPHDR_ACK
);
3184 th
->seq
= htonl(TCP_SKB_CB(skb
)->seq
);
3185 /* XXX data is queued and acked as is. No buffer/window check */
3186 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
3188 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3189 th
->window
= htons(min(req
->rsk_rcv_wnd
, 65535U));
3190 tcp_options_write((__be32
*)(th
+ 1), NULL
, &opts
);
3191 th
->doff
= (tcp_header_size
>> 2);
3192 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
3194 #ifdef CONFIG_TCP_MD5SIG
3195 /* Okay, we have all we need - do the md5 hash if needed */
3197 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
3198 md5
, req_to_sk(req
), skb
);
3202 /* Do not fool tcpdump (if any), clean our debris */
3206 EXPORT_SYMBOL(tcp_make_synack
);
3208 static void tcp_ca_dst_init(struct sock
*sk
, const struct dst_entry
*dst
)
3210 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3211 const struct tcp_congestion_ops
*ca
;
3212 u32 ca_key
= dst_metric(dst
, RTAX_CC_ALGO
);
3214 if (ca_key
== TCP_CA_UNSPEC
)
3218 ca
= tcp_ca_find_key(ca_key
);
3219 if (likely(ca
&& try_module_get(ca
->owner
))) {
3220 module_put(icsk
->icsk_ca_ops
->owner
);
3221 icsk
->icsk_ca_dst_locked
= tcp_ca_dst_locked(dst
);
3222 icsk
->icsk_ca_ops
= ca
;
3227 /* Do all connect socket setups that can be done AF independent. */
3228 static void tcp_connect_init(struct sock
*sk
)
3230 const struct dst_entry
*dst
= __sk_dst_get(sk
);
3231 struct tcp_sock
*tp
= tcp_sk(sk
);
3235 /* We'll fix this up when we get a response from the other end.
3236 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3238 tp
->tcp_header_len
= sizeof(struct tcphdr
);
3239 if (sock_net(sk
)->ipv4
.sysctl_tcp_timestamps
)
3240 tp
->tcp_header_len
+= TCPOLEN_TSTAMP_ALIGNED
;
3242 #ifdef CONFIG_TCP_MD5SIG
3243 if (tp
->af_specific
->md5_lookup(sk
, sk
))
3244 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
3247 /* If user gave his TCP_MAXSEG, record it to clamp */
3248 if (tp
->rx_opt
.user_mss
)
3249 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3252 tcp_sync_mss(sk
, dst_mtu(dst
));
3254 tcp_ca_dst_init(sk
, dst
);
3256 if (!tp
->window_clamp
)
3257 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
3258 tp
->advmss
= tcp_mss_clamp(tp
, dst_metric_advmss(dst
));
3260 tcp_initialize_rcv_mss(sk
);
3262 /* limit the window selection if the user enforce a smaller rx buffer */
3263 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
3264 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
3265 tp
->window_clamp
= tcp_full_space(sk
);
3267 rcv_wnd
= tcp_rwnd_init_bpf(sk
);
3269 rcv_wnd
= dst_metric(dst
, RTAX_INITRWND
);
3271 tcp_select_initial_window(tcp_full_space(sk
),
3272 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
3275 sock_net(sk
)->ipv4
.sysctl_tcp_window_scaling
,
3279 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
3280 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
3283 sock_reset_flag(sk
, SOCK_DONE
);
3286 tp
->snd_una
= tp
->write_seq
;
3287 tp
->snd_sml
= tp
->write_seq
;
3288 tp
->snd_up
= tp
->write_seq
;
3289 tp
->snd_nxt
= tp
->write_seq
;
3291 if (likely(!tp
->repair
))
3294 tp
->rcv_tstamp
= tcp_jiffies32
;
3295 tp
->rcv_wup
= tp
->rcv_nxt
;
3296 tp
->copied_seq
= tp
->rcv_nxt
;
3298 inet_csk(sk
)->icsk_rto
= tcp_timeout_init(sk
);
3299 inet_csk(sk
)->icsk_retransmits
= 0;
3300 tcp_clear_retrans(tp
);
3303 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
3305 struct tcp_sock
*tp
= tcp_sk(sk
);
3306 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
3308 tcb
->end_seq
+= skb
->len
;
3309 __skb_header_release(skb
);
3310 __tcp_add_write_queue_tail(sk
, skb
);
3311 sk
->sk_wmem_queued
+= skb
->truesize
;
3312 sk_mem_charge(sk
, skb
->truesize
);
3313 tp
->write_seq
= tcb
->end_seq
;
3314 tp
->packets_out
+= tcp_skb_pcount(skb
);
3317 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3318 * queue a data-only packet after the regular SYN, such that regular SYNs
3319 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3320 * only the SYN sequence, the data are retransmitted in the first ACK.
3321 * If cookie is not cached or other error occurs, falls back to send a
3322 * regular SYN with Fast Open cookie request option.
3324 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
3326 struct tcp_sock
*tp
= tcp_sk(sk
);
3327 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
3329 struct sk_buff
*syn_data
;
3331 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
3332 if (!tcp_fastopen_cookie_check(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
))
3335 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3336 * user-MSS. Reserve maximum option space for middleboxes that add
3337 * private TCP options. The cost is reduced data space in SYN :(
3339 tp
->rx_opt
.mss_clamp
= tcp_mss_clamp(tp
, tp
->rx_opt
.mss_clamp
);
3341 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
3342 MAX_TCP_OPTION_SPACE
;
3344 space
= min_t(size_t, space
, fo
->size
);
3346 /* limit to order-0 allocations */
3347 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
3349 syn_data
= sk_stream_alloc_skb(sk
, space
, sk
->sk_allocation
, false);
3352 syn_data
->ip_summed
= CHECKSUM_PARTIAL
;
3353 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
3355 int copied
= copy_from_iter(skb_put(syn_data
, space
), space
,
3356 &fo
->data
->msg_iter
);
3357 if (unlikely(!copied
)) {
3358 kfree_skb(syn_data
);
3361 if (copied
!= space
) {
3362 skb_trim(syn_data
, copied
);
3366 /* No more data pending in inet_wait_for_connect() */
3367 if (space
== fo
->size
)
3371 tcp_connect_queue_skb(sk
, syn_data
);
3373 tcp_chrono_start(sk
, TCP_CHRONO_BUSY
);
3375 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
3377 syn
->skb_mstamp
= syn_data
->skb_mstamp
;
3379 /* Now full SYN+DATA was cloned and sent (or not),
3380 * remove the SYN from the original skb (syn_data)
3381 * we keep in write queue in case of a retransmit, as we
3382 * also have the SYN packet (with no data) in the same queue.
3384 TCP_SKB_CB(syn_data
)->seq
++;
3385 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
3387 tp
->syn_data
= (fo
->copied
> 0);
3388 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3392 /* data was not sent, this is our new send_head */
3393 sk
->sk_send_head
= syn_data
;
3394 tp
->packets_out
-= tcp_skb_pcount(syn_data
);
3397 /* Send a regular SYN with Fast Open cookie request option */
3398 if (fo
->cookie
.len
> 0)
3400 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3402 tp
->syn_fastopen
= 0;
3404 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3408 /* Build a SYN and send it off. */
3409 int tcp_connect(struct sock
*sk
)
3411 struct tcp_sock
*tp
= tcp_sk(sk
);
3412 struct sk_buff
*buff
;
3415 tcp_call_bpf(sk
, BPF_SOCK_OPS_TCP_CONNECT_CB
);
3417 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
3418 return -EHOSTUNREACH
; /* Routing failure or similar. */
3420 tcp_connect_init(sk
);
3422 if (unlikely(tp
->repair
)) {
3423 tcp_finish_connect(sk
, NULL
);
3427 buff
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
, true);
3428 if (unlikely(!buff
))
3431 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3432 tcp_mstamp_refresh(tp
);
3433 tp
->retrans_stamp
= tcp_time_stamp(tp
);
3434 tcp_connect_queue_skb(sk
, buff
);
3435 tcp_ecn_send_syn(sk
, buff
);
3437 /* Send off SYN; include data in Fast Open. */
3438 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3439 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3440 if (err
== -ECONNREFUSED
)
3443 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3444 * in order to make this packet get counted in tcpOutSegs.
3446 tp
->snd_nxt
= tp
->write_seq
;
3447 tp
->pushed_seq
= tp
->write_seq
;
3448 buff
= tcp_send_head(sk
);
3449 if (unlikely(buff
)) {
3450 tp
->snd_nxt
= TCP_SKB_CB(buff
)->seq
;
3451 tp
->pushed_seq
= TCP_SKB_CB(buff
)->seq
;
3453 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3455 /* Timer for repeating the SYN until an answer. */
3456 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3457 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3460 EXPORT_SYMBOL(tcp_connect
);
3462 /* Send out a delayed ack, the caller does the policy checking
3463 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3466 void tcp_send_delayed_ack(struct sock
*sk
)
3468 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3469 int ato
= icsk
->icsk_ack
.ato
;
3470 unsigned long timeout
;
3472 tcp_ca_event(sk
, CA_EVENT_DELAYED_ACK
);
3474 if (ato
> TCP_DELACK_MIN
) {
3475 const struct tcp_sock
*tp
= tcp_sk(sk
);
3476 int max_ato
= HZ
/ 2;
3478 if (icsk
->icsk_ack
.pingpong
||
3479 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3480 max_ato
= TCP_DELACK_MAX
;
3482 /* Slow path, intersegment interval is "high". */
3484 /* If some rtt estimate is known, use it to bound delayed ack.
3485 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3489 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3496 ato
= min(ato
, max_ato
);
3499 /* Stay within the limit we were given */
3500 timeout
= jiffies
+ ato
;
3502 /* Use new timeout only if there wasn't a older one earlier. */
3503 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3504 /* If delack timer was blocked or is about to expire,
3507 if (icsk
->icsk_ack
.blocked
||
3508 time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3513 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3514 timeout
= icsk
->icsk_ack
.timeout
;
3516 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3517 icsk
->icsk_ack
.timeout
= timeout
;
3518 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3521 /* This routine sends an ack and also updates the window. */
3522 void tcp_send_ack(struct sock
*sk
)
3524 struct sk_buff
*buff
;
3526 /* If we have been reset, we may not send again. */
3527 if (sk
->sk_state
== TCP_CLOSE
)
3530 tcp_ca_event(sk
, CA_EVENT_NON_DELAYED_ACK
);
3532 /* We are not putting this on the write queue, so
3533 * tcp_transmit_skb() will set the ownership to this
3536 buff
= alloc_skb(MAX_TCP_HEADER
,
3537 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3538 if (unlikely(!buff
)) {
3539 inet_csk_schedule_ack(sk
);
3540 inet_csk(sk
)->icsk_ack
.ato
= TCP_ATO_MIN
;
3541 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
,
3542 TCP_DELACK_MAX
, TCP_RTO_MAX
);
3546 /* Reserve space for headers and prepare control bits. */
3547 skb_reserve(buff
, MAX_TCP_HEADER
);
3548 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3550 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3552 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3554 skb_set_tcp_pure_ack(buff
);
3556 /* Send it off, this clears delayed acks for us. */
3557 tcp_transmit_skb(sk
, buff
, 0, (__force gfp_t
)0);
3559 EXPORT_SYMBOL_GPL(tcp_send_ack
);
3561 /* This routine sends a packet with an out of date sequence
3562 * number. It assumes the other end will try to ack it.
3564 * Question: what should we make while urgent mode?
3565 * 4.4BSD forces sending single byte of data. We cannot send
3566 * out of window data, because we have SND.NXT==SND.MAX...
3568 * Current solution: to send TWO zero-length segments in urgent mode:
3569 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3570 * out-of-date with SND.UNA-1 to probe window.
3572 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
, int mib
)
3574 struct tcp_sock
*tp
= tcp_sk(sk
);
3575 struct sk_buff
*skb
;
3577 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3578 skb
= alloc_skb(MAX_TCP_HEADER
,
3579 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3583 /* Reserve space for headers and set control bits. */
3584 skb_reserve(skb
, MAX_TCP_HEADER
);
3585 /* Use a previous sequence. This should cause the other
3586 * end to send an ack. Don't queue or clone SKB, just
3589 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
3590 NET_INC_STATS(sock_net(sk
), mib
);
3591 return tcp_transmit_skb(sk
, skb
, 0, (__force gfp_t
)0);
3594 /* Called from setsockopt( ... TCP_REPAIR ) */
3595 void tcp_send_window_probe(struct sock
*sk
)
3597 if (sk
->sk_state
== TCP_ESTABLISHED
) {
3598 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
3599 tcp_mstamp_refresh(tcp_sk(sk
));
3600 tcp_xmit_probe_skb(sk
, 0, LINUX_MIB_TCPWINPROBE
);
3604 /* Initiate keepalive or window probe from timer. */
3605 int tcp_write_wakeup(struct sock
*sk
, int mib
)
3607 struct tcp_sock
*tp
= tcp_sk(sk
);
3608 struct sk_buff
*skb
;
3610 if (sk
->sk_state
== TCP_CLOSE
)
3613 skb
= tcp_send_head(sk
);
3614 if (skb
&& before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
3616 unsigned int mss
= tcp_current_mss(sk
);
3617 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
3619 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
3620 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
3622 /* We are probing the opening of a window
3623 * but the window size is != 0
3624 * must have been a result SWS avoidance ( sender )
3626 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
3628 seg_size
= min(seg_size
, mss
);
3629 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3630 if (tcp_fragment(sk
, skb
, seg_size
, mss
, GFP_ATOMIC
))
3632 } else if (!tcp_skb_pcount(skb
))
3633 tcp_set_skb_tso_segs(skb
, mss
);
3635 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
3636 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3638 tcp_event_new_data_sent(sk
, skb
);
3641 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
3642 tcp_xmit_probe_skb(sk
, 1, mib
);
3643 return tcp_xmit_probe_skb(sk
, 0, mib
);
3647 /* A window probe timeout has occurred. If window is not closed send
3648 * a partial packet else a zero probe.
3650 void tcp_send_probe0(struct sock
*sk
)
3652 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3653 struct tcp_sock
*tp
= tcp_sk(sk
);
3654 struct net
*net
= sock_net(sk
);
3655 unsigned long probe_max
;
3658 err
= tcp_write_wakeup(sk
, LINUX_MIB_TCPWINPROBE
);
3660 if (tp
->packets_out
|| !tcp_send_head(sk
)) {
3661 /* Cancel probe timer, if it is not required. */
3662 icsk
->icsk_probes_out
= 0;
3663 icsk
->icsk_backoff
= 0;
3668 if (icsk
->icsk_backoff
< net
->ipv4
.sysctl_tcp_retries2
)
3669 icsk
->icsk_backoff
++;
3670 icsk
->icsk_probes_out
++;
3671 probe_max
= TCP_RTO_MAX
;
3673 /* If packet was not sent due to local congestion,
3674 * do not backoff and do not remember icsk_probes_out.
3675 * Let local senders to fight for local resources.
3677 * Use accumulated backoff yet.
3679 if (!icsk
->icsk_probes_out
)
3680 icsk
->icsk_probes_out
= 1;
3681 probe_max
= TCP_RESOURCE_PROBE_INTERVAL
;
3683 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
,
3684 tcp_probe0_when(sk
, probe_max
),
3688 int tcp_rtx_synack(const struct sock
*sk
, struct request_sock
*req
)
3690 const struct tcp_request_sock_ops
*af_ops
= tcp_rsk(req
)->af_specific
;
3694 tcp_rsk(req
)->txhash
= net_tx_rndhash();
3695 res
= af_ops
->send_synack(sk
, NULL
, &fl
, req
, NULL
, TCP_SYNACK_NORMAL
);
3697 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
3698 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
3699 if (unlikely(tcp_passive_fastopen(sk
)))
3700 tcp_sk(sk
)->total_retrans
++;
3704 EXPORT_SYMBOL(tcp_rtx_synack
);