1 // SPDX-License-Identifier: GPL-2.0-only
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
38 #define pr_fmt(fmt) "TCP: " fmt
41 #include <net/mptcp.h>
43 #include <linux/compiler.h>
44 #include <linux/gfp.h>
45 #include <linux/module.h>
46 #include <linux/static_key.h>
48 #include <trace/events/tcp.h>
50 /* Refresh clocks of a TCP socket,
51 * ensuring monotically increasing values.
53 void tcp_mstamp_refresh(struct tcp_sock
*tp
)
55 u64 val
= tcp_clock_ns();
57 tp
->tcp_clock_cache
= val
;
58 tp
->tcp_mstamp
= div_u64(val
, NSEC_PER_USEC
);
61 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
62 int push_one
, gfp_t gfp
);
64 /* Account for new data that has been sent to the network. */
65 static void tcp_event_new_data_sent(struct sock
*sk
, struct sk_buff
*skb
)
67 struct inet_connection_sock
*icsk
= inet_csk(sk
);
68 struct tcp_sock
*tp
= tcp_sk(sk
);
69 unsigned int prior_packets
= tp
->packets_out
;
71 WRITE_ONCE(tp
->snd_nxt
, TCP_SKB_CB(skb
)->end_seq
);
73 __skb_unlink(skb
, &sk
->sk_write_queue
);
74 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, skb
);
76 if (tp
->highest_sack
== NULL
)
77 tp
->highest_sack
= skb
;
79 tp
->packets_out
+= tcp_skb_pcount(skb
);
80 if (!prior_packets
|| icsk
->icsk_pending
== ICSK_TIME_LOSS_PROBE
)
83 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
,
87 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
88 * window scaling factor due to loss of precision.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
94 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
96 const struct tcp_sock
*tp
= tcp_sk(sk
);
98 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
) ||
99 (tp
->rx_opt
.wscale_ok
&&
100 ((tp
->snd_nxt
- tcp_wnd_end(tp
)) < (1 << tp
->rx_opt
.rcv_wscale
))))
103 return tcp_wnd_end(tp
);
106 /* Calculate mss to advertise in SYN segment.
107 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
109 * 1. It is independent of path mtu.
110 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
111 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
112 * attached devices, because some buggy hosts are confused by
114 * 4. We do not make 3, we advertise MSS, calculated from first
115 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
116 * This may be overridden via information stored in routing table.
117 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
118 * probably even Jumbo".
120 static __u16
tcp_advertise_mss(struct sock
*sk
)
122 struct tcp_sock
*tp
= tcp_sk(sk
);
123 const struct dst_entry
*dst
= __sk_dst_get(sk
);
124 int mss
= tp
->advmss
;
127 unsigned int metric
= dst_metric_advmss(dst
);
138 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
139 * This is the first part of cwnd validation mechanism.
141 void tcp_cwnd_restart(struct sock
*sk
, s32 delta
)
143 struct tcp_sock
*tp
= tcp_sk(sk
);
144 u32 restart_cwnd
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
145 u32 cwnd
= tp
->snd_cwnd
;
147 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
149 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
150 restart_cwnd
= min(restart_cwnd
, cwnd
);
152 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
154 tp
->snd_cwnd
= max(cwnd
, restart_cwnd
);
155 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
156 tp
->snd_cwnd_used
= 0;
159 /* Congestion state accounting after a packet has been sent. */
160 static void tcp_event_data_sent(struct tcp_sock
*tp
,
163 struct inet_connection_sock
*icsk
= inet_csk(sk
);
164 const u32 now
= tcp_jiffies32
;
166 if (tcp_packets_in_flight(tp
) == 0)
167 tcp_ca_event(sk
, CA_EVENT_TX_START
);
169 /* If this is the first data packet sent in response to the
170 * previous received data,
171 * and it is a reply for ato after last received packet,
172 * increase pingpong count.
174 if (before(tp
->lsndtime
, icsk
->icsk_ack
.lrcvtime
) &&
175 (u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
176 inet_csk_inc_pingpong_cnt(sk
);
181 /* Account for an ACK we sent. */
182 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
,
185 struct tcp_sock
*tp
= tcp_sk(sk
);
187 if (unlikely(tp
->compressed_ack
)) {
188 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPACKCOMPRESSED
,
190 tp
->compressed_ack
= 0;
191 if (hrtimer_try_to_cancel(&tp
->compressed_ack_timer
) == 1)
195 if (unlikely(rcv_nxt
!= tp
->rcv_nxt
))
196 return; /* Special ACK sent by DCTCP to reflect ECN */
197 tcp_dec_quickack_mode(sk
, pkts
);
198 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
201 /* Determine a window scaling and initial window to offer.
202 * Based on the assumption that the given amount of space
203 * will be offered. Store the results in the tp structure.
204 * NOTE: for smooth operation initial space offering should
205 * be a multiple of mss if possible. We assume here that mss >= 1.
206 * This MUST be enforced by all callers.
208 void tcp_select_initial_window(const struct sock
*sk
, int __space
, __u32 mss
,
209 __u32
*rcv_wnd
, __u32
*window_clamp
,
210 int wscale_ok
, __u8
*rcv_wscale
,
213 unsigned int space
= (__space
< 0 ? 0 : __space
);
215 /* If no clamp set the clamp to the max possible scaled window */
216 if (*window_clamp
== 0)
217 (*window_clamp
) = (U16_MAX
<< TCP_MAX_WSCALE
);
218 space
= min(*window_clamp
, space
);
220 /* Quantize space offering to a multiple of mss if possible. */
222 space
= rounddown(space
, mss
);
224 /* NOTE: offering an initial window larger than 32767
225 * will break some buggy TCP stacks. If the admin tells us
226 * it is likely we could be speaking with such a buggy stack
227 * we will truncate our initial window offering to 32K-1
228 * unless the remote has sent us a window scaling option,
229 * which we interpret as a sign the remote TCP is not
230 * misinterpreting the window field as a signed quantity.
232 if (sock_net(sk
)->ipv4
.sysctl_tcp_workaround_signed_windows
)
233 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
235 (*rcv_wnd
) = min_t(u32
, space
, U16_MAX
);
238 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
242 /* Set window scaling on max possible window */
243 space
= max_t(u32
, space
, sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[2]);
244 space
= max_t(u32
, space
, sysctl_rmem_max
);
245 space
= min_t(u32
, space
, *window_clamp
);
246 *rcv_wscale
= clamp_t(int, ilog2(space
) - 15,
249 /* Set the clamp no higher than max representable value */
250 (*window_clamp
) = min_t(__u32
, U16_MAX
<< (*rcv_wscale
), *window_clamp
);
252 EXPORT_SYMBOL(tcp_select_initial_window
);
254 /* Chose a new window to advertise, update state in tcp_sock for the
255 * socket, and return result with RFC1323 scaling applied. The return
256 * value can be stuffed directly into th->window for an outgoing
259 static u16
tcp_select_window(struct sock
*sk
)
261 struct tcp_sock
*tp
= tcp_sk(sk
);
262 u32 old_win
= tp
->rcv_wnd
;
263 u32 cur_win
= tcp_receive_window(tp
);
264 u32 new_win
= __tcp_select_window(sk
);
266 /* Never shrink the offered window */
267 if (new_win
< cur_win
) {
268 /* Danger Will Robinson!
269 * Don't update rcv_wup/rcv_wnd here or else
270 * we will not be able to advertise a zero
271 * window in time. --DaveM
273 * Relax Will Robinson.
276 NET_INC_STATS(sock_net(sk
),
277 LINUX_MIB_TCPWANTZEROWINDOWADV
);
278 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
280 tp
->rcv_wnd
= new_win
;
281 tp
->rcv_wup
= tp
->rcv_nxt
;
283 /* Make sure we do not exceed the maximum possible
286 if (!tp
->rx_opt
.rcv_wscale
&&
287 sock_net(sk
)->ipv4
.sysctl_tcp_workaround_signed_windows
)
288 new_win
= min(new_win
, MAX_TCP_WINDOW
);
290 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
292 /* RFC1323 scaling applied */
293 new_win
>>= tp
->rx_opt
.rcv_wscale
;
295 /* If we advertise zero window, disable fast path. */
299 NET_INC_STATS(sock_net(sk
),
300 LINUX_MIB_TCPTOZEROWINDOWADV
);
301 } else if (old_win
== 0) {
302 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFROMZEROWINDOWADV
);
308 /* Packet ECN state for a SYN-ACK */
309 static void tcp_ecn_send_synack(struct sock
*sk
, struct sk_buff
*skb
)
311 const struct tcp_sock
*tp
= tcp_sk(sk
);
313 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
314 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
315 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
316 else if (tcp_ca_needs_ecn(sk
) ||
317 tcp_bpf_ca_needs_ecn(sk
))
321 /* Packet ECN state for a SYN. */
322 static void tcp_ecn_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
324 struct tcp_sock
*tp
= tcp_sk(sk
);
325 bool bpf_needs_ecn
= tcp_bpf_ca_needs_ecn(sk
);
326 bool use_ecn
= sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1 ||
327 tcp_ca_needs_ecn(sk
) || bpf_needs_ecn
;
330 const struct dst_entry
*dst
= __sk_dst_get(sk
);
332 if (dst
&& dst_feature(dst
, RTAX_FEATURE_ECN
))
339 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
340 tp
->ecn_flags
= TCP_ECN_OK
;
341 if (tcp_ca_needs_ecn(sk
) || bpf_needs_ecn
)
346 static void tcp_ecn_clear_syn(struct sock
*sk
, struct sk_buff
*skb
)
348 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn_fallback
)
349 /* tp->ecn_flags are cleared at a later point in time when
350 * SYN ACK is ultimatively being received.
352 TCP_SKB_CB(skb
)->tcp_flags
&= ~(TCPHDR_ECE
| TCPHDR_CWR
);
356 tcp_ecn_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
358 if (inet_rsk(req
)->ecn_ok
)
362 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
365 static void tcp_ecn_send(struct sock
*sk
, struct sk_buff
*skb
,
366 struct tcphdr
*th
, int tcp_header_len
)
368 struct tcp_sock
*tp
= tcp_sk(sk
);
370 if (tp
->ecn_flags
& TCP_ECN_OK
) {
371 /* Not-retransmitted data segment: set ECT and inject CWR. */
372 if (skb
->len
!= tcp_header_len
&&
373 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
375 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
376 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
378 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
380 } else if (!tcp_ca_needs_ecn(sk
)) {
381 /* ACK or retransmitted segment: clear ECT|CE */
382 INET_ECN_dontxmit(sk
);
384 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
389 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
390 * auto increment end seqno.
392 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
394 skb
->ip_summed
= CHECKSUM_PARTIAL
;
396 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
397 TCP_SKB_CB(skb
)->sacked
= 0;
399 tcp_skb_pcount_set(skb
, 1);
401 TCP_SKB_CB(skb
)->seq
= seq
;
402 if (flags
& (TCPHDR_SYN
| TCPHDR_FIN
))
404 TCP_SKB_CB(skb
)->end_seq
= seq
;
407 static inline bool tcp_urg_mode(const struct tcp_sock
*tp
)
409 return tp
->snd_una
!= tp
->snd_up
;
412 #define OPTION_SACK_ADVERTISE (1 << 0)
413 #define OPTION_TS (1 << 1)
414 #define OPTION_MD5 (1 << 2)
415 #define OPTION_WSCALE (1 << 3)
416 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
417 #define OPTION_SMC (1 << 9)
418 #define OPTION_MPTCP (1 << 10)
420 static void smc_options_write(__be32
*ptr
, u16
*options
)
422 #if IS_ENABLED(CONFIG_SMC)
423 if (static_branch_unlikely(&tcp_have_smc
)) {
424 if (unlikely(OPTION_SMC
& *options
)) {
425 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
428 (TCPOLEN_EXP_SMC_BASE
));
429 *ptr
++ = htonl(TCPOPT_SMC_MAGIC
);
435 struct tcp_out_options
{
436 u16 options
; /* bit field of OPTION_* */
437 u16 mss
; /* 0 to disable */
438 u8 ws
; /* window scale, 0 to disable */
439 u8 num_sack_blocks
; /* number of SACK blocks to include */
440 u8 hash_size
; /* bytes in hash_location */
441 u8 bpf_opt_len
; /* length of BPF hdr option */
442 __u8
*hash_location
; /* temporary pointer, overloaded */
443 __u32 tsval
, tsecr
; /* need to include OPTION_TS */
444 struct tcp_fastopen_cookie
*fastopen_cookie
; /* Fast open cookie */
445 struct mptcp_out_options mptcp
;
448 static void mptcp_options_write(__be32
*ptr
, const struct tcp_sock
*tp
,
449 struct tcp_out_options
*opts
)
451 #if IS_ENABLED(CONFIG_MPTCP)
452 if (unlikely(OPTION_MPTCP
& opts
->options
))
453 mptcp_write_options(ptr
, tp
, &opts
->mptcp
);
457 #ifdef CONFIG_CGROUP_BPF
458 static int bpf_skops_write_hdr_opt_arg0(struct sk_buff
*skb
,
459 enum tcp_synack_type synack_type
)
462 return BPF_WRITE_HDR_TCP_CURRENT_MSS
;
464 if (unlikely(synack_type
== TCP_SYNACK_COOKIE
))
465 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE
;
470 /* req, syn_skb and synack_type are used when writing synack */
471 static void bpf_skops_hdr_opt_len(struct sock
*sk
, struct sk_buff
*skb
,
472 struct request_sock
*req
,
473 struct sk_buff
*syn_skb
,
474 enum tcp_synack_type synack_type
,
475 struct tcp_out_options
*opts
,
476 unsigned int *remaining
)
478 struct bpf_sock_ops_kern sock_ops
;
481 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk
),
482 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG
)) ||
486 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
489 memset(&sock_ops
, 0, offsetof(struct bpf_sock_ops_kern
, temp
));
491 sock_ops
.op
= BPF_SOCK_OPS_HDR_OPT_LEN_CB
;
494 /* The listen "sk" cannot be passed here because
495 * it is not locked. It would not make too much
496 * sense to do bpf_setsockopt(listen_sk) based
497 * on individual connection request also.
499 * Thus, "req" is passed here and the cgroup-bpf-progs
500 * of the listen "sk" will be run.
502 * "req" is also used here for fastopen even the "sk" here is
503 * a fullsock "child" sk. It is to keep the behavior
504 * consistent between fastopen and non-fastopen on
505 * the bpf programming side.
507 sock_ops
.sk
= (struct sock
*)req
;
508 sock_ops
.syn_skb
= syn_skb
;
510 sock_owned_by_me(sk
);
512 sock_ops
.is_fullsock
= 1;
516 sock_ops
.args
[0] = bpf_skops_write_hdr_opt_arg0(skb
, synack_type
);
517 sock_ops
.remaining_opt_len
= *remaining
;
518 /* tcp_current_mss() does not pass a skb */
520 bpf_skops_init_skb(&sock_ops
, skb
, 0);
522 err
= BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops
, sk
);
524 if (err
|| sock_ops
.remaining_opt_len
== *remaining
)
527 opts
->bpf_opt_len
= *remaining
- sock_ops
.remaining_opt_len
;
528 /* round up to 4 bytes */
529 opts
->bpf_opt_len
= (opts
->bpf_opt_len
+ 3) & ~3;
531 *remaining
-= opts
->bpf_opt_len
;
534 static void bpf_skops_write_hdr_opt(struct sock
*sk
, struct sk_buff
*skb
,
535 struct request_sock
*req
,
536 struct sk_buff
*syn_skb
,
537 enum tcp_synack_type synack_type
,
538 struct tcp_out_options
*opts
)
540 u8 first_opt_off
, nr_written
, max_opt_len
= opts
->bpf_opt_len
;
541 struct bpf_sock_ops_kern sock_ops
;
544 if (likely(!max_opt_len
))
547 memset(&sock_ops
, 0, offsetof(struct bpf_sock_ops_kern
, temp
));
549 sock_ops
.op
= BPF_SOCK_OPS_WRITE_HDR_OPT_CB
;
552 sock_ops
.sk
= (struct sock
*)req
;
553 sock_ops
.syn_skb
= syn_skb
;
555 sock_owned_by_me(sk
);
557 sock_ops
.is_fullsock
= 1;
561 sock_ops
.args
[0] = bpf_skops_write_hdr_opt_arg0(skb
, synack_type
);
562 sock_ops
.remaining_opt_len
= max_opt_len
;
563 first_opt_off
= tcp_hdrlen(skb
) - max_opt_len
;
564 bpf_skops_init_skb(&sock_ops
, skb
, first_opt_off
);
566 err
= BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops
, sk
);
571 nr_written
= max_opt_len
- sock_ops
.remaining_opt_len
;
573 if (nr_written
< max_opt_len
)
574 memset(skb
->data
+ first_opt_off
+ nr_written
, TCPOPT_NOP
,
575 max_opt_len
- nr_written
);
578 static void bpf_skops_hdr_opt_len(struct sock
*sk
, struct sk_buff
*skb
,
579 struct request_sock
*req
,
580 struct sk_buff
*syn_skb
,
581 enum tcp_synack_type synack_type
,
582 struct tcp_out_options
*opts
,
583 unsigned int *remaining
)
587 static void bpf_skops_write_hdr_opt(struct sock
*sk
, struct sk_buff
*skb
,
588 struct request_sock
*req
,
589 struct sk_buff
*syn_skb
,
590 enum tcp_synack_type synack_type
,
591 struct tcp_out_options
*opts
)
596 /* Write previously computed TCP options to the packet.
598 * Beware: Something in the Internet is very sensitive to the ordering of
599 * TCP options, we learned this through the hard way, so be careful here.
600 * Luckily we can at least blame others for their non-compliance but from
601 * inter-operability perspective it seems that we're somewhat stuck with
602 * the ordering which we have been using if we want to keep working with
603 * those broken things (not that it currently hurts anybody as there isn't
604 * particular reason why the ordering would need to be changed).
606 * At least SACK_PERM as the first option is known to lead to a disaster
607 * (but it may well be that other scenarios fail similarly).
609 static void tcp_options_write(__be32
*ptr
, struct tcp_sock
*tp
,
610 struct tcp_out_options
*opts
)
612 u16 options
= opts
->options
; /* mungable copy */
614 if (unlikely(OPTION_MD5
& options
)) {
615 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
616 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
617 /* overload cookie hash location */
618 opts
->hash_location
= (__u8
*)ptr
;
622 if (unlikely(opts
->mss
)) {
623 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
624 (TCPOLEN_MSS
<< 16) |
628 if (likely(OPTION_TS
& options
)) {
629 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
630 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
631 (TCPOLEN_SACK_PERM
<< 16) |
632 (TCPOPT_TIMESTAMP
<< 8) |
634 options
&= ~OPTION_SACK_ADVERTISE
;
636 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
638 (TCPOPT_TIMESTAMP
<< 8) |
641 *ptr
++ = htonl(opts
->tsval
);
642 *ptr
++ = htonl(opts
->tsecr
);
645 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
646 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
648 (TCPOPT_SACK_PERM
<< 8) |
652 if (unlikely(OPTION_WSCALE
& options
)) {
653 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
654 (TCPOPT_WINDOW
<< 16) |
655 (TCPOLEN_WINDOW
<< 8) |
659 if (unlikely(opts
->num_sack_blocks
)) {
660 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
661 tp
->duplicate_sack
: tp
->selective_acks
;
664 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
667 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
668 TCPOLEN_SACK_PERBLOCK
)));
670 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
672 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
673 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
676 tp
->rx_opt
.dsack
= 0;
679 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
680 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
682 u32 len
; /* Fast Open option length */
685 len
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
686 *ptr
= htonl((TCPOPT_EXP
<< 24) | (len
<< 16) |
687 TCPOPT_FASTOPEN_MAGIC
);
688 p
+= TCPOLEN_EXP_FASTOPEN_BASE
;
690 len
= TCPOLEN_FASTOPEN_BASE
+ foc
->len
;
691 *p
++ = TCPOPT_FASTOPEN
;
695 memcpy(p
, foc
->val
, foc
->len
);
696 if ((len
& 3) == 2) {
697 p
[foc
->len
] = TCPOPT_NOP
;
698 p
[foc
->len
+ 1] = TCPOPT_NOP
;
700 ptr
+= (len
+ 3) >> 2;
703 smc_options_write(ptr
, &options
);
705 mptcp_options_write(ptr
, tp
, opts
);
708 static void smc_set_option(const struct tcp_sock
*tp
,
709 struct tcp_out_options
*opts
,
710 unsigned int *remaining
)
712 #if IS_ENABLED(CONFIG_SMC)
713 if (static_branch_unlikely(&tcp_have_smc
)) {
715 if (*remaining
>= TCPOLEN_EXP_SMC_BASE_ALIGNED
) {
716 opts
->options
|= OPTION_SMC
;
717 *remaining
-= TCPOLEN_EXP_SMC_BASE_ALIGNED
;
724 static void smc_set_option_cond(const struct tcp_sock
*tp
,
725 const struct inet_request_sock
*ireq
,
726 struct tcp_out_options
*opts
,
727 unsigned int *remaining
)
729 #if IS_ENABLED(CONFIG_SMC)
730 if (static_branch_unlikely(&tcp_have_smc
)) {
731 if (tp
->syn_smc
&& ireq
->smc_ok
) {
732 if (*remaining
>= TCPOLEN_EXP_SMC_BASE_ALIGNED
) {
733 opts
->options
|= OPTION_SMC
;
734 *remaining
-= TCPOLEN_EXP_SMC_BASE_ALIGNED
;
741 static void mptcp_set_option_cond(const struct request_sock
*req
,
742 struct tcp_out_options
*opts
,
743 unsigned int *remaining
)
745 if (rsk_is_mptcp(req
)) {
748 if (mptcp_synack_options(req
, &size
, &opts
->mptcp
)) {
749 if (*remaining
>= size
) {
750 opts
->options
|= OPTION_MPTCP
;
757 /* Compute TCP options for SYN packets. This is not the final
758 * network wire format yet.
760 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
761 struct tcp_out_options
*opts
,
762 struct tcp_md5sig_key
**md5
)
764 struct tcp_sock
*tp
= tcp_sk(sk
);
765 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
766 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
769 #ifdef CONFIG_TCP_MD5SIG
770 if (static_branch_unlikely(&tcp_md5_needed
) &&
771 rcu_access_pointer(tp
->md5sig_info
)) {
772 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
774 opts
->options
|= OPTION_MD5
;
775 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
780 /* We always get an MSS option. The option bytes which will be seen in
781 * normal data packets should timestamps be used, must be in the MSS
782 * advertised. But we subtract them from tp->mss_cache so that
783 * calculations in tcp_sendmsg are simpler etc. So account for this
784 * fact here if necessary. If we don't do this correctly, as a
785 * receiver we won't recognize data packets as being full sized when we
786 * should, and thus we won't abide by the delayed ACK rules correctly.
787 * SACKs don't matter, we never delay an ACK when we have any of those
789 opts
->mss
= tcp_advertise_mss(sk
);
790 remaining
-= TCPOLEN_MSS_ALIGNED
;
792 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_timestamps
&& !*md5
)) {
793 opts
->options
|= OPTION_TS
;
794 opts
->tsval
= tcp_skb_timestamp(skb
) + tp
->tsoffset
;
795 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
796 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
798 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_window_scaling
)) {
799 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
800 opts
->options
|= OPTION_WSCALE
;
801 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
803 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_sack
)) {
804 opts
->options
|= OPTION_SACK_ADVERTISE
;
805 if (unlikely(!(OPTION_TS
& opts
->options
)))
806 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
809 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
810 u32 need
= fastopen
->cookie
.len
;
812 need
+= fastopen
->cookie
.exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
813 TCPOLEN_FASTOPEN_BASE
;
814 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
815 if (remaining
>= need
) {
816 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
817 opts
->fastopen_cookie
= &fastopen
->cookie
;
819 tp
->syn_fastopen
= 1;
820 tp
->syn_fastopen_exp
= fastopen
->cookie
.exp
? 1 : 0;
824 smc_set_option(tp
, opts
, &remaining
);
826 if (sk_is_mptcp(sk
)) {
829 if (mptcp_syn_options(sk
, skb
, &size
, &opts
->mptcp
)) {
830 opts
->options
|= OPTION_MPTCP
;
835 bpf_skops_hdr_opt_len(sk
, skb
, NULL
, NULL
, 0, opts
, &remaining
);
837 return MAX_TCP_OPTION_SPACE
- remaining
;
840 /* Set up TCP options for SYN-ACKs. */
841 static unsigned int tcp_synack_options(const struct sock
*sk
,
842 struct request_sock
*req
,
843 unsigned int mss
, struct sk_buff
*skb
,
844 struct tcp_out_options
*opts
,
845 const struct tcp_md5sig_key
*md5
,
846 struct tcp_fastopen_cookie
*foc
,
847 enum tcp_synack_type synack_type
,
848 struct sk_buff
*syn_skb
)
850 struct inet_request_sock
*ireq
= inet_rsk(req
);
851 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
853 #ifdef CONFIG_TCP_MD5SIG
855 opts
->options
|= OPTION_MD5
;
856 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
858 /* We can't fit any SACK blocks in a packet with MD5 + TS
859 * options. There was discussion about disabling SACK
860 * rather than TS in order to fit in better with old,
861 * buggy kernels, but that was deemed to be unnecessary.
863 if (synack_type
!= TCP_SYNACK_COOKIE
)
864 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
868 /* We always send an MSS option. */
870 remaining
-= TCPOLEN_MSS_ALIGNED
;
872 if (likely(ireq
->wscale_ok
)) {
873 opts
->ws
= ireq
->rcv_wscale
;
874 opts
->options
|= OPTION_WSCALE
;
875 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
877 if (likely(ireq
->tstamp_ok
)) {
878 opts
->options
|= OPTION_TS
;
879 opts
->tsval
= tcp_skb_timestamp(skb
) + tcp_rsk(req
)->ts_off
;
880 opts
->tsecr
= req
->ts_recent
;
881 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
883 if (likely(ireq
->sack_ok
)) {
884 opts
->options
|= OPTION_SACK_ADVERTISE
;
885 if (unlikely(!ireq
->tstamp_ok
))
886 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
888 if (foc
!= NULL
&& foc
->len
>= 0) {
891 need
+= foc
->exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
892 TCPOLEN_FASTOPEN_BASE
;
893 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
894 if (remaining
>= need
) {
895 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
896 opts
->fastopen_cookie
= foc
;
901 mptcp_set_option_cond(req
, opts
, &remaining
);
903 smc_set_option_cond(tcp_sk(sk
), ireq
, opts
, &remaining
);
905 bpf_skops_hdr_opt_len((struct sock
*)sk
, skb
, req
, syn_skb
,
906 synack_type
, opts
, &remaining
);
908 return MAX_TCP_OPTION_SPACE
- remaining
;
911 /* Compute TCP options for ESTABLISHED sockets. This is not the
912 * final wire format yet.
914 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
915 struct tcp_out_options
*opts
,
916 struct tcp_md5sig_key
**md5
)
918 struct tcp_sock
*tp
= tcp_sk(sk
);
919 unsigned int size
= 0;
920 unsigned int eff_sacks
;
925 #ifdef CONFIG_TCP_MD5SIG
926 if (static_branch_unlikely(&tcp_md5_needed
) &&
927 rcu_access_pointer(tp
->md5sig_info
)) {
928 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
930 opts
->options
|= OPTION_MD5
;
931 size
+= TCPOLEN_MD5SIG_ALIGNED
;
936 if (likely(tp
->rx_opt
.tstamp_ok
)) {
937 opts
->options
|= OPTION_TS
;
938 opts
->tsval
= skb
? tcp_skb_timestamp(skb
) + tp
->tsoffset
: 0;
939 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
940 size
+= TCPOLEN_TSTAMP_ALIGNED
;
943 /* MPTCP options have precedence over SACK for the limited TCP
944 * option space because a MPTCP connection would be forced to
945 * fall back to regular TCP if a required multipath option is
946 * missing. SACK still gets a chance to use whatever space is
949 if (sk_is_mptcp(sk
)) {
950 unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
951 unsigned int opt_size
= 0;
953 if (mptcp_established_options(sk
, skb
, &opt_size
, remaining
,
955 opts
->options
|= OPTION_MPTCP
;
960 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
961 if (unlikely(eff_sacks
)) {
962 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
963 if (unlikely(remaining
< TCPOLEN_SACK_BASE_ALIGNED
+
964 TCPOLEN_SACK_PERBLOCK
))
967 opts
->num_sack_blocks
=
968 min_t(unsigned int, eff_sacks
,
969 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
970 TCPOLEN_SACK_PERBLOCK
);
972 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
973 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
976 if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp
,
977 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG
))) {
978 unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
980 bpf_skops_hdr_opt_len(sk
, skb
, NULL
, NULL
, 0, opts
, &remaining
);
982 size
= MAX_TCP_OPTION_SPACE
- remaining
;
989 /* TCP SMALL QUEUES (TSQ)
991 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
992 * to reduce RTT and bufferbloat.
993 * We do this using a special skb destructor (tcp_wfree).
995 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
996 * needs to be reallocated in a driver.
997 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
999 * Since transmit from skb destructor is forbidden, we use a tasklet
1000 * to process all sockets that eventually need to send more skbs.
1001 * We use one tasklet per cpu, with its own queue of sockets.
1003 struct tsq_tasklet
{
1004 struct tasklet_struct tasklet
;
1005 struct list_head head
; /* queue of tcp sockets */
1007 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
1009 static void tcp_tsq_write(struct sock
*sk
)
1011 if ((1 << sk
->sk_state
) &
1012 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
1013 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
)) {
1014 struct tcp_sock
*tp
= tcp_sk(sk
);
1016 if (tp
->lost_out
> tp
->retrans_out
&&
1017 tp
->snd_cwnd
> tcp_packets_in_flight(tp
)) {
1018 tcp_mstamp_refresh(tp
);
1019 tcp_xmit_retransmit_queue(sk
);
1022 tcp_write_xmit(sk
, tcp_current_mss(sk
), tp
->nonagle
,
1027 static void tcp_tsq_handler(struct sock
*sk
)
1030 if (!sock_owned_by_user(sk
))
1032 else if (!test_and_set_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
))
1037 * One tasklet per cpu tries to send more skbs.
1038 * We run in tasklet context but need to disable irqs when
1039 * transferring tsq->head because tcp_wfree() might
1040 * interrupt us (non NAPI drivers)
1042 static void tcp_tasklet_func(struct tasklet_struct
*t
)
1044 struct tsq_tasklet
*tsq
= from_tasklet(tsq
, t
, tasklet
);
1046 unsigned long flags
;
1047 struct list_head
*q
, *n
;
1048 struct tcp_sock
*tp
;
1051 local_irq_save(flags
);
1052 list_splice_init(&tsq
->head
, &list
);
1053 local_irq_restore(flags
);
1055 list_for_each_safe(q
, n
, &list
) {
1056 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
1057 list_del(&tp
->tsq_node
);
1059 sk
= (struct sock
*)tp
;
1060 smp_mb__before_atomic();
1061 clear_bit(TSQ_QUEUED
, &sk
->sk_tsq_flags
);
1063 tcp_tsq_handler(sk
);
1068 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
1069 TCPF_WRITE_TIMER_DEFERRED | \
1070 TCPF_DELACK_TIMER_DEFERRED | \
1071 TCPF_MTU_REDUCED_DEFERRED)
1073 * tcp_release_cb - tcp release_sock() callback
1076 * called from release_sock() to perform protocol dependent
1077 * actions before socket release.
1079 void tcp_release_cb(struct sock
*sk
)
1081 unsigned long flags
, nflags
;
1083 /* perform an atomic operation only if at least one flag is set */
1085 flags
= sk
->sk_tsq_flags
;
1086 if (!(flags
& TCP_DEFERRED_ALL
))
1088 nflags
= flags
& ~TCP_DEFERRED_ALL
;
1089 } while (cmpxchg(&sk
->sk_tsq_flags
, flags
, nflags
) != flags
);
1091 if (flags
& TCPF_TSQ_DEFERRED
) {
1095 /* Here begins the tricky part :
1096 * We are called from release_sock() with :
1098 * 2) sk_lock.slock spinlock held
1099 * 3) socket owned by us (sk->sk_lock.owned == 1)
1101 * But following code is meant to be called from BH handlers,
1102 * so we should keep BH disabled, but early release socket ownership
1104 sock_release_ownership(sk
);
1106 if (flags
& TCPF_WRITE_TIMER_DEFERRED
) {
1107 tcp_write_timer_handler(sk
);
1110 if (flags
& TCPF_DELACK_TIMER_DEFERRED
) {
1111 tcp_delack_timer_handler(sk
);
1114 if (flags
& TCPF_MTU_REDUCED_DEFERRED
) {
1115 inet_csk(sk
)->icsk_af_ops
->mtu_reduced(sk
);
1119 EXPORT_SYMBOL(tcp_release_cb
);
1121 void __init
tcp_tasklet_init(void)
1125 for_each_possible_cpu(i
) {
1126 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
1128 INIT_LIST_HEAD(&tsq
->head
);
1129 tasklet_setup(&tsq
->tasklet
, tcp_tasklet_func
);
1134 * Write buffer destructor automatically called from kfree_skb.
1135 * We can't xmit new skbs from this context, as we might already
1138 void tcp_wfree(struct sk_buff
*skb
)
1140 struct sock
*sk
= skb
->sk
;
1141 struct tcp_sock
*tp
= tcp_sk(sk
);
1142 unsigned long flags
, nval
, oval
;
1144 /* Keep one reference on sk_wmem_alloc.
1145 * Will be released by sk_free() from here or tcp_tasklet_func()
1147 WARN_ON(refcount_sub_and_test(skb
->truesize
- 1, &sk
->sk_wmem_alloc
));
1149 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
1150 * Wait until our queues (qdisc + devices) are drained.
1152 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1153 * - chance for incoming ACK (processed by another cpu maybe)
1154 * to migrate this flow (skb->ooo_okay will be eventually set)
1156 if (refcount_read(&sk
->sk_wmem_alloc
) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current
)
1159 for (oval
= READ_ONCE(sk
->sk_tsq_flags
);; oval
= nval
) {
1160 struct tsq_tasklet
*tsq
;
1163 if (!(oval
& TSQF_THROTTLED
) || (oval
& TSQF_QUEUED
))
1166 nval
= (oval
& ~TSQF_THROTTLED
) | TSQF_QUEUED
;
1167 nval
= cmpxchg(&sk
->sk_tsq_flags
, oval
, nval
);
1171 /* queue this socket to tasklet queue */
1172 local_irq_save(flags
);
1173 tsq
= this_cpu_ptr(&tsq_tasklet
);
1174 empty
= list_empty(&tsq
->head
);
1175 list_add(&tp
->tsq_node
, &tsq
->head
);
1177 tasklet_schedule(&tsq
->tasklet
);
1178 local_irq_restore(flags
);
1185 /* Note: Called under soft irq.
1186 * We can call TCP stack right away, unless socket is owned by user.
1188 enum hrtimer_restart
tcp_pace_kick(struct hrtimer
*timer
)
1190 struct tcp_sock
*tp
= container_of(timer
, struct tcp_sock
, pacing_timer
);
1191 struct sock
*sk
= (struct sock
*)tp
;
1193 tcp_tsq_handler(sk
);
1196 return HRTIMER_NORESTART
;
1199 static void tcp_update_skb_after_send(struct sock
*sk
, struct sk_buff
*skb
,
1202 struct tcp_sock
*tp
= tcp_sk(sk
);
1204 if (sk
->sk_pacing_status
!= SK_PACING_NONE
) {
1205 unsigned long rate
= sk
->sk_pacing_rate
;
1207 /* Original sch_fq does not pace first 10 MSS
1208 * Note that tp->data_segs_out overflows after 2^32 packets,
1209 * this is a minor annoyance.
1211 if (rate
!= ~0UL && rate
&& tp
->data_segs_out
>= 10) {
1212 u64 len_ns
= div64_ul((u64
)skb
->len
* NSEC_PER_SEC
, rate
);
1213 u64 credit
= tp
->tcp_wstamp_ns
- prior_wstamp
;
1215 /* take into account OS jitter */
1216 len_ns
-= min_t(u64
, len_ns
/ 2, credit
);
1217 tp
->tcp_wstamp_ns
+= len_ns
;
1220 list_move_tail(&skb
->tcp_tsorted_anchor
, &tp
->tsorted_sent_queue
);
1223 INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock
*sk
, struct sk_buff
*skb
, struct flowi
*fl
));
1224 INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock
*sk
, struct sk_buff
*skb
, struct flowi
*fl
));
1225 INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock
*sk
, struct sk_buff
*skb
));
1227 /* This routine actually transmits TCP packets queued in by
1228 * tcp_do_sendmsg(). This is used by both the initial
1229 * transmission and possible later retransmissions.
1230 * All SKB's seen here are completely headerless. It is our
1231 * job to build the TCP header, and pass the packet down to
1232 * IP so it can do the same plus pass the packet off to the
1235 * We are working here with either a clone of the original
1236 * SKB, or a fresh unique copy made by the retransmit engine.
1238 static int __tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
,
1239 int clone_it
, gfp_t gfp_mask
, u32 rcv_nxt
)
1241 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1242 struct inet_sock
*inet
;
1243 struct tcp_sock
*tp
;
1244 struct tcp_skb_cb
*tcb
;
1245 struct tcp_out_options opts
;
1246 unsigned int tcp_options_size
, tcp_header_size
;
1247 struct sk_buff
*oskb
= NULL
;
1248 struct tcp_md5sig_key
*md5
;
1253 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
1255 prior_wstamp
= tp
->tcp_wstamp_ns
;
1256 tp
->tcp_wstamp_ns
= max(tp
->tcp_wstamp_ns
, tp
->tcp_clock_cache
);
1257 skb
->skb_mstamp_ns
= tp
->tcp_wstamp_ns
;
1259 TCP_SKB_CB(skb
)->tx
.in_flight
= TCP_SKB_CB(skb
)->end_seq
1263 tcp_skb_tsorted_save(oskb
) {
1264 if (unlikely(skb_cloned(oskb
)))
1265 skb
= pskb_copy(oskb
, gfp_mask
);
1267 skb
= skb_clone(oskb
, gfp_mask
);
1268 } tcp_skb_tsorted_restore(oskb
);
1272 /* retransmit skbs might have a non zero value in skb->dev
1273 * because skb->dev is aliased with skb->rbnode.rb_left
1279 tcb
= TCP_SKB_CB(skb
);
1280 memset(&opts
, 0, sizeof(opts
));
1282 if (unlikely(tcb
->tcp_flags
& TCPHDR_SYN
)) {
1283 tcp_options_size
= tcp_syn_options(sk
, skb
, &opts
, &md5
);
1285 tcp_options_size
= tcp_established_options(sk
, skb
, &opts
,
1287 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1288 * at receiver : This slightly improve GRO performance.
1289 * Note that we do not force the PSH flag for non GSO packets,
1290 * because they might be sent under high congestion events,
1291 * and in this case it is better to delay the delivery of 1-MSS
1292 * packets and thus the corresponding ACK packet that would
1293 * release the following packet.
1295 if (tcp_skb_pcount(skb
) > 1)
1296 tcb
->tcp_flags
|= TCPHDR_PSH
;
1298 tcp_header_size
= tcp_options_size
+ sizeof(struct tcphdr
);
1300 /* if no packet is in qdisc/device queue, then allow XPS to select
1301 * another queue. We can be called from tcp_tsq_handler()
1302 * which holds one reference to sk.
1304 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1305 * One way to get this would be to set skb->truesize = 2 on them.
1307 skb
->ooo_okay
= sk_wmem_alloc_get(sk
) < SKB_TRUESIZE(1);
1309 /* If we had to use memory reserve to allocate this skb,
1310 * this might cause drops if packet is looped back :
1311 * Other socket might not have SOCK_MEMALLOC.
1312 * Packets not looped back do not care about pfmemalloc.
1314 skb
->pfmemalloc
= 0;
1316 skb_push(skb
, tcp_header_size
);
1317 skb_reset_transport_header(skb
);
1321 skb
->destructor
= skb_is_tcp_pure_ack(skb
) ? __sock_wfree
: tcp_wfree
;
1322 refcount_add(skb
->truesize
, &sk
->sk_wmem_alloc
);
1324 skb_set_dst_pending_confirm(skb
, sk
->sk_dst_pending_confirm
);
1326 /* Build TCP header and checksum it. */
1327 th
= (struct tcphdr
*)skb
->data
;
1328 th
->source
= inet
->inet_sport
;
1329 th
->dest
= inet
->inet_dport
;
1330 th
->seq
= htonl(tcb
->seq
);
1331 th
->ack_seq
= htonl(rcv_nxt
);
1332 *(((__be16
*)th
) + 6) = htons(((tcp_header_size
>> 2) << 12) |
1338 /* The urg_mode check is necessary during a below snd_una win probe */
1339 if (unlikely(tcp_urg_mode(tp
) && before(tcb
->seq
, tp
->snd_up
))) {
1340 if (before(tp
->snd_up
, tcb
->seq
+ 0x10000)) {
1341 th
->urg_ptr
= htons(tp
->snd_up
- tcb
->seq
);
1343 } else if (after(tcb
->seq
+ 0xFFFF, tp
->snd_nxt
)) {
1344 th
->urg_ptr
= htons(0xFFFF);
1349 skb_shinfo(skb
)->gso_type
= sk
->sk_gso_type
;
1350 if (likely(!(tcb
->tcp_flags
& TCPHDR_SYN
))) {
1351 th
->window
= htons(tcp_select_window(sk
));
1352 tcp_ecn_send(sk
, skb
, th
, tcp_header_size
);
1354 /* RFC1323: The window in SYN & SYN/ACK segments
1357 th
->window
= htons(min(tp
->rcv_wnd
, 65535U));
1360 tcp_options_write((__be32
*)(th
+ 1), tp
, &opts
);
1362 #ifdef CONFIG_TCP_MD5SIG
1363 /* Calculate the MD5 hash, as we have all we need now */
1365 sk_nocaps_add(sk
, NETIF_F_GSO_MASK
);
1366 tp
->af_specific
->calc_md5_hash(opts
.hash_location
,
1371 /* BPF prog is the last one writing header option */
1372 bpf_skops_write_hdr_opt(sk
, skb
, NULL
, NULL
, 0, &opts
);
1374 INDIRECT_CALL_INET(icsk
->icsk_af_ops
->send_check
,
1375 tcp_v6_send_check
, tcp_v4_send_check
,
1378 if (likely(tcb
->tcp_flags
& TCPHDR_ACK
))
1379 tcp_event_ack_sent(sk
, tcp_skb_pcount(skb
), rcv_nxt
);
1381 if (skb
->len
!= tcp_header_size
) {
1382 tcp_event_data_sent(tp
, sk
);
1383 tp
->data_segs_out
+= tcp_skb_pcount(skb
);
1384 tp
->bytes_sent
+= skb
->len
- tcp_header_size
;
1387 if (after(tcb
->end_seq
, tp
->snd_nxt
) || tcb
->seq
== tcb
->end_seq
)
1388 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
,
1389 tcp_skb_pcount(skb
));
1391 tp
->segs_out
+= tcp_skb_pcount(skb
);
1392 skb_set_hash_from_sk(skb
, sk
);
1393 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1394 skb_shinfo(skb
)->gso_segs
= tcp_skb_pcount(skb
);
1395 skb_shinfo(skb
)->gso_size
= tcp_skb_mss(skb
);
1397 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1399 /* Cleanup our debris for IP stacks */
1400 memset(skb
->cb
, 0, max(sizeof(struct inet_skb_parm
),
1401 sizeof(struct inet6_skb_parm
)));
1403 tcp_add_tx_delay(skb
, tp
);
1405 err
= INDIRECT_CALL_INET(icsk
->icsk_af_ops
->queue_xmit
,
1406 inet6_csk_xmit
, ip_queue_xmit
,
1407 sk
, skb
, &inet
->cork
.fl
);
1409 if (unlikely(err
> 0)) {
1411 err
= net_xmit_eval(err
);
1414 tcp_update_skb_after_send(sk
, oskb
, prior_wstamp
);
1415 tcp_rate_skb_sent(sk
, oskb
);
1420 static int tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int clone_it
,
1423 return __tcp_transmit_skb(sk
, skb
, clone_it
, gfp_mask
,
1424 tcp_sk(sk
)->rcv_nxt
);
1427 /* This routine just queues the buffer for sending.
1429 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1430 * otherwise socket can stall.
1432 static void tcp_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
1434 struct tcp_sock
*tp
= tcp_sk(sk
);
1436 /* Advance write_seq and place onto the write_queue. */
1437 WRITE_ONCE(tp
->write_seq
, TCP_SKB_CB(skb
)->end_seq
);
1438 __skb_header_release(skb
);
1439 tcp_add_write_queue_tail(sk
, skb
);
1440 sk_wmem_queued_add(sk
, skb
->truesize
);
1441 sk_mem_charge(sk
, skb
->truesize
);
1444 /* Initialize TSO segments for a packet. */
1445 static void tcp_set_skb_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
1447 if (skb
->len
<= mss_now
) {
1448 /* Avoid the costly divide in the normal
1451 tcp_skb_pcount_set(skb
, 1);
1452 TCP_SKB_CB(skb
)->tcp_gso_size
= 0;
1454 tcp_skb_pcount_set(skb
, DIV_ROUND_UP(skb
->len
, mss_now
));
1455 TCP_SKB_CB(skb
)->tcp_gso_size
= mss_now
;
1459 /* Pcount in the middle of the write queue got changed, we need to do various
1460 * tweaks to fix counters
1462 static void tcp_adjust_pcount(struct sock
*sk
, const struct sk_buff
*skb
, int decr
)
1464 struct tcp_sock
*tp
= tcp_sk(sk
);
1466 tp
->packets_out
-= decr
;
1468 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
1469 tp
->sacked_out
-= decr
;
1470 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
)
1471 tp
->retrans_out
-= decr
;
1472 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_LOST
)
1473 tp
->lost_out
-= decr
;
1475 /* Reno case is special. Sigh... */
1476 if (tcp_is_reno(tp
) && decr
> 0)
1477 tp
->sacked_out
-= min_t(u32
, tp
->sacked_out
, decr
);
1479 if (tp
->lost_skb_hint
&&
1480 before(TCP_SKB_CB(skb
)->seq
, TCP_SKB_CB(tp
->lost_skb_hint
)->seq
) &&
1481 (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
))
1482 tp
->lost_cnt_hint
-= decr
;
1484 tcp_verify_left_out(tp
);
1487 static bool tcp_has_tx_tstamp(const struct sk_buff
*skb
)
1489 return TCP_SKB_CB(skb
)->txstamp_ack
||
1490 (skb_shinfo(skb
)->tx_flags
& SKBTX_ANY_TSTAMP
);
1493 static void tcp_fragment_tstamp(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1495 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
1497 if (unlikely(tcp_has_tx_tstamp(skb
)) &&
1498 !before(shinfo
->tskey
, TCP_SKB_CB(skb2
)->seq
)) {
1499 struct skb_shared_info
*shinfo2
= skb_shinfo(skb2
);
1500 u8 tsflags
= shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
1502 shinfo
->tx_flags
&= ~tsflags
;
1503 shinfo2
->tx_flags
|= tsflags
;
1504 swap(shinfo
->tskey
, shinfo2
->tskey
);
1505 TCP_SKB_CB(skb2
)->txstamp_ack
= TCP_SKB_CB(skb
)->txstamp_ack
;
1506 TCP_SKB_CB(skb
)->txstamp_ack
= 0;
1510 static void tcp_skb_fragment_eor(struct sk_buff
*skb
, struct sk_buff
*skb2
)
1512 TCP_SKB_CB(skb2
)->eor
= TCP_SKB_CB(skb
)->eor
;
1513 TCP_SKB_CB(skb
)->eor
= 0;
1516 /* Insert buff after skb on the write or rtx queue of sk. */
1517 static void tcp_insert_write_queue_after(struct sk_buff
*skb
,
1518 struct sk_buff
*buff
,
1520 enum tcp_queue tcp_queue
)
1522 if (tcp_queue
== TCP_FRAG_IN_WRITE_QUEUE
)
1523 __skb_queue_after(&sk
->sk_write_queue
, skb
, buff
);
1525 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, buff
);
1528 /* Function to create two new TCP segments. Shrinks the given segment
1529 * to the specified size and appends a new segment with the rest of the
1530 * packet to the list. This won't be called frequently, I hope.
1531 * Remember, these are still headerless SKBs at this point.
1533 int tcp_fragment(struct sock
*sk
, enum tcp_queue tcp_queue
,
1534 struct sk_buff
*skb
, u32 len
,
1535 unsigned int mss_now
, gfp_t gfp
)
1537 struct tcp_sock
*tp
= tcp_sk(sk
);
1538 struct sk_buff
*buff
;
1539 int nsize
, old_factor
;
1544 if (WARN_ON(len
> skb
->len
))
1547 nsize
= skb_headlen(skb
) - len
;
1551 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1552 * We need some allowance to not penalize applications setting small
1554 * Also allow first and last skb in retransmit queue to be split.
1556 limit
= sk
->sk_sndbuf
+ 2 * SKB_TRUESIZE(GSO_MAX_SIZE
);
1557 if (unlikely((sk
->sk_wmem_queued
>> 1) > limit
&&
1558 tcp_queue
!= TCP_FRAG_IN_WRITE_QUEUE
&&
1559 skb
!= tcp_rtx_queue_head(sk
) &&
1560 skb
!= tcp_rtx_queue_tail(sk
))) {
1561 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPWQUEUETOOBIG
);
1565 if (skb_unclone(skb
, gfp
))
1568 /* Get a new skb... force flag on. */
1569 buff
= sk_stream_alloc_skb(sk
, nsize
, gfp
, true);
1571 return -ENOMEM
; /* We'll just try again later. */
1572 skb_copy_decrypted(buff
, skb
);
1573 mptcp_skb_ext_copy(buff
, skb
);
1575 sk_wmem_queued_add(sk
, buff
->truesize
);
1576 sk_mem_charge(sk
, buff
->truesize
);
1577 nlen
= skb
->len
- len
- nsize
;
1578 buff
->truesize
+= nlen
;
1579 skb
->truesize
-= nlen
;
1581 /* Correct the sequence numbers. */
1582 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
1583 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
1584 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
1586 /* PSH and FIN should only be set in the second packet. */
1587 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
1588 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
1589 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
1590 TCP_SKB_CB(buff
)->sacked
= TCP_SKB_CB(skb
)->sacked
;
1591 tcp_skb_fragment_eor(skb
, buff
);
1593 skb_split(skb
, buff
, len
);
1595 buff
->ip_summed
= CHECKSUM_PARTIAL
;
1597 buff
->tstamp
= skb
->tstamp
;
1598 tcp_fragment_tstamp(skb
, buff
);
1600 old_factor
= tcp_skb_pcount(skb
);
1602 /* Fix up tso_factor for both original and new SKB. */
1603 tcp_set_skb_tso_segs(skb
, mss_now
);
1604 tcp_set_skb_tso_segs(buff
, mss_now
);
1606 /* Update delivered info for the new segment */
1607 TCP_SKB_CB(buff
)->tx
= TCP_SKB_CB(skb
)->tx
;
1609 /* If this packet has been sent out already, we must
1610 * adjust the various packet counters.
1612 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1613 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1614 tcp_skb_pcount(buff
);
1617 tcp_adjust_pcount(sk
, skb
, diff
);
1620 /* Link BUFF into the send queue. */
1621 __skb_header_release(buff
);
1622 tcp_insert_write_queue_after(skb
, buff
, sk
, tcp_queue
);
1623 if (tcp_queue
== TCP_FRAG_IN_RTX_QUEUE
)
1624 list_add(&buff
->tcp_tsorted_anchor
, &skb
->tcp_tsorted_anchor
);
1629 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1630 * data is not copied, but immediately discarded.
1632 static int __pskb_trim_head(struct sk_buff
*skb
, int len
)
1634 struct skb_shared_info
*shinfo
;
1637 eat
= min_t(int, len
, skb_headlen(skb
));
1639 __skb_pull(skb
, eat
);
1646 shinfo
= skb_shinfo(skb
);
1647 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1648 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1651 skb_frag_unref(skb
, i
);
1654 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1656 skb_frag_off_add(&shinfo
->frags
[k
], eat
);
1657 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1663 shinfo
->nr_frags
= k
;
1665 skb
->data_len
-= len
;
1666 skb
->len
= skb
->data_len
;
1670 /* Remove acked data from a packet in the transmit queue. */
1671 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1675 if (skb_unclone(skb
, GFP_ATOMIC
))
1678 delta_truesize
= __pskb_trim_head(skb
, len
);
1680 TCP_SKB_CB(skb
)->seq
+= len
;
1681 skb
->ip_summed
= CHECKSUM_PARTIAL
;
1683 if (delta_truesize
) {
1684 skb
->truesize
-= delta_truesize
;
1685 sk_wmem_queued_add(sk
, -delta_truesize
);
1686 sk_mem_uncharge(sk
, delta_truesize
);
1689 /* Any change of skb->len requires recalculation of tso factor. */
1690 if (tcp_skb_pcount(skb
) > 1)
1691 tcp_set_skb_tso_segs(skb
, tcp_skb_mss(skb
));
1696 /* Calculate MSS not accounting any TCP options. */
1697 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1699 const struct tcp_sock
*tp
= tcp_sk(sk
);
1700 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1703 /* Calculate base mss without TCP options:
1704 It is MMS_S - sizeof(tcphdr) of rfc1122
1706 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1708 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1709 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1710 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1712 if (dst
&& dst_allfrag(dst
))
1713 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1716 /* Clamp it (mss_clamp does not include tcp options) */
1717 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1718 mss_now
= tp
->rx_opt
.mss_clamp
;
1720 /* Now subtract optional transport overhead */
1721 mss_now
-= icsk
->icsk_ext_hdr_len
;
1723 /* Then reserve room for full set of TCP options and 8 bytes of data */
1724 mss_now
= max(mss_now
, sock_net(sk
)->ipv4
.sysctl_tcp_min_snd_mss
);
1728 /* Calculate MSS. Not accounting for SACKs here. */
1729 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1731 /* Subtract TCP options size, not including SACKs */
1732 return __tcp_mtu_to_mss(sk
, pmtu
) -
1733 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1735 EXPORT_SYMBOL(tcp_mtu_to_mss
);
1737 /* Inverse of above */
1738 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1740 const struct tcp_sock
*tp
= tcp_sk(sk
);
1741 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1745 tp
->tcp_header_len
+
1746 icsk
->icsk_ext_hdr_len
+
1747 icsk
->icsk_af_ops
->net_header_len
;
1749 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1750 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1751 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1753 if (dst
&& dst_allfrag(dst
))
1754 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1758 EXPORT_SYMBOL(tcp_mss_to_mtu
);
1760 /* MTU probing init per socket */
1761 void tcp_mtup_init(struct sock
*sk
)
1763 struct tcp_sock
*tp
= tcp_sk(sk
);
1764 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1765 struct net
*net
= sock_net(sk
);
1767 icsk
->icsk_mtup
.enabled
= net
->ipv4
.sysctl_tcp_mtu_probing
> 1;
1768 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1769 icsk
->icsk_af_ops
->net_header_len
;
1770 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, net
->ipv4
.sysctl_tcp_base_mss
);
1771 icsk
->icsk_mtup
.probe_size
= 0;
1772 if (icsk
->icsk_mtup
.enabled
)
1773 icsk
->icsk_mtup
.probe_timestamp
= tcp_jiffies32
;
1775 EXPORT_SYMBOL(tcp_mtup_init
);
1777 /* This function synchronize snd mss to current pmtu/exthdr set.
1779 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1780 for TCP options, but includes only bare TCP header.
1782 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1783 It is minimum of user_mss and mss received with SYN.
1784 It also does not include TCP options.
1786 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1788 tp->mss_cache is current effective sending mss, including
1789 all tcp options except for SACKs. It is evaluated,
1790 taking into account current pmtu, but never exceeds
1791 tp->rx_opt.mss_clamp.
1793 NOTE1. rfc1122 clearly states that advertised MSS
1794 DOES NOT include either tcp or ip options.
1796 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1797 are READ ONLY outside this function. --ANK (980731)
1799 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1801 struct tcp_sock
*tp
= tcp_sk(sk
);
1802 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1805 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1806 icsk
->icsk_mtup
.search_high
= pmtu
;
1808 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1809 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1811 /* And store cached results */
1812 icsk
->icsk_pmtu_cookie
= pmtu
;
1813 if (icsk
->icsk_mtup
.enabled
)
1814 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1815 tp
->mss_cache
= mss_now
;
1819 EXPORT_SYMBOL(tcp_sync_mss
);
1821 /* Compute the current effective MSS, taking SACKs and IP options,
1822 * and even PMTU discovery events into account.
1824 unsigned int tcp_current_mss(struct sock
*sk
)
1826 const struct tcp_sock
*tp
= tcp_sk(sk
);
1827 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1829 unsigned int header_len
;
1830 struct tcp_out_options opts
;
1831 struct tcp_md5sig_key
*md5
;
1833 mss_now
= tp
->mss_cache
;
1836 u32 mtu
= dst_mtu(dst
);
1837 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1838 mss_now
= tcp_sync_mss(sk
, mtu
);
1841 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1842 sizeof(struct tcphdr
);
1843 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1844 * some common options. If this is an odd packet (because we have SACK
1845 * blocks etc) then our calculated header_len will be different, and
1846 * we have to adjust mss_now correspondingly */
1847 if (header_len
!= tp
->tcp_header_len
) {
1848 int delta
= (int) header_len
- tp
->tcp_header_len
;
1855 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1856 * As additional protections, we do not touch cwnd in retransmission phases,
1857 * and if application hit its sndbuf limit recently.
1859 static void tcp_cwnd_application_limited(struct sock
*sk
)
1861 struct tcp_sock
*tp
= tcp_sk(sk
);
1863 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1864 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1865 /* Limited by application or receiver window. */
1866 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1867 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1868 if (win_used
< tp
->snd_cwnd
) {
1869 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1870 tp
->snd_cwnd
= (tp
->snd_cwnd
+ win_used
) >> 1;
1872 tp
->snd_cwnd_used
= 0;
1874 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1877 static void tcp_cwnd_validate(struct sock
*sk
, bool is_cwnd_limited
)
1879 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1880 struct tcp_sock
*tp
= tcp_sk(sk
);
1882 /* Track the maximum number of outstanding packets in each
1883 * window, and remember whether we were cwnd-limited then.
1885 if (!before(tp
->snd_una
, tp
->max_packets_seq
) ||
1886 tp
->packets_out
> tp
->max_packets_out
||
1888 tp
->max_packets_out
= tp
->packets_out
;
1889 tp
->max_packets_seq
= tp
->snd_nxt
;
1890 tp
->is_cwnd_limited
= is_cwnd_limited
;
1893 if (tcp_is_cwnd_limited(sk
)) {
1894 /* Network is feed fully. */
1895 tp
->snd_cwnd_used
= 0;
1896 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1898 /* Network starves. */
1899 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1900 tp
->snd_cwnd_used
= tp
->packets_out
;
1902 if (sock_net(sk
)->ipv4
.sysctl_tcp_slow_start_after_idle
&&
1903 (s32
)(tcp_jiffies32
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
&&
1904 !ca_ops
->cong_control
)
1905 tcp_cwnd_application_limited(sk
);
1907 /* The following conditions together indicate the starvation
1908 * is caused by insufficient sender buffer:
1909 * 1) just sent some data (see tcp_write_xmit)
1910 * 2) not cwnd limited (this else condition)
1911 * 3) no more data to send (tcp_write_queue_empty())
1912 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1914 if (tcp_write_queue_empty(sk
) && sk
->sk_socket
&&
1915 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
) &&
1916 (1 << sk
->sk_state
) & (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
1917 tcp_chrono_start(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1921 /* Minshall's variant of the Nagle send check. */
1922 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1924 return after(tp
->snd_sml
, tp
->snd_una
) &&
1925 !after(tp
->snd_sml
, tp
->snd_nxt
);
1928 /* Update snd_sml if this skb is under mss
1929 * Note that a TSO packet might end with a sub-mss segment
1930 * The test is really :
1931 * if ((skb->len % mss) != 0)
1932 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1933 * But we can avoid doing the divide again given we already have
1934 * skb_pcount = skb->len / mss_now
1936 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1937 const struct sk_buff
*skb
)
1939 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1940 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1943 /* Return false, if packet can be sent now without violation Nagle's rules:
1944 * 1. It is full sized. (provided by caller in %partial bool)
1945 * 2. Or it contains FIN. (already checked by caller)
1946 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1947 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1948 * With Minshall's modification: all sent small packets are ACKed.
1950 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1954 ((nonagle
& TCP_NAGLE_CORK
) ||
1955 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1958 /* Return how many segs we'd like on a TSO packet,
1959 * to send one TSO packet per ms
1961 static u32
tcp_tso_autosize(const struct sock
*sk
, unsigned int mss_now
,
1966 bytes
= min_t(unsigned long,
1967 sk
->sk_pacing_rate
>> READ_ONCE(sk
->sk_pacing_shift
),
1968 sk
->sk_gso_max_size
- 1 - MAX_TCP_HEADER
);
1970 /* Goal is to send at least one packet per ms,
1971 * not one big TSO packet every 100 ms.
1972 * This preserves ACK clocking and is consistent
1973 * with tcp_tso_should_defer() heuristic.
1975 segs
= max_t(u32
, bytes
/ mss_now
, min_tso_segs
);
1980 /* Return the number of segments we want in the skb we are transmitting.
1981 * See if congestion control module wants to decide; otherwise, autosize.
1983 static u32
tcp_tso_segs(struct sock
*sk
, unsigned int mss_now
)
1985 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1986 u32 min_tso
, tso_segs
;
1988 min_tso
= ca_ops
->min_tso_segs
?
1989 ca_ops
->min_tso_segs(sk
) :
1990 sock_net(sk
)->ipv4
.sysctl_tcp_min_tso_segs
;
1992 tso_segs
= tcp_tso_autosize(sk
, mss_now
, min_tso
);
1993 return min_t(u32
, tso_segs
, sk
->sk_gso_max_segs
);
1996 /* Returns the portion of skb which can be sent right away */
1997 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
1998 const struct sk_buff
*skb
,
1999 unsigned int mss_now
,
2000 unsigned int max_segs
,
2003 const struct tcp_sock
*tp
= tcp_sk(sk
);
2004 u32 partial
, needed
, window
, max_len
;
2006 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2007 max_len
= mss_now
* max_segs
;
2009 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
2012 needed
= min(skb
->len
, window
);
2014 if (max_len
<= needed
)
2017 partial
= needed
% mss_now
;
2018 /* If last segment is not a full MSS, check if Nagle rules allow us
2019 * to include this last segment in this skb.
2020 * Otherwise, we'll split the skb at last MSS boundary
2022 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
2023 return needed
- partial
;
2028 /* Can at least one segment of SKB be sent right now, according to the
2029 * congestion window rules? If so, return how many segments are allowed.
2031 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
2032 const struct sk_buff
*skb
)
2034 u32 in_flight
, cwnd
, halfcwnd
;
2036 /* Don't be strict about the congestion window for the final FIN. */
2037 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
2038 tcp_skb_pcount(skb
) == 1)
2041 in_flight
= tcp_packets_in_flight(tp
);
2042 cwnd
= tp
->snd_cwnd
;
2043 if (in_flight
>= cwnd
)
2046 /* For better scheduling, ensure we have at least
2047 * 2 GSO packets in flight.
2049 halfcwnd
= max(cwnd
>> 1, 1U);
2050 return min(halfcwnd
, cwnd
- in_flight
);
2053 /* Initialize TSO state of a skb.
2054 * This must be invoked the first time we consider transmitting
2055 * SKB onto the wire.
2057 static int tcp_init_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
2059 int tso_segs
= tcp_skb_pcount(skb
);
2061 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
2062 tcp_set_skb_tso_segs(skb
, mss_now
);
2063 tso_segs
= tcp_skb_pcount(skb
);
2069 /* Return true if the Nagle test allows this packet to be
2072 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
2073 unsigned int cur_mss
, int nonagle
)
2075 /* Nagle rule does not apply to frames, which sit in the middle of the
2076 * write_queue (they have no chances to get new data).
2078 * This is implemented in the callers, where they modify the 'nonagle'
2079 * argument based upon the location of SKB in the send queue.
2081 if (nonagle
& TCP_NAGLE_PUSH
)
2084 /* Don't use the nagle rule for urgent data (or for the final FIN). */
2085 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
2088 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
2094 /* Does at least the first segment of SKB fit into the send window? */
2095 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
2096 const struct sk_buff
*skb
,
2097 unsigned int cur_mss
)
2099 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
2101 if (skb
->len
> cur_mss
)
2102 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
2104 return !after(end_seq
, tcp_wnd_end(tp
));
2107 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2108 * which is put after SKB on the list. It is very much like
2109 * tcp_fragment() except that it may make several kinds of assumptions
2110 * in order to speed up the splitting operation. In particular, we
2111 * know that all the data is in scatter-gather pages, and that the
2112 * packet has never been sent out before (and thus is not cloned).
2114 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
2115 unsigned int mss_now
, gfp_t gfp
)
2117 int nlen
= skb
->len
- len
;
2118 struct sk_buff
*buff
;
2121 /* All of a TSO frame must be composed of paged data. */
2122 if (skb
->len
!= skb
->data_len
)
2123 return tcp_fragment(sk
, TCP_FRAG_IN_WRITE_QUEUE
,
2124 skb
, len
, mss_now
, gfp
);
2126 buff
= sk_stream_alloc_skb(sk
, 0, gfp
, true);
2127 if (unlikely(!buff
))
2129 skb_copy_decrypted(buff
, skb
);
2130 mptcp_skb_ext_copy(buff
, skb
);
2132 sk_wmem_queued_add(sk
, buff
->truesize
);
2133 sk_mem_charge(sk
, buff
->truesize
);
2134 buff
->truesize
+= nlen
;
2135 skb
->truesize
-= nlen
;
2137 /* Correct the sequence numbers. */
2138 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
2139 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
2140 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
2142 /* PSH and FIN should only be set in the second packet. */
2143 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
2144 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
2145 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
2147 /* This packet was never sent out yet, so no SACK bits. */
2148 TCP_SKB_CB(buff
)->sacked
= 0;
2150 tcp_skb_fragment_eor(skb
, buff
);
2152 buff
->ip_summed
= CHECKSUM_PARTIAL
;
2153 skb_split(skb
, buff
, len
);
2154 tcp_fragment_tstamp(skb
, buff
);
2156 /* Fix up tso_factor for both original and new SKB. */
2157 tcp_set_skb_tso_segs(skb
, mss_now
);
2158 tcp_set_skb_tso_segs(buff
, mss_now
);
2160 /* Link BUFF into the send queue. */
2161 __skb_header_release(buff
);
2162 tcp_insert_write_queue_after(skb
, buff
, sk
, TCP_FRAG_IN_WRITE_QUEUE
);
2167 /* Try to defer sending, if possible, in order to minimize the amount
2168 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2170 * This algorithm is from John Heffner.
2172 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
,
2173 bool *is_cwnd_limited
,
2174 bool *is_rwnd_limited
,
2177 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2178 u32 send_win
, cong_win
, limit
, in_flight
;
2179 struct tcp_sock
*tp
= tcp_sk(sk
);
2180 struct sk_buff
*head
;
2184 if (icsk
->icsk_ca_state
>= TCP_CA_Recovery
)
2187 /* Avoid bursty behavior by allowing defer
2188 * only if the last write was recent (1 ms).
2189 * Note that tp->tcp_wstamp_ns can be in the future if we have
2190 * packets waiting in a qdisc or device for EDT delivery.
2192 delta
= tp
->tcp_clock_cache
- tp
->tcp_wstamp_ns
- NSEC_PER_MSEC
;
2196 in_flight
= tcp_packets_in_flight(tp
);
2198 BUG_ON(tcp_skb_pcount(skb
) <= 1);
2199 BUG_ON(tp
->snd_cwnd
<= in_flight
);
2201 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2203 /* From in_flight test above, we know that cwnd > in_flight. */
2204 cong_win
= (tp
->snd_cwnd
- in_flight
) * tp
->mss_cache
;
2206 limit
= min(send_win
, cong_win
);
2208 /* If a full-sized TSO skb can be sent, do it. */
2209 if (limit
>= max_segs
* tp
->mss_cache
)
2212 /* Middle in queue won't get any more data, full sendable already? */
2213 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
2216 win_divisor
= READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_tso_win_divisor
);
2218 u32 chunk
= min(tp
->snd_wnd
, tp
->snd_cwnd
* tp
->mss_cache
);
2220 /* If at least some fraction of a window is available,
2223 chunk
/= win_divisor
;
2227 /* Different approach, try not to defer past a single
2228 * ACK. Receiver should ACK every other full sized
2229 * frame, so if we have space for more than 3 frames
2232 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
2236 /* TODO : use tsorted_sent_queue ? */
2237 head
= tcp_rtx_queue_head(sk
);
2240 delta
= tp
->tcp_clock_cache
- head
->tstamp
;
2241 /* If next ACK is likely to come too late (half srtt), do not defer */
2242 if ((s64
)(delta
- (u64
)NSEC_PER_USEC
* (tp
->srtt_us
>> 4)) < 0)
2245 /* Ok, it looks like it is advisable to defer.
2246 * Three cases are tracked :
2247 * 1) We are cwnd-limited
2248 * 2) We are rwnd-limited
2249 * 3) We are application limited.
2251 if (cong_win
< send_win
) {
2252 if (cong_win
<= skb
->len
) {
2253 *is_cwnd_limited
= true;
2257 if (send_win
<= skb
->len
) {
2258 *is_rwnd_limited
= true;
2263 /* If this packet won't get more data, do not wait. */
2264 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) ||
2265 TCP_SKB_CB(skb
)->eor
)
2274 static inline void tcp_mtu_check_reprobe(struct sock
*sk
)
2276 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2277 struct tcp_sock
*tp
= tcp_sk(sk
);
2278 struct net
*net
= sock_net(sk
);
2282 interval
= net
->ipv4
.sysctl_tcp_probe_interval
;
2283 delta
= tcp_jiffies32
- icsk
->icsk_mtup
.probe_timestamp
;
2284 if (unlikely(delta
>= interval
* HZ
)) {
2285 int mss
= tcp_current_mss(sk
);
2287 /* Update current search range */
2288 icsk
->icsk_mtup
.probe_size
= 0;
2289 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+
2290 sizeof(struct tcphdr
) +
2291 icsk
->icsk_af_ops
->net_header_len
;
2292 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
2294 /* Update probe time stamp */
2295 icsk
->icsk_mtup
.probe_timestamp
= tcp_jiffies32
;
2299 static bool tcp_can_coalesce_send_queue_head(struct sock
*sk
, int len
)
2301 struct sk_buff
*skb
, *next
;
2303 skb
= tcp_send_head(sk
);
2304 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
2305 if (len
<= skb
->len
)
2308 if (unlikely(TCP_SKB_CB(skb
)->eor
) || tcp_has_tx_tstamp(skb
))
2317 /* Create a new MTU probe if we are ready.
2318 * MTU probe is regularly attempting to increase the path MTU by
2319 * deliberately sending larger packets. This discovers routing
2320 * changes resulting in larger path MTUs.
2322 * Returns 0 if we should wait to probe (no cwnd available),
2323 * 1 if a probe was sent,
2326 static int tcp_mtu_probe(struct sock
*sk
)
2328 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2329 struct tcp_sock
*tp
= tcp_sk(sk
);
2330 struct sk_buff
*skb
, *nskb
, *next
;
2331 struct net
*net
= sock_net(sk
);
2338 /* Not currently probing/verifying,
2340 * have enough cwnd, and
2341 * not SACKing (the variable headers throw things off)
2343 if (likely(!icsk
->icsk_mtup
.enabled
||
2344 icsk
->icsk_mtup
.probe_size
||
2345 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
2346 tp
->snd_cwnd
< 11 ||
2347 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
))
2350 /* Use binary search for probe_size between tcp_mss_base,
2351 * and current mss_clamp. if (search_high - search_low)
2352 * smaller than a threshold, backoff from probing.
2354 mss_now
= tcp_current_mss(sk
);
2355 probe_size
= tcp_mtu_to_mss(sk
, (icsk
->icsk_mtup
.search_high
+
2356 icsk
->icsk_mtup
.search_low
) >> 1);
2357 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
2358 interval
= icsk
->icsk_mtup
.search_high
- icsk
->icsk_mtup
.search_low
;
2359 /* When misfortune happens, we are reprobing actively,
2360 * and then reprobe timer has expired. We stick with current
2361 * probing process by not resetting search range to its orignal.
2363 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
) ||
2364 interval
< net
->ipv4
.sysctl_tcp_probe_threshold
) {
2365 /* Check whether enough time has elaplased for
2366 * another round of probing.
2368 tcp_mtu_check_reprobe(sk
);
2372 /* Have enough data in the send queue to probe? */
2373 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
2376 if (tp
->snd_wnd
< size_needed
)
2378 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
2381 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2382 if (tcp_packets_in_flight(tp
) + 2 > tp
->snd_cwnd
) {
2383 if (!tcp_packets_in_flight(tp
))
2389 if (!tcp_can_coalesce_send_queue_head(sk
, probe_size
))
2392 /* We're allowed to probe. Build it now. */
2393 nskb
= sk_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
, false);
2396 sk_wmem_queued_add(sk
, nskb
->truesize
);
2397 sk_mem_charge(sk
, nskb
->truesize
);
2399 skb
= tcp_send_head(sk
);
2400 skb_copy_decrypted(nskb
, skb
);
2401 mptcp_skb_ext_copy(nskb
, skb
);
2403 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
2404 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
2405 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
2406 TCP_SKB_CB(nskb
)->sacked
= 0;
2408 nskb
->ip_summed
= CHECKSUM_PARTIAL
;
2410 tcp_insert_write_queue_before(nskb
, skb
, sk
);
2411 tcp_highest_sack_replace(sk
, skb
, nskb
);
2414 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
2415 copy
= min_t(int, skb
->len
, probe_size
- len
);
2416 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
2418 if (skb
->len
<= copy
) {
2419 /* We've eaten all the data from this skb.
2421 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
2422 /* If this is the last SKB we copy and eor is set
2423 * we need to propagate it to the new skb.
2425 TCP_SKB_CB(nskb
)->eor
= TCP_SKB_CB(skb
)->eor
;
2426 tcp_skb_collapse_tstamp(nskb
, skb
);
2427 tcp_unlink_write_queue(skb
, sk
);
2428 sk_wmem_free_skb(sk
, skb
);
2430 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
2431 ~(TCPHDR_FIN
|TCPHDR_PSH
);
2432 if (!skb_shinfo(skb
)->nr_frags
) {
2433 skb_pull(skb
, copy
);
2435 __pskb_trim_head(skb
, copy
);
2436 tcp_set_skb_tso_segs(skb
, mss_now
);
2438 TCP_SKB_CB(skb
)->seq
+= copy
;
2443 if (len
>= probe_size
)
2446 tcp_init_tso_segs(nskb
, nskb
->len
);
2448 /* We're ready to send. If this fails, the probe will
2449 * be resegmented into mss-sized pieces by tcp_write_xmit().
2451 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
2452 /* Decrement cwnd here because we are sending
2453 * effectively two packets. */
2455 tcp_event_new_data_sent(sk
, nskb
);
2457 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
2458 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
2459 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
2467 static bool tcp_pacing_check(struct sock
*sk
)
2469 struct tcp_sock
*tp
= tcp_sk(sk
);
2471 if (!tcp_needs_internal_pacing(sk
))
2474 if (tp
->tcp_wstamp_ns
<= tp
->tcp_clock_cache
)
2477 if (!hrtimer_is_queued(&tp
->pacing_timer
)) {
2478 hrtimer_start(&tp
->pacing_timer
,
2479 ns_to_ktime(tp
->tcp_wstamp_ns
),
2480 HRTIMER_MODE_ABS_PINNED_SOFT
);
2486 /* TCP Small Queues :
2487 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2488 * (These limits are doubled for retransmits)
2490 * - better RTT estimation and ACK scheduling
2493 * Alas, some drivers / subsystems require a fair amount
2494 * of queued bytes to ensure line rate.
2495 * One example is wifi aggregation (802.11 AMPDU)
2497 static bool tcp_small_queue_check(struct sock
*sk
, const struct sk_buff
*skb
,
2498 unsigned int factor
)
2500 unsigned long limit
;
2502 limit
= max_t(unsigned long,
2504 sk
->sk_pacing_rate
>> READ_ONCE(sk
->sk_pacing_shift
));
2505 if (sk
->sk_pacing_status
== SK_PACING_NONE
)
2506 limit
= min_t(unsigned long, limit
,
2507 sock_net(sk
)->ipv4
.sysctl_tcp_limit_output_bytes
);
2510 if (static_branch_unlikely(&tcp_tx_delay_enabled
) &&
2511 tcp_sk(sk
)->tcp_tx_delay
) {
2512 u64 extra_bytes
= (u64
)sk
->sk_pacing_rate
* tcp_sk(sk
)->tcp_tx_delay
;
2514 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2515 * approximate our needs assuming an ~100% skb->truesize overhead.
2516 * USEC_PER_SEC is approximated by 2^20.
2517 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2519 extra_bytes
>>= (20 - 1);
2520 limit
+= extra_bytes
;
2522 if (refcount_read(&sk
->sk_wmem_alloc
) > limit
) {
2523 /* Always send skb if rtx queue is empty.
2524 * No need to wait for TX completion to call us back,
2525 * after softirq/tasklet schedule.
2526 * This helps when TX completions are delayed too much.
2528 if (tcp_rtx_queue_empty(sk
))
2531 set_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
);
2532 /* It is possible TX completion already happened
2533 * before we set TSQ_THROTTLED, so we must
2534 * test again the condition.
2536 smp_mb__after_atomic();
2537 if (refcount_read(&sk
->sk_wmem_alloc
) > limit
)
2543 static void tcp_chrono_set(struct tcp_sock
*tp
, const enum tcp_chrono
new)
2545 const u32 now
= tcp_jiffies32
;
2546 enum tcp_chrono old
= tp
->chrono_type
;
2548 if (old
> TCP_CHRONO_UNSPEC
)
2549 tp
->chrono_stat
[old
- 1] += now
- tp
->chrono_start
;
2550 tp
->chrono_start
= now
;
2551 tp
->chrono_type
= new;
2554 void tcp_chrono_start(struct sock
*sk
, const enum tcp_chrono type
)
2556 struct tcp_sock
*tp
= tcp_sk(sk
);
2558 /* If there are multiple conditions worthy of tracking in a
2559 * chronograph then the highest priority enum takes precedence
2560 * over the other conditions. So that if something "more interesting"
2561 * starts happening, stop the previous chrono and start a new one.
2563 if (type
> tp
->chrono_type
)
2564 tcp_chrono_set(tp
, type
);
2567 void tcp_chrono_stop(struct sock
*sk
, const enum tcp_chrono type
)
2569 struct tcp_sock
*tp
= tcp_sk(sk
);
2572 /* There are multiple conditions worthy of tracking in a
2573 * chronograph, so that the highest priority enum takes
2574 * precedence over the other conditions (see tcp_chrono_start).
2575 * If a condition stops, we only stop chrono tracking if
2576 * it's the "most interesting" or current chrono we are
2577 * tracking and starts busy chrono if we have pending data.
2579 if (tcp_rtx_and_write_queues_empty(sk
))
2580 tcp_chrono_set(tp
, TCP_CHRONO_UNSPEC
);
2581 else if (type
== tp
->chrono_type
)
2582 tcp_chrono_set(tp
, TCP_CHRONO_BUSY
);
2585 /* This routine writes packets to the network. It advances the
2586 * send_head. This happens as incoming acks open up the remote
2589 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2590 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2591 * account rare use of URG, this is not a big flaw.
2593 * Send at most one packet when push_one > 0. Temporarily ignore
2594 * cwnd limit to force at most one packet out when push_one == 2.
2596 * Returns true, if no segments are in flight and we have queued segments,
2597 * but cannot send anything now because of SWS or another problem.
2599 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
2600 int push_one
, gfp_t gfp
)
2602 struct tcp_sock
*tp
= tcp_sk(sk
);
2603 struct sk_buff
*skb
;
2604 unsigned int tso_segs
, sent_pkts
;
2607 bool is_cwnd_limited
= false, is_rwnd_limited
= false;
2612 tcp_mstamp_refresh(tp
);
2614 /* Do MTU probing. */
2615 result
= tcp_mtu_probe(sk
);
2618 } else if (result
> 0) {
2623 max_segs
= tcp_tso_segs(sk
, mss_now
);
2624 while ((skb
= tcp_send_head(sk
))) {
2627 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
) {
2628 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2629 skb
->skb_mstamp_ns
= tp
->tcp_wstamp_ns
= tp
->tcp_clock_cache
;
2630 list_move_tail(&skb
->tcp_tsorted_anchor
, &tp
->tsorted_sent_queue
);
2631 tcp_init_tso_segs(skb
, mss_now
);
2632 goto repair
; /* Skip network transmission */
2635 if (tcp_pacing_check(sk
))
2638 tso_segs
= tcp_init_tso_segs(skb
, mss_now
);
2641 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2644 /* Force out a loss probe pkt. */
2650 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
))) {
2651 is_rwnd_limited
= true;
2655 if (tso_segs
== 1) {
2656 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2657 (tcp_skb_is_last(sk
, skb
) ?
2658 nonagle
: TCP_NAGLE_PUSH
))))
2662 tcp_tso_should_defer(sk
, skb
, &is_cwnd_limited
,
2663 &is_rwnd_limited
, max_segs
))
2668 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
2669 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2675 if (skb
->len
> limit
&&
2676 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
2679 if (tcp_small_queue_check(sk
, skb
, 0))
2682 /* Argh, we hit an empty skb(), presumably a thread
2683 * is sleeping in sendmsg()/sk_stream_wait_memory().
2684 * We do not want to send a pure-ack packet and have
2685 * a strange looking rtx queue with empty packet(s).
2687 if (TCP_SKB_CB(skb
)->end_seq
== TCP_SKB_CB(skb
)->seq
)
2690 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2694 /* Advance the send_head. This one is sent out.
2695 * This call will increment packets_out.
2697 tcp_event_new_data_sent(sk
, skb
);
2699 tcp_minshall_update(tp
, mss_now
, skb
);
2700 sent_pkts
+= tcp_skb_pcount(skb
);
2706 if (is_rwnd_limited
)
2707 tcp_chrono_start(sk
, TCP_CHRONO_RWND_LIMITED
);
2709 tcp_chrono_stop(sk
, TCP_CHRONO_RWND_LIMITED
);
2711 is_cwnd_limited
|= (tcp_packets_in_flight(tp
) >= tp
->snd_cwnd
);
2712 if (likely(sent_pkts
|| is_cwnd_limited
))
2713 tcp_cwnd_validate(sk
, is_cwnd_limited
);
2715 if (likely(sent_pkts
)) {
2716 if (tcp_in_cwnd_reduction(sk
))
2717 tp
->prr_out
+= sent_pkts
;
2719 /* Send one loss probe per tail loss episode. */
2721 tcp_schedule_loss_probe(sk
, false);
2724 return !tp
->packets_out
&& !tcp_write_queue_empty(sk
);
2727 bool tcp_schedule_loss_probe(struct sock
*sk
, bool advancing_rto
)
2729 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2730 struct tcp_sock
*tp
= tcp_sk(sk
);
2731 u32 timeout
, rto_delta_us
;
2734 /* Don't do any loss probe on a Fast Open connection before 3WHS
2737 if (rcu_access_pointer(tp
->fastopen_rsk
))
2740 early_retrans
= sock_net(sk
)->ipv4
.sysctl_tcp_early_retrans
;
2741 /* Schedule a loss probe in 2*RTT for SACK capable connections
2742 * not in loss recovery, that are either limited by cwnd or application.
2744 if ((early_retrans
!= 3 && early_retrans
!= 4) ||
2745 !tp
->packets_out
|| !tcp_is_sack(tp
) ||
2746 (icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2747 icsk
->icsk_ca_state
!= TCP_CA_CWR
))
2750 /* Probe timeout is 2*rtt. Add minimum RTO to account
2751 * for delayed ack when there's one outstanding packet. If no RTT
2752 * sample is available then probe after TCP_TIMEOUT_INIT.
2755 timeout
= usecs_to_jiffies(tp
->srtt_us
>> 2);
2756 if (tp
->packets_out
== 1)
2757 timeout
+= TCP_RTO_MIN
;
2759 timeout
+= TCP_TIMEOUT_MIN
;
2761 timeout
= TCP_TIMEOUT_INIT
;
2764 /* If the RTO formula yields an earlier time, then use that time. */
2765 rto_delta_us
= advancing_rto
?
2766 jiffies_to_usecs(inet_csk(sk
)->icsk_rto
) :
2767 tcp_rto_delta_us(sk
); /* How far in future is RTO? */
2768 if (rto_delta_us
> 0)
2769 timeout
= min_t(u32
, timeout
, usecs_to_jiffies(rto_delta_us
));
2771 tcp_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
, TCP_RTO_MAX
);
2775 /* Thanks to skb fast clones, we can detect if a prior transmit of
2776 * a packet is still in a qdisc or driver queue.
2777 * In this case, there is very little point doing a retransmit !
2779 static bool skb_still_in_host_queue(struct sock
*sk
,
2780 const struct sk_buff
*skb
)
2782 if (unlikely(skb_fclone_busy(sk
, skb
))) {
2783 set_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
);
2784 smp_mb__after_atomic();
2785 if (skb_fclone_busy(sk
, skb
)) {
2786 NET_INC_STATS(sock_net(sk
),
2787 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
2794 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2795 * retransmit the last segment.
2797 void tcp_send_loss_probe(struct sock
*sk
)
2799 struct tcp_sock
*tp
= tcp_sk(sk
);
2800 struct sk_buff
*skb
;
2802 int mss
= tcp_current_mss(sk
);
2804 /* At most one outstanding TLP */
2805 if (tp
->tlp_high_seq
)
2808 tp
->tlp_retrans
= 0;
2809 skb
= tcp_send_head(sk
);
2810 if (skb
&& tcp_snd_wnd_test(tp
, skb
, mss
)) {
2811 pcount
= tp
->packets_out
;
2812 tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2813 if (tp
->packets_out
> pcount
)
2817 skb
= skb_rb_last(&sk
->tcp_rtx_queue
);
2818 if (unlikely(!skb
)) {
2819 WARN_ONCE(tp
->packets_out
,
2820 "invalid inflight: %u state %u cwnd %u mss %d\n",
2821 tp
->packets_out
, sk
->sk_state
, tp
->snd_cwnd
, mss
);
2822 inet_csk(sk
)->icsk_pending
= 0;
2826 if (skb_still_in_host_queue(sk
, skb
))
2829 pcount
= tcp_skb_pcount(skb
);
2830 if (WARN_ON(!pcount
))
2833 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2834 if (unlikely(tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
,
2835 (pcount
- 1) * mss
, mss
,
2838 skb
= skb_rb_next(skb
);
2841 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2844 if (__tcp_retransmit_skb(sk
, skb
, 1))
2847 tp
->tlp_retrans
= 1;
2850 /* Record snd_nxt for loss detection. */
2851 tp
->tlp_high_seq
= tp
->snd_nxt
;
2853 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSPROBES
);
2854 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2855 inet_csk(sk
)->icsk_pending
= 0;
2860 /* Push out any pending frames which were held back due to
2861 * TCP_CORK or attempt at coalescing tiny packets.
2862 * The socket must be locked by the caller.
2864 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2867 /* If we are closed, the bytes will have to remain here.
2868 * In time closedown will finish, we empty the write queue and
2869 * all will be happy.
2871 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2874 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2875 sk_gfp_mask(sk
, GFP_ATOMIC
)))
2876 tcp_check_probe_timer(sk
);
2879 /* Send _single_ skb sitting at the send head. This function requires
2880 * true push pending frames to setup probe timer etc.
2882 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2884 struct sk_buff
*skb
= tcp_send_head(sk
);
2886 BUG_ON(!skb
|| skb
->len
< mss_now
);
2888 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2891 /* This function returns the amount that we can raise the
2892 * usable window based on the following constraints
2894 * 1. The window can never be shrunk once it is offered (RFC 793)
2895 * 2. We limit memory per socket
2898 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2899 * RECV.NEXT + RCV.WIN fixed until:
2900 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2902 * i.e. don't raise the right edge of the window until you can raise
2903 * it at least MSS bytes.
2905 * Unfortunately, the recommended algorithm breaks header prediction,
2906 * since header prediction assumes th->window stays fixed.
2908 * Strictly speaking, keeping th->window fixed violates the receiver
2909 * side SWS prevention criteria. The problem is that under this rule
2910 * a stream of single byte packets will cause the right side of the
2911 * window to always advance by a single byte.
2913 * Of course, if the sender implements sender side SWS prevention
2914 * then this will not be a problem.
2916 * BSD seems to make the following compromise:
2918 * If the free space is less than the 1/4 of the maximum
2919 * space available and the free space is less than 1/2 mss,
2920 * then set the window to 0.
2921 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2922 * Otherwise, just prevent the window from shrinking
2923 * and from being larger than the largest representable value.
2925 * This prevents incremental opening of the window in the regime
2926 * where TCP is limited by the speed of the reader side taking
2927 * data out of the TCP receive queue. It does nothing about
2928 * those cases where the window is constrained on the sender side
2929 * because the pipeline is full.
2931 * BSD also seems to "accidentally" limit itself to windows that are a
2932 * multiple of MSS, at least until the free space gets quite small.
2933 * This would appear to be a side effect of the mbuf implementation.
2934 * Combining these two algorithms results in the observed behavior
2935 * of having a fixed window size at almost all times.
2937 * Below we obtain similar behavior by forcing the offered window to
2938 * a multiple of the mss when it is feasible to do so.
2940 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2941 * Regular options like TIMESTAMP are taken into account.
2943 u32
__tcp_select_window(struct sock
*sk
)
2945 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2946 struct tcp_sock
*tp
= tcp_sk(sk
);
2947 /* MSS for the peer's data. Previous versions used mss_clamp
2948 * here. I don't know if the value based on our guesses
2949 * of peer's MSS is better for the performance. It's more correct
2950 * but may be worse for the performance because of rcv_mss
2951 * fluctuations. --SAW 1998/11/1
2953 int mss
= icsk
->icsk_ack
.rcv_mss
;
2954 int free_space
= tcp_space(sk
);
2955 int allowed_space
= tcp_full_space(sk
);
2956 int full_space
, window
;
2958 if (sk_is_mptcp(sk
))
2959 mptcp_space(sk
, &free_space
, &allowed_space
);
2961 full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2963 if (unlikely(mss
> full_space
)) {
2968 if (free_space
< (full_space
>> 1)) {
2969 icsk
->icsk_ack
.quick
= 0;
2971 if (tcp_under_memory_pressure(sk
))
2972 tp
->rcv_ssthresh
= min(tp
->rcv_ssthresh
,
2975 /* free_space might become our new window, make sure we don't
2976 * increase it due to wscale.
2978 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2980 /* if free space is less than mss estimate, or is below 1/16th
2981 * of the maximum allowed, try to move to zero-window, else
2982 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2983 * new incoming data is dropped due to memory limits.
2984 * With large window, mss test triggers way too late in order
2985 * to announce zero window in time before rmem limit kicks in.
2987 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2991 if (free_space
> tp
->rcv_ssthresh
)
2992 free_space
= tp
->rcv_ssthresh
;
2994 /* Don't do rounding if we are using window scaling, since the
2995 * scaled window will not line up with the MSS boundary anyway.
2997 if (tp
->rx_opt
.rcv_wscale
) {
2998 window
= free_space
;
3000 /* Advertise enough space so that it won't get scaled away.
3001 * Import case: prevent zero window announcement if
3002 * 1<<rcv_wscale > mss.
3004 window
= ALIGN(window
, (1 << tp
->rx_opt
.rcv_wscale
));
3006 window
= tp
->rcv_wnd
;
3007 /* Get the largest window that is a nice multiple of mss.
3008 * Window clamp already applied above.
3009 * If our current window offering is within 1 mss of the
3010 * free space we just keep it. This prevents the divide
3011 * and multiply from happening most of the time.
3012 * We also don't do any window rounding when the free space
3015 if (window
<= free_space
- mss
|| window
> free_space
)
3016 window
= rounddown(free_space
, mss
);
3017 else if (mss
== full_space
&&
3018 free_space
> window
+ (full_space
>> 1))
3019 window
= free_space
;
3025 void tcp_skb_collapse_tstamp(struct sk_buff
*skb
,
3026 const struct sk_buff
*next_skb
)
3028 if (unlikely(tcp_has_tx_tstamp(next_skb
))) {
3029 const struct skb_shared_info
*next_shinfo
=
3030 skb_shinfo(next_skb
);
3031 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
3033 shinfo
->tx_flags
|= next_shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
3034 shinfo
->tskey
= next_shinfo
->tskey
;
3035 TCP_SKB_CB(skb
)->txstamp_ack
|=
3036 TCP_SKB_CB(next_skb
)->txstamp_ack
;
3040 /* Collapses two adjacent SKB's during retransmission. */
3041 static bool tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
3043 struct tcp_sock
*tp
= tcp_sk(sk
);
3044 struct sk_buff
*next_skb
= skb_rb_next(skb
);
3047 next_skb_size
= next_skb
->len
;
3049 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
3051 if (next_skb_size
) {
3052 if (next_skb_size
<= skb_availroom(skb
))
3053 skb_copy_bits(next_skb
, 0, skb_put(skb
, next_skb_size
),
3055 else if (!tcp_skb_shift(skb
, next_skb
, 1, next_skb_size
))
3058 tcp_highest_sack_replace(sk
, next_skb
, skb
);
3060 /* Update sequence range on original skb. */
3061 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
3063 /* Merge over control information. This moves PSH/FIN etc. over */
3064 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
3066 /* All done, get rid of second SKB and account for it so
3067 * packet counting does not break.
3069 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
3070 TCP_SKB_CB(skb
)->eor
= TCP_SKB_CB(next_skb
)->eor
;
3072 /* changed transmit queue under us so clear hints */
3073 tcp_clear_retrans_hints_partial(tp
);
3074 if (next_skb
== tp
->retransmit_skb_hint
)
3075 tp
->retransmit_skb_hint
= skb
;
3077 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
3079 tcp_skb_collapse_tstamp(skb
, next_skb
);
3081 tcp_rtx_queue_unlink_and_free(next_skb
, sk
);
3085 /* Check if coalescing SKBs is legal. */
3086 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
3088 if (tcp_skb_pcount(skb
) > 1)
3090 if (skb_cloned(skb
))
3092 /* Some heuristics for collapsing over SACK'd could be invented */
3093 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
3099 /* Collapse packets in the retransmit queue to make to create
3100 * less packets on the wire. This is only done on retransmission.
3102 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
3105 struct tcp_sock
*tp
= tcp_sk(sk
);
3106 struct sk_buff
*skb
= to
, *tmp
;
3109 if (!sock_net(sk
)->ipv4
.sysctl_tcp_retrans_collapse
)
3111 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
3114 skb_rbtree_walk_from_safe(skb
, tmp
) {
3115 if (!tcp_can_collapse(sk
, skb
))
3118 if (!tcp_skb_can_collapse(to
, skb
))
3131 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
3134 if (!tcp_collapse_retrans(sk
, to
))
3139 /* This retransmits one SKB. Policy decisions and retransmit queue
3140 * state updates are done by the caller. Returns non-zero if an
3141 * error occurred which prevented the send.
3143 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
3145 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3146 struct tcp_sock
*tp
= tcp_sk(sk
);
3147 unsigned int cur_mss
;
3151 /* Inconclusive MTU probe */
3152 if (icsk
->icsk_mtup
.probe_size
)
3153 icsk
->icsk_mtup
.probe_size
= 0;
3155 if (skb_still_in_host_queue(sk
, skb
))
3158 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
3159 if (unlikely(before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))) {
3163 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3167 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
3168 return -EHOSTUNREACH
; /* Routing failure or similar. */
3170 cur_mss
= tcp_current_mss(sk
);
3172 /* If receiver has shrunk his window, and skb is out of
3173 * new window, do not retransmit it. The exception is the
3174 * case, when window is shrunk to zero. In this case
3175 * our retransmit serves as a zero window probe.
3177 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
3178 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
3181 len
= cur_mss
* segs
;
3182 if (skb
->len
> len
) {
3183 if (tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
, len
,
3184 cur_mss
, GFP_ATOMIC
))
3185 return -ENOMEM
; /* We'll try again later. */
3187 if (skb_unclone(skb
, GFP_ATOMIC
))
3190 diff
= tcp_skb_pcount(skb
);
3191 tcp_set_skb_tso_segs(skb
, cur_mss
);
3192 diff
-= tcp_skb_pcount(skb
);
3194 tcp_adjust_pcount(sk
, skb
, diff
);
3195 if (skb
->len
< cur_mss
)
3196 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
3199 /* RFC3168, section 6.1.1.1. ECN fallback */
3200 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN_ECN
) == TCPHDR_SYN_ECN
)
3201 tcp_ecn_clear_syn(sk
, skb
);
3203 /* Update global and local TCP statistics. */
3204 segs
= tcp_skb_pcount(skb
);
3205 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
, segs
);
3206 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
3207 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
3208 tp
->total_retrans
+= segs
;
3209 tp
->bytes_retrans
+= skb
->len
;
3211 /* make sure skb->data is aligned on arches that require it
3212 * and check if ack-trimming & collapsing extended the headroom
3213 * beyond what csum_start can cover.
3215 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
3216 skb_headroom(skb
) >= 0xFFFF)) {
3217 struct sk_buff
*nskb
;
3219 tcp_skb_tsorted_save(skb
) {
3220 nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
, GFP_ATOMIC
);
3223 err
= tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
);
3227 } tcp_skb_tsorted_restore(skb
);
3230 tcp_update_skb_after_send(sk
, skb
, tp
->tcp_wstamp_ns
);
3231 tcp_rate_skb_sent(sk
, skb
);
3234 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3237 /* To avoid taking spuriously low RTT samples based on a timestamp
3238 * for a transmit that never happened, always mark EVER_RETRANS
3240 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
3242 if (BPF_SOCK_OPS_TEST_FLAG(tp
, BPF_SOCK_OPS_RETRANS_CB_FLAG
))
3243 tcp_call_bpf_3arg(sk
, BPF_SOCK_OPS_RETRANS_CB
,
3244 TCP_SKB_CB(skb
)->seq
, segs
, err
);
3247 trace_tcp_retransmit_skb(sk
, skb
);
3248 } else if (err
!= -EBUSY
) {
3249 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
, segs
);
3254 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
3256 struct tcp_sock
*tp
= tcp_sk(sk
);
3257 int err
= __tcp_retransmit_skb(sk
, skb
, segs
);
3260 #if FASTRETRANS_DEBUG > 0
3261 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
3262 net_dbg_ratelimited("retrans_out leaked\n");
3265 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
3266 tp
->retrans_out
+= tcp_skb_pcount(skb
);
3269 /* Save stamp of the first (attempted) retransmit. */
3270 if (!tp
->retrans_stamp
)
3271 tp
->retrans_stamp
= tcp_skb_timestamp(skb
);
3273 if (tp
->undo_retrans
< 0)
3274 tp
->undo_retrans
= 0;
3275 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
3279 /* This gets called after a retransmit timeout, and the initially
3280 * retransmitted data is acknowledged. It tries to continue
3281 * resending the rest of the retransmit queue, until either
3282 * we've sent it all or the congestion window limit is reached.
3284 void tcp_xmit_retransmit_queue(struct sock
*sk
)
3286 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3287 struct sk_buff
*skb
, *rtx_head
, *hole
= NULL
;
3288 struct tcp_sock
*tp
= tcp_sk(sk
);
3289 bool rearm_timer
= false;
3293 if (!tp
->packets_out
)
3296 rtx_head
= tcp_rtx_queue_head(sk
);
3297 skb
= tp
->retransmit_skb_hint
?: rtx_head
;
3298 max_segs
= tcp_tso_segs(sk
, tcp_current_mss(sk
));
3299 skb_rbtree_walk_from(skb
) {
3303 if (tcp_pacing_check(sk
))
3306 /* we could do better than to assign each time */
3308 tp
->retransmit_skb_hint
= skb
;
3310 segs
= tp
->snd_cwnd
- tcp_packets_in_flight(tp
);
3313 sacked
= TCP_SKB_CB(skb
)->sacked
;
3314 /* In case tcp_shift_skb_data() have aggregated large skbs,
3315 * we need to make sure not sending too bigs TSO packets
3317 segs
= min_t(int, segs
, max_segs
);
3319 if (tp
->retrans_out
>= tp
->lost_out
) {
3321 } else if (!(sacked
& TCPCB_LOST
)) {
3322 if (!hole
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
3327 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
3328 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
3330 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
3333 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
3336 if (tcp_small_queue_check(sk
, skb
, 1))
3339 if (tcp_retransmit_skb(sk
, skb
, segs
))
3342 NET_ADD_STATS(sock_net(sk
), mib_idx
, tcp_skb_pcount(skb
));
3344 if (tcp_in_cwnd_reduction(sk
))
3345 tp
->prr_out
+= tcp_skb_pcount(skb
);
3347 if (skb
== rtx_head
&&
3348 icsk
->icsk_pending
!= ICSK_TIME_REO_TIMEOUT
)
3353 tcp_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3354 inet_csk(sk
)->icsk_rto
,
3358 /* We allow to exceed memory limits for FIN packets to expedite
3359 * connection tear down and (memory) recovery.
3360 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3361 * or even be forced to close flow without any FIN.
3362 * In general, we want to allow one skb per socket to avoid hangs
3363 * with edge trigger epoll()
3365 void sk_forced_mem_schedule(struct sock
*sk
, int size
)
3369 if (size
<= sk
->sk_forward_alloc
)
3371 amt
= sk_mem_pages(size
);
3372 sk
->sk_forward_alloc
+= amt
* SK_MEM_QUANTUM
;
3373 sk_memory_allocated_add(sk
, amt
);
3375 if (mem_cgroup_sockets_enabled
&& sk
->sk_memcg
)
3376 mem_cgroup_charge_skmem(sk
->sk_memcg
, amt
,
3377 gfp_memcg_charge() | __GFP_NOFAIL
);
3380 /* Send a FIN. The caller locks the socket for us.
3381 * We should try to send a FIN packet really hard, but eventually give up.
3383 void tcp_send_fin(struct sock
*sk
)
3385 struct sk_buff
*skb
, *tskb
, *tail
= tcp_write_queue_tail(sk
);
3386 struct tcp_sock
*tp
= tcp_sk(sk
);
3388 /* Optimization, tack on the FIN if we have one skb in write queue and
3389 * this skb was not yet sent, or we are under memory pressure.
3390 * Note: in the latter case, FIN packet will be sent after a timeout,
3391 * as TCP stack thinks it has already been transmitted.
3394 if (!tskb
&& tcp_under_memory_pressure(sk
))
3395 tskb
= skb_rb_last(&sk
->tcp_rtx_queue
);
3398 TCP_SKB_CB(tskb
)->tcp_flags
|= TCPHDR_FIN
;
3399 TCP_SKB_CB(tskb
)->end_seq
++;
3402 /* This means tskb was already sent.
3403 * Pretend we included the FIN on previous transmit.
3404 * We need to set tp->snd_nxt to the value it would have
3405 * if FIN had been sent. This is because retransmit path
3406 * does not change tp->snd_nxt.
3408 WRITE_ONCE(tp
->snd_nxt
, tp
->snd_nxt
+ 1);
3412 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, sk
->sk_allocation
);
3416 INIT_LIST_HEAD(&skb
->tcp_tsorted_anchor
);
3417 skb_reserve(skb
, MAX_TCP_HEADER
);
3418 sk_forced_mem_schedule(sk
, skb
->truesize
);
3419 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3420 tcp_init_nondata_skb(skb
, tp
->write_seq
,
3421 TCPHDR_ACK
| TCPHDR_FIN
);
3422 tcp_queue_skb(sk
, skb
);
3424 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
), TCP_NAGLE_OFF
);
3427 /* We get here when a process closes a file descriptor (either due to
3428 * an explicit close() or as a byproduct of exit()'ing) and there
3429 * was unread data in the receive queue. This behavior is recommended
3430 * by RFC 2525, section 2.17. -DaveM
3432 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
3434 struct sk_buff
*skb
;
3436 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
3438 /* NOTE: No TCP options attached and we never retransmit this. */
3439 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
3441 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3445 /* Reserve space for headers and prepare control bits. */
3446 skb_reserve(skb
, MAX_TCP_HEADER
);
3447 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
3448 TCPHDR_ACK
| TCPHDR_RST
);
3449 tcp_mstamp_refresh(tcp_sk(sk
));
3451 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
3452 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3454 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3455 * skb here is different to the troublesome skb, so use NULL
3457 trace_tcp_send_reset(sk
, NULL
);
3460 /* Send a crossed SYN-ACK during socket establishment.
3461 * WARNING: This routine must only be called when we have already sent
3462 * a SYN packet that crossed the incoming SYN that caused this routine
3463 * to get called. If this assumption fails then the initial rcv_wnd
3464 * and rcv_wscale values will not be correct.
3466 int tcp_send_synack(struct sock
*sk
)
3468 struct sk_buff
*skb
;
3470 skb
= tcp_rtx_queue_head(sk
);
3471 if (!skb
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
3472 pr_err("%s: wrong queue state\n", __func__
);
3475 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
3476 if (skb_cloned(skb
)) {
3477 struct sk_buff
*nskb
;
3479 tcp_skb_tsorted_save(skb
) {
3480 nskb
= skb_copy(skb
, GFP_ATOMIC
);
3481 } tcp_skb_tsorted_restore(skb
);
3484 INIT_LIST_HEAD(&nskb
->tcp_tsorted_anchor
);
3485 tcp_highest_sack_replace(sk
, skb
, nskb
);
3486 tcp_rtx_queue_unlink_and_free(skb
, sk
);
3487 __skb_header_release(nskb
);
3488 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, nskb
);
3489 sk_wmem_queued_add(sk
, nskb
->truesize
);
3490 sk_mem_charge(sk
, nskb
->truesize
);
3494 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
3495 tcp_ecn_send_synack(sk
, skb
);
3497 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3501 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3502 * @sk: listener socket
3503 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3504 * should not use it again.
3505 * @req: request_sock pointer
3506 * @foc: cookie for tcp fast open
3507 * @synack_type: Type of synack to prepare
3508 * @syn_skb: SYN packet just received. It could be NULL for rtx case.
3510 struct sk_buff
*tcp_make_synack(const struct sock
*sk
, struct dst_entry
*dst
,
3511 struct request_sock
*req
,
3512 struct tcp_fastopen_cookie
*foc
,
3513 enum tcp_synack_type synack_type
,
3514 struct sk_buff
*syn_skb
)
3516 struct inet_request_sock
*ireq
= inet_rsk(req
);
3517 const struct tcp_sock
*tp
= tcp_sk(sk
);
3518 struct tcp_md5sig_key
*md5
= NULL
;
3519 struct tcp_out_options opts
;
3520 struct sk_buff
*skb
;
3521 int tcp_header_size
;
3526 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
3527 if (unlikely(!skb
)) {
3531 /* Reserve space for headers. */
3532 skb_reserve(skb
, MAX_TCP_HEADER
);
3534 switch (synack_type
) {
3535 case TCP_SYNACK_NORMAL
:
3536 skb_set_owner_w(skb
, req_to_sk(req
));
3538 case TCP_SYNACK_COOKIE
:
3539 /* Under synflood, we do not attach skb to a socket,
3540 * to avoid false sharing.
3543 case TCP_SYNACK_FASTOPEN
:
3544 /* sk is a const pointer, because we want to express multiple
3545 * cpu might call us concurrently.
3546 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3548 skb_set_owner_w(skb
, (struct sock
*)sk
);
3551 skb_dst_set(skb
, dst
);
3553 mss
= tcp_mss_clamp(tp
, dst_metric_advmss(dst
));
3555 memset(&opts
, 0, sizeof(opts
));
3556 now
= tcp_clock_ns();
3557 #ifdef CONFIG_SYN_COOKIES
3558 if (unlikely(synack_type
== TCP_SYNACK_COOKIE
&& ireq
->tstamp_ok
))
3559 skb
->skb_mstamp_ns
= cookie_init_timestamp(req
, now
);
3563 skb
->skb_mstamp_ns
= now
;
3564 if (!tcp_rsk(req
)->snt_synack
) /* Timestamp first SYNACK */
3565 tcp_rsk(req
)->snt_synack
= tcp_skb_timestamp_us(skb
);
3568 #ifdef CONFIG_TCP_MD5SIG
3570 md5
= tcp_rsk(req
)->af_specific
->req_md5_lookup(sk
, req_to_sk(req
));
3572 skb_set_hash(skb
, tcp_rsk(req
)->txhash
, PKT_HASH_TYPE_L4
);
3573 /* bpf program will be interested in the tcp_flags */
3574 TCP_SKB_CB(skb
)->tcp_flags
= TCPHDR_SYN
| TCPHDR_ACK
;
3575 tcp_header_size
= tcp_synack_options(sk
, req
, mss
, skb
, &opts
, md5
,
3577 syn_skb
) + sizeof(*th
);
3579 skb_push(skb
, tcp_header_size
);
3580 skb_reset_transport_header(skb
);
3582 th
= (struct tcphdr
*)skb
->data
;
3583 memset(th
, 0, sizeof(struct tcphdr
));
3586 tcp_ecn_make_synack(req
, th
);
3587 th
->source
= htons(ireq
->ir_num
);
3588 th
->dest
= ireq
->ir_rmt_port
;
3589 skb
->mark
= ireq
->ir_mark
;
3590 skb
->ip_summed
= CHECKSUM_PARTIAL
;
3591 th
->seq
= htonl(tcp_rsk(req
)->snt_isn
);
3592 /* XXX data is queued and acked as is. No buffer/window check */
3593 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
3595 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3596 th
->window
= htons(min(req
->rsk_rcv_wnd
, 65535U));
3597 tcp_options_write((__be32
*)(th
+ 1), NULL
, &opts
);
3598 th
->doff
= (tcp_header_size
>> 2);
3599 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
3601 #ifdef CONFIG_TCP_MD5SIG
3602 /* Okay, we have all we need - do the md5 hash if needed */
3604 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
3605 md5
, req_to_sk(req
), skb
);
3609 bpf_skops_write_hdr_opt((struct sock
*)sk
, skb
, req
, syn_skb
,
3610 synack_type
, &opts
);
3612 skb
->skb_mstamp_ns
= now
;
3613 tcp_add_tx_delay(skb
, tp
);
3617 EXPORT_SYMBOL(tcp_make_synack
);
3619 static void tcp_ca_dst_init(struct sock
*sk
, const struct dst_entry
*dst
)
3621 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3622 const struct tcp_congestion_ops
*ca
;
3623 u32 ca_key
= dst_metric(dst
, RTAX_CC_ALGO
);
3625 if (ca_key
== TCP_CA_UNSPEC
)
3629 ca
= tcp_ca_find_key(ca_key
);
3630 if (likely(ca
&& bpf_try_module_get(ca
, ca
->owner
))) {
3631 bpf_module_put(icsk
->icsk_ca_ops
, icsk
->icsk_ca_ops
->owner
);
3632 icsk
->icsk_ca_dst_locked
= tcp_ca_dst_locked(dst
);
3633 icsk
->icsk_ca_ops
= ca
;
3638 /* Do all connect socket setups that can be done AF independent. */
3639 static void tcp_connect_init(struct sock
*sk
)
3641 const struct dst_entry
*dst
= __sk_dst_get(sk
);
3642 struct tcp_sock
*tp
= tcp_sk(sk
);
3646 /* We'll fix this up when we get a response from the other end.
3647 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3649 tp
->tcp_header_len
= sizeof(struct tcphdr
);
3650 if (sock_net(sk
)->ipv4
.sysctl_tcp_timestamps
)
3651 tp
->tcp_header_len
+= TCPOLEN_TSTAMP_ALIGNED
;
3653 #ifdef CONFIG_TCP_MD5SIG
3654 if (tp
->af_specific
->md5_lookup(sk
, sk
))
3655 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
3658 /* If user gave his TCP_MAXSEG, record it to clamp */
3659 if (tp
->rx_opt
.user_mss
)
3660 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3663 tcp_sync_mss(sk
, dst_mtu(dst
));
3665 tcp_ca_dst_init(sk
, dst
);
3667 if (!tp
->window_clamp
)
3668 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
3669 tp
->advmss
= tcp_mss_clamp(tp
, dst_metric_advmss(dst
));
3671 tcp_initialize_rcv_mss(sk
);
3673 /* limit the window selection if the user enforce a smaller rx buffer */
3674 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
3675 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
3676 tp
->window_clamp
= tcp_full_space(sk
);
3678 rcv_wnd
= tcp_rwnd_init_bpf(sk
);
3680 rcv_wnd
= dst_metric(dst
, RTAX_INITRWND
);
3682 tcp_select_initial_window(sk
, tcp_full_space(sk
),
3683 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
3686 sock_net(sk
)->ipv4
.sysctl_tcp_window_scaling
,
3690 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
3691 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
3694 sock_reset_flag(sk
, SOCK_DONE
);
3697 tcp_write_queue_purge(sk
);
3698 tp
->snd_una
= tp
->write_seq
;
3699 tp
->snd_sml
= tp
->write_seq
;
3700 tp
->snd_up
= tp
->write_seq
;
3701 WRITE_ONCE(tp
->snd_nxt
, tp
->write_seq
);
3703 if (likely(!tp
->repair
))
3706 tp
->rcv_tstamp
= tcp_jiffies32
;
3707 tp
->rcv_wup
= tp
->rcv_nxt
;
3708 WRITE_ONCE(tp
->copied_seq
, tp
->rcv_nxt
);
3710 inet_csk(sk
)->icsk_rto
= tcp_timeout_init(sk
);
3711 inet_csk(sk
)->icsk_retransmits
= 0;
3712 tcp_clear_retrans(tp
);
3715 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
3717 struct tcp_sock
*tp
= tcp_sk(sk
);
3718 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
3720 tcb
->end_seq
+= skb
->len
;
3721 __skb_header_release(skb
);
3722 sk_wmem_queued_add(sk
, skb
->truesize
);
3723 sk_mem_charge(sk
, skb
->truesize
);
3724 WRITE_ONCE(tp
->write_seq
, tcb
->end_seq
);
3725 tp
->packets_out
+= tcp_skb_pcount(skb
);
3728 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3729 * queue a data-only packet after the regular SYN, such that regular SYNs
3730 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3731 * only the SYN sequence, the data are retransmitted in the first ACK.
3732 * If cookie is not cached or other error occurs, falls back to send a
3733 * regular SYN with Fast Open cookie request option.
3735 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
3737 struct tcp_sock
*tp
= tcp_sk(sk
);
3738 struct tcp_fastopen_request
*fo
= tp
->fastopen_req
;
3740 struct sk_buff
*syn_data
;
3742 tp
->rx_opt
.mss_clamp
= tp
->advmss
; /* If MSS is not cached */
3743 if (!tcp_fastopen_cookie_check(sk
, &tp
->rx_opt
.mss_clamp
, &fo
->cookie
))
3746 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3747 * user-MSS. Reserve maximum option space for middleboxes that add
3748 * private TCP options. The cost is reduced data space in SYN :(
3750 tp
->rx_opt
.mss_clamp
= tcp_mss_clamp(tp
, tp
->rx_opt
.mss_clamp
);
3752 space
= __tcp_mtu_to_mss(sk
, inet_csk(sk
)->icsk_pmtu_cookie
) -
3753 MAX_TCP_OPTION_SPACE
;
3755 space
= min_t(size_t, space
, fo
->size
);
3757 /* limit to order-0 allocations */
3758 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
3760 syn_data
= sk_stream_alloc_skb(sk
, space
, sk
->sk_allocation
, false);
3763 syn_data
->ip_summed
= CHECKSUM_PARTIAL
;
3764 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
3766 int copied
= copy_from_iter(skb_put(syn_data
, space
), space
,
3767 &fo
->data
->msg_iter
);
3768 if (unlikely(!copied
)) {
3769 tcp_skb_tsorted_anchor_cleanup(syn_data
);
3770 kfree_skb(syn_data
);
3773 if (copied
!= space
) {
3774 skb_trim(syn_data
, copied
);
3777 skb_zcopy_set(syn_data
, fo
->uarg
, NULL
);
3779 /* No more data pending in inet_wait_for_connect() */
3780 if (space
== fo
->size
)
3784 tcp_connect_queue_skb(sk
, syn_data
);
3786 tcp_chrono_start(sk
, TCP_CHRONO_BUSY
);
3788 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
3790 syn
->skb_mstamp_ns
= syn_data
->skb_mstamp_ns
;
3792 /* Now full SYN+DATA was cloned and sent (or not),
3793 * remove the SYN from the original skb (syn_data)
3794 * we keep in write queue in case of a retransmit, as we
3795 * also have the SYN packet (with no data) in the same queue.
3797 TCP_SKB_CB(syn_data
)->seq
++;
3798 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
3800 tp
->syn_data
= (fo
->copied
> 0);
3801 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, syn_data
);
3802 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3806 /* data was not sent, put it in write_queue */
3807 __skb_queue_tail(&sk
->sk_write_queue
, syn_data
);
3808 tp
->packets_out
-= tcp_skb_pcount(syn_data
);
3811 /* Send a regular SYN with Fast Open cookie request option */
3812 if (fo
->cookie
.len
> 0)
3814 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3816 tp
->syn_fastopen
= 0;
3818 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3822 /* Build a SYN and send it off. */
3823 int tcp_connect(struct sock
*sk
)
3825 struct tcp_sock
*tp
= tcp_sk(sk
);
3826 struct sk_buff
*buff
;
3829 tcp_call_bpf(sk
, BPF_SOCK_OPS_TCP_CONNECT_CB
, 0, NULL
);
3831 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
3832 return -EHOSTUNREACH
; /* Routing failure or similar. */
3834 tcp_connect_init(sk
);
3836 if (unlikely(tp
->repair
)) {
3837 tcp_finish_connect(sk
, NULL
);
3841 buff
= sk_stream_alloc_skb(sk
, 0, sk
->sk_allocation
, true);
3842 if (unlikely(!buff
))
3845 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3846 tcp_mstamp_refresh(tp
);
3847 tp
->retrans_stamp
= tcp_time_stamp(tp
);
3848 tcp_connect_queue_skb(sk
, buff
);
3849 tcp_ecn_send_syn(sk
, buff
);
3850 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, buff
);
3852 /* Send off SYN; include data in Fast Open. */
3853 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3854 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3855 if (err
== -ECONNREFUSED
)
3858 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3859 * in order to make this packet get counted in tcpOutSegs.
3861 WRITE_ONCE(tp
->snd_nxt
, tp
->write_seq
);
3862 tp
->pushed_seq
= tp
->write_seq
;
3863 buff
= tcp_send_head(sk
);
3864 if (unlikely(buff
)) {
3865 WRITE_ONCE(tp
->snd_nxt
, TCP_SKB_CB(buff
)->seq
);
3866 tp
->pushed_seq
= TCP_SKB_CB(buff
)->seq
;
3868 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3870 /* Timer for repeating the SYN until an answer. */
3871 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3872 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3875 EXPORT_SYMBOL(tcp_connect
);
3877 /* Send out a delayed ack, the caller does the policy checking
3878 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3881 void tcp_send_delayed_ack(struct sock
*sk
)
3883 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3884 int ato
= icsk
->icsk_ack
.ato
;
3885 unsigned long timeout
;
3887 if (ato
> TCP_DELACK_MIN
) {
3888 const struct tcp_sock
*tp
= tcp_sk(sk
);
3889 int max_ato
= HZ
/ 2;
3891 if (inet_csk_in_pingpong_mode(sk
) ||
3892 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3893 max_ato
= TCP_DELACK_MAX
;
3895 /* Slow path, intersegment interval is "high". */
3897 /* If some rtt estimate is known, use it to bound delayed ack.
3898 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3902 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3909 ato
= min(ato
, max_ato
);
3912 ato
= min_t(u32
, ato
, inet_csk(sk
)->icsk_delack_max
);
3914 /* Stay within the limit we were given */
3915 timeout
= jiffies
+ ato
;
3917 /* Use new timeout only if there wasn't a older one earlier. */
3918 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3919 /* If delack timer is about to expire, send ACK now. */
3920 if (time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3925 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3926 timeout
= icsk
->icsk_ack
.timeout
;
3928 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3929 icsk
->icsk_ack
.timeout
= timeout
;
3930 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3933 /* This routine sends an ack and also updates the window. */
3934 void __tcp_send_ack(struct sock
*sk
, u32 rcv_nxt
)
3936 struct sk_buff
*buff
;
3938 /* If we have been reset, we may not send again. */
3939 if (sk
->sk_state
== TCP_CLOSE
)
3942 /* We are not putting this on the write queue, so
3943 * tcp_transmit_skb() will set the ownership to this
3946 buff
= alloc_skb(MAX_TCP_HEADER
,
3947 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3948 if (unlikely(!buff
)) {
3949 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3950 unsigned long delay
;
3952 delay
= TCP_DELACK_MAX
<< icsk
->icsk_ack
.retry
;
3953 if (delay
< TCP_RTO_MAX
)
3954 icsk
->icsk_ack
.retry
++;
3955 inet_csk_schedule_ack(sk
);
3956 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
3957 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
, delay
, TCP_RTO_MAX
);
3961 /* Reserve space for headers and prepare control bits. */
3962 skb_reserve(buff
, MAX_TCP_HEADER
);
3963 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3965 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3967 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3969 skb_set_tcp_pure_ack(buff
);
3971 /* Send it off, this clears delayed acks for us. */
3972 __tcp_transmit_skb(sk
, buff
, 0, (__force gfp_t
)0, rcv_nxt
);
3974 EXPORT_SYMBOL_GPL(__tcp_send_ack
);
3976 void tcp_send_ack(struct sock
*sk
)
3978 __tcp_send_ack(sk
, tcp_sk(sk
)->rcv_nxt
);
3981 /* This routine sends a packet with an out of date sequence
3982 * number. It assumes the other end will try to ack it.
3984 * Question: what should we make while urgent mode?
3985 * 4.4BSD forces sending single byte of data. We cannot send
3986 * out of window data, because we have SND.NXT==SND.MAX...
3988 * Current solution: to send TWO zero-length segments in urgent mode:
3989 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3990 * out-of-date with SND.UNA-1 to probe window.
3992 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
, int mib
)
3994 struct tcp_sock
*tp
= tcp_sk(sk
);
3995 struct sk_buff
*skb
;
3997 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3998 skb
= alloc_skb(MAX_TCP_HEADER
,
3999 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
4003 /* Reserve space for headers and set control bits. */
4004 skb_reserve(skb
, MAX_TCP_HEADER
);
4005 /* Use a previous sequence. This should cause the other
4006 * end to send an ack. Don't queue or clone SKB, just
4009 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
4010 NET_INC_STATS(sock_net(sk
), mib
);
4011 return tcp_transmit_skb(sk
, skb
, 0, (__force gfp_t
)0);
4014 /* Called from setsockopt( ... TCP_REPAIR ) */
4015 void tcp_send_window_probe(struct sock
*sk
)
4017 if (sk
->sk_state
== TCP_ESTABLISHED
) {
4018 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
4019 tcp_mstamp_refresh(tcp_sk(sk
));
4020 tcp_xmit_probe_skb(sk
, 0, LINUX_MIB_TCPWINPROBE
);
4024 /* Initiate keepalive or window probe from timer. */
4025 int tcp_write_wakeup(struct sock
*sk
, int mib
)
4027 struct tcp_sock
*tp
= tcp_sk(sk
);
4028 struct sk_buff
*skb
;
4030 if (sk
->sk_state
== TCP_CLOSE
)
4033 skb
= tcp_send_head(sk
);
4034 if (skb
&& before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
4036 unsigned int mss
= tcp_current_mss(sk
);
4037 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
4039 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
4040 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
4042 /* We are probing the opening of a window
4043 * but the window size is != 0
4044 * must have been a result SWS avoidance ( sender )
4046 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
4048 seg_size
= min(seg_size
, mss
);
4049 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
4050 if (tcp_fragment(sk
, TCP_FRAG_IN_WRITE_QUEUE
,
4051 skb
, seg_size
, mss
, GFP_ATOMIC
))
4053 } else if (!tcp_skb_pcount(skb
))
4054 tcp_set_skb_tso_segs(skb
, mss
);
4056 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
4057 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
4059 tcp_event_new_data_sent(sk
, skb
);
4062 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
4063 tcp_xmit_probe_skb(sk
, 1, mib
);
4064 return tcp_xmit_probe_skb(sk
, 0, mib
);
4068 /* A window probe timeout has occurred. If window is not closed send
4069 * a partial packet else a zero probe.
4071 void tcp_send_probe0(struct sock
*sk
)
4073 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4074 struct tcp_sock
*tp
= tcp_sk(sk
);
4075 struct net
*net
= sock_net(sk
);
4076 unsigned long timeout
;
4079 err
= tcp_write_wakeup(sk
, LINUX_MIB_TCPWINPROBE
);
4081 if (tp
->packets_out
|| tcp_write_queue_empty(sk
)) {
4082 /* Cancel probe timer, if it is not required. */
4083 icsk
->icsk_probes_out
= 0;
4084 icsk
->icsk_backoff
= 0;
4085 icsk
->icsk_probes_tstamp
= 0;
4089 icsk
->icsk_probes_out
++;
4091 if (icsk
->icsk_backoff
< net
->ipv4
.sysctl_tcp_retries2
)
4092 icsk
->icsk_backoff
++;
4093 timeout
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
4095 /* If packet was not sent due to local congestion,
4096 * Let senders fight for local resources conservatively.
4098 timeout
= TCP_RESOURCE_PROBE_INTERVAL
;
4101 timeout
= tcp_clamp_probe0_to_user_timeout(sk
, timeout
);
4102 tcp_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
, timeout
, TCP_RTO_MAX
);
4105 int tcp_rtx_synack(const struct sock
*sk
, struct request_sock
*req
)
4107 const struct tcp_request_sock_ops
*af_ops
= tcp_rsk(req
)->af_specific
;
4111 tcp_rsk(req
)->txhash
= net_tx_rndhash();
4112 res
= af_ops
->send_synack(sk
, NULL
, &fl
, req
, NULL
, TCP_SYNACK_NORMAL
,
4115 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
4116 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
4117 if (unlikely(tcp_passive_fastopen(sk
)))
4118 tcp_sk(sk
)->total_retrans
++;
4119 trace_tcp_retransmit_synack(sk
, req
);
4123 EXPORT_SYMBOL(tcp_rtx_synack
);