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
,
88 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
89 * window scaling factor due to loss of precision.
90 * If window has been shrunk, what should we make? It is not clear at all.
91 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
92 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
93 * invalid. OK, let's make this for now:
95 static inline __u32
tcp_acceptable_seq(const struct sock
*sk
)
97 const struct tcp_sock
*tp
= tcp_sk(sk
);
99 if (!before(tcp_wnd_end(tp
), tp
->snd_nxt
) ||
100 (tp
->rx_opt
.wscale_ok
&&
101 ((tp
->snd_nxt
- tcp_wnd_end(tp
)) < (1 << tp
->rx_opt
.rcv_wscale
))))
104 return tcp_wnd_end(tp
);
107 /* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
121 static __u16
tcp_advertise_mss(struct sock
*sk
)
123 struct tcp_sock
*tp
= tcp_sk(sk
);
124 const struct dst_entry
*dst
= __sk_dst_get(sk
);
125 int mss
= tp
->advmss
;
128 unsigned int metric
= dst_metric_advmss(dst
);
139 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism.
142 void tcp_cwnd_restart(struct sock
*sk
, s32 delta
)
144 struct tcp_sock
*tp
= tcp_sk(sk
);
145 u32 restart_cwnd
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
146 u32 cwnd
= tcp_snd_cwnd(tp
);
148 tcp_ca_event(sk
, CA_EVENT_CWND_RESTART
);
150 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
151 restart_cwnd
= min(restart_cwnd
, cwnd
);
153 while ((delta
-= inet_csk(sk
)->icsk_rto
) > 0 && cwnd
> restart_cwnd
)
155 tcp_snd_cwnd_set(tp
, max(cwnd
, restart_cwnd
));
156 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
157 tp
->snd_cwnd_used
= 0;
160 /* Congestion state accounting after a packet has been sent. */
161 static void tcp_event_data_sent(struct tcp_sock
*tp
,
164 struct inet_connection_sock
*icsk
= inet_csk(sk
);
165 const u32 now
= tcp_jiffies32
;
167 if (tcp_packets_in_flight(tp
) == 0)
168 tcp_ca_event(sk
, CA_EVENT_TX_START
);
170 /* If this is the first data packet sent in response to the
171 * previous received data,
172 * and it is a reply for ato after last received packet,
173 * increase pingpong count.
175 if (before(tp
->lsndtime
, icsk
->icsk_ack
.lrcvtime
) &&
176 (u32
)(now
- icsk
->icsk_ack
.lrcvtime
) < icsk
->icsk_ack
.ato
)
177 inet_csk_inc_pingpong_cnt(sk
);
182 /* Account for an ACK we sent. */
183 static inline void tcp_event_ack_sent(struct sock
*sk
, unsigned int pkts
,
186 struct tcp_sock
*tp
= tcp_sk(sk
);
188 if (unlikely(tp
->compressed_ack
)) {
189 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPACKCOMPRESSED
,
191 tp
->compressed_ack
= 0;
192 if (hrtimer_try_to_cancel(&tp
->compressed_ack_timer
) == 1)
196 if (unlikely(rcv_nxt
!= tp
->rcv_nxt
))
197 return; /* Special ACK sent by DCTCP to reflect ECN */
198 tcp_dec_quickack_mode(sk
, pkts
);
199 inet_csk_clear_xmit_timer(sk
, ICSK_TIME_DACK
);
202 /* Determine a window scaling and initial window to offer.
203 * Based on the assumption that the given amount of space
204 * will be offered. Store the results in the tp structure.
205 * NOTE: for smooth operation initial space offering should
206 * be a multiple of mss if possible. We assume here that mss >= 1.
207 * This MUST be enforced by all callers.
209 void tcp_select_initial_window(const struct sock
*sk
, int __space
, __u32 mss
,
210 __u32
*rcv_wnd
, __u32
*window_clamp
,
211 int wscale_ok
, __u8
*rcv_wscale
,
214 unsigned int space
= (__space
< 0 ? 0 : __space
);
216 /* If no clamp set the clamp to the max possible scaled window */
217 if (*window_clamp
== 0)
218 (*window_clamp
) = (U16_MAX
<< TCP_MAX_WSCALE
);
219 space
= min(*window_clamp
, space
);
221 /* Quantize space offering to a multiple of mss if possible. */
223 space
= rounddown(space
, mss
);
225 /* NOTE: offering an initial window larger than 32767
226 * will break some buggy TCP stacks. If the admin tells us
227 * it is likely we could be speaking with such a buggy stack
228 * we will truncate our initial window offering to 32K-1
229 * unless the remote has sent us a window scaling option,
230 * which we interpret as a sign the remote TCP is not
231 * misinterpreting the window field as a signed quantity.
233 if (sock_net(sk
)->ipv4
.sysctl_tcp_workaround_signed_windows
)
234 (*rcv_wnd
) = min(space
, MAX_TCP_WINDOW
);
236 (*rcv_wnd
) = min_t(u32
, space
, U16_MAX
);
239 *rcv_wnd
= min(*rcv_wnd
, init_rcv_wnd
* mss
);
243 /* Set window scaling on max possible window */
244 space
= max_t(u32
, space
, sock_net(sk
)->ipv4
.sysctl_tcp_rmem
[2]);
245 space
= max_t(u32
, space
, sysctl_rmem_max
);
246 space
= min_t(u32
, space
, *window_clamp
);
247 *rcv_wscale
= clamp_t(int, ilog2(space
) - 15,
250 /* Set the clamp no higher than max representable value */
251 (*window_clamp
) = min_t(__u32
, U16_MAX
<< (*rcv_wscale
), *window_clamp
);
253 EXPORT_SYMBOL(tcp_select_initial_window
);
255 /* Chose a new window to advertise, update state in tcp_sock for the
256 * socket, and return result with RFC1323 scaling applied. The return
257 * value can be stuffed directly into th->window for an outgoing
260 static u16
tcp_select_window(struct sock
*sk
)
262 struct tcp_sock
*tp
= tcp_sk(sk
);
263 u32 old_win
= tp
->rcv_wnd
;
264 u32 cur_win
= tcp_receive_window(tp
);
265 u32 new_win
= __tcp_select_window(sk
);
267 /* Never shrink the offered window */
268 if (new_win
< cur_win
) {
269 /* Danger Will Robinson!
270 * Don't update rcv_wup/rcv_wnd here or else
271 * we will not be able to advertise a zero
272 * window in time. --DaveM
274 * Relax Will Robinson.
277 NET_INC_STATS(sock_net(sk
),
278 LINUX_MIB_TCPWANTZEROWINDOWADV
);
279 new_win
= ALIGN(cur_win
, 1 << tp
->rx_opt
.rcv_wscale
);
281 tp
->rcv_wnd
= new_win
;
282 tp
->rcv_wup
= tp
->rcv_nxt
;
284 /* Make sure we do not exceed the maximum possible
287 if (!tp
->rx_opt
.rcv_wscale
&&
288 sock_net(sk
)->ipv4
.sysctl_tcp_workaround_signed_windows
)
289 new_win
= min(new_win
, MAX_TCP_WINDOW
);
291 new_win
= min(new_win
, (65535U << tp
->rx_opt
.rcv_wscale
));
293 /* RFC1323 scaling applied */
294 new_win
>>= tp
->rx_opt
.rcv_wscale
;
296 /* If we advertise zero window, disable fast path. */
300 NET_INC_STATS(sock_net(sk
),
301 LINUX_MIB_TCPTOZEROWINDOWADV
);
302 } else if (old_win
== 0) {
303 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPFROMZEROWINDOWADV
);
309 /* Packet ECN state for a SYN-ACK */
310 static void tcp_ecn_send_synack(struct sock
*sk
, struct sk_buff
*skb
)
312 const struct tcp_sock
*tp
= tcp_sk(sk
);
314 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_CWR
;
315 if (!(tp
->ecn_flags
& TCP_ECN_OK
))
316 TCP_SKB_CB(skb
)->tcp_flags
&= ~TCPHDR_ECE
;
317 else if (tcp_ca_needs_ecn(sk
) ||
318 tcp_bpf_ca_needs_ecn(sk
))
322 /* Packet ECN state for a SYN. */
323 static void tcp_ecn_send_syn(struct sock
*sk
, struct sk_buff
*skb
)
325 struct tcp_sock
*tp
= tcp_sk(sk
);
326 bool bpf_needs_ecn
= tcp_bpf_ca_needs_ecn(sk
);
327 bool use_ecn
= sock_net(sk
)->ipv4
.sysctl_tcp_ecn
== 1 ||
328 tcp_ca_needs_ecn(sk
) || bpf_needs_ecn
;
331 const struct dst_entry
*dst
= __sk_dst_get(sk
);
333 if (dst
&& dst_feature(dst
, RTAX_FEATURE_ECN
))
340 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ECE
| TCPHDR_CWR
;
341 tp
->ecn_flags
= TCP_ECN_OK
;
342 if (tcp_ca_needs_ecn(sk
) || bpf_needs_ecn
)
347 static void tcp_ecn_clear_syn(struct sock
*sk
, struct sk_buff
*skb
)
349 if (sock_net(sk
)->ipv4
.sysctl_tcp_ecn_fallback
)
350 /* tp->ecn_flags are cleared at a later point in time when
351 * SYN ACK is ultimatively being received.
353 TCP_SKB_CB(skb
)->tcp_flags
&= ~(TCPHDR_ECE
| TCPHDR_CWR
);
357 tcp_ecn_make_synack(const struct request_sock
*req
, struct tcphdr
*th
)
359 if (inet_rsk(req
)->ecn_ok
)
363 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
366 static void tcp_ecn_send(struct sock
*sk
, struct sk_buff
*skb
,
367 struct tcphdr
*th
, int tcp_header_len
)
369 struct tcp_sock
*tp
= tcp_sk(sk
);
371 if (tp
->ecn_flags
& TCP_ECN_OK
) {
372 /* Not-retransmitted data segment: set ECT and inject CWR. */
373 if (skb
->len
!= tcp_header_len
&&
374 !before(TCP_SKB_CB(skb
)->seq
, tp
->snd_nxt
)) {
376 if (tp
->ecn_flags
& TCP_ECN_QUEUE_CWR
) {
377 tp
->ecn_flags
&= ~TCP_ECN_QUEUE_CWR
;
379 skb_shinfo(skb
)->gso_type
|= SKB_GSO_TCP_ECN
;
381 } else if (!tcp_ca_needs_ecn(sk
)) {
382 /* ACK or retransmitted segment: clear ECT|CE */
383 INET_ECN_dontxmit(sk
);
385 if (tp
->ecn_flags
& TCP_ECN_DEMAND_CWR
)
390 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
391 * auto increment end seqno.
393 static void tcp_init_nondata_skb(struct sk_buff
*skb
, u32 seq
, u8 flags
)
395 skb
->ip_summed
= CHECKSUM_PARTIAL
;
397 TCP_SKB_CB(skb
)->tcp_flags
= flags
;
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 BIT(0)
413 #define OPTION_TS BIT(1)
414 #define OPTION_MD5 BIT(2)
415 #define OPTION_WSCALE BIT(3)
416 #define OPTION_FAST_OPEN_COOKIE BIT(8)
417 #define OPTION_SMC BIT(9)
418 #define OPTION_MPTCP BIT(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(struct tcphdr
*th
, __be32
*ptr
,
450 struct tcp_out_options
*opts
)
452 #if IS_ENABLED(CONFIG_MPTCP)
453 if (unlikely(OPTION_MPTCP
& opts
->options
))
454 mptcp_write_options(th
, ptr
, tp
, &opts
->mptcp
);
458 #ifdef CONFIG_CGROUP_BPF
459 static int bpf_skops_write_hdr_opt_arg0(struct sk_buff
*skb
,
460 enum tcp_synack_type synack_type
)
463 return BPF_WRITE_HDR_TCP_CURRENT_MSS
;
465 if (unlikely(synack_type
== TCP_SYNACK_COOKIE
))
466 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE
;
471 /* req, syn_skb and synack_type are used when writing synack */
472 static void bpf_skops_hdr_opt_len(struct sock
*sk
, struct sk_buff
*skb
,
473 struct request_sock
*req
,
474 struct sk_buff
*syn_skb
,
475 enum tcp_synack_type synack_type
,
476 struct tcp_out_options
*opts
,
477 unsigned int *remaining
)
479 struct bpf_sock_ops_kern sock_ops
;
482 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk
),
483 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG
)) ||
487 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
490 memset(&sock_ops
, 0, offsetof(struct bpf_sock_ops_kern
, temp
));
492 sock_ops
.op
= BPF_SOCK_OPS_HDR_OPT_LEN_CB
;
495 /* The listen "sk" cannot be passed here because
496 * it is not locked. It would not make too much
497 * sense to do bpf_setsockopt(listen_sk) based
498 * on individual connection request also.
500 * Thus, "req" is passed here and the cgroup-bpf-progs
501 * of the listen "sk" will be run.
503 * "req" is also used here for fastopen even the "sk" here is
504 * a fullsock "child" sk. It is to keep the behavior
505 * consistent between fastopen and non-fastopen on
506 * the bpf programming side.
508 sock_ops
.sk
= (struct sock
*)req
;
509 sock_ops
.syn_skb
= syn_skb
;
511 sock_owned_by_me(sk
);
513 sock_ops
.is_fullsock
= 1;
517 sock_ops
.args
[0] = bpf_skops_write_hdr_opt_arg0(skb
, synack_type
);
518 sock_ops
.remaining_opt_len
= *remaining
;
519 /* tcp_current_mss() does not pass a skb */
521 bpf_skops_init_skb(&sock_ops
, skb
, 0);
523 err
= BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops
, sk
);
525 if (err
|| sock_ops
.remaining_opt_len
== *remaining
)
528 opts
->bpf_opt_len
= *remaining
- sock_ops
.remaining_opt_len
;
529 /* round up to 4 bytes */
530 opts
->bpf_opt_len
= (opts
->bpf_opt_len
+ 3) & ~3;
532 *remaining
-= opts
->bpf_opt_len
;
535 static void bpf_skops_write_hdr_opt(struct sock
*sk
, struct sk_buff
*skb
,
536 struct request_sock
*req
,
537 struct sk_buff
*syn_skb
,
538 enum tcp_synack_type synack_type
,
539 struct tcp_out_options
*opts
)
541 u8 first_opt_off
, nr_written
, max_opt_len
= opts
->bpf_opt_len
;
542 struct bpf_sock_ops_kern sock_ops
;
545 if (likely(!max_opt_len
))
548 memset(&sock_ops
, 0, offsetof(struct bpf_sock_ops_kern
, temp
));
550 sock_ops
.op
= BPF_SOCK_OPS_WRITE_HDR_OPT_CB
;
553 sock_ops
.sk
= (struct sock
*)req
;
554 sock_ops
.syn_skb
= syn_skb
;
556 sock_owned_by_me(sk
);
558 sock_ops
.is_fullsock
= 1;
562 sock_ops
.args
[0] = bpf_skops_write_hdr_opt_arg0(skb
, synack_type
);
563 sock_ops
.remaining_opt_len
= max_opt_len
;
564 first_opt_off
= tcp_hdrlen(skb
) - max_opt_len
;
565 bpf_skops_init_skb(&sock_ops
, skb
, first_opt_off
);
567 err
= BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops
, sk
);
572 nr_written
= max_opt_len
- sock_ops
.remaining_opt_len
;
574 if (nr_written
< max_opt_len
)
575 memset(skb
->data
+ first_opt_off
+ nr_written
, TCPOPT_NOP
,
576 max_opt_len
- nr_written
);
579 static void bpf_skops_hdr_opt_len(struct sock
*sk
, struct sk_buff
*skb
,
580 struct request_sock
*req
,
581 struct sk_buff
*syn_skb
,
582 enum tcp_synack_type synack_type
,
583 struct tcp_out_options
*opts
,
584 unsigned int *remaining
)
588 static void bpf_skops_write_hdr_opt(struct sock
*sk
, struct sk_buff
*skb
,
589 struct request_sock
*req
,
590 struct sk_buff
*syn_skb
,
591 enum tcp_synack_type synack_type
,
592 struct tcp_out_options
*opts
)
597 /* Write previously computed TCP options to the packet.
599 * Beware: Something in the Internet is very sensitive to the ordering of
600 * TCP options, we learned this through the hard way, so be careful here.
601 * Luckily we can at least blame others for their non-compliance but from
602 * inter-operability perspective it seems that we're somewhat stuck with
603 * the ordering which we have been using if we want to keep working with
604 * those broken things (not that it currently hurts anybody as there isn't
605 * particular reason why the ordering would need to be changed).
607 * At least SACK_PERM as the first option is known to lead to a disaster
608 * (but it may well be that other scenarios fail similarly).
610 static void tcp_options_write(struct tcphdr
*th
, struct tcp_sock
*tp
,
611 struct tcp_out_options
*opts
)
613 __be32
*ptr
= (__be32
*)(th
+ 1);
614 u16 options
= opts
->options
; /* mungable copy */
616 if (unlikely(OPTION_MD5
& options
)) {
617 *ptr
++ = htonl((TCPOPT_NOP
<< 24) | (TCPOPT_NOP
<< 16) |
618 (TCPOPT_MD5SIG
<< 8) | TCPOLEN_MD5SIG
);
619 /* overload cookie hash location */
620 opts
->hash_location
= (__u8
*)ptr
;
624 if (unlikely(opts
->mss
)) {
625 *ptr
++ = htonl((TCPOPT_MSS
<< 24) |
626 (TCPOLEN_MSS
<< 16) |
630 if (likely(OPTION_TS
& options
)) {
631 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
632 *ptr
++ = htonl((TCPOPT_SACK_PERM
<< 24) |
633 (TCPOLEN_SACK_PERM
<< 16) |
634 (TCPOPT_TIMESTAMP
<< 8) |
636 options
&= ~OPTION_SACK_ADVERTISE
;
638 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
640 (TCPOPT_TIMESTAMP
<< 8) |
643 *ptr
++ = htonl(opts
->tsval
);
644 *ptr
++ = htonl(opts
->tsecr
);
647 if (unlikely(OPTION_SACK_ADVERTISE
& options
)) {
648 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
650 (TCPOPT_SACK_PERM
<< 8) |
654 if (unlikely(OPTION_WSCALE
& options
)) {
655 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
656 (TCPOPT_WINDOW
<< 16) |
657 (TCPOLEN_WINDOW
<< 8) |
661 if (unlikely(opts
->num_sack_blocks
)) {
662 struct tcp_sack_block
*sp
= tp
->rx_opt
.dsack
?
663 tp
->duplicate_sack
: tp
->selective_acks
;
666 *ptr
++ = htonl((TCPOPT_NOP
<< 24) |
669 (TCPOLEN_SACK_BASE
+ (opts
->num_sack_blocks
*
670 TCPOLEN_SACK_PERBLOCK
)));
672 for (this_sack
= 0; this_sack
< opts
->num_sack_blocks
;
674 *ptr
++ = htonl(sp
[this_sack
].start_seq
);
675 *ptr
++ = htonl(sp
[this_sack
].end_seq
);
678 tp
->rx_opt
.dsack
= 0;
681 if (unlikely(OPTION_FAST_OPEN_COOKIE
& options
)) {
682 struct tcp_fastopen_cookie
*foc
= opts
->fastopen_cookie
;
684 u32 len
; /* Fast Open option length */
687 len
= TCPOLEN_EXP_FASTOPEN_BASE
+ foc
->len
;
688 *ptr
= htonl((TCPOPT_EXP
<< 24) | (len
<< 16) |
689 TCPOPT_FASTOPEN_MAGIC
);
690 p
+= TCPOLEN_EXP_FASTOPEN_BASE
;
692 len
= TCPOLEN_FASTOPEN_BASE
+ foc
->len
;
693 *p
++ = TCPOPT_FASTOPEN
;
697 memcpy(p
, foc
->val
, foc
->len
);
698 if ((len
& 3) == 2) {
699 p
[foc
->len
] = TCPOPT_NOP
;
700 p
[foc
->len
+ 1] = TCPOPT_NOP
;
702 ptr
+= (len
+ 3) >> 2;
705 smc_options_write(ptr
, &options
);
707 mptcp_options_write(th
, ptr
, tp
, opts
);
710 static void smc_set_option(const struct tcp_sock
*tp
,
711 struct tcp_out_options
*opts
,
712 unsigned int *remaining
)
714 #if IS_ENABLED(CONFIG_SMC)
715 if (static_branch_unlikely(&tcp_have_smc
)) {
717 if (*remaining
>= TCPOLEN_EXP_SMC_BASE_ALIGNED
) {
718 opts
->options
|= OPTION_SMC
;
719 *remaining
-= TCPOLEN_EXP_SMC_BASE_ALIGNED
;
726 static void smc_set_option_cond(const struct tcp_sock
*tp
,
727 const struct inet_request_sock
*ireq
,
728 struct tcp_out_options
*opts
,
729 unsigned int *remaining
)
731 #if IS_ENABLED(CONFIG_SMC)
732 if (static_branch_unlikely(&tcp_have_smc
)) {
733 if (tp
->syn_smc
&& ireq
->smc_ok
) {
734 if (*remaining
>= TCPOLEN_EXP_SMC_BASE_ALIGNED
) {
735 opts
->options
|= OPTION_SMC
;
736 *remaining
-= TCPOLEN_EXP_SMC_BASE_ALIGNED
;
743 static void mptcp_set_option_cond(const struct request_sock
*req
,
744 struct tcp_out_options
*opts
,
745 unsigned int *remaining
)
747 if (rsk_is_mptcp(req
)) {
750 if (mptcp_synack_options(req
, &size
, &opts
->mptcp
)) {
751 if (*remaining
>= size
) {
752 opts
->options
|= OPTION_MPTCP
;
759 /* Compute TCP options for SYN packets. This is not the final
760 * network wire format yet.
762 static unsigned int tcp_syn_options(struct sock
*sk
, struct sk_buff
*skb
,
763 struct tcp_out_options
*opts
,
764 struct tcp_md5sig_key
**md5
)
766 struct tcp_sock
*tp
= tcp_sk(sk
);
767 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
768 struct tcp_fastopen_request
*fastopen
= tp
->fastopen_req
;
771 #ifdef CONFIG_TCP_MD5SIG
772 if (static_branch_unlikely(&tcp_md5_needed
) &&
773 rcu_access_pointer(tp
->md5sig_info
)) {
774 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
776 opts
->options
|= OPTION_MD5
;
777 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
782 /* We always get an MSS option. The option bytes which will be seen in
783 * normal data packets should timestamps be used, must be in the MSS
784 * advertised. But we subtract them from tp->mss_cache so that
785 * calculations in tcp_sendmsg are simpler etc. So account for this
786 * fact here if necessary. If we don't do this correctly, as a
787 * receiver we won't recognize data packets as being full sized when we
788 * should, and thus we won't abide by the delayed ACK rules correctly.
789 * SACKs don't matter, we never delay an ACK when we have any of those
791 opts
->mss
= tcp_advertise_mss(sk
);
792 remaining
-= TCPOLEN_MSS_ALIGNED
;
794 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_timestamps
&& !*md5
)) {
795 opts
->options
|= OPTION_TS
;
796 opts
->tsval
= tcp_skb_timestamp(skb
) + tp
->tsoffset
;
797 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
798 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
800 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_window_scaling
)) {
801 opts
->ws
= tp
->rx_opt
.rcv_wscale
;
802 opts
->options
|= OPTION_WSCALE
;
803 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
805 if (likely(sock_net(sk
)->ipv4
.sysctl_tcp_sack
)) {
806 opts
->options
|= OPTION_SACK_ADVERTISE
;
807 if (unlikely(!(OPTION_TS
& opts
->options
)))
808 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
811 if (fastopen
&& fastopen
->cookie
.len
>= 0) {
812 u32 need
= fastopen
->cookie
.len
;
814 need
+= fastopen
->cookie
.exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
815 TCPOLEN_FASTOPEN_BASE
;
816 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
817 if (remaining
>= need
) {
818 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
819 opts
->fastopen_cookie
= &fastopen
->cookie
;
821 tp
->syn_fastopen
= 1;
822 tp
->syn_fastopen_exp
= fastopen
->cookie
.exp
? 1 : 0;
826 smc_set_option(tp
, opts
, &remaining
);
828 if (sk_is_mptcp(sk
)) {
831 if (mptcp_syn_options(sk
, skb
, &size
, &opts
->mptcp
)) {
832 opts
->options
|= OPTION_MPTCP
;
837 bpf_skops_hdr_opt_len(sk
, skb
, NULL
, NULL
, 0, opts
, &remaining
);
839 return MAX_TCP_OPTION_SPACE
- remaining
;
842 /* Set up TCP options for SYN-ACKs. */
843 static unsigned int tcp_synack_options(const struct sock
*sk
,
844 struct request_sock
*req
,
845 unsigned int mss
, struct sk_buff
*skb
,
846 struct tcp_out_options
*opts
,
847 const struct tcp_md5sig_key
*md5
,
848 struct tcp_fastopen_cookie
*foc
,
849 enum tcp_synack_type synack_type
,
850 struct sk_buff
*syn_skb
)
852 struct inet_request_sock
*ireq
= inet_rsk(req
);
853 unsigned int remaining
= MAX_TCP_OPTION_SPACE
;
855 #ifdef CONFIG_TCP_MD5SIG
857 opts
->options
|= OPTION_MD5
;
858 remaining
-= TCPOLEN_MD5SIG_ALIGNED
;
860 /* We can't fit any SACK blocks in a packet with MD5 + TS
861 * options. There was discussion about disabling SACK
862 * rather than TS in order to fit in better with old,
863 * buggy kernels, but that was deemed to be unnecessary.
865 if (synack_type
!= TCP_SYNACK_COOKIE
)
866 ireq
->tstamp_ok
&= !ireq
->sack_ok
;
870 /* We always send an MSS option. */
872 remaining
-= TCPOLEN_MSS_ALIGNED
;
874 if (likely(ireq
->wscale_ok
)) {
875 opts
->ws
= ireq
->rcv_wscale
;
876 opts
->options
|= OPTION_WSCALE
;
877 remaining
-= TCPOLEN_WSCALE_ALIGNED
;
879 if (likely(ireq
->tstamp_ok
)) {
880 opts
->options
|= OPTION_TS
;
881 opts
->tsval
= tcp_skb_timestamp(skb
) + tcp_rsk(req
)->ts_off
;
882 opts
->tsecr
= req
->ts_recent
;
883 remaining
-= TCPOLEN_TSTAMP_ALIGNED
;
885 if (likely(ireq
->sack_ok
)) {
886 opts
->options
|= OPTION_SACK_ADVERTISE
;
887 if (unlikely(!ireq
->tstamp_ok
))
888 remaining
-= TCPOLEN_SACKPERM_ALIGNED
;
890 if (foc
!= NULL
&& foc
->len
>= 0) {
893 need
+= foc
->exp
? TCPOLEN_EXP_FASTOPEN_BASE
:
894 TCPOLEN_FASTOPEN_BASE
;
895 need
= (need
+ 3) & ~3U; /* Align to 32 bits */
896 if (remaining
>= need
) {
897 opts
->options
|= OPTION_FAST_OPEN_COOKIE
;
898 opts
->fastopen_cookie
= foc
;
903 mptcp_set_option_cond(req
, opts
, &remaining
);
905 smc_set_option_cond(tcp_sk(sk
), ireq
, opts
, &remaining
);
907 bpf_skops_hdr_opt_len((struct sock
*)sk
, skb
, req
, syn_skb
,
908 synack_type
, opts
, &remaining
);
910 return MAX_TCP_OPTION_SPACE
- remaining
;
913 /* Compute TCP options for ESTABLISHED sockets. This is not the
914 * final wire format yet.
916 static unsigned int tcp_established_options(struct sock
*sk
, struct sk_buff
*skb
,
917 struct tcp_out_options
*opts
,
918 struct tcp_md5sig_key
**md5
)
920 struct tcp_sock
*tp
= tcp_sk(sk
);
921 unsigned int size
= 0;
922 unsigned int eff_sacks
;
927 #ifdef CONFIG_TCP_MD5SIG
928 if (static_branch_unlikely(&tcp_md5_needed
) &&
929 rcu_access_pointer(tp
->md5sig_info
)) {
930 *md5
= tp
->af_specific
->md5_lookup(sk
, sk
);
932 opts
->options
|= OPTION_MD5
;
933 size
+= TCPOLEN_MD5SIG_ALIGNED
;
938 if (likely(tp
->rx_opt
.tstamp_ok
)) {
939 opts
->options
|= OPTION_TS
;
940 opts
->tsval
= skb
? tcp_skb_timestamp(skb
) + tp
->tsoffset
: 0;
941 opts
->tsecr
= tp
->rx_opt
.ts_recent
;
942 size
+= TCPOLEN_TSTAMP_ALIGNED
;
945 /* MPTCP options have precedence over SACK for the limited TCP
946 * option space because a MPTCP connection would be forced to
947 * fall back to regular TCP if a required multipath option is
948 * missing. SACK still gets a chance to use whatever space is
951 if (sk_is_mptcp(sk
)) {
952 unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
953 unsigned int opt_size
= 0;
955 if (mptcp_established_options(sk
, skb
, &opt_size
, remaining
,
957 opts
->options
|= OPTION_MPTCP
;
962 eff_sacks
= tp
->rx_opt
.num_sacks
+ tp
->rx_opt
.dsack
;
963 if (unlikely(eff_sacks
)) {
964 const unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
965 if (unlikely(remaining
< TCPOLEN_SACK_BASE_ALIGNED
+
966 TCPOLEN_SACK_PERBLOCK
))
969 opts
->num_sack_blocks
=
970 min_t(unsigned int, eff_sacks
,
971 (remaining
- TCPOLEN_SACK_BASE_ALIGNED
) /
972 TCPOLEN_SACK_PERBLOCK
);
974 size
+= TCPOLEN_SACK_BASE_ALIGNED
+
975 opts
->num_sack_blocks
* TCPOLEN_SACK_PERBLOCK
;
978 if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp
,
979 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG
))) {
980 unsigned int remaining
= MAX_TCP_OPTION_SPACE
- size
;
982 bpf_skops_hdr_opt_len(sk
, skb
, NULL
, NULL
, 0, opts
, &remaining
);
984 size
= MAX_TCP_OPTION_SPACE
- remaining
;
991 /* TCP SMALL QUEUES (TSQ)
993 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
994 * to reduce RTT and bufferbloat.
995 * We do this using a special skb destructor (tcp_wfree).
997 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
998 * needs to be reallocated in a driver.
999 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
1001 * Since transmit from skb destructor is forbidden, we use a tasklet
1002 * to process all sockets that eventually need to send more skbs.
1003 * We use one tasklet per cpu, with its own queue of sockets.
1005 struct tsq_tasklet
{
1006 struct tasklet_struct tasklet
;
1007 struct list_head head
; /* queue of tcp sockets */
1009 static DEFINE_PER_CPU(struct tsq_tasklet
, tsq_tasklet
);
1011 static void tcp_tsq_write(struct sock
*sk
)
1013 if ((1 << sk
->sk_state
) &
1014 (TCPF_ESTABLISHED
| TCPF_FIN_WAIT1
| TCPF_CLOSING
|
1015 TCPF_CLOSE_WAIT
| TCPF_LAST_ACK
)) {
1016 struct tcp_sock
*tp
= tcp_sk(sk
);
1018 if (tp
->lost_out
> tp
->retrans_out
&&
1019 tcp_snd_cwnd(tp
) > tcp_packets_in_flight(tp
)) {
1020 tcp_mstamp_refresh(tp
);
1021 tcp_xmit_retransmit_queue(sk
);
1024 tcp_write_xmit(sk
, tcp_current_mss(sk
), tp
->nonagle
,
1029 static void tcp_tsq_handler(struct sock
*sk
)
1032 if (!sock_owned_by_user(sk
))
1034 else if (!test_and_set_bit(TCP_TSQ_DEFERRED
, &sk
->sk_tsq_flags
))
1039 * One tasklet per cpu tries to send more skbs.
1040 * We run in tasklet context but need to disable irqs when
1041 * transferring tsq->head because tcp_wfree() might
1042 * interrupt us (non NAPI drivers)
1044 static void tcp_tasklet_func(struct tasklet_struct
*t
)
1046 struct tsq_tasklet
*tsq
= from_tasklet(tsq
, t
, tasklet
);
1048 unsigned long flags
;
1049 struct list_head
*q
, *n
;
1050 struct tcp_sock
*tp
;
1053 local_irq_save(flags
);
1054 list_splice_init(&tsq
->head
, &list
);
1055 local_irq_restore(flags
);
1057 list_for_each_safe(q
, n
, &list
) {
1058 tp
= list_entry(q
, struct tcp_sock
, tsq_node
);
1059 list_del(&tp
->tsq_node
);
1061 sk
= (struct sock
*)tp
;
1062 smp_mb__before_atomic();
1063 clear_bit(TSQ_QUEUED
, &sk
->sk_tsq_flags
);
1065 tcp_tsq_handler(sk
);
1070 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
1071 TCPF_WRITE_TIMER_DEFERRED | \
1072 TCPF_DELACK_TIMER_DEFERRED | \
1073 TCPF_MTU_REDUCED_DEFERRED)
1075 * tcp_release_cb - tcp release_sock() callback
1078 * called from release_sock() to perform protocol dependent
1079 * actions before socket release.
1081 void tcp_release_cb(struct sock
*sk
)
1083 unsigned long flags
, nflags
;
1085 /* perform an atomic operation only if at least one flag is set */
1087 flags
= sk
->sk_tsq_flags
;
1088 if (!(flags
& TCP_DEFERRED_ALL
))
1090 nflags
= flags
& ~TCP_DEFERRED_ALL
;
1091 } while (cmpxchg(&sk
->sk_tsq_flags
, flags
, nflags
) != flags
);
1093 if (flags
& TCPF_TSQ_DEFERRED
) {
1097 /* Here begins the tricky part :
1098 * We are called from release_sock() with :
1100 * 2) sk_lock.slock spinlock held
1101 * 3) socket owned by us (sk->sk_lock.owned == 1)
1103 * But following code is meant to be called from BH handlers,
1104 * so we should keep BH disabled, but early release socket ownership
1106 sock_release_ownership(sk
);
1108 if (flags
& TCPF_WRITE_TIMER_DEFERRED
) {
1109 tcp_write_timer_handler(sk
);
1112 if (flags
& TCPF_DELACK_TIMER_DEFERRED
) {
1113 tcp_delack_timer_handler(sk
);
1116 if (flags
& TCPF_MTU_REDUCED_DEFERRED
) {
1117 inet_csk(sk
)->icsk_af_ops
->mtu_reduced(sk
);
1121 EXPORT_SYMBOL(tcp_release_cb
);
1123 void __init
tcp_tasklet_init(void)
1127 for_each_possible_cpu(i
) {
1128 struct tsq_tasklet
*tsq
= &per_cpu(tsq_tasklet
, i
);
1130 INIT_LIST_HEAD(&tsq
->head
);
1131 tasklet_setup(&tsq
->tasklet
, tcp_tasklet_func
);
1136 * Write buffer destructor automatically called from kfree_skb.
1137 * We can't xmit new skbs from this context, as we might already
1140 void tcp_wfree(struct sk_buff
*skb
)
1142 struct sock
*sk
= skb
->sk
;
1143 struct tcp_sock
*tp
= tcp_sk(sk
);
1144 unsigned long flags
, nval
, oval
;
1146 /* Keep one reference on sk_wmem_alloc.
1147 * Will be released by sk_free() from here or tcp_tasklet_func()
1149 WARN_ON(refcount_sub_and_test(skb
->truesize
- 1, &sk
->sk_wmem_alloc
));
1151 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
1152 * Wait until our queues (qdisc + devices) are drained.
1154 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1155 * - chance for incoming ACK (processed by another cpu maybe)
1156 * to migrate this flow (skb->ooo_okay will be eventually set)
1158 if (refcount_read(&sk
->sk_wmem_alloc
) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current
)
1161 for (oval
= READ_ONCE(sk
->sk_tsq_flags
);; oval
= nval
) {
1162 struct tsq_tasklet
*tsq
;
1165 if (!(oval
& TSQF_THROTTLED
) || (oval
& TSQF_QUEUED
))
1168 nval
= (oval
& ~TSQF_THROTTLED
) | TSQF_QUEUED
;
1169 nval
= cmpxchg(&sk
->sk_tsq_flags
, oval
, nval
);
1173 /* queue this socket to tasklet queue */
1174 local_irq_save(flags
);
1175 tsq
= this_cpu_ptr(&tsq_tasklet
);
1176 empty
= list_empty(&tsq
->head
);
1177 list_add(&tp
->tsq_node
, &tsq
->head
);
1179 tasklet_schedule(&tsq
->tasklet
);
1180 local_irq_restore(flags
);
1187 /* Note: Called under soft irq.
1188 * We can call TCP stack right away, unless socket is owned by user.
1190 enum hrtimer_restart
tcp_pace_kick(struct hrtimer
*timer
)
1192 struct tcp_sock
*tp
= container_of(timer
, struct tcp_sock
, pacing_timer
);
1193 struct sock
*sk
= (struct sock
*)tp
;
1195 tcp_tsq_handler(sk
);
1198 return HRTIMER_NORESTART
;
1201 static void tcp_update_skb_after_send(struct sock
*sk
, struct sk_buff
*skb
,
1204 struct tcp_sock
*tp
= tcp_sk(sk
);
1206 if (sk
->sk_pacing_status
!= SK_PACING_NONE
) {
1207 unsigned long rate
= sk
->sk_pacing_rate
;
1209 /* Original sch_fq does not pace first 10 MSS
1210 * Note that tp->data_segs_out overflows after 2^32 packets,
1211 * this is a minor annoyance.
1213 if (rate
!= ~0UL && rate
&& tp
->data_segs_out
>= 10) {
1214 u64 len_ns
= div64_ul((u64
)skb
->len
* NSEC_PER_SEC
, rate
);
1215 u64 credit
= tp
->tcp_wstamp_ns
- prior_wstamp
;
1217 /* take into account OS jitter */
1218 len_ns
-= min_t(u64
, len_ns
/ 2, credit
);
1219 tp
->tcp_wstamp_ns
+= len_ns
;
1222 list_move_tail(&skb
->tcp_tsorted_anchor
, &tp
->tsorted_sent_queue
);
1225 INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock
*sk
, struct sk_buff
*skb
, struct flowi
*fl
));
1226 INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock
*sk
, struct sk_buff
*skb
, struct flowi
*fl
));
1227 INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock
*sk
, struct sk_buff
*skb
));
1229 /* This routine actually transmits TCP packets queued in by
1230 * tcp_do_sendmsg(). This is used by both the initial
1231 * transmission and possible later retransmissions.
1232 * All SKB's seen here are completely headerless. It is our
1233 * job to build the TCP header, and pass the packet down to
1234 * IP so it can do the same plus pass the packet off to the
1237 * We are working here with either a clone of the original
1238 * SKB, or a fresh unique copy made by the retransmit engine.
1240 static int __tcp_transmit_skb(struct sock
*sk
, struct sk_buff
*skb
,
1241 int clone_it
, gfp_t gfp_mask
, u32 rcv_nxt
)
1243 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1244 struct inet_sock
*inet
;
1245 struct tcp_sock
*tp
;
1246 struct tcp_skb_cb
*tcb
;
1247 struct tcp_out_options opts
;
1248 unsigned int tcp_options_size
, tcp_header_size
;
1249 struct sk_buff
*oskb
= NULL
;
1250 struct tcp_md5sig_key
*md5
;
1255 BUG_ON(!skb
|| !tcp_skb_pcount(skb
));
1257 prior_wstamp
= tp
->tcp_wstamp_ns
;
1258 tp
->tcp_wstamp_ns
= max(tp
->tcp_wstamp_ns
, tp
->tcp_clock_cache
);
1259 skb_set_delivery_time(skb
, tp
->tcp_wstamp_ns
, true);
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(th
, tp
, &opts
);
1362 #ifdef CONFIG_TCP_MD5SIG
1363 /* Calculate the MD5 hash, as we have all we need now */
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_LEGACY_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_keeptruesize(skb
, gfp
))
1568 /* Get a new skb... force flag on. */
1569 buff
= tcp_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 skb_set_delivery_time(buff
, skb
->tstamp
, true);
1596 tcp_fragment_tstamp(skb
, buff
);
1598 old_factor
= tcp_skb_pcount(skb
);
1600 /* Fix up tso_factor for both original and new SKB. */
1601 tcp_set_skb_tso_segs(skb
, mss_now
);
1602 tcp_set_skb_tso_segs(buff
, mss_now
);
1604 /* Update delivered info for the new segment */
1605 TCP_SKB_CB(buff
)->tx
= TCP_SKB_CB(skb
)->tx
;
1607 /* If this packet has been sent out already, we must
1608 * adjust the various packet counters.
1610 if (!before(tp
->snd_nxt
, TCP_SKB_CB(buff
)->end_seq
)) {
1611 int diff
= old_factor
- tcp_skb_pcount(skb
) -
1612 tcp_skb_pcount(buff
);
1615 tcp_adjust_pcount(sk
, skb
, diff
);
1618 /* Link BUFF into the send queue. */
1619 __skb_header_release(buff
);
1620 tcp_insert_write_queue_after(skb
, buff
, sk
, tcp_queue
);
1621 if (tcp_queue
== TCP_FRAG_IN_RTX_QUEUE
)
1622 list_add(&buff
->tcp_tsorted_anchor
, &skb
->tcp_tsorted_anchor
);
1627 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1628 * data is not copied, but immediately discarded.
1630 static int __pskb_trim_head(struct sk_buff
*skb
, int len
)
1632 struct skb_shared_info
*shinfo
;
1635 eat
= min_t(int, len
, skb_headlen(skb
));
1637 __skb_pull(skb
, eat
);
1644 shinfo
= skb_shinfo(skb
);
1645 for (i
= 0; i
< shinfo
->nr_frags
; i
++) {
1646 int size
= skb_frag_size(&shinfo
->frags
[i
]);
1649 skb_frag_unref(skb
, i
);
1652 shinfo
->frags
[k
] = shinfo
->frags
[i
];
1654 skb_frag_off_add(&shinfo
->frags
[k
], eat
);
1655 skb_frag_size_sub(&shinfo
->frags
[k
], eat
);
1661 shinfo
->nr_frags
= k
;
1663 skb
->data_len
-= len
;
1664 skb
->len
= skb
->data_len
;
1668 /* Remove acked data from a packet in the transmit queue. */
1669 int tcp_trim_head(struct sock
*sk
, struct sk_buff
*skb
, u32 len
)
1673 if (skb_unclone_keeptruesize(skb
, GFP_ATOMIC
))
1676 delta_truesize
= __pskb_trim_head(skb
, len
);
1678 TCP_SKB_CB(skb
)->seq
+= len
;
1680 if (delta_truesize
) {
1681 skb
->truesize
-= delta_truesize
;
1682 sk_wmem_queued_add(sk
, -delta_truesize
);
1683 if (!skb_zcopy_pure(skb
))
1684 sk_mem_uncharge(sk
, delta_truesize
);
1687 /* Any change of skb->len requires recalculation of tso factor. */
1688 if (tcp_skb_pcount(skb
) > 1)
1689 tcp_set_skb_tso_segs(skb
, tcp_skb_mss(skb
));
1694 /* Calculate MSS not accounting any TCP options. */
1695 static inline int __tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1697 const struct tcp_sock
*tp
= tcp_sk(sk
);
1698 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1701 /* Calculate base mss without TCP options:
1702 It is MMS_S - sizeof(tcphdr) of rfc1122
1704 mss_now
= pmtu
- icsk
->icsk_af_ops
->net_header_len
- sizeof(struct tcphdr
);
1706 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1707 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1708 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1710 if (dst
&& dst_allfrag(dst
))
1711 mss_now
-= icsk
->icsk_af_ops
->net_frag_header_len
;
1714 /* Clamp it (mss_clamp does not include tcp options) */
1715 if (mss_now
> tp
->rx_opt
.mss_clamp
)
1716 mss_now
= tp
->rx_opt
.mss_clamp
;
1718 /* Now subtract optional transport overhead */
1719 mss_now
-= icsk
->icsk_ext_hdr_len
;
1721 /* Then reserve room for full set of TCP options and 8 bytes of data */
1722 mss_now
= max(mss_now
, sock_net(sk
)->ipv4
.sysctl_tcp_min_snd_mss
);
1726 /* Calculate MSS. Not accounting for SACKs here. */
1727 int tcp_mtu_to_mss(struct sock
*sk
, int pmtu
)
1729 /* Subtract TCP options size, not including SACKs */
1730 return __tcp_mtu_to_mss(sk
, pmtu
) -
1731 (tcp_sk(sk
)->tcp_header_len
- sizeof(struct tcphdr
));
1733 EXPORT_SYMBOL(tcp_mtu_to_mss
);
1735 /* Inverse of above */
1736 int tcp_mss_to_mtu(struct sock
*sk
, int mss
)
1738 const struct tcp_sock
*tp
= tcp_sk(sk
);
1739 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
1743 tp
->tcp_header_len
+
1744 icsk
->icsk_ext_hdr_len
+
1745 icsk
->icsk_af_ops
->net_header_len
;
1747 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1748 if (icsk
->icsk_af_ops
->net_frag_header_len
) {
1749 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1751 if (dst
&& dst_allfrag(dst
))
1752 mtu
+= icsk
->icsk_af_ops
->net_frag_header_len
;
1756 EXPORT_SYMBOL(tcp_mss_to_mtu
);
1758 /* MTU probing init per socket */
1759 void tcp_mtup_init(struct sock
*sk
)
1761 struct tcp_sock
*tp
= tcp_sk(sk
);
1762 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1763 struct net
*net
= sock_net(sk
);
1765 icsk
->icsk_mtup
.enabled
= net
->ipv4
.sysctl_tcp_mtu_probing
> 1;
1766 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+ sizeof(struct tcphdr
) +
1767 icsk
->icsk_af_ops
->net_header_len
;
1768 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, net
->ipv4
.sysctl_tcp_base_mss
);
1769 icsk
->icsk_mtup
.probe_size
= 0;
1770 if (icsk
->icsk_mtup
.enabled
)
1771 icsk
->icsk_mtup
.probe_timestamp
= tcp_jiffies32
;
1773 EXPORT_SYMBOL(tcp_mtup_init
);
1775 /* This function synchronize snd mss to current pmtu/exthdr set.
1777 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1778 for TCP options, but includes only bare TCP header.
1780 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1781 It is minimum of user_mss and mss received with SYN.
1782 It also does not include TCP options.
1784 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1786 tp->mss_cache is current effective sending mss, including
1787 all tcp options except for SACKs. It is evaluated,
1788 taking into account current pmtu, but never exceeds
1789 tp->rx_opt.mss_clamp.
1791 NOTE1. rfc1122 clearly states that advertised MSS
1792 DOES NOT include either tcp or ip options.
1794 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1795 are READ ONLY outside this function. --ANK (980731)
1797 unsigned int tcp_sync_mss(struct sock
*sk
, u32 pmtu
)
1799 struct tcp_sock
*tp
= tcp_sk(sk
);
1800 struct inet_connection_sock
*icsk
= inet_csk(sk
);
1803 if (icsk
->icsk_mtup
.search_high
> pmtu
)
1804 icsk
->icsk_mtup
.search_high
= pmtu
;
1806 mss_now
= tcp_mtu_to_mss(sk
, pmtu
);
1807 mss_now
= tcp_bound_to_half_wnd(tp
, mss_now
);
1809 /* And store cached results */
1810 icsk
->icsk_pmtu_cookie
= pmtu
;
1811 if (icsk
->icsk_mtup
.enabled
)
1812 mss_now
= min(mss_now
, tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_low
));
1813 tp
->mss_cache
= mss_now
;
1817 EXPORT_SYMBOL(tcp_sync_mss
);
1819 /* Compute the current effective MSS, taking SACKs and IP options,
1820 * and even PMTU discovery events into account.
1822 unsigned int tcp_current_mss(struct sock
*sk
)
1824 const struct tcp_sock
*tp
= tcp_sk(sk
);
1825 const struct dst_entry
*dst
= __sk_dst_get(sk
);
1827 unsigned int header_len
;
1828 struct tcp_out_options opts
;
1829 struct tcp_md5sig_key
*md5
;
1831 mss_now
= tp
->mss_cache
;
1834 u32 mtu
= dst_mtu(dst
);
1835 if (mtu
!= inet_csk(sk
)->icsk_pmtu_cookie
)
1836 mss_now
= tcp_sync_mss(sk
, mtu
);
1839 header_len
= tcp_established_options(sk
, NULL
, &opts
, &md5
) +
1840 sizeof(struct tcphdr
);
1841 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1842 * some common options. If this is an odd packet (because we have SACK
1843 * blocks etc) then our calculated header_len will be different, and
1844 * we have to adjust mss_now correspondingly */
1845 if (header_len
!= tp
->tcp_header_len
) {
1846 int delta
= (int) header_len
- tp
->tcp_header_len
;
1853 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1854 * As additional protections, we do not touch cwnd in retransmission phases,
1855 * and if application hit its sndbuf limit recently.
1857 static void tcp_cwnd_application_limited(struct sock
*sk
)
1859 struct tcp_sock
*tp
= tcp_sk(sk
);
1861 if (inet_csk(sk
)->icsk_ca_state
== TCP_CA_Open
&&
1862 sk
->sk_socket
&& !test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
)) {
1863 /* Limited by application or receiver window. */
1864 u32 init_win
= tcp_init_cwnd(tp
, __sk_dst_get(sk
));
1865 u32 win_used
= max(tp
->snd_cwnd_used
, init_win
);
1866 if (win_used
< tcp_snd_cwnd(tp
)) {
1867 tp
->snd_ssthresh
= tcp_current_ssthresh(sk
);
1868 tcp_snd_cwnd_set(tp
, (tcp_snd_cwnd(tp
) + win_used
) >> 1);
1870 tp
->snd_cwnd_used
= 0;
1872 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1875 static void tcp_cwnd_validate(struct sock
*sk
, bool is_cwnd_limited
)
1877 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1878 struct tcp_sock
*tp
= tcp_sk(sk
);
1880 /* Track the maximum number of outstanding packets in each
1881 * window, and remember whether we were cwnd-limited then.
1883 if (!before(tp
->snd_una
, tp
->max_packets_seq
) ||
1884 tp
->packets_out
> tp
->max_packets_out
||
1886 tp
->max_packets_out
= tp
->packets_out
;
1887 tp
->max_packets_seq
= tp
->snd_nxt
;
1888 tp
->is_cwnd_limited
= is_cwnd_limited
;
1891 if (tcp_is_cwnd_limited(sk
)) {
1892 /* Network is feed fully. */
1893 tp
->snd_cwnd_used
= 0;
1894 tp
->snd_cwnd_stamp
= tcp_jiffies32
;
1896 /* Network starves. */
1897 if (tp
->packets_out
> tp
->snd_cwnd_used
)
1898 tp
->snd_cwnd_used
= tp
->packets_out
;
1900 if (sock_net(sk
)->ipv4
.sysctl_tcp_slow_start_after_idle
&&
1901 (s32
)(tcp_jiffies32
- tp
->snd_cwnd_stamp
) >= inet_csk(sk
)->icsk_rto
&&
1902 !ca_ops
->cong_control
)
1903 tcp_cwnd_application_limited(sk
);
1905 /* The following conditions together indicate the starvation
1906 * is caused by insufficient sender buffer:
1907 * 1) just sent some data (see tcp_write_xmit)
1908 * 2) not cwnd limited (this else condition)
1909 * 3) no more data to send (tcp_write_queue_empty())
1910 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1912 if (tcp_write_queue_empty(sk
) && sk
->sk_socket
&&
1913 test_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
) &&
1914 (1 << sk
->sk_state
) & (TCPF_ESTABLISHED
| TCPF_CLOSE_WAIT
))
1915 tcp_chrono_start(sk
, TCP_CHRONO_SNDBUF_LIMITED
);
1919 /* Minshall's variant of the Nagle send check. */
1920 static bool tcp_minshall_check(const struct tcp_sock
*tp
)
1922 return after(tp
->snd_sml
, tp
->snd_una
) &&
1923 !after(tp
->snd_sml
, tp
->snd_nxt
);
1926 /* Update snd_sml if this skb is under mss
1927 * Note that a TSO packet might end with a sub-mss segment
1928 * The test is really :
1929 * if ((skb->len % mss) != 0)
1930 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1931 * But we can avoid doing the divide again given we already have
1932 * skb_pcount = skb->len / mss_now
1934 static void tcp_minshall_update(struct tcp_sock
*tp
, unsigned int mss_now
,
1935 const struct sk_buff
*skb
)
1937 if (skb
->len
< tcp_skb_pcount(skb
) * mss_now
)
1938 tp
->snd_sml
= TCP_SKB_CB(skb
)->end_seq
;
1941 /* Return false, if packet can be sent now without violation Nagle's rules:
1942 * 1. It is full sized. (provided by caller in %partial bool)
1943 * 2. Or it contains FIN. (already checked by caller)
1944 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1945 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1946 * With Minshall's modification: all sent small packets are ACKed.
1948 static bool tcp_nagle_check(bool partial
, const struct tcp_sock
*tp
,
1952 ((nonagle
& TCP_NAGLE_CORK
) ||
1953 (!nonagle
&& tp
->packets_out
&& tcp_minshall_check(tp
)));
1956 /* Return how many segs we'd like on a TSO packet,
1957 * depending on current pacing rate, and how close the peer is.
1960 * - For close peers, we rather send bigger packets to reduce
1961 * cpu costs, because occasional losses will be repaired fast.
1962 * - For long distance/rtt flows, we would like to get ACK clocking
1963 * with 1 ACK per ms.
1965 * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting
1966 * in bigger TSO bursts. We we cut the RTT-based allowance in half
1967 * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance
1968 * is below 1500 bytes after 6 * ~500 usec = 3ms.
1970 static u32
tcp_tso_autosize(const struct sock
*sk
, unsigned int mss_now
,
1973 unsigned long bytes
;
1976 bytes
= sk
->sk_pacing_rate
>> READ_ONCE(sk
->sk_pacing_shift
);
1978 r
= tcp_min_rtt(tcp_sk(sk
)) >> sock_net(sk
)->ipv4
.sysctl_tcp_tso_rtt_log
;
1979 if (r
< BITS_PER_TYPE(sk
->sk_gso_max_size
))
1980 bytes
+= sk
->sk_gso_max_size
>> r
;
1982 bytes
= min_t(unsigned long, bytes
, sk
->sk_gso_max_size
);
1984 return max_t(u32
, bytes
/ mss_now
, min_tso_segs
);
1987 /* Return the number of segments we want in the skb we are transmitting.
1988 * See if congestion control module wants to decide; otherwise, autosize.
1990 static u32
tcp_tso_segs(struct sock
*sk
, unsigned int mss_now
)
1992 const struct tcp_congestion_ops
*ca_ops
= inet_csk(sk
)->icsk_ca_ops
;
1993 u32 min_tso
, tso_segs
;
1995 min_tso
= ca_ops
->min_tso_segs
?
1996 ca_ops
->min_tso_segs(sk
) :
1997 sock_net(sk
)->ipv4
.sysctl_tcp_min_tso_segs
;
1999 tso_segs
= tcp_tso_autosize(sk
, mss_now
, min_tso
);
2000 return min_t(u32
, tso_segs
, sk
->sk_gso_max_segs
);
2003 /* Returns the portion of skb which can be sent right away */
2004 static unsigned int tcp_mss_split_point(const struct sock
*sk
,
2005 const struct sk_buff
*skb
,
2006 unsigned int mss_now
,
2007 unsigned int max_segs
,
2010 const struct tcp_sock
*tp
= tcp_sk(sk
);
2011 u32 partial
, needed
, window
, max_len
;
2013 window
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2014 max_len
= mss_now
* max_segs
;
2016 if (likely(max_len
<= window
&& skb
!= tcp_write_queue_tail(sk
)))
2019 needed
= min(skb
->len
, window
);
2021 if (max_len
<= needed
)
2024 partial
= needed
% mss_now
;
2025 /* If last segment is not a full MSS, check if Nagle rules allow us
2026 * to include this last segment in this skb.
2027 * Otherwise, we'll split the skb at last MSS boundary
2029 if (tcp_nagle_check(partial
!= 0, tp
, nonagle
))
2030 return needed
- partial
;
2035 /* Can at least one segment of SKB be sent right now, according to the
2036 * congestion window rules? If so, return how many segments are allowed.
2038 static inline unsigned int tcp_cwnd_test(const struct tcp_sock
*tp
,
2039 const struct sk_buff
*skb
)
2041 u32 in_flight
, cwnd
, halfcwnd
;
2043 /* Don't be strict about the congestion window for the final FIN. */
2044 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) &&
2045 tcp_skb_pcount(skb
) == 1)
2048 in_flight
= tcp_packets_in_flight(tp
);
2049 cwnd
= tcp_snd_cwnd(tp
);
2050 if (in_flight
>= cwnd
)
2053 /* For better scheduling, ensure we have at least
2054 * 2 GSO packets in flight.
2056 halfcwnd
= max(cwnd
>> 1, 1U);
2057 return min(halfcwnd
, cwnd
- in_flight
);
2060 /* Initialize TSO state of a skb.
2061 * This must be invoked the first time we consider transmitting
2062 * SKB onto the wire.
2064 static int tcp_init_tso_segs(struct sk_buff
*skb
, unsigned int mss_now
)
2066 int tso_segs
= tcp_skb_pcount(skb
);
2068 if (!tso_segs
|| (tso_segs
> 1 && tcp_skb_mss(skb
) != mss_now
)) {
2069 tcp_set_skb_tso_segs(skb
, mss_now
);
2070 tso_segs
= tcp_skb_pcount(skb
);
2076 /* Return true if the Nagle test allows this packet to be
2079 static inline bool tcp_nagle_test(const struct tcp_sock
*tp
, const struct sk_buff
*skb
,
2080 unsigned int cur_mss
, int nonagle
)
2082 /* Nagle rule does not apply to frames, which sit in the middle of the
2083 * write_queue (they have no chances to get new data).
2085 * This is implemented in the callers, where they modify the 'nonagle'
2086 * argument based upon the location of SKB in the send queue.
2088 if (nonagle
& TCP_NAGLE_PUSH
)
2091 /* Don't use the nagle rule for urgent data (or for the final FIN). */
2092 if (tcp_urg_mode(tp
) || (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
))
2095 if (!tcp_nagle_check(skb
->len
< cur_mss
, tp
, nonagle
))
2101 /* Does at least the first segment of SKB fit into the send window? */
2102 static bool tcp_snd_wnd_test(const struct tcp_sock
*tp
,
2103 const struct sk_buff
*skb
,
2104 unsigned int cur_mss
)
2106 u32 end_seq
= TCP_SKB_CB(skb
)->end_seq
;
2108 if (skb
->len
> cur_mss
)
2109 end_seq
= TCP_SKB_CB(skb
)->seq
+ cur_mss
;
2111 return !after(end_seq
, tcp_wnd_end(tp
));
2114 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2115 * which is put after SKB on the list. It is very much like
2116 * tcp_fragment() except that it may make several kinds of assumptions
2117 * in order to speed up the splitting operation. In particular, we
2118 * know that all the data is in scatter-gather pages, and that the
2119 * packet has never been sent out before (and thus is not cloned).
2121 static int tso_fragment(struct sock
*sk
, struct sk_buff
*skb
, unsigned int len
,
2122 unsigned int mss_now
, gfp_t gfp
)
2124 int nlen
= skb
->len
- len
;
2125 struct sk_buff
*buff
;
2128 /* All of a TSO frame must be composed of paged data. */
2129 if (skb
->len
!= skb
->data_len
)
2130 return tcp_fragment(sk
, TCP_FRAG_IN_WRITE_QUEUE
,
2131 skb
, len
, mss_now
, gfp
);
2133 buff
= tcp_stream_alloc_skb(sk
, 0, gfp
, true);
2134 if (unlikely(!buff
))
2136 skb_copy_decrypted(buff
, skb
);
2137 mptcp_skb_ext_copy(buff
, skb
);
2139 sk_wmem_queued_add(sk
, buff
->truesize
);
2140 sk_mem_charge(sk
, buff
->truesize
);
2141 buff
->truesize
+= nlen
;
2142 skb
->truesize
-= nlen
;
2144 /* Correct the sequence numbers. */
2145 TCP_SKB_CB(buff
)->seq
= TCP_SKB_CB(skb
)->seq
+ len
;
2146 TCP_SKB_CB(buff
)->end_seq
= TCP_SKB_CB(skb
)->end_seq
;
2147 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(buff
)->seq
;
2149 /* PSH and FIN should only be set in the second packet. */
2150 flags
= TCP_SKB_CB(skb
)->tcp_flags
;
2151 TCP_SKB_CB(skb
)->tcp_flags
= flags
& ~(TCPHDR_FIN
| TCPHDR_PSH
);
2152 TCP_SKB_CB(buff
)->tcp_flags
= flags
;
2154 tcp_skb_fragment_eor(skb
, buff
);
2156 skb_split(skb
, buff
, len
);
2157 tcp_fragment_tstamp(skb
, buff
);
2159 /* Fix up tso_factor for both original and new SKB. */
2160 tcp_set_skb_tso_segs(skb
, mss_now
);
2161 tcp_set_skb_tso_segs(buff
, mss_now
);
2163 /* Link BUFF into the send queue. */
2164 __skb_header_release(buff
);
2165 tcp_insert_write_queue_after(skb
, buff
, sk
, TCP_FRAG_IN_WRITE_QUEUE
);
2170 /* Try to defer sending, if possible, in order to minimize the amount
2171 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2173 * This algorithm is from John Heffner.
2175 static bool tcp_tso_should_defer(struct sock
*sk
, struct sk_buff
*skb
,
2176 bool *is_cwnd_limited
,
2177 bool *is_rwnd_limited
,
2180 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
2181 u32 send_win
, cong_win
, limit
, in_flight
;
2182 struct tcp_sock
*tp
= tcp_sk(sk
);
2183 struct sk_buff
*head
;
2187 if (icsk
->icsk_ca_state
>= TCP_CA_Recovery
)
2190 /* Avoid bursty behavior by allowing defer
2191 * only if the last write was recent (1 ms).
2192 * Note that tp->tcp_wstamp_ns can be in the future if we have
2193 * packets waiting in a qdisc or device for EDT delivery.
2195 delta
= tp
->tcp_clock_cache
- tp
->tcp_wstamp_ns
- NSEC_PER_MSEC
;
2199 in_flight
= tcp_packets_in_flight(tp
);
2201 BUG_ON(tcp_skb_pcount(skb
) <= 1);
2202 BUG_ON(tcp_snd_cwnd(tp
) <= in_flight
);
2204 send_win
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
2206 /* From in_flight test above, we know that cwnd > in_flight. */
2207 cong_win
= (tcp_snd_cwnd(tp
) - in_flight
) * tp
->mss_cache
;
2209 limit
= min(send_win
, cong_win
);
2211 /* If a full-sized TSO skb can be sent, do it. */
2212 if (limit
>= max_segs
* tp
->mss_cache
)
2215 /* Middle in queue won't get any more data, full sendable already? */
2216 if ((skb
!= tcp_write_queue_tail(sk
)) && (limit
>= skb
->len
))
2219 win_divisor
= READ_ONCE(sock_net(sk
)->ipv4
.sysctl_tcp_tso_win_divisor
);
2221 u32 chunk
= min(tp
->snd_wnd
, tcp_snd_cwnd(tp
) * tp
->mss_cache
);
2223 /* If at least some fraction of a window is available,
2226 chunk
/= win_divisor
;
2230 /* Different approach, try not to defer past a single
2231 * ACK. Receiver should ACK every other full sized
2232 * frame, so if we have space for more than 3 frames
2235 if (limit
> tcp_max_tso_deferred_mss(tp
) * tp
->mss_cache
)
2239 /* TODO : use tsorted_sent_queue ? */
2240 head
= tcp_rtx_queue_head(sk
);
2243 delta
= tp
->tcp_clock_cache
- head
->tstamp
;
2244 /* If next ACK is likely to come too late (half srtt), do not defer */
2245 if ((s64
)(delta
- (u64
)NSEC_PER_USEC
* (tp
->srtt_us
>> 4)) < 0)
2248 /* Ok, it looks like it is advisable to defer.
2249 * Three cases are tracked :
2250 * 1) We are cwnd-limited
2251 * 2) We are rwnd-limited
2252 * 3) We are application limited.
2254 if (cong_win
< send_win
) {
2255 if (cong_win
<= skb
->len
) {
2256 *is_cwnd_limited
= true;
2260 if (send_win
<= skb
->len
) {
2261 *is_rwnd_limited
= true;
2266 /* If this packet won't get more data, do not wait. */
2267 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_FIN
) ||
2268 TCP_SKB_CB(skb
)->eor
)
2277 static inline void tcp_mtu_check_reprobe(struct sock
*sk
)
2279 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2280 struct tcp_sock
*tp
= tcp_sk(sk
);
2281 struct net
*net
= sock_net(sk
);
2285 interval
= net
->ipv4
.sysctl_tcp_probe_interval
;
2286 delta
= tcp_jiffies32
- icsk
->icsk_mtup
.probe_timestamp
;
2287 if (unlikely(delta
>= interval
* HZ
)) {
2288 int mss
= tcp_current_mss(sk
);
2290 /* Update current search range */
2291 icsk
->icsk_mtup
.probe_size
= 0;
2292 icsk
->icsk_mtup
.search_high
= tp
->rx_opt
.mss_clamp
+
2293 sizeof(struct tcphdr
) +
2294 icsk
->icsk_af_ops
->net_header_len
;
2295 icsk
->icsk_mtup
.search_low
= tcp_mss_to_mtu(sk
, mss
);
2297 /* Update probe time stamp */
2298 icsk
->icsk_mtup
.probe_timestamp
= tcp_jiffies32
;
2302 static bool tcp_can_coalesce_send_queue_head(struct sock
*sk
, int len
)
2304 struct sk_buff
*skb
, *next
;
2306 skb
= tcp_send_head(sk
);
2307 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
2308 if (len
<= skb
->len
)
2311 if (unlikely(TCP_SKB_CB(skb
)->eor
) ||
2312 tcp_has_tx_tstamp(skb
) ||
2313 !skb_pure_zcopy_same(skb
, next
))
2322 /* Create a new MTU probe if we are ready.
2323 * MTU probe is regularly attempting to increase the path MTU by
2324 * deliberately sending larger packets. This discovers routing
2325 * changes resulting in larger path MTUs.
2327 * Returns 0 if we should wait to probe (no cwnd available),
2328 * 1 if a probe was sent,
2331 static int tcp_mtu_probe(struct sock
*sk
)
2333 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2334 struct tcp_sock
*tp
= tcp_sk(sk
);
2335 struct sk_buff
*skb
, *nskb
, *next
;
2336 struct net
*net
= sock_net(sk
);
2343 /* Not currently probing/verifying,
2345 * have enough cwnd, and
2346 * not SACKing (the variable headers throw things off)
2348 if (likely(!icsk
->icsk_mtup
.enabled
||
2349 icsk
->icsk_mtup
.probe_size
||
2350 inet_csk(sk
)->icsk_ca_state
!= TCP_CA_Open
||
2351 tcp_snd_cwnd(tp
) < 11 ||
2352 tp
->rx_opt
.num_sacks
|| tp
->rx_opt
.dsack
))
2355 /* Use binary search for probe_size between tcp_mss_base,
2356 * and current mss_clamp. if (search_high - search_low)
2357 * smaller than a threshold, backoff from probing.
2359 mss_now
= tcp_current_mss(sk
);
2360 probe_size
= tcp_mtu_to_mss(sk
, (icsk
->icsk_mtup
.search_high
+
2361 icsk
->icsk_mtup
.search_low
) >> 1);
2362 size_needed
= probe_size
+ (tp
->reordering
+ 1) * tp
->mss_cache
;
2363 interval
= icsk
->icsk_mtup
.search_high
- icsk
->icsk_mtup
.search_low
;
2364 /* When misfortune happens, we are reprobing actively,
2365 * and then reprobe timer has expired. We stick with current
2366 * probing process by not resetting search range to its orignal.
2368 if (probe_size
> tcp_mtu_to_mss(sk
, icsk
->icsk_mtup
.search_high
) ||
2369 interval
< net
->ipv4
.sysctl_tcp_probe_threshold
) {
2370 /* Check whether enough time has elaplased for
2371 * another round of probing.
2373 tcp_mtu_check_reprobe(sk
);
2377 /* Have enough data in the send queue to probe? */
2378 if (tp
->write_seq
- tp
->snd_nxt
< size_needed
)
2381 if (tp
->snd_wnd
< size_needed
)
2383 if (after(tp
->snd_nxt
+ size_needed
, tcp_wnd_end(tp
)))
2386 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2387 if (tcp_packets_in_flight(tp
) + 2 > tcp_snd_cwnd(tp
)) {
2388 if (!tcp_packets_in_flight(tp
))
2394 if (!tcp_can_coalesce_send_queue_head(sk
, probe_size
))
2397 /* We're allowed to probe. Build it now. */
2398 nskb
= tcp_stream_alloc_skb(sk
, probe_size
, GFP_ATOMIC
, false);
2401 sk_wmem_queued_add(sk
, nskb
->truesize
);
2402 sk_mem_charge(sk
, nskb
->truesize
);
2404 skb
= tcp_send_head(sk
);
2405 skb_copy_decrypted(nskb
, skb
);
2406 mptcp_skb_ext_copy(nskb
, skb
);
2408 TCP_SKB_CB(nskb
)->seq
= TCP_SKB_CB(skb
)->seq
;
2409 TCP_SKB_CB(nskb
)->end_seq
= TCP_SKB_CB(skb
)->seq
+ probe_size
;
2410 TCP_SKB_CB(nskb
)->tcp_flags
= TCPHDR_ACK
;
2412 tcp_insert_write_queue_before(nskb
, skb
, sk
);
2413 tcp_highest_sack_replace(sk
, skb
, nskb
);
2416 tcp_for_write_queue_from_safe(skb
, next
, sk
) {
2417 copy
= min_t(int, skb
->len
, probe_size
- len
);
2418 skb_copy_bits(skb
, 0, skb_put(nskb
, copy
), copy
);
2420 if (skb
->len
<= copy
) {
2421 /* We've eaten all the data from this skb.
2423 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
;
2424 /* If this is the last SKB we copy and eor is set
2425 * we need to propagate it to the new skb.
2427 TCP_SKB_CB(nskb
)->eor
= TCP_SKB_CB(skb
)->eor
;
2428 tcp_skb_collapse_tstamp(nskb
, skb
);
2429 tcp_unlink_write_queue(skb
, sk
);
2430 tcp_wmem_free_skb(sk
, skb
);
2432 TCP_SKB_CB(nskb
)->tcp_flags
|= TCP_SKB_CB(skb
)->tcp_flags
&
2433 ~(TCPHDR_FIN
|TCPHDR_PSH
);
2434 if (!skb_shinfo(skb
)->nr_frags
) {
2435 skb_pull(skb
, copy
);
2437 __pskb_trim_head(skb
, copy
);
2438 tcp_set_skb_tso_segs(skb
, mss_now
);
2440 TCP_SKB_CB(skb
)->seq
+= copy
;
2445 if (len
>= probe_size
)
2448 tcp_init_tso_segs(nskb
, nskb
->len
);
2450 /* We're ready to send. If this fails, the probe will
2451 * be resegmented into mss-sized pieces by tcp_write_xmit().
2453 if (!tcp_transmit_skb(sk
, nskb
, 1, GFP_ATOMIC
)) {
2454 /* Decrement cwnd here because we are sending
2455 * effectively two packets. */
2456 tcp_snd_cwnd_set(tp
, tcp_snd_cwnd(tp
) - 1);
2457 tcp_event_new_data_sent(sk
, nskb
);
2459 icsk
->icsk_mtup
.probe_size
= tcp_mss_to_mtu(sk
, nskb
->len
);
2460 tp
->mtu_probe
.probe_seq_start
= TCP_SKB_CB(nskb
)->seq
;
2461 tp
->mtu_probe
.probe_seq_end
= TCP_SKB_CB(nskb
)->end_seq
;
2469 static bool tcp_pacing_check(struct sock
*sk
)
2471 struct tcp_sock
*tp
= tcp_sk(sk
);
2473 if (!tcp_needs_internal_pacing(sk
))
2476 if (tp
->tcp_wstamp_ns
<= tp
->tcp_clock_cache
)
2479 if (!hrtimer_is_queued(&tp
->pacing_timer
)) {
2480 hrtimer_start(&tp
->pacing_timer
,
2481 ns_to_ktime(tp
->tcp_wstamp_ns
),
2482 HRTIMER_MODE_ABS_PINNED_SOFT
);
2488 /* TCP Small Queues :
2489 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2490 * (These limits are doubled for retransmits)
2492 * - better RTT estimation and ACK scheduling
2495 * Alas, some drivers / subsystems require a fair amount
2496 * of queued bytes to ensure line rate.
2497 * One example is wifi aggregation (802.11 AMPDU)
2499 static bool tcp_small_queue_check(struct sock
*sk
, const struct sk_buff
*skb
,
2500 unsigned int factor
)
2502 unsigned long limit
;
2504 limit
= max_t(unsigned long,
2506 sk
->sk_pacing_rate
>> READ_ONCE(sk
->sk_pacing_shift
));
2507 if (sk
->sk_pacing_status
== SK_PACING_NONE
)
2508 limit
= min_t(unsigned long, limit
,
2509 sock_net(sk
)->ipv4
.sysctl_tcp_limit_output_bytes
);
2512 if (static_branch_unlikely(&tcp_tx_delay_enabled
) &&
2513 tcp_sk(sk
)->tcp_tx_delay
) {
2514 u64 extra_bytes
= (u64
)sk
->sk_pacing_rate
* tcp_sk(sk
)->tcp_tx_delay
;
2516 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2517 * approximate our needs assuming an ~100% skb->truesize overhead.
2518 * USEC_PER_SEC is approximated by 2^20.
2519 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2521 extra_bytes
>>= (20 - 1);
2522 limit
+= extra_bytes
;
2524 if (refcount_read(&sk
->sk_wmem_alloc
) > limit
) {
2525 /* Always send skb if rtx queue is empty.
2526 * No need to wait for TX completion to call us back,
2527 * after softirq/tasklet schedule.
2528 * This helps when TX completions are delayed too much.
2530 if (tcp_rtx_queue_empty(sk
))
2533 set_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
);
2534 /* It is possible TX completion already happened
2535 * before we set TSQ_THROTTLED, so we must
2536 * test again the condition.
2538 smp_mb__after_atomic();
2539 if (refcount_read(&sk
->sk_wmem_alloc
) > limit
)
2545 static void tcp_chrono_set(struct tcp_sock
*tp
, const enum tcp_chrono
new)
2547 const u32 now
= tcp_jiffies32
;
2548 enum tcp_chrono old
= tp
->chrono_type
;
2550 if (old
> TCP_CHRONO_UNSPEC
)
2551 tp
->chrono_stat
[old
- 1] += now
- tp
->chrono_start
;
2552 tp
->chrono_start
= now
;
2553 tp
->chrono_type
= new;
2556 void tcp_chrono_start(struct sock
*sk
, const enum tcp_chrono type
)
2558 struct tcp_sock
*tp
= tcp_sk(sk
);
2560 /* If there are multiple conditions worthy of tracking in a
2561 * chronograph then the highest priority enum takes precedence
2562 * over the other conditions. So that if something "more interesting"
2563 * starts happening, stop the previous chrono and start a new one.
2565 if (type
> tp
->chrono_type
)
2566 tcp_chrono_set(tp
, type
);
2569 void tcp_chrono_stop(struct sock
*sk
, const enum tcp_chrono type
)
2571 struct tcp_sock
*tp
= tcp_sk(sk
);
2574 /* There are multiple conditions worthy of tracking in a
2575 * chronograph, so that the highest priority enum takes
2576 * precedence over the other conditions (see tcp_chrono_start).
2577 * If a condition stops, we only stop chrono tracking if
2578 * it's the "most interesting" or current chrono we are
2579 * tracking and starts busy chrono if we have pending data.
2581 if (tcp_rtx_and_write_queues_empty(sk
))
2582 tcp_chrono_set(tp
, TCP_CHRONO_UNSPEC
);
2583 else if (type
== tp
->chrono_type
)
2584 tcp_chrono_set(tp
, TCP_CHRONO_BUSY
);
2587 /* This routine writes packets to the network. It advances the
2588 * send_head. This happens as incoming acks open up the remote
2591 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2592 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2593 * account rare use of URG, this is not a big flaw.
2595 * Send at most one packet when push_one > 0. Temporarily ignore
2596 * cwnd limit to force at most one packet out when push_one == 2.
2598 * Returns true, if no segments are in flight and we have queued segments,
2599 * but cannot send anything now because of SWS or another problem.
2601 static bool tcp_write_xmit(struct sock
*sk
, unsigned int mss_now
, int nonagle
,
2602 int push_one
, gfp_t gfp
)
2604 struct tcp_sock
*tp
= tcp_sk(sk
);
2605 struct sk_buff
*skb
;
2606 unsigned int tso_segs
, sent_pkts
;
2609 bool is_cwnd_limited
= false, is_rwnd_limited
= false;
2614 tcp_mstamp_refresh(tp
);
2616 /* Do MTU probing. */
2617 result
= tcp_mtu_probe(sk
);
2620 } else if (result
> 0) {
2625 max_segs
= tcp_tso_segs(sk
, mss_now
);
2626 while ((skb
= tcp_send_head(sk
))) {
2629 if (unlikely(tp
->repair
) && tp
->repair_queue
== TCP_SEND_QUEUE
) {
2630 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2631 tp
->tcp_wstamp_ns
= tp
->tcp_clock_cache
;
2632 skb_set_delivery_time(skb
, tp
->tcp_wstamp_ns
, true);
2633 list_move_tail(&skb
->tcp_tsorted_anchor
, &tp
->tsorted_sent_queue
);
2634 tcp_init_tso_segs(skb
, mss_now
);
2635 goto repair
; /* Skip network transmission */
2638 if (tcp_pacing_check(sk
))
2641 tso_segs
= tcp_init_tso_segs(skb
, mss_now
);
2644 cwnd_quota
= tcp_cwnd_test(tp
, skb
);
2647 /* Force out a loss probe pkt. */
2653 if (unlikely(!tcp_snd_wnd_test(tp
, skb
, mss_now
))) {
2654 is_rwnd_limited
= true;
2658 if (tso_segs
== 1) {
2659 if (unlikely(!tcp_nagle_test(tp
, skb
, mss_now
,
2660 (tcp_skb_is_last(sk
, skb
) ?
2661 nonagle
: TCP_NAGLE_PUSH
))))
2665 tcp_tso_should_defer(sk
, skb
, &is_cwnd_limited
,
2666 &is_rwnd_limited
, max_segs
))
2671 if (tso_segs
> 1 && !tcp_urg_mode(tp
))
2672 limit
= tcp_mss_split_point(sk
, skb
, mss_now
,
2678 if (skb
->len
> limit
&&
2679 unlikely(tso_fragment(sk
, skb
, limit
, mss_now
, gfp
)))
2682 if (tcp_small_queue_check(sk
, skb
, 0))
2685 /* Argh, we hit an empty skb(), presumably a thread
2686 * is sleeping in sendmsg()/sk_stream_wait_memory().
2687 * We do not want to send a pure-ack packet and have
2688 * a strange looking rtx queue with empty packet(s).
2690 if (TCP_SKB_CB(skb
)->end_seq
== TCP_SKB_CB(skb
)->seq
)
2693 if (unlikely(tcp_transmit_skb(sk
, skb
, 1, gfp
)))
2697 /* Advance the send_head. This one is sent out.
2698 * This call will increment packets_out.
2700 tcp_event_new_data_sent(sk
, skb
);
2702 tcp_minshall_update(tp
, mss_now
, skb
);
2703 sent_pkts
+= tcp_skb_pcount(skb
);
2709 if (is_rwnd_limited
)
2710 tcp_chrono_start(sk
, TCP_CHRONO_RWND_LIMITED
);
2712 tcp_chrono_stop(sk
, TCP_CHRONO_RWND_LIMITED
);
2714 is_cwnd_limited
|= (tcp_packets_in_flight(tp
) >= tcp_snd_cwnd(tp
));
2715 if (likely(sent_pkts
|| is_cwnd_limited
))
2716 tcp_cwnd_validate(sk
, is_cwnd_limited
);
2718 if (likely(sent_pkts
)) {
2719 if (tcp_in_cwnd_reduction(sk
))
2720 tp
->prr_out
+= sent_pkts
;
2722 /* Send one loss probe per tail loss episode. */
2724 tcp_schedule_loss_probe(sk
, false);
2727 return !tp
->packets_out
&& !tcp_write_queue_empty(sk
);
2730 bool tcp_schedule_loss_probe(struct sock
*sk
, bool advancing_rto
)
2732 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2733 struct tcp_sock
*tp
= tcp_sk(sk
);
2734 u32 timeout
, rto_delta_us
;
2737 /* Don't do any loss probe on a Fast Open connection before 3WHS
2740 if (rcu_access_pointer(tp
->fastopen_rsk
))
2743 early_retrans
= sock_net(sk
)->ipv4
.sysctl_tcp_early_retrans
;
2744 /* Schedule a loss probe in 2*RTT for SACK capable connections
2745 * not in loss recovery, that are either limited by cwnd or application.
2747 if ((early_retrans
!= 3 && early_retrans
!= 4) ||
2748 !tp
->packets_out
|| !tcp_is_sack(tp
) ||
2749 (icsk
->icsk_ca_state
!= TCP_CA_Open
&&
2750 icsk
->icsk_ca_state
!= TCP_CA_CWR
))
2753 /* Probe timeout is 2*rtt. Add minimum RTO to account
2754 * for delayed ack when there's one outstanding packet. If no RTT
2755 * sample is available then probe after TCP_TIMEOUT_INIT.
2758 timeout
= usecs_to_jiffies(tp
->srtt_us
>> 2);
2759 if (tp
->packets_out
== 1)
2760 timeout
+= TCP_RTO_MIN
;
2762 timeout
+= TCP_TIMEOUT_MIN
;
2764 timeout
= TCP_TIMEOUT_INIT
;
2767 /* If the RTO formula yields an earlier time, then use that time. */
2768 rto_delta_us
= advancing_rto
?
2769 jiffies_to_usecs(inet_csk(sk
)->icsk_rto
) :
2770 tcp_rto_delta_us(sk
); /* How far in future is RTO? */
2771 if (rto_delta_us
> 0)
2772 timeout
= min_t(u32
, timeout
, usecs_to_jiffies(rto_delta_us
));
2774 tcp_reset_xmit_timer(sk
, ICSK_TIME_LOSS_PROBE
, timeout
, TCP_RTO_MAX
);
2778 /* Thanks to skb fast clones, we can detect if a prior transmit of
2779 * a packet is still in a qdisc or driver queue.
2780 * In this case, there is very little point doing a retransmit !
2782 static bool skb_still_in_host_queue(struct sock
*sk
,
2783 const struct sk_buff
*skb
)
2785 if (unlikely(skb_fclone_busy(sk
, skb
))) {
2786 set_bit(TSQ_THROTTLED
, &sk
->sk_tsq_flags
);
2787 smp_mb__after_atomic();
2788 if (skb_fclone_busy(sk
, skb
)) {
2789 NET_INC_STATS(sock_net(sk
),
2790 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES
);
2797 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2798 * retransmit the last segment.
2800 void tcp_send_loss_probe(struct sock
*sk
)
2802 struct tcp_sock
*tp
= tcp_sk(sk
);
2803 struct sk_buff
*skb
;
2805 int mss
= tcp_current_mss(sk
);
2807 /* At most one outstanding TLP */
2808 if (tp
->tlp_high_seq
)
2811 tp
->tlp_retrans
= 0;
2812 skb
= tcp_send_head(sk
);
2813 if (skb
&& tcp_snd_wnd_test(tp
, skb
, mss
)) {
2814 pcount
= tp
->packets_out
;
2815 tcp_write_xmit(sk
, mss
, TCP_NAGLE_OFF
, 2, GFP_ATOMIC
);
2816 if (tp
->packets_out
> pcount
)
2820 skb
= skb_rb_last(&sk
->tcp_rtx_queue
);
2821 if (unlikely(!skb
)) {
2822 WARN_ONCE(tp
->packets_out
,
2823 "invalid inflight: %u state %u cwnd %u mss %d\n",
2824 tp
->packets_out
, sk
->sk_state
, tcp_snd_cwnd(tp
), mss
);
2825 inet_csk(sk
)->icsk_pending
= 0;
2829 if (skb_still_in_host_queue(sk
, skb
))
2832 pcount
= tcp_skb_pcount(skb
);
2833 if (WARN_ON(!pcount
))
2836 if ((pcount
> 1) && (skb
->len
> (pcount
- 1) * mss
)) {
2837 if (unlikely(tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
,
2838 (pcount
- 1) * mss
, mss
,
2841 skb
= skb_rb_next(skb
);
2844 if (WARN_ON(!skb
|| !tcp_skb_pcount(skb
)))
2847 if (__tcp_retransmit_skb(sk
, skb
, 1))
2850 tp
->tlp_retrans
= 1;
2853 /* Record snd_nxt for loss detection. */
2854 tp
->tlp_high_seq
= tp
->snd_nxt
;
2856 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPLOSSPROBES
);
2857 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2858 inet_csk(sk
)->icsk_pending
= 0;
2863 /* Push out any pending frames which were held back due to
2864 * TCP_CORK or attempt at coalescing tiny packets.
2865 * The socket must be locked by the caller.
2867 void __tcp_push_pending_frames(struct sock
*sk
, unsigned int cur_mss
,
2870 /* If we are closed, the bytes will have to remain here.
2871 * In time closedown will finish, we empty the write queue and
2872 * all will be happy.
2874 if (unlikely(sk
->sk_state
== TCP_CLOSE
))
2877 if (tcp_write_xmit(sk
, cur_mss
, nonagle
, 0,
2878 sk_gfp_mask(sk
, GFP_ATOMIC
)))
2879 tcp_check_probe_timer(sk
);
2882 /* Send _single_ skb sitting at the send head. This function requires
2883 * true push pending frames to setup probe timer etc.
2885 void tcp_push_one(struct sock
*sk
, unsigned int mss_now
)
2887 struct sk_buff
*skb
= tcp_send_head(sk
);
2889 BUG_ON(!skb
|| skb
->len
< mss_now
);
2891 tcp_write_xmit(sk
, mss_now
, TCP_NAGLE_PUSH
, 1, sk
->sk_allocation
);
2894 /* This function returns the amount that we can raise the
2895 * usable window based on the following constraints
2897 * 1. The window can never be shrunk once it is offered (RFC 793)
2898 * 2. We limit memory per socket
2901 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2902 * RECV.NEXT + RCV.WIN fixed until:
2903 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2905 * i.e. don't raise the right edge of the window until you can raise
2906 * it at least MSS bytes.
2908 * Unfortunately, the recommended algorithm breaks header prediction,
2909 * since header prediction assumes th->window stays fixed.
2911 * Strictly speaking, keeping th->window fixed violates the receiver
2912 * side SWS prevention criteria. The problem is that under this rule
2913 * a stream of single byte packets will cause the right side of the
2914 * window to always advance by a single byte.
2916 * Of course, if the sender implements sender side SWS prevention
2917 * then this will not be a problem.
2919 * BSD seems to make the following compromise:
2921 * If the free space is less than the 1/4 of the maximum
2922 * space available and the free space is less than 1/2 mss,
2923 * then set the window to 0.
2924 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2925 * Otherwise, just prevent the window from shrinking
2926 * and from being larger than the largest representable value.
2928 * This prevents incremental opening of the window in the regime
2929 * where TCP is limited by the speed of the reader side taking
2930 * data out of the TCP receive queue. It does nothing about
2931 * those cases where the window is constrained on the sender side
2932 * because the pipeline is full.
2934 * BSD also seems to "accidentally" limit itself to windows that are a
2935 * multiple of MSS, at least until the free space gets quite small.
2936 * This would appear to be a side effect of the mbuf implementation.
2937 * Combining these two algorithms results in the observed behavior
2938 * of having a fixed window size at almost all times.
2940 * Below we obtain similar behavior by forcing the offered window to
2941 * a multiple of the mss when it is feasible to do so.
2943 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2944 * Regular options like TIMESTAMP are taken into account.
2946 u32
__tcp_select_window(struct sock
*sk
)
2948 struct inet_connection_sock
*icsk
= inet_csk(sk
);
2949 struct tcp_sock
*tp
= tcp_sk(sk
);
2950 /* MSS for the peer's data. Previous versions used mss_clamp
2951 * here. I don't know if the value based on our guesses
2952 * of peer's MSS is better for the performance. It's more correct
2953 * but may be worse for the performance because of rcv_mss
2954 * fluctuations. --SAW 1998/11/1
2956 int mss
= icsk
->icsk_ack
.rcv_mss
;
2957 int free_space
= tcp_space(sk
);
2958 int allowed_space
= tcp_full_space(sk
);
2959 int full_space
, window
;
2961 if (sk_is_mptcp(sk
))
2962 mptcp_space(sk
, &free_space
, &allowed_space
);
2964 full_space
= min_t(int, tp
->window_clamp
, allowed_space
);
2966 if (unlikely(mss
> full_space
)) {
2971 if (free_space
< (full_space
>> 1)) {
2972 icsk
->icsk_ack
.quick
= 0;
2974 if (tcp_under_memory_pressure(sk
))
2975 tcp_adjust_rcv_ssthresh(sk
);
2977 /* free_space might become our new window, make sure we don't
2978 * increase it due to wscale.
2980 free_space
= round_down(free_space
, 1 << tp
->rx_opt
.rcv_wscale
);
2982 /* if free space is less than mss estimate, or is below 1/16th
2983 * of the maximum allowed, try to move to zero-window, else
2984 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2985 * new incoming data is dropped due to memory limits.
2986 * With large window, mss test triggers way too late in order
2987 * to announce zero window in time before rmem limit kicks in.
2989 if (free_space
< (allowed_space
>> 4) || free_space
< mss
)
2993 if (free_space
> tp
->rcv_ssthresh
)
2994 free_space
= tp
->rcv_ssthresh
;
2996 /* Don't do rounding if we are using window scaling, since the
2997 * scaled window will not line up with the MSS boundary anyway.
2999 if (tp
->rx_opt
.rcv_wscale
) {
3000 window
= free_space
;
3002 /* Advertise enough space so that it won't get scaled away.
3003 * Import case: prevent zero window announcement if
3004 * 1<<rcv_wscale > mss.
3006 window
= ALIGN(window
, (1 << tp
->rx_opt
.rcv_wscale
));
3008 window
= tp
->rcv_wnd
;
3009 /* Get the largest window that is a nice multiple of mss.
3010 * Window clamp already applied above.
3011 * If our current window offering is within 1 mss of the
3012 * free space we just keep it. This prevents the divide
3013 * and multiply from happening most of the time.
3014 * We also don't do any window rounding when the free space
3017 if (window
<= free_space
- mss
|| window
> free_space
)
3018 window
= rounddown(free_space
, mss
);
3019 else if (mss
== full_space
&&
3020 free_space
> window
+ (full_space
>> 1))
3021 window
= free_space
;
3027 void tcp_skb_collapse_tstamp(struct sk_buff
*skb
,
3028 const struct sk_buff
*next_skb
)
3030 if (unlikely(tcp_has_tx_tstamp(next_skb
))) {
3031 const struct skb_shared_info
*next_shinfo
=
3032 skb_shinfo(next_skb
);
3033 struct skb_shared_info
*shinfo
= skb_shinfo(skb
);
3035 shinfo
->tx_flags
|= next_shinfo
->tx_flags
& SKBTX_ANY_TSTAMP
;
3036 shinfo
->tskey
= next_shinfo
->tskey
;
3037 TCP_SKB_CB(skb
)->txstamp_ack
|=
3038 TCP_SKB_CB(next_skb
)->txstamp_ack
;
3042 /* Collapses two adjacent SKB's during retransmission. */
3043 static bool tcp_collapse_retrans(struct sock
*sk
, struct sk_buff
*skb
)
3045 struct tcp_sock
*tp
= tcp_sk(sk
);
3046 struct sk_buff
*next_skb
= skb_rb_next(skb
);
3049 next_skb_size
= next_skb
->len
;
3051 BUG_ON(tcp_skb_pcount(skb
) != 1 || tcp_skb_pcount(next_skb
) != 1);
3053 if (next_skb_size
&& !tcp_skb_shift(skb
, next_skb
, 1, next_skb_size
))
3056 tcp_highest_sack_replace(sk
, next_skb
, skb
);
3058 /* Update sequence range on original skb. */
3059 TCP_SKB_CB(skb
)->end_seq
= TCP_SKB_CB(next_skb
)->end_seq
;
3061 /* Merge over control information. This moves PSH/FIN etc. over */
3062 TCP_SKB_CB(skb
)->tcp_flags
|= TCP_SKB_CB(next_skb
)->tcp_flags
;
3064 /* All done, get rid of second SKB and account for it so
3065 * packet counting does not break.
3067 TCP_SKB_CB(skb
)->sacked
|= TCP_SKB_CB(next_skb
)->sacked
& TCPCB_EVER_RETRANS
;
3068 TCP_SKB_CB(skb
)->eor
= TCP_SKB_CB(next_skb
)->eor
;
3070 /* changed transmit queue under us so clear hints */
3071 tcp_clear_retrans_hints_partial(tp
);
3072 if (next_skb
== tp
->retransmit_skb_hint
)
3073 tp
->retransmit_skb_hint
= skb
;
3075 tcp_adjust_pcount(sk
, next_skb
, tcp_skb_pcount(next_skb
));
3077 tcp_skb_collapse_tstamp(skb
, next_skb
);
3079 tcp_rtx_queue_unlink_and_free(next_skb
, sk
);
3083 /* Check if coalescing SKBs is legal. */
3084 static bool tcp_can_collapse(const struct sock
*sk
, const struct sk_buff
*skb
)
3086 if (tcp_skb_pcount(skb
) > 1)
3088 if (skb_cloned(skb
))
3090 /* Some heuristics for collapsing over SACK'd could be invented */
3091 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_ACKED
)
3097 /* Collapse packets in the retransmit queue to make to create
3098 * less packets on the wire. This is only done on retransmission.
3100 static void tcp_retrans_try_collapse(struct sock
*sk
, struct sk_buff
*to
,
3103 struct tcp_sock
*tp
= tcp_sk(sk
);
3104 struct sk_buff
*skb
= to
, *tmp
;
3107 if (!sock_net(sk
)->ipv4
.sysctl_tcp_retrans_collapse
)
3109 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
3112 skb_rbtree_walk_from_safe(skb
, tmp
) {
3113 if (!tcp_can_collapse(sk
, skb
))
3116 if (!tcp_skb_can_collapse(to
, skb
))
3129 if (after(TCP_SKB_CB(skb
)->end_seq
, tcp_wnd_end(tp
)))
3132 if (!tcp_collapse_retrans(sk
, to
))
3137 /* This retransmits one SKB. Policy decisions and retransmit queue
3138 * state updates are done by the caller. Returns non-zero if an
3139 * error occurred which prevented the send.
3141 int __tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
3143 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3144 struct tcp_sock
*tp
= tcp_sk(sk
);
3145 unsigned int cur_mss
;
3149 /* Inconclusive MTU probe */
3150 if (icsk
->icsk_mtup
.probe_size
)
3151 icsk
->icsk_mtup
.probe_size
= 0;
3153 if (skb_still_in_host_queue(sk
, skb
))
3156 if (before(TCP_SKB_CB(skb
)->seq
, tp
->snd_una
)) {
3157 if (unlikely(before(TCP_SKB_CB(skb
)->end_seq
, tp
->snd_una
))) {
3161 if (tcp_trim_head(sk
, skb
, tp
->snd_una
- TCP_SKB_CB(skb
)->seq
))
3165 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
3166 return -EHOSTUNREACH
; /* Routing failure or similar. */
3168 cur_mss
= tcp_current_mss(sk
);
3170 /* If receiver has shrunk his window, and skb is out of
3171 * new window, do not retransmit it. The exception is the
3172 * case, when window is shrunk to zero. In this case
3173 * our retransmit serves as a zero window probe.
3175 if (!before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
)) &&
3176 TCP_SKB_CB(skb
)->seq
!= tp
->snd_una
)
3179 len
= cur_mss
* segs
;
3180 if (skb
->len
> len
) {
3181 if (tcp_fragment(sk
, TCP_FRAG_IN_RTX_QUEUE
, skb
, len
,
3182 cur_mss
, GFP_ATOMIC
))
3183 return -ENOMEM
; /* We'll try again later. */
3185 if (skb_unclone_keeptruesize(skb
, GFP_ATOMIC
))
3188 diff
= tcp_skb_pcount(skb
);
3189 tcp_set_skb_tso_segs(skb
, cur_mss
);
3190 diff
-= tcp_skb_pcount(skb
);
3192 tcp_adjust_pcount(sk
, skb
, diff
);
3193 if (skb
->len
< cur_mss
)
3194 tcp_retrans_try_collapse(sk
, skb
, cur_mss
);
3197 /* RFC3168, section 6.1.1.1. ECN fallback */
3198 if ((TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN_ECN
) == TCPHDR_SYN_ECN
)
3199 tcp_ecn_clear_syn(sk
, skb
);
3201 /* Update global and local TCP statistics. */
3202 segs
= tcp_skb_pcount(skb
);
3203 TCP_ADD_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
, segs
);
3204 if (TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)
3205 __NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
3206 tp
->total_retrans
+= segs
;
3207 tp
->bytes_retrans
+= skb
->len
;
3209 /* make sure skb->data is aligned on arches that require it
3210 * and check if ack-trimming & collapsing extended the headroom
3211 * beyond what csum_start can cover.
3213 if (unlikely((NET_IP_ALIGN
&& ((unsigned long)skb
->data
& 3)) ||
3214 skb_headroom(skb
) >= 0xFFFF)) {
3215 struct sk_buff
*nskb
;
3217 tcp_skb_tsorted_save(skb
) {
3218 nskb
= __pskb_copy(skb
, MAX_TCP_HEADER
, GFP_ATOMIC
);
3221 err
= tcp_transmit_skb(sk
, nskb
, 0, GFP_ATOMIC
);
3225 } tcp_skb_tsorted_restore(skb
);
3228 tcp_update_skb_after_send(sk
, skb
, tp
->tcp_wstamp_ns
);
3229 tcp_rate_skb_sent(sk
, skb
);
3232 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3235 /* To avoid taking spuriously low RTT samples based on a timestamp
3236 * for a transmit that never happened, always mark EVER_RETRANS
3238 TCP_SKB_CB(skb
)->sacked
|= TCPCB_EVER_RETRANS
;
3240 if (BPF_SOCK_OPS_TEST_FLAG(tp
, BPF_SOCK_OPS_RETRANS_CB_FLAG
))
3241 tcp_call_bpf_3arg(sk
, BPF_SOCK_OPS_RETRANS_CB
,
3242 TCP_SKB_CB(skb
)->seq
, segs
, err
);
3245 trace_tcp_retransmit_skb(sk
, skb
);
3246 } else if (err
!= -EBUSY
) {
3247 NET_ADD_STATS(sock_net(sk
), LINUX_MIB_TCPRETRANSFAIL
, segs
);
3252 int tcp_retransmit_skb(struct sock
*sk
, struct sk_buff
*skb
, int segs
)
3254 struct tcp_sock
*tp
= tcp_sk(sk
);
3255 int err
= __tcp_retransmit_skb(sk
, skb
, segs
);
3258 #if FASTRETRANS_DEBUG > 0
3259 if (TCP_SKB_CB(skb
)->sacked
& TCPCB_SACKED_RETRANS
) {
3260 net_dbg_ratelimited("retrans_out leaked\n");
3263 TCP_SKB_CB(skb
)->sacked
|= TCPCB_RETRANS
;
3264 tp
->retrans_out
+= tcp_skb_pcount(skb
);
3267 /* Save stamp of the first (attempted) retransmit. */
3268 if (!tp
->retrans_stamp
)
3269 tp
->retrans_stamp
= tcp_skb_timestamp(skb
);
3271 if (tp
->undo_retrans
< 0)
3272 tp
->undo_retrans
= 0;
3273 tp
->undo_retrans
+= tcp_skb_pcount(skb
);
3277 /* This gets called after a retransmit timeout, and the initially
3278 * retransmitted data is acknowledged. It tries to continue
3279 * resending the rest of the retransmit queue, until either
3280 * we've sent it all or the congestion window limit is reached.
3282 void tcp_xmit_retransmit_queue(struct sock
*sk
)
3284 const struct inet_connection_sock
*icsk
= inet_csk(sk
);
3285 struct sk_buff
*skb
, *rtx_head
, *hole
= NULL
;
3286 struct tcp_sock
*tp
= tcp_sk(sk
);
3287 bool rearm_timer
= false;
3291 if (!tp
->packets_out
)
3294 rtx_head
= tcp_rtx_queue_head(sk
);
3295 skb
= tp
->retransmit_skb_hint
?: rtx_head
;
3296 max_segs
= tcp_tso_segs(sk
, tcp_current_mss(sk
));
3297 skb_rbtree_walk_from(skb
) {
3301 if (tcp_pacing_check(sk
))
3304 /* we could do better than to assign each time */
3306 tp
->retransmit_skb_hint
= skb
;
3308 segs
= tcp_snd_cwnd(tp
) - tcp_packets_in_flight(tp
);
3311 sacked
= TCP_SKB_CB(skb
)->sacked
;
3312 /* In case tcp_shift_skb_data() have aggregated large skbs,
3313 * we need to make sure not sending too bigs TSO packets
3315 segs
= min_t(int, segs
, max_segs
);
3317 if (tp
->retrans_out
>= tp
->lost_out
) {
3319 } else if (!(sacked
& TCPCB_LOST
)) {
3320 if (!hole
&& !(sacked
& (TCPCB_SACKED_RETRANS
|TCPCB_SACKED_ACKED
)))
3325 if (icsk
->icsk_ca_state
!= TCP_CA_Loss
)
3326 mib_idx
= LINUX_MIB_TCPFASTRETRANS
;
3328 mib_idx
= LINUX_MIB_TCPSLOWSTARTRETRANS
;
3331 if (sacked
& (TCPCB_SACKED_ACKED
|TCPCB_SACKED_RETRANS
))
3334 if (tcp_small_queue_check(sk
, skb
, 1))
3337 if (tcp_retransmit_skb(sk
, skb
, segs
))
3340 NET_ADD_STATS(sock_net(sk
), mib_idx
, tcp_skb_pcount(skb
));
3342 if (tcp_in_cwnd_reduction(sk
))
3343 tp
->prr_out
+= tcp_skb_pcount(skb
);
3345 if (skb
== rtx_head
&&
3346 icsk
->icsk_pending
!= ICSK_TIME_REO_TIMEOUT
)
3351 tcp_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3352 inet_csk(sk
)->icsk_rto
,
3356 /* We allow to exceed memory limits for FIN packets to expedite
3357 * connection tear down and (memory) recovery.
3358 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3359 * or even be forced to close flow without any FIN.
3360 * In general, we want to allow one skb per socket to avoid hangs
3361 * with edge trigger epoll()
3363 void sk_forced_mem_schedule(struct sock
*sk
, int size
)
3367 if (size
<= sk
->sk_forward_alloc
)
3369 amt
= sk_mem_pages(size
);
3370 sk
->sk_forward_alloc
+= amt
* SK_MEM_QUANTUM
;
3371 sk_memory_allocated_add(sk
, amt
);
3373 if (mem_cgroup_sockets_enabled
&& sk
->sk_memcg
)
3374 mem_cgroup_charge_skmem(sk
->sk_memcg
, amt
,
3375 gfp_memcg_charge() | __GFP_NOFAIL
);
3378 /* Send a FIN. The caller locks the socket for us.
3379 * We should try to send a FIN packet really hard, but eventually give up.
3381 void tcp_send_fin(struct sock
*sk
)
3383 struct sk_buff
*skb
, *tskb
, *tail
= tcp_write_queue_tail(sk
);
3384 struct tcp_sock
*tp
= tcp_sk(sk
);
3386 /* Optimization, tack on the FIN if we have one skb in write queue and
3387 * this skb was not yet sent, or we are under memory pressure.
3388 * Note: in the latter case, FIN packet will be sent after a timeout,
3389 * as TCP stack thinks it has already been transmitted.
3392 if (!tskb
&& tcp_under_memory_pressure(sk
))
3393 tskb
= skb_rb_last(&sk
->tcp_rtx_queue
);
3396 TCP_SKB_CB(tskb
)->tcp_flags
|= TCPHDR_FIN
;
3397 TCP_SKB_CB(tskb
)->end_seq
++;
3400 /* This means tskb was already sent.
3401 * Pretend we included the FIN on previous transmit.
3402 * We need to set tp->snd_nxt to the value it would have
3403 * if FIN had been sent. This is because retransmit path
3404 * does not change tp->snd_nxt.
3406 WRITE_ONCE(tp
->snd_nxt
, tp
->snd_nxt
+ 1);
3410 skb
= alloc_skb_fclone(MAX_TCP_HEADER
, sk
->sk_allocation
);
3414 INIT_LIST_HEAD(&skb
->tcp_tsorted_anchor
);
3415 skb_reserve(skb
, MAX_TCP_HEADER
);
3416 sk_forced_mem_schedule(sk
, skb
->truesize
);
3417 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3418 tcp_init_nondata_skb(skb
, tp
->write_seq
,
3419 TCPHDR_ACK
| TCPHDR_FIN
);
3420 tcp_queue_skb(sk
, skb
);
3422 __tcp_push_pending_frames(sk
, tcp_current_mss(sk
), TCP_NAGLE_OFF
);
3425 /* We get here when a process closes a file descriptor (either due to
3426 * an explicit close() or as a byproduct of exit()'ing) and there
3427 * was unread data in the receive queue. This behavior is recommended
3428 * by RFC 2525, section 2.17. -DaveM
3430 void tcp_send_active_reset(struct sock
*sk
, gfp_t priority
)
3432 struct sk_buff
*skb
;
3434 TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTRSTS
);
3436 /* NOTE: No TCP options attached and we never retransmit this. */
3437 skb
= alloc_skb(MAX_TCP_HEADER
, priority
);
3439 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3443 /* Reserve space for headers and prepare control bits. */
3444 skb_reserve(skb
, MAX_TCP_HEADER
);
3445 tcp_init_nondata_skb(skb
, tcp_acceptable_seq(sk
),
3446 TCPHDR_ACK
| TCPHDR_RST
);
3447 tcp_mstamp_refresh(tcp_sk(sk
));
3449 if (tcp_transmit_skb(sk
, skb
, 0, priority
))
3450 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPABORTFAILED
);
3452 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3453 * skb here is different to the troublesome skb, so use NULL
3455 trace_tcp_send_reset(sk
, NULL
);
3458 /* Send a crossed SYN-ACK during socket establishment.
3459 * WARNING: This routine must only be called when we have already sent
3460 * a SYN packet that crossed the incoming SYN that caused this routine
3461 * to get called. If this assumption fails then the initial rcv_wnd
3462 * and rcv_wscale values will not be correct.
3464 int tcp_send_synack(struct sock
*sk
)
3466 struct sk_buff
*skb
;
3468 skb
= tcp_rtx_queue_head(sk
);
3469 if (!skb
|| !(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_SYN
)) {
3470 pr_err("%s: wrong queue state\n", __func__
);
3473 if (!(TCP_SKB_CB(skb
)->tcp_flags
& TCPHDR_ACK
)) {
3474 if (skb_cloned(skb
)) {
3475 struct sk_buff
*nskb
;
3477 tcp_skb_tsorted_save(skb
) {
3478 nskb
= skb_copy(skb
, GFP_ATOMIC
);
3479 } tcp_skb_tsorted_restore(skb
);
3482 INIT_LIST_HEAD(&nskb
->tcp_tsorted_anchor
);
3483 tcp_highest_sack_replace(sk
, skb
, nskb
);
3484 tcp_rtx_queue_unlink_and_free(skb
, sk
);
3485 __skb_header_release(nskb
);
3486 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, nskb
);
3487 sk_wmem_queued_add(sk
, nskb
->truesize
);
3488 sk_mem_charge(sk
, nskb
->truesize
);
3492 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_ACK
;
3493 tcp_ecn_send_synack(sk
, skb
);
3495 return tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
3499 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3500 * @sk: listener socket
3501 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3502 * should not use it again.
3503 * @req: request_sock pointer
3504 * @foc: cookie for tcp fast open
3505 * @synack_type: Type of synack to prepare
3506 * @syn_skb: SYN packet just received. It could be NULL for rtx case.
3508 struct sk_buff
*tcp_make_synack(const struct sock
*sk
, struct dst_entry
*dst
,
3509 struct request_sock
*req
,
3510 struct tcp_fastopen_cookie
*foc
,
3511 enum tcp_synack_type synack_type
,
3512 struct sk_buff
*syn_skb
)
3514 struct inet_request_sock
*ireq
= inet_rsk(req
);
3515 const struct tcp_sock
*tp
= tcp_sk(sk
);
3516 struct tcp_md5sig_key
*md5
= NULL
;
3517 struct tcp_out_options opts
;
3518 struct sk_buff
*skb
;
3519 int tcp_header_size
;
3524 skb
= alloc_skb(MAX_TCP_HEADER
, GFP_ATOMIC
);
3525 if (unlikely(!skb
)) {
3529 /* Reserve space for headers. */
3530 skb_reserve(skb
, MAX_TCP_HEADER
);
3532 switch (synack_type
) {
3533 case TCP_SYNACK_NORMAL
:
3534 skb_set_owner_w(skb
, req_to_sk(req
));
3536 case TCP_SYNACK_COOKIE
:
3537 /* Under synflood, we do not attach skb to a socket,
3538 * to avoid false sharing.
3541 case TCP_SYNACK_FASTOPEN
:
3542 /* sk is a const pointer, because we want to express multiple
3543 * cpu might call us concurrently.
3544 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3546 skb_set_owner_w(skb
, (struct sock
*)sk
);
3549 skb_dst_set(skb
, dst
);
3551 mss
= tcp_mss_clamp(tp
, dst_metric_advmss(dst
));
3553 memset(&opts
, 0, sizeof(opts
));
3554 now
= tcp_clock_ns();
3555 #ifdef CONFIG_SYN_COOKIES
3556 if (unlikely(synack_type
== TCP_SYNACK_COOKIE
&& ireq
->tstamp_ok
))
3557 skb_set_delivery_time(skb
, cookie_init_timestamp(req
, now
),
3562 skb_set_delivery_time(skb
, now
, true);
3563 if (!tcp_rsk(req
)->snt_synack
) /* Timestamp first SYNACK */
3564 tcp_rsk(req
)->snt_synack
= tcp_skb_timestamp_us(skb
);
3567 #ifdef CONFIG_TCP_MD5SIG
3569 md5
= tcp_rsk(req
)->af_specific
->req_md5_lookup(sk
, req_to_sk(req
));
3571 skb_set_hash(skb
, tcp_rsk(req
)->txhash
, PKT_HASH_TYPE_L4
);
3572 /* bpf program will be interested in the tcp_flags */
3573 TCP_SKB_CB(skb
)->tcp_flags
= TCPHDR_SYN
| TCPHDR_ACK
;
3574 tcp_header_size
= tcp_synack_options(sk
, req
, mss
, skb
, &opts
, md5
,
3576 syn_skb
) + sizeof(*th
);
3578 skb_push(skb
, tcp_header_size
);
3579 skb_reset_transport_header(skb
);
3581 th
= (struct tcphdr
*)skb
->data
;
3582 memset(th
, 0, sizeof(struct tcphdr
));
3585 tcp_ecn_make_synack(req
, th
);
3586 th
->source
= htons(ireq
->ir_num
);
3587 th
->dest
= ireq
->ir_rmt_port
;
3588 skb
->mark
= ireq
->ir_mark
;
3589 skb
->ip_summed
= CHECKSUM_PARTIAL
;
3590 th
->seq
= htonl(tcp_rsk(req
)->snt_isn
);
3591 /* XXX data is queued and acked as is. No buffer/window check */
3592 th
->ack_seq
= htonl(tcp_rsk(req
)->rcv_nxt
);
3594 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3595 th
->window
= htons(min(req
->rsk_rcv_wnd
, 65535U));
3596 tcp_options_write(th
, NULL
, &opts
);
3597 th
->doff
= (tcp_header_size
>> 2);
3598 __TCP_INC_STATS(sock_net(sk
), TCP_MIB_OUTSEGS
);
3600 #ifdef CONFIG_TCP_MD5SIG
3601 /* Okay, we have all we need - do the md5 hash if needed */
3603 tcp_rsk(req
)->af_specific
->calc_md5_hash(opts
.hash_location
,
3604 md5
, req_to_sk(req
), skb
);
3608 bpf_skops_write_hdr_opt((struct sock
*)sk
, skb
, req
, syn_skb
,
3609 synack_type
, &opts
);
3611 skb_set_delivery_time(skb
, now
, true);
3612 tcp_add_tx_delay(skb
, tp
);
3616 EXPORT_SYMBOL(tcp_make_synack
);
3618 static void tcp_ca_dst_init(struct sock
*sk
, const struct dst_entry
*dst
)
3620 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3621 const struct tcp_congestion_ops
*ca
;
3622 u32 ca_key
= dst_metric(dst
, RTAX_CC_ALGO
);
3624 if (ca_key
== TCP_CA_UNSPEC
)
3628 ca
= tcp_ca_find_key(ca_key
);
3629 if (likely(ca
&& bpf_try_module_get(ca
, ca
->owner
))) {
3630 bpf_module_put(icsk
->icsk_ca_ops
, icsk
->icsk_ca_ops
->owner
);
3631 icsk
->icsk_ca_dst_locked
= tcp_ca_dst_locked(dst
);
3632 icsk
->icsk_ca_ops
= ca
;
3637 /* Do all connect socket setups that can be done AF independent. */
3638 static void tcp_connect_init(struct sock
*sk
)
3640 const struct dst_entry
*dst
= __sk_dst_get(sk
);
3641 struct tcp_sock
*tp
= tcp_sk(sk
);
3645 /* We'll fix this up when we get a response from the other end.
3646 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3648 tp
->tcp_header_len
= sizeof(struct tcphdr
);
3649 if (sock_net(sk
)->ipv4
.sysctl_tcp_timestamps
)
3650 tp
->tcp_header_len
+= TCPOLEN_TSTAMP_ALIGNED
;
3652 #ifdef CONFIG_TCP_MD5SIG
3653 if (tp
->af_specific
->md5_lookup(sk
, sk
))
3654 tp
->tcp_header_len
+= TCPOLEN_MD5SIG_ALIGNED
;
3657 /* If user gave his TCP_MAXSEG, record it to clamp */
3658 if (tp
->rx_opt
.user_mss
)
3659 tp
->rx_opt
.mss_clamp
= tp
->rx_opt
.user_mss
;
3662 tcp_sync_mss(sk
, dst_mtu(dst
));
3664 tcp_ca_dst_init(sk
, dst
);
3666 if (!tp
->window_clamp
)
3667 tp
->window_clamp
= dst_metric(dst
, RTAX_WINDOW
);
3668 tp
->advmss
= tcp_mss_clamp(tp
, dst_metric_advmss(dst
));
3670 tcp_initialize_rcv_mss(sk
);
3672 /* limit the window selection if the user enforce a smaller rx buffer */
3673 if (sk
->sk_userlocks
& SOCK_RCVBUF_LOCK
&&
3674 (tp
->window_clamp
> tcp_full_space(sk
) || tp
->window_clamp
== 0))
3675 tp
->window_clamp
= tcp_full_space(sk
);
3677 rcv_wnd
= tcp_rwnd_init_bpf(sk
);
3679 rcv_wnd
= dst_metric(dst
, RTAX_INITRWND
);
3681 tcp_select_initial_window(sk
, tcp_full_space(sk
),
3682 tp
->advmss
- (tp
->rx_opt
.ts_recent_stamp
? tp
->tcp_header_len
- sizeof(struct tcphdr
) : 0),
3685 sock_net(sk
)->ipv4
.sysctl_tcp_window_scaling
,
3689 tp
->rx_opt
.rcv_wscale
= rcv_wscale
;
3690 tp
->rcv_ssthresh
= tp
->rcv_wnd
;
3693 sock_reset_flag(sk
, SOCK_DONE
);
3696 tcp_write_queue_purge(sk
);
3697 tp
->snd_una
= tp
->write_seq
;
3698 tp
->snd_sml
= tp
->write_seq
;
3699 tp
->snd_up
= tp
->write_seq
;
3700 WRITE_ONCE(tp
->snd_nxt
, tp
->write_seq
);
3702 if (likely(!tp
->repair
))
3705 tp
->rcv_tstamp
= tcp_jiffies32
;
3706 tp
->rcv_wup
= tp
->rcv_nxt
;
3707 WRITE_ONCE(tp
->copied_seq
, tp
->rcv_nxt
);
3709 inet_csk(sk
)->icsk_rto
= tcp_timeout_init(sk
);
3710 inet_csk(sk
)->icsk_retransmits
= 0;
3711 tcp_clear_retrans(tp
);
3714 static void tcp_connect_queue_skb(struct sock
*sk
, struct sk_buff
*skb
)
3716 struct tcp_sock
*tp
= tcp_sk(sk
);
3717 struct tcp_skb_cb
*tcb
= TCP_SKB_CB(skb
);
3719 tcb
->end_seq
+= skb
->len
;
3720 __skb_header_release(skb
);
3721 sk_wmem_queued_add(sk
, skb
->truesize
);
3722 sk_mem_charge(sk
, skb
->truesize
);
3723 WRITE_ONCE(tp
->write_seq
, tcb
->end_seq
);
3724 tp
->packets_out
+= tcp_skb_pcount(skb
);
3727 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3728 * queue a data-only packet after the regular SYN, such that regular SYNs
3729 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3730 * only the SYN sequence, the data are retransmitted in the first ACK.
3731 * If cookie is not cached or other error occurs, falls back to send a
3732 * regular SYN with Fast Open cookie request option.
3734 static int tcp_send_syn_data(struct sock
*sk
, struct sk_buff
*syn
)
3736 struct inet_connection_sock
*icsk
= inet_csk(sk
);
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
);
3751 /* Sync mss_cache after updating the mss_clamp */
3752 tcp_sync_mss(sk
, icsk
->icsk_pmtu_cookie
);
3754 space
= __tcp_mtu_to_mss(sk
, icsk
->icsk_pmtu_cookie
) -
3755 MAX_TCP_OPTION_SPACE
;
3757 space
= min_t(size_t, space
, fo
->size
);
3759 /* limit to order-0 allocations */
3760 space
= min_t(size_t, space
, SKB_MAX_HEAD(MAX_TCP_HEADER
));
3762 syn_data
= tcp_stream_alloc_skb(sk
, space
, sk
->sk_allocation
, false);
3765 memcpy(syn_data
->cb
, syn
->cb
, sizeof(syn
->cb
));
3767 int copied
= copy_from_iter(skb_put(syn_data
, space
), space
,
3768 &fo
->data
->msg_iter
);
3769 if (unlikely(!copied
)) {
3770 tcp_skb_tsorted_anchor_cleanup(syn_data
);
3771 kfree_skb(syn_data
);
3774 if (copied
!= space
) {
3775 skb_trim(syn_data
, copied
);
3778 skb_zcopy_set(syn_data
, fo
->uarg
, NULL
);
3780 /* No more data pending in inet_wait_for_connect() */
3781 if (space
== fo
->size
)
3785 tcp_connect_queue_skb(sk
, syn_data
);
3787 tcp_chrono_start(sk
, TCP_CHRONO_BUSY
);
3789 err
= tcp_transmit_skb(sk
, syn_data
, 1, sk
->sk_allocation
);
3791 skb_set_delivery_time(syn
, syn_data
->skb_mstamp_ns
, true);
3793 /* Now full SYN+DATA was cloned and sent (or not),
3794 * remove the SYN from the original skb (syn_data)
3795 * we keep in write queue in case of a retransmit, as we
3796 * also have the SYN packet (with no data) in the same queue.
3798 TCP_SKB_CB(syn_data
)->seq
++;
3799 TCP_SKB_CB(syn_data
)->tcp_flags
= TCPHDR_ACK
| TCPHDR_PSH
;
3801 tp
->syn_data
= (fo
->copied
> 0);
3802 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, syn_data
);
3803 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPORIGDATASENT
);
3807 /* data was not sent, put it in write_queue */
3808 __skb_queue_tail(&sk
->sk_write_queue
, syn_data
);
3809 tp
->packets_out
-= tcp_skb_pcount(syn_data
);
3812 /* Send a regular SYN with Fast Open cookie request option */
3813 if (fo
->cookie
.len
> 0)
3815 err
= tcp_transmit_skb(sk
, syn
, 1, sk
->sk_allocation
);
3817 tp
->syn_fastopen
= 0;
3819 fo
->cookie
.len
= -1; /* Exclude Fast Open option for SYN retries */
3823 /* Build a SYN and send it off. */
3824 int tcp_connect(struct sock
*sk
)
3826 struct tcp_sock
*tp
= tcp_sk(sk
);
3827 struct sk_buff
*buff
;
3830 tcp_call_bpf(sk
, BPF_SOCK_OPS_TCP_CONNECT_CB
, 0, NULL
);
3832 if (inet_csk(sk
)->icsk_af_ops
->rebuild_header(sk
))
3833 return -EHOSTUNREACH
; /* Routing failure or similar. */
3835 tcp_connect_init(sk
);
3837 if (unlikely(tp
->repair
)) {
3838 tcp_finish_connect(sk
, NULL
);
3842 buff
= tcp_stream_alloc_skb(sk
, 0, sk
->sk_allocation
, true);
3843 if (unlikely(!buff
))
3846 tcp_init_nondata_skb(buff
, tp
->write_seq
++, TCPHDR_SYN
);
3847 tcp_mstamp_refresh(tp
);
3848 tp
->retrans_stamp
= tcp_time_stamp(tp
);
3849 tcp_connect_queue_skb(sk
, buff
);
3850 tcp_ecn_send_syn(sk
, buff
);
3851 tcp_rbtree_insert(&sk
->tcp_rtx_queue
, buff
);
3853 /* Send off SYN; include data in Fast Open. */
3854 err
= tp
->fastopen_req
? tcp_send_syn_data(sk
, buff
) :
3855 tcp_transmit_skb(sk
, buff
, 1, sk
->sk_allocation
);
3856 if (err
== -ECONNREFUSED
)
3859 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3860 * in order to make this packet get counted in tcpOutSegs.
3862 WRITE_ONCE(tp
->snd_nxt
, tp
->write_seq
);
3863 tp
->pushed_seq
= tp
->write_seq
;
3864 buff
= tcp_send_head(sk
);
3865 if (unlikely(buff
)) {
3866 WRITE_ONCE(tp
->snd_nxt
, TCP_SKB_CB(buff
)->seq
);
3867 tp
->pushed_seq
= TCP_SKB_CB(buff
)->seq
;
3869 TCP_INC_STATS(sock_net(sk
), TCP_MIB_ACTIVEOPENS
);
3871 /* Timer for repeating the SYN until an answer. */
3872 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_RETRANS
,
3873 inet_csk(sk
)->icsk_rto
, TCP_RTO_MAX
);
3876 EXPORT_SYMBOL(tcp_connect
);
3878 /* Send out a delayed ack, the caller does the policy checking
3879 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3882 void tcp_send_delayed_ack(struct sock
*sk
)
3884 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3885 int ato
= icsk
->icsk_ack
.ato
;
3886 unsigned long timeout
;
3888 if (ato
> TCP_DELACK_MIN
) {
3889 const struct tcp_sock
*tp
= tcp_sk(sk
);
3890 int max_ato
= HZ
/ 2;
3892 if (inet_csk_in_pingpong_mode(sk
) ||
3893 (icsk
->icsk_ack
.pending
& ICSK_ACK_PUSHED
))
3894 max_ato
= TCP_DELACK_MAX
;
3896 /* Slow path, intersegment interval is "high". */
3898 /* If some rtt estimate is known, use it to bound delayed ack.
3899 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3903 int rtt
= max_t(int, usecs_to_jiffies(tp
->srtt_us
>> 3),
3910 ato
= min(ato
, max_ato
);
3913 ato
= min_t(u32
, ato
, inet_csk(sk
)->icsk_delack_max
);
3915 /* Stay within the limit we were given */
3916 timeout
= jiffies
+ ato
;
3918 /* Use new timeout only if there wasn't a older one earlier. */
3919 if (icsk
->icsk_ack
.pending
& ICSK_ACK_TIMER
) {
3920 /* If delack timer is about to expire, send ACK now. */
3921 if (time_before_eq(icsk
->icsk_ack
.timeout
, jiffies
+ (ato
>> 2))) {
3926 if (!time_before(timeout
, icsk
->icsk_ack
.timeout
))
3927 timeout
= icsk
->icsk_ack
.timeout
;
3929 icsk
->icsk_ack
.pending
|= ICSK_ACK_SCHED
| ICSK_ACK_TIMER
;
3930 icsk
->icsk_ack
.timeout
= timeout
;
3931 sk_reset_timer(sk
, &icsk
->icsk_delack_timer
, timeout
);
3934 /* This routine sends an ack and also updates the window. */
3935 void __tcp_send_ack(struct sock
*sk
, u32 rcv_nxt
)
3937 struct sk_buff
*buff
;
3939 /* If we have been reset, we may not send again. */
3940 if (sk
->sk_state
== TCP_CLOSE
)
3943 /* We are not putting this on the write queue, so
3944 * tcp_transmit_skb() will set the ownership to this
3947 buff
= alloc_skb(MAX_TCP_HEADER
,
3948 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
3949 if (unlikely(!buff
)) {
3950 struct inet_connection_sock
*icsk
= inet_csk(sk
);
3951 unsigned long delay
;
3953 delay
= TCP_DELACK_MAX
<< icsk
->icsk_ack
.retry
;
3954 if (delay
< TCP_RTO_MAX
)
3955 icsk
->icsk_ack
.retry
++;
3956 inet_csk_schedule_ack(sk
);
3957 icsk
->icsk_ack
.ato
= TCP_ATO_MIN
;
3958 inet_csk_reset_xmit_timer(sk
, ICSK_TIME_DACK
, delay
, TCP_RTO_MAX
);
3962 /* Reserve space for headers and prepare control bits. */
3963 skb_reserve(buff
, MAX_TCP_HEADER
);
3964 tcp_init_nondata_skb(buff
, tcp_acceptable_seq(sk
), TCPHDR_ACK
);
3966 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3968 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3970 skb_set_tcp_pure_ack(buff
);
3972 /* Send it off, this clears delayed acks for us. */
3973 __tcp_transmit_skb(sk
, buff
, 0, (__force gfp_t
)0, rcv_nxt
);
3975 EXPORT_SYMBOL_GPL(__tcp_send_ack
);
3977 void tcp_send_ack(struct sock
*sk
)
3979 __tcp_send_ack(sk
, tcp_sk(sk
)->rcv_nxt
);
3982 /* This routine sends a packet with an out of date sequence
3983 * number. It assumes the other end will try to ack it.
3985 * Question: what should we make while urgent mode?
3986 * 4.4BSD forces sending single byte of data. We cannot send
3987 * out of window data, because we have SND.NXT==SND.MAX...
3989 * Current solution: to send TWO zero-length segments in urgent mode:
3990 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3991 * out-of-date with SND.UNA-1 to probe window.
3993 static int tcp_xmit_probe_skb(struct sock
*sk
, int urgent
, int mib
)
3995 struct tcp_sock
*tp
= tcp_sk(sk
);
3996 struct sk_buff
*skb
;
3998 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3999 skb
= alloc_skb(MAX_TCP_HEADER
,
4000 sk_gfp_mask(sk
, GFP_ATOMIC
| __GFP_NOWARN
));
4004 /* Reserve space for headers and set control bits. */
4005 skb_reserve(skb
, MAX_TCP_HEADER
);
4006 /* Use a previous sequence. This should cause the other
4007 * end to send an ack. Don't queue or clone SKB, just
4010 tcp_init_nondata_skb(skb
, tp
->snd_una
- !urgent
, TCPHDR_ACK
);
4011 NET_INC_STATS(sock_net(sk
), mib
);
4012 return tcp_transmit_skb(sk
, skb
, 0, (__force gfp_t
)0);
4015 /* Called from setsockopt( ... TCP_REPAIR ) */
4016 void tcp_send_window_probe(struct sock
*sk
)
4018 if (sk
->sk_state
== TCP_ESTABLISHED
) {
4019 tcp_sk(sk
)->snd_wl1
= tcp_sk(sk
)->rcv_nxt
- 1;
4020 tcp_mstamp_refresh(tcp_sk(sk
));
4021 tcp_xmit_probe_skb(sk
, 0, LINUX_MIB_TCPWINPROBE
);
4025 /* Initiate keepalive or window probe from timer. */
4026 int tcp_write_wakeup(struct sock
*sk
, int mib
)
4028 struct tcp_sock
*tp
= tcp_sk(sk
);
4029 struct sk_buff
*skb
;
4031 if (sk
->sk_state
== TCP_CLOSE
)
4034 skb
= tcp_send_head(sk
);
4035 if (skb
&& before(TCP_SKB_CB(skb
)->seq
, tcp_wnd_end(tp
))) {
4037 unsigned int mss
= tcp_current_mss(sk
);
4038 unsigned int seg_size
= tcp_wnd_end(tp
) - TCP_SKB_CB(skb
)->seq
;
4040 if (before(tp
->pushed_seq
, TCP_SKB_CB(skb
)->end_seq
))
4041 tp
->pushed_seq
= TCP_SKB_CB(skb
)->end_seq
;
4043 /* We are probing the opening of a window
4044 * but the window size is != 0
4045 * must have been a result SWS avoidance ( sender )
4047 if (seg_size
< TCP_SKB_CB(skb
)->end_seq
- TCP_SKB_CB(skb
)->seq
||
4049 seg_size
= min(seg_size
, mss
);
4050 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
4051 if (tcp_fragment(sk
, TCP_FRAG_IN_WRITE_QUEUE
,
4052 skb
, seg_size
, mss
, GFP_ATOMIC
))
4054 } else if (!tcp_skb_pcount(skb
))
4055 tcp_set_skb_tso_segs(skb
, mss
);
4057 TCP_SKB_CB(skb
)->tcp_flags
|= TCPHDR_PSH
;
4058 err
= tcp_transmit_skb(sk
, skb
, 1, GFP_ATOMIC
);
4060 tcp_event_new_data_sent(sk
, skb
);
4063 if (between(tp
->snd_up
, tp
->snd_una
+ 1, tp
->snd_una
+ 0xFFFF))
4064 tcp_xmit_probe_skb(sk
, 1, mib
);
4065 return tcp_xmit_probe_skb(sk
, 0, mib
);
4069 /* A window probe timeout has occurred. If window is not closed send
4070 * a partial packet else a zero probe.
4072 void tcp_send_probe0(struct sock
*sk
)
4074 struct inet_connection_sock
*icsk
= inet_csk(sk
);
4075 struct tcp_sock
*tp
= tcp_sk(sk
);
4076 struct net
*net
= sock_net(sk
);
4077 unsigned long timeout
;
4080 err
= tcp_write_wakeup(sk
, LINUX_MIB_TCPWINPROBE
);
4082 if (tp
->packets_out
|| tcp_write_queue_empty(sk
)) {
4083 /* Cancel probe timer, if it is not required. */
4084 icsk
->icsk_probes_out
= 0;
4085 icsk
->icsk_backoff
= 0;
4086 icsk
->icsk_probes_tstamp
= 0;
4090 icsk
->icsk_probes_out
++;
4092 if (icsk
->icsk_backoff
< net
->ipv4
.sysctl_tcp_retries2
)
4093 icsk
->icsk_backoff
++;
4094 timeout
= tcp_probe0_when(sk
, TCP_RTO_MAX
);
4096 /* If packet was not sent due to local congestion,
4097 * Let senders fight for local resources conservatively.
4099 timeout
= TCP_RESOURCE_PROBE_INTERVAL
;
4102 timeout
= tcp_clamp_probe0_to_user_timeout(sk
, timeout
);
4103 tcp_reset_xmit_timer(sk
, ICSK_TIME_PROBE0
, timeout
, TCP_RTO_MAX
);
4106 int tcp_rtx_synack(const struct sock
*sk
, struct request_sock
*req
)
4108 const struct tcp_request_sock_ops
*af_ops
= tcp_rsk(req
)->af_specific
;
4112 /* Paired with WRITE_ONCE() in sock_setsockopt() */
4113 if (READ_ONCE(sk
->sk_txrehash
) == SOCK_TXREHASH_ENABLED
)
4114 tcp_rsk(req
)->txhash
= net_tx_rndhash();
4115 res
= af_ops
->send_synack(sk
, NULL
, &fl
, req
, NULL
, TCP_SYNACK_NORMAL
,
4118 TCP_INC_STATS(sock_net(sk
), TCP_MIB_RETRANSSEGS
);
4119 NET_INC_STATS(sock_net(sk
), LINUX_MIB_TCPSYNRETRANS
);
4120 if (unlikely(tcp_passive_fastopen(sk
)))
4121 tcp_sk(sk
)->total_retrans
++;
4122 trace_tcp_retransmit_synack(sk
, req
);
4126 EXPORT_SYMBOL(tcp_rtx_synack
);