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);
172 /* If it is a reply for ato after last received
173 * packet, enter pingpong mode.
175 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
176 inet_csk_enter_pingpong_mode(sk);
179 /* Account for an ACK we sent. */
180 static inline void tcp_event_ack_sent(struct sock *sk, u32 rcv_nxt)
182 struct tcp_sock *tp = tcp_sk(sk);
184 if (unlikely(tp->compressed_ack)) {
185 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
187 tp->compressed_ack = 0;
188 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
192 if (unlikely(rcv_nxt != tp->rcv_nxt))
193 return; /* Special ACK sent by DCTCP to reflect ECN */
194 tcp_dec_quickack_mode(sk);
195 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
198 /* Determine a window scaling and initial window to offer.
199 * Based on the assumption that the given amount of space
200 * will be offered. Store the results in the tp structure.
201 * NOTE: for smooth operation initial space offering should
202 * be a multiple of mss if possible. We assume here that mss >= 1.
203 * This MUST be enforced by all callers.
205 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
206 __u32 *rcv_wnd, __u32 *window_clamp,
207 int wscale_ok, __u8 *rcv_wscale,
210 unsigned int space = (__space < 0 ? 0 : __space);
212 /* If no clamp set the clamp to the max possible scaled window */
213 if (*window_clamp == 0)
214 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
215 space = min(*window_clamp, space);
217 /* Quantize space offering to a multiple of mss if possible. */
219 space = rounddown(space, mss);
221 /* NOTE: offering an initial window larger than 32767
222 * will break some buggy TCP stacks. If the admin tells us
223 * it is likely we could be speaking with such a buggy stack
224 * we will truncate our initial window offering to 32K-1
225 * unless the remote has sent us a window scaling option,
226 * which we interpret as a sign the remote TCP is not
227 * misinterpreting the window field as a signed quantity.
229 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
230 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
232 (*rcv_wnd) = min_t(u32, space, U16_MAX);
235 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
239 /* Set window scaling on max possible window */
240 space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
241 space = max_t(u32, space, READ_ONCE(sysctl_rmem_max));
242 space = min_t(u32, space, *window_clamp);
243 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
246 /* Set the clamp no higher than max representable value */
247 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
249 EXPORT_SYMBOL(tcp_select_initial_window);
251 /* Chose a new window to advertise, update state in tcp_sock for the
252 * socket, and return result with RFC1323 scaling applied. The return
253 * value can be stuffed directly into th->window for an outgoing
256 static u16 tcp_select_window(struct sock *sk)
258 struct tcp_sock *tp = tcp_sk(sk);
259 u32 old_win = tp->rcv_wnd;
260 u32 cur_win = tcp_receive_window(tp);
261 u32 new_win = __tcp_select_window(sk);
262 struct net *net = sock_net(sk);
264 if (new_win < cur_win) {
265 /* Danger Will Robinson!
266 * Don't update rcv_wup/rcv_wnd here or else
267 * we will not be able to advertise a zero
268 * window in time. --DaveM
270 * Relax Will Robinson.
272 if (!READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) || !tp->rx_opt.rcv_wscale) {
273 /* Never shrink the offered window */
275 NET_INC_STATS(net, LINUX_MIB_TCPWANTZEROWINDOWADV);
276 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
280 tp->rcv_wnd = new_win;
281 tp->rcv_wup = tp->rcv_nxt;
283 /* Make sure we do not exceed the maximum possible
286 if (!tp->rx_opt.rcv_wscale &&
287 READ_ONCE(net->ipv4.sysctl_tcp_workaround_signed_windows))
288 new_win = min(new_win, MAX_TCP_WINDOW);
290 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
292 /* RFC1323 scaling applied */
293 new_win >>= tp->rx_opt.rcv_wscale;
295 /* If we advertise zero window, disable fast path. */
299 NET_INC_STATS(net, LINUX_MIB_TCPTOZEROWINDOWADV);
300 } else if (old_win == 0) {
301 NET_INC_STATS(net, LINUX_MIB_TCPFROMZEROWINDOWADV);
307 /* Packet ECN state for a SYN-ACK */
308 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
310 const struct tcp_sock *tp = tcp_sk(sk);
312 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
313 if (!(tp->ecn_flags & TCP_ECN_OK))
314 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
315 else if (tcp_ca_needs_ecn(sk) ||
316 tcp_bpf_ca_needs_ecn(sk))
320 /* Packet ECN state for a SYN. */
321 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
323 struct tcp_sock *tp = tcp_sk(sk);
324 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
325 bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 ||
326 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
329 const struct dst_entry *dst = __sk_dst_get(sk);
331 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
338 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
339 tp->ecn_flags = TCP_ECN_OK;
340 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
345 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
347 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback))
348 /* tp->ecn_flags are cleared at a later point in time when
349 * SYN ACK is ultimatively being received.
351 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
355 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
357 if (inet_rsk(req)->ecn_ok)
361 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
364 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
365 struct tcphdr *th, int tcp_header_len)
367 struct tcp_sock *tp = tcp_sk(sk);
369 if (tp->ecn_flags & TCP_ECN_OK) {
370 /* Not-retransmitted data segment: set ECT and inject CWR. */
371 if (skb->len != tcp_header_len &&
372 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
374 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
375 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
377 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
379 } else if (!tcp_ca_needs_ecn(sk)) {
380 /* ACK or retransmitted segment: clear ECT|CE */
381 INET_ECN_dontxmit(sk);
383 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
388 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
389 * auto increment end seqno.
391 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
393 skb->ip_summed = CHECKSUM_PARTIAL;
395 TCP_SKB_CB(skb)->tcp_flags = flags;
397 tcp_skb_pcount_set(skb, 1);
399 TCP_SKB_CB(skb)->seq = seq;
400 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
402 TCP_SKB_CB(skb)->end_seq = seq;
405 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
407 return tp->snd_una != tp->snd_up;
410 #define OPTION_SACK_ADVERTISE BIT(0)
411 #define OPTION_TS BIT(1)
412 #define OPTION_MD5 BIT(2)
413 #define OPTION_WSCALE BIT(3)
414 #define OPTION_FAST_OPEN_COOKIE BIT(8)
415 #define OPTION_SMC BIT(9)
416 #define OPTION_MPTCP BIT(10)
418 static void smc_options_write(__be32 *ptr, u16 *options)
420 #if IS_ENABLED(CONFIG_SMC)
421 if (static_branch_unlikely(&tcp_have_smc)) {
422 if (unlikely(OPTION_SMC & *options)) {
423 *ptr++ = htonl((TCPOPT_NOP << 24) |
426 (TCPOLEN_EXP_SMC_BASE));
427 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
433 struct tcp_out_options {
434 u16 options; /* bit field of OPTION_* */
435 u16 mss; /* 0 to disable */
436 u8 ws; /* window scale, 0 to disable */
437 u8 num_sack_blocks; /* number of SACK blocks to include */
438 u8 hash_size; /* bytes in hash_location */
439 u8 bpf_opt_len; /* length of BPF hdr option */
440 __u8 *hash_location; /* temporary pointer, overloaded */
441 __u32 tsval, tsecr; /* need to include OPTION_TS */
442 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
443 struct mptcp_out_options mptcp;
446 static void mptcp_options_write(struct tcphdr *th, __be32 *ptr,
448 struct tcp_out_options *opts)
450 #if IS_ENABLED(CONFIG_MPTCP)
451 if (unlikely(OPTION_MPTCP & opts->options))
452 mptcp_write_options(th, ptr, tp, &opts->mptcp);
456 #ifdef CONFIG_CGROUP_BPF
457 static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
458 enum tcp_synack_type synack_type)
461 return BPF_WRITE_HDR_TCP_CURRENT_MSS;
463 if (unlikely(synack_type == TCP_SYNACK_COOKIE))
464 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
469 /* req, syn_skb and synack_type are used when writing synack */
470 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
471 struct request_sock *req,
472 struct sk_buff *syn_skb,
473 enum tcp_synack_type synack_type,
474 struct tcp_out_options *opts,
475 unsigned int *remaining)
477 struct bpf_sock_ops_kern sock_ops;
480 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
481 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
485 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
488 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
490 sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
493 /* The listen "sk" cannot be passed here because
494 * it is not locked. It would not make too much
495 * sense to do bpf_setsockopt(listen_sk) based
496 * on individual connection request also.
498 * Thus, "req" is passed here and the cgroup-bpf-progs
499 * of the listen "sk" will be run.
501 * "req" is also used here for fastopen even the "sk" here is
502 * a fullsock "child" sk. It is to keep the behavior
503 * consistent between fastopen and non-fastopen on
504 * the bpf programming side.
506 sock_ops.sk = (struct sock *)req;
507 sock_ops.syn_skb = syn_skb;
509 sock_owned_by_me(sk);
511 sock_ops.is_fullsock = 1;
515 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
516 sock_ops.remaining_opt_len = *remaining;
517 /* tcp_current_mss() does not pass a skb */
519 bpf_skops_init_skb(&sock_ops, skb, 0);
521 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
523 if (err || sock_ops.remaining_opt_len == *remaining)
526 opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
527 /* round up to 4 bytes */
528 opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
530 *remaining -= opts->bpf_opt_len;
533 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
534 struct request_sock *req,
535 struct sk_buff *syn_skb,
536 enum tcp_synack_type synack_type,
537 struct tcp_out_options *opts)
539 u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
540 struct bpf_sock_ops_kern sock_ops;
543 if (likely(!max_opt_len))
546 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
548 sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
551 sock_ops.sk = (struct sock *)req;
552 sock_ops.syn_skb = syn_skb;
554 sock_owned_by_me(sk);
556 sock_ops.is_fullsock = 1;
560 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
561 sock_ops.remaining_opt_len = max_opt_len;
562 first_opt_off = tcp_hdrlen(skb) - max_opt_len;
563 bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
565 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
570 nr_written = max_opt_len - sock_ops.remaining_opt_len;
572 if (nr_written < max_opt_len)
573 memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
574 max_opt_len - nr_written);
577 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
578 struct request_sock *req,
579 struct sk_buff *syn_skb,
580 enum tcp_synack_type synack_type,
581 struct tcp_out_options *opts,
582 unsigned int *remaining)
586 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
587 struct request_sock *req,
588 struct sk_buff *syn_skb,
589 enum tcp_synack_type synack_type,
590 struct tcp_out_options *opts)
595 /* Write previously computed TCP options to the packet.
597 * Beware: Something in the Internet is very sensitive to the ordering of
598 * TCP options, we learned this through the hard way, so be careful here.
599 * Luckily we can at least blame others for their non-compliance but from
600 * inter-operability perspective it seems that we're somewhat stuck with
601 * the ordering which we have been using if we want to keep working with
602 * those broken things (not that it currently hurts anybody as there isn't
603 * particular reason why the ordering would need to be changed).
605 * At least SACK_PERM as the first option is known to lead to a disaster
606 * (but it may well be that other scenarios fail similarly).
608 static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp,
609 struct tcp_out_options *opts)
611 __be32 *ptr = (__be32 *)(th + 1);
612 u16 options = opts->options; /* mungable copy */
614 if (unlikely(OPTION_MD5 & options)) {
615 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
616 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
617 /* overload cookie hash location */
618 opts->hash_location = (__u8 *)ptr;
622 if (unlikely(opts->mss)) {
623 *ptr++ = htonl((TCPOPT_MSS << 24) |
624 (TCPOLEN_MSS << 16) |
628 if (likely(OPTION_TS & options)) {
629 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
630 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
631 (TCPOLEN_SACK_PERM << 16) |
632 (TCPOPT_TIMESTAMP << 8) |
634 options &= ~OPTION_SACK_ADVERTISE;
636 *ptr++ = htonl((TCPOPT_NOP << 24) |
638 (TCPOPT_TIMESTAMP << 8) |
641 *ptr++ = htonl(opts->tsval);
642 *ptr++ = htonl(opts->tsecr);
645 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
646 *ptr++ = htonl((TCPOPT_NOP << 24) |
648 (TCPOPT_SACK_PERM << 8) |
652 if (unlikely(OPTION_WSCALE & options)) {
653 *ptr++ = htonl((TCPOPT_NOP << 24) |
654 (TCPOPT_WINDOW << 16) |
655 (TCPOLEN_WINDOW << 8) |
659 if (unlikely(opts->num_sack_blocks)) {
660 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
661 tp->duplicate_sack : tp->selective_acks;
664 *ptr++ = htonl((TCPOPT_NOP << 24) |
667 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
668 TCPOLEN_SACK_PERBLOCK)));
670 for (this_sack = 0; this_sack < opts->num_sack_blocks;
672 *ptr++ = htonl(sp[this_sack].start_seq);
673 *ptr++ = htonl(sp[this_sack].end_seq);
676 tp->rx_opt.dsack = 0;
679 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
680 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
682 u32 len; /* Fast Open option length */
685 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
686 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
687 TCPOPT_FASTOPEN_MAGIC);
688 p += TCPOLEN_EXP_FASTOPEN_BASE;
690 len = TCPOLEN_FASTOPEN_BASE + foc->len;
691 *p++ = TCPOPT_FASTOPEN;
695 memcpy(p, foc->val, foc->len);
696 if ((len & 3) == 2) {
697 p[foc->len] = TCPOPT_NOP;
698 p[foc->len + 1] = TCPOPT_NOP;
700 ptr += (len + 3) >> 2;
703 smc_options_write(ptr, &options);
705 mptcp_options_write(th, ptr, tp, opts);
708 static void smc_set_option(const struct tcp_sock *tp,
709 struct tcp_out_options *opts,
710 unsigned int *remaining)
712 #if IS_ENABLED(CONFIG_SMC)
713 if (static_branch_unlikely(&tcp_have_smc)) {
715 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
716 opts->options |= OPTION_SMC;
717 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
724 static void smc_set_option_cond(const struct tcp_sock *tp,
725 const struct inet_request_sock *ireq,
726 struct tcp_out_options *opts,
727 unsigned int *remaining)
729 #if IS_ENABLED(CONFIG_SMC)
730 if (static_branch_unlikely(&tcp_have_smc)) {
731 if (tp->syn_smc && ireq->smc_ok) {
732 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
733 opts->options |= OPTION_SMC;
734 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
741 static void mptcp_set_option_cond(const struct request_sock *req,
742 struct tcp_out_options *opts,
743 unsigned int *remaining)
745 if (rsk_is_mptcp(req)) {
748 if (mptcp_synack_options(req, &size, &opts->mptcp)) {
749 if (*remaining >= size) {
750 opts->options |= OPTION_MPTCP;
757 /* Compute TCP options for SYN packets. This is not the final
758 * network wire format yet.
760 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
761 struct tcp_out_options *opts,
762 struct tcp_md5sig_key **md5)
764 struct tcp_sock *tp = tcp_sk(sk);
765 unsigned int remaining = MAX_TCP_OPTION_SPACE;
766 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
769 #ifdef CONFIG_TCP_MD5SIG
770 if (static_branch_unlikely(&tcp_md5_needed) &&
771 rcu_access_pointer(tp->md5sig_info)) {
772 *md5 = tp->af_specific->md5_lookup(sk, sk);
774 opts->options |= OPTION_MD5;
775 remaining -= TCPOLEN_MD5SIG_ALIGNED;
780 /* We always get an MSS option. The option bytes which will be seen in
781 * normal data packets should timestamps be used, must be in the MSS
782 * advertised. But we subtract them from tp->mss_cache so that
783 * calculations in tcp_sendmsg are simpler etc. So account for this
784 * fact here if necessary. If we don't do this correctly, as a
785 * receiver we won't recognize data packets as being full sized when we
786 * should, and thus we won't abide by the delayed ACK rules correctly.
787 * SACKs don't matter, we never delay an ACK when we have any of those
789 opts->mss = tcp_advertise_mss(sk);
790 remaining -= TCPOLEN_MSS_ALIGNED;
792 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps) && !*md5)) {
793 opts->options |= OPTION_TS;
794 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
795 opts->tsecr = tp->rx_opt.ts_recent;
796 remaining -= TCPOLEN_TSTAMP_ALIGNED;
798 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) {
799 opts->ws = tp->rx_opt.rcv_wscale;
800 opts->options |= OPTION_WSCALE;
801 remaining -= TCPOLEN_WSCALE_ALIGNED;
803 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) {
804 opts->options |= OPTION_SACK_ADVERTISE;
805 if (unlikely(!(OPTION_TS & opts->options)))
806 remaining -= TCPOLEN_SACKPERM_ALIGNED;
809 if (fastopen && fastopen->cookie.len >= 0) {
810 u32 need = fastopen->cookie.len;
812 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
813 TCPOLEN_FASTOPEN_BASE;
814 need = (need + 3) & ~3U; /* Align to 32 bits */
815 if (remaining >= need) {
816 opts->options |= OPTION_FAST_OPEN_COOKIE;
817 opts->fastopen_cookie = &fastopen->cookie;
819 tp->syn_fastopen = 1;
820 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
824 smc_set_option(tp, opts, &remaining);
826 if (sk_is_mptcp(sk)) {
829 if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
830 opts->options |= OPTION_MPTCP;
835 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
837 return MAX_TCP_OPTION_SPACE - remaining;
840 /* Set up TCP options for SYN-ACKs. */
841 static unsigned int tcp_synack_options(const struct sock *sk,
842 struct request_sock *req,
843 unsigned int mss, struct sk_buff *skb,
844 struct tcp_out_options *opts,
845 const struct tcp_md5sig_key *md5,
846 struct tcp_fastopen_cookie *foc,
847 enum tcp_synack_type synack_type,
848 struct sk_buff *syn_skb)
850 struct inet_request_sock *ireq = inet_rsk(req);
851 unsigned int remaining = MAX_TCP_OPTION_SPACE;
853 #ifdef CONFIG_TCP_MD5SIG
855 opts->options |= OPTION_MD5;
856 remaining -= TCPOLEN_MD5SIG_ALIGNED;
858 /* We can't fit any SACK blocks in a packet with MD5 + TS
859 * options. There was discussion about disabling SACK
860 * rather than TS in order to fit in better with old,
861 * buggy kernels, but that was deemed to be unnecessary.
863 if (synack_type != TCP_SYNACK_COOKIE)
864 ireq->tstamp_ok &= !ireq->sack_ok;
868 /* We always send an MSS option. */
870 remaining -= TCPOLEN_MSS_ALIGNED;
872 if (likely(ireq->wscale_ok)) {
873 opts->ws = ireq->rcv_wscale;
874 opts->options |= OPTION_WSCALE;
875 remaining -= TCPOLEN_WSCALE_ALIGNED;
877 if (likely(ireq->tstamp_ok)) {
878 opts->options |= OPTION_TS;
879 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
880 opts->tsecr = READ_ONCE(req->ts_recent);
881 remaining -= TCPOLEN_TSTAMP_ALIGNED;
883 if (likely(ireq->sack_ok)) {
884 opts->options |= OPTION_SACK_ADVERTISE;
885 if (unlikely(!ireq->tstamp_ok))
886 remaining -= TCPOLEN_SACKPERM_ALIGNED;
888 if (foc != NULL && foc->len >= 0) {
891 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
892 TCPOLEN_FASTOPEN_BASE;
893 need = (need + 3) & ~3U; /* Align to 32 bits */
894 if (remaining >= need) {
895 opts->options |= OPTION_FAST_OPEN_COOKIE;
896 opts->fastopen_cookie = foc;
901 mptcp_set_option_cond(req, opts, &remaining);
903 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
905 bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb,
906 synack_type, opts, &remaining);
908 return MAX_TCP_OPTION_SPACE - remaining;
911 /* Compute TCP options for ESTABLISHED sockets. This is not the
912 * final wire format yet.
914 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
915 struct tcp_out_options *opts,
916 struct tcp_md5sig_key **md5)
918 struct tcp_sock *tp = tcp_sk(sk);
919 unsigned int size = 0;
920 unsigned int eff_sacks;
925 #ifdef CONFIG_TCP_MD5SIG
926 if (static_branch_unlikely(&tcp_md5_needed) &&
927 rcu_access_pointer(tp->md5sig_info)) {
928 *md5 = tp->af_specific->md5_lookup(sk, sk);
930 opts->options |= OPTION_MD5;
931 size += TCPOLEN_MD5SIG_ALIGNED;
936 if (likely(tp->rx_opt.tstamp_ok)) {
937 opts->options |= OPTION_TS;
938 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
939 opts->tsecr = tp->rx_opt.ts_recent;
940 size += TCPOLEN_TSTAMP_ALIGNED;
943 /* MPTCP options have precedence over SACK for the limited TCP
944 * option space because a MPTCP connection would be forced to
945 * fall back to regular TCP if a required multipath option is
946 * missing. SACK still gets a chance to use whatever space is
949 if (sk_is_mptcp(sk)) {
950 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
951 unsigned int opt_size = 0;
953 if (mptcp_established_options(sk, skb, &opt_size, remaining,
955 opts->options |= OPTION_MPTCP;
960 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
961 if (unlikely(eff_sacks)) {
962 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
963 if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
964 TCPOLEN_SACK_PERBLOCK))
967 opts->num_sack_blocks =
968 min_t(unsigned int, eff_sacks,
969 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
970 TCPOLEN_SACK_PERBLOCK);
972 size += TCPOLEN_SACK_BASE_ALIGNED +
973 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
976 if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
977 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
978 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
980 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
982 size = MAX_TCP_OPTION_SPACE - remaining;
989 /* TCP SMALL QUEUES (TSQ)
991 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
992 * to reduce RTT and bufferbloat.
993 * We do this using a special skb destructor (tcp_wfree).
995 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
996 * needs to be reallocated in a driver.
997 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
999 * Since transmit from skb destructor is forbidden, we use a tasklet
1000 * to process all sockets that eventually need to send more skbs.
1001 * We use one tasklet per cpu, with its own queue of sockets.
1003 struct tsq_tasklet {
1004 struct tasklet_struct tasklet;
1005 struct list_head head; /* queue of tcp sockets */
1007 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
1009 static void tcp_tsq_write(struct sock *sk)
1011 if ((1 << sk->sk_state) &
1012 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1013 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
1014 struct tcp_sock *tp = tcp_sk(sk);
1016 if (tp->lost_out > tp->retrans_out &&
1017 tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) {
1018 tcp_mstamp_refresh(tp);
1019 tcp_xmit_retransmit_queue(sk);
1022 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
1027 static void tcp_tsq_handler(struct sock *sk)
1030 if (!sock_owned_by_user(sk))
1032 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
1037 * One tasklet per cpu tries to send more skbs.
1038 * We run in tasklet context but need to disable irqs when
1039 * transferring tsq->head because tcp_wfree() might
1040 * interrupt us (non NAPI drivers)
1042 static void tcp_tasklet_func(struct tasklet_struct *t)
1044 struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet);
1046 unsigned long flags;
1047 struct list_head *q, *n;
1048 struct tcp_sock *tp;
1051 local_irq_save(flags);
1052 list_splice_init(&tsq->head, &list);
1053 local_irq_restore(flags);
1055 list_for_each_safe(q, n, &list) {
1056 tp = list_entry(q, struct tcp_sock, tsq_node);
1057 list_del(&tp->tsq_node);
1059 sk = (struct sock *)tp;
1060 smp_mb__before_atomic();
1061 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
1063 tcp_tsq_handler(sk);
1068 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
1069 TCPF_WRITE_TIMER_DEFERRED | \
1070 TCPF_DELACK_TIMER_DEFERRED | \
1071 TCPF_MTU_REDUCED_DEFERRED)
1073 * tcp_release_cb - tcp release_sock() callback
1076 * called from release_sock() to perform protocol dependent
1077 * actions before socket release.
1079 void tcp_release_cb(struct sock *sk)
1081 unsigned long flags, nflags;
1083 /* perform an atomic operation only if at least one flag is set */
1085 flags = sk->sk_tsq_flags;
1086 if (!(flags & TCP_DEFERRED_ALL))
1088 nflags = flags & ~TCP_DEFERRED_ALL;
1089 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
1091 if (flags & TCPF_TSQ_DEFERRED) {
1095 /* Here begins the tricky part :
1096 * We are called from release_sock() with :
1098 * 2) sk_lock.slock spinlock held
1099 * 3) socket owned by us (sk->sk_lock.owned == 1)
1101 * But following code is meant to be called from BH handlers,
1102 * so we should keep BH disabled, but early release socket ownership
1104 sock_release_ownership(sk);
1106 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1107 tcp_write_timer_handler(sk);
1110 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1111 tcp_delack_timer_handler(sk);
1114 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1115 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1119 EXPORT_SYMBOL(tcp_release_cb);
1121 void __init tcp_tasklet_init(void)
1125 for_each_possible_cpu(i) {
1126 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
1128 INIT_LIST_HEAD(&tsq->head);
1129 tasklet_setup(&tsq->tasklet, tcp_tasklet_func);
1134 * Write buffer destructor automatically called from kfree_skb.
1135 * We can't xmit new skbs from this context, as we might already
1138 void tcp_wfree(struct sk_buff *skb)
1140 struct sock *sk = skb->sk;
1141 struct tcp_sock *tp = tcp_sk(sk);
1142 unsigned long flags, nval, oval;
1144 /* Keep one reference on sk_wmem_alloc.
1145 * Will be released by sk_free() from here or tcp_tasklet_func()
1147 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1149 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
1150 * Wait until our queues (qdisc + devices) are drained.
1152 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1153 * - chance for incoming ACK (processed by another cpu maybe)
1154 * to migrate this flow (skb->ooo_okay will be eventually set)
1156 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1159 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
1160 struct tsq_tasklet *tsq;
1163 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1166 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1167 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
1171 /* queue this socket to tasklet queue */
1172 local_irq_save(flags);
1173 tsq = this_cpu_ptr(&tsq_tasklet);
1174 empty = list_empty(&tsq->head);
1175 list_add(&tp->tsq_node, &tsq->head);
1177 tasklet_schedule(&tsq->tasklet);
1178 local_irq_restore(flags);
1185 /* Note: Called under soft irq.
1186 * We can call TCP stack right away, unless socket is owned by user.
1188 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1190 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1191 struct sock *sk = (struct sock *)tp;
1193 tcp_tsq_handler(sk);
1196 return HRTIMER_NORESTART;
1199 static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1202 struct tcp_sock *tp = tcp_sk(sk);
1204 if (sk->sk_pacing_status != SK_PACING_NONE) {
1205 unsigned long rate = sk->sk_pacing_rate;
1207 /* Original sch_fq does not pace first 10 MSS
1208 * Note that tp->data_segs_out overflows after 2^32 packets,
1209 * this is a minor annoyance.
1211 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1212 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1213 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1215 /* take into account OS jitter */
1216 len_ns -= min_t(u64, len_ns / 2, credit);
1217 tp->tcp_wstamp_ns += len_ns;
1220 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1223 INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1224 INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1225 INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1227 /* This routine actually transmits TCP packets queued in by
1228 * tcp_do_sendmsg(). This is used by both the initial
1229 * transmission and possible later retransmissions.
1230 * All SKB's seen here are completely headerless. It is our
1231 * job to build the TCP header, and pass the packet down to
1232 * IP so it can do the same plus pass the packet off to the
1235 * We are working here with either a clone of the original
1236 * SKB, or a fresh unique copy made by the retransmit engine.
1238 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1239 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1241 const struct inet_connection_sock *icsk = inet_csk(sk);
1242 struct inet_sock *inet;
1243 struct tcp_sock *tp;
1244 struct tcp_skb_cb *tcb;
1245 struct tcp_out_options opts;
1246 unsigned int tcp_options_size, tcp_header_size;
1247 struct sk_buff *oskb = NULL;
1248 struct tcp_md5sig_key *md5;
1253 BUG_ON(!skb || !tcp_skb_pcount(skb));
1255 prior_wstamp = tp->tcp_wstamp_ns;
1256 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1257 skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
1261 tcp_skb_tsorted_save(oskb) {
1262 if (unlikely(skb_cloned(oskb)))
1263 skb = pskb_copy(oskb, gfp_mask);
1265 skb = skb_clone(oskb, gfp_mask);
1266 } tcp_skb_tsorted_restore(oskb);
1270 /* retransmit skbs might have a non zero value in skb->dev
1271 * because skb->dev is aliased with skb->rbnode.rb_left
1277 tcb = TCP_SKB_CB(skb);
1278 memset(&opts, 0, sizeof(opts));
1280 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1281 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1283 tcp_options_size = tcp_established_options(sk, skb, &opts,
1285 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1286 * at receiver : This slightly improve GRO performance.
1287 * Note that we do not force the PSH flag for non GSO packets,
1288 * because they might be sent under high congestion events,
1289 * and in this case it is better to delay the delivery of 1-MSS
1290 * packets and thus the corresponding ACK packet that would
1291 * release the following packet.
1293 if (tcp_skb_pcount(skb) > 1)
1294 tcb->tcp_flags |= TCPHDR_PSH;
1296 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1298 /* if no packet is in qdisc/device queue, then allow XPS to select
1299 * another queue. We can be called from tcp_tsq_handler()
1300 * which holds one reference to sk.
1302 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1303 * One way to get this would be to set skb->truesize = 2 on them.
1305 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1307 /* If we had to use memory reserve to allocate this skb,
1308 * this might cause drops if packet is looped back :
1309 * Other socket might not have SOCK_MEMALLOC.
1310 * Packets not looped back do not care about pfmemalloc.
1312 skb->pfmemalloc = 0;
1314 skb_push(skb, tcp_header_size);
1315 skb_reset_transport_header(skb);
1319 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1320 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1322 skb_set_dst_pending_confirm(skb, READ_ONCE(sk->sk_dst_pending_confirm));
1324 /* Build TCP header and checksum it. */
1325 th = (struct tcphdr *)skb->data;
1326 th->source = inet->inet_sport;
1327 th->dest = inet->inet_dport;
1328 th->seq = htonl(tcb->seq);
1329 th->ack_seq = htonl(rcv_nxt);
1330 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1336 /* The urg_mode check is necessary during a below snd_una win probe */
1337 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1338 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1339 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1341 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1342 th->urg_ptr = htons(0xFFFF);
1347 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1348 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1349 th->window = htons(tcp_select_window(sk));
1350 tcp_ecn_send(sk, skb, th, tcp_header_size);
1352 /* RFC1323: The window in SYN & SYN/ACK segments
1355 th->window = htons(min(tp->rcv_wnd, 65535U));
1358 tcp_options_write(th, tp, &opts);
1360 #ifdef CONFIG_TCP_MD5SIG
1361 /* Calculate the MD5 hash, as we have all we need now */
1364 tp->af_specific->calc_md5_hash(opts.hash_location,
1369 /* BPF prog is the last one writing header option */
1370 bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1372 INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1373 tcp_v6_send_check, tcp_v4_send_check,
1376 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1377 tcp_event_ack_sent(sk, rcv_nxt);
1379 if (skb->len != tcp_header_size) {
1380 tcp_event_data_sent(tp, sk);
1381 tp->data_segs_out += tcp_skb_pcount(skb);
1382 tp->bytes_sent += skb->len - tcp_header_size;
1385 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1386 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1387 tcp_skb_pcount(skb));
1389 tp->segs_out += tcp_skb_pcount(skb);
1390 skb_set_hash_from_sk(skb, sk);
1391 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1392 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1393 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1395 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1397 /* Cleanup our debris for IP stacks */
1398 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1399 sizeof(struct inet6_skb_parm)));
1401 tcp_add_tx_delay(skb, tp);
1403 err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1404 inet6_csk_xmit, ip_queue_xmit,
1405 sk, skb, &inet->cork.fl);
1407 if (unlikely(err > 0)) {
1409 err = net_xmit_eval(err);
1412 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1413 tcp_rate_skb_sent(sk, oskb);
1418 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1421 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1422 tcp_sk(sk)->rcv_nxt);
1425 /* This routine just queues the buffer for sending.
1427 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1428 * otherwise socket can stall.
1430 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1432 struct tcp_sock *tp = tcp_sk(sk);
1434 /* Advance write_seq and place onto the write_queue. */
1435 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1436 __skb_header_release(skb);
1437 tcp_add_write_queue_tail(sk, skb);
1438 sk_wmem_queued_add(sk, skb->truesize);
1439 sk_mem_charge(sk, skb->truesize);
1442 /* Initialize TSO segments for a packet. */
1443 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1445 if (skb->len <= mss_now) {
1446 /* Avoid the costly divide in the normal
1449 tcp_skb_pcount_set(skb, 1);
1450 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1452 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1453 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1457 /* Pcount in the middle of the write queue got changed, we need to do various
1458 * tweaks to fix counters
1460 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1462 struct tcp_sock *tp = tcp_sk(sk);
1464 tp->packets_out -= decr;
1466 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1467 tp->sacked_out -= decr;
1468 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1469 tp->retrans_out -= decr;
1470 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1471 tp->lost_out -= decr;
1473 /* Reno case is special. Sigh... */
1474 if (tcp_is_reno(tp) && decr > 0)
1475 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1477 if (tp->lost_skb_hint &&
1478 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1479 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1480 tp->lost_cnt_hint -= decr;
1482 tcp_verify_left_out(tp);
1485 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1487 return TCP_SKB_CB(skb)->txstamp_ack ||
1488 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1491 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1493 struct skb_shared_info *shinfo = skb_shinfo(skb);
1495 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1496 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1497 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1498 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1500 shinfo->tx_flags &= ~tsflags;
1501 shinfo2->tx_flags |= tsflags;
1502 swap(shinfo->tskey, shinfo2->tskey);
1503 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1504 TCP_SKB_CB(skb)->txstamp_ack = 0;
1508 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1510 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1511 TCP_SKB_CB(skb)->eor = 0;
1514 /* Insert buff after skb on the write or rtx queue of sk. */
1515 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1516 struct sk_buff *buff,
1518 enum tcp_queue tcp_queue)
1520 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1521 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1523 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1526 /* Function to create two new TCP segments. Shrinks the given segment
1527 * to the specified size and appends a new segment with the rest of the
1528 * packet to the list. This won't be called frequently, I hope.
1529 * Remember, these are still headerless SKBs at this point.
1531 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1532 struct sk_buff *skb, u32 len,
1533 unsigned int mss_now, gfp_t gfp)
1535 struct tcp_sock *tp = tcp_sk(sk);
1536 struct sk_buff *buff;
1537 int nsize, old_factor;
1542 if (WARN_ON(len > skb->len))
1545 nsize = skb_headlen(skb) - len;
1549 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1550 * We need some allowance to not penalize applications setting small
1552 * Also allow first and last skb in retransmit queue to be split.
1554 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE);
1555 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1556 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1557 skb != tcp_rtx_queue_head(sk) &&
1558 skb != tcp_rtx_queue_tail(sk))) {
1559 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1563 if (skb_unclone_keeptruesize(skb, gfp))
1566 /* Get a new skb... force flag on. */
1567 buff = tcp_stream_alloc_skb(sk, nsize, gfp, true);
1569 return -ENOMEM; /* We'll just try again later. */
1570 skb_copy_decrypted(buff, skb);
1571 mptcp_skb_ext_copy(buff, skb);
1573 sk_wmem_queued_add(sk, buff->truesize);
1574 sk_mem_charge(sk, buff->truesize);
1575 nlen = skb->len - len - nsize;
1576 buff->truesize += nlen;
1577 skb->truesize -= nlen;
1579 /* Correct the sequence numbers. */
1580 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1581 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1582 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1584 /* PSH and FIN should only be set in the second packet. */
1585 flags = TCP_SKB_CB(skb)->tcp_flags;
1586 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1587 TCP_SKB_CB(buff)->tcp_flags = flags;
1588 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1589 tcp_skb_fragment_eor(skb, buff);
1591 skb_split(skb, buff, len);
1593 skb_set_delivery_time(buff, skb->tstamp, true);
1594 tcp_fragment_tstamp(skb, buff);
1596 old_factor = tcp_skb_pcount(skb);
1598 /* Fix up tso_factor for both original and new SKB. */
1599 tcp_set_skb_tso_segs(skb, mss_now);
1600 tcp_set_skb_tso_segs(buff, mss_now);
1602 /* Update delivered info for the new segment */
1603 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1605 /* If this packet has been sent out already, we must
1606 * adjust the various packet counters.
1608 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1609 int diff = old_factor - tcp_skb_pcount(skb) -
1610 tcp_skb_pcount(buff);
1613 tcp_adjust_pcount(sk, skb, diff);
1616 /* Link BUFF into the send queue. */
1617 __skb_header_release(buff);
1618 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1619 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1620 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1625 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1626 * data is not copied, but immediately discarded.
1628 static int __pskb_trim_head(struct sk_buff *skb, int len)
1630 struct skb_shared_info *shinfo;
1633 eat = min_t(int, len, skb_headlen(skb));
1635 __skb_pull(skb, eat);
1642 shinfo = skb_shinfo(skb);
1643 for (i = 0; i < shinfo->nr_frags; i++) {
1644 int size = skb_frag_size(&shinfo->frags[i]);
1647 skb_frag_unref(skb, i);
1650 shinfo->frags[k] = shinfo->frags[i];
1652 skb_frag_off_add(&shinfo->frags[k], eat);
1653 skb_frag_size_sub(&shinfo->frags[k], eat);
1659 shinfo->nr_frags = k;
1661 skb->data_len -= len;
1662 skb->len = skb->data_len;
1666 /* Remove acked data from a packet in the transmit queue. */
1667 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1671 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
1674 delta_truesize = __pskb_trim_head(skb, len);
1676 TCP_SKB_CB(skb)->seq += len;
1678 if (delta_truesize) {
1679 skb->truesize -= delta_truesize;
1680 sk_wmem_queued_add(sk, -delta_truesize);
1681 if (!skb_zcopy_pure(skb))
1682 sk_mem_uncharge(sk, delta_truesize);
1685 /* Any change of skb->len requires recalculation of tso factor. */
1686 if (tcp_skb_pcount(skb) > 1)
1687 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1692 /* Calculate MSS not accounting any TCP options. */
1693 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1695 const struct tcp_sock *tp = tcp_sk(sk);
1696 const struct inet_connection_sock *icsk = inet_csk(sk);
1699 /* Calculate base mss without TCP options:
1700 It is MMS_S - sizeof(tcphdr) of rfc1122
1702 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1704 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1705 if (icsk->icsk_af_ops->net_frag_header_len) {
1706 const struct dst_entry *dst = __sk_dst_get(sk);
1708 if (dst && dst_allfrag(dst))
1709 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1712 /* Clamp it (mss_clamp does not include tcp options) */
1713 if (mss_now > tp->rx_opt.mss_clamp)
1714 mss_now = tp->rx_opt.mss_clamp;
1716 /* Now subtract optional transport overhead */
1717 mss_now -= icsk->icsk_ext_hdr_len;
1719 /* Then reserve room for full set of TCP options and 8 bytes of data */
1720 mss_now = max(mss_now,
1721 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss));
1725 /* Calculate MSS. Not accounting for SACKs here. */
1726 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1728 /* Subtract TCP options size, not including SACKs */
1729 return __tcp_mtu_to_mss(sk, pmtu) -
1730 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1732 EXPORT_SYMBOL(tcp_mtu_to_mss);
1734 /* Inverse of above */
1735 int tcp_mss_to_mtu(struct sock *sk, int mss)
1737 const struct tcp_sock *tp = tcp_sk(sk);
1738 const struct inet_connection_sock *icsk = inet_csk(sk);
1742 tp->tcp_header_len +
1743 icsk->icsk_ext_hdr_len +
1744 icsk->icsk_af_ops->net_header_len;
1746 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1747 if (icsk->icsk_af_ops->net_frag_header_len) {
1748 const struct dst_entry *dst = __sk_dst_get(sk);
1750 if (dst && dst_allfrag(dst))
1751 mtu += icsk->icsk_af_ops->net_frag_header_len;
1755 EXPORT_SYMBOL(tcp_mss_to_mtu);
1757 /* MTU probing init per socket */
1758 void tcp_mtup_init(struct sock *sk)
1760 struct tcp_sock *tp = tcp_sk(sk);
1761 struct inet_connection_sock *icsk = inet_csk(sk);
1762 struct net *net = sock_net(sk);
1764 icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1765 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1766 icsk->icsk_af_ops->net_header_len;
1767 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss));
1768 icsk->icsk_mtup.probe_size = 0;
1769 if (icsk->icsk_mtup.enabled)
1770 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1772 EXPORT_SYMBOL(tcp_mtup_init);
1774 /* This function synchronize snd mss to current pmtu/exthdr set.
1776 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1777 for TCP options, but includes only bare TCP header.
1779 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1780 It is minimum of user_mss and mss received with SYN.
1781 It also does not include TCP options.
1783 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1785 tp->mss_cache is current effective sending mss, including
1786 all tcp options except for SACKs. It is evaluated,
1787 taking into account current pmtu, but never exceeds
1788 tp->rx_opt.mss_clamp.
1790 NOTE1. rfc1122 clearly states that advertised MSS
1791 DOES NOT include either tcp or ip options.
1793 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1794 are READ ONLY outside this function. --ANK (980731)
1796 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1798 struct tcp_sock *tp = tcp_sk(sk);
1799 struct inet_connection_sock *icsk = inet_csk(sk);
1802 if (icsk->icsk_mtup.search_high > pmtu)
1803 icsk->icsk_mtup.search_high = pmtu;
1805 mss_now = tcp_mtu_to_mss(sk, pmtu);
1806 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1808 /* And store cached results */
1809 icsk->icsk_pmtu_cookie = pmtu;
1810 if (icsk->icsk_mtup.enabled)
1811 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1812 tp->mss_cache = mss_now;
1816 EXPORT_SYMBOL(tcp_sync_mss);
1818 /* Compute the current effective MSS, taking SACKs and IP options,
1819 * and even PMTU discovery events into account.
1821 unsigned int tcp_current_mss(struct sock *sk)
1823 const struct tcp_sock *tp = tcp_sk(sk);
1824 const struct dst_entry *dst = __sk_dst_get(sk);
1826 unsigned int header_len;
1827 struct tcp_out_options opts;
1828 struct tcp_md5sig_key *md5;
1830 mss_now = tp->mss_cache;
1833 u32 mtu = dst_mtu(dst);
1834 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1835 mss_now = tcp_sync_mss(sk, mtu);
1838 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1839 sizeof(struct tcphdr);
1840 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1841 * some common options. If this is an odd packet (because we have SACK
1842 * blocks etc) then our calculated header_len will be different, and
1843 * we have to adjust mss_now correspondingly */
1844 if (header_len != tp->tcp_header_len) {
1845 int delta = (int) header_len - tp->tcp_header_len;
1852 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1853 * As additional protections, we do not touch cwnd in retransmission phases,
1854 * and if application hit its sndbuf limit recently.
1856 static void tcp_cwnd_application_limited(struct sock *sk)
1858 struct tcp_sock *tp = tcp_sk(sk);
1860 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1861 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1862 /* Limited by application or receiver window. */
1863 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1864 u32 win_used = max(tp->snd_cwnd_used, init_win);
1865 if (win_used < tcp_snd_cwnd(tp)) {
1866 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1867 tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1);
1869 tp->snd_cwnd_used = 0;
1871 tp->snd_cwnd_stamp = tcp_jiffies32;
1874 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1876 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1877 struct tcp_sock *tp = tcp_sk(sk);
1879 /* Track the strongest available signal of the degree to which the cwnd
1880 * is fully utilized. If cwnd-limited then remember that fact for the
1881 * current window. If not cwnd-limited then track the maximum number of
1882 * outstanding packets in the current window. (If cwnd-limited then we
1883 * chose to not update tp->max_packets_out to avoid an extra else
1884 * clause with no functional impact.)
1886 if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
1888 (!tp->is_cwnd_limited &&
1889 tp->packets_out > tp->max_packets_out)) {
1890 tp->is_cwnd_limited = is_cwnd_limited;
1891 tp->max_packets_out = tp->packets_out;
1892 tp->cwnd_usage_seq = tp->snd_nxt;
1895 if (tcp_is_cwnd_limited(sk)) {
1896 /* Network is feed fully. */
1897 tp->snd_cwnd_used = 0;
1898 tp->snd_cwnd_stamp = tcp_jiffies32;
1900 /* Network starves. */
1901 if (tp->packets_out > tp->snd_cwnd_used)
1902 tp->snd_cwnd_used = tp->packets_out;
1904 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
1905 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1906 !ca_ops->cong_control)
1907 tcp_cwnd_application_limited(sk);
1909 /* The following conditions together indicate the starvation
1910 * is caused by insufficient sender buffer:
1911 * 1) just sent some data (see tcp_write_xmit)
1912 * 2) not cwnd limited (this else condition)
1913 * 3) no more data to send (tcp_write_queue_empty())
1914 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1916 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1917 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1918 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1919 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1923 /* Minshall's variant of the Nagle send check. */
1924 static bool tcp_minshall_check(const struct tcp_sock *tp)
1926 return after(tp->snd_sml, tp->snd_una) &&
1927 !after(tp->snd_sml, tp->snd_nxt);
1930 /* Update snd_sml if this skb is under mss
1931 * Note that a TSO packet might end with a sub-mss segment
1932 * The test is really :
1933 * if ((skb->len % mss) != 0)
1934 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1935 * But we can avoid doing the divide again given we already have
1936 * skb_pcount = skb->len / mss_now
1938 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1939 const struct sk_buff *skb)
1941 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1942 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1945 /* Return false, if packet can be sent now without violation Nagle's rules:
1946 * 1. It is full sized. (provided by caller in %partial bool)
1947 * 2. Or it contains FIN. (already checked by caller)
1948 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1949 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1950 * With Minshall's modification: all sent small packets are ACKed.
1952 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1956 ((nonagle & TCP_NAGLE_CORK) ||
1957 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1960 /* Return how many segs we'd like on a TSO packet,
1961 * depending on current pacing rate, and how close the peer is.
1964 * - For close peers, we rather send bigger packets to reduce
1965 * cpu costs, because occasional losses will be repaired fast.
1966 * - For long distance/rtt flows, we would like to get ACK clocking
1967 * with 1 ACK per ms.
1969 * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting
1970 * in bigger TSO bursts. We we cut the RTT-based allowance in half
1971 * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance
1972 * is below 1500 bytes after 6 * ~500 usec = 3ms.
1974 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1977 unsigned long bytes;
1980 bytes = sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift);
1982 r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log);
1983 if (r < BITS_PER_TYPE(sk->sk_gso_max_size))
1984 bytes += sk->sk_gso_max_size >> r;
1986 bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size);
1988 return max_t(u32, bytes / mss_now, min_tso_segs);
1991 /* Return the number of segments we want in the skb we are transmitting.
1992 * See if congestion control module wants to decide; otherwise, autosize.
1994 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1996 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1997 u32 min_tso, tso_segs;
1999 min_tso = ca_ops->min_tso_segs ?
2000 ca_ops->min_tso_segs(sk) :
2001 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
2003 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
2004 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
2007 /* Returns the portion of skb which can be sent right away */
2008 static unsigned int tcp_mss_split_point(const struct sock *sk,
2009 const struct sk_buff *skb,
2010 unsigned int mss_now,
2011 unsigned int max_segs,
2014 const struct tcp_sock *tp = tcp_sk(sk);
2015 u32 partial, needed, window, max_len;
2017 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2018 max_len = mss_now * max_segs;
2020 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2023 needed = min(skb->len, window);
2025 if (max_len <= needed)
2028 partial = needed % mss_now;
2029 /* If last segment is not a full MSS, check if Nagle rules allow us
2030 * to include this last segment in this skb.
2031 * Otherwise, we'll split the skb at last MSS boundary
2033 if (tcp_nagle_check(partial != 0, tp, nonagle))
2034 return needed - partial;
2039 /* Can at least one segment of SKB be sent right now, according to the
2040 * congestion window rules? If so, return how many segments are allowed.
2042 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
2043 const struct sk_buff *skb)
2045 u32 in_flight, cwnd, halfcwnd;
2047 /* Don't be strict about the congestion window for the final FIN. */
2048 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2049 tcp_skb_pcount(skb) == 1)
2052 in_flight = tcp_packets_in_flight(tp);
2053 cwnd = tcp_snd_cwnd(tp);
2054 if (in_flight >= cwnd)
2057 /* For better scheduling, ensure we have at least
2058 * 2 GSO packets in flight.
2060 halfcwnd = max(cwnd >> 1, 1U);
2061 return min(halfcwnd, cwnd - in_flight);
2064 /* Initialize TSO state of a skb.
2065 * This must be invoked the first time we consider transmitting
2066 * SKB onto the wire.
2068 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
2070 int tso_segs = tcp_skb_pcount(skb);
2072 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
2073 tcp_set_skb_tso_segs(skb, mss_now);
2074 tso_segs = tcp_skb_pcount(skb);
2080 /* Return true if the Nagle test allows this packet to be
2083 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2084 unsigned int cur_mss, int nonagle)
2086 /* Nagle rule does not apply to frames, which sit in the middle of the
2087 * write_queue (they have no chances to get new data).
2089 * This is implemented in the callers, where they modify the 'nonagle'
2090 * argument based upon the location of SKB in the send queue.
2092 if (nonagle & TCP_NAGLE_PUSH)
2095 /* Don't use the nagle rule for urgent data (or for the final FIN). */
2096 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2099 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
2105 /* Does at least the first segment of SKB fit into the send window? */
2106 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2107 const struct sk_buff *skb,
2108 unsigned int cur_mss)
2110 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2112 if (skb->len > cur_mss)
2113 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2115 return !after(end_seq, tcp_wnd_end(tp));
2118 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2119 * which is put after SKB on the list. It is very much like
2120 * tcp_fragment() except that it may make several kinds of assumptions
2121 * in order to speed up the splitting operation. In particular, we
2122 * know that all the data is in scatter-gather pages, and that the
2123 * packet has never been sent out before (and thus is not cloned).
2125 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2126 unsigned int mss_now, gfp_t gfp)
2128 int nlen = skb->len - len;
2129 struct sk_buff *buff;
2132 /* All of a TSO frame must be composed of paged data. */
2133 if (skb->len != skb->data_len)
2134 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2135 skb, len, mss_now, gfp);
2137 buff = tcp_stream_alloc_skb(sk, 0, gfp, true);
2138 if (unlikely(!buff))
2140 skb_copy_decrypted(buff, skb);
2141 mptcp_skb_ext_copy(buff, skb);
2143 sk_wmem_queued_add(sk, buff->truesize);
2144 sk_mem_charge(sk, buff->truesize);
2145 buff->truesize += nlen;
2146 skb->truesize -= nlen;
2148 /* Correct the sequence numbers. */
2149 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2150 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2151 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2153 /* PSH and FIN should only be set in the second packet. */
2154 flags = TCP_SKB_CB(skb)->tcp_flags;
2155 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2156 TCP_SKB_CB(buff)->tcp_flags = flags;
2158 tcp_skb_fragment_eor(skb, buff);
2160 skb_split(skb, buff, len);
2161 tcp_fragment_tstamp(skb, buff);
2163 /* Fix up tso_factor for both original and new SKB. */
2164 tcp_set_skb_tso_segs(skb, mss_now);
2165 tcp_set_skb_tso_segs(buff, mss_now);
2167 /* Link BUFF into the send queue. */
2168 __skb_header_release(buff);
2169 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2174 /* Try to defer sending, if possible, in order to minimize the amount
2175 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2177 * This algorithm is from John Heffner.
2179 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2180 bool *is_cwnd_limited,
2181 bool *is_rwnd_limited,
2184 const struct inet_connection_sock *icsk = inet_csk(sk);
2185 u32 send_win, cong_win, limit, in_flight;
2186 struct tcp_sock *tp = tcp_sk(sk);
2187 struct sk_buff *head;
2191 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2194 /* Avoid bursty behavior by allowing defer
2195 * only if the last write was recent (1 ms).
2196 * Note that tp->tcp_wstamp_ns can be in the future if we have
2197 * packets waiting in a qdisc or device for EDT delivery.
2199 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2203 in_flight = tcp_packets_in_flight(tp);
2205 BUG_ON(tcp_skb_pcount(skb) <= 1);
2206 BUG_ON(tcp_snd_cwnd(tp) <= in_flight);
2208 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2210 /* From in_flight test above, we know that cwnd > in_flight. */
2211 cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache;
2213 limit = min(send_win, cong_win);
2215 /* If a full-sized TSO skb can be sent, do it. */
2216 if (limit >= max_segs * tp->mss_cache)
2219 /* Middle in queue won't get any more data, full sendable already? */
2220 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2223 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2225 u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache);
2227 /* If at least some fraction of a window is available,
2230 chunk /= win_divisor;
2234 /* Different approach, try not to defer past a single
2235 * ACK. Receiver should ACK every other full sized
2236 * frame, so if we have space for more than 3 frames
2239 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2243 /* TODO : use tsorted_sent_queue ? */
2244 head = tcp_rtx_queue_head(sk);
2247 delta = tp->tcp_clock_cache - head->tstamp;
2248 /* If next ACK is likely to come too late (half srtt), do not defer */
2249 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2252 /* Ok, it looks like it is advisable to defer.
2253 * Three cases are tracked :
2254 * 1) We are cwnd-limited
2255 * 2) We are rwnd-limited
2256 * 3) We are application limited.
2258 if (cong_win < send_win) {
2259 if (cong_win <= skb->len) {
2260 *is_cwnd_limited = true;
2264 if (send_win <= skb->len) {
2265 *is_rwnd_limited = true;
2270 /* If this packet won't get more data, do not wait. */
2271 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2272 TCP_SKB_CB(skb)->eor)
2281 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2283 struct inet_connection_sock *icsk = inet_csk(sk);
2284 struct tcp_sock *tp = tcp_sk(sk);
2285 struct net *net = sock_net(sk);
2289 interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2290 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2291 if (unlikely(delta >= interval * HZ)) {
2292 int mss = tcp_current_mss(sk);
2294 /* Update current search range */
2295 icsk->icsk_mtup.probe_size = 0;
2296 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2297 sizeof(struct tcphdr) +
2298 icsk->icsk_af_ops->net_header_len;
2299 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2301 /* Update probe time stamp */
2302 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2306 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2308 struct sk_buff *skb, *next;
2310 skb = tcp_send_head(sk);
2311 tcp_for_write_queue_from_safe(skb, next, sk) {
2312 if (len <= skb->len)
2315 if (unlikely(TCP_SKB_CB(skb)->eor) ||
2316 tcp_has_tx_tstamp(skb) ||
2317 !skb_pure_zcopy_same(skb, next))
2326 /* Create a new MTU probe if we are ready.
2327 * MTU probe is regularly attempting to increase the path MTU by
2328 * deliberately sending larger packets. This discovers routing
2329 * changes resulting in larger path MTUs.
2331 * Returns 0 if we should wait to probe (no cwnd available),
2332 * 1 if a probe was sent,
2335 static int tcp_mtu_probe(struct sock *sk)
2337 struct inet_connection_sock *icsk = inet_csk(sk);
2338 struct tcp_sock *tp = tcp_sk(sk);
2339 struct sk_buff *skb, *nskb, *next;
2340 struct net *net = sock_net(sk);
2347 /* Not currently probing/verifying,
2349 * have enough cwnd, and
2350 * not SACKing (the variable headers throw things off)
2352 if (likely(!icsk->icsk_mtup.enabled ||
2353 icsk->icsk_mtup.probe_size ||
2354 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2355 tcp_snd_cwnd(tp) < 11 ||
2356 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2359 /* Use binary search for probe_size between tcp_mss_base,
2360 * and current mss_clamp. if (search_high - search_low)
2361 * smaller than a threshold, backoff from probing.
2363 mss_now = tcp_current_mss(sk);
2364 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2365 icsk->icsk_mtup.search_low) >> 1);
2366 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2367 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2368 /* When misfortune happens, we are reprobing actively,
2369 * and then reprobe timer has expired. We stick with current
2370 * probing process by not resetting search range to its orignal.
2372 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2373 interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2374 /* Check whether enough time has elaplased for
2375 * another round of probing.
2377 tcp_mtu_check_reprobe(sk);
2381 /* Have enough data in the send queue to probe? */
2382 if (tp->write_seq - tp->snd_nxt < size_needed)
2385 if (tp->snd_wnd < size_needed)
2387 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2390 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2391 if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) {
2392 if (!tcp_packets_in_flight(tp))
2398 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2401 /* We're allowed to probe. Build it now. */
2402 nskb = tcp_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2405 sk_wmem_queued_add(sk, nskb->truesize);
2406 sk_mem_charge(sk, nskb->truesize);
2408 skb = tcp_send_head(sk);
2409 skb_copy_decrypted(nskb, skb);
2410 mptcp_skb_ext_copy(nskb, skb);
2412 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2413 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2414 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2416 tcp_insert_write_queue_before(nskb, skb, sk);
2417 tcp_highest_sack_replace(sk, skb, nskb);
2420 tcp_for_write_queue_from_safe(skb, next, sk) {
2421 copy = min_t(int, skb->len, probe_size - len);
2422 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2424 if (skb->len <= copy) {
2425 /* We've eaten all the data from this skb.
2427 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2428 /* If this is the last SKB we copy and eor is set
2429 * we need to propagate it to the new skb.
2431 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2432 tcp_skb_collapse_tstamp(nskb, skb);
2433 tcp_unlink_write_queue(skb, sk);
2434 tcp_wmem_free_skb(sk, skb);
2436 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2437 ~(TCPHDR_FIN|TCPHDR_PSH);
2438 if (!skb_shinfo(skb)->nr_frags) {
2439 skb_pull(skb, copy);
2441 __pskb_trim_head(skb, copy);
2442 tcp_set_skb_tso_segs(skb, mss_now);
2444 TCP_SKB_CB(skb)->seq += copy;
2449 if (len >= probe_size)
2452 tcp_init_tso_segs(nskb, nskb->len);
2454 /* We're ready to send. If this fails, the probe will
2455 * be resegmented into mss-sized pieces by tcp_write_xmit().
2457 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2458 /* Decrement cwnd here because we are sending
2459 * effectively two packets. */
2460 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1);
2461 tcp_event_new_data_sent(sk, nskb);
2463 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2464 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2465 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2473 static bool tcp_pacing_check(struct sock *sk)
2475 struct tcp_sock *tp = tcp_sk(sk);
2477 if (!tcp_needs_internal_pacing(sk))
2480 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2483 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2484 hrtimer_start(&tp->pacing_timer,
2485 ns_to_ktime(tp->tcp_wstamp_ns),
2486 HRTIMER_MODE_ABS_PINNED_SOFT);
2492 static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk)
2494 const struct rb_node *node = sk->tcp_rtx_queue.rb_node;
2496 /* No skb in the rtx queue. */
2500 /* Only one skb in rtx queue. */
2501 return !node->rb_left && !node->rb_right;
2504 /* TCP Small Queues :
2505 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2506 * (These limits are doubled for retransmits)
2508 * - better RTT estimation and ACK scheduling
2511 * Alas, some drivers / subsystems require a fair amount
2512 * of queued bytes to ensure line rate.
2513 * One example is wifi aggregation (802.11 AMPDU)
2515 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2516 unsigned int factor)
2518 unsigned long limit;
2520 limit = max_t(unsigned long,
2522 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2523 if (sk->sk_pacing_status == SK_PACING_NONE)
2524 limit = min_t(unsigned long, limit,
2525 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
2528 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2529 tcp_sk(sk)->tcp_tx_delay) {
2530 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2532 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2533 * approximate our needs assuming an ~100% skb->truesize overhead.
2534 * USEC_PER_SEC is approximated by 2^20.
2535 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2537 extra_bytes >>= (20 - 1);
2538 limit += extra_bytes;
2540 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2541 /* Always send skb if rtx queue is empty or has one skb.
2542 * No need to wait for TX completion to call us back,
2543 * after softirq/tasklet schedule.
2544 * This helps when TX completions are delayed too much.
2546 if (tcp_rtx_queue_empty_or_single_skb(sk))
2549 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2550 /* It is possible TX completion already happened
2551 * before we set TSQ_THROTTLED, so we must
2552 * test again the condition.
2554 smp_mb__after_atomic();
2555 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2561 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2563 const u32 now = tcp_jiffies32;
2564 enum tcp_chrono old = tp->chrono_type;
2566 if (old > TCP_CHRONO_UNSPEC)
2567 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2568 tp->chrono_start = now;
2569 tp->chrono_type = new;
2572 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2574 struct tcp_sock *tp = tcp_sk(sk);
2576 /* If there are multiple conditions worthy of tracking in a
2577 * chronograph then the highest priority enum takes precedence
2578 * over the other conditions. So that if something "more interesting"
2579 * starts happening, stop the previous chrono and start a new one.
2581 if (type > tp->chrono_type)
2582 tcp_chrono_set(tp, type);
2585 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2587 struct tcp_sock *tp = tcp_sk(sk);
2590 /* There are multiple conditions worthy of tracking in a
2591 * chronograph, so that the highest priority enum takes
2592 * precedence over the other conditions (see tcp_chrono_start).
2593 * If a condition stops, we only stop chrono tracking if
2594 * it's the "most interesting" or current chrono we are
2595 * tracking and starts busy chrono if we have pending data.
2597 if (tcp_rtx_and_write_queues_empty(sk))
2598 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2599 else if (type == tp->chrono_type)
2600 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2603 /* This routine writes packets to the network. It advances the
2604 * send_head. This happens as incoming acks open up the remote
2607 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2608 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2609 * account rare use of URG, this is not a big flaw.
2611 * Send at most one packet when push_one > 0. Temporarily ignore
2612 * cwnd limit to force at most one packet out when push_one == 2.
2614 * Returns true, if no segments are in flight and we have queued segments,
2615 * but cannot send anything now because of SWS or another problem.
2617 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2618 int push_one, gfp_t gfp)
2620 struct tcp_sock *tp = tcp_sk(sk);
2621 struct sk_buff *skb;
2622 unsigned int tso_segs, sent_pkts;
2625 bool is_cwnd_limited = false, is_rwnd_limited = false;
2630 tcp_mstamp_refresh(tp);
2632 /* Do MTU probing. */
2633 result = tcp_mtu_probe(sk);
2636 } else if (result > 0) {
2641 max_segs = tcp_tso_segs(sk, mss_now);
2642 while ((skb = tcp_send_head(sk))) {
2645 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2646 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2647 tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2648 skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
2649 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2650 tcp_init_tso_segs(skb, mss_now);
2651 goto repair; /* Skip network transmission */
2654 if (tcp_pacing_check(sk))
2657 tso_segs = tcp_init_tso_segs(skb, mss_now);
2660 cwnd_quota = tcp_cwnd_test(tp, skb);
2663 /* Force out a loss probe pkt. */
2669 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2670 is_rwnd_limited = true;
2674 if (tso_segs == 1) {
2675 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2676 (tcp_skb_is_last(sk, skb) ?
2677 nonagle : TCP_NAGLE_PUSH))))
2681 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2682 &is_rwnd_limited, max_segs))
2687 if (tso_segs > 1 && !tcp_urg_mode(tp))
2688 limit = tcp_mss_split_point(sk, skb, mss_now,
2694 if (skb->len > limit &&
2695 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2698 if (tcp_small_queue_check(sk, skb, 0))
2701 /* Argh, we hit an empty skb(), presumably a thread
2702 * is sleeping in sendmsg()/sk_stream_wait_memory().
2703 * We do not want to send a pure-ack packet and have
2704 * a strange looking rtx queue with empty packet(s).
2706 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2709 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2713 /* Advance the send_head. This one is sent out.
2714 * This call will increment packets_out.
2716 tcp_event_new_data_sent(sk, skb);
2718 tcp_minshall_update(tp, mss_now, skb);
2719 sent_pkts += tcp_skb_pcount(skb);
2725 if (is_rwnd_limited)
2726 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2728 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2730 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp));
2731 if (likely(sent_pkts || is_cwnd_limited))
2732 tcp_cwnd_validate(sk, is_cwnd_limited);
2734 if (likely(sent_pkts)) {
2735 if (tcp_in_cwnd_reduction(sk))
2736 tp->prr_out += sent_pkts;
2738 /* Send one loss probe per tail loss episode. */
2740 tcp_schedule_loss_probe(sk, false);
2743 return !tp->packets_out && !tcp_write_queue_empty(sk);
2746 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2748 struct inet_connection_sock *icsk = inet_csk(sk);
2749 struct tcp_sock *tp = tcp_sk(sk);
2750 u32 timeout, timeout_us, rto_delta_us;
2753 /* Don't do any loss probe on a Fast Open connection before 3WHS
2756 if (rcu_access_pointer(tp->fastopen_rsk))
2759 early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
2760 /* Schedule a loss probe in 2*RTT for SACK capable connections
2761 * not in loss recovery, that are either limited by cwnd or application.
2763 if ((early_retrans != 3 && early_retrans != 4) ||
2764 !tp->packets_out || !tcp_is_sack(tp) ||
2765 (icsk->icsk_ca_state != TCP_CA_Open &&
2766 icsk->icsk_ca_state != TCP_CA_CWR))
2769 /* Probe timeout is 2*rtt. Add minimum RTO to account
2770 * for delayed ack when there's one outstanding packet. If no RTT
2771 * sample is available then probe after TCP_TIMEOUT_INIT.
2774 timeout_us = tp->srtt_us >> 2;
2775 if (tp->packets_out == 1)
2776 timeout_us += tcp_rto_min_us(sk);
2778 timeout_us += TCP_TIMEOUT_MIN_US;
2779 timeout = usecs_to_jiffies(timeout_us);
2781 timeout = TCP_TIMEOUT_INIT;
2784 /* If the RTO formula yields an earlier time, then use that time. */
2785 rto_delta_us = advancing_rto ?
2786 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2787 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2788 if (rto_delta_us > 0)
2789 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2791 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2795 /* Thanks to skb fast clones, we can detect if a prior transmit of
2796 * a packet is still in a qdisc or driver queue.
2797 * In this case, there is very little point doing a retransmit !
2799 static bool skb_still_in_host_queue(struct sock *sk,
2800 const struct sk_buff *skb)
2802 if (unlikely(skb_fclone_busy(sk, skb))) {
2803 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2804 smp_mb__after_atomic();
2805 if (skb_fclone_busy(sk, skb)) {
2806 NET_INC_STATS(sock_net(sk),
2807 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2814 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2815 * retransmit the last segment.
2817 void tcp_send_loss_probe(struct sock *sk)
2819 struct tcp_sock *tp = tcp_sk(sk);
2820 struct sk_buff *skb;
2822 int mss = tcp_current_mss(sk);
2824 /* At most one outstanding TLP */
2825 if (tp->tlp_high_seq)
2828 tp->tlp_retrans = 0;
2829 skb = tcp_send_head(sk);
2830 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2831 pcount = tp->packets_out;
2832 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2833 if (tp->packets_out > pcount)
2837 skb = skb_rb_last(&sk->tcp_rtx_queue);
2838 if (unlikely(!skb)) {
2839 WARN_ONCE(tp->packets_out,
2840 "invalid inflight: %u state %u cwnd %u mss %d\n",
2841 tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss);
2842 inet_csk(sk)->icsk_pending = 0;
2846 if (skb_still_in_host_queue(sk, skb))
2849 pcount = tcp_skb_pcount(skb);
2850 if (WARN_ON(!pcount))
2853 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2854 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2855 (pcount - 1) * mss, mss,
2858 skb = skb_rb_next(skb);
2861 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2864 if (__tcp_retransmit_skb(sk, skb, 1))
2867 tp->tlp_retrans = 1;
2870 /* Record snd_nxt for loss detection. */
2871 tp->tlp_high_seq = tp->snd_nxt;
2873 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2874 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2875 inet_csk(sk)->icsk_pending = 0;
2880 /* Push out any pending frames which were held back due to
2881 * TCP_CORK or attempt at coalescing tiny packets.
2882 * The socket must be locked by the caller.
2884 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2887 /* If we are closed, the bytes will have to remain here.
2888 * In time closedown will finish, we empty the write queue and
2889 * all will be happy.
2891 if (unlikely(sk->sk_state == TCP_CLOSE))
2894 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2895 sk_gfp_mask(sk, GFP_ATOMIC)))
2896 tcp_check_probe_timer(sk);
2899 /* Send _single_ skb sitting at the send head. This function requires
2900 * true push pending frames to setup probe timer etc.
2902 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2904 struct sk_buff *skb = tcp_send_head(sk);
2906 BUG_ON(!skb || skb->len < mss_now);
2908 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2911 /* This function returns the amount that we can raise the
2912 * usable window based on the following constraints
2914 * 1. The window can never be shrunk once it is offered (RFC 793)
2915 * 2. We limit memory per socket
2918 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2919 * RECV.NEXT + RCV.WIN fixed until:
2920 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2922 * i.e. don't raise the right edge of the window until you can raise
2923 * it at least MSS bytes.
2925 * Unfortunately, the recommended algorithm breaks header prediction,
2926 * since header prediction assumes th->window stays fixed.
2928 * Strictly speaking, keeping th->window fixed violates the receiver
2929 * side SWS prevention criteria. The problem is that under this rule
2930 * a stream of single byte packets will cause the right side of the
2931 * window to always advance by a single byte.
2933 * Of course, if the sender implements sender side SWS prevention
2934 * then this will not be a problem.
2936 * BSD seems to make the following compromise:
2938 * If the free space is less than the 1/4 of the maximum
2939 * space available and the free space is less than 1/2 mss,
2940 * then set the window to 0.
2941 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2942 * Otherwise, just prevent the window from shrinking
2943 * and from being larger than the largest representable value.
2945 * This prevents incremental opening of the window in the regime
2946 * where TCP is limited by the speed of the reader side taking
2947 * data out of the TCP receive queue. It does nothing about
2948 * those cases where the window is constrained on the sender side
2949 * because the pipeline is full.
2951 * BSD also seems to "accidentally" limit itself to windows that are a
2952 * multiple of MSS, at least until the free space gets quite small.
2953 * This would appear to be a side effect of the mbuf implementation.
2954 * Combining these two algorithms results in the observed behavior
2955 * of having a fixed window size at almost all times.
2957 * Below we obtain similar behavior by forcing the offered window to
2958 * a multiple of the mss when it is feasible to do so.
2960 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2961 * Regular options like TIMESTAMP are taken into account.
2963 u32 __tcp_select_window(struct sock *sk)
2965 struct inet_connection_sock *icsk = inet_csk(sk);
2966 struct tcp_sock *tp = tcp_sk(sk);
2967 struct net *net = sock_net(sk);
2968 /* MSS for the peer's data. Previous versions used mss_clamp
2969 * here. I don't know if the value based on our guesses
2970 * of peer's MSS is better for the performance. It's more correct
2971 * but may be worse for the performance because of rcv_mss
2972 * fluctuations. --SAW 1998/11/1
2974 int mss = icsk->icsk_ack.rcv_mss;
2975 int free_space = tcp_space(sk);
2976 int allowed_space = tcp_full_space(sk);
2977 int full_space, window;
2979 if (sk_is_mptcp(sk))
2980 mptcp_space(sk, &free_space, &allowed_space);
2982 full_space = min_t(int, tp->window_clamp, allowed_space);
2984 if (unlikely(mss > full_space)) {
2990 /* Only allow window shrink if the sysctl is enabled and we have
2991 * a non-zero scaling factor in effect.
2993 if (READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) && tp->rx_opt.rcv_wscale)
2994 goto shrink_window_allowed;
2996 /* do not allow window to shrink */
2998 if (free_space < (full_space >> 1)) {
2999 icsk->icsk_ack.quick = 0;
3001 if (tcp_under_memory_pressure(sk))
3002 tcp_adjust_rcv_ssthresh(sk);
3004 /* free_space might become our new window, make sure we don't
3005 * increase it due to wscale.
3007 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3009 /* if free space is less than mss estimate, or is below 1/16th
3010 * of the maximum allowed, try to move to zero-window, else
3011 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
3012 * new incoming data is dropped due to memory limits.
3013 * With large window, mss test triggers way too late in order
3014 * to announce zero window in time before rmem limit kicks in.
3016 if (free_space < (allowed_space >> 4) || free_space < mss)
3020 if (free_space > tp->rcv_ssthresh)
3021 free_space = tp->rcv_ssthresh;
3023 /* Don't do rounding if we are using window scaling, since the
3024 * scaled window will not line up with the MSS boundary anyway.
3026 if (tp->rx_opt.rcv_wscale) {
3027 window = free_space;
3029 /* Advertise enough space so that it won't get scaled away.
3030 * Import case: prevent zero window announcement if
3031 * 1<<rcv_wscale > mss.
3033 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
3035 window = tp->rcv_wnd;
3036 /* Get the largest window that is a nice multiple of mss.
3037 * Window clamp already applied above.
3038 * If our current window offering is within 1 mss of the
3039 * free space we just keep it. This prevents the divide
3040 * and multiply from happening most of the time.
3041 * We also don't do any window rounding when the free space
3044 if (window <= free_space - mss || window > free_space)
3045 window = rounddown(free_space, mss);
3046 else if (mss == full_space &&
3047 free_space > window + (full_space >> 1))
3048 window = free_space;
3053 shrink_window_allowed:
3054 /* new window should always be an exact multiple of scaling factor */
3055 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3057 if (free_space < (full_space >> 1)) {
3058 icsk->icsk_ack.quick = 0;
3060 if (tcp_under_memory_pressure(sk))
3061 tcp_adjust_rcv_ssthresh(sk);
3063 /* if free space is too low, return a zero window */
3064 if (free_space < (allowed_space >> 4) || free_space < mss ||
3065 free_space < (1 << tp->rx_opt.rcv_wscale))
3069 if (free_space > tp->rcv_ssthresh) {
3070 free_space = tp->rcv_ssthresh;
3071 /* new window should always be an exact multiple of scaling factor
3073 * For this case, we ALIGN "up" (increase free_space) because
3074 * we know free_space is not zero here, it has been reduced from
3075 * the memory-based limit, and rcv_ssthresh is not a hard limit
3076 * (unlike sk_rcvbuf).
3078 free_space = ALIGN(free_space, (1 << tp->rx_opt.rcv_wscale));
3084 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3085 const struct sk_buff *next_skb)
3087 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3088 const struct skb_shared_info *next_shinfo =
3089 skb_shinfo(next_skb);
3090 struct skb_shared_info *shinfo = skb_shinfo(skb);
3092 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3093 shinfo->tskey = next_shinfo->tskey;
3094 TCP_SKB_CB(skb)->txstamp_ack |=
3095 TCP_SKB_CB(next_skb)->txstamp_ack;
3099 /* Collapses two adjacent SKB's during retransmission. */
3100 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3102 struct tcp_sock *tp = tcp_sk(sk);
3103 struct sk_buff *next_skb = skb_rb_next(skb);
3106 next_skb_size = next_skb->len;
3108 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3110 if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3113 tcp_highest_sack_replace(sk, next_skb, skb);
3115 /* Update sequence range on original skb. */
3116 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3118 /* Merge over control information. This moves PSH/FIN etc. over */
3119 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3121 /* All done, get rid of second SKB and account for it so
3122 * packet counting does not break.
3124 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3125 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3127 /* changed transmit queue under us so clear hints */
3128 tcp_clear_retrans_hints_partial(tp);
3129 if (next_skb == tp->retransmit_skb_hint)
3130 tp->retransmit_skb_hint = skb;
3132 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3134 tcp_skb_collapse_tstamp(skb, next_skb);
3136 tcp_rtx_queue_unlink_and_free(next_skb, sk);
3140 /* Check if coalescing SKBs is legal. */
3141 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3143 if (tcp_skb_pcount(skb) > 1)
3145 if (skb_cloned(skb))
3147 /* Some heuristics for collapsing over SACK'd could be invented */
3148 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3154 /* Collapse packets in the retransmit queue to make to create
3155 * less packets on the wire. This is only done on retransmission.
3157 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3160 struct tcp_sock *tp = tcp_sk(sk);
3161 struct sk_buff *skb = to, *tmp;
3164 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
3166 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3169 skb_rbtree_walk_from_safe(skb, tmp) {
3170 if (!tcp_can_collapse(sk, skb))
3173 if (!tcp_skb_can_collapse(to, skb))
3186 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3189 if (!tcp_collapse_retrans(sk, to))
3194 /* This retransmits one SKB. Policy decisions and retransmit queue
3195 * state updates are done by the caller. Returns non-zero if an
3196 * error occurred which prevented the send.
3198 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3200 struct inet_connection_sock *icsk = inet_csk(sk);
3201 struct tcp_sock *tp = tcp_sk(sk);
3202 unsigned int cur_mss;
3206 /* Inconclusive MTU probe */
3207 if (icsk->icsk_mtup.probe_size)
3208 icsk->icsk_mtup.probe_size = 0;
3210 if (skb_still_in_host_queue(sk, skb))
3214 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
3215 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3216 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
3217 TCP_SKB_CB(skb)->seq++;
3220 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3224 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3228 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3229 return -EHOSTUNREACH; /* Routing failure or similar. */
3231 cur_mss = tcp_current_mss(sk);
3232 avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3234 /* If receiver has shrunk his window, and skb is out of
3235 * new window, do not retransmit it. The exception is the
3236 * case, when window is shrunk to zero. In this case
3237 * our retransmit of one segment serves as a zero window probe.
3239 if (avail_wnd <= 0) {
3240 if (TCP_SKB_CB(skb)->seq != tp->snd_una)
3242 avail_wnd = cur_mss;
3245 len = cur_mss * segs;
3246 if (len > avail_wnd) {
3247 len = rounddown(avail_wnd, cur_mss);
3251 if (skb->len > len) {
3252 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3253 cur_mss, GFP_ATOMIC))
3254 return -ENOMEM; /* We'll try again later. */
3256 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
3259 diff = tcp_skb_pcount(skb);
3260 tcp_set_skb_tso_segs(skb, cur_mss);
3261 diff -= tcp_skb_pcount(skb);
3263 tcp_adjust_pcount(sk, skb, diff);
3264 avail_wnd = min_t(int, avail_wnd, cur_mss);
3265 if (skb->len < avail_wnd)
3266 tcp_retrans_try_collapse(sk, skb, avail_wnd);
3269 /* RFC3168, section 6.1.1.1. ECN fallback */
3270 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3271 tcp_ecn_clear_syn(sk, skb);
3273 /* Update global and local TCP statistics. */
3274 segs = tcp_skb_pcount(skb);
3275 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3276 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3277 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3278 tp->total_retrans += segs;
3279 tp->bytes_retrans += skb->len;
3281 /* make sure skb->data is aligned on arches that require it
3282 * and check if ack-trimming & collapsing extended the headroom
3283 * beyond what csum_start can cover.
3285 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3286 skb_headroom(skb) >= 0xFFFF)) {
3287 struct sk_buff *nskb;
3289 tcp_skb_tsorted_save(skb) {
3290 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3293 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3297 } tcp_skb_tsorted_restore(skb);
3300 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3301 tcp_rate_skb_sent(sk, skb);
3304 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3307 /* To avoid taking spuriously low RTT samples based on a timestamp
3308 * for a transmit that never happened, always mark EVER_RETRANS
3310 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3312 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3313 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3314 TCP_SKB_CB(skb)->seq, segs, err);
3317 trace_tcp_retransmit_skb(sk, skb);
3318 } else if (err != -EBUSY) {
3319 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3324 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3326 struct tcp_sock *tp = tcp_sk(sk);
3327 int err = __tcp_retransmit_skb(sk, skb, segs);
3330 #if FASTRETRANS_DEBUG > 0
3331 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3332 net_dbg_ratelimited("retrans_out leaked\n");
3335 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3336 tp->retrans_out += tcp_skb_pcount(skb);
3339 /* Save stamp of the first (attempted) retransmit. */
3340 if (!tp->retrans_stamp)
3341 tp->retrans_stamp = tcp_skb_timestamp(skb);
3343 if (tp->undo_retrans < 0)
3344 tp->undo_retrans = 0;
3345 tp->undo_retrans += tcp_skb_pcount(skb);
3349 /* This gets called after a retransmit timeout, and the initially
3350 * retransmitted data is acknowledged. It tries to continue
3351 * resending the rest of the retransmit queue, until either
3352 * we've sent it all or the congestion window limit is reached.
3354 void tcp_xmit_retransmit_queue(struct sock *sk)
3356 const struct inet_connection_sock *icsk = inet_csk(sk);
3357 struct sk_buff *skb, *rtx_head, *hole = NULL;
3358 struct tcp_sock *tp = tcp_sk(sk);
3359 bool rearm_timer = false;
3363 if (!tp->packets_out)
3366 rtx_head = tcp_rtx_queue_head(sk);
3367 skb = tp->retransmit_skb_hint ?: rtx_head;
3368 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3369 skb_rbtree_walk_from(skb) {
3373 if (tcp_pacing_check(sk))
3376 /* we could do better than to assign each time */
3378 tp->retransmit_skb_hint = skb;
3380 segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp);
3383 sacked = TCP_SKB_CB(skb)->sacked;
3384 /* In case tcp_shift_skb_data() have aggregated large skbs,
3385 * we need to make sure not sending too bigs TSO packets
3387 segs = min_t(int, segs, max_segs);
3389 if (tp->retrans_out >= tp->lost_out) {
3391 } else if (!(sacked & TCPCB_LOST)) {
3392 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3397 if (icsk->icsk_ca_state != TCP_CA_Loss)
3398 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3400 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3403 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3406 if (tcp_small_queue_check(sk, skb, 1))
3409 if (tcp_retransmit_skb(sk, skb, segs))
3412 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3414 if (tcp_in_cwnd_reduction(sk))
3415 tp->prr_out += tcp_skb_pcount(skb);
3417 if (skb == rtx_head &&
3418 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3423 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3424 inet_csk(sk)->icsk_rto,
3428 /* We allow to exceed memory limits for FIN packets to expedite
3429 * connection tear down and (memory) recovery.
3430 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3431 * or even be forced to close flow without any FIN.
3432 * In general, we want to allow one skb per socket to avoid hangs
3433 * with edge trigger epoll()
3435 void sk_forced_mem_schedule(struct sock *sk, int size)
3439 delta = size - sk->sk_forward_alloc;
3442 amt = sk_mem_pages(delta);
3443 sk_forward_alloc_add(sk, amt << PAGE_SHIFT);
3444 sk_memory_allocated_add(sk, amt);
3446 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3447 mem_cgroup_charge_skmem(sk->sk_memcg, amt,
3448 gfp_memcg_charge() | __GFP_NOFAIL);
3451 /* Send a FIN. The caller locks the socket for us.
3452 * We should try to send a FIN packet really hard, but eventually give up.
3454 void tcp_send_fin(struct sock *sk)
3456 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3457 struct tcp_sock *tp = tcp_sk(sk);
3459 /* Optimization, tack on the FIN if we have one skb in write queue and
3460 * this skb was not yet sent, or we are under memory pressure.
3461 * Note: in the latter case, FIN packet will be sent after a timeout,
3462 * as TCP stack thinks it has already been transmitted.
3465 if (!tskb && tcp_under_memory_pressure(sk))
3466 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3469 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3470 TCP_SKB_CB(tskb)->end_seq++;
3473 /* This means tskb was already sent.
3474 * Pretend we included the FIN on previous transmit.
3475 * We need to set tp->snd_nxt to the value it would have
3476 * if FIN had been sent. This is because retransmit path
3477 * does not change tp->snd_nxt.
3479 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3483 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3487 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3488 skb_reserve(skb, MAX_TCP_HEADER);
3489 sk_forced_mem_schedule(sk, skb->truesize);
3490 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3491 tcp_init_nondata_skb(skb, tp->write_seq,
3492 TCPHDR_ACK | TCPHDR_FIN);
3493 tcp_queue_skb(sk, skb);
3495 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3498 /* We get here when a process closes a file descriptor (either due to
3499 * an explicit close() or as a byproduct of exit()'ing) and there
3500 * was unread data in the receive queue. This behavior is recommended
3501 * by RFC 2525, section 2.17. -DaveM
3503 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3505 struct sk_buff *skb;
3507 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3509 /* NOTE: No TCP options attached and we never retransmit this. */
3510 skb = alloc_skb(MAX_TCP_HEADER, priority);
3512 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3516 /* Reserve space for headers and prepare control bits. */
3517 skb_reserve(skb, MAX_TCP_HEADER);
3518 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3519 TCPHDR_ACK | TCPHDR_RST);
3520 tcp_mstamp_refresh(tcp_sk(sk));
3522 if (tcp_transmit_skb(sk, skb, 0, priority))
3523 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3525 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3526 * skb here is different to the troublesome skb, so use NULL
3528 trace_tcp_send_reset(sk, NULL);
3531 /* Send a crossed SYN-ACK during socket establishment.
3532 * WARNING: This routine must only be called when we have already sent
3533 * a SYN packet that crossed the incoming SYN that caused this routine
3534 * to get called. If this assumption fails then the initial rcv_wnd
3535 * and rcv_wscale values will not be correct.
3537 int tcp_send_synack(struct sock *sk)
3539 struct sk_buff *skb;
3541 skb = tcp_rtx_queue_head(sk);
3542 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3543 pr_err("%s: wrong queue state\n", __func__);
3546 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3547 if (skb_cloned(skb)) {
3548 struct sk_buff *nskb;
3550 tcp_skb_tsorted_save(skb) {
3551 nskb = skb_copy(skb, GFP_ATOMIC);
3552 } tcp_skb_tsorted_restore(skb);
3555 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3556 tcp_highest_sack_replace(sk, skb, nskb);
3557 tcp_rtx_queue_unlink_and_free(skb, sk);
3558 __skb_header_release(nskb);
3559 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3560 sk_wmem_queued_add(sk, nskb->truesize);
3561 sk_mem_charge(sk, nskb->truesize);
3565 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3566 tcp_ecn_send_synack(sk, skb);
3568 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3572 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3573 * @sk: listener socket
3574 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3575 * should not use it again.
3576 * @req: request_sock pointer
3577 * @foc: cookie for tcp fast open
3578 * @synack_type: Type of synack to prepare
3579 * @syn_skb: SYN packet just received. It could be NULL for rtx case.
3581 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3582 struct request_sock *req,
3583 struct tcp_fastopen_cookie *foc,
3584 enum tcp_synack_type synack_type,
3585 struct sk_buff *syn_skb)
3587 struct inet_request_sock *ireq = inet_rsk(req);
3588 const struct tcp_sock *tp = tcp_sk(sk);
3589 struct tcp_md5sig_key *md5 = NULL;
3590 struct tcp_out_options opts;
3591 struct sk_buff *skb;
3592 int tcp_header_size;
3597 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3598 if (unlikely(!skb)) {
3602 /* Reserve space for headers. */
3603 skb_reserve(skb, MAX_TCP_HEADER);
3605 switch (synack_type) {
3606 case TCP_SYNACK_NORMAL:
3607 skb_set_owner_w(skb, req_to_sk(req));
3609 case TCP_SYNACK_COOKIE:
3610 /* Under synflood, we do not attach skb to a socket,
3611 * to avoid false sharing.
3614 case TCP_SYNACK_FASTOPEN:
3615 /* sk is a const pointer, because we want to express multiple
3616 * cpu might call us concurrently.
3617 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3619 skb_set_owner_w(skb, (struct sock *)sk);
3622 skb_dst_set(skb, dst);
3624 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3626 memset(&opts, 0, sizeof(opts));
3627 now = tcp_clock_ns();
3628 #ifdef CONFIG_SYN_COOKIES
3629 if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3630 skb_set_delivery_time(skb, cookie_init_timestamp(req, now),
3635 skb_set_delivery_time(skb, now, true);
3636 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3637 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3640 #ifdef CONFIG_TCP_MD5SIG
3642 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3644 skb_set_hash(skb, READ_ONCE(tcp_rsk(req)->txhash), PKT_HASH_TYPE_L4);
3645 /* bpf program will be interested in the tcp_flags */
3646 TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3647 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3649 syn_skb) + sizeof(*th);
3651 skb_push(skb, tcp_header_size);
3652 skb_reset_transport_header(skb);
3654 th = (struct tcphdr *)skb->data;
3655 memset(th, 0, sizeof(struct tcphdr));
3658 tcp_ecn_make_synack(req, th);
3659 th->source = htons(ireq->ir_num);
3660 th->dest = ireq->ir_rmt_port;
3661 skb->mark = ireq->ir_mark;
3662 skb->ip_summed = CHECKSUM_PARTIAL;
3663 th->seq = htonl(tcp_rsk(req)->snt_isn);
3664 /* XXX data is queued and acked as is. No buffer/window check */
3665 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3667 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3668 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3669 tcp_options_write(th, NULL, &opts);
3670 th->doff = (tcp_header_size >> 2);
3671 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3673 #ifdef CONFIG_TCP_MD5SIG
3674 /* Okay, we have all we need - do the md5 hash if needed */
3676 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3677 md5, req_to_sk(req), skb);
3681 bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
3682 synack_type, &opts);
3684 skb_set_delivery_time(skb, now, true);
3685 tcp_add_tx_delay(skb, tp);
3689 EXPORT_SYMBOL(tcp_make_synack);
3691 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3693 struct inet_connection_sock *icsk = inet_csk(sk);
3694 const struct tcp_congestion_ops *ca;
3695 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3697 if (ca_key == TCP_CA_UNSPEC)
3701 ca = tcp_ca_find_key(ca_key);
3702 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3703 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3704 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3705 icsk->icsk_ca_ops = ca;
3710 /* Do all connect socket setups that can be done AF independent. */
3711 static void tcp_connect_init(struct sock *sk)
3713 const struct dst_entry *dst = __sk_dst_get(sk);
3714 struct tcp_sock *tp = tcp_sk(sk);
3718 /* We'll fix this up when we get a response from the other end.
3719 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3721 tp->tcp_header_len = sizeof(struct tcphdr);
3722 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps))
3723 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3725 #ifdef CONFIG_TCP_MD5SIG
3726 if (tp->af_specific->md5_lookup(sk, sk))
3727 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3730 /* If user gave his TCP_MAXSEG, record it to clamp */
3731 if (tp->rx_opt.user_mss)
3732 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3735 tcp_sync_mss(sk, dst_mtu(dst));
3737 tcp_ca_dst_init(sk, dst);
3739 if (!tp->window_clamp)
3740 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3741 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3743 tcp_initialize_rcv_mss(sk);
3745 /* limit the window selection if the user enforce a smaller rx buffer */
3746 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3747 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3748 tp->window_clamp = tcp_full_space(sk);
3750 rcv_wnd = tcp_rwnd_init_bpf(sk);
3752 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3754 tcp_select_initial_window(sk, tcp_full_space(sk),
3755 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3758 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling),
3762 tp->rx_opt.rcv_wscale = rcv_wscale;
3763 tp->rcv_ssthresh = tp->rcv_wnd;
3766 sock_reset_flag(sk, SOCK_DONE);
3769 tcp_write_queue_purge(sk);
3770 tp->snd_una = tp->write_seq;
3771 tp->snd_sml = tp->write_seq;
3772 tp->snd_up = tp->write_seq;
3773 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3775 if (likely(!tp->repair))
3778 tp->rcv_tstamp = tcp_jiffies32;
3779 tp->rcv_wup = tp->rcv_nxt;
3780 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3782 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3783 inet_csk(sk)->icsk_retransmits = 0;
3784 tcp_clear_retrans(tp);
3787 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3789 struct tcp_sock *tp = tcp_sk(sk);
3790 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3792 tcb->end_seq += skb->len;
3793 __skb_header_release(skb);
3794 sk_wmem_queued_add(sk, skb->truesize);
3795 sk_mem_charge(sk, skb->truesize);
3796 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3797 tp->packets_out += tcp_skb_pcount(skb);
3800 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3801 * queue a data-only packet after the regular SYN, such that regular SYNs
3802 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3803 * only the SYN sequence, the data are retransmitted in the first ACK.
3804 * If cookie is not cached or other error occurs, falls back to send a
3805 * regular SYN with Fast Open cookie request option.
3807 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3809 struct inet_connection_sock *icsk = inet_csk(sk);
3810 struct tcp_sock *tp = tcp_sk(sk);
3811 struct tcp_fastopen_request *fo = tp->fastopen_req;
3813 struct sk_buff *syn_data;
3815 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3816 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3819 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3820 * user-MSS. Reserve maximum option space for middleboxes that add
3821 * private TCP options. The cost is reduced data space in SYN :(
3823 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3824 /* Sync mss_cache after updating the mss_clamp */
3825 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3827 space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3828 MAX_TCP_OPTION_SPACE;
3830 space = min_t(size_t, space, fo->size);
3832 /* limit to order-0 allocations */
3833 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3835 syn_data = tcp_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3838 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3840 int copied = copy_from_iter(skb_put(syn_data, space), space,
3841 &fo->data->msg_iter);
3842 if (unlikely(!copied)) {
3843 tcp_skb_tsorted_anchor_cleanup(syn_data);
3844 kfree_skb(syn_data);
3847 if (copied != space) {
3848 skb_trim(syn_data, copied);
3851 skb_zcopy_set(syn_data, fo->uarg, NULL);
3853 /* No more data pending in inet_wait_for_connect() */
3854 if (space == fo->size)
3858 tcp_connect_queue_skb(sk, syn_data);
3860 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3862 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3864 skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true);
3866 /* Now full SYN+DATA was cloned and sent (or not),
3867 * remove the SYN from the original skb (syn_data)
3868 * we keep in write queue in case of a retransmit, as we
3869 * also have the SYN packet (with no data) in the same queue.
3871 TCP_SKB_CB(syn_data)->seq++;
3872 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3874 tp->syn_data = (fo->copied > 0);
3875 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3876 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3880 /* data was not sent, put it in write_queue */
3881 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3882 tp->packets_out -= tcp_skb_pcount(syn_data);
3885 /* Send a regular SYN with Fast Open cookie request option */
3886 if (fo->cookie.len > 0)
3888 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3890 tp->syn_fastopen = 0;
3892 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3896 /* Build a SYN and send it off. */
3897 int tcp_connect(struct sock *sk)
3899 struct tcp_sock *tp = tcp_sk(sk);
3900 struct sk_buff *buff;
3903 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3905 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3906 return -EHOSTUNREACH; /* Routing failure or similar. */
3908 tcp_connect_init(sk);
3910 if (unlikely(tp->repair)) {
3911 tcp_finish_connect(sk, NULL);
3915 buff = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3916 if (unlikely(!buff))
3919 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3920 tcp_mstamp_refresh(tp);
3921 tp->retrans_stamp = tcp_time_stamp(tp);
3922 tcp_connect_queue_skb(sk, buff);
3923 tcp_ecn_send_syn(sk, buff);
3924 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3926 /* Send off SYN; include data in Fast Open. */
3927 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3928 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3929 if (err == -ECONNREFUSED)
3932 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3933 * in order to make this packet get counted in tcpOutSegs.
3935 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3936 tp->pushed_seq = tp->write_seq;
3937 buff = tcp_send_head(sk);
3938 if (unlikely(buff)) {
3939 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3940 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3942 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3944 /* Timer for repeating the SYN until an answer. */
3945 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3946 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3949 EXPORT_SYMBOL(tcp_connect);
3951 /* Send out a delayed ack, the caller does the policy checking
3952 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3955 void tcp_send_delayed_ack(struct sock *sk)
3957 struct inet_connection_sock *icsk = inet_csk(sk);
3958 int ato = icsk->icsk_ack.ato;
3959 unsigned long timeout;
3961 if (ato > TCP_DELACK_MIN) {
3962 const struct tcp_sock *tp = tcp_sk(sk);
3963 int max_ato = HZ / 2;
3965 if (inet_csk_in_pingpong_mode(sk) ||
3966 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3967 max_ato = TCP_DELACK_MAX;
3969 /* Slow path, intersegment interval is "high". */
3971 /* If some rtt estimate is known, use it to bound delayed ack.
3972 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3976 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3983 ato = min(ato, max_ato);
3986 ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max);
3988 /* Stay within the limit we were given */
3989 timeout = jiffies + ato;
3991 /* Use new timeout only if there wasn't a older one earlier. */
3992 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3993 /* If delack timer is about to expire, send ACK now. */
3994 if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3999 if (!time_before(timeout, icsk->icsk_ack.timeout))
4000 timeout = icsk->icsk_ack.timeout;
4002 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
4003 icsk->icsk_ack.timeout = timeout;
4004 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
4007 /* This routine sends an ack and also updates the window. */
4008 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
4010 struct sk_buff *buff;
4012 /* If we have been reset, we may not send again. */
4013 if (sk->sk_state == TCP_CLOSE)
4016 /* We are not putting this on the write queue, so
4017 * tcp_transmit_skb() will set the ownership to this
4020 buff = alloc_skb(MAX_TCP_HEADER,
4021 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4022 if (unlikely(!buff)) {
4023 struct inet_connection_sock *icsk = inet_csk(sk);
4024 unsigned long delay;
4026 delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
4027 if (delay < TCP_RTO_MAX)
4028 icsk->icsk_ack.retry++;
4029 inet_csk_schedule_ack(sk);
4030 icsk->icsk_ack.ato = TCP_ATO_MIN;
4031 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
4035 /* Reserve space for headers and prepare control bits. */
4036 skb_reserve(buff, MAX_TCP_HEADER);
4037 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
4039 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
4041 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
4043 skb_set_tcp_pure_ack(buff);
4045 /* Send it off, this clears delayed acks for us. */
4046 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
4048 EXPORT_SYMBOL_GPL(__tcp_send_ack);
4050 void tcp_send_ack(struct sock *sk)
4052 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
4055 /* This routine sends a packet with an out of date sequence
4056 * number. It assumes the other end will try to ack it.
4058 * Question: what should we make while urgent mode?
4059 * 4.4BSD forces sending single byte of data. We cannot send
4060 * out of window data, because we have SND.NXT==SND.MAX...
4062 * Current solution: to send TWO zero-length segments in urgent mode:
4063 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
4064 * out-of-date with SND.UNA-1 to probe window.
4066 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
4068 struct tcp_sock *tp = tcp_sk(sk);
4069 struct sk_buff *skb;
4071 /* We don't queue it, tcp_transmit_skb() sets ownership. */
4072 skb = alloc_skb(MAX_TCP_HEADER,
4073 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4077 /* Reserve space for headers and set control bits. */
4078 skb_reserve(skb, MAX_TCP_HEADER);
4079 /* Use a previous sequence. This should cause the other
4080 * end to send an ack. Don't queue or clone SKB, just
4083 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
4084 NET_INC_STATS(sock_net(sk), mib);
4085 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4088 /* Called from setsockopt( ... TCP_REPAIR ) */
4089 void tcp_send_window_probe(struct sock *sk)
4091 if (sk->sk_state == TCP_ESTABLISHED) {
4092 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4093 tcp_mstamp_refresh(tcp_sk(sk));
4094 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4098 /* Initiate keepalive or window probe from timer. */
4099 int tcp_write_wakeup(struct sock *sk, int mib)
4101 struct tcp_sock *tp = tcp_sk(sk);
4102 struct sk_buff *skb;
4104 if (sk->sk_state == TCP_CLOSE)
4107 skb = tcp_send_head(sk);
4108 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4110 unsigned int mss = tcp_current_mss(sk);
4111 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4113 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4114 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4116 /* We are probing the opening of a window
4117 * but the window size is != 0
4118 * must have been a result SWS avoidance ( sender )
4120 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4122 seg_size = min(seg_size, mss);
4123 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4124 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4125 skb, seg_size, mss, GFP_ATOMIC))
4127 } else if (!tcp_skb_pcount(skb))
4128 tcp_set_skb_tso_segs(skb, mss);
4130 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4131 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4133 tcp_event_new_data_sent(sk, skb);
4136 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4137 tcp_xmit_probe_skb(sk, 1, mib);
4138 return tcp_xmit_probe_skb(sk, 0, mib);
4142 /* A window probe timeout has occurred. If window is not closed send
4143 * a partial packet else a zero probe.
4145 void tcp_send_probe0(struct sock *sk)
4147 struct inet_connection_sock *icsk = inet_csk(sk);
4148 struct tcp_sock *tp = tcp_sk(sk);
4149 struct net *net = sock_net(sk);
4150 unsigned long timeout;
4153 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4155 if (tp->packets_out || tcp_write_queue_empty(sk)) {
4156 /* Cancel probe timer, if it is not required. */
4157 icsk->icsk_probes_out = 0;
4158 icsk->icsk_backoff = 0;
4159 icsk->icsk_probes_tstamp = 0;
4163 icsk->icsk_probes_out++;
4165 if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
4166 icsk->icsk_backoff++;
4167 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
4169 /* If packet was not sent due to local congestion,
4170 * Let senders fight for local resources conservatively.
4172 timeout = TCP_RESOURCE_PROBE_INTERVAL;
4175 timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout);
4176 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
4179 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4181 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4185 /* Paired with WRITE_ONCE() in sock_setsockopt() */
4186 if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED)
4187 WRITE_ONCE(tcp_rsk(req)->txhash, net_tx_rndhash());
4188 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4191 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4192 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4193 if (unlikely(tcp_passive_fastopen(sk)))
4194 tcp_sk(sk)->total_retrans++;
4195 trace_tcp_retransmit_synack(sk, req);
4199 EXPORT_SYMBOL(tcp_rtx_synack);