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, unsigned int pkts,
183 struct tcp_sock *tp = tcp_sk(sk);
185 if (unlikely(tp->compressed_ack)) {
186 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
188 tp->compressed_ack = 0;
189 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
193 if (unlikely(rcv_nxt != tp->rcv_nxt))
194 return; /* Special ACK sent by DCTCP to reflect ECN */
195 tcp_dec_quickack_mode(sk, pkts);
196 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
199 /* Determine a window scaling and initial window to offer.
200 * Based on the assumption that the given amount of space
201 * will be offered. Store the results in the tp structure.
202 * NOTE: for smooth operation initial space offering should
203 * be a multiple of mss if possible. We assume here that mss >= 1.
204 * This MUST be enforced by all callers.
206 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
207 __u32 *rcv_wnd, __u32 *window_clamp,
208 int wscale_ok, __u8 *rcv_wscale,
211 unsigned int space = (__space < 0 ? 0 : __space);
213 /* If no clamp set the clamp to the max possible scaled window */
214 if (*window_clamp == 0)
215 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
216 space = min(*window_clamp, space);
218 /* Quantize space offering to a multiple of mss if possible. */
220 space = rounddown(space, mss);
222 /* NOTE: offering an initial window larger than 32767
223 * will break some buggy TCP stacks. If the admin tells us
224 * it is likely we could be speaking with such a buggy stack
225 * we will truncate our initial window offering to 32K-1
226 * unless the remote has sent us a window scaling option,
227 * which we interpret as a sign the remote TCP is not
228 * misinterpreting the window field as a signed quantity.
230 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
231 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
233 (*rcv_wnd) = min_t(u32, space, U16_MAX);
236 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
240 /* Set window scaling on max possible window */
241 space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
242 space = max_t(u32, space, READ_ONCE(sysctl_rmem_max));
243 space = min_t(u32, space, *window_clamp);
244 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
247 /* Set the clamp no higher than max representable value */
248 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
250 EXPORT_SYMBOL(tcp_select_initial_window);
252 /* Chose a new window to advertise, update state in tcp_sock for the
253 * socket, and return result with RFC1323 scaling applied. The return
254 * value can be stuffed directly into th->window for an outgoing
257 static u16 tcp_select_window(struct sock *sk)
259 struct tcp_sock *tp = tcp_sk(sk);
260 u32 old_win = tp->rcv_wnd;
261 u32 cur_win = tcp_receive_window(tp);
262 u32 new_win = __tcp_select_window(sk);
264 /* Never shrink the offered window */
265 if (new_win < cur_win) {
266 /* Danger Will Robinson!
267 * Don't update rcv_wup/rcv_wnd here or else
268 * we will not be able to advertise a zero
269 * window in time. --DaveM
271 * Relax Will Robinson.
274 NET_INC_STATS(sock_net(sk),
275 LINUX_MIB_TCPWANTZEROWINDOWADV);
276 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
278 tp->rcv_wnd = new_win;
279 tp->rcv_wup = tp->rcv_nxt;
281 /* Make sure we do not exceed the maximum possible
284 if (!tp->rx_opt.rcv_wscale &&
285 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
286 new_win = min(new_win, MAX_TCP_WINDOW);
288 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
290 /* RFC1323 scaling applied */
291 new_win >>= tp->rx_opt.rcv_wscale;
293 /* If we advertise zero window, disable fast path. */
297 NET_INC_STATS(sock_net(sk),
298 LINUX_MIB_TCPTOZEROWINDOWADV);
299 } else if (old_win == 0) {
300 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
306 /* Packet ECN state for a SYN-ACK */
307 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
309 const struct tcp_sock *tp = tcp_sk(sk);
311 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
312 if (!(tp->ecn_flags & TCP_ECN_OK))
313 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
314 else if (tcp_ca_needs_ecn(sk) ||
315 tcp_bpf_ca_needs_ecn(sk))
319 /* Packet ECN state for a SYN. */
320 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
322 struct tcp_sock *tp = tcp_sk(sk);
323 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
324 bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 ||
325 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
328 const struct dst_entry *dst = __sk_dst_get(sk);
330 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
337 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
338 tp->ecn_flags = TCP_ECN_OK;
339 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
344 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
346 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback))
347 /* tp->ecn_flags are cleared at a later point in time when
348 * SYN ACK is ultimatively being received.
350 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
354 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
356 if (inet_rsk(req)->ecn_ok)
360 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
363 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
364 struct tcphdr *th, int tcp_header_len)
366 struct tcp_sock *tp = tcp_sk(sk);
368 if (tp->ecn_flags & TCP_ECN_OK) {
369 /* Not-retransmitted data segment: set ECT and inject CWR. */
370 if (skb->len != tcp_header_len &&
371 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
373 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
374 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
376 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
378 } else if (!tcp_ca_needs_ecn(sk)) {
379 /* ACK or retransmitted segment: clear ECT|CE */
380 INET_ECN_dontxmit(sk);
382 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
387 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
388 * auto increment end seqno.
390 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
392 skb->ip_summed = CHECKSUM_PARTIAL;
394 TCP_SKB_CB(skb)->tcp_flags = flags;
396 tcp_skb_pcount_set(skb, 1);
398 TCP_SKB_CB(skb)->seq = seq;
399 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
401 TCP_SKB_CB(skb)->end_seq = seq;
404 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
406 return tp->snd_una != tp->snd_up;
409 #define OPTION_SACK_ADVERTISE BIT(0)
410 #define OPTION_TS BIT(1)
411 #define OPTION_MD5 BIT(2)
412 #define OPTION_WSCALE BIT(3)
413 #define OPTION_FAST_OPEN_COOKIE BIT(8)
414 #define OPTION_SMC BIT(9)
415 #define OPTION_MPTCP BIT(10)
417 static void smc_options_write(__be32 *ptr, u16 *options)
419 #if IS_ENABLED(CONFIG_SMC)
420 if (static_branch_unlikely(&tcp_have_smc)) {
421 if (unlikely(OPTION_SMC & *options)) {
422 *ptr++ = htonl((TCPOPT_NOP << 24) |
425 (TCPOLEN_EXP_SMC_BASE));
426 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
432 struct tcp_out_options {
433 u16 options; /* bit field of OPTION_* */
434 u16 mss; /* 0 to disable */
435 u8 ws; /* window scale, 0 to disable */
436 u8 num_sack_blocks; /* number of SACK blocks to include */
437 u8 hash_size; /* bytes in hash_location */
438 u8 bpf_opt_len; /* length of BPF hdr option */
439 __u8 *hash_location; /* temporary pointer, overloaded */
440 __u32 tsval, tsecr; /* need to include OPTION_TS */
441 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
442 struct mptcp_out_options mptcp;
445 static void mptcp_options_write(struct tcphdr *th, __be32 *ptr,
447 struct tcp_out_options *opts)
449 #if IS_ENABLED(CONFIG_MPTCP)
450 if (unlikely(OPTION_MPTCP & opts->options))
451 mptcp_write_options(th, ptr, tp, &opts->mptcp);
455 #ifdef CONFIG_CGROUP_BPF
456 static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
457 enum tcp_synack_type synack_type)
460 return BPF_WRITE_HDR_TCP_CURRENT_MSS;
462 if (unlikely(synack_type == TCP_SYNACK_COOKIE))
463 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
468 /* req, syn_skb and synack_type are used when writing synack */
469 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
470 struct request_sock *req,
471 struct sk_buff *syn_skb,
472 enum tcp_synack_type synack_type,
473 struct tcp_out_options *opts,
474 unsigned int *remaining)
476 struct bpf_sock_ops_kern sock_ops;
479 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
480 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
484 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
487 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
489 sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
492 /* The listen "sk" cannot be passed here because
493 * it is not locked. It would not make too much
494 * sense to do bpf_setsockopt(listen_sk) based
495 * on individual connection request also.
497 * Thus, "req" is passed here and the cgroup-bpf-progs
498 * of the listen "sk" will be run.
500 * "req" is also used here for fastopen even the "sk" here is
501 * a fullsock "child" sk. It is to keep the behavior
502 * consistent between fastopen and non-fastopen on
503 * the bpf programming side.
505 sock_ops.sk = (struct sock *)req;
506 sock_ops.syn_skb = syn_skb;
508 sock_owned_by_me(sk);
510 sock_ops.is_fullsock = 1;
514 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
515 sock_ops.remaining_opt_len = *remaining;
516 /* tcp_current_mss() does not pass a skb */
518 bpf_skops_init_skb(&sock_ops, skb, 0);
520 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
522 if (err || sock_ops.remaining_opt_len == *remaining)
525 opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
526 /* round up to 4 bytes */
527 opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
529 *remaining -= opts->bpf_opt_len;
532 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
533 struct request_sock *req,
534 struct sk_buff *syn_skb,
535 enum tcp_synack_type synack_type,
536 struct tcp_out_options *opts)
538 u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
539 struct bpf_sock_ops_kern sock_ops;
542 if (likely(!max_opt_len))
545 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
547 sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
550 sock_ops.sk = (struct sock *)req;
551 sock_ops.syn_skb = syn_skb;
553 sock_owned_by_me(sk);
555 sock_ops.is_fullsock = 1;
559 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
560 sock_ops.remaining_opt_len = max_opt_len;
561 first_opt_off = tcp_hdrlen(skb) - max_opt_len;
562 bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
564 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
569 nr_written = max_opt_len - sock_ops.remaining_opt_len;
571 if (nr_written < max_opt_len)
572 memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
573 max_opt_len - nr_written);
576 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
577 struct request_sock *req,
578 struct sk_buff *syn_skb,
579 enum tcp_synack_type synack_type,
580 struct tcp_out_options *opts,
581 unsigned int *remaining)
585 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
586 struct request_sock *req,
587 struct sk_buff *syn_skb,
588 enum tcp_synack_type synack_type,
589 struct tcp_out_options *opts)
594 /* Write previously computed TCP options to the packet.
596 * Beware: Something in the Internet is very sensitive to the ordering of
597 * TCP options, we learned this through the hard way, so be careful here.
598 * Luckily we can at least blame others for their non-compliance but from
599 * inter-operability perspective it seems that we're somewhat stuck with
600 * the ordering which we have been using if we want to keep working with
601 * those broken things (not that it currently hurts anybody as there isn't
602 * particular reason why the ordering would need to be changed).
604 * At least SACK_PERM as the first option is known to lead to a disaster
605 * (but it may well be that other scenarios fail similarly).
607 static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp,
608 struct tcp_out_options *opts)
610 __be32 *ptr = (__be32 *)(th + 1);
611 u16 options = opts->options; /* mungable copy */
613 if (unlikely(OPTION_MD5 & options)) {
614 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
615 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
616 /* overload cookie hash location */
617 opts->hash_location = (__u8 *)ptr;
621 if (unlikely(opts->mss)) {
622 *ptr++ = htonl((TCPOPT_MSS << 24) |
623 (TCPOLEN_MSS << 16) |
627 if (likely(OPTION_TS & options)) {
628 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
629 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
630 (TCPOLEN_SACK_PERM << 16) |
631 (TCPOPT_TIMESTAMP << 8) |
633 options &= ~OPTION_SACK_ADVERTISE;
635 *ptr++ = htonl((TCPOPT_NOP << 24) |
637 (TCPOPT_TIMESTAMP << 8) |
640 *ptr++ = htonl(opts->tsval);
641 *ptr++ = htonl(opts->tsecr);
644 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
645 *ptr++ = htonl((TCPOPT_NOP << 24) |
647 (TCPOPT_SACK_PERM << 8) |
651 if (unlikely(OPTION_WSCALE & options)) {
652 *ptr++ = htonl((TCPOPT_NOP << 24) |
653 (TCPOPT_WINDOW << 16) |
654 (TCPOLEN_WINDOW << 8) |
658 if (unlikely(opts->num_sack_blocks)) {
659 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
660 tp->duplicate_sack : tp->selective_acks;
663 *ptr++ = htonl((TCPOPT_NOP << 24) |
666 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
667 TCPOLEN_SACK_PERBLOCK)));
669 for (this_sack = 0; this_sack < opts->num_sack_blocks;
671 *ptr++ = htonl(sp[this_sack].start_seq);
672 *ptr++ = htonl(sp[this_sack].end_seq);
675 tp->rx_opt.dsack = 0;
678 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
679 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
681 u32 len; /* Fast Open option length */
684 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
685 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
686 TCPOPT_FASTOPEN_MAGIC);
687 p += TCPOLEN_EXP_FASTOPEN_BASE;
689 len = TCPOLEN_FASTOPEN_BASE + foc->len;
690 *p++ = TCPOPT_FASTOPEN;
694 memcpy(p, foc->val, foc->len);
695 if ((len & 3) == 2) {
696 p[foc->len] = TCPOPT_NOP;
697 p[foc->len + 1] = TCPOPT_NOP;
699 ptr += (len + 3) >> 2;
702 smc_options_write(ptr, &options);
704 mptcp_options_write(th, ptr, tp, opts);
707 static void smc_set_option(const struct tcp_sock *tp,
708 struct tcp_out_options *opts,
709 unsigned int *remaining)
711 #if IS_ENABLED(CONFIG_SMC)
712 if (static_branch_unlikely(&tcp_have_smc)) {
714 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
715 opts->options |= OPTION_SMC;
716 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
723 static void smc_set_option_cond(const struct tcp_sock *tp,
724 const struct inet_request_sock *ireq,
725 struct tcp_out_options *opts,
726 unsigned int *remaining)
728 #if IS_ENABLED(CONFIG_SMC)
729 if (static_branch_unlikely(&tcp_have_smc)) {
730 if (tp->syn_smc && ireq->smc_ok) {
731 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
732 opts->options |= OPTION_SMC;
733 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
740 static void mptcp_set_option_cond(const struct request_sock *req,
741 struct tcp_out_options *opts,
742 unsigned int *remaining)
744 if (rsk_is_mptcp(req)) {
747 if (mptcp_synack_options(req, &size, &opts->mptcp)) {
748 if (*remaining >= size) {
749 opts->options |= OPTION_MPTCP;
756 /* Compute TCP options for SYN packets. This is not the final
757 * network wire format yet.
759 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
760 struct tcp_out_options *opts,
761 struct tcp_md5sig_key **md5)
763 struct tcp_sock *tp = tcp_sk(sk);
764 unsigned int remaining = MAX_TCP_OPTION_SPACE;
765 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
768 #ifdef CONFIG_TCP_MD5SIG
769 if (static_branch_unlikely(&tcp_md5_needed) &&
770 rcu_access_pointer(tp->md5sig_info)) {
771 *md5 = tp->af_specific->md5_lookup(sk, sk);
773 opts->options |= OPTION_MD5;
774 remaining -= TCPOLEN_MD5SIG_ALIGNED;
779 /* We always get an MSS option. The option bytes which will be seen in
780 * normal data packets should timestamps be used, must be in the MSS
781 * advertised. But we subtract them from tp->mss_cache so that
782 * calculations in tcp_sendmsg are simpler etc. So account for this
783 * fact here if necessary. If we don't do this correctly, as a
784 * receiver we won't recognize data packets as being full sized when we
785 * should, and thus we won't abide by the delayed ACK rules correctly.
786 * SACKs don't matter, we never delay an ACK when we have any of those
788 opts->mss = tcp_advertise_mss(sk);
789 remaining -= TCPOLEN_MSS_ALIGNED;
791 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps) && !*md5)) {
792 opts->options |= OPTION_TS;
793 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
794 opts->tsecr = tp->rx_opt.ts_recent;
795 remaining -= TCPOLEN_TSTAMP_ALIGNED;
797 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) {
798 opts->ws = tp->rx_opt.rcv_wscale;
799 opts->options |= OPTION_WSCALE;
800 remaining -= TCPOLEN_WSCALE_ALIGNED;
802 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) {
803 opts->options |= OPTION_SACK_ADVERTISE;
804 if (unlikely(!(OPTION_TS & opts->options)))
805 remaining -= TCPOLEN_SACKPERM_ALIGNED;
808 if (fastopen && fastopen->cookie.len >= 0) {
809 u32 need = fastopen->cookie.len;
811 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
812 TCPOLEN_FASTOPEN_BASE;
813 need = (need + 3) & ~3U; /* Align to 32 bits */
814 if (remaining >= need) {
815 opts->options |= OPTION_FAST_OPEN_COOKIE;
816 opts->fastopen_cookie = &fastopen->cookie;
818 tp->syn_fastopen = 1;
819 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
823 smc_set_option(tp, opts, &remaining);
825 if (sk_is_mptcp(sk)) {
828 if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
829 opts->options |= OPTION_MPTCP;
834 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
836 return MAX_TCP_OPTION_SPACE - remaining;
839 /* Set up TCP options for SYN-ACKs. */
840 static unsigned int tcp_synack_options(const struct sock *sk,
841 struct request_sock *req,
842 unsigned int mss, struct sk_buff *skb,
843 struct tcp_out_options *opts,
844 const struct tcp_md5sig_key *md5,
845 struct tcp_fastopen_cookie *foc,
846 enum tcp_synack_type synack_type,
847 struct sk_buff *syn_skb)
849 struct inet_request_sock *ireq = inet_rsk(req);
850 unsigned int remaining = MAX_TCP_OPTION_SPACE;
852 #ifdef CONFIG_TCP_MD5SIG
854 opts->options |= OPTION_MD5;
855 remaining -= TCPOLEN_MD5SIG_ALIGNED;
857 /* We can't fit any SACK blocks in a packet with MD5 + TS
858 * options. There was discussion about disabling SACK
859 * rather than TS in order to fit in better with old,
860 * buggy kernels, but that was deemed to be unnecessary.
862 if (synack_type != TCP_SYNACK_COOKIE)
863 ireq->tstamp_ok &= !ireq->sack_ok;
867 /* We always send an MSS option. */
869 remaining -= TCPOLEN_MSS_ALIGNED;
871 if (likely(ireq->wscale_ok)) {
872 opts->ws = ireq->rcv_wscale;
873 opts->options |= OPTION_WSCALE;
874 remaining -= TCPOLEN_WSCALE_ALIGNED;
876 if (likely(ireq->tstamp_ok)) {
877 opts->options |= OPTION_TS;
878 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
879 opts->tsecr = req->ts_recent;
880 remaining -= TCPOLEN_TSTAMP_ALIGNED;
882 if (likely(ireq->sack_ok)) {
883 opts->options |= OPTION_SACK_ADVERTISE;
884 if (unlikely(!ireq->tstamp_ok))
885 remaining -= TCPOLEN_SACKPERM_ALIGNED;
887 if (foc != NULL && foc->len >= 0) {
890 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
891 TCPOLEN_FASTOPEN_BASE;
892 need = (need + 3) & ~3U; /* Align to 32 bits */
893 if (remaining >= need) {
894 opts->options |= OPTION_FAST_OPEN_COOKIE;
895 opts->fastopen_cookie = foc;
900 mptcp_set_option_cond(req, opts, &remaining);
902 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
904 bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb,
905 synack_type, opts, &remaining);
907 return MAX_TCP_OPTION_SPACE - remaining;
910 /* Compute TCP options for ESTABLISHED sockets. This is not the
911 * final wire format yet.
913 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
914 struct tcp_out_options *opts,
915 struct tcp_md5sig_key **md5)
917 struct tcp_sock *tp = tcp_sk(sk);
918 unsigned int size = 0;
919 unsigned int eff_sacks;
924 #ifdef CONFIG_TCP_MD5SIG
925 if (static_branch_unlikely(&tcp_md5_needed) &&
926 rcu_access_pointer(tp->md5sig_info)) {
927 *md5 = tp->af_specific->md5_lookup(sk, sk);
929 opts->options |= OPTION_MD5;
930 size += TCPOLEN_MD5SIG_ALIGNED;
935 if (likely(tp->rx_opt.tstamp_ok)) {
936 opts->options |= OPTION_TS;
937 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
938 opts->tsecr = tp->rx_opt.ts_recent;
939 size += TCPOLEN_TSTAMP_ALIGNED;
942 /* MPTCP options have precedence over SACK for the limited TCP
943 * option space because a MPTCP connection would be forced to
944 * fall back to regular TCP if a required multipath option is
945 * missing. SACK still gets a chance to use whatever space is
948 if (sk_is_mptcp(sk)) {
949 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
950 unsigned int opt_size = 0;
952 if (mptcp_established_options(sk, skb, &opt_size, remaining,
954 opts->options |= OPTION_MPTCP;
959 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
960 if (unlikely(eff_sacks)) {
961 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
962 if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
963 TCPOLEN_SACK_PERBLOCK))
966 opts->num_sack_blocks =
967 min_t(unsigned int, eff_sacks,
968 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
969 TCPOLEN_SACK_PERBLOCK);
971 size += TCPOLEN_SACK_BASE_ALIGNED +
972 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
975 if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
976 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
977 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
979 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
981 size = MAX_TCP_OPTION_SPACE - remaining;
988 /* TCP SMALL QUEUES (TSQ)
990 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
991 * to reduce RTT and bufferbloat.
992 * We do this using a special skb destructor (tcp_wfree).
994 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
995 * needs to be reallocated in a driver.
996 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
998 * Since transmit from skb destructor is forbidden, we use a tasklet
999 * to process all sockets that eventually need to send more skbs.
1000 * We use one tasklet per cpu, with its own queue of sockets.
1002 struct tsq_tasklet {
1003 struct tasklet_struct tasklet;
1004 struct list_head head; /* queue of tcp sockets */
1006 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
1008 static void tcp_tsq_write(struct sock *sk)
1010 if ((1 << sk->sk_state) &
1011 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1012 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
1013 struct tcp_sock *tp = tcp_sk(sk);
1015 if (tp->lost_out > tp->retrans_out &&
1016 tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) {
1017 tcp_mstamp_refresh(tp);
1018 tcp_xmit_retransmit_queue(sk);
1021 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
1026 static void tcp_tsq_handler(struct sock *sk)
1029 if (!sock_owned_by_user(sk))
1031 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
1036 * One tasklet per cpu tries to send more skbs.
1037 * We run in tasklet context but need to disable irqs when
1038 * transferring tsq->head because tcp_wfree() might
1039 * interrupt us (non NAPI drivers)
1041 static void tcp_tasklet_func(struct tasklet_struct *t)
1043 struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet);
1045 unsigned long flags;
1046 struct list_head *q, *n;
1047 struct tcp_sock *tp;
1050 local_irq_save(flags);
1051 list_splice_init(&tsq->head, &list);
1052 local_irq_restore(flags);
1054 list_for_each_safe(q, n, &list) {
1055 tp = list_entry(q, struct tcp_sock, tsq_node);
1056 list_del(&tp->tsq_node);
1058 sk = (struct sock *)tp;
1059 smp_mb__before_atomic();
1060 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
1062 tcp_tsq_handler(sk);
1067 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
1068 TCPF_WRITE_TIMER_DEFERRED | \
1069 TCPF_DELACK_TIMER_DEFERRED | \
1070 TCPF_MTU_REDUCED_DEFERRED)
1072 * tcp_release_cb - tcp release_sock() callback
1075 * called from release_sock() to perform protocol dependent
1076 * actions before socket release.
1078 void tcp_release_cb(struct sock *sk)
1080 unsigned long flags, nflags;
1082 /* perform an atomic operation only if at least one flag is set */
1084 flags = sk->sk_tsq_flags;
1085 if (!(flags & TCP_DEFERRED_ALL))
1087 nflags = flags & ~TCP_DEFERRED_ALL;
1088 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
1090 if (flags & TCPF_TSQ_DEFERRED) {
1094 /* Here begins the tricky part :
1095 * We are called from release_sock() with :
1097 * 2) sk_lock.slock spinlock held
1098 * 3) socket owned by us (sk->sk_lock.owned == 1)
1100 * But following code is meant to be called from BH handlers,
1101 * so we should keep BH disabled, but early release socket ownership
1103 sock_release_ownership(sk);
1105 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1106 tcp_write_timer_handler(sk);
1109 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1110 tcp_delack_timer_handler(sk);
1113 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1114 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1118 EXPORT_SYMBOL(tcp_release_cb);
1120 void __init tcp_tasklet_init(void)
1124 for_each_possible_cpu(i) {
1125 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
1127 INIT_LIST_HEAD(&tsq->head);
1128 tasklet_setup(&tsq->tasklet, tcp_tasklet_func);
1133 * Write buffer destructor automatically called from kfree_skb.
1134 * We can't xmit new skbs from this context, as we might already
1137 void tcp_wfree(struct sk_buff *skb)
1139 struct sock *sk = skb->sk;
1140 struct tcp_sock *tp = tcp_sk(sk);
1141 unsigned long flags, nval, oval;
1143 /* Keep one reference on sk_wmem_alloc.
1144 * Will be released by sk_free() from here or tcp_tasklet_func()
1146 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1148 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
1149 * Wait until our queues (qdisc + devices) are drained.
1151 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1152 * - chance for incoming ACK (processed by another cpu maybe)
1153 * to migrate this flow (skb->ooo_okay will be eventually set)
1155 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1158 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
1159 struct tsq_tasklet *tsq;
1162 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1165 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1166 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
1170 /* queue this socket to tasklet queue */
1171 local_irq_save(flags);
1172 tsq = this_cpu_ptr(&tsq_tasklet);
1173 empty = list_empty(&tsq->head);
1174 list_add(&tp->tsq_node, &tsq->head);
1176 tasklet_schedule(&tsq->tasklet);
1177 local_irq_restore(flags);
1184 /* Note: Called under soft irq.
1185 * We can call TCP stack right away, unless socket is owned by user.
1187 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1189 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1190 struct sock *sk = (struct sock *)tp;
1192 tcp_tsq_handler(sk);
1195 return HRTIMER_NORESTART;
1198 static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1201 struct tcp_sock *tp = tcp_sk(sk);
1203 if (sk->sk_pacing_status != SK_PACING_NONE) {
1204 unsigned long rate = sk->sk_pacing_rate;
1206 /* Original sch_fq does not pace first 10 MSS
1207 * Note that tp->data_segs_out overflows after 2^32 packets,
1208 * this is a minor annoyance.
1210 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1211 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1212 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1214 /* take into account OS jitter */
1215 len_ns -= min_t(u64, len_ns / 2, credit);
1216 tp->tcp_wstamp_ns += len_ns;
1219 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1222 INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1223 INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1224 INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1226 /* This routine actually transmits TCP packets queued in by
1227 * tcp_do_sendmsg(). This is used by both the initial
1228 * transmission and possible later retransmissions.
1229 * All SKB's seen here are completely headerless. It is our
1230 * job to build the TCP header, and pass the packet down to
1231 * IP so it can do the same plus pass the packet off to the
1234 * We are working here with either a clone of the original
1235 * SKB, or a fresh unique copy made by the retransmit engine.
1237 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1238 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1240 const struct inet_connection_sock *icsk = inet_csk(sk);
1241 struct inet_sock *inet;
1242 struct tcp_sock *tp;
1243 struct tcp_skb_cb *tcb;
1244 struct tcp_out_options opts;
1245 unsigned int tcp_options_size, tcp_header_size;
1246 struct sk_buff *oskb = NULL;
1247 struct tcp_md5sig_key *md5;
1252 BUG_ON(!skb || !tcp_skb_pcount(skb));
1254 prior_wstamp = tp->tcp_wstamp_ns;
1255 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1256 skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
1260 tcp_skb_tsorted_save(oskb) {
1261 if (unlikely(skb_cloned(oskb)))
1262 skb = pskb_copy(oskb, gfp_mask);
1264 skb = skb_clone(oskb, gfp_mask);
1265 } tcp_skb_tsorted_restore(oskb);
1269 /* retransmit skbs might have a non zero value in skb->dev
1270 * because skb->dev is aliased with skb->rbnode.rb_left
1276 tcb = TCP_SKB_CB(skb);
1277 memset(&opts, 0, sizeof(opts));
1279 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1280 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1282 tcp_options_size = tcp_established_options(sk, skb, &opts,
1284 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1285 * at receiver : This slightly improve GRO performance.
1286 * Note that we do not force the PSH flag for non GSO packets,
1287 * because they might be sent under high congestion events,
1288 * and in this case it is better to delay the delivery of 1-MSS
1289 * packets and thus the corresponding ACK packet that would
1290 * release the following packet.
1292 if (tcp_skb_pcount(skb) > 1)
1293 tcb->tcp_flags |= TCPHDR_PSH;
1295 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1297 /* if no packet is in qdisc/device queue, then allow XPS to select
1298 * another queue. We can be called from tcp_tsq_handler()
1299 * which holds one reference to sk.
1301 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1302 * One way to get this would be to set skb->truesize = 2 on them.
1304 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1306 /* If we had to use memory reserve to allocate this skb,
1307 * this might cause drops if packet is looped back :
1308 * Other socket might not have SOCK_MEMALLOC.
1309 * Packets not looped back do not care about pfmemalloc.
1311 skb->pfmemalloc = 0;
1313 skb_push(skb, tcp_header_size);
1314 skb_reset_transport_header(skb);
1318 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1319 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1321 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1323 /* Build TCP header and checksum it. */
1324 th = (struct tcphdr *)skb->data;
1325 th->source = inet->inet_sport;
1326 th->dest = inet->inet_dport;
1327 th->seq = htonl(tcb->seq);
1328 th->ack_seq = htonl(rcv_nxt);
1329 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1335 /* The urg_mode check is necessary during a below snd_una win probe */
1336 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1337 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1338 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1340 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1341 th->urg_ptr = htons(0xFFFF);
1346 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1347 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1348 th->window = htons(tcp_select_window(sk));
1349 tcp_ecn_send(sk, skb, th, tcp_header_size);
1351 /* RFC1323: The window in SYN & SYN/ACK segments
1354 th->window = htons(min(tp->rcv_wnd, 65535U));
1357 tcp_options_write(th, tp, &opts);
1359 #ifdef CONFIG_TCP_MD5SIG
1360 /* Calculate the MD5 hash, as we have all we need now */
1363 tp->af_specific->calc_md5_hash(opts.hash_location,
1368 /* BPF prog is the last one writing header option */
1369 bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1371 INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1372 tcp_v6_send_check, tcp_v4_send_check,
1375 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1376 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1378 if (skb->len != tcp_header_size) {
1379 tcp_event_data_sent(tp, sk);
1380 tp->data_segs_out += tcp_skb_pcount(skb);
1381 tp->bytes_sent += skb->len - tcp_header_size;
1384 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1385 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1386 tcp_skb_pcount(skb));
1388 tp->segs_out += tcp_skb_pcount(skb);
1389 skb_set_hash_from_sk(skb, sk);
1390 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1391 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1392 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1394 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1396 /* Cleanup our debris for IP stacks */
1397 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1398 sizeof(struct inet6_skb_parm)));
1400 tcp_add_tx_delay(skb, tp);
1402 err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1403 inet6_csk_xmit, ip_queue_xmit,
1404 sk, skb, &inet->cork.fl);
1406 if (unlikely(err > 0)) {
1408 err = net_xmit_eval(err);
1411 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1412 tcp_rate_skb_sent(sk, oskb);
1417 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1420 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1421 tcp_sk(sk)->rcv_nxt);
1424 /* This routine just queues the buffer for sending.
1426 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1427 * otherwise socket can stall.
1429 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1431 struct tcp_sock *tp = tcp_sk(sk);
1433 /* Advance write_seq and place onto the write_queue. */
1434 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1435 __skb_header_release(skb);
1436 tcp_add_write_queue_tail(sk, skb);
1437 sk_wmem_queued_add(sk, skb->truesize);
1438 sk_mem_charge(sk, skb->truesize);
1441 /* Initialize TSO segments for a packet. */
1442 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1444 if (skb->len <= mss_now) {
1445 /* Avoid the costly divide in the normal
1448 tcp_skb_pcount_set(skb, 1);
1449 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1451 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1452 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1456 /* Pcount in the middle of the write queue got changed, we need to do various
1457 * tweaks to fix counters
1459 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1461 struct tcp_sock *tp = tcp_sk(sk);
1463 tp->packets_out -= decr;
1465 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1466 tp->sacked_out -= decr;
1467 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1468 tp->retrans_out -= decr;
1469 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1470 tp->lost_out -= decr;
1472 /* Reno case is special. Sigh... */
1473 if (tcp_is_reno(tp) && decr > 0)
1474 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1476 if (tp->lost_skb_hint &&
1477 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1478 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1479 tp->lost_cnt_hint -= decr;
1481 tcp_verify_left_out(tp);
1484 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1486 return TCP_SKB_CB(skb)->txstamp_ack ||
1487 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1490 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1492 struct skb_shared_info *shinfo = skb_shinfo(skb);
1494 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1495 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1496 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1497 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1499 shinfo->tx_flags &= ~tsflags;
1500 shinfo2->tx_flags |= tsflags;
1501 swap(shinfo->tskey, shinfo2->tskey);
1502 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1503 TCP_SKB_CB(skb)->txstamp_ack = 0;
1507 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1509 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1510 TCP_SKB_CB(skb)->eor = 0;
1513 /* Insert buff after skb on the write or rtx queue of sk. */
1514 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1515 struct sk_buff *buff,
1517 enum tcp_queue tcp_queue)
1519 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1520 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1522 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1525 /* Function to create two new TCP segments. Shrinks the given segment
1526 * to the specified size and appends a new segment with the rest of the
1527 * packet to the list. This won't be called frequently, I hope.
1528 * Remember, these are still headerless SKBs at this point.
1530 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1531 struct sk_buff *skb, u32 len,
1532 unsigned int mss_now, gfp_t gfp)
1534 struct tcp_sock *tp = tcp_sk(sk);
1535 struct sk_buff *buff;
1536 int nsize, old_factor;
1541 if (WARN_ON(len > skb->len))
1544 nsize = skb_headlen(skb) - len;
1548 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1549 * We need some allowance to not penalize applications setting small
1551 * Also allow first and last skb in retransmit queue to be split.
1553 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE);
1554 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1555 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1556 skb != tcp_rtx_queue_head(sk) &&
1557 skb != tcp_rtx_queue_tail(sk))) {
1558 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1562 if (skb_unclone_keeptruesize(skb, gfp))
1565 /* Get a new skb... force flag on. */
1566 buff = tcp_stream_alloc_skb(sk, nsize, gfp, true);
1568 return -ENOMEM; /* We'll just try again later. */
1569 skb_copy_decrypted(buff, skb);
1570 mptcp_skb_ext_copy(buff, skb);
1572 sk_wmem_queued_add(sk, buff->truesize);
1573 sk_mem_charge(sk, buff->truesize);
1574 nlen = skb->len - len - nsize;
1575 buff->truesize += nlen;
1576 skb->truesize -= nlen;
1578 /* Correct the sequence numbers. */
1579 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1580 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1581 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1583 /* PSH and FIN should only be set in the second packet. */
1584 flags = TCP_SKB_CB(skb)->tcp_flags;
1585 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1586 TCP_SKB_CB(buff)->tcp_flags = flags;
1587 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1588 tcp_skb_fragment_eor(skb, buff);
1590 skb_split(skb, buff, len);
1592 skb_set_delivery_time(buff, skb->tstamp, true);
1593 tcp_fragment_tstamp(skb, buff);
1595 old_factor = tcp_skb_pcount(skb);
1597 /* Fix up tso_factor for both original and new SKB. */
1598 tcp_set_skb_tso_segs(skb, mss_now);
1599 tcp_set_skb_tso_segs(buff, mss_now);
1601 /* Update delivered info for the new segment */
1602 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1604 /* If this packet has been sent out already, we must
1605 * adjust the various packet counters.
1607 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1608 int diff = old_factor - tcp_skb_pcount(skb) -
1609 tcp_skb_pcount(buff);
1612 tcp_adjust_pcount(sk, skb, diff);
1615 /* Link BUFF into the send queue. */
1616 __skb_header_release(buff);
1617 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1618 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1619 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1624 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1625 * data is not copied, but immediately discarded.
1627 static int __pskb_trim_head(struct sk_buff *skb, int len)
1629 struct skb_shared_info *shinfo;
1632 eat = min_t(int, len, skb_headlen(skb));
1634 __skb_pull(skb, eat);
1641 shinfo = skb_shinfo(skb);
1642 for (i = 0; i < shinfo->nr_frags; i++) {
1643 int size = skb_frag_size(&shinfo->frags[i]);
1646 skb_frag_unref(skb, i);
1649 shinfo->frags[k] = shinfo->frags[i];
1651 skb_frag_off_add(&shinfo->frags[k], eat);
1652 skb_frag_size_sub(&shinfo->frags[k], eat);
1658 shinfo->nr_frags = k;
1660 skb->data_len -= len;
1661 skb->len = skb->data_len;
1665 /* Remove acked data from a packet in the transmit queue. */
1666 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1670 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
1673 delta_truesize = __pskb_trim_head(skb, len);
1675 TCP_SKB_CB(skb)->seq += len;
1677 if (delta_truesize) {
1678 skb->truesize -= delta_truesize;
1679 sk_wmem_queued_add(sk, -delta_truesize);
1680 if (!skb_zcopy_pure(skb))
1681 sk_mem_uncharge(sk, delta_truesize);
1684 /* Any change of skb->len requires recalculation of tso factor. */
1685 if (tcp_skb_pcount(skb) > 1)
1686 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1691 /* Calculate MSS not accounting any TCP options. */
1692 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1694 const struct tcp_sock *tp = tcp_sk(sk);
1695 const struct inet_connection_sock *icsk = inet_csk(sk);
1698 /* Calculate base mss without TCP options:
1699 It is MMS_S - sizeof(tcphdr) of rfc1122
1701 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1703 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1704 if (icsk->icsk_af_ops->net_frag_header_len) {
1705 const struct dst_entry *dst = __sk_dst_get(sk);
1707 if (dst && dst_allfrag(dst))
1708 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1711 /* Clamp it (mss_clamp does not include tcp options) */
1712 if (mss_now > tp->rx_opt.mss_clamp)
1713 mss_now = tp->rx_opt.mss_clamp;
1715 /* Now subtract optional transport overhead */
1716 mss_now -= icsk->icsk_ext_hdr_len;
1718 /* Then reserve room for full set of TCP options and 8 bytes of data */
1719 mss_now = max(mss_now,
1720 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss));
1724 /* Calculate MSS. Not accounting for SACKs here. */
1725 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1727 /* Subtract TCP options size, not including SACKs */
1728 return __tcp_mtu_to_mss(sk, pmtu) -
1729 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1731 EXPORT_SYMBOL(tcp_mtu_to_mss);
1733 /* Inverse of above */
1734 int tcp_mss_to_mtu(struct sock *sk, int mss)
1736 const struct tcp_sock *tp = tcp_sk(sk);
1737 const struct inet_connection_sock *icsk = inet_csk(sk);
1741 tp->tcp_header_len +
1742 icsk->icsk_ext_hdr_len +
1743 icsk->icsk_af_ops->net_header_len;
1745 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1746 if (icsk->icsk_af_ops->net_frag_header_len) {
1747 const struct dst_entry *dst = __sk_dst_get(sk);
1749 if (dst && dst_allfrag(dst))
1750 mtu += icsk->icsk_af_ops->net_frag_header_len;
1754 EXPORT_SYMBOL(tcp_mss_to_mtu);
1756 /* MTU probing init per socket */
1757 void tcp_mtup_init(struct sock *sk)
1759 struct tcp_sock *tp = tcp_sk(sk);
1760 struct inet_connection_sock *icsk = inet_csk(sk);
1761 struct net *net = sock_net(sk);
1763 icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1764 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1765 icsk->icsk_af_ops->net_header_len;
1766 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss));
1767 icsk->icsk_mtup.probe_size = 0;
1768 if (icsk->icsk_mtup.enabled)
1769 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1771 EXPORT_SYMBOL(tcp_mtup_init);
1773 /* This function synchronize snd mss to current pmtu/exthdr set.
1775 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1776 for TCP options, but includes only bare TCP header.
1778 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1779 It is minimum of user_mss and mss received with SYN.
1780 It also does not include TCP options.
1782 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1784 tp->mss_cache is current effective sending mss, including
1785 all tcp options except for SACKs. It is evaluated,
1786 taking into account current pmtu, but never exceeds
1787 tp->rx_opt.mss_clamp.
1789 NOTE1. rfc1122 clearly states that advertised MSS
1790 DOES NOT include either tcp or ip options.
1792 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1793 are READ ONLY outside this function. --ANK (980731)
1795 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1797 struct tcp_sock *tp = tcp_sk(sk);
1798 struct inet_connection_sock *icsk = inet_csk(sk);
1801 if (icsk->icsk_mtup.search_high > pmtu)
1802 icsk->icsk_mtup.search_high = pmtu;
1804 mss_now = tcp_mtu_to_mss(sk, pmtu);
1805 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1807 /* And store cached results */
1808 icsk->icsk_pmtu_cookie = pmtu;
1809 if (icsk->icsk_mtup.enabled)
1810 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1811 tp->mss_cache = mss_now;
1815 EXPORT_SYMBOL(tcp_sync_mss);
1817 /* Compute the current effective MSS, taking SACKs and IP options,
1818 * and even PMTU discovery events into account.
1820 unsigned int tcp_current_mss(struct sock *sk)
1822 const struct tcp_sock *tp = tcp_sk(sk);
1823 const struct dst_entry *dst = __sk_dst_get(sk);
1825 unsigned int header_len;
1826 struct tcp_out_options opts;
1827 struct tcp_md5sig_key *md5;
1829 mss_now = tp->mss_cache;
1832 u32 mtu = dst_mtu(dst);
1833 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1834 mss_now = tcp_sync_mss(sk, mtu);
1837 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1838 sizeof(struct tcphdr);
1839 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1840 * some common options. If this is an odd packet (because we have SACK
1841 * blocks etc) then our calculated header_len will be different, and
1842 * we have to adjust mss_now correspondingly */
1843 if (header_len != tp->tcp_header_len) {
1844 int delta = (int) header_len - tp->tcp_header_len;
1851 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1852 * As additional protections, we do not touch cwnd in retransmission phases,
1853 * and if application hit its sndbuf limit recently.
1855 static void tcp_cwnd_application_limited(struct sock *sk)
1857 struct tcp_sock *tp = tcp_sk(sk);
1859 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1860 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1861 /* Limited by application or receiver window. */
1862 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1863 u32 win_used = max(tp->snd_cwnd_used, init_win);
1864 if (win_used < tcp_snd_cwnd(tp)) {
1865 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1866 tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1);
1868 tp->snd_cwnd_used = 0;
1870 tp->snd_cwnd_stamp = tcp_jiffies32;
1873 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1875 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1876 struct tcp_sock *tp = tcp_sk(sk);
1878 /* Track the strongest available signal of the degree to which the cwnd
1879 * is fully utilized. If cwnd-limited then remember that fact for the
1880 * current window. If not cwnd-limited then track the maximum number of
1881 * outstanding packets in the current window. (If cwnd-limited then we
1882 * chose to not update tp->max_packets_out to avoid an extra else
1883 * clause with no functional impact.)
1885 if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
1887 (!tp->is_cwnd_limited &&
1888 tp->packets_out > tp->max_packets_out)) {
1889 tp->is_cwnd_limited = is_cwnd_limited;
1890 tp->max_packets_out = tp->packets_out;
1891 tp->cwnd_usage_seq = tp->snd_nxt;
1894 if (tcp_is_cwnd_limited(sk)) {
1895 /* Network is feed fully. */
1896 tp->snd_cwnd_used = 0;
1897 tp->snd_cwnd_stamp = tcp_jiffies32;
1899 /* Network starves. */
1900 if (tp->packets_out > tp->snd_cwnd_used)
1901 tp->snd_cwnd_used = tp->packets_out;
1903 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
1904 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1905 !ca_ops->cong_control)
1906 tcp_cwnd_application_limited(sk);
1908 /* The following conditions together indicate the starvation
1909 * is caused by insufficient sender buffer:
1910 * 1) just sent some data (see tcp_write_xmit)
1911 * 2) not cwnd limited (this else condition)
1912 * 3) no more data to send (tcp_write_queue_empty())
1913 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1915 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1916 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1917 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1918 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1922 /* Minshall's variant of the Nagle send check. */
1923 static bool tcp_minshall_check(const struct tcp_sock *tp)
1925 return after(tp->snd_sml, tp->snd_una) &&
1926 !after(tp->snd_sml, tp->snd_nxt);
1929 /* Update snd_sml if this skb is under mss
1930 * Note that a TSO packet might end with a sub-mss segment
1931 * The test is really :
1932 * if ((skb->len % mss) != 0)
1933 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1934 * But we can avoid doing the divide again given we already have
1935 * skb_pcount = skb->len / mss_now
1937 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1938 const struct sk_buff *skb)
1940 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1941 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1944 /* Return false, if packet can be sent now without violation Nagle's rules:
1945 * 1. It is full sized. (provided by caller in %partial bool)
1946 * 2. Or it contains FIN. (already checked by caller)
1947 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1948 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1949 * With Minshall's modification: all sent small packets are ACKed.
1951 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1955 ((nonagle & TCP_NAGLE_CORK) ||
1956 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1959 /* Return how many segs we'd like on a TSO packet,
1960 * depending on current pacing rate, and how close the peer is.
1963 * - For close peers, we rather send bigger packets to reduce
1964 * cpu costs, because occasional losses will be repaired fast.
1965 * - For long distance/rtt flows, we would like to get ACK clocking
1966 * with 1 ACK per ms.
1968 * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting
1969 * in bigger TSO bursts. We we cut the RTT-based allowance in half
1970 * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance
1971 * is below 1500 bytes after 6 * ~500 usec = 3ms.
1973 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1976 unsigned long bytes;
1979 bytes = sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift);
1981 r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log);
1982 if (r < BITS_PER_TYPE(sk->sk_gso_max_size))
1983 bytes += sk->sk_gso_max_size >> r;
1985 bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size);
1987 return max_t(u32, bytes / mss_now, min_tso_segs);
1990 /* Return the number of segments we want in the skb we are transmitting.
1991 * See if congestion control module wants to decide; otherwise, autosize.
1993 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1995 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1996 u32 min_tso, tso_segs;
1998 min_tso = ca_ops->min_tso_segs ?
1999 ca_ops->min_tso_segs(sk) :
2000 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
2002 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
2003 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
2006 /* Returns the portion of skb which can be sent right away */
2007 static unsigned int tcp_mss_split_point(const struct sock *sk,
2008 const struct sk_buff *skb,
2009 unsigned int mss_now,
2010 unsigned int max_segs,
2013 const struct tcp_sock *tp = tcp_sk(sk);
2014 u32 partial, needed, window, max_len;
2016 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2017 max_len = mss_now * max_segs;
2019 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2022 needed = min(skb->len, window);
2024 if (max_len <= needed)
2027 partial = needed % mss_now;
2028 /* If last segment is not a full MSS, check if Nagle rules allow us
2029 * to include this last segment in this skb.
2030 * Otherwise, we'll split the skb at last MSS boundary
2032 if (tcp_nagle_check(partial != 0, tp, nonagle))
2033 return needed - partial;
2038 /* Can at least one segment of SKB be sent right now, according to the
2039 * congestion window rules? If so, return how many segments are allowed.
2041 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
2042 const struct sk_buff *skb)
2044 u32 in_flight, cwnd, halfcwnd;
2046 /* Don't be strict about the congestion window for the final FIN. */
2047 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2048 tcp_skb_pcount(skb) == 1)
2051 in_flight = tcp_packets_in_flight(tp);
2052 cwnd = tcp_snd_cwnd(tp);
2053 if (in_flight >= cwnd)
2056 /* For better scheduling, ensure we have at least
2057 * 2 GSO packets in flight.
2059 halfcwnd = max(cwnd >> 1, 1U);
2060 return min(halfcwnd, cwnd - in_flight);
2063 /* Initialize TSO state of a skb.
2064 * This must be invoked the first time we consider transmitting
2065 * SKB onto the wire.
2067 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
2069 int tso_segs = tcp_skb_pcount(skb);
2071 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
2072 tcp_set_skb_tso_segs(skb, mss_now);
2073 tso_segs = tcp_skb_pcount(skb);
2079 /* Return true if the Nagle test allows this packet to be
2082 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2083 unsigned int cur_mss, int nonagle)
2085 /* Nagle rule does not apply to frames, which sit in the middle of the
2086 * write_queue (they have no chances to get new data).
2088 * This is implemented in the callers, where they modify the 'nonagle'
2089 * argument based upon the location of SKB in the send queue.
2091 if (nonagle & TCP_NAGLE_PUSH)
2094 /* Don't use the nagle rule for urgent data (or for the final FIN). */
2095 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2098 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
2104 /* Does at least the first segment of SKB fit into the send window? */
2105 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2106 const struct sk_buff *skb,
2107 unsigned int cur_mss)
2109 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2111 if (skb->len > cur_mss)
2112 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2114 return !after(end_seq, tcp_wnd_end(tp));
2117 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2118 * which is put after SKB on the list. It is very much like
2119 * tcp_fragment() except that it may make several kinds of assumptions
2120 * in order to speed up the splitting operation. In particular, we
2121 * know that all the data is in scatter-gather pages, and that the
2122 * packet has never been sent out before (and thus is not cloned).
2124 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2125 unsigned int mss_now, gfp_t gfp)
2127 int nlen = skb->len - len;
2128 struct sk_buff *buff;
2131 /* All of a TSO frame must be composed of paged data. */
2132 if (skb->len != skb->data_len)
2133 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2134 skb, len, mss_now, gfp);
2136 buff = tcp_stream_alloc_skb(sk, 0, gfp, true);
2137 if (unlikely(!buff))
2139 skb_copy_decrypted(buff, skb);
2140 mptcp_skb_ext_copy(buff, skb);
2142 sk_wmem_queued_add(sk, buff->truesize);
2143 sk_mem_charge(sk, buff->truesize);
2144 buff->truesize += nlen;
2145 skb->truesize -= nlen;
2147 /* Correct the sequence numbers. */
2148 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2149 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2150 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2152 /* PSH and FIN should only be set in the second packet. */
2153 flags = TCP_SKB_CB(skb)->tcp_flags;
2154 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2155 TCP_SKB_CB(buff)->tcp_flags = flags;
2157 tcp_skb_fragment_eor(skb, buff);
2159 skb_split(skb, buff, len);
2160 tcp_fragment_tstamp(skb, buff);
2162 /* Fix up tso_factor for both original and new SKB. */
2163 tcp_set_skb_tso_segs(skb, mss_now);
2164 tcp_set_skb_tso_segs(buff, mss_now);
2166 /* Link BUFF into the send queue. */
2167 __skb_header_release(buff);
2168 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2173 /* Try to defer sending, if possible, in order to minimize the amount
2174 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2176 * This algorithm is from John Heffner.
2178 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2179 bool *is_cwnd_limited,
2180 bool *is_rwnd_limited,
2183 const struct inet_connection_sock *icsk = inet_csk(sk);
2184 u32 send_win, cong_win, limit, in_flight;
2185 struct tcp_sock *tp = tcp_sk(sk);
2186 struct sk_buff *head;
2190 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2193 /* Avoid bursty behavior by allowing defer
2194 * only if the last write was recent (1 ms).
2195 * Note that tp->tcp_wstamp_ns can be in the future if we have
2196 * packets waiting in a qdisc or device for EDT delivery.
2198 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2202 in_flight = tcp_packets_in_flight(tp);
2204 BUG_ON(tcp_skb_pcount(skb) <= 1);
2205 BUG_ON(tcp_snd_cwnd(tp) <= in_flight);
2207 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2209 /* From in_flight test above, we know that cwnd > in_flight. */
2210 cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache;
2212 limit = min(send_win, cong_win);
2214 /* If a full-sized TSO skb can be sent, do it. */
2215 if (limit >= max_segs * tp->mss_cache)
2218 /* Middle in queue won't get any more data, full sendable already? */
2219 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2222 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2224 u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache);
2226 /* If at least some fraction of a window is available,
2229 chunk /= win_divisor;
2233 /* Different approach, try not to defer past a single
2234 * ACK. Receiver should ACK every other full sized
2235 * frame, so if we have space for more than 3 frames
2238 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2242 /* TODO : use tsorted_sent_queue ? */
2243 head = tcp_rtx_queue_head(sk);
2246 delta = tp->tcp_clock_cache - head->tstamp;
2247 /* If next ACK is likely to come too late (half srtt), do not defer */
2248 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2251 /* Ok, it looks like it is advisable to defer.
2252 * Three cases are tracked :
2253 * 1) We are cwnd-limited
2254 * 2) We are rwnd-limited
2255 * 3) We are application limited.
2257 if (cong_win < send_win) {
2258 if (cong_win <= skb->len) {
2259 *is_cwnd_limited = true;
2263 if (send_win <= skb->len) {
2264 *is_rwnd_limited = true;
2269 /* If this packet won't get more data, do not wait. */
2270 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2271 TCP_SKB_CB(skb)->eor)
2280 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2282 struct inet_connection_sock *icsk = inet_csk(sk);
2283 struct tcp_sock *tp = tcp_sk(sk);
2284 struct net *net = sock_net(sk);
2288 interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2289 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2290 if (unlikely(delta >= interval * HZ)) {
2291 int mss = tcp_current_mss(sk);
2293 /* Update current search range */
2294 icsk->icsk_mtup.probe_size = 0;
2295 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2296 sizeof(struct tcphdr) +
2297 icsk->icsk_af_ops->net_header_len;
2298 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2300 /* Update probe time stamp */
2301 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2305 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2307 struct sk_buff *skb, *next;
2309 skb = tcp_send_head(sk);
2310 tcp_for_write_queue_from_safe(skb, next, sk) {
2311 if (len <= skb->len)
2314 if (unlikely(TCP_SKB_CB(skb)->eor) ||
2315 tcp_has_tx_tstamp(skb) ||
2316 !skb_pure_zcopy_same(skb, next))
2325 /* Create a new MTU probe if we are ready.
2326 * MTU probe is regularly attempting to increase the path MTU by
2327 * deliberately sending larger packets. This discovers routing
2328 * changes resulting in larger path MTUs.
2330 * Returns 0 if we should wait to probe (no cwnd available),
2331 * 1 if a probe was sent,
2334 static int tcp_mtu_probe(struct sock *sk)
2336 struct inet_connection_sock *icsk = inet_csk(sk);
2337 struct tcp_sock *tp = tcp_sk(sk);
2338 struct sk_buff *skb, *nskb, *next;
2339 struct net *net = sock_net(sk);
2346 /* Not currently probing/verifying,
2348 * have enough cwnd, and
2349 * not SACKing (the variable headers throw things off)
2351 if (likely(!icsk->icsk_mtup.enabled ||
2352 icsk->icsk_mtup.probe_size ||
2353 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2354 tcp_snd_cwnd(tp) < 11 ||
2355 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2358 /* Use binary search for probe_size between tcp_mss_base,
2359 * and current mss_clamp. if (search_high - search_low)
2360 * smaller than a threshold, backoff from probing.
2362 mss_now = tcp_current_mss(sk);
2363 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2364 icsk->icsk_mtup.search_low) >> 1);
2365 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2366 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2367 /* When misfortune happens, we are reprobing actively,
2368 * and then reprobe timer has expired. We stick with current
2369 * probing process by not resetting search range to its orignal.
2371 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2372 interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2373 /* Check whether enough time has elaplased for
2374 * another round of probing.
2376 tcp_mtu_check_reprobe(sk);
2380 /* Have enough data in the send queue to probe? */
2381 if (tp->write_seq - tp->snd_nxt < size_needed)
2384 if (tp->snd_wnd < size_needed)
2386 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2389 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2390 if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) {
2391 if (!tcp_packets_in_flight(tp))
2397 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2400 /* We're allowed to probe. Build it now. */
2401 nskb = tcp_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2404 sk_wmem_queued_add(sk, nskb->truesize);
2405 sk_mem_charge(sk, nskb->truesize);
2407 skb = tcp_send_head(sk);
2408 skb_copy_decrypted(nskb, skb);
2409 mptcp_skb_ext_copy(nskb, skb);
2411 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2412 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2413 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2415 tcp_insert_write_queue_before(nskb, skb, sk);
2416 tcp_highest_sack_replace(sk, skb, nskb);
2419 tcp_for_write_queue_from_safe(skb, next, sk) {
2420 copy = min_t(int, skb->len, probe_size - len);
2421 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2423 if (skb->len <= copy) {
2424 /* We've eaten all the data from this skb.
2426 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2427 /* If this is the last SKB we copy and eor is set
2428 * we need to propagate it to the new skb.
2430 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2431 tcp_skb_collapse_tstamp(nskb, skb);
2432 tcp_unlink_write_queue(skb, sk);
2433 tcp_wmem_free_skb(sk, skb);
2435 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2436 ~(TCPHDR_FIN|TCPHDR_PSH);
2437 if (!skb_shinfo(skb)->nr_frags) {
2438 skb_pull(skb, copy);
2440 __pskb_trim_head(skb, copy);
2441 tcp_set_skb_tso_segs(skb, mss_now);
2443 TCP_SKB_CB(skb)->seq += copy;
2448 if (len >= probe_size)
2451 tcp_init_tso_segs(nskb, nskb->len);
2453 /* We're ready to send. If this fails, the probe will
2454 * be resegmented into mss-sized pieces by tcp_write_xmit().
2456 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2457 /* Decrement cwnd here because we are sending
2458 * effectively two packets. */
2459 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1);
2460 tcp_event_new_data_sent(sk, nskb);
2462 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2463 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2464 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2472 static bool tcp_pacing_check(struct sock *sk)
2474 struct tcp_sock *tp = tcp_sk(sk);
2476 if (!tcp_needs_internal_pacing(sk))
2479 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2482 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2483 hrtimer_start(&tp->pacing_timer,
2484 ns_to_ktime(tp->tcp_wstamp_ns),
2485 HRTIMER_MODE_ABS_PINNED_SOFT);
2491 /* TCP Small Queues :
2492 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2493 * (These limits are doubled for retransmits)
2495 * - better RTT estimation and ACK scheduling
2498 * Alas, some drivers / subsystems require a fair amount
2499 * of queued bytes to ensure line rate.
2500 * One example is wifi aggregation (802.11 AMPDU)
2502 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2503 unsigned int factor)
2505 unsigned long limit;
2507 limit = max_t(unsigned long,
2509 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2510 if (sk->sk_pacing_status == SK_PACING_NONE)
2511 limit = min_t(unsigned long, limit,
2512 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
2515 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2516 tcp_sk(sk)->tcp_tx_delay) {
2517 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2519 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2520 * approximate our needs assuming an ~100% skb->truesize overhead.
2521 * USEC_PER_SEC is approximated by 2^20.
2522 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2524 extra_bytes >>= (20 - 1);
2525 limit += extra_bytes;
2527 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2528 /* Always send skb if rtx queue is empty.
2529 * No need to wait for TX completion to call us back,
2530 * after softirq/tasklet schedule.
2531 * This helps when TX completions are delayed too much.
2533 if (tcp_rtx_queue_empty(sk))
2536 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2537 /* It is possible TX completion already happened
2538 * before we set TSQ_THROTTLED, so we must
2539 * test again the condition.
2541 smp_mb__after_atomic();
2542 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2548 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2550 const u32 now = tcp_jiffies32;
2551 enum tcp_chrono old = tp->chrono_type;
2553 if (old > TCP_CHRONO_UNSPEC)
2554 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2555 tp->chrono_start = now;
2556 tp->chrono_type = new;
2559 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2561 struct tcp_sock *tp = tcp_sk(sk);
2563 /* If there are multiple conditions worthy of tracking in a
2564 * chronograph then the highest priority enum takes precedence
2565 * over the other conditions. So that if something "more interesting"
2566 * starts happening, stop the previous chrono and start a new one.
2568 if (type > tp->chrono_type)
2569 tcp_chrono_set(tp, type);
2572 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2574 struct tcp_sock *tp = tcp_sk(sk);
2577 /* There are multiple conditions worthy of tracking in a
2578 * chronograph, so that the highest priority enum takes
2579 * precedence over the other conditions (see tcp_chrono_start).
2580 * If a condition stops, we only stop chrono tracking if
2581 * it's the "most interesting" or current chrono we are
2582 * tracking and starts busy chrono if we have pending data.
2584 if (tcp_rtx_and_write_queues_empty(sk))
2585 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2586 else if (type == tp->chrono_type)
2587 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2590 /* This routine writes packets to the network. It advances the
2591 * send_head. This happens as incoming acks open up the remote
2594 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2595 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2596 * account rare use of URG, this is not a big flaw.
2598 * Send at most one packet when push_one > 0. Temporarily ignore
2599 * cwnd limit to force at most one packet out when push_one == 2.
2601 * Returns true, if no segments are in flight and we have queued segments,
2602 * but cannot send anything now because of SWS or another problem.
2604 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2605 int push_one, gfp_t gfp)
2607 struct tcp_sock *tp = tcp_sk(sk);
2608 struct sk_buff *skb;
2609 unsigned int tso_segs, sent_pkts;
2612 bool is_cwnd_limited = false, is_rwnd_limited = false;
2617 tcp_mstamp_refresh(tp);
2619 /* Do MTU probing. */
2620 result = tcp_mtu_probe(sk);
2623 } else if (result > 0) {
2628 max_segs = tcp_tso_segs(sk, mss_now);
2629 while ((skb = tcp_send_head(sk))) {
2632 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2633 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2634 tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2635 skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
2636 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2637 tcp_init_tso_segs(skb, mss_now);
2638 goto repair; /* Skip network transmission */
2641 if (tcp_pacing_check(sk))
2644 tso_segs = tcp_init_tso_segs(skb, mss_now);
2647 cwnd_quota = tcp_cwnd_test(tp, skb);
2650 /* Force out a loss probe pkt. */
2656 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2657 is_rwnd_limited = true;
2661 if (tso_segs == 1) {
2662 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2663 (tcp_skb_is_last(sk, skb) ?
2664 nonagle : TCP_NAGLE_PUSH))))
2668 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2669 &is_rwnd_limited, max_segs))
2674 if (tso_segs > 1 && !tcp_urg_mode(tp))
2675 limit = tcp_mss_split_point(sk, skb, mss_now,
2681 if (skb->len > limit &&
2682 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2685 if (tcp_small_queue_check(sk, skb, 0))
2688 /* Argh, we hit an empty skb(), presumably a thread
2689 * is sleeping in sendmsg()/sk_stream_wait_memory().
2690 * We do not want to send a pure-ack packet and have
2691 * a strange looking rtx queue with empty packet(s).
2693 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2696 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2700 /* Advance the send_head. This one is sent out.
2701 * This call will increment packets_out.
2703 tcp_event_new_data_sent(sk, skb);
2705 tcp_minshall_update(tp, mss_now, skb);
2706 sent_pkts += tcp_skb_pcount(skb);
2712 if (is_rwnd_limited)
2713 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2715 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2717 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp));
2718 if (likely(sent_pkts || is_cwnd_limited))
2719 tcp_cwnd_validate(sk, is_cwnd_limited);
2721 if (likely(sent_pkts)) {
2722 if (tcp_in_cwnd_reduction(sk))
2723 tp->prr_out += sent_pkts;
2725 /* Send one loss probe per tail loss episode. */
2727 tcp_schedule_loss_probe(sk, false);
2730 return !tp->packets_out && !tcp_write_queue_empty(sk);
2733 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2735 struct inet_connection_sock *icsk = inet_csk(sk);
2736 struct tcp_sock *tp = tcp_sk(sk);
2737 u32 timeout, rto_delta_us;
2740 /* Don't do any loss probe on a Fast Open connection before 3WHS
2743 if (rcu_access_pointer(tp->fastopen_rsk))
2746 early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
2747 /* Schedule a loss probe in 2*RTT for SACK capable connections
2748 * not in loss recovery, that are either limited by cwnd or application.
2750 if ((early_retrans != 3 && early_retrans != 4) ||
2751 !tp->packets_out || !tcp_is_sack(tp) ||
2752 (icsk->icsk_ca_state != TCP_CA_Open &&
2753 icsk->icsk_ca_state != TCP_CA_CWR))
2756 /* Probe timeout is 2*rtt. Add minimum RTO to account
2757 * for delayed ack when there's one outstanding packet. If no RTT
2758 * sample is available then probe after TCP_TIMEOUT_INIT.
2761 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2762 if (tp->packets_out == 1)
2763 timeout += TCP_RTO_MIN;
2765 timeout += TCP_TIMEOUT_MIN;
2767 timeout = TCP_TIMEOUT_INIT;
2770 /* If the RTO formula yields an earlier time, then use that time. */
2771 rto_delta_us = advancing_rto ?
2772 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2773 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2774 if (rto_delta_us > 0)
2775 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2777 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2781 /* Thanks to skb fast clones, we can detect if a prior transmit of
2782 * a packet is still in a qdisc or driver queue.
2783 * In this case, there is very little point doing a retransmit !
2785 static bool skb_still_in_host_queue(struct sock *sk,
2786 const struct sk_buff *skb)
2788 if (unlikely(skb_fclone_busy(sk, skb))) {
2789 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2790 smp_mb__after_atomic();
2791 if (skb_fclone_busy(sk, skb)) {
2792 NET_INC_STATS(sock_net(sk),
2793 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2800 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2801 * retransmit the last segment.
2803 void tcp_send_loss_probe(struct sock *sk)
2805 struct tcp_sock *tp = tcp_sk(sk);
2806 struct sk_buff *skb;
2808 int mss = tcp_current_mss(sk);
2810 /* At most one outstanding TLP */
2811 if (tp->tlp_high_seq)
2814 tp->tlp_retrans = 0;
2815 skb = tcp_send_head(sk);
2816 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2817 pcount = tp->packets_out;
2818 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2819 if (tp->packets_out > pcount)
2823 skb = skb_rb_last(&sk->tcp_rtx_queue);
2824 if (unlikely(!skb)) {
2825 WARN_ONCE(tp->packets_out,
2826 "invalid inflight: %u state %u cwnd %u mss %d\n",
2827 tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss);
2828 inet_csk(sk)->icsk_pending = 0;
2832 if (skb_still_in_host_queue(sk, skb))
2835 pcount = tcp_skb_pcount(skb);
2836 if (WARN_ON(!pcount))
2839 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2840 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2841 (pcount - 1) * mss, mss,
2844 skb = skb_rb_next(skb);
2847 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2850 if (__tcp_retransmit_skb(sk, skb, 1))
2853 tp->tlp_retrans = 1;
2856 /* Record snd_nxt for loss detection. */
2857 tp->tlp_high_seq = tp->snd_nxt;
2859 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2860 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2861 inet_csk(sk)->icsk_pending = 0;
2866 /* Push out any pending frames which were held back due to
2867 * TCP_CORK or attempt at coalescing tiny packets.
2868 * The socket must be locked by the caller.
2870 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2873 /* If we are closed, the bytes will have to remain here.
2874 * In time closedown will finish, we empty the write queue and
2875 * all will be happy.
2877 if (unlikely(sk->sk_state == TCP_CLOSE))
2880 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2881 sk_gfp_mask(sk, GFP_ATOMIC)))
2882 tcp_check_probe_timer(sk);
2885 /* Send _single_ skb sitting at the send head. This function requires
2886 * true push pending frames to setup probe timer etc.
2888 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2890 struct sk_buff *skb = tcp_send_head(sk);
2892 BUG_ON(!skb || skb->len < mss_now);
2894 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2897 /* This function returns the amount that we can raise the
2898 * usable window based on the following constraints
2900 * 1. The window can never be shrunk once it is offered (RFC 793)
2901 * 2. We limit memory per socket
2904 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2905 * RECV.NEXT + RCV.WIN fixed until:
2906 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2908 * i.e. don't raise the right edge of the window until you can raise
2909 * it at least MSS bytes.
2911 * Unfortunately, the recommended algorithm breaks header prediction,
2912 * since header prediction assumes th->window stays fixed.
2914 * Strictly speaking, keeping th->window fixed violates the receiver
2915 * side SWS prevention criteria. The problem is that under this rule
2916 * a stream of single byte packets will cause the right side of the
2917 * window to always advance by a single byte.
2919 * Of course, if the sender implements sender side SWS prevention
2920 * then this will not be a problem.
2922 * BSD seems to make the following compromise:
2924 * If the free space is less than the 1/4 of the maximum
2925 * space available and the free space is less than 1/2 mss,
2926 * then set the window to 0.
2927 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2928 * Otherwise, just prevent the window from shrinking
2929 * and from being larger than the largest representable value.
2931 * This prevents incremental opening of the window in the regime
2932 * where TCP is limited by the speed of the reader side taking
2933 * data out of the TCP receive queue. It does nothing about
2934 * those cases where the window is constrained on the sender side
2935 * because the pipeline is full.
2937 * BSD also seems to "accidentally" limit itself to windows that are a
2938 * multiple of MSS, at least until the free space gets quite small.
2939 * This would appear to be a side effect of the mbuf implementation.
2940 * Combining these two algorithms results in the observed behavior
2941 * of having a fixed window size at almost all times.
2943 * Below we obtain similar behavior by forcing the offered window to
2944 * a multiple of the mss when it is feasible to do so.
2946 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2947 * Regular options like TIMESTAMP are taken into account.
2949 u32 __tcp_select_window(struct sock *sk)
2951 struct inet_connection_sock *icsk = inet_csk(sk);
2952 struct tcp_sock *tp = tcp_sk(sk);
2953 /* MSS for the peer's data. Previous versions used mss_clamp
2954 * here. I don't know if the value based on our guesses
2955 * of peer's MSS is better for the performance. It's more correct
2956 * but may be worse for the performance because of rcv_mss
2957 * fluctuations. --SAW 1998/11/1
2959 int mss = icsk->icsk_ack.rcv_mss;
2960 int free_space = tcp_space(sk);
2961 int allowed_space = tcp_full_space(sk);
2962 int full_space, window;
2964 if (sk_is_mptcp(sk))
2965 mptcp_space(sk, &free_space, &allowed_space);
2967 full_space = min_t(int, tp->window_clamp, allowed_space);
2969 if (unlikely(mss > full_space)) {
2974 if (free_space < (full_space >> 1)) {
2975 icsk->icsk_ack.quick = 0;
2977 if (tcp_under_memory_pressure(sk))
2978 tcp_adjust_rcv_ssthresh(sk);
2980 /* free_space might become our new window, make sure we don't
2981 * increase it due to wscale.
2983 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2985 /* if free space is less than mss estimate, or is below 1/16th
2986 * of the maximum allowed, try to move to zero-window, else
2987 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2988 * new incoming data is dropped due to memory limits.
2989 * With large window, mss test triggers way too late in order
2990 * to announce zero window in time before rmem limit kicks in.
2992 if (free_space < (allowed_space >> 4) || free_space < mss)
2996 if (free_space > tp->rcv_ssthresh)
2997 free_space = tp->rcv_ssthresh;
2999 /* Don't do rounding if we are using window scaling, since the
3000 * scaled window will not line up with the MSS boundary anyway.
3002 if (tp->rx_opt.rcv_wscale) {
3003 window = free_space;
3005 /* Advertise enough space so that it won't get scaled away.
3006 * Import case: prevent zero window announcement if
3007 * 1<<rcv_wscale > mss.
3009 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
3011 window = tp->rcv_wnd;
3012 /* Get the largest window that is a nice multiple of mss.
3013 * Window clamp already applied above.
3014 * If our current window offering is within 1 mss of the
3015 * free space we just keep it. This prevents the divide
3016 * and multiply from happening most of the time.
3017 * We also don't do any window rounding when the free space
3020 if (window <= free_space - mss || window > free_space)
3021 window = rounddown(free_space, mss);
3022 else if (mss == full_space &&
3023 free_space > window + (full_space >> 1))
3024 window = free_space;
3030 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3031 const struct sk_buff *next_skb)
3033 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3034 const struct skb_shared_info *next_shinfo =
3035 skb_shinfo(next_skb);
3036 struct skb_shared_info *shinfo = skb_shinfo(skb);
3038 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3039 shinfo->tskey = next_shinfo->tskey;
3040 TCP_SKB_CB(skb)->txstamp_ack |=
3041 TCP_SKB_CB(next_skb)->txstamp_ack;
3045 /* Collapses two adjacent SKB's during retransmission. */
3046 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3048 struct tcp_sock *tp = tcp_sk(sk);
3049 struct sk_buff *next_skb = skb_rb_next(skb);
3052 next_skb_size = next_skb->len;
3054 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3056 if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3059 tcp_highest_sack_replace(sk, next_skb, skb);
3061 /* Update sequence range on original skb. */
3062 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3064 /* Merge over control information. This moves PSH/FIN etc. over */
3065 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3067 /* All done, get rid of second SKB and account for it so
3068 * packet counting does not break.
3070 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3071 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3073 /* changed transmit queue under us so clear hints */
3074 tcp_clear_retrans_hints_partial(tp);
3075 if (next_skb == tp->retransmit_skb_hint)
3076 tp->retransmit_skb_hint = skb;
3078 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3080 tcp_skb_collapse_tstamp(skb, next_skb);
3082 tcp_rtx_queue_unlink_and_free(next_skb, sk);
3086 /* Check if coalescing SKBs is legal. */
3087 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3089 if (tcp_skb_pcount(skb) > 1)
3091 if (skb_cloned(skb))
3093 /* Some heuristics for collapsing over SACK'd could be invented */
3094 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3100 /* Collapse packets in the retransmit queue to make to create
3101 * less packets on the wire. This is only done on retransmission.
3103 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3106 struct tcp_sock *tp = tcp_sk(sk);
3107 struct sk_buff *skb = to, *tmp;
3110 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
3112 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3115 skb_rbtree_walk_from_safe(skb, tmp) {
3116 if (!tcp_can_collapse(sk, skb))
3119 if (!tcp_skb_can_collapse(to, skb))
3132 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3135 if (!tcp_collapse_retrans(sk, to))
3140 /* This retransmits one SKB. Policy decisions and retransmit queue
3141 * state updates are done by the caller. Returns non-zero if an
3142 * error occurred which prevented the send.
3144 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3146 struct inet_connection_sock *icsk = inet_csk(sk);
3147 struct tcp_sock *tp = tcp_sk(sk);
3148 unsigned int cur_mss;
3152 /* Inconclusive MTU probe */
3153 if (icsk->icsk_mtup.probe_size)
3154 icsk->icsk_mtup.probe_size = 0;
3156 if (skb_still_in_host_queue(sk, skb))
3159 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
3160 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3164 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3168 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3169 return -EHOSTUNREACH; /* Routing failure or similar. */
3171 cur_mss = tcp_current_mss(sk);
3172 avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3174 /* If receiver has shrunk his window, and skb is out of
3175 * new window, do not retransmit it. The exception is the
3176 * case, when window is shrunk to zero. In this case
3177 * our retransmit of one segment serves as a zero window probe.
3179 if (avail_wnd <= 0) {
3180 if (TCP_SKB_CB(skb)->seq != tp->snd_una)
3182 avail_wnd = cur_mss;
3185 len = cur_mss * segs;
3186 if (len > avail_wnd) {
3187 len = rounddown(avail_wnd, cur_mss);
3191 if (skb->len > len) {
3192 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3193 cur_mss, GFP_ATOMIC))
3194 return -ENOMEM; /* We'll try again later. */
3196 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
3199 diff = tcp_skb_pcount(skb);
3200 tcp_set_skb_tso_segs(skb, cur_mss);
3201 diff -= tcp_skb_pcount(skb);
3203 tcp_adjust_pcount(sk, skb, diff);
3204 avail_wnd = min_t(int, avail_wnd, cur_mss);
3205 if (skb->len < avail_wnd)
3206 tcp_retrans_try_collapse(sk, skb, avail_wnd);
3209 /* RFC3168, section 6.1.1.1. ECN fallback */
3210 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3211 tcp_ecn_clear_syn(sk, skb);
3213 /* Update global and local TCP statistics. */
3214 segs = tcp_skb_pcount(skb);
3215 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3216 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3217 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3218 tp->total_retrans += segs;
3219 tp->bytes_retrans += skb->len;
3221 /* make sure skb->data is aligned on arches that require it
3222 * and check if ack-trimming & collapsing extended the headroom
3223 * beyond what csum_start can cover.
3225 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3226 skb_headroom(skb) >= 0xFFFF)) {
3227 struct sk_buff *nskb;
3229 tcp_skb_tsorted_save(skb) {
3230 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3233 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3237 } tcp_skb_tsorted_restore(skb);
3240 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3241 tcp_rate_skb_sent(sk, skb);
3244 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3247 /* To avoid taking spuriously low RTT samples based on a timestamp
3248 * for a transmit that never happened, always mark EVER_RETRANS
3250 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3252 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3253 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3254 TCP_SKB_CB(skb)->seq, segs, err);
3257 trace_tcp_retransmit_skb(sk, skb);
3258 } else if (err != -EBUSY) {
3259 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3264 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3266 struct tcp_sock *tp = tcp_sk(sk);
3267 int err = __tcp_retransmit_skb(sk, skb, segs);
3270 #if FASTRETRANS_DEBUG > 0
3271 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3272 net_dbg_ratelimited("retrans_out leaked\n");
3275 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3276 tp->retrans_out += tcp_skb_pcount(skb);
3279 /* Save stamp of the first (attempted) retransmit. */
3280 if (!tp->retrans_stamp)
3281 tp->retrans_stamp = tcp_skb_timestamp(skb);
3283 if (tp->undo_retrans < 0)
3284 tp->undo_retrans = 0;
3285 tp->undo_retrans += tcp_skb_pcount(skb);
3289 /* This gets called after a retransmit timeout, and the initially
3290 * retransmitted data is acknowledged. It tries to continue
3291 * resending the rest of the retransmit queue, until either
3292 * we've sent it all or the congestion window limit is reached.
3294 void tcp_xmit_retransmit_queue(struct sock *sk)
3296 const struct inet_connection_sock *icsk = inet_csk(sk);
3297 struct sk_buff *skb, *rtx_head, *hole = NULL;
3298 struct tcp_sock *tp = tcp_sk(sk);
3299 bool rearm_timer = false;
3303 if (!tp->packets_out)
3306 rtx_head = tcp_rtx_queue_head(sk);
3307 skb = tp->retransmit_skb_hint ?: rtx_head;
3308 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3309 skb_rbtree_walk_from(skb) {
3313 if (tcp_pacing_check(sk))
3316 /* we could do better than to assign each time */
3318 tp->retransmit_skb_hint = skb;
3320 segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp);
3323 sacked = TCP_SKB_CB(skb)->sacked;
3324 /* In case tcp_shift_skb_data() have aggregated large skbs,
3325 * we need to make sure not sending too bigs TSO packets
3327 segs = min_t(int, segs, max_segs);
3329 if (tp->retrans_out >= tp->lost_out) {
3331 } else if (!(sacked & TCPCB_LOST)) {
3332 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3337 if (icsk->icsk_ca_state != TCP_CA_Loss)
3338 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3340 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3343 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3346 if (tcp_small_queue_check(sk, skb, 1))
3349 if (tcp_retransmit_skb(sk, skb, segs))
3352 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3354 if (tcp_in_cwnd_reduction(sk))
3355 tp->prr_out += tcp_skb_pcount(skb);
3357 if (skb == rtx_head &&
3358 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3363 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3364 inet_csk(sk)->icsk_rto,
3368 /* We allow to exceed memory limits for FIN packets to expedite
3369 * connection tear down and (memory) recovery.
3370 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3371 * or even be forced to close flow without any FIN.
3372 * In general, we want to allow one skb per socket to avoid hangs
3373 * with edge trigger epoll()
3375 void sk_forced_mem_schedule(struct sock *sk, int size)
3379 delta = size - sk->sk_forward_alloc;
3382 amt = sk_mem_pages(delta);
3383 sk->sk_forward_alloc += amt << PAGE_SHIFT;
3384 sk_memory_allocated_add(sk, amt);
3386 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3387 mem_cgroup_charge_skmem(sk->sk_memcg, amt,
3388 gfp_memcg_charge() | __GFP_NOFAIL);
3391 /* Send a FIN. The caller locks the socket for us.
3392 * We should try to send a FIN packet really hard, but eventually give up.
3394 void tcp_send_fin(struct sock *sk)
3396 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3397 struct tcp_sock *tp = tcp_sk(sk);
3399 /* Optimization, tack on the FIN if we have one skb in write queue and
3400 * this skb was not yet sent, or we are under memory pressure.
3401 * Note: in the latter case, FIN packet will be sent after a timeout,
3402 * as TCP stack thinks it has already been transmitted.
3405 if (!tskb && tcp_under_memory_pressure(sk))
3406 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3409 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3410 TCP_SKB_CB(tskb)->end_seq++;
3413 /* This means tskb was already sent.
3414 * Pretend we included the FIN on previous transmit.
3415 * We need to set tp->snd_nxt to the value it would have
3416 * if FIN had been sent. This is because retransmit path
3417 * does not change tp->snd_nxt.
3419 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3423 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3427 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3428 skb_reserve(skb, MAX_TCP_HEADER);
3429 sk_forced_mem_schedule(sk, skb->truesize);
3430 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3431 tcp_init_nondata_skb(skb, tp->write_seq,
3432 TCPHDR_ACK | TCPHDR_FIN);
3433 tcp_queue_skb(sk, skb);
3435 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3438 /* We get here when a process closes a file descriptor (either due to
3439 * an explicit close() or as a byproduct of exit()'ing) and there
3440 * was unread data in the receive queue. This behavior is recommended
3441 * by RFC 2525, section 2.17. -DaveM
3443 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3445 struct sk_buff *skb;
3447 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3449 /* NOTE: No TCP options attached and we never retransmit this. */
3450 skb = alloc_skb(MAX_TCP_HEADER, priority);
3452 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3456 /* Reserve space for headers and prepare control bits. */
3457 skb_reserve(skb, MAX_TCP_HEADER);
3458 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3459 TCPHDR_ACK | TCPHDR_RST);
3460 tcp_mstamp_refresh(tcp_sk(sk));
3462 if (tcp_transmit_skb(sk, skb, 0, priority))
3463 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3465 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3466 * skb here is different to the troublesome skb, so use NULL
3468 trace_tcp_send_reset(sk, NULL);
3471 /* Send a crossed SYN-ACK during socket establishment.
3472 * WARNING: This routine must only be called when we have already sent
3473 * a SYN packet that crossed the incoming SYN that caused this routine
3474 * to get called. If this assumption fails then the initial rcv_wnd
3475 * and rcv_wscale values will not be correct.
3477 int tcp_send_synack(struct sock *sk)
3479 struct sk_buff *skb;
3481 skb = tcp_rtx_queue_head(sk);
3482 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3483 pr_err("%s: wrong queue state\n", __func__);
3486 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3487 if (skb_cloned(skb)) {
3488 struct sk_buff *nskb;
3490 tcp_skb_tsorted_save(skb) {
3491 nskb = skb_copy(skb, GFP_ATOMIC);
3492 } tcp_skb_tsorted_restore(skb);
3495 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3496 tcp_highest_sack_replace(sk, skb, nskb);
3497 tcp_rtx_queue_unlink_and_free(skb, sk);
3498 __skb_header_release(nskb);
3499 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3500 sk_wmem_queued_add(sk, nskb->truesize);
3501 sk_mem_charge(sk, nskb->truesize);
3505 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3506 tcp_ecn_send_synack(sk, skb);
3508 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3512 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3513 * @sk: listener socket
3514 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3515 * should not use it again.
3516 * @req: request_sock pointer
3517 * @foc: cookie for tcp fast open
3518 * @synack_type: Type of synack to prepare
3519 * @syn_skb: SYN packet just received. It could be NULL for rtx case.
3521 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3522 struct request_sock *req,
3523 struct tcp_fastopen_cookie *foc,
3524 enum tcp_synack_type synack_type,
3525 struct sk_buff *syn_skb)
3527 struct inet_request_sock *ireq = inet_rsk(req);
3528 const struct tcp_sock *tp = tcp_sk(sk);
3529 struct tcp_md5sig_key *md5 = NULL;
3530 struct tcp_out_options opts;
3531 struct sk_buff *skb;
3532 int tcp_header_size;
3537 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3538 if (unlikely(!skb)) {
3542 /* Reserve space for headers. */
3543 skb_reserve(skb, MAX_TCP_HEADER);
3545 switch (synack_type) {
3546 case TCP_SYNACK_NORMAL:
3547 skb_set_owner_w(skb, req_to_sk(req));
3549 case TCP_SYNACK_COOKIE:
3550 /* Under synflood, we do not attach skb to a socket,
3551 * to avoid false sharing.
3554 case TCP_SYNACK_FASTOPEN:
3555 /* sk is a const pointer, because we want to express multiple
3556 * cpu might call us concurrently.
3557 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3559 skb_set_owner_w(skb, (struct sock *)sk);
3562 skb_dst_set(skb, dst);
3564 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3566 memset(&opts, 0, sizeof(opts));
3567 now = tcp_clock_ns();
3568 #ifdef CONFIG_SYN_COOKIES
3569 if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3570 skb_set_delivery_time(skb, cookie_init_timestamp(req, now),
3575 skb_set_delivery_time(skb, now, true);
3576 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3577 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3580 #ifdef CONFIG_TCP_MD5SIG
3582 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3584 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3585 /* bpf program will be interested in the tcp_flags */
3586 TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3587 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3589 syn_skb) + sizeof(*th);
3591 skb_push(skb, tcp_header_size);
3592 skb_reset_transport_header(skb);
3594 th = (struct tcphdr *)skb->data;
3595 memset(th, 0, sizeof(struct tcphdr));
3598 tcp_ecn_make_synack(req, th);
3599 th->source = htons(ireq->ir_num);
3600 th->dest = ireq->ir_rmt_port;
3601 skb->mark = ireq->ir_mark;
3602 skb->ip_summed = CHECKSUM_PARTIAL;
3603 th->seq = htonl(tcp_rsk(req)->snt_isn);
3604 /* XXX data is queued and acked as is. No buffer/window check */
3605 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3607 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3608 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3609 tcp_options_write(th, NULL, &opts);
3610 th->doff = (tcp_header_size >> 2);
3611 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3613 #ifdef CONFIG_TCP_MD5SIG
3614 /* Okay, we have all we need - do the md5 hash if needed */
3616 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3617 md5, req_to_sk(req), skb);
3621 bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
3622 synack_type, &opts);
3624 skb_set_delivery_time(skb, now, true);
3625 tcp_add_tx_delay(skb, tp);
3629 EXPORT_SYMBOL(tcp_make_synack);
3631 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3633 struct inet_connection_sock *icsk = inet_csk(sk);
3634 const struct tcp_congestion_ops *ca;
3635 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3637 if (ca_key == TCP_CA_UNSPEC)
3641 ca = tcp_ca_find_key(ca_key);
3642 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3643 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3644 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3645 icsk->icsk_ca_ops = ca;
3650 /* Do all connect socket setups that can be done AF independent. */
3651 static void tcp_connect_init(struct sock *sk)
3653 const struct dst_entry *dst = __sk_dst_get(sk);
3654 struct tcp_sock *tp = tcp_sk(sk);
3658 /* We'll fix this up when we get a response from the other end.
3659 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3661 tp->tcp_header_len = sizeof(struct tcphdr);
3662 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps))
3663 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3665 #ifdef CONFIG_TCP_MD5SIG
3666 if (tp->af_specific->md5_lookup(sk, sk))
3667 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3670 /* If user gave his TCP_MAXSEG, record it to clamp */
3671 if (tp->rx_opt.user_mss)
3672 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3675 tcp_sync_mss(sk, dst_mtu(dst));
3677 tcp_ca_dst_init(sk, dst);
3679 if (!tp->window_clamp)
3680 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3681 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3683 tcp_initialize_rcv_mss(sk);
3685 /* limit the window selection if the user enforce a smaller rx buffer */
3686 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3687 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3688 tp->window_clamp = tcp_full_space(sk);
3690 rcv_wnd = tcp_rwnd_init_bpf(sk);
3692 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3694 tcp_select_initial_window(sk, tcp_full_space(sk),
3695 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3698 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling),
3702 tp->rx_opt.rcv_wscale = rcv_wscale;
3703 tp->rcv_ssthresh = tp->rcv_wnd;
3706 sock_reset_flag(sk, SOCK_DONE);
3709 tcp_write_queue_purge(sk);
3710 tp->snd_una = tp->write_seq;
3711 tp->snd_sml = tp->write_seq;
3712 tp->snd_up = tp->write_seq;
3713 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3715 if (likely(!tp->repair))
3718 tp->rcv_tstamp = tcp_jiffies32;
3719 tp->rcv_wup = tp->rcv_nxt;
3720 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3722 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3723 inet_csk(sk)->icsk_retransmits = 0;
3724 tcp_clear_retrans(tp);
3727 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3729 struct tcp_sock *tp = tcp_sk(sk);
3730 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3732 tcb->end_seq += skb->len;
3733 __skb_header_release(skb);
3734 sk_wmem_queued_add(sk, skb->truesize);
3735 sk_mem_charge(sk, skb->truesize);
3736 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3737 tp->packets_out += tcp_skb_pcount(skb);
3740 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3741 * queue a data-only packet after the regular SYN, such that regular SYNs
3742 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3743 * only the SYN sequence, the data are retransmitted in the first ACK.
3744 * If cookie is not cached or other error occurs, falls back to send a
3745 * regular SYN with Fast Open cookie request option.
3747 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3749 struct inet_connection_sock *icsk = inet_csk(sk);
3750 struct tcp_sock *tp = tcp_sk(sk);
3751 struct tcp_fastopen_request *fo = tp->fastopen_req;
3753 struct sk_buff *syn_data;
3755 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3756 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3759 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3760 * user-MSS. Reserve maximum option space for middleboxes that add
3761 * private TCP options. The cost is reduced data space in SYN :(
3763 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3764 /* Sync mss_cache after updating the mss_clamp */
3765 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3767 space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3768 MAX_TCP_OPTION_SPACE;
3770 space = min_t(size_t, space, fo->size);
3772 /* limit to order-0 allocations */
3773 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3775 syn_data = tcp_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3778 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3780 int copied = copy_from_iter(skb_put(syn_data, space), space,
3781 &fo->data->msg_iter);
3782 if (unlikely(!copied)) {
3783 tcp_skb_tsorted_anchor_cleanup(syn_data);
3784 kfree_skb(syn_data);
3787 if (copied != space) {
3788 skb_trim(syn_data, copied);
3791 skb_zcopy_set(syn_data, fo->uarg, NULL);
3793 /* No more data pending in inet_wait_for_connect() */
3794 if (space == fo->size)
3798 tcp_connect_queue_skb(sk, syn_data);
3800 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3802 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3804 skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true);
3806 /* Now full SYN+DATA was cloned and sent (or not),
3807 * remove the SYN from the original skb (syn_data)
3808 * we keep in write queue in case of a retransmit, as we
3809 * also have the SYN packet (with no data) in the same queue.
3811 TCP_SKB_CB(syn_data)->seq++;
3812 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3814 tp->syn_data = (fo->copied > 0);
3815 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3816 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3820 /* data was not sent, put it in write_queue */
3821 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3822 tp->packets_out -= tcp_skb_pcount(syn_data);
3825 /* Send a regular SYN with Fast Open cookie request option */
3826 if (fo->cookie.len > 0)
3828 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3830 tp->syn_fastopen = 0;
3832 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3836 /* Build a SYN and send it off. */
3837 int tcp_connect(struct sock *sk)
3839 struct tcp_sock *tp = tcp_sk(sk);
3840 struct sk_buff *buff;
3843 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3845 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3846 return -EHOSTUNREACH; /* Routing failure or similar. */
3848 tcp_connect_init(sk);
3850 if (unlikely(tp->repair)) {
3851 tcp_finish_connect(sk, NULL);
3855 buff = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3856 if (unlikely(!buff))
3859 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3860 tcp_mstamp_refresh(tp);
3861 tp->retrans_stamp = tcp_time_stamp(tp);
3862 tcp_connect_queue_skb(sk, buff);
3863 tcp_ecn_send_syn(sk, buff);
3864 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3866 /* Send off SYN; include data in Fast Open. */
3867 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3868 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3869 if (err == -ECONNREFUSED)
3872 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3873 * in order to make this packet get counted in tcpOutSegs.
3875 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3876 tp->pushed_seq = tp->write_seq;
3877 buff = tcp_send_head(sk);
3878 if (unlikely(buff)) {
3879 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3880 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3882 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3884 /* Timer for repeating the SYN until an answer. */
3885 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3886 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3889 EXPORT_SYMBOL(tcp_connect);
3891 /* Send out a delayed ack, the caller does the policy checking
3892 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3895 void tcp_send_delayed_ack(struct sock *sk)
3897 struct inet_connection_sock *icsk = inet_csk(sk);
3898 int ato = icsk->icsk_ack.ato;
3899 unsigned long timeout;
3901 if (ato > TCP_DELACK_MIN) {
3902 const struct tcp_sock *tp = tcp_sk(sk);
3903 int max_ato = HZ / 2;
3905 if (inet_csk_in_pingpong_mode(sk) ||
3906 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3907 max_ato = TCP_DELACK_MAX;
3909 /* Slow path, intersegment interval is "high". */
3911 /* If some rtt estimate is known, use it to bound delayed ack.
3912 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3916 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3923 ato = min(ato, max_ato);
3926 ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max);
3928 /* Stay within the limit we were given */
3929 timeout = jiffies + ato;
3931 /* Use new timeout only if there wasn't a older one earlier. */
3932 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3933 /* If delack timer is about to expire, send ACK now. */
3934 if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3939 if (!time_before(timeout, icsk->icsk_ack.timeout))
3940 timeout = icsk->icsk_ack.timeout;
3942 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3943 icsk->icsk_ack.timeout = timeout;
3944 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3947 /* This routine sends an ack and also updates the window. */
3948 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3950 struct sk_buff *buff;
3952 /* If we have been reset, we may not send again. */
3953 if (sk->sk_state == TCP_CLOSE)
3956 /* We are not putting this on the write queue, so
3957 * tcp_transmit_skb() will set the ownership to this
3960 buff = alloc_skb(MAX_TCP_HEADER,
3961 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3962 if (unlikely(!buff)) {
3963 struct inet_connection_sock *icsk = inet_csk(sk);
3964 unsigned long delay;
3966 delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
3967 if (delay < TCP_RTO_MAX)
3968 icsk->icsk_ack.retry++;
3969 inet_csk_schedule_ack(sk);
3970 icsk->icsk_ack.ato = TCP_ATO_MIN;
3971 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
3975 /* Reserve space for headers and prepare control bits. */
3976 skb_reserve(buff, MAX_TCP_HEADER);
3977 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3979 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3981 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3983 skb_set_tcp_pure_ack(buff);
3985 /* Send it off, this clears delayed acks for us. */
3986 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3988 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3990 void tcp_send_ack(struct sock *sk)
3992 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3995 /* This routine sends a packet with an out of date sequence
3996 * number. It assumes the other end will try to ack it.
3998 * Question: what should we make while urgent mode?
3999 * 4.4BSD forces sending single byte of data. We cannot send
4000 * out of window data, because we have SND.NXT==SND.MAX...
4002 * Current solution: to send TWO zero-length segments in urgent mode:
4003 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
4004 * out-of-date with SND.UNA-1 to probe window.
4006 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
4008 struct tcp_sock *tp = tcp_sk(sk);
4009 struct sk_buff *skb;
4011 /* We don't queue it, tcp_transmit_skb() sets ownership. */
4012 skb = alloc_skb(MAX_TCP_HEADER,
4013 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4017 /* Reserve space for headers and set control bits. */
4018 skb_reserve(skb, MAX_TCP_HEADER);
4019 /* Use a previous sequence. This should cause the other
4020 * end to send an ack. Don't queue or clone SKB, just
4023 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
4024 NET_INC_STATS(sock_net(sk), mib);
4025 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4028 /* Called from setsockopt( ... TCP_REPAIR ) */
4029 void tcp_send_window_probe(struct sock *sk)
4031 if (sk->sk_state == TCP_ESTABLISHED) {
4032 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4033 tcp_mstamp_refresh(tcp_sk(sk));
4034 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4038 /* Initiate keepalive or window probe from timer. */
4039 int tcp_write_wakeup(struct sock *sk, int mib)
4041 struct tcp_sock *tp = tcp_sk(sk);
4042 struct sk_buff *skb;
4044 if (sk->sk_state == TCP_CLOSE)
4047 skb = tcp_send_head(sk);
4048 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4050 unsigned int mss = tcp_current_mss(sk);
4051 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4053 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4054 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4056 /* We are probing the opening of a window
4057 * but the window size is != 0
4058 * must have been a result SWS avoidance ( sender )
4060 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4062 seg_size = min(seg_size, mss);
4063 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4064 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4065 skb, seg_size, mss, GFP_ATOMIC))
4067 } else if (!tcp_skb_pcount(skb))
4068 tcp_set_skb_tso_segs(skb, mss);
4070 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4071 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4073 tcp_event_new_data_sent(sk, skb);
4076 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4077 tcp_xmit_probe_skb(sk, 1, mib);
4078 return tcp_xmit_probe_skb(sk, 0, mib);
4082 /* A window probe timeout has occurred. If window is not closed send
4083 * a partial packet else a zero probe.
4085 void tcp_send_probe0(struct sock *sk)
4087 struct inet_connection_sock *icsk = inet_csk(sk);
4088 struct tcp_sock *tp = tcp_sk(sk);
4089 struct net *net = sock_net(sk);
4090 unsigned long timeout;
4093 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4095 if (tp->packets_out || tcp_write_queue_empty(sk)) {
4096 /* Cancel probe timer, if it is not required. */
4097 icsk->icsk_probes_out = 0;
4098 icsk->icsk_backoff = 0;
4099 icsk->icsk_probes_tstamp = 0;
4103 icsk->icsk_probes_out++;
4105 if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
4106 icsk->icsk_backoff++;
4107 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
4109 /* If packet was not sent due to local congestion,
4110 * Let senders fight for local resources conservatively.
4112 timeout = TCP_RESOURCE_PROBE_INTERVAL;
4115 timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout);
4116 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
4119 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4121 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4125 /* Paired with WRITE_ONCE() in sock_setsockopt() */
4126 if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED)
4127 tcp_rsk(req)->txhash = net_tx_rndhash();
4128 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4131 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4132 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4133 if (unlikely(tcp_passive_fastopen(sk)))
4134 tcp_sk(sk)->total_retrans++;
4135 trace_tcp_retransmit_synack(sk, req);
4139 EXPORT_SYMBOL(tcp_rtx_synack);