1 // SPDX-License-Identifier: GPL-2.0-only
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
38 #define pr_fmt(fmt) "TCP: " fmt
41 #include <net/mptcp.h>
43 #include <linux/compiler.h>
44 #include <linux/gfp.h>
45 #include <linux/module.h>
46 #include <linux/static_key.h>
48 #include <trace/events/tcp.h>
50 /* Refresh clocks of a TCP socket,
51 * ensuring monotically increasing values.
53 void tcp_mstamp_refresh(struct tcp_sock *tp)
55 u64 val = tcp_clock_ns();
57 tp->tcp_clock_cache = val;
58 tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
61 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
62 int push_one, gfp_t gfp);
64 /* Account for new data that has been sent to the network. */
65 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
67 struct inet_connection_sock *icsk = inet_csk(sk);
68 struct tcp_sock *tp = tcp_sk(sk);
69 unsigned int prior_packets = tp->packets_out;
71 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
73 __skb_unlink(skb, &sk->sk_write_queue);
74 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
76 if (tp->highest_sack == NULL)
77 tp->highest_sack = skb;
79 tp->packets_out += tcp_skb_pcount(skb);
80 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
83 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
87 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
88 * window scaling factor due to loss of precision.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
94 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
96 const struct tcp_sock *tp = tcp_sk(sk);
98 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
99 (tp->rx_opt.wscale_ok &&
100 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
103 return tcp_wnd_end(tp);
106 /* Calculate mss to advertise in SYN segment.
107 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
109 * 1. It is independent of path mtu.
110 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
111 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
112 * attached devices, because some buggy hosts are confused by
114 * 4. We do not make 3, we advertise MSS, calculated from first
115 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
116 * This may be overridden via information stored in routing table.
117 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
118 * probably even Jumbo".
120 static __u16 tcp_advertise_mss(struct sock *sk)
122 struct tcp_sock *tp = tcp_sk(sk);
123 const struct dst_entry *dst = __sk_dst_get(sk);
124 int mss = tp->advmss;
127 unsigned int metric = dst_metric_advmss(dst);
138 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
139 * This is the first part of cwnd validation mechanism.
141 void tcp_cwnd_restart(struct sock *sk, s32 delta)
143 struct tcp_sock *tp = tcp_sk(sk);
144 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
145 u32 cwnd = tp->snd_cwnd;
147 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
149 tp->snd_ssthresh = tcp_current_ssthresh(sk);
150 restart_cwnd = min(restart_cwnd, cwnd);
152 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
154 tp->snd_cwnd = max(cwnd, restart_cwnd);
155 tp->snd_cwnd_stamp = tcp_jiffies32;
156 tp->snd_cwnd_used = 0;
159 /* Congestion state accounting after a packet has been sent. */
160 static void tcp_event_data_sent(struct tcp_sock *tp,
163 struct inet_connection_sock *icsk = inet_csk(sk);
164 const u32 now = tcp_jiffies32;
166 if (tcp_packets_in_flight(tp) == 0)
167 tcp_ca_event(sk, CA_EVENT_TX_START);
169 /* If this is the first data packet sent in response to the
170 * previous received data,
171 * and it is a reply for ato after last received packet,
172 * increase pingpong count.
174 if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) &&
175 (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
176 inet_csk_inc_pingpong_cnt(sk);
181 /* Account for an ACK we sent. */
182 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
185 struct tcp_sock *tp = tcp_sk(sk);
187 if (unlikely(tp->compressed_ack)) {
188 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
190 tp->compressed_ack = 0;
191 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
195 if (unlikely(rcv_nxt != tp->rcv_nxt))
196 return; /* Special ACK sent by DCTCP to reflect ECN */
197 tcp_dec_quickack_mode(sk, pkts);
198 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
201 /* Determine a window scaling and initial window to offer.
202 * Based on the assumption that the given amount of space
203 * will be offered. Store the results in the tp structure.
204 * NOTE: for smooth operation initial space offering should
205 * be a multiple of mss if possible. We assume here that mss >= 1.
206 * This MUST be enforced by all callers.
208 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
209 __u32 *rcv_wnd, __u32 *window_clamp,
210 int wscale_ok, __u8 *rcv_wscale,
213 unsigned int space = (__space < 0 ? 0 : __space);
215 /* If no clamp set the clamp to the max possible scaled window */
216 if (*window_clamp == 0)
217 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
218 space = min(*window_clamp, space);
220 /* Quantize space offering to a multiple of mss if possible. */
222 space = rounddown(space, mss);
224 /* NOTE: offering an initial window larger than 32767
225 * will break some buggy TCP stacks. If the admin tells us
226 * it is likely we could be speaking with such a buggy stack
227 * we will truncate our initial window offering to 32K-1
228 * unless the remote has sent us a window scaling option,
229 * which we interpret as a sign the remote TCP is not
230 * misinterpreting the window field as a signed quantity.
232 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
233 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
235 (*rcv_wnd) = min_t(u32, space, U16_MAX);
238 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
242 /* Set window scaling on max possible window */
243 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
244 space = max_t(u32, space, sysctl_rmem_max);
245 space = min_t(u32, space, *window_clamp);
246 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
249 /* Set the clamp no higher than max representable value */
250 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
252 EXPORT_SYMBOL(tcp_select_initial_window);
254 /* Chose a new window to advertise, update state in tcp_sock for the
255 * socket, and return result with RFC1323 scaling applied. The return
256 * value can be stuffed directly into th->window for an outgoing
259 static u16 tcp_select_window(struct sock *sk)
261 struct tcp_sock *tp = tcp_sk(sk);
262 u32 old_win = tp->rcv_wnd;
263 u32 cur_win = tcp_receive_window(tp);
264 u32 new_win = __tcp_select_window(sk);
266 /* Never shrink the offered window */
267 if (new_win < cur_win) {
268 /* Danger Will Robinson!
269 * Don't update rcv_wup/rcv_wnd here or else
270 * we will not be able to advertise a zero
271 * window in time. --DaveM
273 * Relax Will Robinson.
276 NET_INC_STATS(sock_net(sk),
277 LINUX_MIB_TCPWANTZEROWINDOWADV);
278 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
280 tp->rcv_wnd = new_win;
281 tp->rcv_wup = tp->rcv_nxt;
283 /* Make sure we do not exceed the maximum possible
286 if (!tp->rx_opt.rcv_wscale &&
287 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
288 new_win = min(new_win, MAX_TCP_WINDOW);
290 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
292 /* RFC1323 scaling applied */
293 new_win >>= tp->rx_opt.rcv_wscale;
295 /* If we advertise zero window, disable fast path. */
299 NET_INC_STATS(sock_net(sk),
300 LINUX_MIB_TCPTOZEROWINDOWADV);
301 } else if (old_win == 0) {
302 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
308 /* Packet ECN state for a SYN-ACK */
309 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
311 const struct tcp_sock *tp = tcp_sk(sk);
313 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
314 if (!(tp->ecn_flags & TCP_ECN_OK))
315 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
316 else if (tcp_ca_needs_ecn(sk) ||
317 tcp_bpf_ca_needs_ecn(sk))
321 /* Packet ECN state for a SYN. */
322 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
324 struct tcp_sock *tp = tcp_sk(sk);
325 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
326 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
327 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
330 const struct dst_entry *dst = __sk_dst_get(sk);
332 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
339 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
340 tp->ecn_flags = TCP_ECN_OK;
341 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
346 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
348 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
349 /* tp->ecn_flags are cleared at a later point in time when
350 * SYN ACK is ultimatively being received.
352 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
356 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
358 if (inet_rsk(req)->ecn_ok)
362 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
365 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
366 struct tcphdr *th, int tcp_header_len)
368 struct tcp_sock *tp = tcp_sk(sk);
370 if (tp->ecn_flags & TCP_ECN_OK) {
371 /* Not-retransmitted data segment: set ECT and inject CWR. */
372 if (skb->len != tcp_header_len &&
373 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
375 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
376 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
378 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
380 } else if (!tcp_ca_needs_ecn(sk)) {
381 /* ACK or retransmitted segment: clear ECT|CE */
382 INET_ECN_dontxmit(sk);
384 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
389 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
390 * auto increment end seqno.
392 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
394 skb->ip_summed = CHECKSUM_PARTIAL;
396 TCP_SKB_CB(skb)->tcp_flags = flags;
397 TCP_SKB_CB(skb)->sacked = 0;
399 tcp_skb_pcount_set(skb, 1);
401 TCP_SKB_CB(skb)->seq = seq;
402 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
404 TCP_SKB_CB(skb)->end_seq = seq;
407 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
409 return tp->snd_una != tp->snd_up;
412 #define OPTION_SACK_ADVERTISE (1 << 0)
413 #define OPTION_TS (1 << 1)
414 #define OPTION_MD5 (1 << 2)
415 #define OPTION_WSCALE (1 << 3)
416 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
417 #define OPTION_SMC (1 << 9)
418 #define OPTION_MPTCP (1 << 10)
420 static void smc_options_write(__be32 *ptr, u16 *options)
422 #if IS_ENABLED(CONFIG_SMC)
423 if (static_branch_unlikely(&tcp_have_smc)) {
424 if (unlikely(OPTION_SMC & *options)) {
425 *ptr++ = htonl((TCPOPT_NOP << 24) |
428 (TCPOLEN_EXP_SMC_BASE));
429 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
435 struct tcp_out_options {
436 u16 options; /* bit field of OPTION_* */
437 u16 mss; /* 0 to disable */
438 u8 ws; /* window scale, 0 to disable */
439 u8 num_sack_blocks; /* number of SACK blocks to include */
440 u8 hash_size; /* bytes in hash_location */
441 u8 bpf_opt_len; /* length of BPF hdr option */
442 __u8 *hash_location; /* temporary pointer, overloaded */
443 __u32 tsval, tsecr; /* need to include OPTION_TS */
444 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
445 struct mptcp_out_options mptcp;
448 static void mptcp_options_write(__be32 *ptr, struct tcp_out_options *opts)
450 #if IS_ENABLED(CONFIG_MPTCP)
451 if (unlikely(OPTION_MPTCP & opts->options))
452 mptcp_write_options(ptr, &opts->mptcp);
456 #ifdef CONFIG_CGROUP_BPF
457 static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
458 enum tcp_synack_type synack_type)
461 return BPF_WRITE_HDR_TCP_CURRENT_MSS;
463 if (unlikely(synack_type == TCP_SYNACK_COOKIE))
464 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
469 /* req, syn_skb and synack_type are used when writing synack */
470 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
471 struct request_sock *req,
472 struct sk_buff *syn_skb,
473 enum tcp_synack_type synack_type,
474 struct tcp_out_options *opts,
475 unsigned int *remaining)
477 struct bpf_sock_ops_kern sock_ops;
480 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
481 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
485 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
488 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
490 sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
493 /* The listen "sk" cannot be passed here because
494 * it is not locked. It would not make too much
495 * sense to do bpf_setsockopt(listen_sk) based
496 * on individual connection request also.
498 * Thus, "req" is passed here and the cgroup-bpf-progs
499 * of the listen "sk" will be run.
501 * "req" is also used here for fastopen even the "sk" here is
502 * a fullsock "child" sk. It is to keep the behavior
503 * consistent between fastopen and non-fastopen on
504 * the bpf programming side.
506 sock_ops.sk = (struct sock *)req;
507 sock_ops.syn_skb = syn_skb;
509 sock_owned_by_me(sk);
511 sock_ops.is_fullsock = 1;
515 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
516 sock_ops.remaining_opt_len = *remaining;
517 /* tcp_current_mss() does not pass a skb */
519 bpf_skops_init_skb(&sock_ops, skb, 0);
521 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
523 if (err || sock_ops.remaining_opt_len == *remaining)
526 opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
527 /* round up to 4 bytes */
528 opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
530 *remaining -= opts->bpf_opt_len;
533 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
534 struct request_sock *req,
535 struct sk_buff *syn_skb,
536 enum tcp_synack_type synack_type,
537 struct tcp_out_options *opts)
539 u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
540 struct bpf_sock_ops_kern sock_ops;
543 if (likely(!max_opt_len))
546 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
548 sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
551 sock_ops.sk = (struct sock *)req;
552 sock_ops.syn_skb = syn_skb;
554 sock_owned_by_me(sk);
556 sock_ops.is_fullsock = 1;
560 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
561 sock_ops.remaining_opt_len = max_opt_len;
562 first_opt_off = tcp_hdrlen(skb) - max_opt_len;
563 bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
565 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
570 nr_written = max_opt_len - sock_ops.remaining_opt_len;
572 if (nr_written < max_opt_len)
573 memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
574 max_opt_len - nr_written);
577 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
578 struct request_sock *req,
579 struct sk_buff *syn_skb,
580 enum tcp_synack_type synack_type,
581 struct tcp_out_options *opts,
582 unsigned int *remaining)
586 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
587 struct request_sock *req,
588 struct sk_buff *syn_skb,
589 enum tcp_synack_type synack_type,
590 struct tcp_out_options *opts)
595 /* Write previously computed TCP options to the packet.
597 * Beware: Something in the Internet is very sensitive to the ordering of
598 * TCP options, we learned this through the hard way, so be careful here.
599 * Luckily we can at least blame others for their non-compliance but from
600 * inter-operability perspective it seems that we're somewhat stuck with
601 * the ordering which we have been using if we want to keep working with
602 * those broken things (not that it currently hurts anybody as there isn't
603 * particular reason why the ordering would need to be changed).
605 * At least SACK_PERM as the first option is known to lead to a disaster
606 * (but it may well be that other scenarios fail similarly).
608 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
609 struct tcp_out_options *opts)
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(ptr, 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(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(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(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 tp->snd_cwnd > 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(unsigned long data)
1043 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
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_init(&tsq->tasklet,
1130 (unsigned long)tsq);
1135 * Write buffer destructor automatically called from kfree_skb.
1136 * We can't xmit new skbs from this context, as we might already
1139 void tcp_wfree(struct sk_buff *skb)
1141 struct sock *sk = skb->sk;
1142 struct tcp_sock *tp = tcp_sk(sk);
1143 unsigned long flags, nval, oval;
1145 /* Keep one reference on sk_wmem_alloc.
1146 * Will be released by sk_free() from here or tcp_tasklet_func()
1148 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1150 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
1151 * Wait until our queues (qdisc + devices) are drained.
1153 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1154 * - chance for incoming ACK (processed by another cpu maybe)
1155 * to migrate this flow (skb->ooo_okay will be eventually set)
1157 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1160 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
1161 struct tsq_tasklet *tsq;
1164 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1167 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1168 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
1172 /* queue this socket to tasklet queue */
1173 local_irq_save(flags);
1174 tsq = this_cpu_ptr(&tsq_tasklet);
1175 empty = list_empty(&tsq->head);
1176 list_add(&tp->tsq_node, &tsq->head);
1178 tasklet_schedule(&tsq->tasklet);
1179 local_irq_restore(flags);
1186 /* Note: Called under soft irq.
1187 * We can call TCP stack right away, unless socket is owned by user.
1189 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1191 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1192 struct sock *sk = (struct sock *)tp;
1194 tcp_tsq_handler(sk);
1197 return HRTIMER_NORESTART;
1200 static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1203 struct tcp_sock *tp = tcp_sk(sk);
1205 if (sk->sk_pacing_status != SK_PACING_NONE) {
1206 unsigned long rate = sk->sk_pacing_rate;
1208 /* Original sch_fq does not pace first 10 MSS
1209 * Note that tp->data_segs_out overflows after 2^32 packets,
1210 * this is a minor annoyance.
1212 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1213 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1214 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1216 /* take into account OS jitter */
1217 len_ns -= min_t(u64, len_ns / 2, credit);
1218 tp->tcp_wstamp_ns += len_ns;
1221 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1224 INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1225 INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1226 INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1228 /* This routine actually transmits TCP packets queued in by
1229 * tcp_do_sendmsg(). This is used by both the initial
1230 * transmission and possible later retransmissions.
1231 * All SKB's seen here are completely headerless. It is our
1232 * job to build the TCP header, and pass the packet down to
1233 * IP so it can do the same plus pass the packet off to the
1236 * We are working here with either a clone of the original
1237 * SKB, or a fresh unique copy made by the retransmit engine.
1239 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1240 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1242 const struct inet_connection_sock *icsk = inet_csk(sk);
1243 struct inet_sock *inet;
1244 struct tcp_sock *tp;
1245 struct tcp_skb_cb *tcb;
1246 struct tcp_out_options opts;
1247 unsigned int tcp_options_size, tcp_header_size;
1248 struct sk_buff *oskb = NULL;
1249 struct tcp_md5sig_key *md5;
1254 BUG_ON(!skb || !tcp_skb_pcount(skb));
1256 prior_wstamp = tp->tcp_wstamp_ns;
1257 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1258 skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
1260 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1264 tcp_skb_tsorted_save(oskb) {
1265 if (unlikely(skb_cloned(oskb)))
1266 skb = pskb_copy(oskb, gfp_mask);
1268 skb = skb_clone(oskb, gfp_mask);
1269 } tcp_skb_tsorted_restore(oskb);
1273 /* retransmit skbs might have a non zero value in skb->dev
1274 * because skb->dev is aliased with skb->rbnode.rb_left
1280 tcb = TCP_SKB_CB(skb);
1281 memset(&opts, 0, sizeof(opts));
1283 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1284 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1286 tcp_options_size = tcp_established_options(sk, skb, &opts,
1288 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1289 * at receiver : This slightly improve GRO performance.
1290 * Note that we do not force the PSH flag for non GSO packets,
1291 * because they might be sent under high congestion events,
1292 * and in this case it is better to delay the delivery of 1-MSS
1293 * packets and thus the corresponding ACK packet that would
1294 * release the following packet.
1296 if (tcp_skb_pcount(skb) > 1)
1297 tcb->tcp_flags |= TCPHDR_PSH;
1299 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1301 /* if no packet is in qdisc/device queue, then allow XPS to select
1302 * another queue. We can be called from tcp_tsq_handler()
1303 * which holds one reference to sk.
1305 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1306 * One way to get this would be to set skb->truesize = 2 on them.
1308 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1310 /* If we had to use memory reserve to allocate this skb,
1311 * this might cause drops if packet is looped back :
1312 * Other socket might not have SOCK_MEMALLOC.
1313 * Packets not looped back do not care about pfmemalloc.
1315 skb->pfmemalloc = 0;
1317 skb_push(skb, tcp_header_size);
1318 skb_reset_transport_header(skb);
1322 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1323 skb_set_hash_from_sk(skb, sk);
1324 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1326 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1328 /* Build TCP header and checksum it. */
1329 th = (struct tcphdr *)skb->data;
1330 th->source = inet->inet_sport;
1331 th->dest = inet->inet_dport;
1332 th->seq = htonl(tcb->seq);
1333 th->ack_seq = htonl(rcv_nxt);
1334 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1340 /* The urg_mode check is necessary during a below snd_una win probe */
1341 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1342 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1343 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1345 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1346 th->urg_ptr = htons(0xFFFF);
1351 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1352 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1353 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1354 th->window = htons(tcp_select_window(sk));
1355 tcp_ecn_send(sk, skb, th, tcp_header_size);
1357 /* RFC1323: The window in SYN & SYN/ACK segments
1360 th->window = htons(min(tp->rcv_wnd, 65535U));
1362 #ifdef CONFIG_TCP_MD5SIG
1363 /* Calculate the MD5 hash, as we have all we need now */
1365 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1366 tp->af_specific->calc_md5_hash(opts.hash_location,
1371 /* BPF prog is the last one writing header option */
1372 bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1374 INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1375 tcp_v6_send_check, tcp_v4_send_check,
1378 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1379 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1381 if (skb->len != tcp_header_size) {
1382 tcp_event_data_sent(tp, sk);
1383 tp->data_segs_out += tcp_skb_pcount(skb);
1384 tp->bytes_sent += skb->len - tcp_header_size;
1387 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1388 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1389 tcp_skb_pcount(skb));
1391 tp->segs_out += tcp_skb_pcount(skb);
1392 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1393 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1394 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1396 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1398 /* Cleanup our debris for IP stacks */
1399 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1400 sizeof(struct inet6_skb_parm)));
1402 tcp_add_tx_delay(skb, tp);
1404 err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1405 inet6_csk_xmit, ip_queue_xmit,
1406 sk, skb, &inet->cork.fl);
1408 if (unlikely(err > 0)) {
1410 err = net_xmit_eval(err);
1413 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1414 tcp_rate_skb_sent(sk, oskb);
1419 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1422 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1423 tcp_sk(sk)->rcv_nxt);
1426 /* This routine just queues the buffer for sending.
1428 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1429 * otherwise socket can stall.
1431 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1433 struct tcp_sock *tp = tcp_sk(sk);
1435 /* Advance write_seq and place onto the write_queue. */
1436 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1437 __skb_header_release(skb);
1438 tcp_add_write_queue_tail(sk, skb);
1439 sk_wmem_queued_add(sk, skb->truesize);
1440 sk_mem_charge(sk, skb->truesize);
1443 /* Initialize TSO segments for a packet. */
1444 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1446 if (skb->len <= mss_now) {
1447 /* Avoid the costly divide in the normal
1450 tcp_skb_pcount_set(skb, 1);
1451 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1453 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1454 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1458 /* Pcount in the middle of the write queue got changed, we need to do various
1459 * tweaks to fix counters
1461 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1463 struct tcp_sock *tp = tcp_sk(sk);
1465 tp->packets_out -= decr;
1467 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1468 tp->sacked_out -= decr;
1469 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1470 tp->retrans_out -= decr;
1471 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1472 tp->lost_out -= decr;
1474 /* Reno case is special. Sigh... */
1475 if (tcp_is_reno(tp) && decr > 0)
1476 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1478 if (tp->lost_skb_hint &&
1479 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1480 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1481 tp->lost_cnt_hint -= decr;
1483 tcp_verify_left_out(tp);
1486 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1488 return TCP_SKB_CB(skb)->txstamp_ack ||
1489 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1492 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1494 struct skb_shared_info *shinfo = skb_shinfo(skb);
1496 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1497 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1498 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1499 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1501 shinfo->tx_flags &= ~tsflags;
1502 shinfo2->tx_flags |= tsflags;
1503 swap(shinfo->tskey, shinfo2->tskey);
1504 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1505 TCP_SKB_CB(skb)->txstamp_ack = 0;
1509 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1511 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1512 TCP_SKB_CB(skb)->eor = 0;
1515 /* Insert buff after skb on the write or rtx queue of sk. */
1516 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1517 struct sk_buff *buff,
1519 enum tcp_queue tcp_queue)
1521 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1522 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1524 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1527 /* Function to create two new TCP segments. Shrinks the given segment
1528 * to the specified size and appends a new segment with the rest of the
1529 * packet to the list. This won't be called frequently, I hope.
1530 * Remember, these are still headerless SKBs at this point.
1532 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1533 struct sk_buff *skb, u32 len,
1534 unsigned int mss_now, gfp_t gfp)
1536 struct tcp_sock *tp = tcp_sk(sk);
1537 struct sk_buff *buff;
1538 int nsize, old_factor;
1543 if (WARN_ON(len > skb->len))
1546 nsize = skb_headlen(skb) - len;
1550 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1551 * We need some allowance to not penalize applications setting small
1553 * Also allow first and last skb in retransmit queue to be split.
1555 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1556 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1557 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1558 skb != tcp_rtx_queue_head(sk) &&
1559 skb != tcp_rtx_queue_tail(sk))) {
1560 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1564 if (skb_unclone(skb, gfp))
1567 /* Get a new skb... force flag on. */
1568 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1570 return -ENOMEM; /* We'll just try again later. */
1571 skb_copy_decrypted(buff, skb);
1573 sk_wmem_queued_add(sk, buff->truesize);
1574 sk_mem_charge(sk, buff->truesize);
1575 nlen = skb->len - len - nsize;
1576 buff->truesize += nlen;
1577 skb->truesize -= nlen;
1579 /* Correct the sequence numbers. */
1580 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1581 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1582 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1584 /* PSH and FIN should only be set in the second packet. */
1585 flags = TCP_SKB_CB(skb)->tcp_flags;
1586 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1587 TCP_SKB_CB(buff)->tcp_flags = flags;
1588 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1589 tcp_skb_fragment_eor(skb, buff);
1591 skb_split(skb, buff, len);
1593 buff->ip_summed = CHECKSUM_PARTIAL;
1595 buff->tstamp = skb->tstamp;
1596 tcp_fragment_tstamp(skb, buff);
1598 old_factor = tcp_skb_pcount(skb);
1600 /* Fix up tso_factor for both original and new SKB. */
1601 tcp_set_skb_tso_segs(skb, mss_now);
1602 tcp_set_skb_tso_segs(buff, mss_now);
1604 /* Update delivered info for the new segment */
1605 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1607 /* If this packet has been sent out already, we must
1608 * adjust the various packet counters.
1610 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1611 int diff = old_factor - tcp_skb_pcount(skb) -
1612 tcp_skb_pcount(buff);
1615 tcp_adjust_pcount(sk, skb, diff);
1618 /* Link BUFF into the send queue. */
1619 __skb_header_release(buff);
1620 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1621 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1622 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1627 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1628 * data is not copied, but immediately discarded.
1630 static int __pskb_trim_head(struct sk_buff *skb, int len)
1632 struct skb_shared_info *shinfo;
1635 eat = min_t(int, len, skb_headlen(skb));
1637 __skb_pull(skb, eat);
1644 shinfo = skb_shinfo(skb);
1645 for (i = 0; i < shinfo->nr_frags; i++) {
1646 int size = skb_frag_size(&shinfo->frags[i]);
1649 skb_frag_unref(skb, i);
1652 shinfo->frags[k] = shinfo->frags[i];
1654 skb_frag_off_add(&shinfo->frags[k], eat);
1655 skb_frag_size_sub(&shinfo->frags[k], eat);
1661 shinfo->nr_frags = k;
1663 skb->data_len -= len;
1664 skb->len = skb->data_len;
1668 /* Remove acked data from a packet in the transmit queue. */
1669 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1673 if (skb_unclone(skb, GFP_ATOMIC))
1676 delta_truesize = __pskb_trim_head(skb, len);
1678 TCP_SKB_CB(skb)->seq += len;
1679 skb->ip_summed = CHECKSUM_PARTIAL;
1681 if (delta_truesize) {
1682 skb->truesize -= delta_truesize;
1683 sk_wmem_queued_add(sk, -delta_truesize);
1684 sk_mem_uncharge(sk, delta_truesize);
1687 /* Any change of skb->len requires recalculation of tso factor. */
1688 if (tcp_skb_pcount(skb) > 1)
1689 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1694 /* Calculate MSS not accounting any TCP options. */
1695 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1697 const struct tcp_sock *tp = tcp_sk(sk);
1698 const struct inet_connection_sock *icsk = inet_csk(sk);
1701 /* Calculate base mss without TCP options:
1702 It is MMS_S - sizeof(tcphdr) of rfc1122
1704 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1706 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1707 if (icsk->icsk_af_ops->net_frag_header_len) {
1708 const struct dst_entry *dst = __sk_dst_get(sk);
1710 if (dst && dst_allfrag(dst))
1711 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1714 /* Clamp it (mss_clamp does not include tcp options) */
1715 if (mss_now > tp->rx_opt.mss_clamp)
1716 mss_now = tp->rx_opt.mss_clamp;
1718 /* Now subtract optional transport overhead */
1719 mss_now -= icsk->icsk_ext_hdr_len;
1721 /* Then reserve room for full set of TCP options and 8 bytes of data */
1722 mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1726 /* Calculate MSS. Not accounting for SACKs here. */
1727 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1729 /* Subtract TCP options size, not including SACKs */
1730 return __tcp_mtu_to_mss(sk, pmtu) -
1731 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1734 /* Inverse of above */
1735 int tcp_mss_to_mtu(struct sock *sk, int mss)
1737 const struct tcp_sock *tp = tcp_sk(sk);
1738 const struct inet_connection_sock *icsk = inet_csk(sk);
1742 tp->tcp_header_len +
1743 icsk->icsk_ext_hdr_len +
1744 icsk->icsk_af_ops->net_header_len;
1746 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1747 if (icsk->icsk_af_ops->net_frag_header_len) {
1748 const struct dst_entry *dst = __sk_dst_get(sk);
1750 if (dst && dst_allfrag(dst))
1751 mtu += icsk->icsk_af_ops->net_frag_header_len;
1755 EXPORT_SYMBOL(tcp_mss_to_mtu);
1757 /* MTU probing init per socket */
1758 void tcp_mtup_init(struct sock *sk)
1760 struct tcp_sock *tp = tcp_sk(sk);
1761 struct inet_connection_sock *icsk = inet_csk(sk);
1762 struct net *net = sock_net(sk);
1764 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1765 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1766 icsk->icsk_af_ops->net_header_len;
1767 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1768 icsk->icsk_mtup.probe_size = 0;
1769 if (icsk->icsk_mtup.enabled)
1770 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1772 EXPORT_SYMBOL(tcp_mtup_init);
1774 /* This function synchronize snd mss to current pmtu/exthdr set.
1776 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1777 for TCP options, but includes only bare TCP header.
1779 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1780 It is minimum of user_mss and mss received with SYN.
1781 It also does not include TCP options.
1783 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1785 tp->mss_cache is current effective sending mss, including
1786 all tcp options except for SACKs. It is evaluated,
1787 taking into account current pmtu, but never exceeds
1788 tp->rx_opt.mss_clamp.
1790 NOTE1. rfc1122 clearly states that advertised MSS
1791 DOES NOT include either tcp or ip options.
1793 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1794 are READ ONLY outside this function. --ANK (980731)
1796 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1798 struct tcp_sock *tp = tcp_sk(sk);
1799 struct inet_connection_sock *icsk = inet_csk(sk);
1802 if (icsk->icsk_mtup.search_high > pmtu)
1803 icsk->icsk_mtup.search_high = pmtu;
1805 mss_now = tcp_mtu_to_mss(sk, pmtu);
1806 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1808 /* And store cached results */
1809 icsk->icsk_pmtu_cookie = pmtu;
1810 if (icsk->icsk_mtup.enabled)
1811 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1812 tp->mss_cache = mss_now;
1816 EXPORT_SYMBOL(tcp_sync_mss);
1818 /* Compute the current effective MSS, taking SACKs and IP options,
1819 * and even PMTU discovery events into account.
1821 unsigned int tcp_current_mss(struct sock *sk)
1823 const struct tcp_sock *tp = tcp_sk(sk);
1824 const struct dst_entry *dst = __sk_dst_get(sk);
1826 unsigned int header_len;
1827 struct tcp_out_options opts;
1828 struct tcp_md5sig_key *md5;
1830 mss_now = tp->mss_cache;
1833 u32 mtu = dst_mtu(dst);
1834 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1835 mss_now = tcp_sync_mss(sk, mtu);
1838 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1839 sizeof(struct tcphdr);
1840 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1841 * some common options. If this is an odd packet (because we have SACK
1842 * blocks etc) then our calculated header_len will be different, and
1843 * we have to adjust mss_now correspondingly */
1844 if (header_len != tp->tcp_header_len) {
1845 int delta = (int) header_len - tp->tcp_header_len;
1852 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1853 * As additional protections, we do not touch cwnd in retransmission phases,
1854 * and if application hit its sndbuf limit recently.
1856 static void tcp_cwnd_application_limited(struct sock *sk)
1858 struct tcp_sock *tp = tcp_sk(sk);
1860 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1861 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1862 /* Limited by application or receiver window. */
1863 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1864 u32 win_used = max(tp->snd_cwnd_used, init_win);
1865 if (win_used < tp->snd_cwnd) {
1866 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1867 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1869 tp->snd_cwnd_used = 0;
1871 tp->snd_cwnd_stamp = tcp_jiffies32;
1874 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1876 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1877 struct tcp_sock *tp = tcp_sk(sk);
1879 /* Track the maximum number of outstanding packets in each
1880 * window, and remember whether we were cwnd-limited then.
1882 if (!before(tp->snd_una, tp->max_packets_seq) ||
1883 tp->packets_out > tp->max_packets_out) {
1884 tp->max_packets_out = tp->packets_out;
1885 tp->max_packets_seq = tp->snd_nxt;
1886 tp->is_cwnd_limited = is_cwnd_limited;
1889 if (tcp_is_cwnd_limited(sk)) {
1890 /* Network is feed fully. */
1891 tp->snd_cwnd_used = 0;
1892 tp->snd_cwnd_stamp = tcp_jiffies32;
1894 /* Network starves. */
1895 if (tp->packets_out > tp->snd_cwnd_used)
1896 tp->snd_cwnd_used = tp->packets_out;
1898 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1899 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1900 !ca_ops->cong_control)
1901 tcp_cwnd_application_limited(sk);
1903 /* The following conditions together indicate the starvation
1904 * is caused by insufficient sender buffer:
1905 * 1) just sent some data (see tcp_write_xmit)
1906 * 2) not cwnd limited (this else condition)
1907 * 3) no more data to send (tcp_write_queue_empty())
1908 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1910 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1911 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1912 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1913 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1917 /* Minshall's variant of the Nagle send check. */
1918 static bool tcp_minshall_check(const struct tcp_sock *tp)
1920 return after(tp->snd_sml, tp->snd_una) &&
1921 !after(tp->snd_sml, tp->snd_nxt);
1924 /* Update snd_sml if this skb is under mss
1925 * Note that a TSO packet might end with a sub-mss segment
1926 * The test is really :
1927 * if ((skb->len % mss) != 0)
1928 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1929 * But we can avoid doing the divide again given we already have
1930 * skb_pcount = skb->len / mss_now
1932 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1933 const struct sk_buff *skb)
1935 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1936 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1939 /* Return false, if packet can be sent now without violation Nagle's rules:
1940 * 1. It is full sized. (provided by caller in %partial bool)
1941 * 2. Or it contains FIN. (already checked by caller)
1942 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1943 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1944 * With Minshall's modification: all sent small packets are ACKed.
1946 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1950 ((nonagle & TCP_NAGLE_CORK) ||
1951 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1954 /* Return how many segs we'd like on a TSO packet,
1955 * to send one TSO packet per ms
1957 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1962 bytes = min_t(unsigned long,
1963 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift),
1964 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1966 /* Goal is to send at least one packet per ms,
1967 * not one big TSO packet every 100 ms.
1968 * This preserves ACK clocking and is consistent
1969 * with tcp_tso_should_defer() heuristic.
1971 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1976 /* Return the number of segments we want in the skb we are transmitting.
1977 * See if congestion control module wants to decide; otherwise, autosize.
1979 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1981 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1982 u32 min_tso, tso_segs;
1984 min_tso = ca_ops->min_tso_segs ?
1985 ca_ops->min_tso_segs(sk) :
1986 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1988 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1989 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1992 /* Returns the portion of skb which can be sent right away */
1993 static unsigned int tcp_mss_split_point(const struct sock *sk,
1994 const struct sk_buff *skb,
1995 unsigned int mss_now,
1996 unsigned int max_segs,
1999 const struct tcp_sock *tp = tcp_sk(sk);
2000 u32 partial, needed, window, max_len;
2002 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2003 max_len = mss_now * max_segs;
2005 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2008 needed = min(skb->len, window);
2010 if (max_len <= needed)
2013 partial = needed % mss_now;
2014 /* If last segment is not a full MSS, check if Nagle rules allow us
2015 * to include this last segment in this skb.
2016 * Otherwise, we'll split the skb at last MSS boundary
2018 if (tcp_nagle_check(partial != 0, tp, nonagle))
2019 return needed - partial;
2024 /* Can at least one segment of SKB be sent right now, according to the
2025 * congestion window rules? If so, return how many segments are allowed.
2027 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
2028 const struct sk_buff *skb)
2030 u32 in_flight, cwnd, halfcwnd;
2032 /* Don't be strict about the congestion window for the final FIN. */
2033 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2034 tcp_skb_pcount(skb) == 1)
2037 in_flight = tcp_packets_in_flight(tp);
2038 cwnd = tp->snd_cwnd;
2039 if (in_flight >= cwnd)
2042 /* For better scheduling, ensure we have at least
2043 * 2 GSO packets in flight.
2045 halfcwnd = max(cwnd >> 1, 1U);
2046 return min(halfcwnd, cwnd - in_flight);
2049 /* Initialize TSO state of a skb.
2050 * This must be invoked the first time we consider transmitting
2051 * SKB onto the wire.
2053 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
2055 int tso_segs = tcp_skb_pcount(skb);
2057 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
2058 tcp_set_skb_tso_segs(skb, mss_now);
2059 tso_segs = tcp_skb_pcount(skb);
2065 /* Return true if the Nagle test allows this packet to be
2068 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2069 unsigned int cur_mss, int nonagle)
2071 /* Nagle rule does not apply to frames, which sit in the middle of the
2072 * write_queue (they have no chances to get new data).
2074 * This is implemented in the callers, where they modify the 'nonagle'
2075 * argument based upon the location of SKB in the send queue.
2077 if (nonagle & TCP_NAGLE_PUSH)
2080 /* Don't use the nagle rule for urgent data (or for the final FIN). */
2081 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2084 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
2090 /* Does at least the first segment of SKB fit into the send window? */
2091 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2092 const struct sk_buff *skb,
2093 unsigned int cur_mss)
2095 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2097 if (skb->len > cur_mss)
2098 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2100 return !after(end_seq, tcp_wnd_end(tp));
2103 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2104 * which is put after SKB on the list. It is very much like
2105 * tcp_fragment() except that it may make several kinds of assumptions
2106 * in order to speed up the splitting operation. In particular, we
2107 * know that all the data is in scatter-gather pages, and that the
2108 * packet has never been sent out before (and thus is not cloned).
2110 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2111 unsigned int mss_now, gfp_t gfp)
2113 int nlen = skb->len - len;
2114 struct sk_buff *buff;
2117 /* All of a TSO frame must be composed of paged data. */
2118 if (skb->len != skb->data_len)
2119 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2120 skb, len, mss_now, gfp);
2122 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
2123 if (unlikely(!buff))
2125 skb_copy_decrypted(buff, skb);
2127 sk_wmem_queued_add(sk, buff->truesize);
2128 sk_mem_charge(sk, buff->truesize);
2129 buff->truesize += nlen;
2130 skb->truesize -= nlen;
2132 /* Correct the sequence numbers. */
2133 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2134 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2135 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2137 /* PSH and FIN should only be set in the second packet. */
2138 flags = TCP_SKB_CB(skb)->tcp_flags;
2139 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2140 TCP_SKB_CB(buff)->tcp_flags = flags;
2142 /* This packet was never sent out yet, so no SACK bits. */
2143 TCP_SKB_CB(buff)->sacked = 0;
2145 tcp_skb_fragment_eor(skb, buff);
2147 buff->ip_summed = CHECKSUM_PARTIAL;
2148 skb_split(skb, buff, len);
2149 tcp_fragment_tstamp(skb, buff);
2151 /* Fix up tso_factor for both original and new SKB. */
2152 tcp_set_skb_tso_segs(skb, mss_now);
2153 tcp_set_skb_tso_segs(buff, mss_now);
2155 /* Link BUFF into the send queue. */
2156 __skb_header_release(buff);
2157 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2162 /* Try to defer sending, if possible, in order to minimize the amount
2163 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2165 * This algorithm is from John Heffner.
2167 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2168 bool *is_cwnd_limited,
2169 bool *is_rwnd_limited,
2172 const struct inet_connection_sock *icsk = inet_csk(sk);
2173 u32 send_win, cong_win, limit, in_flight;
2174 struct tcp_sock *tp = tcp_sk(sk);
2175 struct sk_buff *head;
2179 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2182 /* Avoid bursty behavior by allowing defer
2183 * only if the last write was recent (1 ms).
2184 * Note that tp->tcp_wstamp_ns can be in the future if we have
2185 * packets waiting in a qdisc or device for EDT delivery.
2187 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2191 in_flight = tcp_packets_in_flight(tp);
2193 BUG_ON(tcp_skb_pcount(skb) <= 1);
2194 BUG_ON(tp->snd_cwnd <= in_flight);
2196 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2198 /* From in_flight test above, we know that cwnd > in_flight. */
2199 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
2201 limit = min(send_win, cong_win);
2203 /* If a full-sized TSO skb can be sent, do it. */
2204 if (limit >= max_segs * tp->mss_cache)
2207 /* Middle in queue won't get any more data, full sendable already? */
2208 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2211 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2213 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
2215 /* If at least some fraction of a window is available,
2218 chunk /= win_divisor;
2222 /* Different approach, try not to defer past a single
2223 * ACK. Receiver should ACK every other full sized
2224 * frame, so if we have space for more than 3 frames
2227 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2231 /* TODO : use tsorted_sent_queue ? */
2232 head = tcp_rtx_queue_head(sk);
2235 delta = tp->tcp_clock_cache - head->tstamp;
2236 /* If next ACK is likely to come too late (half srtt), do not defer */
2237 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2240 /* Ok, it looks like it is advisable to defer.
2241 * Three cases are tracked :
2242 * 1) We are cwnd-limited
2243 * 2) We are rwnd-limited
2244 * 3) We are application limited.
2246 if (cong_win < send_win) {
2247 if (cong_win <= skb->len) {
2248 *is_cwnd_limited = true;
2252 if (send_win <= skb->len) {
2253 *is_rwnd_limited = true;
2258 /* If this packet won't get more data, do not wait. */
2259 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2260 TCP_SKB_CB(skb)->eor)
2269 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2271 struct inet_connection_sock *icsk = inet_csk(sk);
2272 struct tcp_sock *tp = tcp_sk(sk);
2273 struct net *net = sock_net(sk);
2277 interval = net->ipv4.sysctl_tcp_probe_interval;
2278 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2279 if (unlikely(delta >= interval * HZ)) {
2280 int mss = tcp_current_mss(sk);
2282 /* Update current search range */
2283 icsk->icsk_mtup.probe_size = 0;
2284 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2285 sizeof(struct tcphdr) +
2286 icsk->icsk_af_ops->net_header_len;
2287 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2289 /* Update probe time stamp */
2290 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2294 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2296 struct sk_buff *skb, *next;
2298 skb = tcp_send_head(sk);
2299 tcp_for_write_queue_from_safe(skb, next, sk) {
2300 if (len <= skb->len)
2303 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2312 /* Create a new MTU probe if we are ready.
2313 * MTU probe is regularly attempting to increase the path MTU by
2314 * deliberately sending larger packets. This discovers routing
2315 * changes resulting in larger path MTUs.
2317 * Returns 0 if we should wait to probe (no cwnd available),
2318 * 1 if a probe was sent,
2321 static int tcp_mtu_probe(struct sock *sk)
2323 struct inet_connection_sock *icsk = inet_csk(sk);
2324 struct tcp_sock *tp = tcp_sk(sk);
2325 struct sk_buff *skb, *nskb, *next;
2326 struct net *net = sock_net(sk);
2333 /* Not currently probing/verifying,
2335 * have enough cwnd, and
2336 * not SACKing (the variable headers throw things off)
2338 if (likely(!icsk->icsk_mtup.enabled ||
2339 icsk->icsk_mtup.probe_size ||
2340 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2341 tp->snd_cwnd < 11 ||
2342 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2345 /* Use binary search for probe_size between tcp_mss_base,
2346 * and current mss_clamp. if (search_high - search_low)
2347 * smaller than a threshold, backoff from probing.
2349 mss_now = tcp_current_mss(sk);
2350 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2351 icsk->icsk_mtup.search_low) >> 1);
2352 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2353 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2354 /* When misfortune happens, we are reprobing actively,
2355 * and then reprobe timer has expired. We stick with current
2356 * probing process by not resetting search range to its orignal.
2358 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2359 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2360 /* Check whether enough time has elaplased for
2361 * another round of probing.
2363 tcp_mtu_check_reprobe(sk);
2367 /* Have enough data in the send queue to probe? */
2368 if (tp->write_seq - tp->snd_nxt < size_needed)
2371 if (tp->snd_wnd < size_needed)
2373 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2376 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2377 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2378 if (!tcp_packets_in_flight(tp))
2384 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2387 /* We're allowed to probe. Build it now. */
2388 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2391 sk_wmem_queued_add(sk, nskb->truesize);
2392 sk_mem_charge(sk, nskb->truesize);
2394 skb = tcp_send_head(sk);
2395 skb_copy_decrypted(nskb, skb);
2397 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2398 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2399 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2400 TCP_SKB_CB(nskb)->sacked = 0;
2402 nskb->ip_summed = CHECKSUM_PARTIAL;
2404 tcp_insert_write_queue_before(nskb, skb, sk);
2405 tcp_highest_sack_replace(sk, skb, nskb);
2408 tcp_for_write_queue_from_safe(skb, next, sk) {
2409 copy = min_t(int, skb->len, probe_size - len);
2410 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2412 if (skb->len <= copy) {
2413 /* We've eaten all the data from this skb.
2415 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2416 /* If this is the last SKB we copy and eor is set
2417 * we need to propagate it to the new skb.
2419 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2420 tcp_skb_collapse_tstamp(nskb, skb);
2421 tcp_unlink_write_queue(skb, sk);
2422 sk_wmem_free_skb(sk, skb);
2424 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2425 ~(TCPHDR_FIN|TCPHDR_PSH);
2426 if (!skb_shinfo(skb)->nr_frags) {
2427 skb_pull(skb, copy);
2429 __pskb_trim_head(skb, copy);
2430 tcp_set_skb_tso_segs(skb, mss_now);
2432 TCP_SKB_CB(skb)->seq += copy;
2437 if (len >= probe_size)
2440 tcp_init_tso_segs(nskb, nskb->len);
2442 /* We're ready to send. If this fails, the probe will
2443 * be resegmented into mss-sized pieces by tcp_write_xmit().
2445 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2446 /* Decrement cwnd here because we are sending
2447 * effectively two packets. */
2449 tcp_event_new_data_sent(sk, nskb);
2451 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2452 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2453 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2461 static bool tcp_pacing_check(struct sock *sk)
2463 struct tcp_sock *tp = tcp_sk(sk);
2465 if (!tcp_needs_internal_pacing(sk))
2468 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2471 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2472 hrtimer_start(&tp->pacing_timer,
2473 ns_to_ktime(tp->tcp_wstamp_ns),
2474 HRTIMER_MODE_ABS_PINNED_SOFT);
2480 /* TCP Small Queues :
2481 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2482 * (These limits are doubled for retransmits)
2484 * - better RTT estimation and ACK scheduling
2487 * Alas, some drivers / subsystems require a fair amount
2488 * of queued bytes to ensure line rate.
2489 * One example is wifi aggregation (802.11 AMPDU)
2491 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2492 unsigned int factor)
2494 unsigned long limit;
2496 limit = max_t(unsigned long,
2498 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2499 if (sk->sk_pacing_status == SK_PACING_NONE)
2500 limit = min_t(unsigned long, limit,
2501 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2504 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2505 tcp_sk(sk)->tcp_tx_delay) {
2506 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2508 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2509 * approximate our needs assuming an ~100% skb->truesize overhead.
2510 * USEC_PER_SEC is approximated by 2^20.
2511 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2513 extra_bytes >>= (20 - 1);
2514 limit += extra_bytes;
2516 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2517 /* Always send skb if rtx queue is empty.
2518 * No need to wait for TX completion to call us back,
2519 * after softirq/tasklet schedule.
2520 * This helps when TX completions are delayed too much.
2522 if (tcp_rtx_queue_empty(sk))
2525 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2526 /* It is possible TX completion already happened
2527 * before we set TSQ_THROTTLED, so we must
2528 * test again the condition.
2530 smp_mb__after_atomic();
2531 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2537 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2539 const u32 now = tcp_jiffies32;
2540 enum tcp_chrono old = tp->chrono_type;
2542 if (old > TCP_CHRONO_UNSPEC)
2543 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2544 tp->chrono_start = now;
2545 tp->chrono_type = new;
2548 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2550 struct tcp_sock *tp = tcp_sk(sk);
2552 /* If there are multiple conditions worthy of tracking in a
2553 * chronograph then the highest priority enum takes precedence
2554 * over the other conditions. So that if something "more interesting"
2555 * starts happening, stop the previous chrono and start a new one.
2557 if (type > tp->chrono_type)
2558 tcp_chrono_set(tp, type);
2561 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2563 struct tcp_sock *tp = tcp_sk(sk);
2566 /* There are multiple conditions worthy of tracking in a
2567 * chronograph, so that the highest priority enum takes
2568 * precedence over the other conditions (see tcp_chrono_start).
2569 * If a condition stops, we only stop chrono tracking if
2570 * it's the "most interesting" or current chrono we are
2571 * tracking and starts busy chrono if we have pending data.
2573 if (tcp_rtx_and_write_queues_empty(sk))
2574 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2575 else if (type == tp->chrono_type)
2576 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2579 /* This routine writes packets to the network. It advances the
2580 * send_head. This happens as incoming acks open up the remote
2583 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2584 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2585 * account rare use of URG, this is not a big flaw.
2587 * Send at most one packet when push_one > 0. Temporarily ignore
2588 * cwnd limit to force at most one packet out when push_one == 2.
2590 * Returns true, if no segments are in flight and we have queued segments,
2591 * but cannot send anything now because of SWS or another problem.
2593 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2594 int push_one, gfp_t gfp)
2596 struct tcp_sock *tp = tcp_sk(sk);
2597 struct sk_buff *skb;
2598 unsigned int tso_segs, sent_pkts;
2601 bool is_cwnd_limited = false, is_rwnd_limited = false;
2606 tcp_mstamp_refresh(tp);
2608 /* Do MTU probing. */
2609 result = tcp_mtu_probe(sk);
2612 } else if (result > 0) {
2617 max_segs = tcp_tso_segs(sk, mss_now);
2618 while ((skb = tcp_send_head(sk))) {
2621 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2622 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2623 skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2624 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2625 tcp_init_tso_segs(skb, mss_now);
2626 goto repair; /* Skip network transmission */
2629 if (tcp_pacing_check(sk))
2632 tso_segs = tcp_init_tso_segs(skb, mss_now);
2635 cwnd_quota = tcp_cwnd_test(tp, skb);
2638 /* Force out a loss probe pkt. */
2644 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2645 is_rwnd_limited = true;
2649 if (tso_segs == 1) {
2650 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2651 (tcp_skb_is_last(sk, skb) ?
2652 nonagle : TCP_NAGLE_PUSH))))
2656 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2657 &is_rwnd_limited, max_segs))
2662 if (tso_segs > 1 && !tcp_urg_mode(tp))
2663 limit = tcp_mss_split_point(sk, skb, mss_now,
2669 if (skb->len > limit &&
2670 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2673 if (tcp_small_queue_check(sk, skb, 0))
2676 /* Argh, we hit an empty skb(), presumably a thread
2677 * is sleeping in sendmsg()/sk_stream_wait_memory().
2678 * We do not want to send a pure-ack packet and have
2679 * a strange looking rtx queue with empty packet(s).
2681 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2684 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2688 /* Advance the send_head. This one is sent out.
2689 * This call will increment packets_out.
2691 tcp_event_new_data_sent(sk, skb);
2693 tcp_minshall_update(tp, mss_now, skb);
2694 sent_pkts += tcp_skb_pcount(skb);
2700 if (is_rwnd_limited)
2701 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2703 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2705 if (likely(sent_pkts)) {
2706 if (tcp_in_cwnd_reduction(sk))
2707 tp->prr_out += sent_pkts;
2709 /* Send one loss probe per tail loss episode. */
2711 tcp_schedule_loss_probe(sk, false);
2712 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2713 tcp_cwnd_validate(sk, is_cwnd_limited);
2716 return !tp->packets_out && !tcp_write_queue_empty(sk);
2719 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2721 struct inet_connection_sock *icsk = inet_csk(sk);
2722 struct tcp_sock *tp = tcp_sk(sk);
2723 u32 timeout, rto_delta_us;
2726 /* Don't do any loss probe on a Fast Open connection before 3WHS
2729 if (rcu_access_pointer(tp->fastopen_rsk))
2732 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2733 /* Schedule a loss probe in 2*RTT for SACK capable connections
2734 * not in loss recovery, that are either limited by cwnd or application.
2736 if ((early_retrans != 3 && early_retrans != 4) ||
2737 !tp->packets_out || !tcp_is_sack(tp) ||
2738 (icsk->icsk_ca_state != TCP_CA_Open &&
2739 icsk->icsk_ca_state != TCP_CA_CWR))
2742 /* Probe timeout is 2*rtt. Add minimum RTO to account
2743 * for delayed ack when there's one outstanding packet. If no RTT
2744 * sample is available then probe after TCP_TIMEOUT_INIT.
2747 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2748 if (tp->packets_out == 1)
2749 timeout += TCP_RTO_MIN;
2751 timeout += TCP_TIMEOUT_MIN;
2753 timeout = TCP_TIMEOUT_INIT;
2756 /* If the RTO formula yields an earlier time, then use that time. */
2757 rto_delta_us = advancing_rto ?
2758 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2759 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2760 if (rto_delta_us > 0)
2761 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2763 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2767 /* Thanks to skb fast clones, we can detect if a prior transmit of
2768 * a packet is still in a qdisc or driver queue.
2769 * In this case, there is very little point doing a retransmit !
2771 static bool skb_still_in_host_queue(const struct sock *sk,
2772 const struct sk_buff *skb)
2774 if (unlikely(skb_fclone_busy(sk, skb))) {
2775 NET_INC_STATS(sock_net(sk),
2776 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2782 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2783 * retransmit the last segment.
2785 void tcp_send_loss_probe(struct sock *sk)
2787 struct tcp_sock *tp = tcp_sk(sk);
2788 struct sk_buff *skb;
2790 int mss = tcp_current_mss(sk);
2792 /* At most one outstanding TLP */
2793 if (tp->tlp_high_seq)
2796 tp->tlp_retrans = 0;
2797 skb = tcp_send_head(sk);
2798 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2799 pcount = tp->packets_out;
2800 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2801 if (tp->packets_out > pcount)
2805 skb = skb_rb_last(&sk->tcp_rtx_queue);
2806 if (unlikely(!skb)) {
2807 WARN_ONCE(tp->packets_out,
2808 "invalid inflight: %u state %u cwnd %u mss %d\n",
2809 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2810 inet_csk(sk)->icsk_pending = 0;
2814 if (skb_still_in_host_queue(sk, skb))
2817 pcount = tcp_skb_pcount(skb);
2818 if (WARN_ON(!pcount))
2821 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2822 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2823 (pcount - 1) * mss, mss,
2826 skb = skb_rb_next(skb);
2829 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2832 if (__tcp_retransmit_skb(sk, skb, 1))
2835 tp->tlp_retrans = 1;
2838 /* Record snd_nxt for loss detection. */
2839 tp->tlp_high_seq = tp->snd_nxt;
2841 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2842 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2843 inet_csk(sk)->icsk_pending = 0;
2848 /* Push out any pending frames which were held back due to
2849 * TCP_CORK or attempt at coalescing tiny packets.
2850 * The socket must be locked by the caller.
2852 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2855 /* If we are closed, the bytes will have to remain here.
2856 * In time closedown will finish, we empty the write queue and
2857 * all will be happy.
2859 if (unlikely(sk->sk_state == TCP_CLOSE))
2862 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2863 sk_gfp_mask(sk, GFP_ATOMIC)))
2864 tcp_check_probe_timer(sk);
2867 /* Send _single_ skb sitting at the send head. This function requires
2868 * true push pending frames to setup probe timer etc.
2870 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2872 struct sk_buff *skb = tcp_send_head(sk);
2874 BUG_ON(!skb || skb->len < mss_now);
2876 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2879 /* This function returns the amount that we can raise the
2880 * usable window based on the following constraints
2882 * 1. The window can never be shrunk once it is offered (RFC 793)
2883 * 2. We limit memory per socket
2886 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2887 * RECV.NEXT + RCV.WIN fixed until:
2888 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2890 * i.e. don't raise the right edge of the window until you can raise
2891 * it at least MSS bytes.
2893 * Unfortunately, the recommended algorithm breaks header prediction,
2894 * since header prediction assumes th->window stays fixed.
2896 * Strictly speaking, keeping th->window fixed violates the receiver
2897 * side SWS prevention criteria. The problem is that under this rule
2898 * a stream of single byte packets will cause the right side of the
2899 * window to always advance by a single byte.
2901 * Of course, if the sender implements sender side SWS prevention
2902 * then this will not be a problem.
2904 * BSD seems to make the following compromise:
2906 * If the free space is less than the 1/4 of the maximum
2907 * space available and the free space is less than 1/2 mss,
2908 * then set the window to 0.
2909 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2910 * Otherwise, just prevent the window from shrinking
2911 * and from being larger than the largest representable value.
2913 * This prevents incremental opening of the window in the regime
2914 * where TCP is limited by the speed of the reader side taking
2915 * data out of the TCP receive queue. It does nothing about
2916 * those cases where the window is constrained on the sender side
2917 * because the pipeline is full.
2919 * BSD also seems to "accidentally" limit itself to windows that are a
2920 * multiple of MSS, at least until the free space gets quite small.
2921 * This would appear to be a side effect of the mbuf implementation.
2922 * Combining these two algorithms results in the observed behavior
2923 * of having a fixed window size at almost all times.
2925 * Below we obtain similar behavior by forcing the offered window to
2926 * a multiple of the mss when it is feasible to do so.
2928 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2929 * Regular options like TIMESTAMP are taken into account.
2931 u32 __tcp_select_window(struct sock *sk)
2933 struct inet_connection_sock *icsk = inet_csk(sk);
2934 struct tcp_sock *tp = tcp_sk(sk);
2935 /* MSS for the peer's data. Previous versions used mss_clamp
2936 * here. I don't know if the value based on our guesses
2937 * of peer's MSS is better for the performance. It's more correct
2938 * but may be worse for the performance because of rcv_mss
2939 * fluctuations. --SAW 1998/11/1
2941 int mss = icsk->icsk_ack.rcv_mss;
2942 int free_space = tcp_space(sk);
2943 int allowed_space = tcp_full_space(sk);
2944 int full_space, window;
2946 if (sk_is_mptcp(sk))
2947 mptcp_space(sk, &free_space, &allowed_space);
2949 full_space = min_t(int, tp->window_clamp, allowed_space);
2951 if (unlikely(mss > full_space)) {
2956 if (free_space < (full_space >> 1)) {
2957 icsk->icsk_ack.quick = 0;
2959 if (tcp_under_memory_pressure(sk))
2960 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2963 /* free_space might become our new window, make sure we don't
2964 * increase it due to wscale.
2966 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2968 /* if free space is less than mss estimate, or is below 1/16th
2969 * of the maximum allowed, try to move to zero-window, else
2970 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2971 * new incoming data is dropped due to memory limits.
2972 * With large window, mss test triggers way too late in order
2973 * to announce zero window in time before rmem limit kicks in.
2975 if (free_space < (allowed_space >> 4) || free_space < mss)
2979 if (free_space > tp->rcv_ssthresh)
2980 free_space = tp->rcv_ssthresh;
2982 /* Don't do rounding if we are using window scaling, since the
2983 * scaled window will not line up with the MSS boundary anyway.
2985 if (tp->rx_opt.rcv_wscale) {
2986 window = free_space;
2988 /* Advertise enough space so that it won't get scaled away.
2989 * Import case: prevent zero window announcement if
2990 * 1<<rcv_wscale > mss.
2992 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2994 window = tp->rcv_wnd;
2995 /* Get the largest window that is a nice multiple of mss.
2996 * Window clamp already applied above.
2997 * If our current window offering is within 1 mss of the
2998 * free space we just keep it. This prevents the divide
2999 * and multiply from happening most of the time.
3000 * We also don't do any window rounding when the free space
3003 if (window <= free_space - mss || window > free_space)
3004 window = rounddown(free_space, mss);
3005 else if (mss == full_space &&
3006 free_space > window + (full_space >> 1))
3007 window = free_space;
3013 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3014 const struct sk_buff *next_skb)
3016 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3017 const struct skb_shared_info *next_shinfo =
3018 skb_shinfo(next_skb);
3019 struct skb_shared_info *shinfo = skb_shinfo(skb);
3021 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3022 shinfo->tskey = next_shinfo->tskey;
3023 TCP_SKB_CB(skb)->txstamp_ack |=
3024 TCP_SKB_CB(next_skb)->txstamp_ack;
3028 /* Collapses two adjacent SKB's during retransmission. */
3029 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3031 struct tcp_sock *tp = tcp_sk(sk);
3032 struct sk_buff *next_skb = skb_rb_next(skb);
3035 next_skb_size = next_skb->len;
3037 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3039 if (next_skb_size) {
3040 if (next_skb_size <= skb_availroom(skb))
3041 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
3043 else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3046 tcp_highest_sack_replace(sk, next_skb, skb);
3048 /* Update sequence range on original skb. */
3049 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3051 /* Merge over control information. This moves PSH/FIN etc. over */
3052 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3054 /* All done, get rid of second SKB and account for it so
3055 * packet counting does not break.
3057 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3058 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3060 /* changed transmit queue under us so clear hints */
3061 tcp_clear_retrans_hints_partial(tp);
3062 if (next_skb == tp->retransmit_skb_hint)
3063 tp->retransmit_skb_hint = skb;
3065 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3067 tcp_skb_collapse_tstamp(skb, next_skb);
3069 tcp_rtx_queue_unlink_and_free(next_skb, sk);
3073 /* Check if coalescing SKBs is legal. */
3074 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3076 if (tcp_skb_pcount(skb) > 1)
3078 if (skb_cloned(skb))
3080 /* Some heuristics for collapsing over SACK'd could be invented */
3081 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3087 /* Collapse packets in the retransmit queue to make to create
3088 * less packets on the wire. This is only done on retransmission.
3090 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3093 struct tcp_sock *tp = tcp_sk(sk);
3094 struct sk_buff *skb = to, *tmp;
3097 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
3099 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3102 skb_rbtree_walk_from_safe(skb, tmp) {
3103 if (!tcp_can_collapse(sk, skb))
3106 if (!tcp_skb_can_collapse(to, skb))
3119 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3122 if (!tcp_collapse_retrans(sk, to))
3127 /* This retransmits one SKB. Policy decisions and retransmit queue
3128 * state updates are done by the caller. Returns non-zero if an
3129 * error occurred which prevented the send.
3131 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3133 struct inet_connection_sock *icsk = inet_csk(sk);
3134 struct tcp_sock *tp = tcp_sk(sk);
3135 unsigned int cur_mss;
3139 /* Inconclusive MTU probe */
3140 if (icsk->icsk_mtup.probe_size)
3141 icsk->icsk_mtup.probe_size = 0;
3143 /* Do not sent more than we queued. 1/4 is reserved for possible
3144 * copying overhead: fragmentation, tunneling, mangling etc.
3146 if (refcount_read(&sk->sk_wmem_alloc) >
3147 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
3151 if (skb_still_in_host_queue(sk, skb))
3154 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
3155 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3159 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3163 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3164 return -EHOSTUNREACH; /* Routing failure or similar. */
3166 cur_mss = tcp_current_mss(sk);
3168 /* If receiver has shrunk his window, and skb is out of
3169 * new window, do not retransmit it. The exception is the
3170 * case, when window is shrunk to zero. In this case
3171 * our retransmit serves as a zero window probe.
3173 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
3174 TCP_SKB_CB(skb)->seq != tp->snd_una)
3177 len = cur_mss * segs;
3178 if (skb->len > len) {
3179 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3180 cur_mss, GFP_ATOMIC))
3181 return -ENOMEM; /* We'll try again later. */
3183 if (skb_unclone(skb, GFP_ATOMIC))
3186 diff = tcp_skb_pcount(skb);
3187 tcp_set_skb_tso_segs(skb, cur_mss);
3188 diff -= tcp_skb_pcount(skb);
3190 tcp_adjust_pcount(sk, skb, diff);
3191 if (skb->len < cur_mss)
3192 tcp_retrans_try_collapse(sk, skb, cur_mss);
3195 /* RFC3168, section 6.1.1.1. ECN fallback */
3196 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3197 tcp_ecn_clear_syn(sk, skb);
3199 /* Update global and local TCP statistics. */
3200 segs = tcp_skb_pcount(skb);
3201 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3202 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3203 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3204 tp->total_retrans += segs;
3205 tp->bytes_retrans += skb->len;
3207 /* make sure skb->data is aligned on arches that require it
3208 * and check if ack-trimming & collapsing extended the headroom
3209 * beyond what csum_start can cover.
3211 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3212 skb_headroom(skb) >= 0xFFFF)) {
3213 struct sk_buff *nskb;
3215 tcp_skb_tsorted_save(skb) {
3216 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3219 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3223 } tcp_skb_tsorted_restore(skb);
3226 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3227 tcp_rate_skb_sent(sk, skb);
3230 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3233 /* To avoid taking spuriously low RTT samples based on a timestamp
3234 * for a transmit that never happened, always mark EVER_RETRANS
3236 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3238 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3239 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3240 TCP_SKB_CB(skb)->seq, segs, err);
3243 trace_tcp_retransmit_skb(sk, skb);
3244 } else if (err != -EBUSY) {
3245 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3250 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3252 struct tcp_sock *tp = tcp_sk(sk);
3253 int err = __tcp_retransmit_skb(sk, skb, segs);
3256 #if FASTRETRANS_DEBUG > 0
3257 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3258 net_dbg_ratelimited("retrans_out leaked\n");
3261 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3262 tp->retrans_out += tcp_skb_pcount(skb);
3265 /* Save stamp of the first (attempted) retransmit. */
3266 if (!tp->retrans_stamp)
3267 tp->retrans_stamp = tcp_skb_timestamp(skb);
3269 if (tp->undo_retrans < 0)
3270 tp->undo_retrans = 0;
3271 tp->undo_retrans += tcp_skb_pcount(skb);
3275 /* This gets called after a retransmit timeout, and the initially
3276 * retransmitted data is acknowledged. It tries to continue
3277 * resending the rest of the retransmit queue, until either
3278 * we've sent it all or the congestion window limit is reached.
3280 void tcp_xmit_retransmit_queue(struct sock *sk)
3282 const struct inet_connection_sock *icsk = inet_csk(sk);
3283 struct sk_buff *skb, *rtx_head, *hole = NULL;
3284 struct tcp_sock *tp = tcp_sk(sk);
3285 bool rearm_timer = false;
3289 if (!tp->packets_out)
3292 rtx_head = tcp_rtx_queue_head(sk);
3293 skb = tp->retransmit_skb_hint ?: rtx_head;
3294 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3295 skb_rbtree_walk_from(skb) {
3299 if (tcp_pacing_check(sk))
3302 /* we could do better than to assign each time */
3304 tp->retransmit_skb_hint = skb;
3306 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
3309 sacked = TCP_SKB_CB(skb)->sacked;
3310 /* In case tcp_shift_skb_data() have aggregated large skbs,
3311 * we need to make sure not sending too bigs TSO packets
3313 segs = min_t(int, segs, max_segs);
3315 if (tp->retrans_out >= tp->lost_out) {
3317 } else if (!(sacked & TCPCB_LOST)) {
3318 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3323 if (icsk->icsk_ca_state != TCP_CA_Loss)
3324 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3326 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3329 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3332 if (tcp_small_queue_check(sk, skb, 1))
3335 if (tcp_retransmit_skb(sk, skb, segs))
3338 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3340 if (tcp_in_cwnd_reduction(sk))
3341 tp->prr_out += tcp_skb_pcount(skb);
3343 if (skb == rtx_head &&
3344 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3349 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3350 inet_csk(sk)->icsk_rto,
3354 /* We allow to exceed memory limits for FIN packets to expedite
3355 * connection tear down and (memory) recovery.
3356 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3357 * or even be forced to close flow without any FIN.
3358 * In general, we want to allow one skb per socket to avoid hangs
3359 * with edge trigger epoll()
3361 void sk_forced_mem_schedule(struct sock *sk, int size)
3365 if (size <= sk->sk_forward_alloc)
3367 amt = sk_mem_pages(size);
3368 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3369 sk_memory_allocated_add(sk, amt);
3371 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3372 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3375 /* Send a FIN. The caller locks the socket for us.
3376 * We should try to send a FIN packet really hard, but eventually give up.
3378 void tcp_send_fin(struct sock *sk)
3380 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3381 struct tcp_sock *tp = tcp_sk(sk);
3383 /* Optimization, tack on the FIN if we have one skb in write queue and
3384 * this skb was not yet sent, or we are under memory pressure.
3385 * Note: in the latter case, FIN packet will be sent after a timeout,
3386 * as TCP stack thinks it has already been transmitted.
3389 if (!tskb && tcp_under_memory_pressure(sk))
3390 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3393 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3394 TCP_SKB_CB(tskb)->end_seq++;
3397 /* This means tskb was already sent.
3398 * Pretend we included the FIN on previous transmit.
3399 * We need to set tp->snd_nxt to the value it would have
3400 * if FIN had been sent. This is because retransmit path
3401 * does not change tp->snd_nxt.
3403 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3407 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3411 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3412 skb_reserve(skb, MAX_TCP_HEADER);
3413 sk_forced_mem_schedule(sk, skb->truesize);
3414 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3415 tcp_init_nondata_skb(skb, tp->write_seq,
3416 TCPHDR_ACK | TCPHDR_FIN);
3417 tcp_queue_skb(sk, skb);
3419 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3422 /* We get here when a process closes a file descriptor (either due to
3423 * an explicit close() or as a byproduct of exit()'ing) and there
3424 * was unread data in the receive queue. This behavior is recommended
3425 * by RFC 2525, section 2.17. -DaveM
3427 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3429 struct sk_buff *skb;
3431 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3433 /* NOTE: No TCP options attached and we never retransmit this. */
3434 skb = alloc_skb(MAX_TCP_HEADER, priority);
3436 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3440 /* Reserve space for headers and prepare control bits. */
3441 skb_reserve(skb, MAX_TCP_HEADER);
3442 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3443 TCPHDR_ACK | TCPHDR_RST);
3444 tcp_mstamp_refresh(tcp_sk(sk));
3446 if (tcp_transmit_skb(sk, skb, 0, priority))
3447 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3449 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3450 * skb here is different to the troublesome skb, so use NULL
3452 trace_tcp_send_reset(sk, NULL);
3455 /* Send a crossed SYN-ACK during socket establishment.
3456 * WARNING: This routine must only be called when we have already sent
3457 * a SYN packet that crossed the incoming SYN that caused this routine
3458 * to get called. If this assumption fails then the initial rcv_wnd
3459 * and rcv_wscale values will not be correct.
3461 int tcp_send_synack(struct sock *sk)
3463 struct sk_buff *skb;
3465 skb = tcp_rtx_queue_head(sk);
3466 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3467 pr_err("%s: wrong queue state\n", __func__);
3470 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3471 if (skb_cloned(skb)) {
3472 struct sk_buff *nskb;
3474 tcp_skb_tsorted_save(skb) {
3475 nskb = skb_copy(skb, GFP_ATOMIC);
3476 } tcp_skb_tsorted_restore(skb);
3479 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3480 tcp_highest_sack_replace(sk, skb, nskb);
3481 tcp_rtx_queue_unlink_and_free(skb, sk);
3482 __skb_header_release(nskb);
3483 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3484 sk_wmem_queued_add(sk, nskb->truesize);
3485 sk_mem_charge(sk, nskb->truesize);
3489 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3490 tcp_ecn_send_synack(sk, skb);
3492 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3496 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3497 * @sk: listener socket
3498 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3499 * should not use it again.
3500 * @req: request_sock pointer
3501 * @foc: cookie for tcp fast open
3502 * @synack_type: Type of synack to prepare
3503 * @syn_skb: SYN packet just received. It could be NULL for rtx case.
3505 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3506 struct request_sock *req,
3507 struct tcp_fastopen_cookie *foc,
3508 enum tcp_synack_type synack_type,
3509 struct sk_buff *syn_skb)
3511 struct inet_request_sock *ireq = inet_rsk(req);
3512 const struct tcp_sock *tp = tcp_sk(sk);
3513 struct tcp_md5sig_key *md5 = NULL;
3514 struct tcp_out_options opts;
3515 struct sk_buff *skb;
3516 int tcp_header_size;
3521 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3522 if (unlikely(!skb)) {
3526 /* Reserve space for headers. */
3527 skb_reserve(skb, MAX_TCP_HEADER);
3529 switch (synack_type) {
3530 case TCP_SYNACK_NORMAL:
3531 skb_set_owner_w(skb, req_to_sk(req));
3533 case TCP_SYNACK_COOKIE:
3534 /* Under synflood, we do not attach skb to a socket,
3535 * to avoid false sharing.
3538 case TCP_SYNACK_FASTOPEN:
3539 /* sk is a const pointer, because we want to express multiple
3540 * cpu might call us concurrently.
3541 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3543 skb_set_owner_w(skb, (struct sock *)sk);
3546 skb_dst_set(skb, dst);
3548 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3550 memset(&opts, 0, sizeof(opts));
3551 now = tcp_clock_ns();
3552 #ifdef CONFIG_SYN_COOKIES
3553 if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3554 skb->skb_mstamp_ns = cookie_init_timestamp(req, now);
3558 skb->skb_mstamp_ns = now;
3559 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3560 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3563 #ifdef CONFIG_TCP_MD5SIG
3565 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3567 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3568 /* bpf program will be interested in the tcp_flags */
3569 TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3570 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3572 syn_skb) + sizeof(*th);
3574 skb_push(skb, tcp_header_size);
3575 skb_reset_transport_header(skb);
3577 th = (struct tcphdr *)skb->data;
3578 memset(th, 0, sizeof(struct tcphdr));
3581 tcp_ecn_make_synack(req, th);
3582 th->source = htons(ireq->ir_num);
3583 th->dest = ireq->ir_rmt_port;
3584 skb->mark = ireq->ir_mark;
3585 skb->ip_summed = CHECKSUM_PARTIAL;
3586 th->seq = htonl(tcp_rsk(req)->snt_isn);
3587 /* XXX data is queued and acked as is. No buffer/window check */
3588 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3590 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3591 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3592 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3593 th->doff = (tcp_header_size >> 2);
3594 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3596 #ifdef CONFIG_TCP_MD5SIG
3597 /* Okay, we have all we need - do the md5 hash if needed */
3599 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3600 md5, req_to_sk(req), skb);
3604 bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
3605 synack_type, &opts);
3607 skb->skb_mstamp_ns = now;
3608 tcp_add_tx_delay(skb, tp);
3612 EXPORT_SYMBOL(tcp_make_synack);
3614 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3616 struct inet_connection_sock *icsk = inet_csk(sk);
3617 const struct tcp_congestion_ops *ca;
3618 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3620 if (ca_key == TCP_CA_UNSPEC)
3624 ca = tcp_ca_find_key(ca_key);
3625 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3626 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3627 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3628 icsk->icsk_ca_ops = ca;
3633 /* Do all connect socket setups that can be done AF independent. */
3634 static void tcp_connect_init(struct sock *sk)
3636 const struct dst_entry *dst = __sk_dst_get(sk);
3637 struct tcp_sock *tp = tcp_sk(sk);
3641 /* We'll fix this up when we get a response from the other end.
3642 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3644 tp->tcp_header_len = sizeof(struct tcphdr);
3645 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3646 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3648 #ifdef CONFIG_TCP_MD5SIG
3649 if (tp->af_specific->md5_lookup(sk, sk))
3650 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3653 /* If user gave his TCP_MAXSEG, record it to clamp */
3654 if (tp->rx_opt.user_mss)
3655 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3658 tcp_sync_mss(sk, dst_mtu(dst));
3660 tcp_ca_dst_init(sk, dst);
3662 if (!tp->window_clamp)
3663 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3664 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3666 tcp_initialize_rcv_mss(sk);
3668 /* limit the window selection if the user enforce a smaller rx buffer */
3669 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3670 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3671 tp->window_clamp = tcp_full_space(sk);
3673 rcv_wnd = tcp_rwnd_init_bpf(sk);
3675 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3677 tcp_select_initial_window(sk, tcp_full_space(sk),
3678 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3681 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3685 tp->rx_opt.rcv_wscale = rcv_wscale;
3686 tp->rcv_ssthresh = tp->rcv_wnd;
3689 sock_reset_flag(sk, SOCK_DONE);
3692 tcp_write_queue_purge(sk);
3693 tp->snd_una = tp->write_seq;
3694 tp->snd_sml = tp->write_seq;
3695 tp->snd_up = tp->write_seq;
3696 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3698 if (likely(!tp->repair))
3701 tp->rcv_tstamp = tcp_jiffies32;
3702 tp->rcv_wup = tp->rcv_nxt;
3703 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3705 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3706 inet_csk(sk)->icsk_retransmits = 0;
3707 tcp_clear_retrans(tp);
3710 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3712 struct tcp_sock *tp = tcp_sk(sk);
3713 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3715 tcb->end_seq += skb->len;
3716 __skb_header_release(skb);
3717 sk_wmem_queued_add(sk, skb->truesize);
3718 sk_mem_charge(sk, skb->truesize);
3719 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3720 tp->packets_out += tcp_skb_pcount(skb);
3723 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3724 * queue a data-only packet after the regular SYN, such that regular SYNs
3725 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3726 * only the SYN sequence, the data are retransmitted in the first ACK.
3727 * If cookie is not cached or other error occurs, falls back to send a
3728 * regular SYN with Fast Open cookie request option.
3730 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3732 struct tcp_sock *tp = tcp_sk(sk);
3733 struct tcp_fastopen_request *fo = tp->fastopen_req;
3735 struct sk_buff *syn_data;
3737 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3738 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3741 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3742 * user-MSS. Reserve maximum option space for middleboxes that add
3743 * private TCP options. The cost is reduced data space in SYN :(
3745 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3747 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3748 MAX_TCP_OPTION_SPACE;
3750 space = min_t(size_t, space, fo->size);
3752 /* limit to order-0 allocations */
3753 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3755 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3758 syn_data->ip_summed = CHECKSUM_PARTIAL;
3759 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3761 int copied = copy_from_iter(skb_put(syn_data, space), space,
3762 &fo->data->msg_iter);
3763 if (unlikely(!copied)) {
3764 tcp_skb_tsorted_anchor_cleanup(syn_data);
3765 kfree_skb(syn_data);
3768 if (copied != space) {
3769 skb_trim(syn_data, copied);
3772 skb_zcopy_set(syn_data, fo->uarg, NULL);
3774 /* No more data pending in inet_wait_for_connect() */
3775 if (space == fo->size)
3779 tcp_connect_queue_skb(sk, syn_data);
3781 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3783 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3785 syn->skb_mstamp_ns = syn_data->skb_mstamp_ns;
3787 /* Now full SYN+DATA was cloned and sent (or not),
3788 * remove the SYN from the original skb (syn_data)
3789 * we keep in write queue in case of a retransmit, as we
3790 * also have the SYN packet (with no data) in the same queue.
3792 TCP_SKB_CB(syn_data)->seq++;
3793 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3795 tp->syn_data = (fo->copied > 0);
3796 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3797 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3801 /* data was not sent, put it in write_queue */
3802 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3803 tp->packets_out -= tcp_skb_pcount(syn_data);
3806 /* Send a regular SYN with Fast Open cookie request option */
3807 if (fo->cookie.len > 0)
3809 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3811 tp->syn_fastopen = 0;
3813 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3817 /* Build a SYN and send it off. */
3818 int tcp_connect(struct sock *sk)
3820 struct tcp_sock *tp = tcp_sk(sk);
3821 struct sk_buff *buff;
3824 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3826 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3827 return -EHOSTUNREACH; /* Routing failure or similar. */
3829 tcp_connect_init(sk);
3831 if (unlikely(tp->repair)) {
3832 tcp_finish_connect(sk, NULL);
3836 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3837 if (unlikely(!buff))
3840 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3841 tcp_mstamp_refresh(tp);
3842 tp->retrans_stamp = tcp_time_stamp(tp);
3843 tcp_connect_queue_skb(sk, buff);
3844 tcp_ecn_send_syn(sk, buff);
3845 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3847 /* Send off SYN; include data in Fast Open. */
3848 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3849 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3850 if (err == -ECONNREFUSED)
3853 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3854 * in order to make this packet get counted in tcpOutSegs.
3856 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3857 tp->pushed_seq = tp->write_seq;
3858 buff = tcp_send_head(sk);
3859 if (unlikely(buff)) {
3860 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3861 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3863 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3865 /* Timer for repeating the SYN until an answer. */
3866 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3867 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3870 EXPORT_SYMBOL(tcp_connect);
3872 /* Send out a delayed ack, the caller does the policy checking
3873 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3876 void tcp_send_delayed_ack(struct sock *sk)
3878 struct inet_connection_sock *icsk = inet_csk(sk);
3879 int ato = icsk->icsk_ack.ato;
3880 unsigned long timeout;
3882 if (ato > TCP_DELACK_MIN) {
3883 const struct tcp_sock *tp = tcp_sk(sk);
3884 int max_ato = HZ / 2;
3886 if (inet_csk_in_pingpong_mode(sk) ||
3887 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3888 max_ato = TCP_DELACK_MAX;
3890 /* Slow path, intersegment interval is "high". */
3892 /* If some rtt estimate is known, use it to bound delayed ack.
3893 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3897 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3904 ato = min(ato, max_ato);
3907 ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max);
3909 /* Stay within the limit we were given */
3910 timeout = jiffies + ato;
3912 /* Use new timeout only if there wasn't a older one earlier. */
3913 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3914 /* If delack timer is about to expire, send ACK now. */
3915 if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3920 if (!time_before(timeout, icsk->icsk_ack.timeout))
3921 timeout = icsk->icsk_ack.timeout;
3923 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3924 icsk->icsk_ack.timeout = timeout;
3925 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3928 /* This routine sends an ack and also updates the window. */
3929 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3931 struct sk_buff *buff;
3933 /* If we have been reset, we may not send again. */
3934 if (sk->sk_state == TCP_CLOSE)
3937 /* We are not putting this on the write queue, so
3938 * tcp_transmit_skb() will set the ownership to this
3941 buff = alloc_skb(MAX_TCP_HEADER,
3942 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3943 if (unlikely(!buff)) {
3944 inet_csk_schedule_ack(sk);
3945 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3946 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3947 TCP_DELACK_MAX, TCP_RTO_MAX);
3951 /* Reserve space for headers and prepare control bits. */
3952 skb_reserve(buff, MAX_TCP_HEADER);
3953 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3955 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3957 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3959 skb_set_tcp_pure_ack(buff);
3961 /* Send it off, this clears delayed acks for us. */
3962 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3964 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3966 void tcp_send_ack(struct sock *sk)
3968 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3971 /* This routine sends a packet with an out of date sequence
3972 * number. It assumes the other end will try to ack it.
3974 * Question: what should we make while urgent mode?
3975 * 4.4BSD forces sending single byte of data. We cannot send
3976 * out of window data, because we have SND.NXT==SND.MAX...
3978 * Current solution: to send TWO zero-length segments in urgent mode:
3979 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3980 * out-of-date with SND.UNA-1 to probe window.
3982 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3984 struct tcp_sock *tp = tcp_sk(sk);
3985 struct sk_buff *skb;
3987 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3988 skb = alloc_skb(MAX_TCP_HEADER,
3989 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3993 /* Reserve space for headers and set control bits. */
3994 skb_reserve(skb, MAX_TCP_HEADER);
3995 /* Use a previous sequence. This should cause the other
3996 * end to send an ack. Don't queue or clone SKB, just
3999 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
4000 NET_INC_STATS(sock_net(sk), mib);
4001 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4004 /* Called from setsockopt( ... TCP_REPAIR ) */
4005 void tcp_send_window_probe(struct sock *sk)
4007 if (sk->sk_state == TCP_ESTABLISHED) {
4008 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4009 tcp_mstamp_refresh(tcp_sk(sk));
4010 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4014 /* Initiate keepalive or window probe from timer. */
4015 int tcp_write_wakeup(struct sock *sk, int mib)
4017 struct tcp_sock *tp = tcp_sk(sk);
4018 struct sk_buff *skb;
4020 if (sk->sk_state == TCP_CLOSE)
4023 skb = tcp_send_head(sk);
4024 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4026 unsigned int mss = tcp_current_mss(sk);
4027 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4029 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4030 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4032 /* We are probing the opening of a window
4033 * but the window size is != 0
4034 * must have been a result SWS avoidance ( sender )
4036 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4038 seg_size = min(seg_size, mss);
4039 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4040 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4041 skb, seg_size, mss, GFP_ATOMIC))
4043 } else if (!tcp_skb_pcount(skb))
4044 tcp_set_skb_tso_segs(skb, mss);
4046 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4047 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4049 tcp_event_new_data_sent(sk, skb);
4052 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4053 tcp_xmit_probe_skb(sk, 1, mib);
4054 return tcp_xmit_probe_skb(sk, 0, mib);
4058 /* A window probe timeout has occurred. If window is not closed send
4059 * a partial packet else a zero probe.
4061 void tcp_send_probe0(struct sock *sk)
4063 struct inet_connection_sock *icsk = inet_csk(sk);
4064 struct tcp_sock *tp = tcp_sk(sk);
4065 struct net *net = sock_net(sk);
4066 unsigned long timeout;
4069 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4071 if (tp->packets_out || tcp_write_queue_empty(sk)) {
4072 /* Cancel probe timer, if it is not required. */
4073 icsk->icsk_probes_out = 0;
4074 icsk->icsk_backoff = 0;
4078 icsk->icsk_probes_out++;
4080 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
4081 icsk->icsk_backoff++;
4082 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
4084 /* If packet was not sent due to local congestion,
4085 * Let senders fight for local resources conservatively.
4087 timeout = TCP_RESOURCE_PROBE_INTERVAL;
4089 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
4092 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4094 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4098 tcp_rsk(req)->txhash = net_tx_rndhash();
4099 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4102 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4103 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4104 if (unlikely(tcp_passive_fastopen(sk)))
4105 tcp_sk(sk)->total_retrans++;
4106 trace_tcp_retransmit_synack(sk, req);
4110 EXPORT_SYMBOL(tcp_rtx_synack);