1 // SPDX-License-Identifier: GPL-2.0-only
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
38 #define pr_fmt(fmt) "TCP: " fmt
41 #include <net/mptcp.h>
43 #include <linux/compiler.h>
44 #include <linux/gfp.h>
45 #include <linux/module.h>
46 #include <linux/static_key.h>
48 #include <trace/events/tcp.h>
50 /* Refresh clocks of a TCP socket,
51 * ensuring monotically increasing values.
53 void tcp_mstamp_refresh(struct tcp_sock *tp)
55 u64 val = tcp_clock_ns();
57 tp->tcp_clock_cache = val;
58 tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
61 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
62 int push_one, gfp_t gfp);
64 /* Account for new data that has been sent to the network. */
65 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
67 struct inet_connection_sock *icsk = inet_csk(sk);
68 struct tcp_sock *tp = tcp_sk(sk);
69 unsigned int prior_packets = tp->packets_out;
71 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
73 __skb_unlink(skb, &sk->sk_write_queue);
74 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
76 if (tp->highest_sack == NULL)
77 tp->highest_sack = skb;
79 tp->packets_out += tcp_skb_pcount(skb);
80 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
83 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
87 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
88 * window scaling factor due to loss of precision.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
94 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
96 const struct tcp_sock *tp = tcp_sk(sk);
98 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
99 (tp->rx_opt.wscale_ok &&
100 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
103 return tcp_wnd_end(tp);
106 /* Calculate mss to advertise in SYN segment.
107 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
109 * 1. It is independent of path mtu.
110 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
111 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
112 * attached devices, because some buggy hosts are confused by
114 * 4. We do not make 3, we advertise MSS, calculated from first
115 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
116 * This may be overridden via information stored in routing table.
117 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
118 * probably even Jumbo".
120 static __u16 tcp_advertise_mss(struct sock *sk)
122 struct tcp_sock *tp = tcp_sk(sk);
123 const struct dst_entry *dst = __sk_dst_get(sk);
124 int mss = tp->advmss;
127 unsigned int metric = dst_metric_advmss(dst);
138 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
139 * This is the first part of cwnd validation mechanism.
141 void tcp_cwnd_restart(struct sock *sk, s32 delta)
143 struct tcp_sock *tp = tcp_sk(sk);
144 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
145 u32 cwnd = tp->snd_cwnd;
147 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
149 tp->snd_ssthresh = tcp_current_ssthresh(sk);
150 restart_cwnd = min(restart_cwnd, cwnd);
152 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
154 tp->snd_cwnd = max(cwnd, restart_cwnd);
155 tp->snd_cwnd_stamp = tcp_jiffies32;
156 tp->snd_cwnd_used = 0;
159 /* Congestion state accounting after a packet has been sent. */
160 static void tcp_event_data_sent(struct tcp_sock *tp,
163 struct inet_connection_sock *icsk = inet_csk(sk);
164 const u32 now = tcp_jiffies32;
166 if (tcp_packets_in_flight(tp) == 0)
167 tcp_ca_event(sk, CA_EVENT_TX_START);
169 /* If this is the first data packet sent in response to the
170 * previous received data,
171 * and it is a reply for ato after last received packet,
172 * increase pingpong count.
174 if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) &&
175 (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
176 inet_csk_inc_pingpong_cnt(sk);
181 /* Account for an ACK we sent. */
182 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
185 struct tcp_sock *tp = tcp_sk(sk);
187 if (unlikely(tp->compressed_ack)) {
188 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
190 tp->compressed_ack = 0;
191 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
195 if (unlikely(rcv_nxt != tp->rcv_nxt))
196 return; /* Special ACK sent by DCTCP to reflect ECN */
197 tcp_dec_quickack_mode(sk, pkts);
198 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
201 /* Determine a window scaling and initial window to offer.
202 * Based on the assumption that the given amount of space
203 * will be offered. Store the results in the tp structure.
204 * NOTE: for smooth operation initial space offering should
205 * be a multiple of mss if possible. We assume here that mss >= 1.
206 * This MUST be enforced by all callers.
208 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
209 __u32 *rcv_wnd, __u32 *window_clamp,
210 int wscale_ok, __u8 *rcv_wscale,
213 unsigned int space = (__space < 0 ? 0 : __space);
215 /* If no clamp set the clamp to the max possible scaled window */
216 if (*window_clamp == 0)
217 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
218 space = min(*window_clamp, space);
220 /* Quantize space offering to a multiple of mss if possible. */
222 space = rounddown(space, mss);
224 /* NOTE: offering an initial window larger than 32767
225 * will break some buggy TCP stacks. If the admin tells us
226 * it is likely we could be speaking with such a buggy stack
227 * we will truncate our initial window offering to 32K-1
228 * unless the remote has sent us a window scaling option,
229 * which we interpret as a sign the remote TCP is not
230 * misinterpreting the window field as a signed quantity.
232 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
233 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
235 (*rcv_wnd) = min_t(u32, space, U16_MAX);
238 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
242 /* Set window scaling on max possible window */
243 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
244 space = max_t(u32, space, sysctl_rmem_max);
245 space = min_t(u32, space, *window_clamp);
246 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
249 /* Set the clamp no higher than max representable value */
250 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
252 EXPORT_SYMBOL(tcp_select_initial_window);
254 /* Chose a new window to advertise, update state in tcp_sock for the
255 * socket, and return result with RFC1323 scaling applied. The return
256 * value can be stuffed directly into th->window for an outgoing
259 static u16 tcp_select_window(struct sock *sk)
261 struct tcp_sock *tp = tcp_sk(sk);
262 u32 old_win = tp->rcv_wnd;
263 u32 cur_win = tcp_receive_window(tp);
264 u32 new_win = __tcp_select_window(sk);
266 /* Never shrink the offered window */
267 if (new_win < cur_win) {
268 /* Danger Will Robinson!
269 * Don't update rcv_wup/rcv_wnd here or else
270 * we will not be able to advertise a zero
271 * window in time. --DaveM
273 * Relax Will Robinson.
276 NET_INC_STATS(sock_net(sk),
277 LINUX_MIB_TCPWANTZEROWINDOWADV);
278 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
280 tp->rcv_wnd = new_win;
281 tp->rcv_wup = tp->rcv_nxt;
283 /* Make sure we do not exceed the maximum possible
286 if (!tp->rx_opt.rcv_wscale &&
287 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
288 new_win = min(new_win, MAX_TCP_WINDOW);
290 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
292 /* RFC1323 scaling applied */
293 new_win >>= tp->rx_opt.rcv_wscale;
295 /* If we advertise zero window, disable fast path. */
299 NET_INC_STATS(sock_net(sk),
300 LINUX_MIB_TCPTOZEROWINDOWADV);
301 } else if (old_win == 0) {
302 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
308 /* Packet ECN state for a SYN-ACK */
309 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
311 const struct tcp_sock *tp = tcp_sk(sk);
313 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
314 if (!(tp->ecn_flags & TCP_ECN_OK))
315 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
316 else if (tcp_ca_needs_ecn(sk) ||
317 tcp_bpf_ca_needs_ecn(sk))
321 /* Packet ECN state for a SYN. */
322 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
324 struct tcp_sock *tp = tcp_sk(sk);
325 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
326 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
327 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
330 const struct dst_entry *dst = __sk_dst_get(sk);
332 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
339 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
340 tp->ecn_flags = TCP_ECN_OK;
341 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
346 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
348 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
349 /* tp->ecn_flags are cleared at a later point in time when
350 * SYN ACK is ultimatively being received.
352 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
356 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
358 if (inet_rsk(req)->ecn_ok)
362 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
365 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
366 struct tcphdr *th, int tcp_header_len)
368 struct tcp_sock *tp = tcp_sk(sk);
370 if (tp->ecn_flags & TCP_ECN_OK) {
371 /* Not-retransmitted data segment: set ECT and inject CWR. */
372 if (skb->len != tcp_header_len &&
373 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
375 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
376 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
378 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
380 } else if (!tcp_ca_needs_ecn(sk)) {
381 /* ACK or retransmitted segment: clear ECT|CE */
382 INET_ECN_dontxmit(sk);
384 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
389 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
390 * auto increment end seqno.
392 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
394 skb->ip_summed = CHECKSUM_PARTIAL;
396 TCP_SKB_CB(skb)->tcp_flags = flags;
397 TCP_SKB_CB(skb)->sacked = 0;
399 tcp_skb_pcount_set(skb, 1);
401 TCP_SKB_CB(skb)->seq = seq;
402 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
404 TCP_SKB_CB(skb)->end_seq = seq;
407 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
409 return tp->snd_una != tp->snd_up;
412 #define OPTION_SACK_ADVERTISE (1 << 0)
413 #define OPTION_TS (1 << 1)
414 #define OPTION_MD5 (1 << 2)
415 #define OPTION_WSCALE (1 << 3)
416 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
417 #define OPTION_SMC (1 << 9)
418 #define OPTION_MPTCP (1 << 10)
420 static void smc_options_write(__be32 *ptr, u16 *options)
422 #if IS_ENABLED(CONFIG_SMC)
423 if (static_branch_unlikely(&tcp_have_smc)) {
424 if (unlikely(OPTION_SMC & *options)) {
425 *ptr++ = htonl((TCPOPT_NOP << 24) |
428 (TCPOLEN_EXP_SMC_BASE));
429 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
435 struct tcp_out_options {
436 u16 options; /* bit field of OPTION_* */
437 u16 mss; /* 0 to disable */
438 u8 ws; /* window scale, 0 to disable */
439 u8 num_sack_blocks; /* number of SACK blocks to include */
440 u8 hash_size; /* bytes in hash_location */
441 u8 bpf_opt_len; /* length of BPF hdr option */
442 __u8 *hash_location; /* temporary pointer, overloaded */
443 __u32 tsval, tsecr; /* need to include OPTION_TS */
444 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
445 struct mptcp_out_options mptcp;
448 static void mptcp_options_write(__be32 *ptr, const struct tcp_sock *tp,
449 struct tcp_out_options *opts)
451 #if IS_ENABLED(CONFIG_MPTCP)
452 if (unlikely(OPTION_MPTCP & opts->options))
453 mptcp_write_options(ptr, tp, &opts->mptcp);
457 #ifdef CONFIG_CGROUP_BPF
458 static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
459 enum tcp_synack_type synack_type)
462 return BPF_WRITE_HDR_TCP_CURRENT_MSS;
464 if (unlikely(synack_type == TCP_SYNACK_COOKIE))
465 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
470 /* req, syn_skb and synack_type are used when writing synack */
471 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
472 struct request_sock *req,
473 struct sk_buff *syn_skb,
474 enum tcp_synack_type synack_type,
475 struct tcp_out_options *opts,
476 unsigned int *remaining)
478 struct bpf_sock_ops_kern sock_ops;
481 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
482 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
486 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
489 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
491 sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
494 /* The listen "sk" cannot be passed here because
495 * it is not locked. It would not make too much
496 * sense to do bpf_setsockopt(listen_sk) based
497 * on individual connection request also.
499 * Thus, "req" is passed here and the cgroup-bpf-progs
500 * of the listen "sk" will be run.
502 * "req" is also used here for fastopen even the "sk" here is
503 * a fullsock "child" sk. It is to keep the behavior
504 * consistent between fastopen and non-fastopen on
505 * the bpf programming side.
507 sock_ops.sk = (struct sock *)req;
508 sock_ops.syn_skb = syn_skb;
510 sock_owned_by_me(sk);
512 sock_ops.is_fullsock = 1;
516 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
517 sock_ops.remaining_opt_len = *remaining;
518 /* tcp_current_mss() does not pass a skb */
520 bpf_skops_init_skb(&sock_ops, skb, 0);
522 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
524 if (err || sock_ops.remaining_opt_len == *remaining)
527 opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
528 /* round up to 4 bytes */
529 opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
531 *remaining -= opts->bpf_opt_len;
534 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
535 struct request_sock *req,
536 struct sk_buff *syn_skb,
537 enum tcp_synack_type synack_type,
538 struct tcp_out_options *opts)
540 u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
541 struct bpf_sock_ops_kern sock_ops;
544 if (likely(!max_opt_len))
547 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
549 sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
552 sock_ops.sk = (struct sock *)req;
553 sock_ops.syn_skb = syn_skb;
555 sock_owned_by_me(sk);
557 sock_ops.is_fullsock = 1;
561 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
562 sock_ops.remaining_opt_len = max_opt_len;
563 first_opt_off = tcp_hdrlen(skb) - max_opt_len;
564 bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
566 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
571 nr_written = max_opt_len - sock_ops.remaining_opt_len;
573 if (nr_written < max_opt_len)
574 memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
575 max_opt_len - nr_written);
578 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
579 struct request_sock *req,
580 struct sk_buff *syn_skb,
581 enum tcp_synack_type synack_type,
582 struct tcp_out_options *opts,
583 unsigned int *remaining)
587 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
588 struct request_sock *req,
589 struct sk_buff *syn_skb,
590 enum tcp_synack_type synack_type,
591 struct tcp_out_options *opts)
596 /* Write previously computed TCP options to the packet.
598 * Beware: Something in the Internet is very sensitive to the ordering of
599 * TCP options, we learned this through the hard way, so be careful here.
600 * Luckily we can at least blame others for their non-compliance but from
601 * inter-operability perspective it seems that we're somewhat stuck with
602 * the ordering which we have been using if we want to keep working with
603 * those broken things (not that it currently hurts anybody as there isn't
604 * particular reason why the ordering would need to be changed).
606 * At least SACK_PERM as the first option is known to lead to a disaster
607 * (but it may well be that other scenarios fail similarly).
609 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
610 struct tcp_out_options *opts)
612 u16 options = opts->options; /* mungable copy */
614 if (unlikely(OPTION_MD5 & options)) {
615 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
616 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
617 /* overload cookie hash location */
618 opts->hash_location = (__u8 *)ptr;
622 if (unlikely(opts->mss)) {
623 *ptr++ = htonl((TCPOPT_MSS << 24) |
624 (TCPOLEN_MSS << 16) |
628 if (likely(OPTION_TS & options)) {
629 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
630 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
631 (TCPOLEN_SACK_PERM << 16) |
632 (TCPOPT_TIMESTAMP << 8) |
634 options &= ~OPTION_SACK_ADVERTISE;
636 *ptr++ = htonl((TCPOPT_NOP << 24) |
638 (TCPOPT_TIMESTAMP << 8) |
641 *ptr++ = htonl(opts->tsval);
642 *ptr++ = htonl(opts->tsecr);
645 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
646 *ptr++ = htonl((TCPOPT_NOP << 24) |
648 (TCPOPT_SACK_PERM << 8) |
652 if (unlikely(OPTION_WSCALE & options)) {
653 *ptr++ = htonl((TCPOPT_NOP << 24) |
654 (TCPOPT_WINDOW << 16) |
655 (TCPOLEN_WINDOW << 8) |
659 if (unlikely(opts->num_sack_blocks)) {
660 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
661 tp->duplicate_sack : tp->selective_acks;
664 *ptr++ = htonl((TCPOPT_NOP << 24) |
667 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
668 TCPOLEN_SACK_PERBLOCK)));
670 for (this_sack = 0; this_sack < opts->num_sack_blocks;
672 *ptr++ = htonl(sp[this_sack].start_seq);
673 *ptr++ = htonl(sp[this_sack].end_seq);
676 tp->rx_opt.dsack = 0;
679 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
680 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
682 u32 len; /* Fast Open option length */
685 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
686 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
687 TCPOPT_FASTOPEN_MAGIC);
688 p += TCPOLEN_EXP_FASTOPEN_BASE;
690 len = TCPOLEN_FASTOPEN_BASE + foc->len;
691 *p++ = TCPOPT_FASTOPEN;
695 memcpy(p, foc->val, foc->len);
696 if ((len & 3) == 2) {
697 p[foc->len] = TCPOPT_NOP;
698 p[foc->len + 1] = TCPOPT_NOP;
700 ptr += (len + 3) >> 2;
703 smc_options_write(ptr, &options);
705 mptcp_options_write(ptr, tp, opts);
708 static void smc_set_option(const struct tcp_sock *tp,
709 struct tcp_out_options *opts,
710 unsigned int *remaining)
712 #if IS_ENABLED(CONFIG_SMC)
713 if (static_branch_unlikely(&tcp_have_smc)) {
715 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
716 opts->options |= OPTION_SMC;
717 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
724 static void smc_set_option_cond(const struct tcp_sock *tp,
725 const struct inet_request_sock *ireq,
726 struct tcp_out_options *opts,
727 unsigned int *remaining)
729 #if IS_ENABLED(CONFIG_SMC)
730 if (static_branch_unlikely(&tcp_have_smc)) {
731 if (tp->syn_smc && ireq->smc_ok) {
732 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
733 opts->options |= OPTION_SMC;
734 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
741 static void mptcp_set_option_cond(const struct request_sock *req,
742 struct tcp_out_options *opts,
743 unsigned int *remaining)
745 if (rsk_is_mptcp(req)) {
748 if (mptcp_synack_options(req, &size, &opts->mptcp)) {
749 if (*remaining >= size) {
750 opts->options |= OPTION_MPTCP;
757 /* Compute TCP options for SYN packets. This is not the final
758 * network wire format yet.
760 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
761 struct tcp_out_options *opts,
762 struct tcp_md5sig_key **md5)
764 struct tcp_sock *tp = tcp_sk(sk);
765 unsigned int remaining = MAX_TCP_OPTION_SPACE;
766 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
769 #ifdef CONFIG_TCP_MD5SIG
770 if (static_branch_unlikely(&tcp_md5_needed) &&
771 rcu_access_pointer(tp->md5sig_info)) {
772 *md5 = tp->af_specific->md5_lookup(sk, sk);
774 opts->options |= OPTION_MD5;
775 remaining -= TCPOLEN_MD5SIG_ALIGNED;
780 /* We always get an MSS option. The option bytes which will be seen in
781 * normal data packets should timestamps be used, must be in the MSS
782 * advertised. But we subtract them from tp->mss_cache so that
783 * calculations in tcp_sendmsg are simpler etc. So account for this
784 * fact here if necessary. If we don't do this correctly, as a
785 * receiver we won't recognize data packets as being full sized when we
786 * should, and thus we won't abide by the delayed ACK rules correctly.
787 * SACKs don't matter, we never delay an ACK when we have any of those
789 opts->mss = tcp_advertise_mss(sk);
790 remaining -= TCPOLEN_MSS_ALIGNED;
792 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
793 opts->options |= OPTION_TS;
794 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
795 opts->tsecr = tp->rx_opt.ts_recent;
796 remaining -= TCPOLEN_TSTAMP_ALIGNED;
798 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
799 opts->ws = tp->rx_opt.rcv_wscale;
800 opts->options |= OPTION_WSCALE;
801 remaining -= TCPOLEN_WSCALE_ALIGNED;
803 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
804 opts->options |= OPTION_SACK_ADVERTISE;
805 if (unlikely(!(OPTION_TS & opts->options)))
806 remaining -= TCPOLEN_SACKPERM_ALIGNED;
809 if (fastopen && fastopen->cookie.len >= 0) {
810 u32 need = fastopen->cookie.len;
812 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
813 TCPOLEN_FASTOPEN_BASE;
814 need = (need + 3) & ~3U; /* Align to 32 bits */
815 if (remaining >= need) {
816 opts->options |= OPTION_FAST_OPEN_COOKIE;
817 opts->fastopen_cookie = &fastopen->cookie;
819 tp->syn_fastopen = 1;
820 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
824 smc_set_option(tp, opts, &remaining);
826 if (sk_is_mptcp(sk)) {
829 if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
830 opts->options |= OPTION_MPTCP;
835 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
837 return MAX_TCP_OPTION_SPACE - remaining;
840 /* Set up TCP options for SYN-ACKs. */
841 static unsigned int tcp_synack_options(const struct sock *sk,
842 struct request_sock *req,
843 unsigned int mss, struct sk_buff *skb,
844 struct tcp_out_options *opts,
845 const struct tcp_md5sig_key *md5,
846 struct tcp_fastopen_cookie *foc,
847 enum tcp_synack_type synack_type,
848 struct sk_buff *syn_skb)
850 struct inet_request_sock *ireq = inet_rsk(req);
851 unsigned int remaining = MAX_TCP_OPTION_SPACE;
853 #ifdef CONFIG_TCP_MD5SIG
855 opts->options |= OPTION_MD5;
856 remaining -= TCPOLEN_MD5SIG_ALIGNED;
858 /* We can't fit any SACK blocks in a packet with MD5 + TS
859 * options. There was discussion about disabling SACK
860 * rather than TS in order to fit in better with old,
861 * buggy kernels, but that was deemed to be unnecessary.
863 if (synack_type != TCP_SYNACK_COOKIE)
864 ireq->tstamp_ok &= !ireq->sack_ok;
868 /* We always send an MSS option. */
870 remaining -= TCPOLEN_MSS_ALIGNED;
872 if (likely(ireq->wscale_ok)) {
873 opts->ws = ireq->rcv_wscale;
874 opts->options |= OPTION_WSCALE;
875 remaining -= TCPOLEN_WSCALE_ALIGNED;
877 if (likely(ireq->tstamp_ok)) {
878 opts->options |= OPTION_TS;
879 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
880 opts->tsecr = req->ts_recent;
881 remaining -= TCPOLEN_TSTAMP_ALIGNED;
883 if (likely(ireq->sack_ok)) {
884 opts->options |= OPTION_SACK_ADVERTISE;
885 if (unlikely(!ireq->tstamp_ok))
886 remaining -= TCPOLEN_SACKPERM_ALIGNED;
888 if (foc != NULL && foc->len >= 0) {
891 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
892 TCPOLEN_FASTOPEN_BASE;
893 need = (need + 3) & ~3U; /* Align to 32 bits */
894 if (remaining >= need) {
895 opts->options |= OPTION_FAST_OPEN_COOKIE;
896 opts->fastopen_cookie = foc;
901 mptcp_set_option_cond(req, opts, &remaining);
903 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
905 bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb,
906 synack_type, opts, &remaining);
908 return MAX_TCP_OPTION_SPACE - remaining;
911 /* Compute TCP options for ESTABLISHED sockets. This is not the
912 * final wire format yet.
914 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
915 struct tcp_out_options *opts,
916 struct tcp_md5sig_key **md5)
918 struct tcp_sock *tp = tcp_sk(sk);
919 unsigned int size = 0;
920 unsigned int eff_sacks;
925 #ifdef CONFIG_TCP_MD5SIG
926 if (static_branch_unlikely(&tcp_md5_needed) &&
927 rcu_access_pointer(tp->md5sig_info)) {
928 *md5 = tp->af_specific->md5_lookup(sk, sk);
930 opts->options |= OPTION_MD5;
931 size += TCPOLEN_MD5SIG_ALIGNED;
936 if (likely(tp->rx_opt.tstamp_ok)) {
937 opts->options |= OPTION_TS;
938 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
939 opts->tsecr = tp->rx_opt.ts_recent;
940 size += TCPOLEN_TSTAMP_ALIGNED;
943 /* MPTCP options have precedence over SACK for the limited TCP
944 * option space because a MPTCP connection would be forced to
945 * fall back to regular TCP if a required multipath option is
946 * missing. SACK still gets a chance to use whatever space is
949 if (sk_is_mptcp(sk)) {
950 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
951 unsigned int opt_size = 0;
953 if (mptcp_established_options(sk, skb, &opt_size, remaining,
955 opts->options |= OPTION_MPTCP;
960 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
961 if (unlikely(eff_sacks)) {
962 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
963 if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
964 TCPOLEN_SACK_PERBLOCK))
967 opts->num_sack_blocks =
968 min_t(unsigned int, eff_sacks,
969 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
970 TCPOLEN_SACK_PERBLOCK);
972 size += TCPOLEN_SACK_BASE_ALIGNED +
973 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
976 if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
977 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
978 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
980 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
982 size = MAX_TCP_OPTION_SPACE - remaining;
989 /* TCP SMALL QUEUES (TSQ)
991 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
992 * to reduce RTT and bufferbloat.
993 * We do this using a special skb destructor (tcp_wfree).
995 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
996 * needs to be reallocated in a driver.
997 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
999 * Since transmit from skb destructor is forbidden, we use a tasklet
1000 * to process all sockets that eventually need to send more skbs.
1001 * We use one tasklet per cpu, with its own queue of sockets.
1003 struct tsq_tasklet {
1004 struct tasklet_struct tasklet;
1005 struct list_head head; /* queue of tcp sockets */
1007 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
1009 static void tcp_tsq_write(struct sock *sk)
1011 if ((1 << sk->sk_state) &
1012 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1013 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
1014 struct tcp_sock *tp = tcp_sk(sk);
1016 if (tp->lost_out > tp->retrans_out &&
1017 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
1018 tcp_mstamp_refresh(tp);
1019 tcp_xmit_retransmit_queue(sk);
1022 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
1027 static void tcp_tsq_handler(struct sock *sk)
1030 if (!sock_owned_by_user(sk))
1032 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
1037 * One tasklet per cpu tries to send more skbs.
1038 * We run in tasklet context but need to disable irqs when
1039 * transferring tsq->head because tcp_wfree() might
1040 * interrupt us (non NAPI drivers)
1042 static void tcp_tasklet_func(struct tasklet_struct *t)
1044 struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet);
1046 unsigned long flags;
1047 struct list_head *q, *n;
1048 struct tcp_sock *tp;
1051 local_irq_save(flags);
1052 list_splice_init(&tsq->head, &list);
1053 local_irq_restore(flags);
1055 list_for_each_safe(q, n, &list) {
1056 tp = list_entry(q, struct tcp_sock, tsq_node);
1057 list_del(&tp->tsq_node);
1059 sk = (struct sock *)tp;
1060 smp_mb__before_atomic();
1061 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
1063 tcp_tsq_handler(sk);
1068 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
1069 TCPF_WRITE_TIMER_DEFERRED | \
1070 TCPF_DELACK_TIMER_DEFERRED | \
1071 TCPF_MTU_REDUCED_DEFERRED)
1073 * tcp_release_cb - tcp release_sock() callback
1076 * called from release_sock() to perform protocol dependent
1077 * actions before socket release.
1079 void tcp_release_cb(struct sock *sk)
1081 unsigned long flags, nflags;
1083 /* perform an atomic operation only if at least one flag is set */
1085 flags = sk->sk_tsq_flags;
1086 if (!(flags & TCP_DEFERRED_ALL))
1088 nflags = flags & ~TCP_DEFERRED_ALL;
1089 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
1091 if (flags & TCPF_TSQ_DEFERRED) {
1095 /* Here begins the tricky part :
1096 * We are called from release_sock() with :
1098 * 2) sk_lock.slock spinlock held
1099 * 3) socket owned by us (sk->sk_lock.owned == 1)
1101 * But following code is meant to be called from BH handlers,
1102 * so we should keep BH disabled, but early release socket ownership
1104 sock_release_ownership(sk);
1106 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1107 tcp_write_timer_handler(sk);
1110 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1111 tcp_delack_timer_handler(sk);
1114 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1115 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1119 EXPORT_SYMBOL(tcp_release_cb);
1121 void __init tcp_tasklet_init(void)
1125 for_each_possible_cpu(i) {
1126 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
1128 INIT_LIST_HEAD(&tsq->head);
1129 tasklet_setup(&tsq->tasklet, tcp_tasklet_func);
1134 * Write buffer destructor automatically called from kfree_skb.
1135 * We can't xmit new skbs from this context, as we might already
1138 void tcp_wfree(struct sk_buff *skb)
1140 struct sock *sk = skb->sk;
1141 struct tcp_sock *tp = tcp_sk(sk);
1142 unsigned long flags, nval, oval;
1144 /* Keep one reference on sk_wmem_alloc.
1145 * Will be released by sk_free() from here or tcp_tasklet_func()
1147 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1149 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
1150 * Wait until our queues (qdisc + devices) are drained.
1152 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1153 * - chance for incoming ACK (processed by another cpu maybe)
1154 * to migrate this flow (skb->ooo_okay will be eventually set)
1156 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1159 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
1160 struct tsq_tasklet *tsq;
1163 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1166 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1167 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
1171 /* queue this socket to tasklet queue */
1172 local_irq_save(flags);
1173 tsq = this_cpu_ptr(&tsq_tasklet);
1174 empty = list_empty(&tsq->head);
1175 list_add(&tp->tsq_node, &tsq->head);
1177 tasklet_schedule(&tsq->tasklet);
1178 local_irq_restore(flags);
1185 /* Note: Called under soft irq.
1186 * We can call TCP stack right away, unless socket is owned by user.
1188 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1190 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1191 struct sock *sk = (struct sock *)tp;
1193 tcp_tsq_handler(sk);
1196 return HRTIMER_NORESTART;
1199 static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1202 struct tcp_sock *tp = tcp_sk(sk);
1204 if (sk->sk_pacing_status != SK_PACING_NONE) {
1205 unsigned long rate = sk->sk_pacing_rate;
1207 /* Original sch_fq does not pace first 10 MSS
1208 * Note that tp->data_segs_out overflows after 2^32 packets,
1209 * this is a minor annoyance.
1211 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1212 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1213 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1215 /* take into account OS jitter */
1216 len_ns -= min_t(u64, len_ns / 2, credit);
1217 tp->tcp_wstamp_ns += len_ns;
1220 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1223 INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1224 INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1225 INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1227 /* This routine actually transmits TCP packets queued in by
1228 * tcp_do_sendmsg(). This is used by both the initial
1229 * transmission and possible later retransmissions.
1230 * All SKB's seen here are completely headerless. It is our
1231 * job to build the TCP header, and pass the packet down to
1232 * IP so it can do the same plus pass the packet off to the
1235 * We are working here with either a clone of the original
1236 * SKB, or a fresh unique copy made by the retransmit engine.
1238 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1239 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1241 const struct inet_connection_sock *icsk = inet_csk(sk);
1242 struct inet_sock *inet;
1243 struct tcp_sock *tp;
1244 struct tcp_skb_cb *tcb;
1245 struct tcp_out_options opts;
1246 unsigned int tcp_options_size, tcp_header_size;
1247 struct sk_buff *oskb = NULL;
1248 struct tcp_md5sig_key *md5;
1253 BUG_ON(!skb || !tcp_skb_pcount(skb));
1255 prior_wstamp = tp->tcp_wstamp_ns;
1256 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1257 skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
1259 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1263 tcp_skb_tsorted_save(oskb) {
1264 if (unlikely(skb_cloned(oskb)))
1265 skb = pskb_copy(oskb, gfp_mask);
1267 skb = skb_clone(oskb, gfp_mask);
1268 } tcp_skb_tsorted_restore(oskb);
1272 /* retransmit skbs might have a non zero value in skb->dev
1273 * because skb->dev is aliased with skb->rbnode.rb_left
1279 tcb = TCP_SKB_CB(skb);
1280 memset(&opts, 0, sizeof(opts));
1282 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1283 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1285 tcp_options_size = tcp_established_options(sk, skb, &opts,
1287 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1288 * at receiver : This slightly improve GRO performance.
1289 * Note that we do not force the PSH flag for non GSO packets,
1290 * because they might be sent under high congestion events,
1291 * and in this case it is better to delay the delivery of 1-MSS
1292 * packets and thus the corresponding ACK packet that would
1293 * release the following packet.
1295 if (tcp_skb_pcount(skb) > 1)
1296 tcb->tcp_flags |= TCPHDR_PSH;
1298 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1300 /* if no packet is in qdisc/device queue, then allow XPS to select
1301 * another queue. We can be called from tcp_tsq_handler()
1302 * which holds one reference to sk.
1304 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1305 * One way to get this would be to set skb->truesize = 2 on them.
1307 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1309 /* If we had to use memory reserve to allocate this skb,
1310 * this might cause drops if packet is looped back :
1311 * Other socket might not have SOCK_MEMALLOC.
1312 * Packets not looped back do not care about pfmemalloc.
1314 skb->pfmemalloc = 0;
1316 skb_push(skb, tcp_header_size);
1317 skb_reset_transport_header(skb);
1321 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1322 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1324 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1326 /* Build TCP header and checksum it. */
1327 th = (struct tcphdr *)skb->data;
1328 th->source = inet->inet_sport;
1329 th->dest = inet->inet_dport;
1330 th->seq = htonl(tcb->seq);
1331 th->ack_seq = htonl(rcv_nxt);
1332 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1338 /* The urg_mode check is necessary during a below snd_una win probe */
1339 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1340 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1341 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1343 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1344 th->urg_ptr = htons(0xFFFF);
1349 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1350 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1351 th->window = htons(tcp_select_window(sk));
1352 tcp_ecn_send(sk, skb, th, tcp_header_size);
1354 /* RFC1323: The window in SYN & SYN/ACK segments
1357 th->window = htons(min(tp->rcv_wnd, 65535U));
1360 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1362 #ifdef CONFIG_TCP_MD5SIG
1363 /* Calculate the MD5 hash, as we have all we need now */
1365 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1366 tp->af_specific->calc_md5_hash(opts.hash_location,
1371 /* BPF prog is the last one writing header option */
1372 bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1374 INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1375 tcp_v6_send_check, tcp_v4_send_check,
1378 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1379 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1381 if (skb->len != tcp_header_size) {
1382 tcp_event_data_sent(tp, sk);
1383 tp->data_segs_out += tcp_skb_pcount(skb);
1384 tp->bytes_sent += skb->len - tcp_header_size;
1387 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1388 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1389 tcp_skb_pcount(skb));
1391 tp->segs_out += tcp_skb_pcount(skb);
1392 skb_set_hash_from_sk(skb, sk);
1393 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1394 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1395 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1397 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1399 /* Cleanup our debris for IP stacks */
1400 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1401 sizeof(struct inet6_skb_parm)));
1403 tcp_add_tx_delay(skb, tp);
1405 err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1406 inet6_csk_xmit, ip_queue_xmit,
1407 sk, skb, &inet->cork.fl);
1409 if (unlikely(err > 0)) {
1411 err = net_xmit_eval(err);
1414 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1415 tcp_rate_skb_sent(sk, oskb);
1420 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1423 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1424 tcp_sk(sk)->rcv_nxt);
1427 /* This routine just queues the buffer for sending.
1429 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1430 * otherwise socket can stall.
1432 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1434 struct tcp_sock *tp = tcp_sk(sk);
1436 /* Advance write_seq and place onto the write_queue. */
1437 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1438 __skb_header_release(skb);
1439 tcp_add_write_queue_tail(sk, skb);
1440 sk_wmem_queued_add(sk, skb->truesize);
1441 sk_mem_charge(sk, skb->truesize);
1444 /* Initialize TSO segments for a packet. */
1445 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1447 if (skb->len <= mss_now) {
1448 /* Avoid the costly divide in the normal
1451 tcp_skb_pcount_set(skb, 1);
1452 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1454 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1455 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1459 /* Pcount in the middle of the write queue got changed, we need to do various
1460 * tweaks to fix counters
1462 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1464 struct tcp_sock *tp = tcp_sk(sk);
1466 tp->packets_out -= decr;
1468 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1469 tp->sacked_out -= decr;
1470 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1471 tp->retrans_out -= decr;
1472 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1473 tp->lost_out -= decr;
1475 /* Reno case is special. Sigh... */
1476 if (tcp_is_reno(tp) && decr > 0)
1477 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1479 if (tp->lost_skb_hint &&
1480 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1481 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1482 tp->lost_cnt_hint -= decr;
1484 tcp_verify_left_out(tp);
1487 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1489 return TCP_SKB_CB(skb)->txstamp_ack ||
1490 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1493 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1495 struct skb_shared_info *shinfo = skb_shinfo(skb);
1497 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1498 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1499 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1500 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1502 shinfo->tx_flags &= ~tsflags;
1503 shinfo2->tx_flags |= tsflags;
1504 swap(shinfo->tskey, shinfo2->tskey);
1505 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1506 TCP_SKB_CB(skb)->txstamp_ack = 0;
1510 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1512 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1513 TCP_SKB_CB(skb)->eor = 0;
1516 /* Insert buff after skb on the write or rtx queue of sk. */
1517 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1518 struct sk_buff *buff,
1520 enum tcp_queue tcp_queue)
1522 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1523 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1525 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1528 /* Function to create two new TCP segments. Shrinks the given segment
1529 * to the specified size and appends a new segment with the rest of the
1530 * packet to the list. This won't be called frequently, I hope.
1531 * Remember, these are still headerless SKBs at this point.
1533 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1534 struct sk_buff *skb, u32 len,
1535 unsigned int mss_now, gfp_t gfp)
1537 struct tcp_sock *tp = tcp_sk(sk);
1538 struct sk_buff *buff;
1539 int nsize, old_factor;
1544 if (WARN_ON(len > skb->len))
1547 nsize = skb_headlen(skb) - len;
1551 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1552 * We need some allowance to not penalize applications setting small
1554 * Also allow first and last skb in retransmit queue to be split.
1556 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1557 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1558 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1559 skb != tcp_rtx_queue_head(sk) &&
1560 skb != tcp_rtx_queue_tail(sk))) {
1561 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1565 if (skb_unclone(skb, gfp))
1568 /* Get a new skb... force flag on. */
1569 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1571 return -ENOMEM; /* We'll just try again later. */
1572 skb_copy_decrypted(buff, skb);
1573 mptcp_skb_ext_copy(buff, skb);
1575 sk_wmem_queued_add(sk, buff->truesize);
1576 sk_mem_charge(sk, buff->truesize);
1577 nlen = skb->len - len - nsize;
1578 buff->truesize += nlen;
1579 skb->truesize -= nlen;
1581 /* Correct the sequence numbers. */
1582 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1583 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1584 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1586 /* PSH and FIN should only be set in the second packet. */
1587 flags = TCP_SKB_CB(skb)->tcp_flags;
1588 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1589 TCP_SKB_CB(buff)->tcp_flags = flags;
1590 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1591 tcp_skb_fragment_eor(skb, buff);
1593 skb_split(skb, buff, len);
1595 buff->ip_summed = CHECKSUM_PARTIAL;
1597 buff->tstamp = skb->tstamp;
1598 tcp_fragment_tstamp(skb, buff);
1600 old_factor = tcp_skb_pcount(skb);
1602 /* Fix up tso_factor for both original and new SKB. */
1603 tcp_set_skb_tso_segs(skb, mss_now);
1604 tcp_set_skb_tso_segs(buff, mss_now);
1606 /* Update delivered info for the new segment */
1607 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1609 /* If this packet has been sent out already, we must
1610 * adjust the various packet counters.
1612 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1613 int diff = old_factor - tcp_skb_pcount(skb) -
1614 tcp_skb_pcount(buff);
1617 tcp_adjust_pcount(sk, skb, diff);
1620 /* Link BUFF into the send queue. */
1621 __skb_header_release(buff);
1622 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1623 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1624 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1629 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1630 * data is not copied, but immediately discarded.
1632 static int __pskb_trim_head(struct sk_buff *skb, int len)
1634 struct skb_shared_info *shinfo;
1637 eat = min_t(int, len, skb_headlen(skb));
1639 __skb_pull(skb, eat);
1646 shinfo = skb_shinfo(skb);
1647 for (i = 0; i < shinfo->nr_frags; i++) {
1648 int size = skb_frag_size(&shinfo->frags[i]);
1651 skb_frag_unref(skb, i);
1654 shinfo->frags[k] = shinfo->frags[i];
1656 skb_frag_off_add(&shinfo->frags[k], eat);
1657 skb_frag_size_sub(&shinfo->frags[k], eat);
1663 shinfo->nr_frags = k;
1665 skb->data_len -= len;
1666 skb->len = skb->data_len;
1670 /* Remove acked data from a packet in the transmit queue. */
1671 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1675 if (skb_unclone(skb, GFP_ATOMIC))
1678 delta_truesize = __pskb_trim_head(skb, len);
1680 TCP_SKB_CB(skb)->seq += len;
1681 skb->ip_summed = CHECKSUM_PARTIAL;
1683 if (delta_truesize) {
1684 skb->truesize -= delta_truesize;
1685 sk_wmem_queued_add(sk, -delta_truesize);
1686 sk_mem_uncharge(sk, delta_truesize);
1689 /* Any change of skb->len requires recalculation of tso factor. */
1690 if (tcp_skb_pcount(skb) > 1)
1691 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1696 /* Calculate MSS not accounting any TCP options. */
1697 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1699 const struct tcp_sock *tp = tcp_sk(sk);
1700 const struct inet_connection_sock *icsk = inet_csk(sk);
1703 /* Calculate base mss without TCP options:
1704 It is MMS_S - sizeof(tcphdr) of rfc1122
1706 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1708 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1709 if (icsk->icsk_af_ops->net_frag_header_len) {
1710 const struct dst_entry *dst = __sk_dst_get(sk);
1712 if (dst && dst_allfrag(dst))
1713 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1716 /* Clamp it (mss_clamp does not include tcp options) */
1717 if (mss_now > tp->rx_opt.mss_clamp)
1718 mss_now = tp->rx_opt.mss_clamp;
1720 /* Now subtract optional transport overhead */
1721 mss_now -= icsk->icsk_ext_hdr_len;
1723 /* Then reserve room for full set of TCP options and 8 bytes of data */
1724 mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1728 /* Calculate MSS. Not accounting for SACKs here. */
1729 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1731 /* Subtract TCP options size, not including SACKs */
1732 return __tcp_mtu_to_mss(sk, pmtu) -
1733 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1736 /* Inverse of above */
1737 int tcp_mss_to_mtu(struct sock *sk, int mss)
1739 const struct tcp_sock *tp = tcp_sk(sk);
1740 const struct inet_connection_sock *icsk = inet_csk(sk);
1744 tp->tcp_header_len +
1745 icsk->icsk_ext_hdr_len +
1746 icsk->icsk_af_ops->net_header_len;
1748 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1749 if (icsk->icsk_af_ops->net_frag_header_len) {
1750 const struct dst_entry *dst = __sk_dst_get(sk);
1752 if (dst && dst_allfrag(dst))
1753 mtu += icsk->icsk_af_ops->net_frag_header_len;
1757 EXPORT_SYMBOL(tcp_mss_to_mtu);
1759 /* MTU probing init per socket */
1760 void tcp_mtup_init(struct sock *sk)
1762 struct tcp_sock *tp = tcp_sk(sk);
1763 struct inet_connection_sock *icsk = inet_csk(sk);
1764 struct net *net = sock_net(sk);
1766 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1767 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1768 icsk->icsk_af_ops->net_header_len;
1769 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1770 icsk->icsk_mtup.probe_size = 0;
1771 if (icsk->icsk_mtup.enabled)
1772 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1774 EXPORT_SYMBOL(tcp_mtup_init);
1776 /* This function synchronize snd mss to current pmtu/exthdr set.
1778 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1779 for TCP options, but includes only bare TCP header.
1781 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1782 It is minimum of user_mss and mss received with SYN.
1783 It also does not include TCP options.
1785 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1787 tp->mss_cache is current effective sending mss, including
1788 all tcp options except for SACKs. It is evaluated,
1789 taking into account current pmtu, but never exceeds
1790 tp->rx_opt.mss_clamp.
1792 NOTE1. rfc1122 clearly states that advertised MSS
1793 DOES NOT include either tcp or ip options.
1795 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1796 are READ ONLY outside this function. --ANK (980731)
1798 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1800 struct tcp_sock *tp = tcp_sk(sk);
1801 struct inet_connection_sock *icsk = inet_csk(sk);
1804 if (icsk->icsk_mtup.search_high > pmtu)
1805 icsk->icsk_mtup.search_high = pmtu;
1807 mss_now = tcp_mtu_to_mss(sk, pmtu);
1808 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1810 /* And store cached results */
1811 icsk->icsk_pmtu_cookie = pmtu;
1812 if (icsk->icsk_mtup.enabled)
1813 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1814 tp->mss_cache = mss_now;
1818 EXPORT_SYMBOL(tcp_sync_mss);
1820 /* Compute the current effective MSS, taking SACKs and IP options,
1821 * and even PMTU discovery events into account.
1823 unsigned int tcp_current_mss(struct sock *sk)
1825 const struct tcp_sock *tp = tcp_sk(sk);
1826 const struct dst_entry *dst = __sk_dst_get(sk);
1828 unsigned int header_len;
1829 struct tcp_out_options opts;
1830 struct tcp_md5sig_key *md5;
1832 mss_now = tp->mss_cache;
1835 u32 mtu = dst_mtu(dst);
1836 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1837 mss_now = tcp_sync_mss(sk, mtu);
1840 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1841 sizeof(struct tcphdr);
1842 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1843 * some common options. If this is an odd packet (because we have SACK
1844 * blocks etc) then our calculated header_len will be different, and
1845 * we have to adjust mss_now correspondingly */
1846 if (header_len != tp->tcp_header_len) {
1847 int delta = (int) header_len - tp->tcp_header_len;
1854 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1855 * As additional protections, we do not touch cwnd in retransmission phases,
1856 * and if application hit its sndbuf limit recently.
1858 static void tcp_cwnd_application_limited(struct sock *sk)
1860 struct tcp_sock *tp = tcp_sk(sk);
1862 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1863 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1864 /* Limited by application or receiver window. */
1865 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1866 u32 win_used = max(tp->snd_cwnd_used, init_win);
1867 if (win_used < tp->snd_cwnd) {
1868 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1869 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1871 tp->snd_cwnd_used = 0;
1873 tp->snd_cwnd_stamp = tcp_jiffies32;
1876 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1878 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1879 struct tcp_sock *tp = tcp_sk(sk);
1881 /* Track the maximum number of outstanding packets in each
1882 * window, and remember whether we were cwnd-limited then.
1884 if (!before(tp->snd_una, tp->max_packets_seq) ||
1885 tp->packets_out > tp->max_packets_out ||
1887 tp->max_packets_out = tp->packets_out;
1888 tp->max_packets_seq = tp->snd_nxt;
1889 tp->is_cwnd_limited = is_cwnd_limited;
1892 if (tcp_is_cwnd_limited(sk)) {
1893 /* Network is feed fully. */
1894 tp->snd_cwnd_used = 0;
1895 tp->snd_cwnd_stamp = tcp_jiffies32;
1897 /* Network starves. */
1898 if (tp->packets_out > tp->snd_cwnd_used)
1899 tp->snd_cwnd_used = tp->packets_out;
1901 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1902 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1903 !ca_ops->cong_control)
1904 tcp_cwnd_application_limited(sk);
1906 /* The following conditions together indicate the starvation
1907 * is caused by insufficient sender buffer:
1908 * 1) just sent some data (see tcp_write_xmit)
1909 * 2) not cwnd limited (this else condition)
1910 * 3) no more data to send (tcp_write_queue_empty())
1911 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1913 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1914 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1915 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1916 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1920 /* Minshall's variant of the Nagle send check. */
1921 static bool tcp_minshall_check(const struct tcp_sock *tp)
1923 return after(tp->snd_sml, tp->snd_una) &&
1924 !after(tp->snd_sml, tp->snd_nxt);
1927 /* Update snd_sml if this skb is under mss
1928 * Note that a TSO packet might end with a sub-mss segment
1929 * The test is really :
1930 * if ((skb->len % mss) != 0)
1931 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1932 * But we can avoid doing the divide again given we already have
1933 * skb_pcount = skb->len / mss_now
1935 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1936 const struct sk_buff *skb)
1938 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1939 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1942 /* Return false, if packet can be sent now without violation Nagle's rules:
1943 * 1. It is full sized. (provided by caller in %partial bool)
1944 * 2. Or it contains FIN. (already checked by caller)
1945 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1946 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1947 * With Minshall's modification: all sent small packets are ACKed.
1949 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1953 ((nonagle & TCP_NAGLE_CORK) ||
1954 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1957 /* Return how many segs we'd like on a TSO packet,
1958 * to send one TSO packet per ms
1960 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1965 bytes = min_t(unsigned long,
1966 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift),
1967 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1969 /* Goal is to send at least one packet per ms,
1970 * not one big TSO packet every 100 ms.
1971 * This preserves ACK clocking and is consistent
1972 * with tcp_tso_should_defer() heuristic.
1974 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1979 /* Return the number of segments we want in the skb we are transmitting.
1980 * See if congestion control module wants to decide; otherwise, autosize.
1982 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1984 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1985 u32 min_tso, tso_segs;
1987 min_tso = ca_ops->min_tso_segs ?
1988 ca_ops->min_tso_segs(sk) :
1989 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1991 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1992 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1995 /* Returns the portion of skb which can be sent right away */
1996 static unsigned int tcp_mss_split_point(const struct sock *sk,
1997 const struct sk_buff *skb,
1998 unsigned int mss_now,
1999 unsigned int max_segs,
2002 const struct tcp_sock *tp = tcp_sk(sk);
2003 u32 partial, needed, window, max_len;
2005 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2006 max_len = mss_now * max_segs;
2008 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2011 needed = min(skb->len, window);
2013 if (max_len <= needed)
2016 partial = needed % mss_now;
2017 /* If last segment is not a full MSS, check if Nagle rules allow us
2018 * to include this last segment in this skb.
2019 * Otherwise, we'll split the skb at last MSS boundary
2021 if (tcp_nagle_check(partial != 0, tp, nonagle))
2022 return needed - partial;
2027 /* Can at least one segment of SKB be sent right now, according to the
2028 * congestion window rules? If so, return how many segments are allowed.
2030 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
2031 const struct sk_buff *skb)
2033 u32 in_flight, cwnd, halfcwnd;
2035 /* Don't be strict about the congestion window for the final FIN. */
2036 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2037 tcp_skb_pcount(skb) == 1)
2040 in_flight = tcp_packets_in_flight(tp);
2041 cwnd = tp->snd_cwnd;
2042 if (in_flight >= cwnd)
2045 /* For better scheduling, ensure we have at least
2046 * 2 GSO packets in flight.
2048 halfcwnd = max(cwnd >> 1, 1U);
2049 return min(halfcwnd, cwnd - in_flight);
2052 /* Initialize TSO state of a skb.
2053 * This must be invoked the first time we consider transmitting
2054 * SKB onto the wire.
2056 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
2058 int tso_segs = tcp_skb_pcount(skb);
2060 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
2061 tcp_set_skb_tso_segs(skb, mss_now);
2062 tso_segs = tcp_skb_pcount(skb);
2068 /* Return true if the Nagle test allows this packet to be
2071 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2072 unsigned int cur_mss, int nonagle)
2074 /* Nagle rule does not apply to frames, which sit in the middle of the
2075 * write_queue (they have no chances to get new data).
2077 * This is implemented in the callers, where they modify the 'nonagle'
2078 * argument based upon the location of SKB in the send queue.
2080 if (nonagle & TCP_NAGLE_PUSH)
2083 /* Don't use the nagle rule for urgent data (or for the final FIN). */
2084 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2087 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
2093 /* Does at least the first segment of SKB fit into the send window? */
2094 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2095 const struct sk_buff *skb,
2096 unsigned int cur_mss)
2098 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2100 if (skb->len > cur_mss)
2101 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2103 return !after(end_seq, tcp_wnd_end(tp));
2106 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2107 * which is put after SKB on the list. It is very much like
2108 * tcp_fragment() except that it may make several kinds of assumptions
2109 * in order to speed up the splitting operation. In particular, we
2110 * know that all the data is in scatter-gather pages, and that the
2111 * packet has never been sent out before (and thus is not cloned).
2113 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2114 unsigned int mss_now, gfp_t gfp)
2116 int nlen = skb->len - len;
2117 struct sk_buff *buff;
2120 /* All of a TSO frame must be composed of paged data. */
2121 if (skb->len != skb->data_len)
2122 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2123 skb, len, mss_now, gfp);
2125 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
2126 if (unlikely(!buff))
2128 skb_copy_decrypted(buff, skb);
2129 mptcp_skb_ext_copy(buff, skb);
2131 sk_wmem_queued_add(sk, buff->truesize);
2132 sk_mem_charge(sk, buff->truesize);
2133 buff->truesize += nlen;
2134 skb->truesize -= nlen;
2136 /* Correct the sequence numbers. */
2137 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2138 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2139 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2141 /* PSH and FIN should only be set in the second packet. */
2142 flags = TCP_SKB_CB(skb)->tcp_flags;
2143 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2144 TCP_SKB_CB(buff)->tcp_flags = flags;
2146 /* This packet was never sent out yet, so no SACK bits. */
2147 TCP_SKB_CB(buff)->sacked = 0;
2149 tcp_skb_fragment_eor(skb, buff);
2151 buff->ip_summed = CHECKSUM_PARTIAL;
2152 skb_split(skb, buff, len);
2153 tcp_fragment_tstamp(skb, buff);
2155 /* Fix up tso_factor for both original and new SKB. */
2156 tcp_set_skb_tso_segs(skb, mss_now);
2157 tcp_set_skb_tso_segs(buff, mss_now);
2159 /* Link BUFF into the send queue. */
2160 __skb_header_release(buff);
2161 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2166 /* Try to defer sending, if possible, in order to minimize the amount
2167 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2169 * This algorithm is from John Heffner.
2171 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2172 bool *is_cwnd_limited,
2173 bool *is_rwnd_limited,
2176 const struct inet_connection_sock *icsk = inet_csk(sk);
2177 u32 send_win, cong_win, limit, in_flight;
2178 struct tcp_sock *tp = tcp_sk(sk);
2179 struct sk_buff *head;
2183 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2186 /* Avoid bursty behavior by allowing defer
2187 * only if the last write was recent (1 ms).
2188 * Note that tp->tcp_wstamp_ns can be in the future if we have
2189 * packets waiting in a qdisc or device for EDT delivery.
2191 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2195 in_flight = tcp_packets_in_flight(tp);
2197 BUG_ON(tcp_skb_pcount(skb) <= 1);
2198 BUG_ON(tp->snd_cwnd <= in_flight);
2200 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2202 /* From in_flight test above, we know that cwnd > in_flight. */
2203 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
2205 limit = min(send_win, cong_win);
2207 /* If a full-sized TSO skb can be sent, do it. */
2208 if (limit >= max_segs * tp->mss_cache)
2211 /* Middle in queue won't get any more data, full sendable already? */
2212 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2215 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2217 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
2219 /* If at least some fraction of a window is available,
2222 chunk /= win_divisor;
2226 /* Different approach, try not to defer past a single
2227 * ACK. Receiver should ACK every other full sized
2228 * frame, so if we have space for more than 3 frames
2231 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2235 /* TODO : use tsorted_sent_queue ? */
2236 head = tcp_rtx_queue_head(sk);
2239 delta = tp->tcp_clock_cache - head->tstamp;
2240 /* If next ACK is likely to come too late (half srtt), do not defer */
2241 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2244 /* Ok, it looks like it is advisable to defer.
2245 * Three cases are tracked :
2246 * 1) We are cwnd-limited
2247 * 2) We are rwnd-limited
2248 * 3) We are application limited.
2250 if (cong_win < send_win) {
2251 if (cong_win <= skb->len) {
2252 *is_cwnd_limited = true;
2256 if (send_win <= skb->len) {
2257 *is_rwnd_limited = true;
2262 /* If this packet won't get more data, do not wait. */
2263 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2264 TCP_SKB_CB(skb)->eor)
2273 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2275 struct inet_connection_sock *icsk = inet_csk(sk);
2276 struct tcp_sock *tp = tcp_sk(sk);
2277 struct net *net = sock_net(sk);
2281 interval = net->ipv4.sysctl_tcp_probe_interval;
2282 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2283 if (unlikely(delta >= interval * HZ)) {
2284 int mss = tcp_current_mss(sk);
2286 /* Update current search range */
2287 icsk->icsk_mtup.probe_size = 0;
2288 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2289 sizeof(struct tcphdr) +
2290 icsk->icsk_af_ops->net_header_len;
2291 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2293 /* Update probe time stamp */
2294 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2298 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2300 struct sk_buff *skb, *next;
2302 skb = tcp_send_head(sk);
2303 tcp_for_write_queue_from_safe(skb, next, sk) {
2304 if (len <= skb->len)
2307 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2316 /* Create a new MTU probe if we are ready.
2317 * MTU probe is regularly attempting to increase the path MTU by
2318 * deliberately sending larger packets. This discovers routing
2319 * changes resulting in larger path MTUs.
2321 * Returns 0 if we should wait to probe (no cwnd available),
2322 * 1 if a probe was sent,
2325 static int tcp_mtu_probe(struct sock *sk)
2327 struct inet_connection_sock *icsk = inet_csk(sk);
2328 struct tcp_sock *tp = tcp_sk(sk);
2329 struct sk_buff *skb, *nskb, *next;
2330 struct net *net = sock_net(sk);
2337 /* Not currently probing/verifying,
2339 * have enough cwnd, and
2340 * not SACKing (the variable headers throw things off)
2342 if (likely(!icsk->icsk_mtup.enabled ||
2343 icsk->icsk_mtup.probe_size ||
2344 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2345 tp->snd_cwnd < 11 ||
2346 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2349 /* Use binary search for probe_size between tcp_mss_base,
2350 * and current mss_clamp. if (search_high - search_low)
2351 * smaller than a threshold, backoff from probing.
2353 mss_now = tcp_current_mss(sk);
2354 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2355 icsk->icsk_mtup.search_low) >> 1);
2356 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2357 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2358 /* When misfortune happens, we are reprobing actively,
2359 * and then reprobe timer has expired. We stick with current
2360 * probing process by not resetting search range to its orignal.
2362 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2363 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2364 /* Check whether enough time has elaplased for
2365 * another round of probing.
2367 tcp_mtu_check_reprobe(sk);
2371 /* Have enough data in the send queue to probe? */
2372 if (tp->write_seq - tp->snd_nxt < size_needed)
2375 if (tp->snd_wnd < size_needed)
2377 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2380 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2381 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2382 if (!tcp_packets_in_flight(tp))
2388 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2391 /* We're allowed to probe. Build it now. */
2392 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2395 sk_wmem_queued_add(sk, nskb->truesize);
2396 sk_mem_charge(sk, nskb->truesize);
2398 skb = tcp_send_head(sk);
2399 skb_copy_decrypted(nskb, skb);
2400 mptcp_skb_ext_copy(nskb, skb);
2402 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2403 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2404 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2405 TCP_SKB_CB(nskb)->sacked = 0;
2407 nskb->ip_summed = CHECKSUM_PARTIAL;
2409 tcp_insert_write_queue_before(nskb, skb, sk);
2410 tcp_highest_sack_replace(sk, skb, nskb);
2413 tcp_for_write_queue_from_safe(skb, next, sk) {
2414 copy = min_t(int, skb->len, probe_size - len);
2415 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2417 if (skb->len <= copy) {
2418 /* We've eaten all the data from this skb.
2420 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2421 /* If this is the last SKB we copy and eor is set
2422 * we need to propagate it to the new skb.
2424 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2425 tcp_skb_collapse_tstamp(nskb, skb);
2426 tcp_unlink_write_queue(skb, sk);
2427 sk_wmem_free_skb(sk, skb);
2429 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2430 ~(TCPHDR_FIN|TCPHDR_PSH);
2431 if (!skb_shinfo(skb)->nr_frags) {
2432 skb_pull(skb, copy);
2434 __pskb_trim_head(skb, copy);
2435 tcp_set_skb_tso_segs(skb, mss_now);
2437 TCP_SKB_CB(skb)->seq += copy;
2442 if (len >= probe_size)
2445 tcp_init_tso_segs(nskb, nskb->len);
2447 /* We're ready to send. If this fails, the probe will
2448 * be resegmented into mss-sized pieces by tcp_write_xmit().
2450 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2451 /* Decrement cwnd here because we are sending
2452 * effectively two packets. */
2454 tcp_event_new_data_sent(sk, nskb);
2456 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2457 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2458 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2466 static bool tcp_pacing_check(struct sock *sk)
2468 struct tcp_sock *tp = tcp_sk(sk);
2470 if (!tcp_needs_internal_pacing(sk))
2473 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2476 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2477 hrtimer_start(&tp->pacing_timer,
2478 ns_to_ktime(tp->tcp_wstamp_ns),
2479 HRTIMER_MODE_ABS_PINNED_SOFT);
2485 /* TCP Small Queues :
2486 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2487 * (These limits are doubled for retransmits)
2489 * - better RTT estimation and ACK scheduling
2492 * Alas, some drivers / subsystems require a fair amount
2493 * of queued bytes to ensure line rate.
2494 * One example is wifi aggregation (802.11 AMPDU)
2496 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2497 unsigned int factor)
2499 unsigned long limit;
2501 limit = max_t(unsigned long,
2503 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2504 if (sk->sk_pacing_status == SK_PACING_NONE)
2505 limit = min_t(unsigned long, limit,
2506 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2509 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2510 tcp_sk(sk)->tcp_tx_delay) {
2511 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2513 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2514 * approximate our needs assuming an ~100% skb->truesize overhead.
2515 * USEC_PER_SEC is approximated by 2^20.
2516 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2518 extra_bytes >>= (20 - 1);
2519 limit += extra_bytes;
2521 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2522 /* Always send skb if rtx queue is empty.
2523 * No need to wait for TX completion to call us back,
2524 * after softirq/tasklet schedule.
2525 * This helps when TX completions are delayed too much.
2527 if (tcp_rtx_queue_empty(sk))
2530 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2531 /* It is possible TX completion already happened
2532 * before we set TSQ_THROTTLED, so we must
2533 * test again the condition.
2535 smp_mb__after_atomic();
2536 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2542 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2544 const u32 now = tcp_jiffies32;
2545 enum tcp_chrono old = tp->chrono_type;
2547 if (old > TCP_CHRONO_UNSPEC)
2548 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2549 tp->chrono_start = now;
2550 tp->chrono_type = new;
2553 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2555 struct tcp_sock *tp = tcp_sk(sk);
2557 /* If there are multiple conditions worthy of tracking in a
2558 * chronograph then the highest priority enum takes precedence
2559 * over the other conditions. So that if something "more interesting"
2560 * starts happening, stop the previous chrono and start a new one.
2562 if (type > tp->chrono_type)
2563 tcp_chrono_set(tp, type);
2566 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2568 struct tcp_sock *tp = tcp_sk(sk);
2571 /* There are multiple conditions worthy of tracking in a
2572 * chronograph, so that the highest priority enum takes
2573 * precedence over the other conditions (see tcp_chrono_start).
2574 * If a condition stops, we only stop chrono tracking if
2575 * it's the "most interesting" or current chrono we are
2576 * tracking and starts busy chrono if we have pending data.
2578 if (tcp_rtx_and_write_queues_empty(sk))
2579 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2580 else if (type == tp->chrono_type)
2581 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2584 /* This routine writes packets to the network. It advances the
2585 * send_head. This happens as incoming acks open up the remote
2588 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2589 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2590 * account rare use of URG, this is not a big flaw.
2592 * Send at most one packet when push_one > 0. Temporarily ignore
2593 * cwnd limit to force at most one packet out when push_one == 2.
2595 * Returns true, if no segments are in flight and we have queued segments,
2596 * but cannot send anything now because of SWS or another problem.
2598 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2599 int push_one, gfp_t gfp)
2601 struct tcp_sock *tp = tcp_sk(sk);
2602 struct sk_buff *skb;
2603 unsigned int tso_segs, sent_pkts;
2606 bool is_cwnd_limited = false, is_rwnd_limited = false;
2611 tcp_mstamp_refresh(tp);
2613 /* Do MTU probing. */
2614 result = tcp_mtu_probe(sk);
2617 } else if (result > 0) {
2622 max_segs = tcp_tso_segs(sk, mss_now);
2623 while ((skb = tcp_send_head(sk))) {
2626 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2627 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2628 skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2629 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2630 tcp_init_tso_segs(skb, mss_now);
2631 goto repair; /* Skip network transmission */
2634 if (tcp_pacing_check(sk))
2637 tso_segs = tcp_init_tso_segs(skb, mss_now);
2640 cwnd_quota = tcp_cwnd_test(tp, skb);
2643 /* Force out a loss probe pkt. */
2649 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2650 is_rwnd_limited = true;
2654 if (tso_segs == 1) {
2655 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2656 (tcp_skb_is_last(sk, skb) ?
2657 nonagle : TCP_NAGLE_PUSH))))
2661 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2662 &is_rwnd_limited, max_segs))
2667 if (tso_segs > 1 && !tcp_urg_mode(tp))
2668 limit = tcp_mss_split_point(sk, skb, mss_now,
2674 if (skb->len > limit &&
2675 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2678 if (tcp_small_queue_check(sk, skb, 0))
2681 /* Argh, we hit an empty skb(), presumably a thread
2682 * is sleeping in sendmsg()/sk_stream_wait_memory().
2683 * We do not want to send a pure-ack packet and have
2684 * a strange looking rtx queue with empty packet(s).
2686 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2689 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2693 /* Advance the send_head. This one is sent out.
2694 * This call will increment packets_out.
2696 tcp_event_new_data_sent(sk, skb);
2698 tcp_minshall_update(tp, mss_now, skb);
2699 sent_pkts += tcp_skb_pcount(skb);
2705 if (is_rwnd_limited)
2706 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2708 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2710 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2711 if (likely(sent_pkts || is_cwnd_limited))
2712 tcp_cwnd_validate(sk, is_cwnd_limited);
2714 if (likely(sent_pkts)) {
2715 if (tcp_in_cwnd_reduction(sk))
2716 tp->prr_out += sent_pkts;
2718 /* Send one loss probe per tail loss episode. */
2720 tcp_schedule_loss_probe(sk, false);
2723 return !tp->packets_out && !tcp_write_queue_empty(sk);
2726 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2728 struct inet_connection_sock *icsk = inet_csk(sk);
2729 struct tcp_sock *tp = tcp_sk(sk);
2730 u32 timeout, rto_delta_us;
2733 /* Don't do any loss probe on a Fast Open connection before 3WHS
2736 if (rcu_access_pointer(tp->fastopen_rsk))
2739 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2740 /* Schedule a loss probe in 2*RTT for SACK capable connections
2741 * not in loss recovery, that are either limited by cwnd or application.
2743 if ((early_retrans != 3 && early_retrans != 4) ||
2744 !tp->packets_out || !tcp_is_sack(tp) ||
2745 (icsk->icsk_ca_state != TCP_CA_Open &&
2746 icsk->icsk_ca_state != TCP_CA_CWR))
2749 /* Probe timeout is 2*rtt. Add minimum RTO to account
2750 * for delayed ack when there's one outstanding packet. If no RTT
2751 * sample is available then probe after TCP_TIMEOUT_INIT.
2754 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2755 if (tp->packets_out == 1)
2756 timeout += TCP_RTO_MIN;
2758 timeout += TCP_TIMEOUT_MIN;
2760 timeout = TCP_TIMEOUT_INIT;
2763 /* If the RTO formula yields an earlier time, then use that time. */
2764 rto_delta_us = advancing_rto ?
2765 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2766 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2767 if (rto_delta_us > 0)
2768 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2770 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2774 /* Thanks to skb fast clones, we can detect if a prior transmit of
2775 * a packet is still in a qdisc or driver queue.
2776 * In this case, there is very little point doing a retransmit !
2778 static bool skb_still_in_host_queue(const struct sock *sk,
2779 const struct sk_buff *skb)
2781 if (unlikely(skb_fclone_busy(sk, skb))) {
2782 NET_INC_STATS(sock_net(sk),
2783 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2789 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2790 * retransmit the last segment.
2792 void tcp_send_loss_probe(struct sock *sk)
2794 struct tcp_sock *tp = tcp_sk(sk);
2795 struct sk_buff *skb;
2797 int mss = tcp_current_mss(sk);
2799 /* At most one outstanding TLP */
2800 if (tp->tlp_high_seq)
2803 tp->tlp_retrans = 0;
2804 skb = tcp_send_head(sk);
2805 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2806 pcount = tp->packets_out;
2807 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2808 if (tp->packets_out > pcount)
2812 skb = skb_rb_last(&sk->tcp_rtx_queue);
2813 if (unlikely(!skb)) {
2814 WARN_ONCE(tp->packets_out,
2815 "invalid inflight: %u state %u cwnd %u mss %d\n",
2816 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2817 inet_csk(sk)->icsk_pending = 0;
2821 if (skb_still_in_host_queue(sk, skb))
2824 pcount = tcp_skb_pcount(skb);
2825 if (WARN_ON(!pcount))
2828 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2829 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2830 (pcount - 1) * mss, mss,
2833 skb = skb_rb_next(skb);
2836 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2839 if (__tcp_retransmit_skb(sk, skb, 1))
2842 tp->tlp_retrans = 1;
2845 /* Record snd_nxt for loss detection. */
2846 tp->tlp_high_seq = tp->snd_nxt;
2848 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2849 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2850 inet_csk(sk)->icsk_pending = 0;
2855 /* Push out any pending frames which were held back due to
2856 * TCP_CORK or attempt at coalescing tiny packets.
2857 * The socket must be locked by the caller.
2859 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2862 /* If we are closed, the bytes will have to remain here.
2863 * In time closedown will finish, we empty the write queue and
2864 * all will be happy.
2866 if (unlikely(sk->sk_state == TCP_CLOSE))
2869 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2870 sk_gfp_mask(sk, GFP_ATOMIC)))
2871 tcp_check_probe_timer(sk);
2874 /* Send _single_ skb sitting at the send head. This function requires
2875 * true push pending frames to setup probe timer etc.
2877 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2879 struct sk_buff *skb = tcp_send_head(sk);
2881 BUG_ON(!skb || skb->len < mss_now);
2883 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2886 /* This function returns the amount that we can raise the
2887 * usable window based on the following constraints
2889 * 1. The window can never be shrunk once it is offered (RFC 793)
2890 * 2. We limit memory per socket
2893 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2894 * RECV.NEXT + RCV.WIN fixed until:
2895 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2897 * i.e. don't raise the right edge of the window until you can raise
2898 * it at least MSS bytes.
2900 * Unfortunately, the recommended algorithm breaks header prediction,
2901 * since header prediction assumes th->window stays fixed.
2903 * Strictly speaking, keeping th->window fixed violates the receiver
2904 * side SWS prevention criteria. The problem is that under this rule
2905 * a stream of single byte packets will cause the right side of the
2906 * window to always advance by a single byte.
2908 * Of course, if the sender implements sender side SWS prevention
2909 * then this will not be a problem.
2911 * BSD seems to make the following compromise:
2913 * If the free space is less than the 1/4 of the maximum
2914 * space available and the free space is less than 1/2 mss,
2915 * then set the window to 0.
2916 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2917 * Otherwise, just prevent the window from shrinking
2918 * and from being larger than the largest representable value.
2920 * This prevents incremental opening of the window in the regime
2921 * where TCP is limited by the speed of the reader side taking
2922 * data out of the TCP receive queue. It does nothing about
2923 * those cases where the window is constrained on the sender side
2924 * because the pipeline is full.
2926 * BSD also seems to "accidentally" limit itself to windows that are a
2927 * multiple of MSS, at least until the free space gets quite small.
2928 * This would appear to be a side effect of the mbuf implementation.
2929 * Combining these two algorithms results in the observed behavior
2930 * of having a fixed window size at almost all times.
2932 * Below we obtain similar behavior by forcing the offered window to
2933 * a multiple of the mss when it is feasible to do so.
2935 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2936 * Regular options like TIMESTAMP are taken into account.
2938 u32 __tcp_select_window(struct sock *sk)
2940 struct inet_connection_sock *icsk = inet_csk(sk);
2941 struct tcp_sock *tp = tcp_sk(sk);
2942 /* MSS for the peer's data. Previous versions used mss_clamp
2943 * here. I don't know if the value based on our guesses
2944 * of peer's MSS is better for the performance. It's more correct
2945 * but may be worse for the performance because of rcv_mss
2946 * fluctuations. --SAW 1998/11/1
2948 int mss = icsk->icsk_ack.rcv_mss;
2949 int free_space = tcp_space(sk);
2950 int allowed_space = tcp_full_space(sk);
2951 int full_space, window;
2953 if (sk_is_mptcp(sk))
2954 mptcp_space(sk, &free_space, &allowed_space);
2956 full_space = min_t(int, tp->window_clamp, allowed_space);
2958 if (unlikely(mss > full_space)) {
2963 if (free_space < (full_space >> 1)) {
2964 icsk->icsk_ack.quick = 0;
2966 if (tcp_under_memory_pressure(sk))
2967 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2970 /* free_space might become our new window, make sure we don't
2971 * increase it due to wscale.
2973 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2975 /* if free space is less than mss estimate, or is below 1/16th
2976 * of the maximum allowed, try to move to zero-window, else
2977 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2978 * new incoming data is dropped due to memory limits.
2979 * With large window, mss test triggers way too late in order
2980 * to announce zero window in time before rmem limit kicks in.
2982 if (free_space < (allowed_space >> 4) || free_space < mss)
2986 if (free_space > tp->rcv_ssthresh)
2987 free_space = tp->rcv_ssthresh;
2989 /* Don't do rounding if we are using window scaling, since the
2990 * scaled window will not line up with the MSS boundary anyway.
2992 if (tp->rx_opt.rcv_wscale) {
2993 window = free_space;
2995 /* Advertise enough space so that it won't get scaled away.
2996 * Import case: prevent zero window announcement if
2997 * 1<<rcv_wscale > mss.
2999 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
3001 window = tp->rcv_wnd;
3002 /* Get the largest window that is a nice multiple of mss.
3003 * Window clamp already applied above.
3004 * If our current window offering is within 1 mss of the
3005 * free space we just keep it. This prevents the divide
3006 * and multiply from happening most of the time.
3007 * We also don't do any window rounding when the free space
3010 if (window <= free_space - mss || window > free_space)
3011 window = rounddown(free_space, mss);
3012 else if (mss == full_space &&
3013 free_space > window + (full_space >> 1))
3014 window = free_space;
3020 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3021 const struct sk_buff *next_skb)
3023 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3024 const struct skb_shared_info *next_shinfo =
3025 skb_shinfo(next_skb);
3026 struct skb_shared_info *shinfo = skb_shinfo(skb);
3028 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3029 shinfo->tskey = next_shinfo->tskey;
3030 TCP_SKB_CB(skb)->txstamp_ack |=
3031 TCP_SKB_CB(next_skb)->txstamp_ack;
3035 /* Collapses two adjacent SKB's during retransmission. */
3036 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3038 struct tcp_sock *tp = tcp_sk(sk);
3039 struct sk_buff *next_skb = skb_rb_next(skb);
3042 next_skb_size = next_skb->len;
3044 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3046 if (next_skb_size) {
3047 if (next_skb_size <= skb_availroom(skb))
3048 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
3050 else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3053 tcp_highest_sack_replace(sk, next_skb, skb);
3055 /* Update sequence range on original skb. */
3056 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3058 /* Merge over control information. This moves PSH/FIN etc. over */
3059 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3061 /* All done, get rid of second SKB and account for it so
3062 * packet counting does not break.
3064 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3065 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3067 /* changed transmit queue under us so clear hints */
3068 tcp_clear_retrans_hints_partial(tp);
3069 if (next_skb == tp->retransmit_skb_hint)
3070 tp->retransmit_skb_hint = skb;
3072 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3074 tcp_skb_collapse_tstamp(skb, next_skb);
3076 tcp_rtx_queue_unlink_and_free(next_skb, sk);
3080 /* Check if coalescing SKBs is legal. */
3081 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3083 if (tcp_skb_pcount(skb) > 1)
3085 if (skb_cloned(skb))
3087 /* Some heuristics for collapsing over SACK'd could be invented */
3088 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3094 /* Collapse packets in the retransmit queue to make to create
3095 * less packets on the wire. This is only done on retransmission.
3097 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3100 struct tcp_sock *tp = tcp_sk(sk);
3101 struct sk_buff *skb = to, *tmp;
3104 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
3106 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3109 skb_rbtree_walk_from_safe(skb, tmp) {
3110 if (!tcp_can_collapse(sk, skb))
3113 if (!tcp_skb_can_collapse(to, skb))
3126 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3129 if (!tcp_collapse_retrans(sk, to))
3134 /* This retransmits one SKB. Policy decisions and retransmit queue
3135 * state updates are done by the caller. Returns non-zero if an
3136 * error occurred which prevented the send.
3138 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3140 struct inet_connection_sock *icsk = inet_csk(sk);
3141 struct tcp_sock *tp = tcp_sk(sk);
3142 unsigned int cur_mss;
3146 /* Inconclusive MTU probe */
3147 if (icsk->icsk_mtup.probe_size)
3148 icsk->icsk_mtup.probe_size = 0;
3150 /* Do not sent more than we queued. 1/4 is reserved for possible
3151 * copying overhead: fragmentation, tunneling, mangling etc.
3153 if (refcount_read(&sk->sk_wmem_alloc) >
3154 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
3158 if (skb_still_in_host_queue(sk, skb))
3161 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
3162 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3166 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3170 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3171 return -EHOSTUNREACH; /* Routing failure or similar. */
3173 cur_mss = tcp_current_mss(sk);
3175 /* If receiver has shrunk his window, and skb is out of
3176 * new window, do not retransmit it. The exception is the
3177 * case, when window is shrunk to zero. In this case
3178 * our retransmit serves as a zero window probe.
3180 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
3181 TCP_SKB_CB(skb)->seq != tp->snd_una)
3184 len = cur_mss * segs;
3185 if (skb->len > len) {
3186 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3187 cur_mss, GFP_ATOMIC))
3188 return -ENOMEM; /* We'll try again later. */
3190 if (skb_unclone(skb, GFP_ATOMIC))
3193 diff = tcp_skb_pcount(skb);
3194 tcp_set_skb_tso_segs(skb, cur_mss);
3195 diff -= tcp_skb_pcount(skb);
3197 tcp_adjust_pcount(sk, skb, diff);
3198 if (skb->len < cur_mss)
3199 tcp_retrans_try_collapse(sk, skb, cur_mss);
3202 /* RFC3168, section 6.1.1.1. ECN fallback */
3203 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3204 tcp_ecn_clear_syn(sk, skb);
3206 /* Update global and local TCP statistics. */
3207 segs = tcp_skb_pcount(skb);
3208 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3209 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3210 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3211 tp->total_retrans += segs;
3212 tp->bytes_retrans += skb->len;
3214 /* make sure skb->data is aligned on arches that require it
3215 * and check if ack-trimming & collapsing extended the headroom
3216 * beyond what csum_start can cover.
3218 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3219 skb_headroom(skb) >= 0xFFFF)) {
3220 struct sk_buff *nskb;
3222 tcp_skb_tsorted_save(skb) {
3223 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3226 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3230 } tcp_skb_tsorted_restore(skb);
3233 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3234 tcp_rate_skb_sent(sk, skb);
3237 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3240 /* To avoid taking spuriously low RTT samples based on a timestamp
3241 * for a transmit that never happened, always mark EVER_RETRANS
3243 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3245 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3246 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3247 TCP_SKB_CB(skb)->seq, segs, err);
3250 trace_tcp_retransmit_skb(sk, skb);
3251 } else if (err != -EBUSY) {
3252 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3257 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3259 struct tcp_sock *tp = tcp_sk(sk);
3260 int err = __tcp_retransmit_skb(sk, skb, segs);
3263 #if FASTRETRANS_DEBUG > 0
3264 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3265 net_dbg_ratelimited("retrans_out leaked\n");
3268 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3269 tp->retrans_out += tcp_skb_pcount(skb);
3272 /* Save stamp of the first (attempted) retransmit. */
3273 if (!tp->retrans_stamp)
3274 tp->retrans_stamp = tcp_skb_timestamp(skb);
3276 if (tp->undo_retrans < 0)
3277 tp->undo_retrans = 0;
3278 tp->undo_retrans += tcp_skb_pcount(skb);
3282 /* This gets called after a retransmit timeout, and the initially
3283 * retransmitted data is acknowledged. It tries to continue
3284 * resending the rest of the retransmit queue, until either
3285 * we've sent it all or the congestion window limit is reached.
3287 void tcp_xmit_retransmit_queue(struct sock *sk)
3289 const struct inet_connection_sock *icsk = inet_csk(sk);
3290 struct sk_buff *skb, *rtx_head, *hole = NULL;
3291 struct tcp_sock *tp = tcp_sk(sk);
3292 bool rearm_timer = false;
3296 if (!tp->packets_out)
3299 rtx_head = tcp_rtx_queue_head(sk);
3300 skb = tp->retransmit_skb_hint ?: rtx_head;
3301 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3302 skb_rbtree_walk_from(skb) {
3306 if (tcp_pacing_check(sk))
3309 /* we could do better than to assign each time */
3311 tp->retransmit_skb_hint = skb;
3313 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
3316 sacked = TCP_SKB_CB(skb)->sacked;
3317 /* In case tcp_shift_skb_data() have aggregated large skbs,
3318 * we need to make sure not sending too bigs TSO packets
3320 segs = min_t(int, segs, max_segs);
3322 if (tp->retrans_out >= tp->lost_out) {
3324 } else if (!(sacked & TCPCB_LOST)) {
3325 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3330 if (icsk->icsk_ca_state != TCP_CA_Loss)
3331 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3333 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3336 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3339 if (tcp_small_queue_check(sk, skb, 1))
3342 if (tcp_retransmit_skb(sk, skb, segs))
3345 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3347 if (tcp_in_cwnd_reduction(sk))
3348 tp->prr_out += tcp_skb_pcount(skb);
3350 if (skb == rtx_head &&
3351 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3356 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3357 inet_csk(sk)->icsk_rto,
3361 /* We allow to exceed memory limits for FIN packets to expedite
3362 * connection tear down and (memory) recovery.
3363 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3364 * or even be forced to close flow without any FIN.
3365 * In general, we want to allow one skb per socket to avoid hangs
3366 * with edge trigger epoll()
3368 void sk_forced_mem_schedule(struct sock *sk, int size)
3372 if (size <= sk->sk_forward_alloc)
3374 amt = sk_mem_pages(size);
3375 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3376 sk_memory_allocated_add(sk, amt);
3378 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3379 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3382 /* Send a FIN. The caller locks the socket for us.
3383 * We should try to send a FIN packet really hard, but eventually give up.
3385 void tcp_send_fin(struct sock *sk)
3387 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3388 struct tcp_sock *tp = tcp_sk(sk);
3390 /* Optimization, tack on the FIN if we have one skb in write queue and
3391 * this skb was not yet sent, or we are under memory pressure.
3392 * Note: in the latter case, FIN packet will be sent after a timeout,
3393 * as TCP stack thinks it has already been transmitted.
3396 if (!tskb && tcp_under_memory_pressure(sk))
3397 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3400 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3401 TCP_SKB_CB(tskb)->end_seq++;
3404 /* This means tskb was already sent.
3405 * Pretend we included the FIN on previous transmit.
3406 * We need to set tp->snd_nxt to the value it would have
3407 * if FIN had been sent. This is because retransmit path
3408 * does not change tp->snd_nxt.
3410 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3414 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3418 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3419 skb_reserve(skb, MAX_TCP_HEADER);
3420 sk_forced_mem_schedule(sk, skb->truesize);
3421 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3422 tcp_init_nondata_skb(skb, tp->write_seq,
3423 TCPHDR_ACK | TCPHDR_FIN);
3424 tcp_queue_skb(sk, skb);
3426 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3429 /* We get here when a process closes a file descriptor (either due to
3430 * an explicit close() or as a byproduct of exit()'ing) and there
3431 * was unread data in the receive queue. This behavior is recommended
3432 * by RFC 2525, section 2.17. -DaveM
3434 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3436 struct sk_buff *skb;
3438 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3440 /* NOTE: No TCP options attached and we never retransmit this. */
3441 skb = alloc_skb(MAX_TCP_HEADER, priority);
3443 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3447 /* Reserve space for headers and prepare control bits. */
3448 skb_reserve(skb, MAX_TCP_HEADER);
3449 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3450 TCPHDR_ACK | TCPHDR_RST);
3451 tcp_mstamp_refresh(tcp_sk(sk));
3453 if (tcp_transmit_skb(sk, skb, 0, priority))
3454 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3456 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3457 * skb here is different to the troublesome skb, so use NULL
3459 trace_tcp_send_reset(sk, NULL);
3462 /* Send a crossed SYN-ACK during socket establishment.
3463 * WARNING: This routine must only be called when we have already sent
3464 * a SYN packet that crossed the incoming SYN that caused this routine
3465 * to get called. If this assumption fails then the initial rcv_wnd
3466 * and rcv_wscale values will not be correct.
3468 int tcp_send_synack(struct sock *sk)
3470 struct sk_buff *skb;
3472 skb = tcp_rtx_queue_head(sk);
3473 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3474 pr_err("%s: wrong queue state\n", __func__);
3477 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3478 if (skb_cloned(skb)) {
3479 struct sk_buff *nskb;
3481 tcp_skb_tsorted_save(skb) {
3482 nskb = skb_copy(skb, GFP_ATOMIC);
3483 } tcp_skb_tsorted_restore(skb);
3486 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3487 tcp_highest_sack_replace(sk, skb, nskb);
3488 tcp_rtx_queue_unlink_and_free(skb, sk);
3489 __skb_header_release(nskb);
3490 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3491 sk_wmem_queued_add(sk, nskb->truesize);
3492 sk_mem_charge(sk, nskb->truesize);
3496 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3497 tcp_ecn_send_synack(sk, skb);
3499 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3503 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3504 * @sk: listener socket
3505 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3506 * should not use it again.
3507 * @req: request_sock pointer
3508 * @foc: cookie for tcp fast open
3509 * @synack_type: Type of synack to prepare
3510 * @syn_skb: SYN packet just received. It could be NULL for rtx case.
3512 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3513 struct request_sock *req,
3514 struct tcp_fastopen_cookie *foc,
3515 enum tcp_synack_type synack_type,
3516 struct sk_buff *syn_skb)
3518 struct inet_request_sock *ireq = inet_rsk(req);
3519 const struct tcp_sock *tp = tcp_sk(sk);
3520 struct tcp_md5sig_key *md5 = NULL;
3521 struct tcp_out_options opts;
3522 struct sk_buff *skb;
3523 int tcp_header_size;
3528 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3529 if (unlikely(!skb)) {
3533 /* Reserve space for headers. */
3534 skb_reserve(skb, MAX_TCP_HEADER);
3536 switch (synack_type) {
3537 case TCP_SYNACK_NORMAL:
3538 skb_set_owner_w(skb, req_to_sk(req));
3540 case TCP_SYNACK_COOKIE:
3541 /* Under synflood, we do not attach skb to a socket,
3542 * to avoid false sharing.
3545 case TCP_SYNACK_FASTOPEN:
3546 /* sk is a const pointer, because we want to express multiple
3547 * cpu might call us concurrently.
3548 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3550 skb_set_owner_w(skb, (struct sock *)sk);
3553 skb_dst_set(skb, dst);
3555 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3557 memset(&opts, 0, sizeof(opts));
3558 now = tcp_clock_ns();
3559 #ifdef CONFIG_SYN_COOKIES
3560 if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3561 skb->skb_mstamp_ns = cookie_init_timestamp(req, now);
3565 skb->skb_mstamp_ns = now;
3566 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3567 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3570 #ifdef CONFIG_TCP_MD5SIG
3572 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3574 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3575 /* bpf program will be interested in the tcp_flags */
3576 TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3577 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3579 syn_skb) + sizeof(*th);
3581 skb_push(skb, tcp_header_size);
3582 skb_reset_transport_header(skb);
3584 th = (struct tcphdr *)skb->data;
3585 memset(th, 0, sizeof(struct tcphdr));
3588 tcp_ecn_make_synack(req, th);
3589 th->source = htons(ireq->ir_num);
3590 th->dest = ireq->ir_rmt_port;
3591 skb->mark = ireq->ir_mark;
3592 skb->ip_summed = CHECKSUM_PARTIAL;
3593 th->seq = htonl(tcp_rsk(req)->snt_isn);
3594 /* XXX data is queued and acked as is. No buffer/window check */
3595 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3597 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3598 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3599 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3600 th->doff = (tcp_header_size >> 2);
3601 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3603 #ifdef CONFIG_TCP_MD5SIG
3604 /* Okay, we have all we need - do the md5 hash if needed */
3606 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3607 md5, req_to_sk(req), skb);
3611 bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
3612 synack_type, &opts);
3614 skb->skb_mstamp_ns = now;
3615 tcp_add_tx_delay(skb, tp);
3619 EXPORT_SYMBOL(tcp_make_synack);
3621 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3623 struct inet_connection_sock *icsk = inet_csk(sk);
3624 const struct tcp_congestion_ops *ca;
3625 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3627 if (ca_key == TCP_CA_UNSPEC)
3631 ca = tcp_ca_find_key(ca_key);
3632 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3633 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3634 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3635 icsk->icsk_ca_ops = ca;
3640 /* Do all connect socket setups that can be done AF independent. */
3641 static void tcp_connect_init(struct sock *sk)
3643 const struct dst_entry *dst = __sk_dst_get(sk);
3644 struct tcp_sock *tp = tcp_sk(sk);
3648 /* We'll fix this up when we get a response from the other end.
3649 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3651 tp->tcp_header_len = sizeof(struct tcphdr);
3652 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3653 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3655 #ifdef CONFIG_TCP_MD5SIG
3656 if (tp->af_specific->md5_lookup(sk, sk))
3657 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3660 /* If user gave his TCP_MAXSEG, record it to clamp */
3661 if (tp->rx_opt.user_mss)
3662 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3665 tcp_sync_mss(sk, dst_mtu(dst));
3667 tcp_ca_dst_init(sk, dst);
3669 if (!tp->window_clamp)
3670 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3671 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3673 tcp_initialize_rcv_mss(sk);
3675 /* limit the window selection if the user enforce a smaller rx buffer */
3676 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3677 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3678 tp->window_clamp = tcp_full_space(sk);
3680 rcv_wnd = tcp_rwnd_init_bpf(sk);
3682 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3684 tcp_select_initial_window(sk, tcp_full_space(sk),
3685 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3688 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3692 tp->rx_opt.rcv_wscale = rcv_wscale;
3693 tp->rcv_ssthresh = tp->rcv_wnd;
3696 sock_reset_flag(sk, SOCK_DONE);
3699 tcp_write_queue_purge(sk);
3700 tp->snd_una = tp->write_seq;
3701 tp->snd_sml = tp->write_seq;
3702 tp->snd_up = tp->write_seq;
3703 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3705 if (likely(!tp->repair))
3708 tp->rcv_tstamp = tcp_jiffies32;
3709 tp->rcv_wup = tp->rcv_nxt;
3710 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3712 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3713 inet_csk(sk)->icsk_retransmits = 0;
3714 tcp_clear_retrans(tp);
3717 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3719 struct tcp_sock *tp = tcp_sk(sk);
3720 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3722 tcb->end_seq += skb->len;
3723 __skb_header_release(skb);
3724 sk_wmem_queued_add(sk, skb->truesize);
3725 sk_mem_charge(sk, skb->truesize);
3726 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3727 tp->packets_out += tcp_skb_pcount(skb);
3730 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3731 * queue a data-only packet after the regular SYN, such that regular SYNs
3732 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3733 * only the SYN sequence, the data are retransmitted in the first ACK.
3734 * If cookie is not cached or other error occurs, falls back to send a
3735 * regular SYN with Fast Open cookie request option.
3737 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3739 struct tcp_sock *tp = tcp_sk(sk);
3740 struct tcp_fastopen_request *fo = tp->fastopen_req;
3742 struct sk_buff *syn_data;
3744 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3745 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3748 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3749 * user-MSS. Reserve maximum option space for middleboxes that add
3750 * private TCP options. The cost is reduced data space in SYN :(
3752 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3754 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3755 MAX_TCP_OPTION_SPACE;
3757 space = min_t(size_t, space, fo->size);
3759 /* limit to order-0 allocations */
3760 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3762 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3765 syn_data->ip_summed = CHECKSUM_PARTIAL;
3766 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3768 int copied = copy_from_iter(skb_put(syn_data, space), space,
3769 &fo->data->msg_iter);
3770 if (unlikely(!copied)) {
3771 tcp_skb_tsorted_anchor_cleanup(syn_data);
3772 kfree_skb(syn_data);
3775 if (copied != space) {
3776 skb_trim(syn_data, copied);
3779 skb_zcopy_set(syn_data, fo->uarg, NULL);
3781 /* No more data pending in inet_wait_for_connect() */
3782 if (space == fo->size)
3786 tcp_connect_queue_skb(sk, syn_data);
3788 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3790 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3792 syn->skb_mstamp_ns = syn_data->skb_mstamp_ns;
3794 /* Now full SYN+DATA was cloned and sent (or not),
3795 * remove the SYN from the original skb (syn_data)
3796 * we keep in write queue in case of a retransmit, as we
3797 * also have the SYN packet (with no data) in the same queue.
3799 TCP_SKB_CB(syn_data)->seq++;
3800 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3802 tp->syn_data = (fo->copied > 0);
3803 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3804 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3808 /* data was not sent, put it in write_queue */
3809 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3810 tp->packets_out -= tcp_skb_pcount(syn_data);
3813 /* Send a regular SYN with Fast Open cookie request option */
3814 if (fo->cookie.len > 0)
3816 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3818 tp->syn_fastopen = 0;
3820 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3824 /* Build a SYN and send it off. */
3825 int tcp_connect(struct sock *sk)
3827 struct tcp_sock *tp = tcp_sk(sk);
3828 struct sk_buff *buff;
3831 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3833 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3834 return -EHOSTUNREACH; /* Routing failure or similar. */
3836 tcp_connect_init(sk);
3838 if (unlikely(tp->repair)) {
3839 tcp_finish_connect(sk, NULL);
3843 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3844 if (unlikely(!buff))
3847 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3848 tcp_mstamp_refresh(tp);
3849 tp->retrans_stamp = tcp_time_stamp(tp);
3850 tcp_connect_queue_skb(sk, buff);
3851 tcp_ecn_send_syn(sk, buff);
3852 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3854 /* Send off SYN; include data in Fast Open. */
3855 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3856 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3857 if (err == -ECONNREFUSED)
3860 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3861 * in order to make this packet get counted in tcpOutSegs.
3863 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3864 tp->pushed_seq = tp->write_seq;
3865 buff = tcp_send_head(sk);
3866 if (unlikely(buff)) {
3867 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3868 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3870 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3872 /* Timer for repeating the SYN until an answer. */
3873 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3874 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3877 EXPORT_SYMBOL(tcp_connect);
3879 /* Send out a delayed ack, the caller does the policy checking
3880 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3883 void tcp_send_delayed_ack(struct sock *sk)
3885 struct inet_connection_sock *icsk = inet_csk(sk);
3886 int ato = icsk->icsk_ack.ato;
3887 unsigned long timeout;
3889 if (ato > TCP_DELACK_MIN) {
3890 const struct tcp_sock *tp = tcp_sk(sk);
3891 int max_ato = HZ / 2;
3893 if (inet_csk_in_pingpong_mode(sk) ||
3894 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3895 max_ato = TCP_DELACK_MAX;
3897 /* Slow path, intersegment interval is "high". */
3899 /* If some rtt estimate is known, use it to bound delayed ack.
3900 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3904 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3911 ato = min(ato, max_ato);
3914 ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max);
3916 /* Stay within the limit we were given */
3917 timeout = jiffies + ato;
3919 /* Use new timeout only if there wasn't a older one earlier. */
3920 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3921 /* If delack timer is about to expire, send ACK now. */
3922 if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3927 if (!time_before(timeout, icsk->icsk_ack.timeout))
3928 timeout = icsk->icsk_ack.timeout;
3930 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3931 icsk->icsk_ack.timeout = timeout;
3932 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3935 /* This routine sends an ack and also updates the window. */
3936 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3938 struct sk_buff *buff;
3940 /* If we have been reset, we may not send again. */
3941 if (sk->sk_state == TCP_CLOSE)
3944 /* We are not putting this on the write queue, so
3945 * tcp_transmit_skb() will set the ownership to this
3948 buff = alloc_skb(MAX_TCP_HEADER,
3949 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3950 if (unlikely(!buff)) {
3951 struct inet_connection_sock *icsk = inet_csk(sk);
3952 unsigned long delay;
3954 delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
3955 if (delay < TCP_RTO_MAX)
3956 icsk->icsk_ack.retry++;
3957 inet_csk_schedule_ack(sk);
3958 icsk->icsk_ack.ato = TCP_ATO_MIN;
3959 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
3963 /* Reserve space for headers and prepare control bits. */
3964 skb_reserve(buff, MAX_TCP_HEADER);
3965 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3967 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3969 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3971 skb_set_tcp_pure_ack(buff);
3973 /* Send it off, this clears delayed acks for us. */
3974 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3976 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3978 void tcp_send_ack(struct sock *sk)
3980 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3983 /* This routine sends a packet with an out of date sequence
3984 * number. It assumes the other end will try to ack it.
3986 * Question: what should we make while urgent mode?
3987 * 4.4BSD forces sending single byte of data. We cannot send
3988 * out of window data, because we have SND.NXT==SND.MAX...
3990 * Current solution: to send TWO zero-length segments in urgent mode:
3991 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3992 * out-of-date with SND.UNA-1 to probe window.
3994 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3996 struct tcp_sock *tp = tcp_sk(sk);
3997 struct sk_buff *skb;
3999 /* We don't queue it, tcp_transmit_skb() sets ownership. */
4000 skb = alloc_skb(MAX_TCP_HEADER,
4001 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4005 /* Reserve space for headers and set control bits. */
4006 skb_reserve(skb, MAX_TCP_HEADER);
4007 /* Use a previous sequence. This should cause the other
4008 * end to send an ack. Don't queue or clone SKB, just
4011 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
4012 NET_INC_STATS(sock_net(sk), mib);
4013 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4016 /* Called from setsockopt( ... TCP_REPAIR ) */
4017 void tcp_send_window_probe(struct sock *sk)
4019 if (sk->sk_state == TCP_ESTABLISHED) {
4020 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4021 tcp_mstamp_refresh(tcp_sk(sk));
4022 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4026 /* Initiate keepalive or window probe from timer. */
4027 int tcp_write_wakeup(struct sock *sk, int mib)
4029 struct tcp_sock *tp = tcp_sk(sk);
4030 struct sk_buff *skb;
4032 if (sk->sk_state == TCP_CLOSE)
4035 skb = tcp_send_head(sk);
4036 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4038 unsigned int mss = tcp_current_mss(sk);
4039 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4041 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4042 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4044 /* We are probing the opening of a window
4045 * but the window size is != 0
4046 * must have been a result SWS avoidance ( sender )
4048 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4050 seg_size = min(seg_size, mss);
4051 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4052 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4053 skb, seg_size, mss, GFP_ATOMIC))
4055 } else if (!tcp_skb_pcount(skb))
4056 tcp_set_skb_tso_segs(skb, mss);
4058 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4059 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4061 tcp_event_new_data_sent(sk, skb);
4064 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4065 tcp_xmit_probe_skb(sk, 1, mib);
4066 return tcp_xmit_probe_skb(sk, 0, mib);
4070 /* A window probe timeout has occurred. If window is not closed send
4071 * a partial packet else a zero probe.
4073 void tcp_send_probe0(struct sock *sk)
4075 struct inet_connection_sock *icsk = inet_csk(sk);
4076 struct tcp_sock *tp = tcp_sk(sk);
4077 struct net *net = sock_net(sk);
4078 unsigned long timeout;
4081 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4083 if (tp->packets_out || tcp_write_queue_empty(sk)) {
4084 /* Cancel probe timer, if it is not required. */
4085 icsk->icsk_probes_out = 0;
4086 icsk->icsk_backoff = 0;
4087 icsk->icsk_probes_tstamp = 0;
4091 icsk->icsk_probes_out++;
4093 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
4094 icsk->icsk_backoff++;
4095 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
4097 /* If packet was not sent due to local congestion,
4098 * Let senders fight for local resources conservatively.
4100 timeout = TCP_RESOURCE_PROBE_INTERVAL;
4102 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
4105 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4107 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4111 tcp_rsk(req)->txhash = net_tx_rndhash();
4112 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4115 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4116 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4117 if (unlikely(tcp_passive_fastopen(sk)))
4118 tcp_sk(sk)->total_retrans++;
4119 trace_tcp_retransmit_synack(sk, req);
4123 EXPORT_SYMBOL(tcp_rtx_synack);