1 // SPDX-License-Identifier: GPL-2.0-only
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
38 #define pr_fmt(fmt) "TCP: " fmt
41 #include <net/mptcp.h>
43 #include <linux/compiler.h>
44 #include <linux/gfp.h>
45 #include <linux/module.h>
46 #include <linux/static_key.h>
48 #include <trace/events/tcp.h>
50 /* Refresh clocks of a TCP socket,
51 * ensuring monotically increasing values.
53 void tcp_mstamp_refresh(struct tcp_sock *tp)
55 u64 val = tcp_clock_ns();
57 tp->tcp_clock_cache = val;
58 tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
61 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
62 int push_one, gfp_t gfp);
64 /* Account for new data that has been sent to the network. */
65 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
67 struct inet_connection_sock *icsk = inet_csk(sk);
68 struct tcp_sock *tp = tcp_sk(sk);
69 unsigned int prior_packets = tp->packets_out;
71 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
73 __skb_unlink(skb, &sk->sk_write_queue);
74 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
76 if (tp->highest_sack == NULL)
77 tp->highest_sack = skb;
79 tp->packets_out += tcp_skb_pcount(skb);
80 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
83 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
88 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
89 * window scaling factor due to loss of precision.
90 * If window has been shrunk, what should we make? It is not clear at all.
91 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
92 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
93 * invalid. OK, let's make this for now:
95 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
97 const struct tcp_sock *tp = tcp_sk(sk);
99 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
100 (tp->rx_opt.wscale_ok &&
101 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
104 return tcp_wnd_end(tp);
107 /* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
121 static __u16 tcp_advertise_mss(struct sock *sk)
123 struct tcp_sock *tp = tcp_sk(sk);
124 const struct dst_entry *dst = __sk_dst_get(sk);
125 int mss = tp->advmss;
128 unsigned int metric = dst_metric_advmss(dst);
139 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism.
142 void tcp_cwnd_restart(struct sock *sk, s32 delta)
144 struct tcp_sock *tp = tcp_sk(sk);
145 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
146 u32 cwnd = tcp_snd_cwnd(tp);
148 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
150 tp->snd_ssthresh = tcp_current_ssthresh(sk);
151 restart_cwnd = min(restart_cwnd, cwnd);
153 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
155 tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd));
156 tp->snd_cwnd_stamp = tcp_jiffies32;
157 tp->snd_cwnd_used = 0;
160 /* Congestion state accounting after a packet has been sent. */
161 static void tcp_event_data_sent(struct tcp_sock *tp,
164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const u32 now = tcp_jiffies32;
167 if (tcp_packets_in_flight(tp) == 0)
168 tcp_ca_event(sk, CA_EVENT_TX_START);
172 /* If it is a reply for ato after last received
173 * packet, enter pingpong mode.
175 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
176 inet_csk_enter_pingpong_mode(sk);
179 /* Account for an ACK we sent. */
180 static inline void tcp_event_ack_sent(struct sock *sk, u32 rcv_nxt)
182 struct tcp_sock *tp = tcp_sk(sk);
184 if (unlikely(tp->compressed_ack)) {
185 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
187 tp->compressed_ack = 0;
188 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
192 if (unlikely(rcv_nxt != tp->rcv_nxt))
193 return; /* Special ACK sent by DCTCP to reflect ECN */
194 tcp_dec_quickack_mode(sk);
195 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
198 /* Determine a window scaling and initial window to offer.
199 * Based on the assumption that the given amount of space
200 * will be offered. Store the results in the tp structure.
201 * NOTE: for smooth operation initial space offering should
202 * be a multiple of mss if possible. We assume here that mss >= 1.
203 * This MUST be enforced by all callers.
205 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
206 __u32 *rcv_wnd, __u32 *window_clamp,
207 int wscale_ok, __u8 *rcv_wscale,
210 unsigned int space = (__space < 0 ? 0 : __space);
212 /* If no clamp set the clamp to the max possible scaled window */
213 if (*window_clamp == 0)
214 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
215 space = min(*window_clamp, space);
217 /* Quantize space offering to a multiple of mss if possible. */
219 space = rounddown(space, mss);
221 /* NOTE: offering an initial window larger than 32767
222 * will break some buggy TCP stacks. If the admin tells us
223 * it is likely we could be speaking with such a buggy stack
224 * we will truncate our initial window offering to 32K-1
225 * unless the remote has sent us a window scaling option,
226 * which we interpret as a sign the remote TCP is not
227 * misinterpreting the window field as a signed quantity.
229 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
230 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
232 (*rcv_wnd) = min_t(u32, space, U16_MAX);
235 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
239 /* Set window scaling on max possible window */
240 space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
241 space = max_t(u32, space, READ_ONCE(sysctl_rmem_max));
242 space = min_t(u32, space, *window_clamp);
243 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
246 /* Set the clamp no higher than max representable value */
247 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
249 EXPORT_SYMBOL(tcp_select_initial_window);
251 /* Chose a new window to advertise, update state in tcp_sock for the
252 * socket, and return result with RFC1323 scaling applied. The return
253 * value can be stuffed directly into th->window for an outgoing
256 static u16 tcp_select_window(struct sock *sk)
258 struct tcp_sock *tp = tcp_sk(sk);
259 u32 old_win = tp->rcv_wnd;
260 u32 cur_win = tcp_receive_window(tp);
261 u32 new_win = __tcp_select_window(sk);
263 /* Never shrink the offered window */
264 if (new_win < cur_win) {
265 /* Danger Will Robinson!
266 * Don't update rcv_wup/rcv_wnd here or else
267 * we will not be able to advertise a zero
268 * window in time. --DaveM
270 * Relax Will Robinson.
273 NET_INC_STATS(sock_net(sk),
274 LINUX_MIB_TCPWANTZEROWINDOWADV);
275 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
277 tp->rcv_wnd = new_win;
278 tp->rcv_wup = tp->rcv_nxt;
280 /* Make sure we do not exceed the maximum possible
283 if (!tp->rx_opt.rcv_wscale &&
284 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
285 new_win = min(new_win, MAX_TCP_WINDOW);
287 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
289 /* RFC1323 scaling applied */
290 new_win >>= tp->rx_opt.rcv_wscale;
292 /* If we advertise zero window, disable fast path. */
296 NET_INC_STATS(sock_net(sk),
297 LINUX_MIB_TCPTOZEROWINDOWADV);
298 } else if (old_win == 0) {
299 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
305 /* Packet ECN state for a SYN-ACK */
306 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
308 const struct tcp_sock *tp = tcp_sk(sk);
310 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
311 if (!(tp->ecn_flags & TCP_ECN_OK))
312 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
313 else if (tcp_ca_needs_ecn(sk) ||
314 tcp_bpf_ca_needs_ecn(sk))
318 /* Packet ECN state for a SYN. */
319 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
321 struct tcp_sock *tp = tcp_sk(sk);
322 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
323 bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 ||
324 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
327 const struct dst_entry *dst = __sk_dst_get(sk);
329 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
336 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
337 tp->ecn_flags = TCP_ECN_OK;
338 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
343 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
345 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback))
346 /* tp->ecn_flags are cleared at a later point in time when
347 * SYN ACK is ultimatively being received.
349 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
353 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
355 if (inet_rsk(req)->ecn_ok)
359 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
362 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
363 struct tcphdr *th, int tcp_header_len)
365 struct tcp_sock *tp = tcp_sk(sk);
367 if (tp->ecn_flags & TCP_ECN_OK) {
368 /* Not-retransmitted data segment: set ECT and inject CWR. */
369 if (skb->len != tcp_header_len &&
370 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
372 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
373 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
375 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
377 } else if (!tcp_ca_needs_ecn(sk)) {
378 /* ACK or retransmitted segment: clear ECT|CE */
379 INET_ECN_dontxmit(sk);
381 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
386 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
387 * auto increment end seqno.
389 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
391 skb->ip_summed = CHECKSUM_PARTIAL;
393 TCP_SKB_CB(skb)->tcp_flags = flags;
394 TCP_SKB_CB(skb)->sacked = 0;
396 tcp_skb_pcount_set(skb, 1);
398 TCP_SKB_CB(skb)->seq = seq;
399 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
401 TCP_SKB_CB(skb)->end_seq = seq;
404 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
406 return tp->snd_una != tp->snd_up;
409 #define OPTION_SACK_ADVERTISE (1 << 0)
410 #define OPTION_TS (1 << 1)
411 #define OPTION_MD5 (1 << 2)
412 #define OPTION_WSCALE (1 << 3)
413 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
414 #define OPTION_SMC (1 << 9)
415 #define OPTION_MPTCP (1 << 10)
417 static void smc_options_write(__be32 *ptr, u16 *options)
419 #if IS_ENABLED(CONFIG_SMC)
420 if (static_branch_unlikely(&tcp_have_smc)) {
421 if (unlikely(OPTION_SMC & *options)) {
422 *ptr++ = htonl((TCPOPT_NOP << 24) |
425 (TCPOLEN_EXP_SMC_BASE));
426 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
432 struct tcp_out_options {
433 u16 options; /* bit field of OPTION_* */
434 u16 mss; /* 0 to disable */
435 u8 ws; /* window scale, 0 to disable */
436 u8 num_sack_blocks; /* number of SACK blocks to include */
437 u8 hash_size; /* bytes in hash_location */
438 u8 bpf_opt_len; /* length of BPF hdr option */
439 __u8 *hash_location; /* temporary pointer, overloaded */
440 __u32 tsval, tsecr; /* need to include OPTION_TS */
441 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
442 struct mptcp_out_options mptcp;
445 static void mptcp_options_write(__be32 *ptr, const struct tcp_sock *tp,
446 struct tcp_out_options *opts)
448 #if IS_ENABLED(CONFIG_MPTCP)
449 if (unlikely(OPTION_MPTCP & opts->options))
450 mptcp_write_options(ptr, tp, &opts->mptcp);
454 #ifdef CONFIG_CGROUP_BPF
455 static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
456 enum tcp_synack_type synack_type)
459 return BPF_WRITE_HDR_TCP_CURRENT_MSS;
461 if (unlikely(synack_type == TCP_SYNACK_COOKIE))
462 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
467 /* req, syn_skb and synack_type are used when writing synack */
468 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
469 struct request_sock *req,
470 struct sk_buff *syn_skb,
471 enum tcp_synack_type synack_type,
472 struct tcp_out_options *opts,
473 unsigned int *remaining)
475 struct bpf_sock_ops_kern sock_ops;
478 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
479 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
483 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
486 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
488 sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
491 /* The listen "sk" cannot be passed here because
492 * it is not locked. It would not make too much
493 * sense to do bpf_setsockopt(listen_sk) based
494 * on individual connection request also.
496 * Thus, "req" is passed here and the cgroup-bpf-progs
497 * of the listen "sk" will be run.
499 * "req" is also used here for fastopen even the "sk" here is
500 * a fullsock "child" sk. It is to keep the behavior
501 * consistent between fastopen and non-fastopen on
502 * the bpf programming side.
504 sock_ops.sk = (struct sock *)req;
505 sock_ops.syn_skb = syn_skb;
507 sock_owned_by_me(sk);
509 sock_ops.is_fullsock = 1;
513 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
514 sock_ops.remaining_opt_len = *remaining;
515 /* tcp_current_mss() does not pass a skb */
517 bpf_skops_init_skb(&sock_ops, skb, 0);
519 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
521 if (err || sock_ops.remaining_opt_len == *remaining)
524 opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
525 /* round up to 4 bytes */
526 opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
528 *remaining -= opts->bpf_opt_len;
531 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
532 struct request_sock *req,
533 struct sk_buff *syn_skb,
534 enum tcp_synack_type synack_type,
535 struct tcp_out_options *opts)
537 u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
538 struct bpf_sock_ops_kern sock_ops;
541 if (likely(!max_opt_len))
544 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
546 sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
549 sock_ops.sk = (struct sock *)req;
550 sock_ops.syn_skb = syn_skb;
552 sock_owned_by_me(sk);
554 sock_ops.is_fullsock = 1;
558 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
559 sock_ops.remaining_opt_len = max_opt_len;
560 first_opt_off = tcp_hdrlen(skb) - max_opt_len;
561 bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
563 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
568 nr_written = max_opt_len - sock_ops.remaining_opt_len;
570 if (nr_written < max_opt_len)
571 memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
572 max_opt_len - nr_written);
575 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
576 struct request_sock *req,
577 struct sk_buff *syn_skb,
578 enum tcp_synack_type synack_type,
579 struct tcp_out_options *opts,
580 unsigned int *remaining)
584 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
585 struct request_sock *req,
586 struct sk_buff *syn_skb,
587 enum tcp_synack_type synack_type,
588 struct tcp_out_options *opts)
593 /* Write previously computed TCP options to the packet.
595 * Beware: Something in the Internet is very sensitive to the ordering of
596 * TCP options, we learned this through the hard way, so be careful here.
597 * Luckily we can at least blame others for their non-compliance but from
598 * inter-operability perspective it seems that we're somewhat stuck with
599 * the ordering which we have been using if we want to keep working with
600 * those broken things (not that it currently hurts anybody as there isn't
601 * particular reason why the ordering would need to be changed).
603 * At least SACK_PERM as the first option is known to lead to a disaster
604 * (but it may well be that other scenarios fail similarly).
606 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
607 struct tcp_out_options *opts)
609 u16 options = opts->options; /* mungable copy */
611 if (unlikely(OPTION_MD5 & options)) {
612 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
613 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
614 /* overload cookie hash location */
615 opts->hash_location = (__u8 *)ptr;
619 if (unlikely(opts->mss)) {
620 *ptr++ = htonl((TCPOPT_MSS << 24) |
621 (TCPOLEN_MSS << 16) |
625 if (likely(OPTION_TS & options)) {
626 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
627 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
628 (TCPOLEN_SACK_PERM << 16) |
629 (TCPOPT_TIMESTAMP << 8) |
631 options &= ~OPTION_SACK_ADVERTISE;
633 *ptr++ = htonl((TCPOPT_NOP << 24) |
635 (TCPOPT_TIMESTAMP << 8) |
638 *ptr++ = htonl(opts->tsval);
639 *ptr++ = htonl(opts->tsecr);
642 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
643 *ptr++ = htonl((TCPOPT_NOP << 24) |
645 (TCPOPT_SACK_PERM << 8) |
649 if (unlikely(OPTION_WSCALE & options)) {
650 *ptr++ = htonl((TCPOPT_NOP << 24) |
651 (TCPOPT_WINDOW << 16) |
652 (TCPOLEN_WINDOW << 8) |
656 if (unlikely(opts->num_sack_blocks)) {
657 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
658 tp->duplicate_sack : tp->selective_acks;
661 *ptr++ = htonl((TCPOPT_NOP << 24) |
664 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
665 TCPOLEN_SACK_PERBLOCK)));
667 for (this_sack = 0; this_sack < opts->num_sack_blocks;
669 *ptr++ = htonl(sp[this_sack].start_seq);
670 *ptr++ = htonl(sp[this_sack].end_seq);
673 tp->rx_opt.dsack = 0;
676 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
677 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
679 u32 len; /* Fast Open option length */
682 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
683 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
684 TCPOPT_FASTOPEN_MAGIC);
685 p += TCPOLEN_EXP_FASTOPEN_BASE;
687 len = TCPOLEN_FASTOPEN_BASE + foc->len;
688 *p++ = TCPOPT_FASTOPEN;
692 memcpy(p, foc->val, foc->len);
693 if ((len & 3) == 2) {
694 p[foc->len] = TCPOPT_NOP;
695 p[foc->len + 1] = TCPOPT_NOP;
697 ptr += (len + 3) >> 2;
700 smc_options_write(ptr, &options);
702 mptcp_options_write(ptr, tp, opts);
705 static void smc_set_option(const struct tcp_sock *tp,
706 struct tcp_out_options *opts,
707 unsigned int *remaining)
709 #if IS_ENABLED(CONFIG_SMC)
710 if (static_branch_unlikely(&tcp_have_smc)) {
712 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
713 opts->options |= OPTION_SMC;
714 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
721 static void smc_set_option_cond(const struct tcp_sock *tp,
722 const struct inet_request_sock *ireq,
723 struct tcp_out_options *opts,
724 unsigned int *remaining)
726 #if IS_ENABLED(CONFIG_SMC)
727 if (static_branch_unlikely(&tcp_have_smc)) {
728 if (tp->syn_smc && ireq->smc_ok) {
729 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
730 opts->options |= OPTION_SMC;
731 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
738 static void mptcp_set_option_cond(const struct request_sock *req,
739 struct tcp_out_options *opts,
740 unsigned int *remaining)
742 if (rsk_is_mptcp(req)) {
745 if (mptcp_synack_options(req, &size, &opts->mptcp)) {
746 if (*remaining >= size) {
747 opts->options |= OPTION_MPTCP;
754 /* Compute TCP options for SYN packets. This is not the final
755 * network wire format yet.
757 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
758 struct tcp_out_options *opts,
759 struct tcp_md5sig_key **md5)
761 struct tcp_sock *tp = tcp_sk(sk);
762 unsigned int remaining = MAX_TCP_OPTION_SPACE;
763 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
766 #ifdef CONFIG_TCP_MD5SIG
767 if (static_branch_unlikely(&tcp_md5_needed) &&
768 rcu_access_pointer(tp->md5sig_info)) {
769 *md5 = tp->af_specific->md5_lookup(sk, sk);
771 opts->options |= OPTION_MD5;
772 remaining -= TCPOLEN_MD5SIG_ALIGNED;
777 /* We always get an MSS option. The option bytes which will be seen in
778 * normal data packets should timestamps be used, must be in the MSS
779 * advertised. But we subtract them from tp->mss_cache so that
780 * calculations in tcp_sendmsg are simpler etc. So account for this
781 * fact here if necessary. If we don't do this correctly, as a
782 * receiver we won't recognize data packets as being full sized when we
783 * should, and thus we won't abide by the delayed ACK rules correctly.
784 * SACKs don't matter, we never delay an ACK when we have any of those
786 opts->mss = tcp_advertise_mss(sk);
787 remaining -= TCPOLEN_MSS_ALIGNED;
789 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps) && !*md5)) {
790 opts->options |= OPTION_TS;
791 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
792 opts->tsecr = tp->rx_opt.ts_recent;
793 remaining -= TCPOLEN_TSTAMP_ALIGNED;
795 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) {
796 opts->ws = tp->rx_opt.rcv_wscale;
797 opts->options |= OPTION_WSCALE;
798 remaining -= TCPOLEN_WSCALE_ALIGNED;
800 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) {
801 opts->options |= OPTION_SACK_ADVERTISE;
802 if (unlikely(!(OPTION_TS & opts->options)))
803 remaining -= TCPOLEN_SACKPERM_ALIGNED;
806 if (fastopen && fastopen->cookie.len >= 0) {
807 u32 need = fastopen->cookie.len;
809 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
810 TCPOLEN_FASTOPEN_BASE;
811 need = (need + 3) & ~3U; /* Align to 32 bits */
812 if (remaining >= need) {
813 opts->options |= OPTION_FAST_OPEN_COOKIE;
814 opts->fastopen_cookie = &fastopen->cookie;
816 tp->syn_fastopen = 1;
817 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
821 smc_set_option(tp, opts, &remaining);
823 if (sk_is_mptcp(sk)) {
826 if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
827 opts->options |= OPTION_MPTCP;
832 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
834 return MAX_TCP_OPTION_SPACE - remaining;
837 /* Set up TCP options for SYN-ACKs. */
838 static unsigned int tcp_synack_options(const struct sock *sk,
839 struct request_sock *req,
840 unsigned int mss, struct sk_buff *skb,
841 struct tcp_out_options *opts,
842 const struct tcp_md5sig_key *md5,
843 struct tcp_fastopen_cookie *foc,
844 enum tcp_synack_type synack_type,
845 struct sk_buff *syn_skb)
847 struct inet_request_sock *ireq = inet_rsk(req);
848 unsigned int remaining = MAX_TCP_OPTION_SPACE;
850 #ifdef CONFIG_TCP_MD5SIG
852 opts->options |= OPTION_MD5;
853 remaining -= TCPOLEN_MD5SIG_ALIGNED;
855 /* We can't fit any SACK blocks in a packet with MD5 + TS
856 * options. There was discussion about disabling SACK
857 * rather than TS in order to fit in better with old,
858 * buggy kernels, but that was deemed to be unnecessary.
860 if (synack_type != TCP_SYNACK_COOKIE)
861 ireq->tstamp_ok &= !ireq->sack_ok;
865 /* We always send an MSS option. */
867 remaining -= TCPOLEN_MSS_ALIGNED;
869 if (likely(ireq->wscale_ok)) {
870 opts->ws = ireq->rcv_wscale;
871 opts->options |= OPTION_WSCALE;
872 remaining -= TCPOLEN_WSCALE_ALIGNED;
874 if (likely(ireq->tstamp_ok)) {
875 opts->options |= OPTION_TS;
876 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
877 opts->tsecr = READ_ONCE(req->ts_recent);
878 remaining -= TCPOLEN_TSTAMP_ALIGNED;
880 if (likely(ireq->sack_ok)) {
881 opts->options |= OPTION_SACK_ADVERTISE;
882 if (unlikely(!ireq->tstamp_ok))
883 remaining -= TCPOLEN_SACKPERM_ALIGNED;
885 if (foc != NULL && foc->len >= 0) {
888 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
889 TCPOLEN_FASTOPEN_BASE;
890 need = (need + 3) & ~3U; /* Align to 32 bits */
891 if (remaining >= need) {
892 opts->options |= OPTION_FAST_OPEN_COOKIE;
893 opts->fastopen_cookie = foc;
898 mptcp_set_option_cond(req, opts, &remaining);
900 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
902 bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb,
903 synack_type, opts, &remaining);
905 return MAX_TCP_OPTION_SPACE - remaining;
908 /* Compute TCP options for ESTABLISHED sockets. This is not the
909 * final wire format yet.
911 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
912 struct tcp_out_options *opts,
913 struct tcp_md5sig_key **md5)
915 struct tcp_sock *tp = tcp_sk(sk);
916 unsigned int size = 0;
917 unsigned int eff_sacks;
922 #ifdef CONFIG_TCP_MD5SIG
923 if (static_branch_unlikely(&tcp_md5_needed) &&
924 rcu_access_pointer(tp->md5sig_info)) {
925 *md5 = tp->af_specific->md5_lookup(sk, sk);
927 opts->options |= OPTION_MD5;
928 size += TCPOLEN_MD5SIG_ALIGNED;
933 if (likely(tp->rx_opt.tstamp_ok)) {
934 opts->options |= OPTION_TS;
935 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
936 opts->tsecr = tp->rx_opt.ts_recent;
937 size += TCPOLEN_TSTAMP_ALIGNED;
940 /* MPTCP options have precedence over SACK for the limited TCP
941 * option space because a MPTCP connection would be forced to
942 * fall back to regular TCP if a required multipath option is
943 * missing. SACK still gets a chance to use whatever space is
946 if (sk_is_mptcp(sk)) {
947 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
948 unsigned int opt_size = 0;
950 if (mptcp_established_options(sk, skb, &opt_size, remaining,
952 opts->options |= OPTION_MPTCP;
957 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
958 if (unlikely(eff_sacks)) {
959 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
960 if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
961 TCPOLEN_SACK_PERBLOCK))
964 opts->num_sack_blocks =
965 min_t(unsigned int, eff_sacks,
966 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
967 TCPOLEN_SACK_PERBLOCK);
969 size += TCPOLEN_SACK_BASE_ALIGNED +
970 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
973 if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
974 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
975 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
977 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
979 size = MAX_TCP_OPTION_SPACE - remaining;
986 /* TCP SMALL QUEUES (TSQ)
988 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
989 * to reduce RTT and bufferbloat.
990 * We do this using a special skb destructor (tcp_wfree).
992 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
993 * needs to be reallocated in a driver.
994 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
996 * Since transmit from skb destructor is forbidden, we use a tasklet
997 * to process all sockets that eventually need to send more skbs.
998 * We use one tasklet per cpu, with its own queue of sockets.
1000 struct tsq_tasklet {
1001 struct tasklet_struct tasklet;
1002 struct list_head head; /* queue of tcp sockets */
1004 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
1006 static void tcp_tsq_write(struct sock *sk)
1008 if ((1 << sk->sk_state) &
1009 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1010 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
1011 struct tcp_sock *tp = tcp_sk(sk);
1013 if (tp->lost_out > tp->retrans_out &&
1014 tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) {
1015 tcp_mstamp_refresh(tp);
1016 tcp_xmit_retransmit_queue(sk);
1019 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
1024 static void tcp_tsq_handler(struct sock *sk)
1027 if (!sock_owned_by_user(sk))
1029 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
1034 * One tasklet per cpu tries to send more skbs.
1035 * We run in tasklet context but need to disable irqs when
1036 * transferring tsq->head because tcp_wfree() might
1037 * interrupt us (non NAPI drivers)
1039 static void tcp_tasklet_func(struct tasklet_struct *t)
1041 struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet);
1043 unsigned long flags;
1044 struct list_head *q, *n;
1045 struct tcp_sock *tp;
1048 local_irq_save(flags);
1049 list_splice_init(&tsq->head, &list);
1050 local_irq_restore(flags);
1052 list_for_each_safe(q, n, &list) {
1053 tp = list_entry(q, struct tcp_sock, tsq_node);
1054 list_del(&tp->tsq_node);
1056 sk = (struct sock *)tp;
1057 smp_mb__before_atomic();
1058 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
1060 tcp_tsq_handler(sk);
1065 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
1066 TCPF_WRITE_TIMER_DEFERRED | \
1067 TCPF_DELACK_TIMER_DEFERRED | \
1068 TCPF_MTU_REDUCED_DEFERRED)
1070 * tcp_release_cb - tcp release_sock() callback
1073 * called from release_sock() to perform protocol dependent
1074 * actions before socket release.
1076 void tcp_release_cb(struct sock *sk)
1078 unsigned long flags, nflags;
1080 /* perform an atomic operation only if at least one flag is set */
1082 flags = sk->sk_tsq_flags;
1083 if (!(flags & TCP_DEFERRED_ALL))
1085 nflags = flags & ~TCP_DEFERRED_ALL;
1086 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
1088 if (flags & TCPF_TSQ_DEFERRED) {
1092 /* Here begins the tricky part :
1093 * We are called from release_sock() with :
1095 * 2) sk_lock.slock spinlock held
1096 * 3) socket owned by us (sk->sk_lock.owned == 1)
1098 * But following code is meant to be called from BH handlers,
1099 * so we should keep BH disabled, but early release socket ownership
1101 sock_release_ownership(sk);
1103 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1104 tcp_write_timer_handler(sk);
1107 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1108 tcp_delack_timer_handler(sk);
1111 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1112 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1116 EXPORT_SYMBOL(tcp_release_cb);
1118 void __init tcp_tasklet_init(void)
1122 for_each_possible_cpu(i) {
1123 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
1125 INIT_LIST_HEAD(&tsq->head);
1126 tasklet_setup(&tsq->tasklet, tcp_tasklet_func);
1131 * Write buffer destructor automatically called from kfree_skb.
1132 * We can't xmit new skbs from this context, as we might already
1135 void tcp_wfree(struct sk_buff *skb)
1137 struct sock *sk = skb->sk;
1138 struct tcp_sock *tp = tcp_sk(sk);
1139 unsigned long flags, nval, oval;
1141 /* Keep one reference on sk_wmem_alloc.
1142 * Will be released by sk_free() from here or tcp_tasklet_func()
1144 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1146 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
1147 * Wait until our queues (qdisc + devices) are drained.
1149 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1150 * - chance for incoming ACK (processed by another cpu maybe)
1151 * to migrate this flow (skb->ooo_okay will be eventually set)
1153 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1156 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
1157 struct tsq_tasklet *tsq;
1160 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1163 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1164 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
1168 /* queue this socket to tasklet queue */
1169 local_irq_save(flags);
1170 tsq = this_cpu_ptr(&tsq_tasklet);
1171 empty = list_empty(&tsq->head);
1172 list_add(&tp->tsq_node, &tsq->head);
1174 tasklet_schedule(&tsq->tasklet);
1175 local_irq_restore(flags);
1182 /* Note: Called under soft irq.
1183 * We can call TCP stack right away, unless socket is owned by user.
1185 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1187 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1188 struct sock *sk = (struct sock *)tp;
1190 tcp_tsq_handler(sk);
1193 return HRTIMER_NORESTART;
1196 static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1199 struct tcp_sock *tp = tcp_sk(sk);
1201 if (sk->sk_pacing_status != SK_PACING_NONE) {
1202 unsigned long rate = sk->sk_pacing_rate;
1204 /* Original sch_fq does not pace first 10 MSS
1205 * Note that tp->data_segs_out overflows after 2^32 packets,
1206 * this is a minor annoyance.
1208 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1209 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1210 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1212 /* take into account OS jitter */
1213 len_ns -= min_t(u64, len_ns / 2, credit);
1214 tp->tcp_wstamp_ns += len_ns;
1217 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1220 INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1221 INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1222 INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1224 /* This routine actually transmits TCP packets queued in by
1225 * tcp_do_sendmsg(). This is used by both the initial
1226 * transmission and possible later retransmissions.
1227 * All SKB's seen here are completely headerless. It is our
1228 * job to build the TCP header, and pass the packet down to
1229 * IP so it can do the same plus pass the packet off to the
1232 * We are working here with either a clone of the original
1233 * SKB, or a fresh unique copy made by the retransmit engine.
1235 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1236 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1238 const struct inet_connection_sock *icsk = inet_csk(sk);
1239 struct inet_sock *inet;
1240 struct tcp_sock *tp;
1241 struct tcp_skb_cb *tcb;
1242 struct tcp_out_options opts;
1243 unsigned int tcp_options_size, tcp_header_size;
1244 struct sk_buff *oskb = NULL;
1245 struct tcp_md5sig_key *md5;
1250 BUG_ON(!skb || !tcp_skb_pcount(skb));
1252 prior_wstamp = tp->tcp_wstamp_ns;
1253 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1254 skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
1256 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1260 tcp_skb_tsorted_save(oskb) {
1261 if (unlikely(skb_cloned(oskb)))
1262 skb = pskb_copy(oskb, gfp_mask);
1264 skb = skb_clone(oskb, gfp_mask);
1265 } tcp_skb_tsorted_restore(oskb);
1269 /* retransmit skbs might have a non zero value in skb->dev
1270 * because skb->dev is aliased with skb->rbnode.rb_left
1276 tcb = TCP_SKB_CB(skb);
1277 memset(&opts, 0, sizeof(opts));
1279 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1280 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1282 tcp_options_size = tcp_established_options(sk, skb, &opts,
1284 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1285 * at receiver : This slightly improve GRO performance.
1286 * Note that we do not force the PSH flag for non GSO packets,
1287 * because they might be sent under high congestion events,
1288 * and in this case it is better to delay the delivery of 1-MSS
1289 * packets and thus the corresponding ACK packet that would
1290 * release the following packet.
1292 if (tcp_skb_pcount(skb) > 1)
1293 tcb->tcp_flags |= TCPHDR_PSH;
1295 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1297 /* if no packet is in qdisc/device queue, then allow XPS to select
1298 * another queue. We can be called from tcp_tsq_handler()
1299 * which holds one reference to sk.
1301 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1302 * One way to get this would be to set skb->truesize = 2 on them.
1304 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1306 /* If we had to use memory reserve to allocate this skb,
1307 * this might cause drops if packet is looped back :
1308 * Other socket might not have SOCK_MEMALLOC.
1309 * Packets not looped back do not care about pfmemalloc.
1311 skb->pfmemalloc = 0;
1313 skb_push(skb, tcp_header_size);
1314 skb_reset_transport_header(skb);
1318 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1319 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1321 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1323 /* Build TCP header and checksum it. */
1324 th = (struct tcphdr *)skb->data;
1325 th->source = inet->inet_sport;
1326 th->dest = inet->inet_dport;
1327 th->seq = htonl(tcb->seq);
1328 th->ack_seq = htonl(rcv_nxt);
1329 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1335 /* The urg_mode check is necessary during a below snd_una win probe */
1336 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1337 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1338 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1340 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1341 th->urg_ptr = htons(0xFFFF);
1346 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1347 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1348 th->window = htons(tcp_select_window(sk));
1349 tcp_ecn_send(sk, skb, th, tcp_header_size);
1351 /* RFC1323: The window in SYN & SYN/ACK segments
1354 th->window = htons(min(tp->rcv_wnd, 65535U));
1357 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1359 #ifdef CONFIG_TCP_MD5SIG
1360 /* Calculate the MD5 hash, as we have all we need now */
1362 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1363 tp->af_specific->calc_md5_hash(opts.hash_location,
1368 /* BPF prog is the last one writing header option */
1369 bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1371 INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1372 tcp_v6_send_check, tcp_v4_send_check,
1375 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1376 tcp_event_ack_sent(sk, rcv_nxt);
1378 if (skb->len != tcp_header_size) {
1379 tcp_event_data_sent(tp, sk);
1380 tp->data_segs_out += tcp_skb_pcount(skb);
1381 tp->bytes_sent += skb->len - tcp_header_size;
1384 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1385 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1386 tcp_skb_pcount(skb));
1388 tp->segs_out += tcp_skb_pcount(skb);
1389 skb_set_hash_from_sk(skb, sk);
1390 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1391 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1392 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1394 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1396 /* Cleanup our debris for IP stacks */
1397 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1398 sizeof(struct inet6_skb_parm)));
1400 tcp_add_tx_delay(skb, tp);
1402 err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1403 inet6_csk_xmit, ip_queue_xmit,
1404 sk, skb, &inet->cork.fl);
1406 if (unlikely(err > 0)) {
1408 err = net_xmit_eval(err);
1411 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1412 tcp_rate_skb_sent(sk, oskb);
1417 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1420 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1421 tcp_sk(sk)->rcv_nxt);
1424 /* This routine just queues the buffer for sending.
1426 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1427 * otherwise socket can stall.
1429 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1431 struct tcp_sock *tp = tcp_sk(sk);
1433 /* Advance write_seq and place onto the write_queue. */
1434 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1435 __skb_header_release(skb);
1436 tcp_add_write_queue_tail(sk, skb);
1437 sk_wmem_queued_add(sk, skb->truesize);
1438 sk_mem_charge(sk, skb->truesize);
1441 /* Initialize TSO segments for a packet. */
1442 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1444 if (skb->len <= mss_now) {
1445 /* Avoid the costly divide in the normal
1448 tcp_skb_pcount_set(skb, 1);
1449 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1451 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1452 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1456 /* Pcount in the middle of the write queue got changed, we need to do various
1457 * tweaks to fix counters
1459 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1461 struct tcp_sock *tp = tcp_sk(sk);
1463 tp->packets_out -= decr;
1465 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1466 tp->sacked_out -= decr;
1467 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1468 tp->retrans_out -= decr;
1469 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1470 tp->lost_out -= decr;
1472 /* Reno case is special. Sigh... */
1473 if (tcp_is_reno(tp) && decr > 0)
1474 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1476 if (tp->lost_skb_hint &&
1477 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1478 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1479 tp->lost_cnt_hint -= decr;
1481 tcp_verify_left_out(tp);
1484 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1486 return TCP_SKB_CB(skb)->txstamp_ack ||
1487 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1490 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1492 struct skb_shared_info *shinfo = skb_shinfo(skb);
1494 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1495 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1496 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1497 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1499 shinfo->tx_flags &= ~tsflags;
1500 shinfo2->tx_flags |= tsflags;
1501 swap(shinfo->tskey, shinfo2->tskey);
1502 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1503 TCP_SKB_CB(skb)->txstamp_ack = 0;
1507 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1509 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1510 TCP_SKB_CB(skb)->eor = 0;
1513 /* Insert buff after skb on the write or rtx queue of sk. */
1514 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1515 struct sk_buff *buff,
1517 enum tcp_queue tcp_queue)
1519 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1520 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1522 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1525 /* Function to create two new TCP segments. Shrinks the given segment
1526 * to the specified size and appends a new segment with the rest of the
1527 * packet to the list. This won't be called frequently, I hope.
1528 * Remember, these are still headerless SKBs at this point.
1530 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1531 struct sk_buff *skb, u32 len,
1532 unsigned int mss_now, gfp_t gfp)
1534 struct tcp_sock *tp = tcp_sk(sk);
1535 struct sk_buff *buff;
1536 int nsize, old_factor;
1541 if (WARN_ON(len > skb->len))
1544 nsize = skb_headlen(skb) - len;
1548 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1549 * We need some allowance to not penalize applications setting small
1551 * Also allow first and last skb in retransmit queue to be split.
1553 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1554 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1555 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1556 skb != tcp_rtx_queue_head(sk) &&
1557 skb != tcp_rtx_queue_tail(sk))) {
1558 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1562 if (skb_unclone_keeptruesize(skb, gfp))
1565 /* Get a new skb... force flag on. */
1566 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1568 return -ENOMEM; /* We'll just try again later. */
1569 skb_copy_decrypted(buff, skb);
1570 mptcp_skb_ext_copy(buff, skb);
1572 sk_wmem_queued_add(sk, buff->truesize);
1573 sk_mem_charge(sk, buff->truesize);
1574 nlen = skb->len - len - nsize;
1575 buff->truesize += nlen;
1576 skb->truesize -= nlen;
1578 /* Correct the sequence numbers. */
1579 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1580 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1581 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1583 /* PSH and FIN should only be set in the second packet. */
1584 flags = TCP_SKB_CB(skb)->tcp_flags;
1585 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1586 TCP_SKB_CB(buff)->tcp_flags = flags;
1587 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1588 tcp_skb_fragment_eor(skb, buff);
1590 skb_split(skb, buff, len);
1592 buff->ip_summed = CHECKSUM_PARTIAL;
1594 buff->tstamp = skb->tstamp;
1595 tcp_fragment_tstamp(skb, buff);
1597 old_factor = tcp_skb_pcount(skb);
1599 /* Fix up tso_factor for both original and new SKB. */
1600 tcp_set_skb_tso_segs(skb, mss_now);
1601 tcp_set_skb_tso_segs(buff, mss_now);
1603 /* Update delivered info for the new segment */
1604 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1606 /* If this packet has been sent out already, we must
1607 * adjust the various packet counters.
1609 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1610 int diff = old_factor - tcp_skb_pcount(skb) -
1611 tcp_skb_pcount(buff);
1614 tcp_adjust_pcount(sk, skb, diff);
1617 /* Link BUFF into the send queue. */
1618 __skb_header_release(buff);
1619 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1620 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1621 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1626 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1627 * data is not copied, but immediately discarded.
1629 static int __pskb_trim_head(struct sk_buff *skb, int len)
1631 struct skb_shared_info *shinfo;
1634 eat = min_t(int, len, skb_headlen(skb));
1636 __skb_pull(skb, eat);
1643 shinfo = skb_shinfo(skb);
1644 for (i = 0; i < shinfo->nr_frags; i++) {
1645 int size = skb_frag_size(&shinfo->frags[i]);
1648 skb_frag_unref(skb, i);
1651 shinfo->frags[k] = shinfo->frags[i];
1653 skb_frag_off_add(&shinfo->frags[k], eat);
1654 skb_frag_size_sub(&shinfo->frags[k], eat);
1660 shinfo->nr_frags = k;
1662 skb->data_len -= len;
1663 skb->len = skb->data_len;
1667 /* Remove acked data from a packet in the transmit queue. */
1668 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1672 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
1675 delta_truesize = __pskb_trim_head(skb, len);
1677 TCP_SKB_CB(skb)->seq += len;
1678 skb->ip_summed = CHECKSUM_PARTIAL;
1680 if (delta_truesize) {
1681 skb->truesize -= delta_truesize;
1682 sk_wmem_queued_add(sk, -delta_truesize);
1683 sk_mem_uncharge(sk, delta_truesize);
1686 /* Any change of skb->len requires recalculation of tso factor. */
1687 if (tcp_skb_pcount(skb) > 1)
1688 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1693 /* Calculate MSS not accounting any TCP options. */
1694 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1696 const struct tcp_sock *tp = tcp_sk(sk);
1697 const struct inet_connection_sock *icsk = inet_csk(sk);
1700 /* Calculate base mss without TCP options:
1701 It is MMS_S - sizeof(tcphdr) of rfc1122
1703 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1705 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1706 if (icsk->icsk_af_ops->net_frag_header_len) {
1707 const struct dst_entry *dst = __sk_dst_get(sk);
1709 if (dst && dst_allfrag(dst))
1710 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1713 /* Clamp it (mss_clamp does not include tcp options) */
1714 if (mss_now > tp->rx_opt.mss_clamp)
1715 mss_now = tp->rx_opt.mss_clamp;
1717 /* Now subtract optional transport overhead */
1718 mss_now -= icsk->icsk_ext_hdr_len;
1720 /* Then reserve room for full set of TCP options and 8 bytes of data */
1721 mss_now = max(mss_now,
1722 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss));
1726 /* Calculate MSS. Not accounting for SACKs here. */
1727 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1729 /* Subtract TCP options size, not including SACKs */
1730 return __tcp_mtu_to_mss(sk, pmtu) -
1731 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1733 EXPORT_SYMBOL(tcp_mtu_to_mss);
1735 /* Inverse of above */
1736 int tcp_mss_to_mtu(struct sock *sk, int mss)
1738 const struct tcp_sock *tp = tcp_sk(sk);
1739 const struct inet_connection_sock *icsk = inet_csk(sk);
1743 tp->tcp_header_len +
1744 icsk->icsk_ext_hdr_len +
1745 icsk->icsk_af_ops->net_header_len;
1747 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1748 if (icsk->icsk_af_ops->net_frag_header_len) {
1749 const struct dst_entry *dst = __sk_dst_get(sk);
1751 if (dst && dst_allfrag(dst))
1752 mtu += icsk->icsk_af_ops->net_frag_header_len;
1756 EXPORT_SYMBOL(tcp_mss_to_mtu);
1758 /* MTU probing init per socket */
1759 void tcp_mtup_init(struct sock *sk)
1761 struct tcp_sock *tp = tcp_sk(sk);
1762 struct inet_connection_sock *icsk = inet_csk(sk);
1763 struct net *net = sock_net(sk);
1765 icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1766 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1767 icsk->icsk_af_ops->net_header_len;
1768 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss));
1769 icsk->icsk_mtup.probe_size = 0;
1770 if (icsk->icsk_mtup.enabled)
1771 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1773 EXPORT_SYMBOL(tcp_mtup_init);
1775 /* This function synchronize snd mss to current pmtu/exthdr set.
1777 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1778 for TCP options, but includes only bare TCP header.
1780 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1781 It is minimum of user_mss and mss received with SYN.
1782 It also does not include TCP options.
1784 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1786 tp->mss_cache is current effective sending mss, including
1787 all tcp options except for SACKs. It is evaluated,
1788 taking into account current pmtu, but never exceeds
1789 tp->rx_opt.mss_clamp.
1791 NOTE1. rfc1122 clearly states that advertised MSS
1792 DOES NOT include either tcp or ip options.
1794 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1795 are READ ONLY outside this function. --ANK (980731)
1797 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1799 struct tcp_sock *tp = tcp_sk(sk);
1800 struct inet_connection_sock *icsk = inet_csk(sk);
1803 if (icsk->icsk_mtup.search_high > pmtu)
1804 icsk->icsk_mtup.search_high = pmtu;
1806 mss_now = tcp_mtu_to_mss(sk, pmtu);
1807 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1809 /* And store cached results */
1810 icsk->icsk_pmtu_cookie = pmtu;
1811 if (icsk->icsk_mtup.enabled)
1812 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1813 tp->mss_cache = mss_now;
1817 EXPORT_SYMBOL(tcp_sync_mss);
1819 /* Compute the current effective MSS, taking SACKs and IP options,
1820 * and even PMTU discovery events into account.
1822 unsigned int tcp_current_mss(struct sock *sk)
1824 const struct tcp_sock *tp = tcp_sk(sk);
1825 const struct dst_entry *dst = __sk_dst_get(sk);
1827 unsigned int header_len;
1828 struct tcp_out_options opts;
1829 struct tcp_md5sig_key *md5;
1831 mss_now = tp->mss_cache;
1834 u32 mtu = dst_mtu(dst);
1835 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1836 mss_now = tcp_sync_mss(sk, mtu);
1839 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1840 sizeof(struct tcphdr);
1841 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1842 * some common options. If this is an odd packet (because we have SACK
1843 * blocks etc) then our calculated header_len will be different, and
1844 * we have to adjust mss_now correspondingly */
1845 if (header_len != tp->tcp_header_len) {
1846 int delta = (int) header_len - tp->tcp_header_len;
1853 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1854 * As additional protections, we do not touch cwnd in retransmission phases,
1855 * and if application hit its sndbuf limit recently.
1857 static void tcp_cwnd_application_limited(struct sock *sk)
1859 struct tcp_sock *tp = tcp_sk(sk);
1861 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1862 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1863 /* Limited by application or receiver window. */
1864 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1865 u32 win_used = max(tp->snd_cwnd_used, init_win);
1866 if (win_used < tcp_snd_cwnd(tp)) {
1867 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1868 tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1);
1870 tp->snd_cwnd_used = 0;
1872 tp->snd_cwnd_stamp = tcp_jiffies32;
1875 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1877 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1878 struct tcp_sock *tp = tcp_sk(sk);
1880 /* Track the strongest available signal of the degree to which the cwnd
1881 * is fully utilized. If cwnd-limited then remember that fact for the
1882 * current window. If not cwnd-limited then track the maximum number of
1883 * outstanding packets in the current window. (If cwnd-limited then we
1884 * chose to not update tp->max_packets_out to avoid an extra else
1885 * clause with no functional impact.)
1887 if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
1889 (!tp->is_cwnd_limited &&
1890 tp->packets_out > tp->max_packets_out)) {
1891 tp->is_cwnd_limited = is_cwnd_limited;
1892 tp->max_packets_out = tp->packets_out;
1893 tp->cwnd_usage_seq = tp->snd_nxt;
1896 if (tcp_is_cwnd_limited(sk)) {
1897 /* Network is feed fully. */
1898 tp->snd_cwnd_used = 0;
1899 tp->snd_cwnd_stamp = tcp_jiffies32;
1901 /* Network starves. */
1902 if (tp->packets_out > tp->snd_cwnd_used)
1903 tp->snd_cwnd_used = tp->packets_out;
1905 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
1906 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1907 !ca_ops->cong_control)
1908 tcp_cwnd_application_limited(sk);
1910 /* The following conditions together indicate the starvation
1911 * is caused by insufficient sender buffer:
1912 * 1) just sent some data (see tcp_write_xmit)
1913 * 2) not cwnd limited (this else condition)
1914 * 3) no more data to send (tcp_write_queue_empty())
1915 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1917 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1918 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1919 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1920 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1924 /* Minshall's variant of the Nagle send check. */
1925 static bool tcp_minshall_check(const struct tcp_sock *tp)
1927 return after(tp->snd_sml, tp->snd_una) &&
1928 !after(tp->snd_sml, tp->snd_nxt);
1931 /* Update snd_sml if this skb is under mss
1932 * Note that a TSO packet might end with a sub-mss segment
1933 * The test is really :
1934 * if ((skb->len % mss) != 0)
1935 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1936 * But we can avoid doing the divide again given we already have
1937 * skb_pcount = skb->len / mss_now
1939 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1940 const struct sk_buff *skb)
1942 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1943 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1946 /* Return false, if packet can be sent now without violation Nagle's rules:
1947 * 1. It is full sized. (provided by caller in %partial bool)
1948 * 2. Or it contains FIN. (already checked by caller)
1949 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1950 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1951 * With Minshall's modification: all sent small packets are ACKed.
1953 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1957 ((nonagle & TCP_NAGLE_CORK) ||
1958 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1961 /* Return how many segs we'd like on a TSO packet,
1962 * to send one TSO packet per ms
1964 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1969 bytes = min_t(unsigned long,
1970 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift),
1971 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1973 /* Goal is to send at least one packet per ms,
1974 * not one big TSO packet every 100 ms.
1975 * This preserves ACK clocking and is consistent
1976 * with tcp_tso_should_defer() heuristic.
1978 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1983 /* Return the number of segments we want in the skb we are transmitting.
1984 * See if congestion control module wants to decide; otherwise, autosize.
1986 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1988 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1989 u32 min_tso, tso_segs;
1991 min_tso = ca_ops->min_tso_segs ?
1992 ca_ops->min_tso_segs(sk) :
1993 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
1995 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1996 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1999 /* Returns the portion of skb which can be sent right away */
2000 static unsigned int tcp_mss_split_point(const struct sock *sk,
2001 const struct sk_buff *skb,
2002 unsigned int mss_now,
2003 unsigned int max_segs,
2006 const struct tcp_sock *tp = tcp_sk(sk);
2007 u32 partial, needed, window, max_len;
2009 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2010 max_len = mss_now * max_segs;
2012 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2015 needed = min(skb->len, window);
2017 if (max_len <= needed)
2020 partial = needed % mss_now;
2021 /* If last segment is not a full MSS, check if Nagle rules allow us
2022 * to include this last segment in this skb.
2023 * Otherwise, we'll split the skb at last MSS boundary
2025 if (tcp_nagle_check(partial != 0, tp, nonagle))
2026 return needed - partial;
2031 /* Can at least one segment of SKB be sent right now, according to the
2032 * congestion window rules? If so, return how many segments are allowed.
2034 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
2035 const struct sk_buff *skb)
2037 u32 in_flight, cwnd, halfcwnd;
2039 /* Don't be strict about the congestion window for the final FIN. */
2040 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2041 tcp_skb_pcount(skb) == 1)
2044 in_flight = tcp_packets_in_flight(tp);
2045 cwnd = tcp_snd_cwnd(tp);
2046 if (in_flight >= cwnd)
2049 /* For better scheduling, ensure we have at least
2050 * 2 GSO packets in flight.
2052 halfcwnd = max(cwnd >> 1, 1U);
2053 return min(halfcwnd, cwnd - in_flight);
2056 /* Initialize TSO state of a skb.
2057 * This must be invoked the first time we consider transmitting
2058 * SKB onto the wire.
2060 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
2062 int tso_segs = tcp_skb_pcount(skb);
2064 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
2065 tcp_set_skb_tso_segs(skb, mss_now);
2066 tso_segs = tcp_skb_pcount(skb);
2072 /* Return true if the Nagle test allows this packet to be
2075 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2076 unsigned int cur_mss, int nonagle)
2078 /* Nagle rule does not apply to frames, which sit in the middle of the
2079 * write_queue (they have no chances to get new data).
2081 * This is implemented in the callers, where they modify the 'nonagle'
2082 * argument based upon the location of SKB in the send queue.
2084 if (nonagle & TCP_NAGLE_PUSH)
2087 /* Don't use the nagle rule for urgent data (or for the final FIN). */
2088 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2091 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
2097 /* Does at least the first segment of SKB fit into the send window? */
2098 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2099 const struct sk_buff *skb,
2100 unsigned int cur_mss)
2102 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2104 if (skb->len > cur_mss)
2105 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2107 return !after(end_seq, tcp_wnd_end(tp));
2110 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2111 * which is put after SKB on the list. It is very much like
2112 * tcp_fragment() except that it may make several kinds of assumptions
2113 * in order to speed up the splitting operation. In particular, we
2114 * know that all the data is in scatter-gather pages, and that the
2115 * packet has never been sent out before (and thus is not cloned).
2117 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2118 unsigned int mss_now, gfp_t gfp)
2120 int nlen = skb->len - len;
2121 struct sk_buff *buff;
2124 /* All of a TSO frame must be composed of paged data. */
2125 if (skb->len != skb->data_len)
2126 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2127 skb, len, mss_now, gfp);
2129 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
2130 if (unlikely(!buff))
2132 skb_copy_decrypted(buff, skb);
2133 mptcp_skb_ext_copy(buff, skb);
2135 sk_wmem_queued_add(sk, buff->truesize);
2136 sk_mem_charge(sk, buff->truesize);
2137 buff->truesize += nlen;
2138 skb->truesize -= nlen;
2140 /* Correct the sequence numbers. */
2141 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2142 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2143 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2145 /* PSH and FIN should only be set in the second packet. */
2146 flags = TCP_SKB_CB(skb)->tcp_flags;
2147 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2148 TCP_SKB_CB(buff)->tcp_flags = flags;
2150 /* This packet was never sent out yet, so no SACK bits. */
2151 TCP_SKB_CB(buff)->sacked = 0;
2153 tcp_skb_fragment_eor(skb, buff);
2155 buff->ip_summed = CHECKSUM_PARTIAL;
2156 skb_split(skb, buff, len);
2157 tcp_fragment_tstamp(skb, buff);
2159 /* Fix up tso_factor for both original and new SKB. */
2160 tcp_set_skb_tso_segs(skb, mss_now);
2161 tcp_set_skb_tso_segs(buff, mss_now);
2163 /* Link BUFF into the send queue. */
2164 __skb_header_release(buff);
2165 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2170 /* Try to defer sending, if possible, in order to minimize the amount
2171 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2173 * This algorithm is from John Heffner.
2175 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2176 bool *is_cwnd_limited,
2177 bool *is_rwnd_limited,
2180 const struct inet_connection_sock *icsk = inet_csk(sk);
2181 u32 send_win, cong_win, limit, in_flight;
2182 struct tcp_sock *tp = tcp_sk(sk);
2183 struct sk_buff *head;
2187 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2190 /* Avoid bursty behavior by allowing defer
2191 * only if the last write was recent (1 ms).
2192 * Note that tp->tcp_wstamp_ns can be in the future if we have
2193 * packets waiting in a qdisc or device for EDT delivery.
2195 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2199 in_flight = tcp_packets_in_flight(tp);
2201 BUG_ON(tcp_skb_pcount(skb) <= 1);
2202 BUG_ON(tcp_snd_cwnd(tp) <= in_flight);
2204 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2206 /* From in_flight test above, we know that cwnd > in_flight. */
2207 cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache;
2209 limit = min(send_win, cong_win);
2211 /* If a full-sized TSO skb can be sent, do it. */
2212 if (limit >= max_segs * tp->mss_cache)
2215 /* Middle in queue won't get any more data, full sendable already? */
2216 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2219 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2221 u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache);
2223 /* If at least some fraction of a window is available,
2226 chunk /= win_divisor;
2230 /* Different approach, try not to defer past a single
2231 * ACK. Receiver should ACK every other full sized
2232 * frame, so if we have space for more than 3 frames
2235 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2239 /* TODO : use tsorted_sent_queue ? */
2240 head = tcp_rtx_queue_head(sk);
2243 delta = tp->tcp_clock_cache - head->tstamp;
2244 /* If next ACK is likely to come too late (half srtt), do not defer */
2245 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2248 /* Ok, it looks like it is advisable to defer.
2249 * Three cases are tracked :
2250 * 1) We are cwnd-limited
2251 * 2) We are rwnd-limited
2252 * 3) We are application limited.
2254 if (cong_win < send_win) {
2255 if (cong_win <= skb->len) {
2256 *is_cwnd_limited = true;
2260 if (send_win <= skb->len) {
2261 *is_rwnd_limited = true;
2266 /* If this packet won't get more data, do not wait. */
2267 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2268 TCP_SKB_CB(skb)->eor)
2277 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2279 struct inet_connection_sock *icsk = inet_csk(sk);
2280 struct tcp_sock *tp = tcp_sk(sk);
2281 struct net *net = sock_net(sk);
2285 interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2286 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2287 if (unlikely(delta >= interval * HZ)) {
2288 int mss = tcp_current_mss(sk);
2290 /* Update current search range */
2291 icsk->icsk_mtup.probe_size = 0;
2292 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2293 sizeof(struct tcphdr) +
2294 icsk->icsk_af_ops->net_header_len;
2295 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2297 /* Update probe time stamp */
2298 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2302 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2304 struct sk_buff *skb, *next;
2306 skb = tcp_send_head(sk);
2307 tcp_for_write_queue_from_safe(skb, next, sk) {
2308 if (len <= skb->len)
2311 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2320 /* Create a new MTU probe if we are ready.
2321 * MTU probe is regularly attempting to increase the path MTU by
2322 * deliberately sending larger packets. This discovers routing
2323 * changes resulting in larger path MTUs.
2325 * Returns 0 if we should wait to probe (no cwnd available),
2326 * 1 if a probe was sent,
2329 static int tcp_mtu_probe(struct sock *sk)
2331 struct inet_connection_sock *icsk = inet_csk(sk);
2332 struct tcp_sock *tp = tcp_sk(sk);
2333 struct sk_buff *skb, *nskb, *next;
2334 struct net *net = sock_net(sk);
2341 /* Not currently probing/verifying,
2343 * have enough cwnd, and
2344 * not SACKing (the variable headers throw things off)
2346 if (likely(!icsk->icsk_mtup.enabled ||
2347 icsk->icsk_mtup.probe_size ||
2348 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2349 tcp_snd_cwnd(tp) < 11 ||
2350 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2353 /* Use binary search for probe_size between tcp_mss_base,
2354 * and current mss_clamp. if (search_high - search_low)
2355 * smaller than a threshold, backoff from probing.
2357 mss_now = tcp_current_mss(sk);
2358 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2359 icsk->icsk_mtup.search_low) >> 1);
2360 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2361 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2362 /* When misfortune happens, we are reprobing actively,
2363 * and then reprobe timer has expired. We stick with current
2364 * probing process by not resetting search range to its orignal.
2366 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2367 interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2368 /* Check whether enough time has elaplased for
2369 * another round of probing.
2371 tcp_mtu_check_reprobe(sk);
2375 /* Have enough data in the send queue to probe? */
2376 if (tp->write_seq - tp->snd_nxt < size_needed)
2379 if (tp->snd_wnd < size_needed)
2381 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2384 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2385 if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) {
2386 if (!tcp_packets_in_flight(tp))
2392 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2395 /* We're allowed to probe. Build it now. */
2396 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2399 sk_wmem_queued_add(sk, nskb->truesize);
2400 sk_mem_charge(sk, nskb->truesize);
2402 skb = tcp_send_head(sk);
2403 skb_copy_decrypted(nskb, skb);
2404 mptcp_skb_ext_copy(nskb, skb);
2406 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2407 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2408 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2409 TCP_SKB_CB(nskb)->sacked = 0;
2411 nskb->ip_summed = CHECKSUM_PARTIAL;
2413 tcp_insert_write_queue_before(nskb, skb, sk);
2414 tcp_highest_sack_replace(sk, skb, nskb);
2417 tcp_for_write_queue_from_safe(skb, next, sk) {
2418 copy = min_t(int, skb->len, probe_size - len);
2419 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2421 if (skb->len <= copy) {
2422 /* We've eaten all the data from this skb.
2424 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2425 /* If this is the last SKB we copy and eor is set
2426 * we need to propagate it to the new skb.
2428 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2429 tcp_skb_collapse_tstamp(nskb, skb);
2430 tcp_unlink_write_queue(skb, sk);
2431 sk_wmem_free_skb(sk, skb);
2433 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2434 ~(TCPHDR_FIN|TCPHDR_PSH);
2435 if (!skb_shinfo(skb)->nr_frags) {
2436 skb_pull(skb, copy);
2438 __pskb_trim_head(skb, copy);
2439 tcp_set_skb_tso_segs(skb, mss_now);
2441 TCP_SKB_CB(skb)->seq += copy;
2446 if (len >= probe_size)
2449 tcp_init_tso_segs(nskb, nskb->len);
2451 /* We're ready to send. If this fails, the probe will
2452 * be resegmented into mss-sized pieces by tcp_write_xmit().
2454 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2455 /* Decrement cwnd here because we are sending
2456 * effectively two packets. */
2457 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1);
2458 tcp_event_new_data_sent(sk, nskb);
2460 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2461 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2462 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2470 static bool tcp_pacing_check(struct sock *sk)
2472 struct tcp_sock *tp = tcp_sk(sk);
2474 if (!tcp_needs_internal_pacing(sk))
2477 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2480 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2481 hrtimer_start(&tp->pacing_timer,
2482 ns_to_ktime(tp->tcp_wstamp_ns),
2483 HRTIMER_MODE_ABS_PINNED_SOFT);
2489 static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk)
2491 const struct rb_node *node = sk->tcp_rtx_queue.rb_node;
2493 /* No skb in the rtx queue. */
2497 /* Only one skb in rtx queue. */
2498 return !node->rb_left && !node->rb_right;
2501 /* TCP Small Queues :
2502 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2503 * (These limits are doubled for retransmits)
2505 * - better RTT estimation and ACK scheduling
2508 * Alas, some drivers / subsystems require a fair amount
2509 * of queued bytes to ensure line rate.
2510 * One example is wifi aggregation (802.11 AMPDU)
2512 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2513 unsigned int factor)
2515 unsigned long limit;
2517 limit = max_t(unsigned long,
2519 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2520 if (sk->sk_pacing_status == SK_PACING_NONE)
2521 limit = min_t(unsigned long, limit,
2522 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
2525 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2526 tcp_sk(sk)->tcp_tx_delay) {
2527 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2529 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2530 * approximate our needs assuming an ~100% skb->truesize overhead.
2531 * USEC_PER_SEC is approximated by 2^20.
2532 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2534 extra_bytes >>= (20 - 1);
2535 limit += extra_bytes;
2537 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2538 /* Always send skb if rtx queue is empty or has one skb.
2539 * No need to wait for TX completion to call us back,
2540 * after softirq/tasklet schedule.
2541 * This helps when TX completions are delayed too much.
2543 if (tcp_rtx_queue_empty_or_single_skb(sk))
2546 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2547 /* It is possible TX completion already happened
2548 * before we set TSQ_THROTTLED, so we must
2549 * test again the condition.
2551 smp_mb__after_atomic();
2552 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2558 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2560 const u32 now = tcp_jiffies32;
2561 enum tcp_chrono old = tp->chrono_type;
2563 if (old > TCP_CHRONO_UNSPEC)
2564 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2565 tp->chrono_start = now;
2566 tp->chrono_type = new;
2569 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2571 struct tcp_sock *tp = tcp_sk(sk);
2573 /* If there are multiple conditions worthy of tracking in a
2574 * chronograph then the highest priority enum takes precedence
2575 * over the other conditions. So that if something "more interesting"
2576 * starts happening, stop the previous chrono and start a new one.
2578 if (type > tp->chrono_type)
2579 tcp_chrono_set(tp, type);
2582 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2584 struct tcp_sock *tp = tcp_sk(sk);
2587 /* There are multiple conditions worthy of tracking in a
2588 * chronograph, so that the highest priority enum takes
2589 * precedence over the other conditions (see tcp_chrono_start).
2590 * If a condition stops, we only stop chrono tracking if
2591 * it's the "most interesting" or current chrono we are
2592 * tracking and starts busy chrono if we have pending data.
2594 if (tcp_rtx_and_write_queues_empty(sk))
2595 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2596 else if (type == tp->chrono_type)
2597 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2600 /* This routine writes packets to the network. It advances the
2601 * send_head. This happens as incoming acks open up the remote
2604 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2605 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2606 * account rare use of URG, this is not a big flaw.
2608 * Send at most one packet when push_one > 0. Temporarily ignore
2609 * cwnd limit to force at most one packet out when push_one == 2.
2611 * Returns true, if no segments are in flight and we have queued segments,
2612 * but cannot send anything now because of SWS or another problem.
2614 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2615 int push_one, gfp_t gfp)
2617 struct tcp_sock *tp = tcp_sk(sk);
2618 struct sk_buff *skb;
2619 unsigned int tso_segs, sent_pkts;
2622 bool is_cwnd_limited = false, is_rwnd_limited = false;
2627 tcp_mstamp_refresh(tp);
2629 /* Do MTU probing. */
2630 result = tcp_mtu_probe(sk);
2633 } else if (result > 0) {
2638 max_segs = tcp_tso_segs(sk, mss_now);
2639 while ((skb = tcp_send_head(sk))) {
2642 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2643 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2644 skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2645 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2646 tcp_init_tso_segs(skb, mss_now);
2647 goto repair; /* Skip network transmission */
2650 if (tcp_pacing_check(sk))
2653 tso_segs = tcp_init_tso_segs(skb, mss_now);
2656 cwnd_quota = tcp_cwnd_test(tp, skb);
2659 /* Force out a loss probe pkt. */
2665 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2666 is_rwnd_limited = true;
2670 if (tso_segs == 1) {
2671 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2672 (tcp_skb_is_last(sk, skb) ?
2673 nonagle : TCP_NAGLE_PUSH))))
2677 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2678 &is_rwnd_limited, max_segs))
2683 if (tso_segs > 1 && !tcp_urg_mode(tp))
2684 limit = tcp_mss_split_point(sk, skb, mss_now,
2690 if (skb->len > limit &&
2691 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2694 if (tcp_small_queue_check(sk, skb, 0))
2697 /* Argh, we hit an empty skb(), presumably a thread
2698 * is sleeping in sendmsg()/sk_stream_wait_memory().
2699 * We do not want to send a pure-ack packet and have
2700 * a strange looking rtx queue with empty packet(s).
2702 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2705 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2709 /* Advance the send_head. This one is sent out.
2710 * This call will increment packets_out.
2712 tcp_event_new_data_sent(sk, skb);
2714 tcp_minshall_update(tp, mss_now, skb);
2715 sent_pkts += tcp_skb_pcount(skb);
2721 if (is_rwnd_limited)
2722 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2724 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2726 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp));
2727 if (likely(sent_pkts || is_cwnd_limited))
2728 tcp_cwnd_validate(sk, is_cwnd_limited);
2730 if (likely(sent_pkts)) {
2731 if (tcp_in_cwnd_reduction(sk))
2732 tp->prr_out += sent_pkts;
2734 /* Send one loss probe per tail loss episode. */
2736 tcp_schedule_loss_probe(sk, false);
2739 return !tp->packets_out && !tcp_write_queue_empty(sk);
2742 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2744 struct inet_connection_sock *icsk = inet_csk(sk);
2745 struct tcp_sock *tp = tcp_sk(sk);
2746 u32 timeout, timeout_us, rto_delta_us;
2749 /* Don't do any loss probe on a Fast Open connection before 3WHS
2752 if (rcu_access_pointer(tp->fastopen_rsk))
2755 early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
2756 /* Schedule a loss probe in 2*RTT for SACK capable connections
2757 * not in loss recovery, that are either limited by cwnd or application.
2759 if ((early_retrans != 3 && early_retrans != 4) ||
2760 !tp->packets_out || !tcp_is_sack(tp) ||
2761 (icsk->icsk_ca_state != TCP_CA_Open &&
2762 icsk->icsk_ca_state != TCP_CA_CWR))
2765 /* Probe timeout is 2*rtt. Add minimum RTO to account
2766 * for delayed ack when there's one outstanding packet. If no RTT
2767 * sample is available then probe after TCP_TIMEOUT_INIT.
2770 timeout_us = tp->srtt_us >> 2;
2771 if (tp->packets_out == 1)
2772 timeout_us += tcp_rto_min_us(sk);
2774 timeout_us += TCP_TIMEOUT_MIN_US;
2775 timeout = usecs_to_jiffies(timeout_us);
2777 timeout = TCP_TIMEOUT_INIT;
2780 /* If the RTO formula yields an earlier time, then use that time. */
2781 rto_delta_us = advancing_rto ?
2782 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2783 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2784 if (rto_delta_us > 0)
2785 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2787 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2791 /* Thanks to skb fast clones, we can detect if a prior transmit of
2792 * a packet is still in a qdisc or driver queue.
2793 * In this case, there is very little point doing a retransmit !
2795 static bool skb_still_in_host_queue(struct sock *sk,
2796 const struct sk_buff *skb)
2798 if (unlikely(skb_fclone_busy(sk, skb))) {
2799 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2800 smp_mb__after_atomic();
2801 if (skb_fclone_busy(sk, skb)) {
2802 NET_INC_STATS(sock_net(sk),
2803 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2810 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2811 * retransmit the last segment.
2813 void tcp_send_loss_probe(struct sock *sk)
2815 struct tcp_sock *tp = tcp_sk(sk);
2816 struct sk_buff *skb;
2818 int mss = tcp_current_mss(sk);
2820 /* At most one outstanding TLP */
2821 if (tp->tlp_high_seq)
2824 tp->tlp_retrans = 0;
2825 skb = tcp_send_head(sk);
2826 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2827 pcount = tp->packets_out;
2828 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2829 if (tp->packets_out > pcount)
2833 skb = skb_rb_last(&sk->tcp_rtx_queue);
2834 if (unlikely(!skb)) {
2835 WARN_ONCE(tp->packets_out,
2836 "invalid inflight: %u state %u cwnd %u mss %d\n",
2837 tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss);
2838 inet_csk(sk)->icsk_pending = 0;
2842 if (skb_still_in_host_queue(sk, skb))
2845 pcount = tcp_skb_pcount(skb);
2846 if (WARN_ON(!pcount))
2849 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2850 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2851 (pcount - 1) * mss, mss,
2854 skb = skb_rb_next(skb);
2857 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2860 if (__tcp_retransmit_skb(sk, skb, 1))
2863 tp->tlp_retrans = 1;
2866 /* Record snd_nxt for loss detection. */
2867 tp->tlp_high_seq = tp->snd_nxt;
2869 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2870 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2871 inet_csk(sk)->icsk_pending = 0;
2876 /* Push out any pending frames which were held back due to
2877 * TCP_CORK or attempt at coalescing tiny packets.
2878 * The socket must be locked by the caller.
2880 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2883 /* If we are closed, the bytes will have to remain here.
2884 * In time closedown will finish, we empty the write queue and
2885 * all will be happy.
2887 if (unlikely(sk->sk_state == TCP_CLOSE))
2890 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2891 sk_gfp_mask(sk, GFP_ATOMIC)))
2892 tcp_check_probe_timer(sk);
2895 /* Send _single_ skb sitting at the send head. This function requires
2896 * true push pending frames to setup probe timer etc.
2898 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2900 struct sk_buff *skb = tcp_send_head(sk);
2902 BUG_ON(!skb || skb->len < mss_now);
2904 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2907 /* This function returns the amount that we can raise the
2908 * usable window based on the following constraints
2910 * 1. The window can never be shrunk once it is offered (RFC 793)
2911 * 2. We limit memory per socket
2914 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2915 * RECV.NEXT + RCV.WIN fixed until:
2916 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2918 * i.e. don't raise the right edge of the window until you can raise
2919 * it at least MSS bytes.
2921 * Unfortunately, the recommended algorithm breaks header prediction,
2922 * since header prediction assumes th->window stays fixed.
2924 * Strictly speaking, keeping th->window fixed violates the receiver
2925 * side SWS prevention criteria. The problem is that under this rule
2926 * a stream of single byte packets will cause the right side of the
2927 * window to always advance by a single byte.
2929 * Of course, if the sender implements sender side SWS prevention
2930 * then this will not be a problem.
2932 * BSD seems to make the following compromise:
2934 * If the free space is less than the 1/4 of the maximum
2935 * space available and the free space is less than 1/2 mss,
2936 * then set the window to 0.
2937 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2938 * Otherwise, just prevent the window from shrinking
2939 * and from being larger than the largest representable value.
2941 * This prevents incremental opening of the window in the regime
2942 * where TCP is limited by the speed of the reader side taking
2943 * data out of the TCP receive queue. It does nothing about
2944 * those cases where the window is constrained on the sender side
2945 * because the pipeline is full.
2947 * BSD also seems to "accidentally" limit itself to windows that are a
2948 * multiple of MSS, at least until the free space gets quite small.
2949 * This would appear to be a side effect of the mbuf implementation.
2950 * Combining these two algorithms results in the observed behavior
2951 * of having a fixed window size at almost all times.
2953 * Below we obtain similar behavior by forcing the offered window to
2954 * a multiple of the mss when it is feasible to do so.
2956 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2957 * Regular options like TIMESTAMP are taken into account.
2959 u32 __tcp_select_window(struct sock *sk)
2961 struct inet_connection_sock *icsk = inet_csk(sk);
2962 struct tcp_sock *tp = tcp_sk(sk);
2963 /* MSS for the peer's data. Previous versions used mss_clamp
2964 * here. I don't know if the value based on our guesses
2965 * of peer's MSS is better for the performance. It's more correct
2966 * but may be worse for the performance because of rcv_mss
2967 * fluctuations. --SAW 1998/11/1
2969 int mss = icsk->icsk_ack.rcv_mss;
2970 int free_space = tcp_space(sk);
2971 int allowed_space = tcp_full_space(sk);
2972 int full_space, window;
2974 if (sk_is_mptcp(sk))
2975 mptcp_space(sk, &free_space, &allowed_space);
2977 full_space = min_t(int, tp->window_clamp, allowed_space);
2979 if (unlikely(mss > full_space)) {
2984 if (free_space < (full_space >> 1)) {
2985 icsk->icsk_ack.quick = 0;
2987 if (tcp_under_memory_pressure(sk))
2988 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2991 /* free_space might become our new window, make sure we don't
2992 * increase it due to wscale.
2994 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2996 /* if free space is less than mss estimate, or is below 1/16th
2997 * of the maximum allowed, try to move to zero-window, else
2998 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2999 * new incoming data is dropped due to memory limits.
3000 * With large window, mss test triggers way too late in order
3001 * to announce zero window in time before rmem limit kicks in.
3003 if (free_space < (allowed_space >> 4) || free_space < mss)
3007 if (free_space > tp->rcv_ssthresh)
3008 free_space = tp->rcv_ssthresh;
3010 /* Don't do rounding if we are using window scaling, since the
3011 * scaled window will not line up with the MSS boundary anyway.
3013 if (tp->rx_opt.rcv_wscale) {
3014 window = free_space;
3016 /* Advertise enough space so that it won't get scaled away.
3017 * Import case: prevent zero window announcement if
3018 * 1<<rcv_wscale > mss.
3020 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
3022 window = tp->rcv_wnd;
3023 /* Get the largest window that is a nice multiple of mss.
3024 * Window clamp already applied above.
3025 * If our current window offering is within 1 mss of the
3026 * free space we just keep it. This prevents the divide
3027 * and multiply from happening most of the time.
3028 * We also don't do any window rounding when the free space
3031 if (window <= free_space - mss || window > free_space)
3032 window = rounddown(free_space, mss);
3033 else if (mss == full_space &&
3034 free_space > window + (full_space >> 1))
3035 window = free_space;
3041 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3042 const struct sk_buff *next_skb)
3044 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3045 const struct skb_shared_info *next_shinfo =
3046 skb_shinfo(next_skb);
3047 struct skb_shared_info *shinfo = skb_shinfo(skb);
3049 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3050 shinfo->tskey = next_shinfo->tskey;
3051 TCP_SKB_CB(skb)->txstamp_ack |=
3052 TCP_SKB_CB(next_skb)->txstamp_ack;
3056 /* Collapses two adjacent SKB's during retransmission. */
3057 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3059 struct tcp_sock *tp = tcp_sk(sk);
3060 struct sk_buff *next_skb = skb_rb_next(skb);
3063 next_skb_size = next_skb->len;
3065 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3067 if (next_skb_size) {
3068 if (next_skb_size <= skb_availroom(skb))
3069 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
3071 else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3074 tcp_highest_sack_replace(sk, next_skb, skb);
3076 /* Update sequence range on original skb. */
3077 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3079 /* Merge over control information. This moves PSH/FIN etc. over */
3080 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3082 /* All done, get rid of second SKB and account for it so
3083 * packet counting does not break.
3085 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3086 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3088 /* changed transmit queue under us so clear hints */
3089 tcp_clear_retrans_hints_partial(tp);
3090 if (next_skb == tp->retransmit_skb_hint)
3091 tp->retransmit_skb_hint = skb;
3093 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3095 tcp_skb_collapse_tstamp(skb, next_skb);
3097 tcp_rtx_queue_unlink_and_free(next_skb, sk);
3101 /* Check if coalescing SKBs is legal. */
3102 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3104 if (tcp_skb_pcount(skb) > 1)
3106 if (skb_cloned(skb))
3108 /* Some heuristics for collapsing over SACK'd could be invented */
3109 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3115 /* Collapse packets in the retransmit queue to make to create
3116 * less packets on the wire. This is only done on retransmission.
3118 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3121 struct tcp_sock *tp = tcp_sk(sk);
3122 struct sk_buff *skb = to, *tmp;
3125 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
3127 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3130 skb_rbtree_walk_from_safe(skb, tmp) {
3131 if (!tcp_can_collapse(sk, skb))
3134 if (!tcp_skb_can_collapse(to, skb))
3147 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3150 if (!tcp_collapse_retrans(sk, to))
3155 /* This retransmits one SKB. Policy decisions and retransmit queue
3156 * state updates are done by the caller. Returns non-zero if an
3157 * error occurred which prevented the send.
3159 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3161 struct inet_connection_sock *icsk = inet_csk(sk);
3162 struct tcp_sock *tp = tcp_sk(sk);
3163 unsigned int cur_mss;
3167 /* Inconclusive MTU probe */
3168 if (icsk->icsk_mtup.probe_size)
3169 icsk->icsk_mtup.probe_size = 0;
3171 if (skb_still_in_host_queue(sk, skb))
3174 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
3175 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3179 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3183 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3184 return -EHOSTUNREACH; /* Routing failure or similar. */
3186 cur_mss = tcp_current_mss(sk);
3187 avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3189 /* If receiver has shrunk his window, and skb is out of
3190 * new window, do not retransmit it. The exception is the
3191 * case, when window is shrunk to zero. In this case
3192 * our retransmit of one segment serves as a zero window probe.
3194 if (avail_wnd <= 0) {
3195 if (TCP_SKB_CB(skb)->seq != tp->snd_una)
3197 avail_wnd = cur_mss;
3200 len = cur_mss * segs;
3201 if (len > avail_wnd) {
3202 len = rounddown(avail_wnd, cur_mss);
3206 if (skb->len > len) {
3207 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3208 cur_mss, GFP_ATOMIC))
3209 return -ENOMEM; /* We'll try again later. */
3211 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
3214 diff = tcp_skb_pcount(skb);
3215 tcp_set_skb_tso_segs(skb, cur_mss);
3216 diff -= tcp_skb_pcount(skb);
3218 tcp_adjust_pcount(sk, skb, diff);
3219 avail_wnd = min_t(int, avail_wnd, cur_mss);
3220 if (skb->len < avail_wnd)
3221 tcp_retrans_try_collapse(sk, skb, avail_wnd);
3224 /* RFC3168, section 6.1.1.1. ECN fallback */
3225 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3226 tcp_ecn_clear_syn(sk, skb);
3228 /* Update global and local TCP statistics. */
3229 segs = tcp_skb_pcount(skb);
3230 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3231 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3232 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3233 tp->total_retrans += segs;
3234 tp->bytes_retrans += skb->len;
3236 /* make sure skb->data is aligned on arches that require it
3237 * and check if ack-trimming & collapsing extended the headroom
3238 * beyond what csum_start can cover.
3240 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3241 skb_headroom(skb) >= 0xFFFF)) {
3242 struct sk_buff *nskb;
3244 tcp_skb_tsorted_save(skb) {
3245 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3248 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3252 } tcp_skb_tsorted_restore(skb);
3255 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3256 tcp_rate_skb_sent(sk, skb);
3259 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3262 /* To avoid taking spuriously low RTT samples based on a timestamp
3263 * for a transmit that never happened, always mark EVER_RETRANS
3265 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3267 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3268 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3269 TCP_SKB_CB(skb)->seq, segs, err);
3272 trace_tcp_retransmit_skb(sk, skb);
3273 } else if (err != -EBUSY) {
3274 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3279 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3281 struct tcp_sock *tp = tcp_sk(sk);
3282 int err = __tcp_retransmit_skb(sk, skb, segs);
3285 #if FASTRETRANS_DEBUG > 0
3286 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3287 net_dbg_ratelimited("retrans_out leaked\n");
3290 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3291 tp->retrans_out += tcp_skb_pcount(skb);
3294 /* Save stamp of the first (attempted) retransmit. */
3295 if (!tp->retrans_stamp)
3296 tp->retrans_stamp = tcp_skb_timestamp(skb);
3298 if (tp->undo_retrans < 0)
3299 tp->undo_retrans = 0;
3300 tp->undo_retrans += tcp_skb_pcount(skb);
3304 /* This gets called after a retransmit timeout, and the initially
3305 * retransmitted data is acknowledged. It tries to continue
3306 * resending the rest of the retransmit queue, until either
3307 * we've sent it all or the congestion window limit is reached.
3309 void tcp_xmit_retransmit_queue(struct sock *sk)
3311 const struct inet_connection_sock *icsk = inet_csk(sk);
3312 struct sk_buff *skb, *rtx_head, *hole = NULL;
3313 struct tcp_sock *tp = tcp_sk(sk);
3314 bool rearm_timer = false;
3318 if (!tp->packets_out)
3321 rtx_head = tcp_rtx_queue_head(sk);
3322 skb = tp->retransmit_skb_hint ?: rtx_head;
3323 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3324 skb_rbtree_walk_from(skb) {
3328 if (tcp_pacing_check(sk))
3331 /* we could do better than to assign each time */
3333 tp->retransmit_skb_hint = skb;
3335 segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp);
3338 sacked = TCP_SKB_CB(skb)->sacked;
3339 /* In case tcp_shift_skb_data() have aggregated large skbs,
3340 * we need to make sure not sending too bigs TSO packets
3342 segs = min_t(int, segs, max_segs);
3344 if (tp->retrans_out >= tp->lost_out) {
3346 } else if (!(sacked & TCPCB_LOST)) {
3347 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3352 if (icsk->icsk_ca_state != TCP_CA_Loss)
3353 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3355 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3358 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3361 if (tcp_small_queue_check(sk, skb, 1))
3364 if (tcp_retransmit_skb(sk, skb, segs))
3367 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3369 if (tcp_in_cwnd_reduction(sk))
3370 tp->prr_out += tcp_skb_pcount(skb);
3372 if (skb == rtx_head &&
3373 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3378 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3379 inet_csk(sk)->icsk_rto,
3383 /* We allow to exceed memory limits for FIN packets to expedite
3384 * connection tear down and (memory) recovery.
3385 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3386 * or even be forced to close flow without any FIN.
3387 * In general, we want to allow one skb per socket to avoid hangs
3388 * with edge trigger epoll()
3390 void sk_forced_mem_schedule(struct sock *sk, int size)
3394 delta = size - sk->sk_forward_alloc;
3397 amt = sk_mem_pages(delta);
3398 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3399 sk_memory_allocated_add(sk, amt);
3401 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3402 mem_cgroup_charge_skmem(sk->sk_memcg, amt,
3403 gfp_memcg_charge() | __GFP_NOFAIL);
3406 /* Send a FIN. The caller locks the socket for us.
3407 * We should try to send a FIN packet really hard, but eventually give up.
3409 void tcp_send_fin(struct sock *sk)
3411 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3412 struct tcp_sock *tp = tcp_sk(sk);
3414 /* Optimization, tack on the FIN if we have one skb in write queue and
3415 * this skb was not yet sent, or we are under memory pressure.
3416 * Note: in the latter case, FIN packet will be sent after a timeout,
3417 * as TCP stack thinks it has already been transmitted.
3420 if (!tskb && tcp_under_memory_pressure(sk))
3421 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3424 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3425 TCP_SKB_CB(tskb)->end_seq++;
3428 /* This means tskb was already sent.
3429 * Pretend we included the FIN on previous transmit.
3430 * We need to set tp->snd_nxt to the value it would have
3431 * if FIN had been sent. This is because retransmit path
3432 * does not change tp->snd_nxt.
3434 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3438 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3442 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3443 skb_reserve(skb, MAX_TCP_HEADER);
3444 sk_forced_mem_schedule(sk, skb->truesize);
3445 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3446 tcp_init_nondata_skb(skb, tp->write_seq,
3447 TCPHDR_ACK | TCPHDR_FIN);
3448 tcp_queue_skb(sk, skb);
3450 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3453 /* We get here when a process closes a file descriptor (either due to
3454 * an explicit close() or as a byproduct of exit()'ing) and there
3455 * was unread data in the receive queue. This behavior is recommended
3456 * by RFC 2525, section 2.17. -DaveM
3458 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3460 struct sk_buff *skb;
3462 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3464 /* NOTE: No TCP options attached and we never retransmit this. */
3465 skb = alloc_skb(MAX_TCP_HEADER, priority);
3467 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3471 /* Reserve space for headers and prepare control bits. */
3472 skb_reserve(skb, MAX_TCP_HEADER);
3473 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3474 TCPHDR_ACK | TCPHDR_RST);
3475 tcp_mstamp_refresh(tcp_sk(sk));
3477 if (tcp_transmit_skb(sk, skb, 0, priority))
3478 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3480 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3481 * skb here is different to the troublesome skb, so use NULL
3483 trace_tcp_send_reset(sk, NULL);
3486 /* Send a crossed SYN-ACK during socket establishment.
3487 * WARNING: This routine must only be called when we have already sent
3488 * a SYN packet that crossed the incoming SYN that caused this routine
3489 * to get called. If this assumption fails then the initial rcv_wnd
3490 * and rcv_wscale values will not be correct.
3492 int tcp_send_synack(struct sock *sk)
3494 struct sk_buff *skb;
3496 skb = tcp_rtx_queue_head(sk);
3497 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3498 pr_err("%s: wrong queue state\n", __func__);
3501 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3502 if (skb_cloned(skb)) {
3503 struct sk_buff *nskb;
3505 tcp_skb_tsorted_save(skb) {
3506 nskb = skb_copy(skb, GFP_ATOMIC);
3507 } tcp_skb_tsorted_restore(skb);
3510 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3511 tcp_highest_sack_replace(sk, skb, nskb);
3512 tcp_rtx_queue_unlink_and_free(skb, sk);
3513 __skb_header_release(nskb);
3514 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3515 sk_wmem_queued_add(sk, nskb->truesize);
3516 sk_mem_charge(sk, nskb->truesize);
3520 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3521 tcp_ecn_send_synack(sk, skb);
3523 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3527 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3528 * @sk: listener socket
3529 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3530 * should not use it again.
3531 * @req: request_sock pointer
3532 * @foc: cookie for tcp fast open
3533 * @synack_type: Type of synack to prepare
3534 * @syn_skb: SYN packet just received. It could be NULL for rtx case.
3536 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3537 struct request_sock *req,
3538 struct tcp_fastopen_cookie *foc,
3539 enum tcp_synack_type synack_type,
3540 struct sk_buff *syn_skb)
3542 struct inet_request_sock *ireq = inet_rsk(req);
3543 const struct tcp_sock *tp = tcp_sk(sk);
3544 struct tcp_md5sig_key *md5 = NULL;
3545 struct tcp_out_options opts;
3546 struct sk_buff *skb;
3547 int tcp_header_size;
3552 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3553 if (unlikely(!skb)) {
3557 /* Reserve space for headers. */
3558 skb_reserve(skb, MAX_TCP_HEADER);
3560 switch (synack_type) {
3561 case TCP_SYNACK_NORMAL:
3562 skb_set_owner_w(skb, req_to_sk(req));
3564 case TCP_SYNACK_COOKIE:
3565 /* Under synflood, we do not attach skb to a socket,
3566 * to avoid false sharing.
3569 case TCP_SYNACK_FASTOPEN:
3570 /* sk is a const pointer, because we want to express multiple
3571 * cpu might call us concurrently.
3572 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3574 skb_set_owner_w(skb, (struct sock *)sk);
3577 skb_dst_set(skb, dst);
3579 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3581 memset(&opts, 0, sizeof(opts));
3582 now = tcp_clock_ns();
3583 #ifdef CONFIG_SYN_COOKIES
3584 if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3585 skb->skb_mstamp_ns = cookie_init_timestamp(req, now);
3589 skb->skb_mstamp_ns = now;
3590 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3591 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3594 #ifdef CONFIG_TCP_MD5SIG
3596 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3598 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3599 /* bpf program will be interested in the tcp_flags */
3600 TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3601 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3603 syn_skb) + sizeof(*th);
3605 skb_push(skb, tcp_header_size);
3606 skb_reset_transport_header(skb);
3608 th = (struct tcphdr *)skb->data;
3609 memset(th, 0, sizeof(struct tcphdr));
3612 tcp_ecn_make_synack(req, th);
3613 th->source = htons(ireq->ir_num);
3614 th->dest = ireq->ir_rmt_port;
3615 skb->mark = ireq->ir_mark;
3616 skb->ip_summed = CHECKSUM_PARTIAL;
3617 th->seq = htonl(tcp_rsk(req)->snt_isn);
3618 /* XXX data is queued and acked as is. No buffer/window check */
3619 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3621 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3622 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3623 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3624 th->doff = (tcp_header_size >> 2);
3625 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3627 #ifdef CONFIG_TCP_MD5SIG
3628 /* Okay, we have all we need - do the md5 hash if needed */
3630 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3631 md5, req_to_sk(req), skb);
3635 bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
3636 synack_type, &opts);
3638 skb->skb_mstamp_ns = now;
3639 tcp_add_tx_delay(skb, tp);
3643 EXPORT_SYMBOL(tcp_make_synack);
3645 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3647 struct inet_connection_sock *icsk = inet_csk(sk);
3648 const struct tcp_congestion_ops *ca;
3649 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3651 if (ca_key == TCP_CA_UNSPEC)
3655 ca = tcp_ca_find_key(ca_key);
3656 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3657 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3658 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3659 icsk->icsk_ca_ops = ca;
3664 /* Do all connect socket setups that can be done AF independent. */
3665 static void tcp_connect_init(struct sock *sk)
3667 const struct dst_entry *dst = __sk_dst_get(sk);
3668 struct tcp_sock *tp = tcp_sk(sk);
3672 /* We'll fix this up when we get a response from the other end.
3673 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3675 tp->tcp_header_len = sizeof(struct tcphdr);
3676 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps))
3677 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3679 #ifdef CONFIG_TCP_MD5SIG
3680 if (tp->af_specific->md5_lookup(sk, sk))
3681 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3684 /* If user gave his TCP_MAXSEG, record it to clamp */
3685 if (tp->rx_opt.user_mss)
3686 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3689 tcp_sync_mss(sk, dst_mtu(dst));
3691 tcp_ca_dst_init(sk, dst);
3693 if (!tp->window_clamp)
3694 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3695 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3697 tcp_initialize_rcv_mss(sk);
3699 /* limit the window selection if the user enforce a smaller rx buffer */
3700 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3701 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3702 tp->window_clamp = tcp_full_space(sk);
3704 rcv_wnd = tcp_rwnd_init_bpf(sk);
3706 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3708 tcp_select_initial_window(sk, tcp_full_space(sk),
3709 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3712 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling),
3716 tp->rx_opt.rcv_wscale = rcv_wscale;
3717 tp->rcv_ssthresh = tp->rcv_wnd;
3720 sock_reset_flag(sk, SOCK_DONE);
3723 tcp_write_queue_purge(sk);
3724 tp->snd_una = tp->write_seq;
3725 tp->snd_sml = tp->write_seq;
3726 tp->snd_up = tp->write_seq;
3727 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3729 if (likely(!tp->repair))
3732 tp->rcv_tstamp = tcp_jiffies32;
3733 tp->rcv_wup = tp->rcv_nxt;
3734 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3736 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3737 inet_csk(sk)->icsk_retransmits = 0;
3738 tcp_clear_retrans(tp);
3741 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3743 struct tcp_sock *tp = tcp_sk(sk);
3744 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3746 tcb->end_seq += skb->len;
3747 __skb_header_release(skb);
3748 sk_wmem_queued_add(sk, skb->truesize);
3749 sk_mem_charge(sk, skb->truesize);
3750 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3751 tp->packets_out += tcp_skb_pcount(skb);
3754 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3755 * queue a data-only packet after the regular SYN, such that regular SYNs
3756 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3757 * only the SYN sequence, the data are retransmitted in the first ACK.
3758 * If cookie is not cached or other error occurs, falls back to send a
3759 * regular SYN with Fast Open cookie request option.
3761 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3763 struct inet_connection_sock *icsk = inet_csk(sk);
3764 struct tcp_sock *tp = tcp_sk(sk);
3765 struct tcp_fastopen_request *fo = tp->fastopen_req;
3767 struct sk_buff *syn_data;
3769 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3770 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3773 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3774 * user-MSS. Reserve maximum option space for middleboxes that add
3775 * private TCP options. The cost is reduced data space in SYN :(
3777 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3778 /* Sync mss_cache after updating the mss_clamp */
3779 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3781 space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3782 MAX_TCP_OPTION_SPACE;
3784 space = min_t(size_t, space, fo->size);
3786 /* limit to order-0 allocations */
3787 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3789 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3792 syn_data->ip_summed = CHECKSUM_PARTIAL;
3793 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3795 int copied = copy_from_iter(skb_put(syn_data, space), space,
3796 &fo->data->msg_iter);
3797 if (unlikely(!copied)) {
3798 tcp_skb_tsorted_anchor_cleanup(syn_data);
3799 kfree_skb(syn_data);
3802 if (copied != space) {
3803 skb_trim(syn_data, copied);
3806 skb_zcopy_set(syn_data, fo->uarg, NULL);
3808 /* No more data pending in inet_wait_for_connect() */
3809 if (space == fo->size)
3813 tcp_connect_queue_skb(sk, syn_data);
3815 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3817 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3819 syn->skb_mstamp_ns = syn_data->skb_mstamp_ns;
3821 /* Now full SYN+DATA was cloned and sent (or not),
3822 * remove the SYN from the original skb (syn_data)
3823 * we keep in write queue in case of a retransmit, as we
3824 * also have the SYN packet (with no data) in the same queue.
3826 TCP_SKB_CB(syn_data)->seq++;
3827 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3829 tp->syn_data = (fo->copied > 0);
3830 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3831 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3835 /* data was not sent, put it in write_queue */
3836 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3837 tp->packets_out -= tcp_skb_pcount(syn_data);
3840 /* Send a regular SYN with Fast Open cookie request option */
3841 if (fo->cookie.len > 0)
3843 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3845 tp->syn_fastopen = 0;
3847 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3851 /* Build a SYN and send it off. */
3852 int tcp_connect(struct sock *sk)
3854 struct tcp_sock *tp = tcp_sk(sk);
3855 struct sk_buff *buff;
3858 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3860 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3861 return -EHOSTUNREACH; /* Routing failure or similar. */
3863 tcp_connect_init(sk);
3865 if (unlikely(tp->repair)) {
3866 tcp_finish_connect(sk, NULL);
3870 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3871 if (unlikely(!buff))
3874 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3875 tcp_mstamp_refresh(tp);
3876 tp->retrans_stamp = tcp_time_stamp(tp);
3877 tcp_connect_queue_skb(sk, buff);
3878 tcp_ecn_send_syn(sk, buff);
3879 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3881 /* Send off SYN; include data in Fast Open. */
3882 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3883 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3884 if (err == -ECONNREFUSED)
3887 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3888 * in order to make this packet get counted in tcpOutSegs.
3890 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3891 tp->pushed_seq = tp->write_seq;
3892 buff = tcp_send_head(sk);
3893 if (unlikely(buff)) {
3894 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3895 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3897 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3899 /* Timer for repeating the SYN until an answer. */
3900 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3901 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3904 EXPORT_SYMBOL(tcp_connect);
3906 /* Send out a delayed ack, the caller does the policy checking
3907 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3910 void tcp_send_delayed_ack(struct sock *sk)
3912 struct inet_connection_sock *icsk = inet_csk(sk);
3913 int ato = icsk->icsk_ack.ato;
3914 unsigned long timeout;
3916 if (ato > TCP_DELACK_MIN) {
3917 const struct tcp_sock *tp = tcp_sk(sk);
3918 int max_ato = HZ / 2;
3920 if (inet_csk_in_pingpong_mode(sk) ||
3921 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3922 max_ato = TCP_DELACK_MAX;
3924 /* Slow path, intersegment interval is "high". */
3926 /* If some rtt estimate is known, use it to bound delayed ack.
3927 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3931 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3938 ato = min(ato, max_ato);
3941 ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max);
3943 /* Stay within the limit we were given */
3944 timeout = jiffies + ato;
3946 /* Use new timeout only if there wasn't a older one earlier. */
3947 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3948 /* If delack timer is about to expire, send ACK now. */
3949 if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3954 if (!time_before(timeout, icsk->icsk_ack.timeout))
3955 timeout = icsk->icsk_ack.timeout;
3957 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3958 icsk->icsk_ack.timeout = timeout;
3959 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3962 /* This routine sends an ack and also updates the window. */
3963 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3965 struct sk_buff *buff;
3967 /* If we have been reset, we may not send again. */
3968 if (sk->sk_state == TCP_CLOSE)
3971 /* We are not putting this on the write queue, so
3972 * tcp_transmit_skb() will set the ownership to this
3975 buff = alloc_skb(MAX_TCP_HEADER,
3976 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3977 if (unlikely(!buff)) {
3978 struct inet_connection_sock *icsk = inet_csk(sk);
3979 unsigned long delay;
3981 delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
3982 if (delay < TCP_RTO_MAX)
3983 icsk->icsk_ack.retry++;
3984 inet_csk_schedule_ack(sk);
3985 icsk->icsk_ack.ato = TCP_ATO_MIN;
3986 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
3990 /* Reserve space for headers and prepare control bits. */
3991 skb_reserve(buff, MAX_TCP_HEADER);
3992 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3994 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3996 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3998 skb_set_tcp_pure_ack(buff);
4000 /* Send it off, this clears delayed acks for us. */
4001 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
4003 EXPORT_SYMBOL_GPL(__tcp_send_ack);
4005 void tcp_send_ack(struct sock *sk)
4007 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
4010 /* This routine sends a packet with an out of date sequence
4011 * number. It assumes the other end will try to ack it.
4013 * Question: what should we make while urgent mode?
4014 * 4.4BSD forces sending single byte of data. We cannot send
4015 * out of window data, because we have SND.NXT==SND.MAX...
4017 * Current solution: to send TWO zero-length segments in urgent mode:
4018 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
4019 * out-of-date with SND.UNA-1 to probe window.
4021 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
4023 struct tcp_sock *tp = tcp_sk(sk);
4024 struct sk_buff *skb;
4026 /* We don't queue it, tcp_transmit_skb() sets ownership. */
4027 skb = alloc_skb(MAX_TCP_HEADER,
4028 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4032 /* Reserve space for headers and set control bits. */
4033 skb_reserve(skb, MAX_TCP_HEADER);
4034 /* Use a previous sequence. This should cause the other
4035 * end to send an ack. Don't queue or clone SKB, just
4038 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
4039 NET_INC_STATS(sock_net(sk), mib);
4040 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4043 /* Called from setsockopt( ... TCP_REPAIR ) */
4044 void tcp_send_window_probe(struct sock *sk)
4046 if (sk->sk_state == TCP_ESTABLISHED) {
4047 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4048 tcp_mstamp_refresh(tcp_sk(sk));
4049 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4053 /* Initiate keepalive or window probe from timer. */
4054 int tcp_write_wakeup(struct sock *sk, int mib)
4056 struct tcp_sock *tp = tcp_sk(sk);
4057 struct sk_buff *skb;
4059 if (sk->sk_state == TCP_CLOSE)
4062 skb = tcp_send_head(sk);
4063 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4065 unsigned int mss = tcp_current_mss(sk);
4066 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4068 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4069 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4071 /* We are probing the opening of a window
4072 * but the window size is != 0
4073 * must have been a result SWS avoidance ( sender )
4075 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4077 seg_size = min(seg_size, mss);
4078 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4079 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4080 skb, seg_size, mss, GFP_ATOMIC))
4082 } else if (!tcp_skb_pcount(skb))
4083 tcp_set_skb_tso_segs(skb, mss);
4085 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4086 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4088 tcp_event_new_data_sent(sk, skb);
4091 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4092 tcp_xmit_probe_skb(sk, 1, mib);
4093 return tcp_xmit_probe_skb(sk, 0, mib);
4097 /* A window probe timeout has occurred. If window is not closed send
4098 * a partial packet else a zero probe.
4100 void tcp_send_probe0(struct sock *sk)
4102 struct inet_connection_sock *icsk = inet_csk(sk);
4103 struct tcp_sock *tp = tcp_sk(sk);
4104 struct net *net = sock_net(sk);
4105 unsigned long timeout;
4108 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4110 if (tp->packets_out || tcp_write_queue_empty(sk)) {
4111 /* Cancel probe timer, if it is not required. */
4112 icsk->icsk_probes_out = 0;
4113 icsk->icsk_backoff = 0;
4114 icsk->icsk_probes_tstamp = 0;
4118 icsk->icsk_probes_out++;
4120 if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
4121 icsk->icsk_backoff++;
4122 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
4124 /* If packet was not sent due to local congestion,
4125 * Let senders fight for local resources conservatively.
4127 timeout = TCP_RESOURCE_PROBE_INTERVAL;
4130 timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout);
4131 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
4134 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4136 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4140 tcp_rsk(req)->txhash = net_tx_rndhash();
4141 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4144 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4145 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4146 if (unlikely(tcp_passive_fastopen(sk)))
4147 tcp_sk(sk)->total_retrans++;
4148 trace_tcp_retransmit_synack(sk, req);
4152 EXPORT_SYMBOL(tcp_rtx_synack);