1 // SPDX-License-Identifier: GPL-2.0-only
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
38 #define pr_fmt(fmt) "TCP: " fmt
41 #include <net/mptcp.h>
43 #include <linux/compiler.h>
44 #include <linux/gfp.h>
45 #include <linux/module.h>
46 #include <linux/static_key.h>
48 #include <trace/events/tcp.h>
50 /* Refresh clocks of a TCP socket,
51 * ensuring monotically increasing values.
53 void tcp_mstamp_refresh(struct tcp_sock *tp)
55 u64 val = tcp_clock_ns();
57 tp->tcp_clock_cache = val;
58 tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
61 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
62 int push_one, gfp_t gfp);
64 /* Account for new data that has been sent to the network. */
65 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
67 struct inet_connection_sock *icsk = inet_csk(sk);
68 struct tcp_sock *tp = tcp_sk(sk);
69 unsigned int prior_packets = tp->packets_out;
71 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
73 __skb_unlink(skb, &sk->sk_write_queue);
74 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
76 if (tp->highest_sack == NULL)
77 tp->highest_sack = skb;
79 tp->packets_out += tcp_skb_pcount(skb);
80 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
83 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
88 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
89 * window scaling factor due to loss of precision.
90 * If window has been shrunk, what should we make? It is not clear at all.
91 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
92 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
93 * invalid. OK, let's make this for now:
95 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
97 const struct tcp_sock *tp = tcp_sk(sk);
99 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
100 (tp->rx_opt.wscale_ok &&
101 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
104 return tcp_wnd_end(tp);
107 /* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
121 static __u16 tcp_advertise_mss(struct sock *sk)
123 struct tcp_sock *tp = tcp_sk(sk);
124 const struct dst_entry *dst = __sk_dst_get(sk);
125 int mss = tp->advmss;
128 unsigned int metric = dst_metric_advmss(dst);
139 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism.
142 void tcp_cwnd_restart(struct sock *sk, s32 delta)
144 struct tcp_sock *tp = tcp_sk(sk);
145 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
146 u32 cwnd = tcp_snd_cwnd(tp);
148 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
150 tp->snd_ssthresh = tcp_current_ssthresh(sk);
151 restart_cwnd = min(restart_cwnd, cwnd);
153 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
155 tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd));
156 tp->snd_cwnd_stamp = tcp_jiffies32;
157 tp->snd_cwnd_used = 0;
160 /* Congestion state accounting after a packet has been sent. */
161 static void tcp_event_data_sent(struct tcp_sock *tp,
164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const u32 now = tcp_jiffies32;
167 if (tcp_packets_in_flight(tp) == 0)
168 tcp_ca_event(sk, CA_EVENT_TX_START);
172 /* If it is a reply for ato after last received
173 * packet, enter pingpong mode.
175 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
176 inet_csk_enter_pingpong_mode(sk);
179 /* Account for an ACK we sent. */
180 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
183 struct tcp_sock *tp = tcp_sk(sk);
185 if (unlikely(tp->compressed_ack)) {
186 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
188 tp->compressed_ack = 0;
189 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
193 if (unlikely(rcv_nxt != tp->rcv_nxt))
194 return; /* Special ACK sent by DCTCP to reflect ECN */
195 tcp_dec_quickack_mode(sk, pkts);
196 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
199 /* Determine a window scaling and initial window to offer.
200 * Based on the assumption that the given amount of space
201 * will be offered. Store the results in the tp structure.
202 * NOTE: for smooth operation initial space offering should
203 * be a multiple of mss if possible. We assume here that mss >= 1.
204 * This MUST be enforced by all callers.
206 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
207 __u32 *rcv_wnd, __u32 *window_clamp,
208 int wscale_ok, __u8 *rcv_wscale,
211 unsigned int space = (__space < 0 ? 0 : __space);
213 /* If no clamp set the clamp to the max possible scaled window */
214 if (*window_clamp == 0)
215 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
216 space = min(*window_clamp, space);
218 /* Quantize space offering to a multiple of mss if possible. */
220 space = rounddown(space, mss);
222 /* NOTE: offering an initial window larger than 32767
223 * will break some buggy TCP stacks. If the admin tells us
224 * it is likely we could be speaking with such a buggy stack
225 * we will truncate our initial window offering to 32K-1
226 * unless the remote has sent us a window scaling option,
227 * which we interpret as a sign the remote TCP is not
228 * misinterpreting the window field as a signed quantity.
230 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
231 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
233 (*rcv_wnd) = min_t(u32, space, U16_MAX);
236 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
240 /* Set window scaling on max possible window */
241 space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
242 space = max_t(u32, space, READ_ONCE(sysctl_rmem_max));
243 space = min_t(u32, space, *window_clamp);
244 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
247 /* Set the clamp no higher than max representable value */
248 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
250 EXPORT_SYMBOL(tcp_select_initial_window);
252 /* Chose a new window to advertise, update state in tcp_sock for the
253 * socket, and return result with RFC1323 scaling applied. The return
254 * value can be stuffed directly into th->window for an outgoing
257 static u16 tcp_select_window(struct sock *sk)
259 struct tcp_sock *tp = tcp_sk(sk);
260 struct net *net = sock_net(sk);
261 u32 old_win = tp->rcv_wnd;
262 u32 cur_win, new_win;
264 /* Make the window 0 if we failed to queue the data because we
265 * are out of memory. The window is temporary, so we don't store
268 if (unlikely(inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOMEM))
271 cur_win = tcp_receive_window(tp);
272 new_win = __tcp_select_window(sk);
273 if (new_win < cur_win) {
274 /* Danger Will Robinson!
275 * Don't update rcv_wup/rcv_wnd here or else
276 * we will not be able to advertise a zero
277 * window in time. --DaveM
279 * Relax Will Robinson.
281 if (!READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) || !tp->rx_opt.rcv_wscale) {
282 /* Never shrink the offered window */
284 NET_INC_STATS(net, LINUX_MIB_TCPWANTZEROWINDOWADV);
285 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
289 tp->rcv_wnd = new_win;
290 tp->rcv_wup = tp->rcv_nxt;
292 /* Make sure we do not exceed the maximum possible
295 if (!tp->rx_opt.rcv_wscale &&
296 READ_ONCE(net->ipv4.sysctl_tcp_workaround_signed_windows))
297 new_win = min(new_win, MAX_TCP_WINDOW);
299 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
301 /* RFC1323 scaling applied */
302 new_win >>= tp->rx_opt.rcv_wscale;
304 /* If we advertise zero window, disable fast path. */
308 NET_INC_STATS(net, LINUX_MIB_TCPTOZEROWINDOWADV);
309 } else if (old_win == 0) {
310 NET_INC_STATS(net, LINUX_MIB_TCPFROMZEROWINDOWADV);
316 /* Packet ECN state for a SYN-ACK */
317 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
319 const struct tcp_sock *tp = tcp_sk(sk);
321 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
322 if (!(tp->ecn_flags & TCP_ECN_OK))
323 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
324 else if (tcp_ca_needs_ecn(sk) ||
325 tcp_bpf_ca_needs_ecn(sk))
329 /* Packet ECN state for a SYN. */
330 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
332 struct tcp_sock *tp = tcp_sk(sk);
333 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
334 bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 ||
335 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
338 const struct dst_entry *dst = __sk_dst_get(sk);
340 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
347 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
348 tp->ecn_flags = TCP_ECN_OK;
349 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
354 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
356 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback))
357 /* tp->ecn_flags are cleared at a later point in time when
358 * SYN ACK is ultimatively being received.
360 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
364 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
366 if (inet_rsk(req)->ecn_ok)
370 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
373 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
374 struct tcphdr *th, int tcp_header_len)
376 struct tcp_sock *tp = tcp_sk(sk);
378 if (tp->ecn_flags & TCP_ECN_OK) {
379 /* Not-retransmitted data segment: set ECT and inject CWR. */
380 if (skb->len != tcp_header_len &&
381 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
383 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
384 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
386 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
388 } else if (!tcp_ca_needs_ecn(sk)) {
389 /* ACK or retransmitted segment: clear ECT|CE */
390 INET_ECN_dontxmit(sk);
392 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
397 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
398 * auto increment end seqno.
400 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
402 skb->ip_summed = CHECKSUM_PARTIAL;
404 TCP_SKB_CB(skb)->tcp_flags = flags;
406 tcp_skb_pcount_set(skb, 1);
408 TCP_SKB_CB(skb)->seq = seq;
409 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
411 TCP_SKB_CB(skb)->end_seq = seq;
414 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
416 return tp->snd_una != tp->snd_up;
419 #define OPTION_SACK_ADVERTISE BIT(0)
420 #define OPTION_TS BIT(1)
421 #define OPTION_MD5 BIT(2)
422 #define OPTION_WSCALE BIT(3)
423 #define OPTION_FAST_OPEN_COOKIE BIT(8)
424 #define OPTION_SMC BIT(9)
425 #define OPTION_MPTCP BIT(10)
427 static void smc_options_write(__be32 *ptr, u16 *options)
429 #if IS_ENABLED(CONFIG_SMC)
430 if (static_branch_unlikely(&tcp_have_smc)) {
431 if (unlikely(OPTION_SMC & *options)) {
432 *ptr++ = htonl((TCPOPT_NOP << 24) |
435 (TCPOLEN_EXP_SMC_BASE));
436 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
442 struct tcp_out_options {
443 u16 options; /* bit field of OPTION_* */
444 u16 mss; /* 0 to disable */
445 u8 ws; /* window scale, 0 to disable */
446 u8 num_sack_blocks; /* number of SACK blocks to include */
447 u8 hash_size; /* bytes in hash_location */
448 u8 bpf_opt_len; /* length of BPF hdr option */
449 __u8 *hash_location; /* temporary pointer, overloaded */
450 __u32 tsval, tsecr; /* need to include OPTION_TS */
451 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
452 struct mptcp_out_options mptcp;
455 static void mptcp_options_write(struct tcphdr *th, __be32 *ptr,
457 struct tcp_out_options *opts)
459 #if IS_ENABLED(CONFIG_MPTCP)
460 if (unlikely(OPTION_MPTCP & opts->options))
461 mptcp_write_options(th, ptr, tp, &opts->mptcp);
465 #ifdef CONFIG_CGROUP_BPF
466 static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
467 enum tcp_synack_type synack_type)
470 return BPF_WRITE_HDR_TCP_CURRENT_MSS;
472 if (unlikely(synack_type == TCP_SYNACK_COOKIE))
473 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
478 /* req, syn_skb and synack_type are used when writing synack */
479 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
480 struct request_sock *req,
481 struct sk_buff *syn_skb,
482 enum tcp_synack_type synack_type,
483 struct tcp_out_options *opts,
484 unsigned int *remaining)
486 struct bpf_sock_ops_kern sock_ops;
489 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
490 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
494 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
497 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
499 sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
502 /* The listen "sk" cannot be passed here because
503 * it is not locked. It would not make too much
504 * sense to do bpf_setsockopt(listen_sk) based
505 * on individual connection request also.
507 * Thus, "req" is passed here and the cgroup-bpf-progs
508 * of the listen "sk" will be run.
510 * "req" is also used here for fastopen even the "sk" here is
511 * a fullsock "child" sk. It is to keep the behavior
512 * consistent between fastopen and non-fastopen on
513 * the bpf programming side.
515 sock_ops.sk = (struct sock *)req;
516 sock_ops.syn_skb = syn_skb;
518 sock_owned_by_me(sk);
520 sock_ops.is_fullsock = 1;
524 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
525 sock_ops.remaining_opt_len = *remaining;
526 /* tcp_current_mss() does not pass a skb */
528 bpf_skops_init_skb(&sock_ops, skb, 0);
530 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
532 if (err || sock_ops.remaining_opt_len == *remaining)
535 opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
536 /* round up to 4 bytes */
537 opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
539 *remaining -= opts->bpf_opt_len;
542 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
543 struct request_sock *req,
544 struct sk_buff *syn_skb,
545 enum tcp_synack_type synack_type,
546 struct tcp_out_options *opts)
548 u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
549 struct bpf_sock_ops_kern sock_ops;
552 if (likely(!max_opt_len))
555 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
557 sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
560 sock_ops.sk = (struct sock *)req;
561 sock_ops.syn_skb = syn_skb;
563 sock_owned_by_me(sk);
565 sock_ops.is_fullsock = 1;
569 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
570 sock_ops.remaining_opt_len = max_opt_len;
571 first_opt_off = tcp_hdrlen(skb) - max_opt_len;
572 bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
574 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
579 nr_written = max_opt_len - sock_ops.remaining_opt_len;
581 if (nr_written < max_opt_len)
582 memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
583 max_opt_len - nr_written);
586 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
587 struct request_sock *req,
588 struct sk_buff *syn_skb,
589 enum tcp_synack_type synack_type,
590 struct tcp_out_options *opts,
591 unsigned int *remaining)
595 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
596 struct request_sock *req,
597 struct sk_buff *syn_skb,
598 enum tcp_synack_type synack_type,
599 struct tcp_out_options *opts)
604 /* Write previously computed TCP options to the packet.
606 * Beware: Something in the Internet is very sensitive to the ordering of
607 * TCP options, we learned this through the hard way, so be careful here.
608 * Luckily we can at least blame others for their non-compliance but from
609 * inter-operability perspective it seems that we're somewhat stuck with
610 * the ordering which we have been using if we want to keep working with
611 * those broken things (not that it currently hurts anybody as there isn't
612 * particular reason why the ordering would need to be changed).
614 * At least SACK_PERM as the first option is known to lead to a disaster
615 * (but it may well be that other scenarios fail similarly).
617 static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp,
618 struct tcp_out_options *opts)
620 __be32 *ptr = (__be32 *)(th + 1);
621 u16 options = opts->options; /* mungable copy */
623 if (unlikely(OPTION_MD5 & options)) {
624 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
625 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
626 /* overload cookie hash location */
627 opts->hash_location = (__u8 *)ptr;
631 if (unlikely(opts->mss)) {
632 *ptr++ = htonl((TCPOPT_MSS << 24) |
633 (TCPOLEN_MSS << 16) |
637 if (likely(OPTION_TS & options)) {
638 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
639 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
640 (TCPOLEN_SACK_PERM << 16) |
641 (TCPOPT_TIMESTAMP << 8) |
643 options &= ~OPTION_SACK_ADVERTISE;
645 *ptr++ = htonl((TCPOPT_NOP << 24) |
647 (TCPOPT_TIMESTAMP << 8) |
650 *ptr++ = htonl(opts->tsval);
651 *ptr++ = htonl(opts->tsecr);
654 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
655 *ptr++ = htonl((TCPOPT_NOP << 24) |
657 (TCPOPT_SACK_PERM << 8) |
661 if (unlikely(OPTION_WSCALE & options)) {
662 *ptr++ = htonl((TCPOPT_NOP << 24) |
663 (TCPOPT_WINDOW << 16) |
664 (TCPOLEN_WINDOW << 8) |
668 if (unlikely(opts->num_sack_blocks)) {
669 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
670 tp->duplicate_sack : tp->selective_acks;
673 *ptr++ = htonl((TCPOPT_NOP << 24) |
676 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
677 TCPOLEN_SACK_PERBLOCK)));
679 for (this_sack = 0; this_sack < opts->num_sack_blocks;
681 *ptr++ = htonl(sp[this_sack].start_seq);
682 *ptr++ = htonl(sp[this_sack].end_seq);
685 tp->rx_opt.dsack = 0;
688 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
689 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
691 u32 len; /* Fast Open option length */
694 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
695 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
696 TCPOPT_FASTOPEN_MAGIC);
697 p += TCPOLEN_EXP_FASTOPEN_BASE;
699 len = TCPOLEN_FASTOPEN_BASE + foc->len;
700 *p++ = TCPOPT_FASTOPEN;
704 memcpy(p, foc->val, foc->len);
705 if ((len & 3) == 2) {
706 p[foc->len] = TCPOPT_NOP;
707 p[foc->len + 1] = TCPOPT_NOP;
709 ptr += (len + 3) >> 2;
712 smc_options_write(ptr, &options);
714 mptcp_options_write(th, ptr, tp, opts);
717 static void smc_set_option(const struct tcp_sock *tp,
718 struct tcp_out_options *opts,
719 unsigned int *remaining)
721 #if IS_ENABLED(CONFIG_SMC)
722 if (static_branch_unlikely(&tcp_have_smc)) {
724 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
725 opts->options |= OPTION_SMC;
726 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
733 static void smc_set_option_cond(const struct tcp_sock *tp,
734 const struct inet_request_sock *ireq,
735 struct tcp_out_options *opts,
736 unsigned int *remaining)
738 #if IS_ENABLED(CONFIG_SMC)
739 if (static_branch_unlikely(&tcp_have_smc)) {
740 if (tp->syn_smc && ireq->smc_ok) {
741 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
742 opts->options |= OPTION_SMC;
743 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
750 static void mptcp_set_option_cond(const struct request_sock *req,
751 struct tcp_out_options *opts,
752 unsigned int *remaining)
754 if (rsk_is_mptcp(req)) {
757 if (mptcp_synack_options(req, &size, &opts->mptcp)) {
758 if (*remaining >= size) {
759 opts->options |= OPTION_MPTCP;
766 /* Compute TCP options for SYN packets. This is not the final
767 * network wire format yet.
769 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
770 struct tcp_out_options *opts,
771 struct tcp_md5sig_key **md5)
773 struct tcp_sock *tp = tcp_sk(sk);
774 unsigned int remaining = MAX_TCP_OPTION_SPACE;
775 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
778 #ifdef CONFIG_TCP_MD5SIG
779 if (static_branch_unlikely(&tcp_md5_needed.key) &&
780 rcu_access_pointer(tp->md5sig_info)) {
781 *md5 = tp->af_specific->md5_lookup(sk, sk);
783 opts->options |= OPTION_MD5;
784 remaining -= TCPOLEN_MD5SIG_ALIGNED;
789 /* We always get an MSS option. The option bytes which will be seen in
790 * normal data packets should timestamps be used, must be in the MSS
791 * advertised. But we subtract them from tp->mss_cache so that
792 * calculations in tcp_sendmsg are simpler etc. So account for this
793 * fact here if necessary. If we don't do this correctly, as a
794 * receiver we won't recognize data packets as being full sized when we
795 * should, and thus we won't abide by the delayed ACK rules correctly.
796 * SACKs don't matter, we never delay an ACK when we have any of those
798 opts->mss = tcp_advertise_mss(sk);
799 remaining -= TCPOLEN_MSS_ALIGNED;
801 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps) && !*md5)) {
802 opts->options |= OPTION_TS;
803 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
804 opts->tsecr = tp->rx_opt.ts_recent;
805 remaining -= TCPOLEN_TSTAMP_ALIGNED;
807 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) {
808 opts->ws = tp->rx_opt.rcv_wscale;
809 opts->options |= OPTION_WSCALE;
810 remaining -= TCPOLEN_WSCALE_ALIGNED;
812 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) {
813 opts->options |= OPTION_SACK_ADVERTISE;
814 if (unlikely(!(OPTION_TS & opts->options)))
815 remaining -= TCPOLEN_SACKPERM_ALIGNED;
818 if (fastopen && fastopen->cookie.len >= 0) {
819 u32 need = fastopen->cookie.len;
821 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
822 TCPOLEN_FASTOPEN_BASE;
823 need = (need + 3) & ~3U; /* Align to 32 bits */
824 if (remaining >= need) {
825 opts->options |= OPTION_FAST_OPEN_COOKIE;
826 opts->fastopen_cookie = &fastopen->cookie;
828 tp->syn_fastopen = 1;
829 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
833 smc_set_option(tp, opts, &remaining);
835 if (sk_is_mptcp(sk)) {
838 if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
839 opts->options |= OPTION_MPTCP;
844 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
846 return MAX_TCP_OPTION_SPACE - remaining;
849 /* Set up TCP options for SYN-ACKs. */
850 static unsigned int tcp_synack_options(const struct sock *sk,
851 struct request_sock *req,
852 unsigned int mss, struct sk_buff *skb,
853 struct tcp_out_options *opts,
854 const struct tcp_md5sig_key *md5,
855 struct tcp_fastopen_cookie *foc,
856 enum tcp_synack_type synack_type,
857 struct sk_buff *syn_skb)
859 struct inet_request_sock *ireq = inet_rsk(req);
860 unsigned int remaining = MAX_TCP_OPTION_SPACE;
862 #ifdef CONFIG_TCP_MD5SIG
864 opts->options |= OPTION_MD5;
865 remaining -= TCPOLEN_MD5SIG_ALIGNED;
867 /* We can't fit any SACK blocks in a packet with MD5 + TS
868 * options. There was discussion about disabling SACK
869 * rather than TS in order to fit in better with old,
870 * buggy kernels, but that was deemed to be unnecessary.
872 if (synack_type != TCP_SYNACK_COOKIE)
873 ireq->tstamp_ok &= !ireq->sack_ok;
877 /* We always send an MSS option. */
879 remaining -= TCPOLEN_MSS_ALIGNED;
881 if (likely(ireq->wscale_ok)) {
882 opts->ws = ireq->rcv_wscale;
883 opts->options |= OPTION_WSCALE;
884 remaining -= TCPOLEN_WSCALE_ALIGNED;
886 if (likely(ireq->tstamp_ok)) {
887 opts->options |= OPTION_TS;
888 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
889 opts->tsecr = READ_ONCE(req->ts_recent);
890 remaining -= TCPOLEN_TSTAMP_ALIGNED;
892 if (likely(ireq->sack_ok)) {
893 opts->options |= OPTION_SACK_ADVERTISE;
894 if (unlikely(!ireq->tstamp_ok))
895 remaining -= TCPOLEN_SACKPERM_ALIGNED;
897 if (foc != NULL && foc->len >= 0) {
900 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
901 TCPOLEN_FASTOPEN_BASE;
902 need = (need + 3) & ~3U; /* Align to 32 bits */
903 if (remaining >= need) {
904 opts->options |= OPTION_FAST_OPEN_COOKIE;
905 opts->fastopen_cookie = foc;
910 mptcp_set_option_cond(req, opts, &remaining);
912 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
914 bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb,
915 synack_type, opts, &remaining);
917 return MAX_TCP_OPTION_SPACE - remaining;
920 /* Compute TCP options for ESTABLISHED sockets. This is not the
921 * final wire format yet.
923 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
924 struct tcp_out_options *opts,
925 struct tcp_md5sig_key **md5)
927 struct tcp_sock *tp = tcp_sk(sk);
928 unsigned int size = 0;
929 unsigned int eff_sacks;
934 #ifdef CONFIG_TCP_MD5SIG
935 if (static_branch_unlikely(&tcp_md5_needed.key) &&
936 rcu_access_pointer(tp->md5sig_info)) {
937 *md5 = tp->af_specific->md5_lookup(sk, sk);
939 opts->options |= OPTION_MD5;
940 size += TCPOLEN_MD5SIG_ALIGNED;
945 if (likely(tp->rx_opt.tstamp_ok)) {
946 opts->options |= OPTION_TS;
947 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
948 opts->tsecr = tp->rx_opt.ts_recent;
949 size += TCPOLEN_TSTAMP_ALIGNED;
952 /* MPTCP options have precedence over SACK for the limited TCP
953 * option space because a MPTCP connection would be forced to
954 * fall back to regular TCP if a required multipath option is
955 * missing. SACK still gets a chance to use whatever space is
958 if (sk_is_mptcp(sk)) {
959 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
960 unsigned int opt_size = 0;
962 if (mptcp_established_options(sk, skb, &opt_size, remaining,
964 opts->options |= OPTION_MPTCP;
969 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
970 if (unlikely(eff_sacks)) {
971 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
972 if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
973 TCPOLEN_SACK_PERBLOCK))
976 opts->num_sack_blocks =
977 min_t(unsigned int, eff_sacks,
978 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
979 TCPOLEN_SACK_PERBLOCK);
981 size += TCPOLEN_SACK_BASE_ALIGNED +
982 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
985 if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
986 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
987 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
989 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
991 size = MAX_TCP_OPTION_SPACE - remaining;
998 /* TCP SMALL QUEUES (TSQ)
1000 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
1001 * to reduce RTT and bufferbloat.
1002 * We do this using a special skb destructor (tcp_wfree).
1004 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
1005 * needs to be reallocated in a driver.
1006 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
1008 * Since transmit from skb destructor is forbidden, we use a tasklet
1009 * to process all sockets that eventually need to send more skbs.
1010 * We use one tasklet per cpu, with its own queue of sockets.
1012 struct tsq_tasklet {
1013 struct tasklet_struct tasklet;
1014 struct list_head head; /* queue of tcp sockets */
1016 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
1018 static void tcp_tsq_write(struct sock *sk)
1020 if ((1 << sk->sk_state) &
1021 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1022 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
1023 struct tcp_sock *tp = tcp_sk(sk);
1025 if (tp->lost_out > tp->retrans_out &&
1026 tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) {
1027 tcp_mstamp_refresh(tp);
1028 tcp_xmit_retransmit_queue(sk);
1031 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
1036 static void tcp_tsq_handler(struct sock *sk)
1039 if (!sock_owned_by_user(sk))
1041 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
1046 * One tasklet per cpu tries to send more skbs.
1047 * We run in tasklet context but need to disable irqs when
1048 * transferring tsq->head because tcp_wfree() might
1049 * interrupt us (non NAPI drivers)
1051 static void tcp_tasklet_func(struct tasklet_struct *t)
1053 struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet);
1055 unsigned long flags;
1056 struct list_head *q, *n;
1057 struct tcp_sock *tp;
1060 local_irq_save(flags);
1061 list_splice_init(&tsq->head, &list);
1062 local_irq_restore(flags);
1064 list_for_each_safe(q, n, &list) {
1065 tp = list_entry(q, struct tcp_sock, tsq_node);
1066 list_del(&tp->tsq_node);
1068 sk = (struct sock *)tp;
1069 smp_mb__before_atomic();
1070 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
1072 tcp_tsq_handler(sk);
1077 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
1078 TCPF_WRITE_TIMER_DEFERRED | \
1079 TCPF_DELACK_TIMER_DEFERRED | \
1080 TCPF_MTU_REDUCED_DEFERRED)
1082 * tcp_release_cb - tcp release_sock() callback
1085 * called from release_sock() to perform protocol dependent
1086 * actions before socket release.
1088 void tcp_release_cb(struct sock *sk)
1090 unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags);
1091 unsigned long nflags;
1093 /* perform an atomic operation only if at least one flag is set */
1095 if (!(flags & TCP_DEFERRED_ALL))
1097 nflags = flags & ~TCP_DEFERRED_ALL;
1098 } while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags));
1100 if (flags & TCPF_TSQ_DEFERRED) {
1104 /* Here begins the tricky part :
1105 * We are called from release_sock() with :
1107 * 2) sk_lock.slock spinlock held
1108 * 3) socket owned by us (sk->sk_lock.owned == 1)
1110 * But following code is meant to be called from BH handlers,
1111 * so we should keep BH disabled, but early release socket ownership
1113 sock_release_ownership(sk);
1115 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1116 tcp_write_timer_handler(sk);
1119 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1120 tcp_delack_timer_handler(sk);
1123 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1124 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1128 EXPORT_SYMBOL(tcp_release_cb);
1130 void __init tcp_tasklet_init(void)
1134 for_each_possible_cpu(i) {
1135 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
1137 INIT_LIST_HEAD(&tsq->head);
1138 tasklet_setup(&tsq->tasklet, tcp_tasklet_func);
1143 * Write buffer destructor automatically called from kfree_skb.
1144 * We can't xmit new skbs from this context, as we might already
1147 void tcp_wfree(struct sk_buff *skb)
1149 struct sock *sk = skb->sk;
1150 struct tcp_sock *tp = tcp_sk(sk);
1151 unsigned long flags, nval, oval;
1152 struct tsq_tasklet *tsq;
1155 /* Keep one reference on sk_wmem_alloc.
1156 * Will be released by sk_free() from here or tcp_tasklet_func()
1158 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1160 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
1161 * Wait until our queues (qdisc + devices) are drained.
1163 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1164 * - chance for incoming ACK (processed by another cpu maybe)
1165 * to migrate this flow (skb->ooo_okay will be eventually set)
1167 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1170 oval = smp_load_acquire(&sk->sk_tsq_flags);
1172 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1175 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1176 } while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval));
1178 /* queue this socket to tasklet queue */
1179 local_irq_save(flags);
1180 tsq = this_cpu_ptr(&tsq_tasklet);
1181 empty = list_empty(&tsq->head);
1182 list_add(&tp->tsq_node, &tsq->head);
1184 tasklet_schedule(&tsq->tasklet);
1185 local_irq_restore(flags);
1191 /* Note: Called under soft irq.
1192 * We can call TCP stack right away, unless socket is owned by user.
1194 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1196 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1197 struct sock *sk = (struct sock *)tp;
1199 tcp_tsq_handler(sk);
1202 return HRTIMER_NORESTART;
1205 static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1208 struct tcp_sock *tp = tcp_sk(sk);
1210 if (sk->sk_pacing_status != SK_PACING_NONE) {
1211 unsigned long rate = sk->sk_pacing_rate;
1213 /* Original sch_fq does not pace first 10 MSS
1214 * Note that tp->data_segs_out overflows after 2^32 packets,
1215 * this is a minor annoyance.
1217 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1218 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1219 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1221 /* take into account OS jitter */
1222 len_ns -= min_t(u64, len_ns / 2, credit);
1223 tp->tcp_wstamp_ns += len_ns;
1226 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1229 INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1230 INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1231 INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1233 /* This routine actually transmits TCP packets queued in by
1234 * tcp_do_sendmsg(). This is used by both the initial
1235 * transmission and possible later retransmissions.
1236 * All SKB's seen here are completely headerless. It is our
1237 * job to build the TCP header, and pass the packet down to
1238 * IP so it can do the same plus pass the packet off to the
1241 * We are working here with either a clone of the original
1242 * SKB, or a fresh unique copy made by the retransmit engine.
1244 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1245 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1247 const struct inet_connection_sock *icsk = inet_csk(sk);
1248 struct inet_sock *inet;
1249 struct tcp_sock *tp;
1250 struct tcp_skb_cb *tcb;
1251 struct tcp_out_options opts;
1252 unsigned int tcp_options_size, tcp_header_size;
1253 struct sk_buff *oskb = NULL;
1254 struct tcp_md5sig_key *md5;
1259 BUG_ON(!skb || !tcp_skb_pcount(skb));
1261 prior_wstamp = tp->tcp_wstamp_ns;
1262 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1263 skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
1267 tcp_skb_tsorted_save(oskb) {
1268 if (unlikely(skb_cloned(oskb)))
1269 skb = pskb_copy(oskb, gfp_mask);
1271 skb = skb_clone(oskb, gfp_mask);
1272 } tcp_skb_tsorted_restore(oskb);
1276 /* retransmit skbs might have a non zero value in skb->dev
1277 * because skb->dev is aliased with skb->rbnode.rb_left
1283 tcb = TCP_SKB_CB(skb);
1284 memset(&opts, 0, sizeof(opts));
1286 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1287 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1289 tcp_options_size = tcp_established_options(sk, skb, &opts,
1291 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1292 * at receiver : This slightly improve GRO performance.
1293 * Note that we do not force the PSH flag for non GSO packets,
1294 * because they might be sent under high congestion events,
1295 * and in this case it is better to delay the delivery of 1-MSS
1296 * packets and thus the corresponding ACK packet that would
1297 * release the following packet.
1299 if (tcp_skb_pcount(skb) > 1)
1300 tcb->tcp_flags |= TCPHDR_PSH;
1302 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1304 /* if no packet is in qdisc/device queue, then allow XPS to select
1305 * another queue. We can be called from tcp_tsq_handler()
1306 * which holds one reference to sk.
1308 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1309 * One way to get this would be to set skb->truesize = 2 on them.
1311 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1313 /* If we had to use memory reserve to allocate this skb,
1314 * this might cause drops if packet is looped back :
1315 * Other socket might not have SOCK_MEMALLOC.
1316 * Packets not looped back do not care about pfmemalloc.
1318 skb->pfmemalloc = 0;
1320 skb_push(skb, tcp_header_size);
1321 skb_reset_transport_header(skb);
1325 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1326 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1328 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1330 /* Build TCP header and checksum it. */
1331 th = (struct tcphdr *)skb->data;
1332 th->source = inet->inet_sport;
1333 th->dest = inet->inet_dport;
1334 th->seq = htonl(tcb->seq);
1335 th->ack_seq = htonl(rcv_nxt);
1336 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1342 /* The urg_mode check is necessary during a below snd_una win probe */
1343 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1344 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1345 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1347 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1348 th->urg_ptr = htons(0xFFFF);
1353 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1354 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1355 th->window = htons(tcp_select_window(sk));
1356 tcp_ecn_send(sk, skb, th, tcp_header_size);
1358 /* RFC1323: The window in SYN & SYN/ACK segments
1361 th->window = htons(min(tp->rcv_wnd, 65535U));
1364 tcp_options_write(th, tp, &opts);
1366 #ifdef CONFIG_TCP_MD5SIG
1367 /* Calculate the MD5 hash, as we have all we need now */
1370 tp->af_specific->calc_md5_hash(opts.hash_location,
1375 /* BPF prog is the last one writing header option */
1376 bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1378 INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1379 tcp_v6_send_check, tcp_v4_send_check,
1382 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1383 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1385 if (skb->len != tcp_header_size) {
1386 tcp_event_data_sent(tp, sk);
1387 tp->data_segs_out += tcp_skb_pcount(skb);
1388 tp->bytes_sent += skb->len - tcp_header_size;
1391 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1392 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1393 tcp_skb_pcount(skb));
1395 tp->segs_out += tcp_skb_pcount(skb);
1396 skb_set_hash_from_sk(skb, sk);
1397 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1398 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1399 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1401 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1403 /* Cleanup our debris for IP stacks */
1404 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1405 sizeof(struct inet6_skb_parm)));
1407 tcp_add_tx_delay(skb, tp);
1409 err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1410 inet6_csk_xmit, ip_queue_xmit,
1411 sk, skb, &inet->cork.fl);
1413 if (unlikely(err > 0)) {
1415 err = net_xmit_eval(err);
1418 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1419 tcp_rate_skb_sent(sk, oskb);
1424 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1427 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1428 tcp_sk(sk)->rcv_nxt);
1431 /* This routine just queues the buffer for sending.
1433 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1434 * otherwise socket can stall.
1436 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1438 struct tcp_sock *tp = tcp_sk(sk);
1440 /* Advance write_seq and place onto the write_queue. */
1441 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1442 __skb_header_release(skb);
1443 tcp_add_write_queue_tail(sk, skb);
1444 sk_wmem_queued_add(sk, skb->truesize);
1445 sk_mem_charge(sk, skb->truesize);
1448 /* Initialize TSO segments for a packet. */
1449 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1451 if (skb->len <= mss_now) {
1452 /* Avoid the costly divide in the normal
1455 tcp_skb_pcount_set(skb, 1);
1456 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1458 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1459 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1463 /* Pcount in the middle of the write queue got changed, we need to do various
1464 * tweaks to fix counters
1466 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1468 struct tcp_sock *tp = tcp_sk(sk);
1470 tp->packets_out -= decr;
1472 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1473 tp->sacked_out -= decr;
1474 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1475 tp->retrans_out -= decr;
1476 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1477 tp->lost_out -= decr;
1479 /* Reno case is special. Sigh... */
1480 if (tcp_is_reno(tp) && decr > 0)
1481 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1483 if (tp->lost_skb_hint &&
1484 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1485 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1486 tp->lost_cnt_hint -= decr;
1488 tcp_verify_left_out(tp);
1491 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1493 return TCP_SKB_CB(skb)->txstamp_ack ||
1494 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1497 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1499 struct skb_shared_info *shinfo = skb_shinfo(skb);
1501 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1502 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1503 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1504 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1506 shinfo->tx_flags &= ~tsflags;
1507 shinfo2->tx_flags |= tsflags;
1508 swap(shinfo->tskey, shinfo2->tskey);
1509 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1510 TCP_SKB_CB(skb)->txstamp_ack = 0;
1514 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1516 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1517 TCP_SKB_CB(skb)->eor = 0;
1520 /* Insert buff after skb on the write or rtx queue of sk. */
1521 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1522 struct sk_buff *buff,
1524 enum tcp_queue tcp_queue)
1526 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1527 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1529 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1532 /* Function to create two new TCP segments. Shrinks the given segment
1533 * to the specified size and appends a new segment with the rest of the
1534 * packet to the list. This won't be called frequently, I hope.
1535 * Remember, these are still headerless SKBs at this point.
1537 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1538 struct sk_buff *skb, u32 len,
1539 unsigned int mss_now, gfp_t gfp)
1541 struct tcp_sock *tp = tcp_sk(sk);
1542 struct sk_buff *buff;
1548 if (WARN_ON(len > skb->len))
1551 DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb));
1553 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1554 * We need some allowance to not penalize applications setting small
1556 * Also allow first and last skb in retransmit queue to be split.
1558 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE);
1559 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1560 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1561 skb != tcp_rtx_queue_head(sk) &&
1562 skb != tcp_rtx_queue_tail(sk))) {
1563 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1567 if (skb_unclone_keeptruesize(skb, gfp))
1570 /* Get a new skb... force flag on. */
1571 buff = tcp_stream_alloc_skb(sk, gfp, true);
1573 return -ENOMEM; /* We'll just try again later. */
1574 skb_copy_decrypted(buff, skb);
1575 mptcp_skb_ext_copy(buff, skb);
1577 sk_wmem_queued_add(sk, buff->truesize);
1578 sk_mem_charge(sk, buff->truesize);
1579 nlen = skb->len - len;
1580 buff->truesize += nlen;
1581 skb->truesize -= nlen;
1583 /* Correct the sequence numbers. */
1584 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1585 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1586 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1588 /* PSH and FIN should only be set in the second packet. */
1589 flags = TCP_SKB_CB(skb)->tcp_flags;
1590 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1591 TCP_SKB_CB(buff)->tcp_flags = flags;
1592 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1593 tcp_skb_fragment_eor(skb, buff);
1595 skb_split(skb, buff, len);
1597 skb_set_delivery_time(buff, skb->tstamp, true);
1598 tcp_fragment_tstamp(skb, buff);
1600 old_factor = tcp_skb_pcount(skb);
1602 /* Fix up tso_factor for both original and new SKB. */
1603 tcp_set_skb_tso_segs(skb, mss_now);
1604 tcp_set_skb_tso_segs(buff, mss_now);
1606 /* Update delivered info for the new segment */
1607 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1609 /* If this packet has been sent out already, we must
1610 * adjust the various packet counters.
1612 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1613 int diff = old_factor - tcp_skb_pcount(skb) -
1614 tcp_skb_pcount(buff);
1617 tcp_adjust_pcount(sk, skb, diff);
1620 /* Link BUFF into the send queue. */
1621 __skb_header_release(buff);
1622 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1623 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1624 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1629 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1630 * data is not copied, but immediately discarded.
1632 static int __pskb_trim_head(struct sk_buff *skb, int len)
1634 struct skb_shared_info *shinfo;
1637 DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb));
1640 shinfo = skb_shinfo(skb);
1641 for (i = 0; i < shinfo->nr_frags; i++) {
1642 int size = skb_frag_size(&shinfo->frags[i]);
1645 skb_frag_unref(skb, i);
1648 shinfo->frags[k] = shinfo->frags[i];
1650 skb_frag_off_add(&shinfo->frags[k], eat);
1651 skb_frag_size_sub(&shinfo->frags[k], eat);
1657 shinfo->nr_frags = k;
1659 skb->data_len -= len;
1660 skb->len = skb->data_len;
1664 /* Remove acked data from a packet in the transmit queue. */
1665 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1669 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
1672 delta_truesize = __pskb_trim_head(skb, len);
1674 TCP_SKB_CB(skb)->seq += len;
1676 skb->truesize -= delta_truesize;
1677 sk_wmem_queued_add(sk, -delta_truesize);
1678 if (!skb_zcopy_pure(skb))
1679 sk_mem_uncharge(sk, delta_truesize);
1681 /* Any change of skb->len requires recalculation of tso factor. */
1682 if (tcp_skb_pcount(skb) > 1)
1683 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1688 /* Calculate MSS not accounting any TCP options. */
1689 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1691 const struct tcp_sock *tp = tcp_sk(sk);
1692 const struct inet_connection_sock *icsk = inet_csk(sk);
1695 /* Calculate base mss without TCP options:
1696 It is MMS_S - sizeof(tcphdr) of rfc1122
1698 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1700 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1701 if (icsk->icsk_af_ops->net_frag_header_len) {
1702 const struct dst_entry *dst = __sk_dst_get(sk);
1704 if (dst && dst_allfrag(dst))
1705 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1708 /* Clamp it (mss_clamp does not include tcp options) */
1709 if (mss_now > tp->rx_opt.mss_clamp)
1710 mss_now = tp->rx_opt.mss_clamp;
1712 /* Now subtract optional transport overhead */
1713 mss_now -= icsk->icsk_ext_hdr_len;
1715 /* Then reserve room for full set of TCP options and 8 bytes of data */
1716 mss_now = max(mss_now,
1717 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss));
1721 /* Calculate MSS. Not accounting for SACKs here. */
1722 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1724 /* Subtract TCP options size, not including SACKs */
1725 return __tcp_mtu_to_mss(sk, pmtu) -
1726 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1728 EXPORT_SYMBOL(tcp_mtu_to_mss);
1730 /* Inverse of above */
1731 int tcp_mss_to_mtu(struct sock *sk, int mss)
1733 const struct tcp_sock *tp = tcp_sk(sk);
1734 const struct inet_connection_sock *icsk = inet_csk(sk);
1738 tp->tcp_header_len +
1739 icsk->icsk_ext_hdr_len +
1740 icsk->icsk_af_ops->net_header_len;
1742 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1743 if (icsk->icsk_af_ops->net_frag_header_len) {
1744 const struct dst_entry *dst = __sk_dst_get(sk);
1746 if (dst && dst_allfrag(dst))
1747 mtu += icsk->icsk_af_ops->net_frag_header_len;
1751 EXPORT_SYMBOL(tcp_mss_to_mtu);
1753 /* MTU probing init per socket */
1754 void tcp_mtup_init(struct sock *sk)
1756 struct tcp_sock *tp = tcp_sk(sk);
1757 struct inet_connection_sock *icsk = inet_csk(sk);
1758 struct net *net = sock_net(sk);
1760 icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1761 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1762 icsk->icsk_af_ops->net_header_len;
1763 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss));
1764 icsk->icsk_mtup.probe_size = 0;
1765 if (icsk->icsk_mtup.enabled)
1766 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1768 EXPORT_SYMBOL(tcp_mtup_init);
1770 /* This function synchronize snd mss to current pmtu/exthdr set.
1772 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1773 for TCP options, but includes only bare TCP header.
1775 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1776 It is minimum of user_mss and mss received with SYN.
1777 It also does not include TCP options.
1779 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1781 tp->mss_cache is current effective sending mss, including
1782 all tcp options except for SACKs. It is evaluated,
1783 taking into account current pmtu, but never exceeds
1784 tp->rx_opt.mss_clamp.
1786 NOTE1. rfc1122 clearly states that advertised MSS
1787 DOES NOT include either tcp or ip options.
1789 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1790 are READ ONLY outside this function. --ANK (980731)
1792 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1794 struct tcp_sock *tp = tcp_sk(sk);
1795 struct inet_connection_sock *icsk = inet_csk(sk);
1798 if (icsk->icsk_mtup.search_high > pmtu)
1799 icsk->icsk_mtup.search_high = pmtu;
1801 mss_now = tcp_mtu_to_mss(sk, pmtu);
1802 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1804 /* And store cached results */
1805 icsk->icsk_pmtu_cookie = pmtu;
1806 if (icsk->icsk_mtup.enabled)
1807 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1808 tp->mss_cache = mss_now;
1812 EXPORT_SYMBOL(tcp_sync_mss);
1814 /* Compute the current effective MSS, taking SACKs and IP options,
1815 * and even PMTU discovery events into account.
1817 unsigned int tcp_current_mss(struct sock *sk)
1819 const struct tcp_sock *tp = tcp_sk(sk);
1820 const struct dst_entry *dst = __sk_dst_get(sk);
1822 unsigned int header_len;
1823 struct tcp_out_options opts;
1824 struct tcp_md5sig_key *md5;
1826 mss_now = tp->mss_cache;
1829 u32 mtu = dst_mtu(dst);
1830 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1831 mss_now = tcp_sync_mss(sk, mtu);
1834 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1835 sizeof(struct tcphdr);
1836 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1837 * some common options. If this is an odd packet (because we have SACK
1838 * blocks etc) then our calculated header_len will be different, and
1839 * we have to adjust mss_now correspondingly */
1840 if (header_len != tp->tcp_header_len) {
1841 int delta = (int) header_len - tp->tcp_header_len;
1848 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1849 * As additional protections, we do not touch cwnd in retransmission phases,
1850 * and if application hit its sndbuf limit recently.
1852 static void tcp_cwnd_application_limited(struct sock *sk)
1854 struct tcp_sock *tp = tcp_sk(sk);
1856 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1857 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1858 /* Limited by application or receiver window. */
1859 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1860 u32 win_used = max(tp->snd_cwnd_used, init_win);
1861 if (win_used < tcp_snd_cwnd(tp)) {
1862 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1863 tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1);
1865 tp->snd_cwnd_used = 0;
1867 tp->snd_cwnd_stamp = tcp_jiffies32;
1870 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1872 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1873 struct tcp_sock *tp = tcp_sk(sk);
1875 /* Track the strongest available signal of the degree to which the cwnd
1876 * is fully utilized. If cwnd-limited then remember that fact for the
1877 * current window. If not cwnd-limited then track the maximum number of
1878 * outstanding packets in the current window. (If cwnd-limited then we
1879 * chose to not update tp->max_packets_out to avoid an extra else
1880 * clause with no functional impact.)
1882 if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
1884 (!tp->is_cwnd_limited &&
1885 tp->packets_out > tp->max_packets_out)) {
1886 tp->is_cwnd_limited = is_cwnd_limited;
1887 tp->max_packets_out = tp->packets_out;
1888 tp->cwnd_usage_seq = tp->snd_nxt;
1891 if (tcp_is_cwnd_limited(sk)) {
1892 /* Network is feed fully. */
1893 tp->snd_cwnd_used = 0;
1894 tp->snd_cwnd_stamp = tcp_jiffies32;
1896 /* Network starves. */
1897 if (tp->packets_out > tp->snd_cwnd_used)
1898 tp->snd_cwnd_used = tp->packets_out;
1900 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
1901 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1902 !ca_ops->cong_control)
1903 tcp_cwnd_application_limited(sk);
1905 /* The following conditions together indicate the starvation
1906 * is caused by insufficient sender buffer:
1907 * 1) just sent some data (see tcp_write_xmit)
1908 * 2) not cwnd limited (this else condition)
1909 * 3) no more data to send (tcp_write_queue_empty())
1910 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1912 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1913 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1914 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1915 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1919 /* Minshall's variant of the Nagle send check. */
1920 static bool tcp_minshall_check(const struct tcp_sock *tp)
1922 return after(tp->snd_sml, tp->snd_una) &&
1923 !after(tp->snd_sml, tp->snd_nxt);
1926 /* Update snd_sml if this skb is under mss
1927 * Note that a TSO packet might end with a sub-mss segment
1928 * The test is really :
1929 * if ((skb->len % mss) != 0)
1930 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1931 * But we can avoid doing the divide again given we already have
1932 * skb_pcount = skb->len / mss_now
1934 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1935 const struct sk_buff *skb)
1937 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1938 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1941 /* Return false, if packet can be sent now without violation Nagle's rules:
1942 * 1. It is full sized. (provided by caller in %partial bool)
1943 * 2. Or it contains FIN. (already checked by caller)
1944 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1945 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1946 * With Minshall's modification: all sent small packets are ACKed.
1948 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1952 ((nonagle & TCP_NAGLE_CORK) ||
1953 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1956 /* Return how many segs we'd like on a TSO packet,
1957 * depending on current pacing rate, and how close the peer is.
1960 * - For close peers, we rather send bigger packets to reduce
1961 * cpu costs, because occasional losses will be repaired fast.
1962 * - For long distance/rtt flows, we would like to get ACK clocking
1963 * with 1 ACK per ms.
1965 * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting
1966 * in bigger TSO bursts. We we cut the RTT-based allowance in half
1967 * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance
1968 * is below 1500 bytes after 6 * ~500 usec = 3ms.
1970 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1973 unsigned long bytes;
1976 bytes = sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift);
1978 r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log);
1979 if (r < BITS_PER_TYPE(sk->sk_gso_max_size))
1980 bytes += sk->sk_gso_max_size >> r;
1982 bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size);
1984 return max_t(u32, bytes / mss_now, min_tso_segs);
1987 /* Return the number of segments we want in the skb we are transmitting.
1988 * See if congestion control module wants to decide; otherwise, autosize.
1990 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1992 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1993 u32 min_tso, tso_segs;
1995 min_tso = ca_ops->min_tso_segs ?
1996 ca_ops->min_tso_segs(sk) :
1997 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
1999 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
2000 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
2003 /* Returns the portion of skb which can be sent right away */
2004 static unsigned int tcp_mss_split_point(const struct sock *sk,
2005 const struct sk_buff *skb,
2006 unsigned int mss_now,
2007 unsigned int max_segs,
2010 const struct tcp_sock *tp = tcp_sk(sk);
2011 u32 partial, needed, window, max_len;
2013 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2014 max_len = mss_now * max_segs;
2016 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2019 needed = min(skb->len, window);
2021 if (max_len <= needed)
2024 partial = needed % mss_now;
2025 /* If last segment is not a full MSS, check if Nagle rules allow us
2026 * to include this last segment in this skb.
2027 * Otherwise, we'll split the skb at last MSS boundary
2029 if (tcp_nagle_check(partial != 0, tp, nonagle))
2030 return needed - partial;
2035 /* Can at least one segment of SKB be sent right now, according to the
2036 * congestion window rules? If so, return how many segments are allowed.
2038 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
2039 const struct sk_buff *skb)
2041 u32 in_flight, cwnd, halfcwnd;
2043 /* Don't be strict about the congestion window for the final FIN. */
2044 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2045 tcp_skb_pcount(skb) == 1)
2048 in_flight = tcp_packets_in_flight(tp);
2049 cwnd = tcp_snd_cwnd(tp);
2050 if (in_flight >= cwnd)
2053 /* For better scheduling, ensure we have at least
2054 * 2 GSO packets in flight.
2056 halfcwnd = max(cwnd >> 1, 1U);
2057 return min(halfcwnd, cwnd - in_flight);
2060 /* Initialize TSO state of a skb.
2061 * This must be invoked the first time we consider transmitting
2062 * SKB onto the wire.
2064 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
2066 int tso_segs = tcp_skb_pcount(skb);
2068 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
2069 tcp_set_skb_tso_segs(skb, mss_now);
2070 tso_segs = tcp_skb_pcount(skb);
2076 /* Return true if the Nagle test allows this packet to be
2079 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2080 unsigned int cur_mss, int nonagle)
2082 /* Nagle rule does not apply to frames, which sit in the middle of the
2083 * write_queue (they have no chances to get new data).
2085 * This is implemented in the callers, where they modify the 'nonagle'
2086 * argument based upon the location of SKB in the send queue.
2088 if (nonagle & TCP_NAGLE_PUSH)
2091 /* Don't use the nagle rule for urgent data (or for the final FIN). */
2092 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2095 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
2101 /* Does at least the first segment of SKB fit into the send window? */
2102 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2103 const struct sk_buff *skb,
2104 unsigned int cur_mss)
2106 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2108 if (skb->len > cur_mss)
2109 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2111 return !after(end_seq, tcp_wnd_end(tp));
2114 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2115 * which is put after SKB on the list. It is very much like
2116 * tcp_fragment() except that it may make several kinds of assumptions
2117 * in order to speed up the splitting operation. In particular, we
2118 * know that all the data is in scatter-gather pages, and that the
2119 * packet has never been sent out before (and thus is not cloned).
2121 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2122 unsigned int mss_now, gfp_t gfp)
2124 int nlen = skb->len - len;
2125 struct sk_buff *buff;
2128 /* All of a TSO frame must be composed of paged data. */
2129 DEBUG_NET_WARN_ON_ONCE(skb->len != skb->data_len);
2131 buff = tcp_stream_alloc_skb(sk, gfp, true);
2132 if (unlikely(!buff))
2134 skb_copy_decrypted(buff, skb);
2135 mptcp_skb_ext_copy(buff, skb);
2137 sk_wmem_queued_add(sk, buff->truesize);
2138 sk_mem_charge(sk, buff->truesize);
2139 buff->truesize += nlen;
2140 skb->truesize -= nlen;
2142 /* Correct the sequence numbers. */
2143 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2144 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2145 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2147 /* PSH and FIN should only be set in the second packet. */
2148 flags = TCP_SKB_CB(skb)->tcp_flags;
2149 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2150 TCP_SKB_CB(buff)->tcp_flags = flags;
2152 tcp_skb_fragment_eor(skb, buff);
2154 skb_split(skb, buff, len);
2155 tcp_fragment_tstamp(skb, buff);
2157 /* Fix up tso_factor for both original and new SKB. */
2158 tcp_set_skb_tso_segs(skb, mss_now);
2159 tcp_set_skb_tso_segs(buff, mss_now);
2161 /* Link BUFF into the send queue. */
2162 __skb_header_release(buff);
2163 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2168 /* Try to defer sending, if possible, in order to minimize the amount
2169 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2171 * This algorithm is from John Heffner.
2173 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2174 bool *is_cwnd_limited,
2175 bool *is_rwnd_limited,
2178 const struct inet_connection_sock *icsk = inet_csk(sk);
2179 u32 send_win, cong_win, limit, in_flight;
2180 struct tcp_sock *tp = tcp_sk(sk);
2181 struct sk_buff *head;
2185 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2188 /* Avoid bursty behavior by allowing defer
2189 * only if the last write was recent (1 ms).
2190 * Note that tp->tcp_wstamp_ns can be in the future if we have
2191 * packets waiting in a qdisc or device for EDT delivery.
2193 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2197 in_flight = tcp_packets_in_flight(tp);
2199 BUG_ON(tcp_skb_pcount(skb) <= 1);
2200 BUG_ON(tcp_snd_cwnd(tp) <= in_flight);
2202 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2204 /* From in_flight test above, we know that cwnd > in_flight. */
2205 cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache;
2207 limit = min(send_win, cong_win);
2209 /* If a full-sized TSO skb can be sent, do it. */
2210 if (limit >= max_segs * tp->mss_cache)
2213 /* Middle in queue won't get any more data, full sendable already? */
2214 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2217 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2219 u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache);
2221 /* If at least some fraction of a window is available,
2224 chunk /= win_divisor;
2228 /* Different approach, try not to defer past a single
2229 * ACK. Receiver should ACK every other full sized
2230 * frame, so if we have space for more than 3 frames
2233 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2237 /* TODO : use tsorted_sent_queue ? */
2238 head = tcp_rtx_queue_head(sk);
2241 delta = tp->tcp_clock_cache - head->tstamp;
2242 /* If next ACK is likely to come too late (half srtt), do not defer */
2243 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2246 /* Ok, it looks like it is advisable to defer.
2247 * Three cases are tracked :
2248 * 1) We are cwnd-limited
2249 * 2) We are rwnd-limited
2250 * 3) We are application limited.
2252 if (cong_win < send_win) {
2253 if (cong_win <= skb->len) {
2254 *is_cwnd_limited = true;
2258 if (send_win <= skb->len) {
2259 *is_rwnd_limited = true;
2264 /* If this packet won't get more data, do not wait. */
2265 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2266 TCP_SKB_CB(skb)->eor)
2275 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2277 struct inet_connection_sock *icsk = inet_csk(sk);
2278 struct tcp_sock *tp = tcp_sk(sk);
2279 struct net *net = sock_net(sk);
2283 interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2284 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2285 if (unlikely(delta >= interval * HZ)) {
2286 int mss = tcp_current_mss(sk);
2288 /* Update current search range */
2289 icsk->icsk_mtup.probe_size = 0;
2290 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2291 sizeof(struct tcphdr) +
2292 icsk->icsk_af_ops->net_header_len;
2293 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2295 /* Update probe time stamp */
2296 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2300 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2302 struct sk_buff *skb, *next;
2304 skb = tcp_send_head(sk);
2305 tcp_for_write_queue_from_safe(skb, next, sk) {
2306 if (len <= skb->len)
2309 if (unlikely(TCP_SKB_CB(skb)->eor) ||
2310 tcp_has_tx_tstamp(skb) ||
2311 !skb_pure_zcopy_same(skb, next))
2320 static int tcp_clone_payload(struct sock *sk, struct sk_buff *to,
2323 skb_frag_t *lastfrag = NULL, *fragto = skb_shinfo(to)->frags;
2324 int i, todo, len = 0, nr_frags = 0;
2325 const struct sk_buff *skb;
2327 if (!sk_wmem_schedule(sk, to->truesize + probe_size))
2330 skb_queue_walk(&sk->sk_write_queue, skb) {
2331 const skb_frag_t *fragfrom = skb_shinfo(skb)->frags;
2333 if (skb_headlen(skb))
2336 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, fragfrom++) {
2337 if (len >= probe_size)
2339 todo = min_t(int, skb_frag_size(fragfrom),
2343 skb_frag_page(fragfrom) == skb_frag_page(lastfrag) &&
2344 skb_frag_off(fragfrom) == skb_frag_off(lastfrag) +
2345 skb_frag_size(lastfrag)) {
2346 skb_frag_size_add(lastfrag, todo);
2349 if (unlikely(nr_frags == MAX_SKB_FRAGS))
2351 skb_frag_page_copy(fragto, fragfrom);
2352 skb_frag_off_copy(fragto, fragfrom);
2353 skb_frag_size_set(fragto, todo);
2355 lastfrag = fragto++;
2359 WARN_ON_ONCE(len != probe_size);
2360 for (i = 0; i < nr_frags; i++)
2361 skb_frag_ref(to, i);
2363 skb_shinfo(to)->nr_frags = nr_frags;
2364 to->truesize += probe_size;
2365 to->len += probe_size;
2366 to->data_len += probe_size;
2367 __skb_header_release(to);
2371 /* Create a new MTU probe if we are ready.
2372 * MTU probe is regularly attempting to increase the path MTU by
2373 * deliberately sending larger packets. This discovers routing
2374 * changes resulting in larger path MTUs.
2376 * Returns 0 if we should wait to probe (no cwnd available),
2377 * 1 if a probe was sent,
2380 static int tcp_mtu_probe(struct sock *sk)
2382 struct inet_connection_sock *icsk = inet_csk(sk);
2383 struct tcp_sock *tp = tcp_sk(sk);
2384 struct sk_buff *skb, *nskb, *next;
2385 struct net *net = sock_net(sk);
2392 /* Not currently probing/verifying,
2394 * have enough cwnd, and
2395 * not SACKing (the variable headers throw things off)
2397 if (likely(!icsk->icsk_mtup.enabled ||
2398 icsk->icsk_mtup.probe_size ||
2399 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2400 tcp_snd_cwnd(tp) < 11 ||
2401 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2404 /* Use binary search for probe_size between tcp_mss_base,
2405 * and current mss_clamp. if (search_high - search_low)
2406 * smaller than a threshold, backoff from probing.
2408 mss_now = tcp_current_mss(sk);
2409 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2410 icsk->icsk_mtup.search_low) >> 1);
2411 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2412 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2413 /* When misfortune happens, we are reprobing actively,
2414 * and then reprobe timer has expired. We stick with current
2415 * probing process by not resetting search range to its orignal.
2417 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2418 interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2419 /* Check whether enough time has elaplased for
2420 * another round of probing.
2422 tcp_mtu_check_reprobe(sk);
2426 /* Have enough data in the send queue to probe? */
2427 if (tp->write_seq - tp->snd_nxt < size_needed)
2430 if (tp->snd_wnd < size_needed)
2432 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2435 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2436 if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) {
2437 if (!tcp_packets_in_flight(tp))
2443 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2446 /* We're allowed to probe. Build it now. */
2447 nskb = tcp_stream_alloc_skb(sk, GFP_ATOMIC, false);
2451 /* build the payload, and be prepared to abort if this fails. */
2452 if (tcp_clone_payload(sk, nskb, probe_size)) {
2456 sk_wmem_queued_add(sk, nskb->truesize);
2457 sk_mem_charge(sk, nskb->truesize);
2459 skb = tcp_send_head(sk);
2460 skb_copy_decrypted(nskb, skb);
2461 mptcp_skb_ext_copy(nskb, skb);
2463 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2464 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2465 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2467 tcp_insert_write_queue_before(nskb, skb, sk);
2468 tcp_highest_sack_replace(sk, skb, nskb);
2471 tcp_for_write_queue_from_safe(skb, next, sk) {
2472 copy = min_t(int, skb->len, probe_size - len);
2474 if (skb->len <= copy) {
2475 /* We've eaten all the data from this skb.
2477 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2478 /* If this is the last SKB we copy and eor is set
2479 * we need to propagate it to the new skb.
2481 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2482 tcp_skb_collapse_tstamp(nskb, skb);
2483 tcp_unlink_write_queue(skb, sk);
2484 tcp_wmem_free_skb(sk, skb);
2486 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2487 ~(TCPHDR_FIN|TCPHDR_PSH);
2488 __pskb_trim_head(skb, copy);
2489 tcp_set_skb_tso_segs(skb, mss_now);
2490 TCP_SKB_CB(skb)->seq += copy;
2495 if (len >= probe_size)
2498 tcp_init_tso_segs(nskb, nskb->len);
2500 /* We're ready to send. If this fails, the probe will
2501 * be resegmented into mss-sized pieces by tcp_write_xmit().
2503 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2504 /* Decrement cwnd here because we are sending
2505 * effectively two packets. */
2506 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1);
2507 tcp_event_new_data_sent(sk, nskb);
2509 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2510 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2511 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2519 static bool tcp_pacing_check(struct sock *sk)
2521 struct tcp_sock *tp = tcp_sk(sk);
2523 if (!tcp_needs_internal_pacing(sk))
2526 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2529 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2530 hrtimer_start(&tp->pacing_timer,
2531 ns_to_ktime(tp->tcp_wstamp_ns),
2532 HRTIMER_MODE_ABS_PINNED_SOFT);
2538 /* TCP Small Queues :
2539 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2540 * (These limits are doubled for retransmits)
2542 * - better RTT estimation and ACK scheduling
2545 * Alas, some drivers / subsystems require a fair amount
2546 * of queued bytes to ensure line rate.
2547 * One example is wifi aggregation (802.11 AMPDU)
2549 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2550 unsigned int factor)
2552 unsigned long limit;
2554 limit = max_t(unsigned long,
2556 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2557 if (sk->sk_pacing_status == SK_PACING_NONE)
2558 limit = min_t(unsigned long, limit,
2559 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
2562 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2563 tcp_sk(sk)->tcp_tx_delay) {
2564 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2566 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2567 * approximate our needs assuming an ~100% skb->truesize overhead.
2568 * USEC_PER_SEC is approximated by 2^20.
2569 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2571 extra_bytes >>= (20 - 1);
2572 limit += extra_bytes;
2574 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2575 /* Always send skb if rtx queue is empty.
2576 * No need to wait for TX completion to call us back,
2577 * after softirq/tasklet schedule.
2578 * This helps when TX completions are delayed too much.
2580 if (tcp_rtx_queue_empty(sk))
2583 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2584 /* It is possible TX completion already happened
2585 * before we set TSQ_THROTTLED, so we must
2586 * test again the condition.
2588 smp_mb__after_atomic();
2589 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2595 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2597 const u32 now = tcp_jiffies32;
2598 enum tcp_chrono old = tp->chrono_type;
2600 if (old > TCP_CHRONO_UNSPEC)
2601 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2602 tp->chrono_start = now;
2603 tp->chrono_type = new;
2606 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2608 struct tcp_sock *tp = tcp_sk(sk);
2610 /* If there are multiple conditions worthy of tracking in a
2611 * chronograph then the highest priority enum takes precedence
2612 * over the other conditions. So that if something "more interesting"
2613 * starts happening, stop the previous chrono and start a new one.
2615 if (type > tp->chrono_type)
2616 tcp_chrono_set(tp, type);
2619 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2621 struct tcp_sock *tp = tcp_sk(sk);
2624 /* There are multiple conditions worthy of tracking in a
2625 * chronograph, so that the highest priority enum takes
2626 * precedence over the other conditions (see tcp_chrono_start).
2627 * If a condition stops, we only stop chrono tracking if
2628 * it's the "most interesting" or current chrono we are
2629 * tracking and starts busy chrono if we have pending data.
2631 if (tcp_rtx_and_write_queues_empty(sk))
2632 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2633 else if (type == tp->chrono_type)
2634 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2637 /* This routine writes packets to the network. It advances the
2638 * send_head. This happens as incoming acks open up the remote
2641 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2642 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2643 * account rare use of URG, this is not a big flaw.
2645 * Send at most one packet when push_one > 0. Temporarily ignore
2646 * cwnd limit to force at most one packet out when push_one == 2.
2648 * Returns true, if no segments are in flight and we have queued segments,
2649 * but cannot send anything now because of SWS or another problem.
2651 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2652 int push_one, gfp_t gfp)
2654 struct tcp_sock *tp = tcp_sk(sk);
2655 struct sk_buff *skb;
2656 unsigned int tso_segs, sent_pkts;
2659 bool is_cwnd_limited = false, is_rwnd_limited = false;
2664 tcp_mstamp_refresh(tp);
2666 /* Do MTU probing. */
2667 result = tcp_mtu_probe(sk);
2670 } else if (result > 0) {
2675 max_segs = tcp_tso_segs(sk, mss_now);
2676 while ((skb = tcp_send_head(sk))) {
2679 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2680 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2681 tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2682 skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
2683 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2684 tcp_init_tso_segs(skb, mss_now);
2685 goto repair; /* Skip network transmission */
2688 if (tcp_pacing_check(sk))
2691 tso_segs = tcp_init_tso_segs(skb, mss_now);
2694 cwnd_quota = tcp_cwnd_test(tp, skb);
2697 /* Force out a loss probe pkt. */
2703 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2704 is_rwnd_limited = true;
2708 if (tso_segs == 1) {
2709 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2710 (tcp_skb_is_last(sk, skb) ?
2711 nonagle : TCP_NAGLE_PUSH))))
2715 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2716 &is_rwnd_limited, max_segs))
2721 if (tso_segs > 1 && !tcp_urg_mode(tp))
2722 limit = tcp_mss_split_point(sk, skb, mss_now,
2728 if (skb->len > limit &&
2729 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2732 if (tcp_small_queue_check(sk, skb, 0))
2735 /* Argh, we hit an empty skb(), presumably a thread
2736 * is sleeping in sendmsg()/sk_stream_wait_memory().
2737 * We do not want to send a pure-ack packet and have
2738 * a strange looking rtx queue with empty packet(s).
2740 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2743 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2747 /* Advance the send_head. This one is sent out.
2748 * This call will increment packets_out.
2750 tcp_event_new_data_sent(sk, skb);
2752 tcp_minshall_update(tp, mss_now, skb);
2753 sent_pkts += tcp_skb_pcount(skb);
2759 if (is_rwnd_limited)
2760 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2762 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2764 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp));
2765 if (likely(sent_pkts || is_cwnd_limited))
2766 tcp_cwnd_validate(sk, is_cwnd_limited);
2768 if (likely(sent_pkts)) {
2769 if (tcp_in_cwnd_reduction(sk))
2770 tp->prr_out += sent_pkts;
2772 /* Send one loss probe per tail loss episode. */
2774 tcp_schedule_loss_probe(sk, false);
2777 return !tp->packets_out && !tcp_write_queue_empty(sk);
2780 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2782 struct inet_connection_sock *icsk = inet_csk(sk);
2783 struct tcp_sock *tp = tcp_sk(sk);
2784 u32 timeout, rto_delta_us;
2787 /* Don't do any loss probe on a Fast Open connection before 3WHS
2790 if (rcu_access_pointer(tp->fastopen_rsk))
2793 early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
2794 /* Schedule a loss probe in 2*RTT for SACK capable connections
2795 * not in loss recovery, that are either limited by cwnd or application.
2797 if ((early_retrans != 3 && early_retrans != 4) ||
2798 !tp->packets_out || !tcp_is_sack(tp) ||
2799 (icsk->icsk_ca_state != TCP_CA_Open &&
2800 icsk->icsk_ca_state != TCP_CA_CWR))
2803 /* Probe timeout is 2*rtt. Add minimum RTO to account
2804 * for delayed ack when there's one outstanding packet. If no RTT
2805 * sample is available then probe after TCP_TIMEOUT_INIT.
2808 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2809 if (tp->packets_out == 1)
2810 timeout += TCP_RTO_MIN;
2812 timeout += TCP_TIMEOUT_MIN;
2814 timeout = TCP_TIMEOUT_INIT;
2817 /* If the RTO formula yields an earlier time, then use that time. */
2818 rto_delta_us = advancing_rto ?
2819 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2820 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2821 if (rto_delta_us > 0)
2822 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2824 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2828 /* Thanks to skb fast clones, we can detect if a prior transmit of
2829 * a packet is still in a qdisc or driver queue.
2830 * In this case, there is very little point doing a retransmit !
2832 static bool skb_still_in_host_queue(struct sock *sk,
2833 const struct sk_buff *skb)
2835 if (unlikely(skb_fclone_busy(sk, skb))) {
2836 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2837 smp_mb__after_atomic();
2838 if (skb_fclone_busy(sk, skb)) {
2839 NET_INC_STATS(sock_net(sk),
2840 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2847 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2848 * retransmit the last segment.
2850 void tcp_send_loss_probe(struct sock *sk)
2852 struct tcp_sock *tp = tcp_sk(sk);
2853 struct sk_buff *skb;
2855 int mss = tcp_current_mss(sk);
2857 /* At most one outstanding TLP */
2858 if (tp->tlp_high_seq)
2861 tp->tlp_retrans = 0;
2862 skb = tcp_send_head(sk);
2863 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2864 pcount = tp->packets_out;
2865 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2866 if (tp->packets_out > pcount)
2870 skb = skb_rb_last(&sk->tcp_rtx_queue);
2871 if (unlikely(!skb)) {
2872 WARN_ONCE(tp->packets_out,
2873 "invalid inflight: %u state %u cwnd %u mss %d\n",
2874 tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss);
2875 inet_csk(sk)->icsk_pending = 0;
2879 if (skb_still_in_host_queue(sk, skb))
2882 pcount = tcp_skb_pcount(skb);
2883 if (WARN_ON(!pcount))
2886 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2887 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2888 (pcount - 1) * mss, mss,
2891 skb = skb_rb_next(skb);
2894 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2897 if (__tcp_retransmit_skb(sk, skb, 1))
2900 tp->tlp_retrans = 1;
2903 /* Record snd_nxt for loss detection. */
2904 tp->tlp_high_seq = tp->snd_nxt;
2906 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2907 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2908 inet_csk(sk)->icsk_pending = 0;
2913 /* Push out any pending frames which were held back due to
2914 * TCP_CORK or attempt at coalescing tiny packets.
2915 * The socket must be locked by the caller.
2917 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2920 /* If we are closed, the bytes will have to remain here.
2921 * In time closedown will finish, we empty the write queue and
2922 * all will be happy.
2924 if (unlikely(sk->sk_state == TCP_CLOSE))
2927 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2928 sk_gfp_mask(sk, GFP_ATOMIC)))
2929 tcp_check_probe_timer(sk);
2932 /* Send _single_ skb sitting at the send head. This function requires
2933 * true push pending frames to setup probe timer etc.
2935 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2937 struct sk_buff *skb = tcp_send_head(sk);
2939 BUG_ON(!skb || skb->len < mss_now);
2941 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2944 /* This function returns the amount that we can raise the
2945 * usable window based on the following constraints
2947 * 1. The window can never be shrunk once it is offered (RFC 793)
2948 * 2. We limit memory per socket
2951 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2952 * RECV.NEXT + RCV.WIN fixed until:
2953 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2955 * i.e. don't raise the right edge of the window until you can raise
2956 * it at least MSS bytes.
2958 * Unfortunately, the recommended algorithm breaks header prediction,
2959 * since header prediction assumes th->window stays fixed.
2961 * Strictly speaking, keeping th->window fixed violates the receiver
2962 * side SWS prevention criteria. The problem is that under this rule
2963 * a stream of single byte packets will cause the right side of the
2964 * window to always advance by a single byte.
2966 * Of course, if the sender implements sender side SWS prevention
2967 * then this will not be a problem.
2969 * BSD seems to make the following compromise:
2971 * If the free space is less than the 1/4 of the maximum
2972 * space available and the free space is less than 1/2 mss,
2973 * then set the window to 0.
2974 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2975 * Otherwise, just prevent the window from shrinking
2976 * and from being larger than the largest representable value.
2978 * This prevents incremental opening of the window in the regime
2979 * where TCP is limited by the speed of the reader side taking
2980 * data out of the TCP receive queue. It does nothing about
2981 * those cases where the window is constrained on the sender side
2982 * because the pipeline is full.
2984 * BSD also seems to "accidentally" limit itself to windows that are a
2985 * multiple of MSS, at least until the free space gets quite small.
2986 * This would appear to be a side effect of the mbuf implementation.
2987 * Combining these two algorithms results in the observed behavior
2988 * of having a fixed window size at almost all times.
2990 * Below we obtain similar behavior by forcing the offered window to
2991 * a multiple of the mss when it is feasible to do so.
2993 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2994 * Regular options like TIMESTAMP are taken into account.
2996 u32 __tcp_select_window(struct sock *sk)
2998 struct inet_connection_sock *icsk = inet_csk(sk);
2999 struct tcp_sock *tp = tcp_sk(sk);
3000 struct net *net = sock_net(sk);
3001 /* MSS for the peer's data. Previous versions used mss_clamp
3002 * here. I don't know if the value based on our guesses
3003 * of peer's MSS is better for the performance. It's more correct
3004 * but may be worse for the performance because of rcv_mss
3005 * fluctuations. --SAW 1998/11/1
3007 int mss = icsk->icsk_ack.rcv_mss;
3008 int free_space = tcp_space(sk);
3009 int allowed_space = tcp_full_space(sk);
3010 int full_space, window;
3012 if (sk_is_mptcp(sk))
3013 mptcp_space(sk, &free_space, &allowed_space);
3015 full_space = min_t(int, tp->window_clamp, allowed_space);
3017 if (unlikely(mss > full_space)) {
3023 /* Only allow window shrink if the sysctl is enabled and we have
3024 * a non-zero scaling factor in effect.
3026 if (READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) && tp->rx_opt.rcv_wscale)
3027 goto shrink_window_allowed;
3029 /* do not allow window to shrink */
3031 if (free_space < (full_space >> 1)) {
3032 icsk->icsk_ack.quick = 0;
3034 if (tcp_under_memory_pressure(sk))
3035 tcp_adjust_rcv_ssthresh(sk);
3037 /* free_space might become our new window, make sure we don't
3038 * increase it due to wscale.
3040 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3042 /* if free space is less than mss estimate, or is below 1/16th
3043 * of the maximum allowed, try to move to zero-window, else
3044 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
3045 * new incoming data is dropped due to memory limits.
3046 * With large window, mss test triggers way too late in order
3047 * to announce zero window in time before rmem limit kicks in.
3049 if (free_space < (allowed_space >> 4) || free_space < mss)
3053 if (free_space > tp->rcv_ssthresh)
3054 free_space = tp->rcv_ssthresh;
3056 /* Don't do rounding if we are using window scaling, since the
3057 * scaled window will not line up with the MSS boundary anyway.
3059 if (tp->rx_opt.rcv_wscale) {
3060 window = free_space;
3062 /* Advertise enough space so that it won't get scaled away.
3063 * Import case: prevent zero window announcement if
3064 * 1<<rcv_wscale > mss.
3066 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
3068 window = tp->rcv_wnd;
3069 /* Get the largest window that is a nice multiple of mss.
3070 * Window clamp already applied above.
3071 * If our current window offering is within 1 mss of the
3072 * free space we just keep it. This prevents the divide
3073 * and multiply from happening most of the time.
3074 * We also don't do any window rounding when the free space
3077 if (window <= free_space - mss || window > free_space)
3078 window = rounddown(free_space, mss);
3079 else if (mss == full_space &&
3080 free_space > window + (full_space >> 1))
3081 window = free_space;
3086 shrink_window_allowed:
3087 /* new window should always be an exact multiple of scaling factor */
3088 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3090 if (free_space < (full_space >> 1)) {
3091 icsk->icsk_ack.quick = 0;
3093 if (tcp_under_memory_pressure(sk))
3094 tcp_adjust_rcv_ssthresh(sk);
3096 /* if free space is too low, return a zero window */
3097 if (free_space < (allowed_space >> 4) || free_space < mss ||
3098 free_space < (1 << tp->rx_opt.rcv_wscale))
3102 if (free_space > tp->rcv_ssthresh) {
3103 free_space = tp->rcv_ssthresh;
3104 /* new window should always be an exact multiple of scaling factor
3106 * For this case, we ALIGN "up" (increase free_space) because
3107 * we know free_space is not zero here, it has been reduced from
3108 * the memory-based limit, and rcv_ssthresh is not a hard limit
3109 * (unlike sk_rcvbuf).
3111 free_space = ALIGN(free_space, (1 << tp->rx_opt.rcv_wscale));
3117 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3118 const struct sk_buff *next_skb)
3120 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3121 const struct skb_shared_info *next_shinfo =
3122 skb_shinfo(next_skb);
3123 struct skb_shared_info *shinfo = skb_shinfo(skb);
3125 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3126 shinfo->tskey = next_shinfo->tskey;
3127 TCP_SKB_CB(skb)->txstamp_ack |=
3128 TCP_SKB_CB(next_skb)->txstamp_ack;
3132 /* Collapses two adjacent SKB's during retransmission. */
3133 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3135 struct tcp_sock *tp = tcp_sk(sk);
3136 struct sk_buff *next_skb = skb_rb_next(skb);
3139 next_skb_size = next_skb->len;
3141 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3143 if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3146 tcp_highest_sack_replace(sk, next_skb, skb);
3148 /* Update sequence range on original skb. */
3149 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3151 /* Merge over control information. This moves PSH/FIN etc. over */
3152 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3154 /* All done, get rid of second SKB and account for it so
3155 * packet counting does not break.
3157 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3158 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3160 /* changed transmit queue under us so clear hints */
3161 tcp_clear_retrans_hints_partial(tp);
3162 if (next_skb == tp->retransmit_skb_hint)
3163 tp->retransmit_skb_hint = skb;
3165 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3167 tcp_skb_collapse_tstamp(skb, next_skb);
3169 tcp_rtx_queue_unlink_and_free(next_skb, sk);
3173 /* Check if coalescing SKBs is legal. */
3174 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3176 if (tcp_skb_pcount(skb) > 1)
3178 if (skb_cloned(skb))
3180 /* Some heuristics for collapsing over SACK'd could be invented */
3181 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3187 /* Collapse packets in the retransmit queue to make to create
3188 * less packets on the wire. This is only done on retransmission.
3190 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3193 struct tcp_sock *tp = tcp_sk(sk);
3194 struct sk_buff *skb = to, *tmp;
3197 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
3199 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3202 skb_rbtree_walk_from_safe(skb, tmp) {
3203 if (!tcp_can_collapse(sk, skb))
3206 if (!tcp_skb_can_collapse(to, skb))
3219 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3222 if (!tcp_collapse_retrans(sk, to))
3227 /* This retransmits one SKB. Policy decisions and retransmit queue
3228 * state updates are done by the caller. Returns non-zero if an
3229 * error occurred which prevented the send.
3231 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3233 struct inet_connection_sock *icsk = inet_csk(sk);
3234 struct tcp_sock *tp = tcp_sk(sk);
3235 unsigned int cur_mss;
3239 /* Inconclusive MTU probe */
3240 if (icsk->icsk_mtup.probe_size)
3241 icsk->icsk_mtup.probe_size = 0;
3243 if (skb_still_in_host_queue(sk, skb))
3246 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
3247 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3251 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3255 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3256 return -EHOSTUNREACH; /* Routing failure or similar. */
3258 cur_mss = tcp_current_mss(sk);
3259 avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3261 /* If receiver has shrunk his window, and skb is out of
3262 * new window, do not retransmit it. The exception is the
3263 * case, when window is shrunk to zero. In this case
3264 * our retransmit of one segment serves as a zero window probe.
3266 if (avail_wnd <= 0) {
3267 if (TCP_SKB_CB(skb)->seq != tp->snd_una)
3269 avail_wnd = cur_mss;
3272 len = cur_mss * segs;
3273 if (len > avail_wnd) {
3274 len = rounddown(avail_wnd, cur_mss);
3278 if (skb->len > len) {
3279 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3280 cur_mss, GFP_ATOMIC))
3281 return -ENOMEM; /* We'll try again later. */
3283 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
3286 diff = tcp_skb_pcount(skb);
3287 tcp_set_skb_tso_segs(skb, cur_mss);
3288 diff -= tcp_skb_pcount(skb);
3290 tcp_adjust_pcount(sk, skb, diff);
3291 avail_wnd = min_t(int, avail_wnd, cur_mss);
3292 if (skb->len < avail_wnd)
3293 tcp_retrans_try_collapse(sk, skb, avail_wnd);
3296 /* RFC3168, section 6.1.1.1. ECN fallback */
3297 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3298 tcp_ecn_clear_syn(sk, skb);
3300 /* Update global and local TCP statistics. */
3301 segs = tcp_skb_pcount(skb);
3302 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3303 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3304 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3305 tp->total_retrans += segs;
3306 tp->bytes_retrans += skb->len;
3308 /* make sure skb->data is aligned on arches that require it
3309 * and check if ack-trimming & collapsing extended the headroom
3310 * beyond what csum_start can cover.
3312 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3313 skb_headroom(skb) >= 0xFFFF)) {
3314 struct sk_buff *nskb;
3316 tcp_skb_tsorted_save(skb) {
3317 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3320 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3324 } tcp_skb_tsorted_restore(skb);
3327 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3328 tcp_rate_skb_sent(sk, skb);
3331 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3334 /* To avoid taking spuriously low RTT samples based on a timestamp
3335 * for a transmit that never happened, always mark EVER_RETRANS
3337 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3339 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3340 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3341 TCP_SKB_CB(skb)->seq, segs, err);
3344 trace_tcp_retransmit_skb(sk, skb);
3345 } else if (err != -EBUSY) {
3346 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3351 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3353 struct tcp_sock *tp = tcp_sk(sk);
3354 int err = __tcp_retransmit_skb(sk, skb, segs);
3357 #if FASTRETRANS_DEBUG > 0
3358 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3359 net_dbg_ratelimited("retrans_out leaked\n");
3362 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3363 tp->retrans_out += tcp_skb_pcount(skb);
3366 /* Save stamp of the first (attempted) retransmit. */
3367 if (!tp->retrans_stamp)
3368 tp->retrans_stamp = tcp_skb_timestamp(skb);
3370 if (tp->undo_retrans < 0)
3371 tp->undo_retrans = 0;
3372 tp->undo_retrans += tcp_skb_pcount(skb);
3376 /* This gets called after a retransmit timeout, and the initially
3377 * retransmitted data is acknowledged. It tries to continue
3378 * resending the rest of the retransmit queue, until either
3379 * we've sent it all or the congestion window limit is reached.
3381 void tcp_xmit_retransmit_queue(struct sock *sk)
3383 const struct inet_connection_sock *icsk = inet_csk(sk);
3384 struct sk_buff *skb, *rtx_head, *hole = NULL;
3385 struct tcp_sock *tp = tcp_sk(sk);
3386 bool rearm_timer = false;
3390 if (!tp->packets_out)
3393 rtx_head = tcp_rtx_queue_head(sk);
3394 skb = tp->retransmit_skb_hint ?: rtx_head;
3395 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3396 skb_rbtree_walk_from(skb) {
3400 if (tcp_pacing_check(sk))
3403 /* we could do better than to assign each time */
3405 tp->retransmit_skb_hint = skb;
3407 segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp);
3410 sacked = TCP_SKB_CB(skb)->sacked;
3411 /* In case tcp_shift_skb_data() have aggregated large skbs,
3412 * we need to make sure not sending too bigs TSO packets
3414 segs = min_t(int, segs, max_segs);
3416 if (tp->retrans_out >= tp->lost_out) {
3418 } else if (!(sacked & TCPCB_LOST)) {
3419 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3424 if (icsk->icsk_ca_state != TCP_CA_Loss)
3425 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3427 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3430 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3433 if (tcp_small_queue_check(sk, skb, 1))
3436 if (tcp_retransmit_skb(sk, skb, segs))
3439 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3441 if (tcp_in_cwnd_reduction(sk))
3442 tp->prr_out += tcp_skb_pcount(skb);
3444 if (skb == rtx_head &&
3445 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3450 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3451 inet_csk(sk)->icsk_rto,
3455 /* We allow to exceed memory limits for FIN packets to expedite
3456 * connection tear down and (memory) recovery.
3457 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3458 * or even be forced to close flow without any FIN.
3459 * In general, we want to allow one skb per socket to avoid hangs
3460 * with edge trigger epoll()
3462 void sk_forced_mem_schedule(struct sock *sk, int size)
3466 delta = size - sk->sk_forward_alloc;
3469 amt = sk_mem_pages(delta);
3470 sk->sk_forward_alloc += amt << PAGE_SHIFT;
3471 sk_memory_allocated_add(sk, amt);
3473 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3474 mem_cgroup_charge_skmem(sk->sk_memcg, amt,
3475 gfp_memcg_charge() | __GFP_NOFAIL);
3478 /* Send a FIN. The caller locks the socket for us.
3479 * We should try to send a FIN packet really hard, but eventually give up.
3481 void tcp_send_fin(struct sock *sk)
3483 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3484 struct tcp_sock *tp = tcp_sk(sk);
3486 /* Optimization, tack on the FIN if we have one skb in write queue and
3487 * this skb was not yet sent, or we are under memory pressure.
3488 * Note: in the latter case, FIN packet will be sent after a timeout,
3489 * as TCP stack thinks it has already been transmitted.
3492 if (!tskb && tcp_under_memory_pressure(sk))
3493 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3496 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3497 TCP_SKB_CB(tskb)->end_seq++;
3500 /* This means tskb was already sent.
3501 * Pretend we included the FIN on previous transmit.
3502 * We need to set tp->snd_nxt to the value it would have
3503 * if FIN had been sent. This is because retransmit path
3504 * does not change tp->snd_nxt.
3506 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3510 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3514 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3515 skb_reserve(skb, MAX_TCP_HEADER);
3516 sk_forced_mem_schedule(sk, skb->truesize);
3517 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3518 tcp_init_nondata_skb(skb, tp->write_seq,
3519 TCPHDR_ACK | TCPHDR_FIN);
3520 tcp_queue_skb(sk, skb);
3522 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3525 /* We get here when a process closes a file descriptor (either due to
3526 * an explicit close() or as a byproduct of exit()'ing) and there
3527 * was unread data in the receive queue. This behavior is recommended
3528 * by RFC 2525, section 2.17. -DaveM
3530 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3532 struct sk_buff *skb;
3534 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3536 /* NOTE: No TCP options attached and we never retransmit this. */
3537 skb = alloc_skb(MAX_TCP_HEADER, priority);
3539 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3543 /* Reserve space for headers and prepare control bits. */
3544 skb_reserve(skb, MAX_TCP_HEADER);
3545 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3546 TCPHDR_ACK | TCPHDR_RST);
3547 tcp_mstamp_refresh(tcp_sk(sk));
3549 if (tcp_transmit_skb(sk, skb, 0, priority))
3550 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3552 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3553 * skb here is different to the troublesome skb, so use NULL
3555 trace_tcp_send_reset(sk, NULL);
3558 /* Send a crossed SYN-ACK during socket establishment.
3559 * WARNING: This routine must only be called when we have already sent
3560 * a SYN packet that crossed the incoming SYN that caused this routine
3561 * to get called. If this assumption fails then the initial rcv_wnd
3562 * and rcv_wscale values will not be correct.
3564 int tcp_send_synack(struct sock *sk)
3566 struct sk_buff *skb;
3568 skb = tcp_rtx_queue_head(sk);
3569 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3570 pr_err("%s: wrong queue state\n", __func__);
3573 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3574 if (skb_cloned(skb)) {
3575 struct sk_buff *nskb;
3577 tcp_skb_tsorted_save(skb) {
3578 nskb = skb_copy(skb, GFP_ATOMIC);
3579 } tcp_skb_tsorted_restore(skb);
3582 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3583 tcp_highest_sack_replace(sk, skb, nskb);
3584 tcp_rtx_queue_unlink_and_free(skb, sk);
3585 __skb_header_release(nskb);
3586 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3587 sk_wmem_queued_add(sk, nskb->truesize);
3588 sk_mem_charge(sk, nskb->truesize);
3592 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3593 tcp_ecn_send_synack(sk, skb);
3595 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3599 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3600 * @sk: listener socket
3601 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3602 * should not use it again.
3603 * @req: request_sock pointer
3604 * @foc: cookie for tcp fast open
3605 * @synack_type: Type of synack to prepare
3606 * @syn_skb: SYN packet just received. It could be NULL for rtx case.
3608 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3609 struct request_sock *req,
3610 struct tcp_fastopen_cookie *foc,
3611 enum tcp_synack_type synack_type,
3612 struct sk_buff *syn_skb)
3614 struct inet_request_sock *ireq = inet_rsk(req);
3615 const struct tcp_sock *tp = tcp_sk(sk);
3616 struct tcp_md5sig_key *md5 = NULL;
3617 struct tcp_out_options opts;
3618 struct sk_buff *skb;
3619 int tcp_header_size;
3624 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3625 if (unlikely(!skb)) {
3629 /* Reserve space for headers. */
3630 skb_reserve(skb, MAX_TCP_HEADER);
3632 switch (synack_type) {
3633 case TCP_SYNACK_NORMAL:
3634 skb_set_owner_w(skb, req_to_sk(req));
3636 case TCP_SYNACK_COOKIE:
3637 /* Under synflood, we do not attach skb to a socket,
3638 * to avoid false sharing.
3641 case TCP_SYNACK_FASTOPEN:
3642 /* sk is a const pointer, because we want to express multiple
3643 * cpu might call us concurrently.
3644 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3646 skb_set_owner_w(skb, (struct sock *)sk);
3649 skb_dst_set(skb, dst);
3651 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3653 memset(&opts, 0, sizeof(opts));
3654 now = tcp_clock_ns();
3655 #ifdef CONFIG_SYN_COOKIES
3656 if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3657 skb_set_delivery_time(skb, cookie_init_timestamp(req, now),
3662 skb_set_delivery_time(skb, now, true);
3663 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3664 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3667 #ifdef CONFIG_TCP_MD5SIG
3669 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3671 skb_set_hash(skb, READ_ONCE(tcp_rsk(req)->txhash), PKT_HASH_TYPE_L4);
3672 /* bpf program will be interested in the tcp_flags */
3673 TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3674 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3676 syn_skb) + sizeof(*th);
3678 skb_push(skb, tcp_header_size);
3679 skb_reset_transport_header(skb);
3681 th = (struct tcphdr *)skb->data;
3682 memset(th, 0, sizeof(struct tcphdr));
3685 tcp_ecn_make_synack(req, th);
3686 th->source = htons(ireq->ir_num);
3687 th->dest = ireq->ir_rmt_port;
3688 skb->mark = ireq->ir_mark;
3689 skb->ip_summed = CHECKSUM_PARTIAL;
3690 th->seq = htonl(tcp_rsk(req)->snt_isn);
3691 /* XXX data is queued and acked as is. No buffer/window check */
3692 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3694 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3695 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3696 tcp_options_write(th, NULL, &opts);
3697 th->doff = (tcp_header_size >> 2);
3698 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3700 #ifdef CONFIG_TCP_MD5SIG
3701 /* Okay, we have all we need - do the md5 hash if needed */
3703 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3704 md5, req_to_sk(req), skb);
3708 bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
3709 synack_type, &opts);
3711 skb_set_delivery_time(skb, now, true);
3712 tcp_add_tx_delay(skb, tp);
3716 EXPORT_SYMBOL(tcp_make_synack);
3718 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3720 struct inet_connection_sock *icsk = inet_csk(sk);
3721 const struct tcp_congestion_ops *ca;
3722 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3724 if (ca_key == TCP_CA_UNSPEC)
3728 ca = tcp_ca_find_key(ca_key);
3729 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3730 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3731 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3732 icsk->icsk_ca_ops = ca;
3737 /* Do all connect socket setups that can be done AF independent. */
3738 static void tcp_connect_init(struct sock *sk)
3740 const struct dst_entry *dst = __sk_dst_get(sk);
3741 struct tcp_sock *tp = tcp_sk(sk);
3745 /* We'll fix this up when we get a response from the other end.
3746 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3748 tp->tcp_header_len = sizeof(struct tcphdr);
3749 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps))
3750 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3752 /* If user gave his TCP_MAXSEG, record it to clamp */
3753 if (tp->rx_opt.user_mss)
3754 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3757 tcp_sync_mss(sk, dst_mtu(dst));
3759 tcp_ca_dst_init(sk, dst);
3761 if (!tp->window_clamp)
3762 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3763 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3765 tcp_initialize_rcv_mss(sk);
3767 /* limit the window selection if the user enforce a smaller rx buffer */
3768 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3769 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3770 tp->window_clamp = tcp_full_space(sk);
3772 rcv_wnd = tcp_rwnd_init_bpf(sk);
3774 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3776 tcp_select_initial_window(sk, tcp_full_space(sk),
3777 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3780 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling),
3784 tp->rx_opt.rcv_wscale = rcv_wscale;
3785 tp->rcv_ssthresh = tp->rcv_wnd;
3787 WRITE_ONCE(sk->sk_err, 0);
3788 sock_reset_flag(sk, SOCK_DONE);
3791 tcp_write_queue_purge(sk);
3792 tp->snd_una = tp->write_seq;
3793 tp->snd_sml = tp->write_seq;
3794 tp->snd_up = tp->write_seq;
3795 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3797 if (likely(!tp->repair))
3800 tp->rcv_tstamp = tcp_jiffies32;
3801 tp->rcv_wup = tp->rcv_nxt;
3802 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3804 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3805 inet_csk(sk)->icsk_retransmits = 0;
3806 tcp_clear_retrans(tp);
3809 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3811 struct tcp_sock *tp = tcp_sk(sk);
3812 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3814 tcb->end_seq += skb->len;
3815 __skb_header_release(skb);
3816 sk_wmem_queued_add(sk, skb->truesize);
3817 sk_mem_charge(sk, skb->truesize);
3818 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3819 tp->packets_out += tcp_skb_pcount(skb);
3822 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3823 * queue a data-only packet after the regular SYN, such that regular SYNs
3824 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3825 * only the SYN sequence, the data are retransmitted in the first ACK.
3826 * If cookie is not cached or other error occurs, falls back to send a
3827 * regular SYN with Fast Open cookie request option.
3829 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3831 struct inet_connection_sock *icsk = inet_csk(sk);
3832 struct tcp_sock *tp = tcp_sk(sk);
3833 struct tcp_fastopen_request *fo = tp->fastopen_req;
3834 struct page_frag *pfrag = sk_page_frag(sk);
3835 struct sk_buff *syn_data;
3838 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3839 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3842 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3843 * user-MSS. Reserve maximum option space for middleboxes that add
3844 * private TCP options. The cost is reduced data space in SYN :(
3846 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3847 /* Sync mss_cache after updating the mss_clamp */
3848 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3850 space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3851 MAX_TCP_OPTION_SPACE;
3853 space = min_t(size_t, space, fo->size);
3856 !skb_page_frag_refill(min_t(size_t, space, PAGE_SIZE),
3857 pfrag, sk->sk_allocation))
3859 syn_data = tcp_stream_alloc_skb(sk, sk->sk_allocation, false);
3862 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3864 space = min_t(size_t, space, pfrag->size - pfrag->offset);
3865 space = tcp_wmem_schedule(sk, space);
3868 space = copy_page_from_iter(pfrag->page, pfrag->offset,
3869 space, &fo->data->msg_iter);
3870 if (unlikely(!space)) {
3871 tcp_skb_tsorted_anchor_cleanup(syn_data);
3872 kfree_skb(syn_data);
3875 skb_fill_page_desc(syn_data, 0, pfrag->page,
3876 pfrag->offset, space);
3877 page_ref_inc(pfrag->page);
3878 pfrag->offset += space;
3879 skb_len_add(syn_data, space);
3880 skb_zcopy_set(syn_data, fo->uarg, NULL);
3882 /* No more data pending in inet_wait_for_connect() */
3883 if (space == fo->size)
3887 tcp_connect_queue_skb(sk, syn_data);
3889 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3891 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3893 skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true);
3895 /* Now full SYN+DATA was cloned and sent (or not),
3896 * remove the SYN from the original skb (syn_data)
3897 * we keep in write queue in case of a retransmit, as we
3898 * also have the SYN packet (with no data) in the same queue.
3900 TCP_SKB_CB(syn_data)->seq++;
3901 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3903 tp->syn_data = (fo->copied > 0);
3904 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3905 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3909 /* data was not sent, put it in write_queue */
3910 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3911 tp->packets_out -= tcp_skb_pcount(syn_data);
3914 /* Send a regular SYN with Fast Open cookie request option */
3915 if (fo->cookie.len > 0)
3917 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3919 tp->syn_fastopen = 0;
3921 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3925 /* Build a SYN and send it off. */
3926 int tcp_connect(struct sock *sk)
3928 struct tcp_sock *tp = tcp_sk(sk);
3929 struct sk_buff *buff;
3932 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3934 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3935 return -EHOSTUNREACH; /* Routing failure or similar. */
3937 tcp_connect_init(sk);
3939 if (unlikely(tp->repair)) {
3940 tcp_finish_connect(sk, NULL);
3944 buff = tcp_stream_alloc_skb(sk, sk->sk_allocation, true);
3945 if (unlikely(!buff))
3948 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3949 tcp_mstamp_refresh(tp);
3950 tp->retrans_stamp = tcp_time_stamp(tp);
3951 tcp_connect_queue_skb(sk, buff);
3952 tcp_ecn_send_syn(sk, buff);
3953 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3955 /* Send off SYN; include data in Fast Open. */
3956 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3957 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3958 if (err == -ECONNREFUSED)
3961 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3962 * in order to make this packet get counted in tcpOutSegs.
3964 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3965 tp->pushed_seq = tp->write_seq;
3966 buff = tcp_send_head(sk);
3967 if (unlikely(buff)) {
3968 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3969 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3971 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3973 /* Timer for repeating the SYN until an answer. */
3974 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3975 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3978 EXPORT_SYMBOL(tcp_connect);
3980 /* Send out a delayed ack, the caller does the policy checking
3981 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3984 void tcp_send_delayed_ack(struct sock *sk)
3986 struct inet_connection_sock *icsk = inet_csk(sk);
3987 int ato = icsk->icsk_ack.ato;
3988 unsigned long timeout;
3990 if (ato > TCP_DELACK_MIN) {
3991 const struct tcp_sock *tp = tcp_sk(sk);
3992 int max_ato = HZ / 2;
3994 if (inet_csk_in_pingpong_mode(sk) ||
3995 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3996 max_ato = TCP_DELACK_MAX;
3998 /* Slow path, intersegment interval is "high". */
4000 /* If some rtt estimate is known, use it to bound delayed ack.
4001 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
4005 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
4012 ato = min(ato, max_ato);
4015 ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max);
4017 /* Stay within the limit we were given */
4018 timeout = jiffies + ato;
4020 /* Use new timeout only if there wasn't a older one earlier. */
4021 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
4022 /* If delack timer is about to expire, send ACK now. */
4023 if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
4028 if (!time_before(timeout, icsk->icsk_ack.timeout))
4029 timeout = icsk->icsk_ack.timeout;
4031 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
4032 icsk->icsk_ack.timeout = timeout;
4033 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
4036 /* This routine sends an ack and also updates the window. */
4037 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
4039 struct sk_buff *buff;
4041 /* If we have been reset, we may not send again. */
4042 if (sk->sk_state == TCP_CLOSE)
4045 /* We are not putting this on the write queue, so
4046 * tcp_transmit_skb() will set the ownership to this
4049 buff = alloc_skb(MAX_TCP_HEADER,
4050 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4051 if (unlikely(!buff)) {
4052 struct inet_connection_sock *icsk = inet_csk(sk);
4053 unsigned long delay;
4055 delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
4056 if (delay < TCP_RTO_MAX)
4057 icsk->icsk_ack.retry++;
4058 inet_csk_schedule_ack(sk);
4059 icsk->icsk_ack.ato = TCP_ATO_MIN;
4060 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
4064 /* Reserve space for headers and prepare control bits. */
4065 skb_reserve(buff, MAX_TCP_HEADER);
4066 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
4068 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
4070 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
4072 skb_set_tcp_pure_ack(buff);
4074 /* Send it off, this clears delayed acks for us. */
4075 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
4077 EXPORT_SYMBOL_GPL(__tcp_send_ack);
4079 void tcp_send_ack(struct sock *sk)
4081 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
4084 /* This routine sends a packet with an out of date sequence
4085 * number. It assumes the other end will try to ack it.
4087 * Question: what should we make while urgent mode?
4088 * 4.4BSD forces sending single byte of data. We cannot send
4089 * out of window data, because we have SND.NXT==SND.MAX...
4091 * Current solution: to send TWO zero-length segments in urgent mode:
4092 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
4093 * out-of-date with SND.UNA-1 to probe window.
4095 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
4097 struct tcp_sock *tp = tcp_sk(sk);
4098 struct sk_buff *skb;
4100 /* We don't queue it, tcp_transmit_skb() sets ownership. */
4101 skb = alloc_skb(MAX_TCP_HEADER,
4102 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4106 /* Reserve space for headers and set control bits. */
4107 skb_reserve(skb, MAX_TCP_HEADER);
4108 /* Use a previous sequence. This should cause the other
4109 * end to send an ack. Don't queue or clone SKB, just
4112 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
4113 NET_INC_STATS(sock_net(sk), mib);
4114 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4117 /* Called from setsockopt( ... TCP_REPAIR ) */
4118 void tcp_send_window_probe(struct sock *sk)
4120 if (sk->sk_state == TCP_ESTABLISHED) {
4121 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4122 tcp_mstamp_refresh(tcp_sk(sk));
4123 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4127 /* Initiate keepalive or window probe from timer. */
4128 int tcp_write_wakeup(struct sock *sk, int mib)
4130 struct tcp_sock *tp = tcp_sk(sk);
4131 struct sk_buff *skb;
4133 if (sk->sk_state == TCP_CLOSE)
4136 skb = tcp_send_head(sk);
4137 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4139 unsigned int mss = tcp_current_mss(sk);
4140 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4142 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4143 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4145 /* We are probing the opening of a window
4146 * but the window size is != 0
4147 * must have been a result SWS avoidance ( sender )
4149 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4151 seg_size = min(seg_size, mss);
4152 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4153 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4154 skb, seg_size, mss, GFP_ATOMIC))
4156 } else if (!tcp_skb_pcount(skb))
4157 tcp_set_skb_tso_segs(skb, mss);
4159 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4160 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4162 tcp_event_new_data_sent(sk, skb);
4165 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4166 tcp_xmit_probe_skb(sk, 1, mib);
4167 return tcp_xmit_probe_skb(sk, 0, mib);
4171 /* A window probe timeout has occurred. If window is not closed send
4172 * a partial packet else a zero probe.
4174 void tcp_send_probe0(struct sock *sk)
4176 struct inet_connection_sock *icsk = inet_csk(sk);
4177 struct tcp_sock *tp = tcp_sk(sk);
4178 struct net *net = sock_net(sk);
4179 unsigned long timeout;
4182 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4184 if (tp->packets_out || tcp_write_queue_empty(sk)) {
4185 /* Cancel probe timer, if it is not required. */
4186 icsk->icsk_probes_out = 0;
4187 icsk->icsk_backoff = 0;
4188 icsk->icsk_probes_tstamp = 0;
4192 icsk->icsk_probes_out++;
4194 if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
4195 icsk->icsk_backoff++;
4196 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
4198 /* If packet was not sent due to local congestion,
4199 * Let senders fight for local resources conservatively.
4201 timeout = TCP_RESOURCE_PROBE_INTERVAL;
4204 timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout);
4205 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
4208 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4210 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4214 /* Paired with WRITE_ONCE() in sock_setsockopt() */
4215 if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED)
4216 WRITE_ONCE(tcp_rsk(req)->txhash, net_tx_rndhash());
4217 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4220 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4221 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4222 if (unlikely(tcp_passive_fastopen(sk))) {
4223 /* sk has const attribute because listeners are lockless.
4224 * However in this case, we are dealing with a passive fastopen
4225 * socket thus we can change total_retrans value.
4227 tcp_sk_rw(sk)->total_retrans++;
4229 trace_tcp_retransmit_synack(sk, req);
4233 EXPORT_SYMBOL(tcp_rtx_synack);