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 = tp->snd_cwnd;
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 tp->snd_cwnd = max(cwnd, restart_cwnd);
156 tp->snd_cwnd_stamp = tcp_jiffies32;
157 tp->snd_cwnd_used = 0;
160 /* Congestion state accounting after a packet has been sent. */
161 static void tcp_event_data_sent(struct tcp_sock *tp,
164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const u32 now = tcp_jiffies32;
167 if (tcp_packets_in_flight(tp) == 0)
168 tcp_ca_event(sk, CA_EVENT_TX_START);
172 /* If it is a reply for ato after last received
173 * packet, enter pingpong mode.
175 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
176 inet_csk_enter_pingpong_mode(sk);
179 /* Account for an ACK we sent. */
180 static inline void tcp_event_ack_sent(struct sock *sk, u32 rcv_nxt)
182 struct tcp_sock *tp = tcp_sk(sk);
184 if (unlikely(tp->compressed_ack)) {
185 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
187 tp->compressed_ack = 0;
188 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
192 if (unlikely(rcv_nxt != tp->rcv_nxt))
193 return; /* Special ACK sent by DCTCP to reflect ECN */
194 tcp_dec_quickack_mode(sk);
195 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
198 /* Determine a window scaling and initial window to offer.
199 * Based on the assumption that the given amount of space
200 * will be offered. Store the results in the tp structure.
201 * NOTE: for smooth operation initial space offering should
202 * be a multiple of mss if possible. We assume here that mss >= 1.
203 * This MUST be enforced by all callers.
205 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
206 __u32 *rcv_wnd, __u32 *window_clamp,
207 int wscale_ok, __u8 *rcv_wscale,
210 unsigned int space = (__space < 0 ? 0 : __space);
212 /* If no clamp set the clamp to the max possible scaled window */
213 if (*window_clamp == 0)
214 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
215 space = min(*window_clamp, space);
217 /* Quantize space offering to a multiple of mss if possible. */
219 space = rounddown(space, mss);
221 /* NOTE: offering an initial window larger than 32767
222 * will break some buggy TCP stacks. If the admin tells us
223 * it is likely we could be speaking with such a buggy stack
224 * we will truncate our initial window offering to 32K-1
225 * unless the remote has sent us a window scaling option,
226 * which we interpret as a sign the remote TCP is not
227 * misinterpreting the window field as a signed quantity.
229 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
230 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
232 (*rcv_wnd) = min_t(u32, space, U16_MAX);
235 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
239 /* Set window scaling on max possible window */
240 space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
241 space = max_t(u32, space, READ_ONCE(sysctl_rmem_max));
242 space = min_t(u32, space, *window_clamp);
243 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
246 /* Set the clamp no higher than max representable value */
247 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
249 EXPORT_SYMBOL(tcp_select_initial_window);
251 /* Chose a new window to advertise, update state in tcp_sock for the
252 * socket, and return result with RFC1323 scaling applied. The return
253 * value can be stuffed directly into th->window for an outgoing
256 static u16 tcp_select_window(struct sock *sk)
258 struct tcp_sock *tp = tcp_sk(sk);
259 u32 old_win = tp->rcv_wnd;
260 u32 cur_win = tcp_receive_window(tp);
261 u32 new_win = __tcp_select_window(sk);
263 /* Never shrink the offered window */
264 if (new_win < cur_win) {
265 /* Danger Will Robinson!
266 * Don't update rcv_wup/rcv_wnd here or else
267 * we will not be able to advertise a zero
268 * window in time. --DaveM
270 * Relax Will Robinson.
273 NET_INC_STATS(sock_net(sk),
274 LINUX_MIB_TCPWANTZEROWINDOWADV);
275 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
277 tp->rcv_wnd = new_win;
278 tp->rcv_wup = tp->rcv_nxt;
280 /* Make sure we do not exceed the maximum possible
283 if (!tp->rx_opt.rcv_wscale &&
284 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
285 new_win = min(new_win, MAX_TCP_WINDOW);
287 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
289 /* RFC1323 scaling applied */
290 new_win >>= tp->rx_opt.rcv_wscale;
292 /* If we advertise zero window, disable fast path. */
296 NET_INC_STATS(sock_net(sk),
297 LINUX_MIB_TCPTOZEROWINDOWADV);
298 } else if (old_win == 0) {
299 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
305 /* Packet ECN state for a SYN-ACK */
306 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
308 const struct tcp_sock *tp = tcp_sk(sk);
310 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
311 if (!(tp->ecn_flags & TCP_ECN_OK))
312 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
313 else if (tcp_ca_needs_ecn(sk) ||
314 tcp_bpf_ca_needs_ecn(sk))
318 /* Packet ECN state for a SYN. */
319 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
321 struct tcp_sock *tp = tcp_sk(sk);
322 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
323 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
324 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
327 const struct dst_entry *dst = __sk_dst_get(sk);
329 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
336 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
337 tp->ecn_flags = TCP_ECN_OK;
338 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
343 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
345 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
346 /* tp->ecn_flags are cleared at a later point in time when
347 * SYN ACK is ultimatively being received.
349 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
353 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
355 if (inet_rsk(req)->ecn_ok)
359 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
362 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
363 struct tcphdr *th, int tcp_header_len)
365 struct tcp_sock *tp = tcp_sk(sk);
367 if (tp->ecn_flags & TCP_ECN_OK) {
368 /* Not-retransmitted data segment: set ECT and inject CWR. */
369 if (skb->len != tcp_header_len &&
370 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
372 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
373 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
375 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
377 } else if (!tcp_ca_needs_ecn(sk)) {
378 /* ACK or retransmitted segment: clear ECT|CE */
379 INET_ECN_dontxmit(sk);
381 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
386 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
387 * auto increment end seqno.
389 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
391 skb->ip_summed = CHECKSUM_PARTIAL;
393 TCP_SKB_CB(skb)->tcp_flags = flags;
394 TCP_SKB_CB(skb)->sacked = 0;
396 tcp_skb_pcount_set(skb, 1);
398 TCP_SKB_CB(skb)->seq = seq;
399 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
401 TCP_SKB_CB(skb)->end_seq = seq;
404 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
406 return tp->snd_una != tp->snd_up;
409 #define OPTION_SACK_ADVERTISE (1 << 0)
410 #define OPTION_TS (1 << 1)
411 #define OPTION_MD5 (1 << 2)
412 #define OPTION_WSCALE (1 << 3)
413 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
414 #define OPTION_SMC (1 << 9)
415 #define OPTION_MPTCP (1 << 10)
417 static void smc_options_write(__be32 *ptr, u16 *options)
419 #if IS_ENABLED(CONFIG_SMC)
420 if (static_branch_unlikely(&tcp_have_smc)) {
421 if (unlikely(OPTION_SMC & *options)) {
422 *ptr++ = htonl((TCPOPT_NOP << 24) |
425 (TCPOLEN_EXP_SMC_BASE));
426 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
432 struct tcp_out_options {
433 u16 options; /* bit field of OPTION_* */
434 u16 mss; /* 0 to disable */
435 u8 ws; /* window scale, 0 to disable */
436 u8 num_sack_blocks; /* number of SACK blocks to include */
437 u8 hash_size; /* bytes in hash_location */
438 u8 bpf_opt_len; /* length of BPF hdr option */
439 __u8 *hash_location; /* temporary pointer, overloaded */
440 __u32 tsval, tsecr; /* need to include OPTION_TS */
441 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
442 struct mptcp_out_options mptcp;
445 static void mptcp_options_write(__be32 *ptr, struct tcp_out_options *opts)
447 #if IS_ENABLED(CONFIG_MPTCP)
448 if (unlikely(OPTION_MPTCP & opts->options))
449 mptcp_write_options(ptr, &opts->mptcp);
453 #ifdef CONFIG_CGROUP_BPF
454 static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
455 enum tcp_synack_type synack_type)
458 return BPF_WRITE_HDR_TCP_CURRENT_MSS;
460 if (unlikely(synack_type == TCP_SYNACK_COOKIE))
461 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
466 /* req, syn_skb and synack_type are used when writing synack */
467 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
468 struct request_sock *req,
469 struct sk_buff *syn_skb,
470 enum tcp_synack_type synack_type,
471 struct tcp_out_options *opts,
472 unsigned int *remaining)
474 struct bpf_sock_ops_kern sock_ops;
477 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
478 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
482 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
485 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
487 sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
490 /* The listen "sk" cannot be passed here because
491 * it is not locked. It would not make too much
492 * sense to do bpf_setsockopt(listen_sk) based
493 * on individual connection request also.
495 * Thus, "req" is passed here and the cgroup-bpf-progs
496 * of the listen "sk" will be run.
498 * "req" is also used here for fastopen even the "sk" here is
499 * a fullsock "child" sk. It is to keep the behavior
500 * consistent between fastopen and non-fastopen on
501 * the bpf programming side.
503 sock_ops.sk = (struct sock *)req;
504 sock_ops.syn_skb = syn_skb;
506 sock_owned_by_me(sk);
508 sock_ops.is_fullsock = 1;
512 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
513 sock_ops.remaining_opt_len = *remaining;
514 /* tcp_current_mss() does not pass a skb */
516 bpf_skops_init_skb(&sock_ops, skb, 0);
518 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
520 if (err || sock_ops.remaining_opt_len == *remaining)
523 opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
524 /* round up to 4 bytes */
525 opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
527 *remaining -= opts->bpf_opt_len;
530 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
531 struct request_sock *req,
532 struct sk_buff *syn_skb,
533 enum tcp_synack_type synack_type,
534 struct tcp_out_options *opts)
536 u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
537 struct bpf_sock_ops_kern sock_ops;
540 if (likely(!max_opt_len))
543 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
545 sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
548 sock_ops.sk = (struct sock *)req;
549 sock_ops.syn_skb = syn_skb;
551 sock_owned_by_me(sk);
553 sock_ops.is_fullsock = 1;
557 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
558 sock_ops.remaining_opt_len = max_opt_len;
559 first_opt_off = tcp_hdrlen(skb) - max_opt_len;
560 bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
562 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
567 nr_written = max_opt_len - sock_ops.remaining_opt_len;
569 if (nr_written < max_opt_len)
570 memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
571 max_opt_len - nr_written);
574 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
575 struct request_sock *req,
576 struct sk_buff *syn_skb,
577 enum tcp_synack_type synack_type,
578 struct tcp_out_options *opts,
579 unsigned int *remaining)
583 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
584 struct request_sock *req,
585 struct sk_buff *syn_skb,
586 enum tcp_synack_type synack_type,
587 struct tcp_out_options *opts)
592 /* Write previously computed TCP options to the packet.
594 * Beware: Something in the Internet is very sensitive to the ordering of
595 * TCP options, we learned this through the hard way, so be careful here.
596 * Luckily we can at least blame others for their non-compliance but from
597 * inter-operability perspective it seems that we're somewhat stuck with
598 * the ordering which we have been using if we want to keep working with
599 * those broken things (not that it currently hurts anybody as there isn't
600 * particular reason why the ordering would need to be changed).
602 * At least SACK_PERM as the first option is known to lead to a disaster
603 * (but it may well be that other scenarios fail similarly).
605 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
606 struct tcp_out_options *opts)
608 u16 options = opts->options; /* mungable copy */
610 if (unlikely(OPTION_MD5 & options)) {
611 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
612 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
613 /* overload cookie hash location */
614 opts->hash_location = (__u8 *)ptr;
618 if (unlikely(opts->mss)) {
619 *ptr++ = htonl((TCPOPT_MSS << 24) |
620 (TCPOLEN_MSS << 16) |
624 if (likely(OPTION_TS & options)) {
625 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
626 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
627 (TCPOLEN_SACK_PERM << 16) |
628 (TCPOPT_TIMESTAMP << 8) |
630 options &= ~OPTION_SACK_ADVERTISE;
632 *ptr++ = htonl((TCPOPT_NOP << 24) |
634 (TCPOPT_TIMESTAMP << 8) |
637 *ptr++ = htonl(opts->tsval);
638 *ptr++ = htonl(opts->tsecr);
641 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
642 *ptr++ = htonl((TCPOPT_NOP << 24) |
644 (TCPOPT_SACK_PERM << 8) |
648 if (unlikely(OPTION_WSCALE & options)) {
649 *ptr++ = htonl((TCPOPT_NOP << 24) |
650 (TCPOPT_WINDOW << 16) |
651 (TCPOLEN_WINDOW << 8) |
655 if (unlikely(opts->num_sack_blocks)) {
656 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
657 tp->duplicate_sack : tp->selective_acks;
660 *ptr++ = htonl((TCPOPT_NOP << 24) |
663 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
664 TCPOLEN_SACK_PERBLOCK)));
666 for (this_sack = 0; this_sack < opts->num_sack_blocks;
668 *ptr++ = htonl(sp[this_sack].start_seq);
669 *ptr++ = htonl(sp[this_sack].end_seq);
672 tp->rx_opt.dsack = 0;
675 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
676 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
678 u32 len; /* Fast Open option length */
681 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
682 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
683 TCPOPT_FASTOPEN_MAGIC);
684 p += TCPOLEN_EXP_FASTOPEN_BASE;
686 len = TCPOLEN_FASTOPEN_BASE + foc->len;
687 *p++ = TCPOPT_FASTOPEN;
691 memcpy(p, foc->val, foc->len);
692 if ((len & 3) == 2) {
693 p[foc->len] = TCPOPT_NOP;
694 p[foc->len + 1] = TCPOPT_NOP;
696 ptr += (len + 3) >> 2;
699 smc_options_write(ptr, &options);
701 mptcp_options_write(ptr, opts);
704 static void smc_set_option(const struct tcp_sock *tp,
705 struct tcp_out_options *opts,
706 unsigned int *remaining)
708 #if IS_ENABLED(CONFIG_SMC)
709 if (static_branch_unlikely(&tcp_have_smc)) {
711 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
712 opts->options |= OPTION_SMC;
713 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
720 static void smc_set_option_cond(const struct tcp_sock *tp,
721 const struct inet_request_sock *ireq,
722 struct tcp_out_options *opts,
723 unsigned int *remaining)
725 #if IS_ENABLED(CONFIG_SMC)
726 if (static_branch_unlikely(&tcp_have_smc)) {
727 if (tp->syn_smc && ireq->smc_ok) {
728 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
729 opts->options |= OPTION_SMC;
730 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
737 static void mptcp_set_option_cond(const struct request_sock *req,
738 struct tcp_out_options *opts,
739 unsigned int *remaining)
741 if (rsk_is_mptcp(req)) {
744 if (mptcp_synack_options(req, &size, &opts->mptcp)) {
745 if (*remaining >= size) {
746 opts->options |= OPTION_MPTCP;
753 /* Compute TCP options for SYN packets. This is not the final
754 * network wire format yet.
756 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
757 struct tcp_out_options *opts,
758 struct tcp_md5sig_key **md5)
760 struct tcp_sock *tp = tcp_sk(sk);
761 unsigned int remaining = MAX_TCP_OPTION_SPACE;
762 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
765 #ifdef CONFIG_TCP_MD5SIG
766 if (static_branch_unlikely(&tcp_md5_needed) &&
767 rcu_access_pointer(tp->md5sig_info)) {
768 *md5 = tp->af_specific->md5_lookup(sk, sk);
770 opts->options |= OPTION_MD5;
771 remaining -= TCPOLEN_MD5SIG_ALIGNED;
776 /* We always get an MSS option. The option bytes which will be seen in
777 * normal data packets should timestamps be used, must be in the MSS
778 * advertised. But we subtract them from tp->mss_cache so that
779 * calculations in tcp_sendmsg are simpler etc. So account for this
780 * fact here if necessary. If we don't do this correctly, as a
781 * receiver we won't recognize data packets as being full sized when we
782 * should, and thus we won't abide by the delayed ACK rules correctly.
783 * SACKs don't matter, we never delay an ACK when we have any of those
785 opts->mss = tcp_advertise_mss(sk);
786 remaining -= TCPOLEN_MSS_ALIGNED;
788 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps) && !*md5)) {
789 opts->options |= OPTION_TS;
790 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
791 opts->tsecr = tp->rx_opt.ts_recent;
792 remaining -= TCPOLEN_TSTAMP_ALIGNED;
794 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) {
795 opts->ws = tp->rx_opt.rcv_wscale;
796 opts->options |= OPTION_WSCALE;
797 remaining -= TCPOLEN_WSCALE_ALIGNED;
799 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) {
800 opts->options |= OPTION_SACK_ADVERTISE;
801 if (unlikely(!(OPTION_TS & opts->options)))
802 remaining -= TCPOLEN_SACKPERM_ALIGNED;
805 if (fastopen && fastopen->cookie.len >= 0) {
806 u32 need = fastopen->cookie.len;
808 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
809 TCPOLEN_FASTOPEN_BASE;
810 need = (need + 3) & ~3U; /* Align to 32 bits */
811 if (remaining >= need) {
812 opts->options |= OPTION_FAST_OPEN_COOKIE;
813 opts->fastopen_cookie = &fastopen->cookie;
815 tp->syn_fastopen = 1;
816 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
820 smc_set_option(tp, opts, &remaining);
822 if (sk_is_mptcp(sk)) {
825 if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
826 opts->options |= OPTION_MPTCP;
831 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
833 return MAX_TCP_OPTION_SPACE - remaining;
836 /* Set up TCP options for SYN-ACKs. */
837 static unsigned int tcp_synack_options(const struct sock *sk,
838 struct request_sock *req,
839 unsigned int mss, struct sk_buff *skb,
840 struct tcp_out_options *opts,
841 const struct tcp_md5sig_key *md5,
842 struct tcp_fastopen_cookie *foc,
843 enum tcp_synack_type synack_type,
844 struct sk_buff *syn_skb)
846 struct inet_request_sock *ireq = inet_rsk(req);
847 unsigned int remaining = MAX_TCP_OPTION_SPACE;
849 #ifdef CONFIG_TCP_MD5SIG
851 opts->options |= OPTION_MD5;
852 remaining -= TCPOLEN_MD5SIG_ALIGNED;
854 /* We can't fit any SACK blocks in a packet with MD5 + TS
855 * options. There was discussion about disabling SACK
856 * rather than TS in order to fit in better with old,
857 * buggy kernels, but that was deemed to be unnecessary.
859 if (synack_type != TCP_SYNACK_COOKIE)
860 ireq->tstamp_ok &= !ireq->sack_ok;
864 /* We always send an MSS option. */
866 remaining -= TCPOLEN_MSS_ALIGNED;
868 if (likely(ireq->wscale_ok)) {
869 opts->ws = ireq->rcv_wscale;
870 opts->options |= OPTION_WSCALE;
871 remaining -= TCPOLEN_WSCALE_ALIGNED;
873 if (likely(ireq->tstamp_ok)) {
874 opts->options |= OPTION_TS;
875 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
876 opts->tsecr = READ_ONCE(req->ts_recent);
877 remaining -= TCPOLEN_TSTAMP_ALIGNED;
879 if (likely(ireq->sack_ok)) {
880 opts->options |= OPTION_SACK_ADVERTISE;
881 if (unlikely(!ireq->tstamp_ok))
882 remaining -= TCPOLEN_SACKPERM_ALIGNED;
884 if (foc != NULL && foc->len >= 0) {
887 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
888 TCPOLEN_FASTOPEN_BASE;
889 need = (need + 3) & ~3U; /* Align to 32 bits */
890 if (remaining >= need) {
891 opts->options |= OPTION_FAST_OPEN_COOKIE;
892 opts->fastopen_cookie = foc;
897 mptcp_set_option_cond(req, opts, &remaining);
899 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
901 bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb,
902 synack_type, opts, &remaining);
904 return MAX_TCP_OPTION_SPACE - remaining;
907 /* Compute TCP options for ESTABLISHED sockets. This is not the
908 * final wire format yet.
910 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
911 struct tcp_out_options *opts,
912 struct tcp_md5sig_key **md5)
914 struct tcp_sock *tp = tcp_sk(sk);
915 unsigned int size = 0;
916 unsigned int eff_sacks;
921 #ifdef CONFIG_TCP_MD5SIG
922 if (static_branch_unlikely(&tcp_md5_needed) &&
923 rcu_access_pointer(tp->md5sig_info)) {
924 *md5 = tp->af_specific->md5_lookup(sk, sk);
926 opts->options |= OPTION_MD5;
927 size += TCPOLEN_MD5SIG_ALIGNED;
932 if (likely(tp->rx_opt.tstamp_ok)) {
933 opts->options |= OPTION_TS;
934 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
935 opts->tsecr = tp->rx_opt.ts_recent;
936 size += TCPOLEN_TSTAMP_ALIGNED;
939 /* MPTCP options have precedence over SACK for the limited TCP
940 * option space because a MPTCP connection would be forced to
941 * fall back to regular TCP if a required multipath option is
942 * missing. SACK still gets a chance to use whatever space is
945 if (sk_is_mptcp(sk)) {
946 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
947 unsigned int opt_size = 0;
949 if (mptcp_established_options(sk, skb, &opt_size, remaining,
951 opts->options |= OPTION_MPTCP;
956 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
957 if (unlikely(eff_sacks)) {
958 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
959 if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
960 TCPOLEN_SACK_PERBLOCK))
963 opts->num_sack_blocks =
964 min_t(unsigned int, eff_sacks,
965 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
966 TCPOLEN_SACK_PERBLOCK);
968 size += TCPOLEN_SACK_BASE_ALIGNED +
969 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
972 if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
973 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
974 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
976 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
978 size = MAX_TCP_OPTION_SPACE - remaining;
985 /* TCP SMALL QUEUES (TSQ)
987 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
988 * to reduce RTT and bufferbloat.
989 * We do this using a special skb destructor (tcp_wfree).
991 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
992 * needs to be reallocated in a driver.
993 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
995 * Since transmit from skb destructor is forbidden, we use a tasklet
996 * to process all sockets that eventually need to send more skbs.
997 * We use one tasklet per cpu, with its own queue of sockets.
1000 struct tasklet_struct tasklet;
1001 struct list_head head; /* queue of tcp sockets */
1003 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
1005 static void tcp_tsq_write(struct sock *sk)
1007 if ((1 << sk->sk_state) &
1008 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1009 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
1010 struct tcp_sock *tp = tcp_sk(sk);
1012 if (tp->lost_out > tp->retrans_out &&
1013 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
1014 tcp_mstamp_refresh(tp);
1015 tcp_xmit_retransmit_queue(sk);
1018 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
1023 static void tcp_tsq_handler(struct sock *sk)
1026 if (!sock_owned_by_user(sk))
1028 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
1033 * One tasklet per cpu tries to send more skbs.
1034 * We run in tasklet context but need to disable irqs when
1035 * transferring tsq->head because tcp_wfree() might
1036 * interrupt us (non NAPI drivers)
1038 static void tcp_tasklet_func(unsigned long data)
1040 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
1042 unsigned long flags;
1043 struct list_head *q, *n;
1044 struct tcp_sock *tp;
1047 local_irq_save(flags);
1048 list_splice_init(&tsq->head, &list);
1049 local_irq_restore(flags);
1051 list_for_each_safe(q, n, &list) {
1052 tp = list_entry(q, struct tcp_sock, tsq_node);
1053 list_del(&tp->tsq_node);
1055 sk = (struct sock *)tp;
1056 smp_mb__before_atomic();
1057 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
1059 tcp_tsq_handler(sk);
1064 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
1065 TCPF_WRITE_TIMER_DEFERRED | \
1066 TCPF_DELACK_TIMER_DEFERRED | \
1067 TCPF_MTU_REDUCED_DEFERRED)
1069 * tcp_release_cb - tcp release_sock() callback
1072 * called from release_sock() to perform protocol dependent
1073 * actions before socket release.
1075 void tcp_release_cb(struct sock *sk)
1077 unsigned long flags, nflags;
1079 /* perform an atomic operation only if at least one flag is set */
1081 flags = sk->sk_tsq_flags;
1082 if (!(flags & TCP_DEFERRED_ALL))
1084 nflags = flags & ~TCP_DEFERRED_ALL;
1085 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
1087 if (flags & TCPF_TSQ_DEFERRED) {
1091 /* Here begins the tricky part :
1092 * We are called from release_sock() with :
1094 * 2) sk_lock.slock spinlock held
1095 * 3) socket owned by us (sk->sk_lock.owned == 1)
1097 * But following code is meant to be called from BH handlers,
1098 * so we should keep BH disabled, but early release socket ownership
1100 sock_release_ownership(sk);
1102 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1103 tcp_write_timer_handler(sk);
1106 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1107 tcp_delack_timer_handler(sk);
1110 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1111 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1115 EXPORT_SYMBOL(tcp_release_cb);
1117 void __init tcp_tasklet_init(void)
1121 for_each_possible_cpu(i) {
1122 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
1124 INIT_LIST_HEAD(&tsq->head);
1125 tasklet_init(&tsq->tasklet,
1127 (unsigned long)tsq);
1132 * Write buffer destructor automatically called from kfree_skb.
1133 * We can't xmit new skbs from this context, as we might already
1136 void tcp_wfree(struct sk_buff *skb)
1138 struct sock *sk = skb->sk;
1139 struct tcp_sock *tp = tcp_sk(sk);
1140 unsigned long flags, nval, oval;
1142 /* Keep one reference on sk_wmem_alloc.
1143 * Will be released by sk_free() from here or tcp_tasklet_func()
1145 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1147 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
1148 * Wait until our queues (qdisc + devices) are drained.
1150 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1151 * - chance for incoming ACK (processed by another cpu maybe)
1152 * to migrate this flow (skb->ooo_okay will be eventually set)
1154 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1157 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
1158 struct tsq_tasklet *tsq;
1161 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1164 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1165 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
1169 /* queue this socket to tasklet queue */
1170 local_irq_save(flags);
1171 tsq = this_cpu_ptr(&tsq_tasklet);
1172 empty = list_empty(&tsq->head);
1173 list_add(&tp->tsq_node, &tsq->head);
1175 tasklet_schedule(&tsq->tasklet);
1176 local_irq_restore(flags);
1183 /* Note: Called under soft irq.
1184 * We can call TCP stack right away, unless socket is owned by user.
1186 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1188 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1189 struct sock *sk = (struct sock *)tp;
1191 tcp_tsq_handler(sk);
1194 return HRTIMER_NORESTART;
1197 static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1200 struct tcp_sock *tp = tcp_sk(sk);
1202 if (sk->sk_pacing_status != SK_PACING_NONE) {
1203 unsigned long rate = sk->sk_pacing_rate;
1205 /* Original sch_fq does not pace first 10 MSS
1206 * Note that tp->data_segs_out overflows after 2^32 packets,
1207 * this is a minor annoyance.
1209 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1210 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1211 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1213 /* take into account OS jitter */
1214 len_ns -= min_t(u64, len_ns / 2, credit);
1215 tp->tcp_wstamp_ns += len_ns;
1218 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1221 INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1222 INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1223 INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1225 /* This routine actually transmits TCP packets queued in by
1226 * tcp_do_sendmsg(). This is used by both the initial
1227 * transmission and possible later retransmissions.
1228 * All SKB's seen here are completely headerless. It is our
1229 * job to build the TCP header, and pass the packet down to
1230 * IP so it can do the same plus pass the packet off to the
1233 * We are working here with either a clone of the original
1234 * SKB, or a fresh unique copy made by the retransmit engine.
1236 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1237 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1239 const struct inet_connection_sock *icsk = inet_csk(sk);
1240 struct inet_sock *inet;
1241 struct tcp_sock *tp;
1242 struct tcp_skb_cb *tcb;
1243 struct tcp_out_options opts;
1244 unsigned int tcp_options_size, tcp_header_size;
1245 struct sk_buff *oskb = NULL;
1246 struct tcp_md5sig_key *md5;
1251 BUG_ON(!skb || !tcp_skb_pcount(skb));
1253 prior_wstamp = tp->tcp_wstamp_ns;
1254 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1255 skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
1257 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1261 tcp_skb_tsorted_save(oskb) {
1262 if (unlikely(skb_cloned(oskb)))
1263 skb = pskb_copy(oskb, gfp_mask);
1265 skb = skb_clone(oskb, gfp_mask);
1266 } tcp_skb_tsorted_restore(oskb);
1270 /* retransmit skbs might have a non zero value in skb->dev
1271 * because skb->dev is aliased with skb->rbnode.rb_left
1277 tcb = TCP_SKB_CB(skb);
1278 memset(&opts, 0, sizeof(opts));
1280 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1281 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1283 tcp_options_size = tcp_established_options(sk, skb, &opts,
1285 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1286 * at receiver : This slightly improve GRO performance.
1287 * Note that we do not force the PSH flag for non GSO packets,
1288 * because they might be sent under high congestion events,
1289 * and in this case it is better to delay the delivery of 1-MSS
1290 * packets and thus the corresponding ACK packet that would
1291 * release the following packet.
1293 if (tcp_skb_pcount(skb) > 1)
1294 tcb->tcp_flags |= TCPHDR_PSH;
1296 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1298 /* if no packet is in qdisc/device queue, then allow XPS to select
1299 * another queue. We can be called from tcp_tsq_handler()
1300 * which holds one reference to sk.
1302 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1303 * One way to get this would be to set skb->truesize = 2 on them.
1305 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1307 /* If we had to use memory reserve to allocate this skb,
1308 * this might cause drops if packet is looped back :
1309 * Other socket might not have SOCK_MEMALLOC.
1310 * Packets not looped back do not care about pfmemalloc.
1312 skb->pfmemalloc = 0;
1314 skb_push(skb, tcp_header_size);
1315 skb_reset_transport_header(skb);
1319 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1320 skb_set_hash_from_sk(skb, sk);
1321 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1323 skb_set_dst_pending_confirm(skb, READ_ONCE(sk->sk_dst_pending_confirm));
1325 /* Build TCP header and checksum it. */
1326 th = (struct tcphdr *)skb->data;
1327 th->source = inet->inet_sport;
1328 th->dest = inet->inet_dport;
1329 th->seq = htonl(tcb->seq);
1330 th->ack_seq = htonl(rcv_nxt);
1331 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1337 /* The urg_mode check is necessary during a below snd_una win probe */
1338 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1339 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1340 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1342 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1343 th->urg_ptr = htons(0xFFFF);
1348 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1349 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1350 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1351 th->window = htons(tcp_select_window(sk));
1352 tcp_ecn_send(sk, skb, th, tcp_header_size);
1354 /* RFC1323: The window in SYN & SYN/ACK segments
1357 th->window = htons(min(tp->rcv_wnd, 65535U));
1359 #ifdef CONFIG_TCP_MD5SIG
1360 /* Calculate the MD5 hash, as we have all we need now */
1362 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1363 tp->af_specific->calc_md5_hash(opts.hash_location,
1368 /* BPF prog is the last one writing header option */
1369 bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1371 INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1372 tcp_v6_send_check, tcp_v4_send_check,
1375 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1376 tcp_event_ack_sent(sk, rcv_nxt);
1378 if (skb->len != tcp_header_size) {
1379 tcp_event_data_sent(tp, sk);
1380 tp->data_segs_out += tcp_skb_pcount(skb);
1381 tp->bytes_sent += skb->len - tcp_header_size;
1384 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1385 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1386 tcp_skb_pcount(skb));
1388 tp->segs_out += tcp_skb_pcount(skb);
1389 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1390 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1391 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1393 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1395 /* Cleanup our debris for IP stacks */
1396 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1397 sizeof(struct inet6_skb_parm)));
1399 tcp_add_tx_delay(skb, tp);
1401 err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1402 inet6_csk_xmit, ip_queue_xmit,
1403 sk, skb, &inet->cork.fl);
1405 if (unlikely(err > 0)) {
1407 err = net_xmit_eval(err);
1410 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1411 tcp_rate_skb_sent(sk, oskb);
1416 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1419 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1420 tcp_sk(sk)->rcv_nxt);
1423 /* This routine just queues the buffer for sending.
1425 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1426 * otherwise socket can stall.
1428 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1430 struct tcp_sock *tp = tcp_sk(sk);
1432 /* Advance write_seq and place onto the write_queue. */
1433 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1434 __skb_header_release(skb);
1435 tcp_add_write_queue_tail(sk, skb);
1436 sk_wmem_queued_add(sk, skb->truesize);
1437 sk_mem_charge(sk, skb->truesize);
1440 /* Initialize TSO segments for a packet. */
1441 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1443 if (skb->len <= mss_now) {
1444 /* Avoid the costly divide in the normal
1447 tcp_skb_pcount_set(skb, 1);
1448 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1450 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1451 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1455 /* Pcount in the middle of the write queue got changed, we need to do various
1456 * tweaks to fix counters
1458 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1460 struct tcp_sock *tp = tcp_sk(sk);
1462 tp->packets_out -= decr;
1464 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1465 tp->sacked_out -= decr;
1466 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1467 tp->retrans_out -= decr;
1468 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1469 tp->lost_out -= decr;
1471 /* Reno case is special. Sigh... */
1472 if (tcp_is_reno(tp) && decr > 0)
1473 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1475 if (tp->lost_skb_hint &&
1476 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1477 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1478 tp->lost_cnt_hint -= decr;
1480 tcp_verify_left_out(tp);
1483 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1485 return TCP_SKB_CB(skb)->txstamp_ack ||
1486 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1489 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1491 struct skb_shared_info *shinfo = skb_shinfo(skb);
1493 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1494 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1495 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1496 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1498 shinfo->tx_flags &= ~tsflags;
1499 shinfo2->tx_flags |= tsflags;
1500 swap(shinfo->tskey, shinfo2->tskey);
1501 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1502 TCP_SKB_CB(skb)->txstamp_ack = 0;
1506 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1508 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1509 TCP_SKB_CB(skb)->eor = 0;
1512 /* Insert buff after skb on the write or rtx queue of sk. */
1513 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1514 struct sk_buff *buff,
1516 enum tcp_queue tcp_queue)
1518 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1519 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1521 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1524 /* Function to create two new TCP segments. Shrinks the given segment
1525 * to the specified size and appends a new segment with the rest of the
1526 * packet to the list. This won't be called frequently, I hope.
1527 * Remember, these are still headerless SKBs at this point.
1529 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1530 struct sk_buff *skb, u32 len,
1531 unsigned int mss_now, gfp_t gfp)
1533 struct tcp_sock *tp = tcp_sk(sk);
1534 struct sk_buff *buff;
1535 int nsize, old_factor;
1540 if (WARN_ON(len > skb->len))
1543 nsize = skb_headlen(skb) - len;
1547 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1548 * We need some allowance to not penalize applications setting small
1550 * Also allow first and last skb in retransmit queue to be split.
1552 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1553 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1554 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1555 skb != tcp_rtx_queue_head(sk) &&
1556 skb != tcp_rtx_queue_tail(sk))) {
1557 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1561 if (skb_unclone(skb, gfp))
1564 /* Get a new skb... force flag on. */
1565 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1567 return -ENOMEM; /* We'll just try again later. */
1568 skb_copy_decrypted(buff, skb);
1570 sk_wmem_queued_add(sk, buff->truesize);
1571 sk_mem_charge(sk, buff->truesize);
1572 nlen = skb->len - len - nsize;
1573 buff->truesize += nlen;
1574 skb->truesize -= nlen;
1576 /* Correct the sequence numbers. */
1577 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1578 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1579 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1581 /* PSH and FIN should only be set in the second packet. */
1582 flags = TCP_SKB_CB(skb)->tcp_flags;
1583 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1584 TCP_SKB_CB(buff)->tcp_flags = flags;
1585 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1586 tcp_skb_fragment_eor(skb, buff);
1588 skb_split(skb, buff, len);
1590 buff->ip_summed = CHECKSUM_PARTIAL;
1592 buff->tstamp = skb->tstamp;
1593 tcp_fragment_tstamp(skb, buff);
1595 old_factor = tcp_skb_pcount(skb);
1597 /* Fix up tso_factor for both original and new SKB. */
1598 tcp_set_skb_tso_segs(skb, mss_now);
1599 tcp_set_skb_tso_segs(buff, mss_now);
1601 /* Update delivered info for the new segment */
1602 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1604 /* If this packet has been sent out already, we must
1605 * adjust the various packet counters.
1607 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1608 int diff = old_factor - tcp_skb_pcount(skb) -
1609 tcp_skb_pcount(buff);
1612 tcp_adjust_pcount(sk, skb, diff);
1615 /* Link BUFF into the send queue. */
1616 __skb_header_release(buff);
1617 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1618 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1619 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1624 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1625 * data is not copied, but immediately discarded.
1627 static int __pskb_trim_head(struct sk_buff *skb, int len)
1629 struct skb_shared_info *shinfo;
1632 eat = min_t(int, len, skb_headlen(skb));
1634 __skb_pull(skb, eat);
1641 shinfo = skb_shinfo(skb);
1642 for (i = 0; i < shinfo->nr_frags; i++) {
1643 int size = skb_frag_size(&shinfo->frags[i]);
1646 skb_frag_unref(skb, i);
1649 shinfo->frags[k] = shinfo->frags[i];
1651 skb_frag_off_add(&shinfo->frags[k], eat);
1652 skb_frag_size_sub(&shinfo->frags[k], eat);
1658 shinfo->nr_frags = k;
1660 skb->data_len -= len;
1661 skb->len = skb->data_len;
1665 /* Remove acked data from a packet in the transmit queue. */
1666 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1670 if (skb_unclone(skb, GFP_ATOMIC))
1673 delta_truesize = __pskb_trim_head(skb, len);
1675 TCP_SKB_CB(skb)->seq += len;
1676 skb->ip_summed = CHECKSUM_PARTIAL;
1678 if (delta_truesize) {
1679 skb->truesize -= delta_truesize;
1680 sk_wmem_queued_add(sk, -delta_truesize);
1681 sk_mem_uncharge(sk, delta_truesize);
1684 /* Any change of skb->len requires recalculation of tso factor. */
1685 if (tcp_skb_pcount(skb) > 1)
1686 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1691 /* Calculate MSS not accounting any TCP options. */
1692 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1694 const struct tcp_sock *tp = tcp_sk(sk);
1695 const struct inet_connection_sock *icsk = inet_csk(sk);
1698 /* Calculate base mss without TCP options:
1699 It is MMS_S - sizeof(tcphdr) of rfc1122
1701 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1703 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1704 if (icsk->icsk_af_ops->net_frag_header_len) {
1705 const struct dst_entry *dst = __sk_dst_get(sk);
1707 if (dst && dst_allfrag(dst))
1708 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1711 /* Clamp it (mss_clamp does not include tcp options) */
1712 if (mss_now > tp->rx_opt.mss_clamp)
1713 mss_now = tp->rx_opt.mss_clamp;
1715 /* Now subtract optional transport overhead */
1716 mss_now -= icsk->icsk_ext_hdr_len;
1718 /* Then reserve room for full set of TCP options and 8 bytes of data */
1719 mss_now = max(mss_now,
1720 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss));
1724 /* Calculate MSS. Not accounting for SACKs here. */
1725 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1727 /* Subtract TCP options size, not including SACKs */
1728 return __tcp_mtu_to_mss(sk, pmtu) -
1729 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1731 EXPORT_SYMBOL(tcp_mtu_to_mss);
1733 /* Inverse of above */
1734 int tcp_mss_to_mtu(struct sock *sk, int mss)
1736 const struct tcp_sock *tp = tcp_sk(sk);
1737 const struct inet_connection_sock *icsk = inet_csk(sk);
1741 tp->tcp_header_len +
1742 icsk->icsk_ext_hdr_len +
1743 icsk->icsk_af_ops->net_header_len;
1745 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1746 if (icsk->icsk_af_ops->net_frag_header_len) {
1747 const struct dst_entry *dst = __sk_dst_get(sk);
1749 if (dst && dst_allfrag(dst))
1750 mtu += icsk->icsk_af_ops->net_frag_header_len;
1754 EXPORT_SYMBOL(tcp_mss_to_mtu);
1756 /* MTU probing init per socket */
1757 void tcp_mtup_init(struct sock *sk)
1759 struct tcp_sock *tp = tcp_sk(sk);
1760 struct inet_connection_sock *icsk = inet_csk(sk);
1761 struct net *net = sock_net(sk);
1763 icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1764 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1765 icsk->icsk_af_ops->net_header_len;
1766 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss));
1767 icsk->icsk_mtup.probe_size = 0;
1768 if (icsk->icsk_mtup.enabled)
1769 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1771 EXPORT_SYMBOL(tcp_mtup_init);
1773 /* This function synchronize snd mss to current pmtu/exthdr set.
1775 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1776 for TCP options, but includes only bare TCP header.
1778 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1779 It is minimum of user_mss and mss received with SYN.
1780 It also does not include TCP options.
1782 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1784 tp->mss_cache is current effective sending mss, including
1785 all tcp options except for SACKs. It is evaluated,
1786 taking into account current pmtu, but never exceeds
1787 tp->rx_opt.mss_clamp.
1789 NOTE1. rfc1122 clearly states that advertised MSS
1790 DOES NOT include either tcp or ip options.
1792 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1793 are READ ONLY outside this function. --ANK (980731)
1795 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1797 struct tcp_sock *tp = tcp_sk(sk);
1798 struct inet_connection_sock *icsk = inet_csk(sk);
1801 if (icsk->icsk_mtup.search_high > pmtu)
1802 icsk->icsk_mtup.search_high = pmtu;
1804 mss_now = tcp_mtu_to_mss(sk, pmtu);
1805 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1807 /* And store cached results */
1808 icsk->icsk_pmtu_cookie = pmtu;
1809 if (icsk->icsk_mtup.enabled)
1810 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1811 tp->mss_cache = mss_now;
1815 EXPORT_SYMBOL(tcp_sync_mss);
1817 /* Compute the current effective MSS, taking SACKs and IP options,
1818 * and even PMTU discovery events into account.
1820 unsigned int tcp_current_mss(struct sock *sk)
1822 const struct tcp_sock *tp = tcp_sk(sk);
1823 const struct dst_entry *dst = __sk_dst_get(sk);
1825 unsigned int header_len;
1826 struct tcp_out_options opts;
1827 struct tcp_md5sig_key *md5;
1829 mss_now = tp->mss_cache;
1832 u32 mtu = dst_mtu(dst);
1833 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1834 mss_now = tcp_sync_mss(sk, mtu);
1837 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1838 sizeof(struct tcphdr);
1839 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1840 * some common options. If this is an odd packet (because we have SACK
1841 * blocks etc) then our calculated header_len will be different, and
1842 * we have to adjust mss_now correspondingly */
1843 if (header_len != tp->tcp_header_len) {
1844 int delta = (int) header_len - tp->tcp_header_len;
1851 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1852 * As additional protections, we do not touch cwnd in retransmission phases,
1853 * and if application hit its sndbuf limit recently.
1855 static void tcp_cwnd_application_limited(struct sock *sk)
1857 struct tcp_sock *tp = tcp_sk(sk);
1859 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1860 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1861 /* Limited by application or receiver window. */
1862 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1863 u32 win_used = max(tp->snd_cwnd_used, init_win);
1864 if (win_used < tp->snd_cwnd) {
1865 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1866 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1868 tp->snd_cwnd_used = 0;
1870 tp->snd_cwnd_stamp = tcp_jiffies32;
1873 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1875 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1876 struct tcp_sock *tp = tcp_sk(sk);
1878 /* Track the strongest available signal of the degree to which the cwnd
1879 * is fully utilized. If cwnd-limited then remember that fact for the
1880 * current window. If not cwnd-limited then track the maximum number of
1881 * outstanding packets in the current window. (If cwnd-limited then we
1882 * chose to not update tp->max_packets_out to avoid an extra else
1883 * clause with no functional impact.)
1885 if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
1887 (!tp->is_cwnd_limited &&
1888 tp->packets_out > tp->max_packets_out)) {
1889 tp->is_cwnd_limited = is_cwnd_limited;
1890 tp->max_packets_out = tp->packets_out;
1891 tp->cwnd_usage_seq = tp->snd_nxt;
1894 if (tcp_is_cwnd_limited(sk)) {
1895 /* Network is feed fully. */
1896 tp->snd_cwnd_used = 0;
1897 tp->snd_cwnd_stamp = tcp_jiffies32;
1899 /* Network starves. */
1900 if (tp->packets_out > tp->snd_cwnd_used)
1901 tp->snd_cwnd_used = tp->packets_out;
1903 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
1904 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1905 !ca_ops->cong_control)
1906 tcp_cwnd_application_limited(sk);
1908 /* The following conditions together indicate the starvation
1909 * is caused by insufficient sender buffer:
1910 * 1) just sent some data (see tcp_write_xmit)
1911 * 2) not cwnd limited (this else condition)
1912 * 3) no more data to send (tcp_write_queue_empty())
1913 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1915 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1916 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1917 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1918 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1922 /* Minshall's variant of the Nagle send check. */
1923 static bool tcp_minshall_check(const struct tcp_sock *tp)
1925 return after(tp->snd_sml, tp->snd_una) &&
1926 !after(tp->snd_sml, tp->snd_nxt);
1929 /* Update snd_sml if this skb is under mss
1930 * Note that a TSO packet might end with a sub-mss segment
1931 * The test is really :
1932 * if ((skb->len % mss) != 0)
1933 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1934 * But we can avoid doing the divide again given we already have
1935 * skb_pcount = skb->len / mss_now
1937 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1938 const struct sk_buff *skb)
1940 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1941 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1944 /* Return false, if packet can be sent now without violation Nagle's rules:
1945 * 1. It is full sized. (provided by caller in %partial bool)
1946 * 2. Or it contains FIN. (already checked by caller)
1947 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1948 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1949 * With Minshall's modification: all sent small packets are ACKed.
1951 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1955 ((nonagle & TCP_NAGLE_CORK) ||
1956 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1959 /* Return how many segs we'd like on a TSO packet,
1960 * to send one TSO packet per ms
1962 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1967 bytes = min_t(unsigned long,
1968 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift),
1969 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1971 /* Goal is to send at least one packet per ms,
1972 * not one big TSO packet every 100 ms.
1973 * This preserves ACK clocking and is consistent
1974 * with tcp_tso_should_defer() heuristic.
1976 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1981 /* Return the number of segments we want in the skb we are transmitting.
1982 * See if congestion control module wants to decide; otherwise, autosize.
1984 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1986 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1987 u32 min_tso, tso_segs;
1989 min_tso = ca_ops->min_tso_segs ?
1990 ca_ops->min_tso_segs(sk) :
1991 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
1993 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1994 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1997 /* Returns the portion of skb which can be sent right away */
1998 static unsigned int tcp_mss_split_point(const struct sock *sk,
1999 const struct sk_buff *skb,
2000 unsigned int mss_now,
2001 unsigned int max_segs,
2004 const struct tcp_sock *tp = tcp_sk(sk);
2005 u32 partial, needed, window, max_len;
2007 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2008 max_len = mss_now * max_segs;
2010 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2013 needed = min(skb->len, window);
2015 if (max_len <= needed)
2018 partial = needed % mss_now;
2019 /* If last segment is not a full MSS, check if Nagle rules allow us
2020 * to include this last segment in this skb.
2021 * Otherwise, we'll split the skb at last MSS boundary
2023 if (tcp_nagle_check(partial != 0, tp, nonagle))
2024 return needed - partial;
2029 /* Can at least one segment of SKB be sent right now, according to the
2030 * congestion window rules? If so, return how many segments are allowed.
2032 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
2033 const struct sk_buff *skb)
2035 u32 in_flight, cwnd, halfcwnd;
2037 /* Don't be strict about the congestion window for the final FIN. */
2038 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2039 tcp_skb_pcount(skb) == 1)
2042 in_flight = tcp_packets_in_flight(tp);
2043 cwnd = tp->snd_cwnd;
2044 if (in_flight >= cwnd)
2047 /* For better scheduling, ensure we have at least
2048 * 2 GSO packets in flight.
2050 halfcwnd = max(cwnd >> 1, 1U);
2051 return min(halfcwnd, cwnd - in_flight);
2054 /* Initialize TSO state of a skb.
2055 * This must be invoked the first time we consider transmitting
2056 * SKB onto the wire.
2058 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
2060 int tso_segs = tcp_skb_pcount(skb);
2062 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
2063 tcp_set_skb_tso_segs(skb, mss_now);
2064 tso_segs = tcp_skb_pcount(skb);
2070 /* Return true if the Nagle test allows this packet to be
2073 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2074 unsigned int cur_mss, int nonagle)
2076 /* Nagle rule does not apply to frames, which sit in the middle of the
2077 * write_queue (they have no chances to get new data).
2079 * This is implemented in the callers, where they modify the 'nonagle'
2080 * argument based upon the location of SKB in the send queue.
2082 if (nonagle & TCP_NAGLE_PUSH)
2085 /* Don't use the nagle rule for urgent data (or for the final FIN). */
2086 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2089 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
2095 /* Does at least the first segment of SKB fit into the send window? */
2096 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2097 const struct sk_buff *skb,
2098 unsigned int cur_mss)
2100 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2102 if (skb->len > cur_mss)
2103 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2105 return !after(end_seq, tcp_wnd_end(tp));
2108 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2109 * which is put after SKB on the list. It is very much like
2110 * tcp_fragment() except that it may make several kinds of assumptions
2111 * in order to speed up the splitting operation. In particular, we
2112 * know that all the data is in scatter-gather pages, and that the
2113 * packet has never been sent out before (and thus is not cloned).
2115 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2116 unsigned int mss_now, gfp_t gfp)
2118 int nlen = skb->len - len;
2119 struct sk_buff *buff;
2122 /* All of a TSO frame must be composed of paged data. */
2123 if (skb->len != skb->data_len)
2124 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2125 skb, len, mss_now, gfp);
2127 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
2128 if (unlikely(!buff))
2130 skb_copy_decrypted(buff, skb);
2132 sk_wmem_queued_add(sk, buff->truesize);
2133 sk_mem_charge(sk, buff->truesize);
2134 buff->truesize += nlen;
2135 skb->truesize -= nlen;
2137 /* Correct the sequence numbers. */
2138 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2139 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2140 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2142 /* PSH and FIN should only be set in the second packet. */
2143 flags = TCP_SKB_CB(skb)->tcp_flags;
2144 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2145 TCP_SKB_CB(buff)->tcp_flags = flags;
2147 /* This packet was never sent out yet, so no SACK bits. */
2148 TCP_SKB_CB(buff)->sacked = 0;
2150 tcp_skb_fragment_eor(skb, buff);
2152 buff->ip_summed = CHECKSUM_PARTIAL;
2153 skb_split(skb, buff, len);
2154 tcp_fragment_tstamp(skb, buff);
2156 /* Fix up tso_factor for both original and new SKB. */
2157 tcp_set_skb_tso_segs(skb, mss_now);
2158 tcp_set_skb_tso_segs(buff, mss_now);
2160 /* Link BUFF into the send queue. */
2161 __skb_header_release(buff);
2162 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2167 /* Try to defer sending, if possible, in order to minimize the amount
2168 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2170 * This algorithm is from John Heffner.
2172 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2173 bool *is_cwnd_limited,
2174 bool *is_rwnd_limited,
2177 const struct inet_connection_sock *icsk = inet_csk(sk);
2178 u32 send_win, cong_win, limit, in_flight;
2179 struct tcp_sock *tp = tcp_sk(sk);
2180 struct sk_buff *head;
2184 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2187 /* Avoid bursty behavior by allowing defer
2188 * only if the last write was recent (1 ms).
2189 * Note that tp->tcp_wstamp_ns can be in the future if we have
2190 * packets waiting in a qdisc or device for EDT delivery.
2192 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2196 in_flight = tcp_packets_in_flight(tp);
2198 BUG_ON(tcp_skb_pcount(skb) <= 1);
2199 BUG_ON(tp->snd_cwnd <= in_flight);
2201 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2203 /* From in_flight test above, we know that cwnd > in_flight. */
2204 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
2206 limit = min(send_win, cong_win);
2208 /* If a full-sized TSO skb can be sent, do it. */
2209 if (limit >= max_segs * tp->mss_cache)
2212 /* Middle in queue won't get any more data, full sendable already? */
2213 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2216 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2218 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
2220 /* If at least some fraction of a window is available,
2223 chunk /= win_divisor;
2227 /* Different approach, try not to defer past a single
2228 * ACK. Receiver should ACK every other full sized
2229 * frame, so if we have space for more than 3 frames
2232 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2236 /* TODO : use tsorted_sent_queue ? */
2237 head = tcp_rtx_queue_head(sk);
2240 delta = tp->tcp_clock_cache - head->tstamp;
2241 /* If next ACK is likely to come too late (half srtt), do not defer */
2242 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2245 /* Ok, it looks like it is advisable to defer.
2246 * Three cases are tracked :
2247 * 1) We are cwnd-limited
2248 * 2) We are rwnd-limited
2249 * 3) We are application limited.
2251 if (cong_win < send_win) {
2252 if (cong_win <= skb->len) {
2253 *is_cwnd_limited = true;
2257 if (send_win <= skb->len) {
2258 *is_rwnd_limited = true;
2263 /* If this packet won't get more data, do not wait. */
2264 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2265 TCP_SKB_CB(skb)->eor)
2274 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2276 struct inet_connection_sock *icsk = inet_csk(sk);
2277 struct tcp_sock *tp = tcp_sk(sk);
2278 struct net *net = sock_net(sk);
2282 interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2283 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2284 if (unlikely(delta >= interval * HZ)) {
2285 int mss = tcp_current_mss(sk);
2287 /* Update current search range */
2288 icsk->icsk_mtup.probe_size = 0;
2289 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2290 sizeof(struct tcphdr) +
2291 icsk->icsk_af_ops->net_header_len;
2292 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2294 /* Update probe time stamp */
2295 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2299 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2301 struct sk_buff *skb, *next;
2303 skb = tcp_send_head(sk);
2304 tcp_for_write_queue_from_safe(skb, next, sk) {
2305 if (len <= skb->len)
2308 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2317 /* Create a new MTU probe if we are ready.
2318 * MTU probe is regularly attempting to increase the path MTU by
2319 * deliberately sending larger packets. This discovers routing
2320 * changes resulting in larger path MTUs.
2322 * Returns 0 if we should wait to probe (no cwnd available),
2323 * 1 if a probe was sent,
2326 static int tcp_mtu_probe(struct sock *sk)
2328 struct inet_connection_sock *icsk = inet_csk(sk);
2329 struct tcp_sock *tp = tcp_sk(sk);
2330 struct sk_buff *skb, *nskb, *next;
2331 struct net *net = sock_net(sk);
2338 /* Not currently probing/verifying,
2340 * have enough cwnd, and
2341 * not SACKing (the variable headers throw things off)
2343 if (likely(!icsk->icsk_mtup.enabled ||
2344 icsk->icsk_mtup.probe_size ||
2345 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2346 tp->snd_cwnd < 11 ||
2347 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2350 /* Use binary search for probe_size between tcp_mss_base,
2351 * and current mss_clamp. if (search_high - search_low)
2352 * smaller than a threshold, backoff from probing.
2354 mss_now = tcp_current_mss(sk);
2355 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2356 icsk->icsk_mtup.search_low) >> 1);
2357 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2358 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2359 /* When misfortune happens, we are reprobing actively,
2360 * and then reprobe timer has expired. We stick with current
2361 * probing process by not resetting search range to its orignal.
2363 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2364 interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2365 /* Check whether enough time has elaplased for
2366 * another round of probing.
2368 tcp_mtu_check_reprobe(sk);
2372 /* Have enough data in the send queue to probe? */
2373 if (tp->write_seq - tp->snd_nxt < size_needed)
2376 if (tp->snd_wnd < size_needed)
2378 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2381 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2382 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2383 if (!tcp_packets_in_flight(tp))
2389 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2392 /* We're allowed to probe. Build it now. */
2393 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2396 sk_wmem_queued_add(sk, nskb->truesize);
2397 sk_mem_charge(sk, nskb->truesize);
2399 skb = tcp_send_head(sk);
2400 skb_copy_decrypted(nskb, skb);
2402 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2403 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2404 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2405 TCP_SKB_CB(nskb)->sacked = 0;
2407 nskb->ip_summed = CHECKSUM_PARTIAL;
2409 tcp_insert_write_queue_before(nskb, skb, sk);
2410 tcp_highest_sack_replace(sk, skb, nskb);
2413 tcp_for_write_queue_from_safe(skb, next, sk) {
2414 copy = min_t(int, skb->len, probe_size - len);
2415 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2417 if (skb->len <= copy) {
2418 /* We've eaten all the data from this skb.
2420 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2421 /* If this is the last SKB we copy and eor is set
2422 * we need to propagate it to the new skb.
2424 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2425 tcp_skb_collapse_tstamp(nskb, skb);
2426 tcp_unlink_write_queue(skb, sk);
2427 sk_wmem_free_skb(sk, skb);
2429 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2430 ~(TCPHDR_FIN|TCPHDR_PSH);
2431 if (!skb_shinfo(skb)->nr_frags) {
2432 skb_pull(skb, copy);
2434 __pskb_trim_head(skb, copy);
2435 tcp_set_skb_tso_segs(skb, mss_now);
2437 TCP_SKB_CB(skb)->seq += copy;
2442 if (len >= probe_size)
2445 tcp_init_tso_segs(nskb, nskb->len);
2447 /* We're ready to send. If this fails, the probe will
2448 * be resegmented into mss-sized pieces by tcp_write_xmit().
2450 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2451 /* Decrement cwnd here because we are sending
2452 * effectively two packets. */
2454 tcp_event_new_data_sent(sk, nskb);
2456 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2457 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2458 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2466 static bool tcp_pacing_check(struct sock *sk)
2468 struct tcp_sock *tp = tcp_sk(sk);
2470 if (!tcp_needs_internal_pacing(sk))
2473 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2476 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2477 hrtimer_start(&tp->pacing_timer,
2478 ns_to_ktime(tp->tcp_wstamp_ns),
2479 HRTIMER_MODE_ABS_PINNED_SOFT);
2485 static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk)
2487 const struct rb_node *node = sk->tcp_rtx_queue.rb_node;
2489 /* No skb in the rtx queue. */
2493 /* Only one skb in rtx queue. */
2494 return !node->rb_left && !node->rb_right;
2497 /* TCP Small Queues :
2498 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2499 * (These limits are doubled for retransmits)
2501 * - better RTT estimation and ACK scheduling
2504 * Alas, some drivers / subsystems require a fair amount
2505 * of queued bytes to ensure line rate.
2506 * One example is wifi aggregation (802.11 AMPDU)
2508 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2509 unsigned int factor)
2511 unsigned long limit;
2513 limit = max_t(unsigned long,
2515 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2516 if (sk->sk_pacing_status == SK_PACING_NONE)
2517 limit = min_t(unsigned long, limit,
2518 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
2521 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2522 tcp_sk(sk)->tcp_tx_delay) {
2523 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2525 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2526 * approximate our needs assuming an ~100% skb->truesize overhead.
2527 * USEC_PER_SEC is approximated by 2^20.
2528 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2530 extra_bytes >>= (20 - 1);
2531 limit += extra_bytes;
2533 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2534 /* Always send skb if rtx queue is empty or has one skb.
2535 * No need to wait for TX completion to call us back,
2536 * after softirq/tasklet schedule.
2537 * This helps when TX completions are delayed too much.
2539 if (tcp_rtx_queue_empty_or_single_skb(sk))
2542 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2543 /* It is possible TX completion already happened
2544 * before we set TSQ_THROTTLED, so we must
2545 * test again the condition.
2547 smp_mb__after_atomic();
2548 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2554 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2556 const u32 now = tcp_jiffies32;
2557 enum tcp_chrono old = tp->chrono_type;
2559 if (old > TCP_CHRONO_UNSPEC)
2560 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2561 tp->chrono_start = now;
2562 tp->chrono_type = new;
2565 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2567 struct tcp_sock *tp = tcp_sk(sk);
2569 /* If there are multiple conditions worthy of tracking in a
2570 * chronograph then the highest priority enum takes precedence
2571 * over the other conditions. So that if something "more interesting"
2572 * starts happening, stop the previous chrono and start a new one.
2574 if (type > tp->chrono_type)
2575 tcp_chrono_set(tp, type);
2578 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2580 struct tcp_sock *tp = tcp_sk(sk);
2583 /* There are multiple conditions worthy of tracking in a
2584 * chronograph, so that the highest priority enum takes
2585 * precedence over the other conditions (see tcp_chrono_start).
2586 * If a condition stops, we only stop chrono tracking if
2587 * it's the "most interesting" or current chrono we are
2588 * tracking and starts busy chrono if we have pending data.
2590 if (tcp_rtx_and_write_queues_empty(sk))
2591 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2592 else if (type == tp->chrono_type)
2593 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2596 /* This routine writes packets to the network. It advances the
2597 * send_head. This happens as incoming acks open up the remote
2600 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2601 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2602 * account rare use of URG, this is not a big flaw.
2604 * Send at most one packet when push_one > 0. Temporarily ignore
2605 * cwnd limit to force at most one packet out when push_one == 2.
2607 * Returns true, if no segments are in flight and we have queued segments,
2608 * but cannot send anything now because of SWS or another problem.
2610 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2611 int push_one, gfp_t gfp)
2613 struct tcp_sock *tp = tcp_sk(sk);
2614 struct sk_buff *skb;
2615 unsigned int tso_segs, sent_pkts;
2618 bool is_cwnd_limited = false, is_rwnd_limited = false;
2623 tcp_mstamp_refresh(tp);
2625 /* Do MTU probing. */
2626 result = tcp_mtu_probe(sk);
2629 } else if (result > 0) {
2634 max_segs = tcp_tso_segs(sk, mss_now);
2635 while ((skb = tcp_send_head(sk))) {
2638 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2639 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2640 skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2641 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2642 tcp_init_tso_segs(skb, mss_now);
2643 goto repair; /* Skip network transmission */
2646 if (tcp_pacing_check(sk))
2649 tso_segs = tcp_init_tso_segs(skb, mss_now);
2652 cwnd_quota = tcp_cwnd_test(tp, skb);
2655 /* Force out a loss probe pkt. */
2661 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2662 is_rwnd_limited = true;
2666 if (tso_segs == 1) {
2667 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2668 (tcp_skb_is_last(sk, skb) ?
2669 nonagle : TCP_NAGLE_PUSH))))
2673 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2674 &is_rwnd_limited, max_segs))
2679 if (tso_segs > 1 && !tcp_urg_mode(tp))
2680 limit = tcp_mss_split_point(sk, skb, mss_now,
2686 if (skb->len > limit &&
2687 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2690 if (tcp_small_queue_check(sk, skb, 0))
2693 /* Argh, we hit an empty skb(), presumably a thread
2694 * is sleeping in sendmsg()/sk_stream_wait_memory().
2695 * We do not want to send a pure-ack packet and have
2696 * a strange looking rtx queue with empty packet(s).
2698 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2701 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2705 /* Advance the send_head. This one is sent out.
2706 * This call will increment packets_out.
2708 tcp_event_new_data_sent(sk, skb);
2710 tcp_minshall_update(tp, mss_now, skb);
2711 sent_pkts += tcp_skb_pcount(skb);
2717 if (is_rwnd_limited)
2718 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2720 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2722 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2723 if (likely(sent_pkts || is_cwnd_limited))
2724 tcp_cwnd_validate(sk, is_cwnd_limited);
2726 if (likely(sent_pkts)) {
2727 if (tcp_in_cwnd_reduction(sk))
2728 tp->prr_out += sent_pkts;
2730 /* Send one loss probe per tail loss episode. */
2732 tcp_schedule_loss_probe(sk, false);
2735 return !tp->packets_out && !tcp_write_queue_empty(sk);
2738 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2740 struct inet_connection_sock *icsk = inet_csk(sk);
2741 struct tcp_sock *tp = tcp_sk(sk);
2742 u32 timeout, timeout_us, rto_delta_us;
2745 /* Don't do any loss probe on a Fast Open connection before 3WHS
2748 if (rcu_access_pointer(tp->fastopen_rsk))
2751 early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
2752 /* Schedule a loss probe in 2*RTT for SACK capable connections
2753 * not in loss recovery, that are either limited by cwnd or application.
2755 if ((early_retrans != 3 && early_retrans != 4) ||
2756 !tp->packets_out || !tcp_is_sack(tp) ||
2757 (icsk->icsk_ca_state != TCP_CA_Open &&
2758 icsk->icsk_ca_state != TCP_CA_CWR))
2761 /* Probe timeout is 2*rtt. Add minimum RTO to account
2762 * for delayed ack when there's one outstanding packet. If no RTT
2763 * sample is available then probe after TCP_TIMEOUT_INIT.
2766 timeout_us = tp->srtt_us >> 2;
2767 if (tp->packets_out == 1)
2768 timeout_us += tcp_rto_min_us(sk);
2770 timeout_us += TCP_TIMEOUT_MIN_US;
2771 timeout = usecs_to_jiffies(timeout_us);
2773 timeout = TCP_TIMEOUT_INIT;
2776 /* If the RTO formula yields an earlier time, then use that time. */
2777 rto_delta_us = advancing_rto ?
2778 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2779 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2780 if (rto_delta_us > 0)
2781 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2783 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2787 /* Thanks to skb fast clones, we can detect if a prior transmit of
2788 * a packet is still in a qdisc or driver queue.
2789 * In this case, there is very little point doing a retransmit !
2791 static bool skb_still_in_host_queue(const struct sock *sk,
2792 const struct sk_buff *skb)
2794 if (unlikely(skb_fclone_busy(sk, skb))) {
2795 NET_INC_STATS(sock_net(sk),
2796 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2802 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2803 * retransmit the last segment.
2805 void tcp_send_loss_probe(struct sock *sk)
2807 struct tcp_sock *tp = tcp_sk(sk);
2808 struct sk_buff *skb;
2810 int mss = tcp_current_mss(sk);
2812 /* At most one outstanding TLP */
2813 if (tp->tlp_high_seq)
2816 tp->tlp_retrans = 0;
2817 skb = tcp_send_head(sk);
2818 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2819 pcount = tp->packets_out;
2820 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2821 if (tp->packets_out > pcount)
2825 skb = skb_rb_last(&sk->tcp_rtx_queue);
2826 if (unlikely(!skb)) {
2827 WARN_ONCE(tp->packets_out,
2828 "invalid inflight: %u state %u cwnd %u mss %d\n",
2829 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2830 inet_csk(sk)->icsk_pending = 0;
2834 if (skb_still_in_host_queue(sk, skb))
2837 pcount = tcp_skb_pcount(skb);
2838 if (WARN_ON(!pcount))
2841 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2842 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2843 (pcount - 1) * mss, mss,
2846 skb = skb_rb_next(skb);
2849 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2852 if (__tcp_retransmit_skb(sk, skb, 1))
2855 tp->tlp_retrans = 1;
2858 /* Record snd_nxt for loss detection. */
2859 tp->tlp_high_seq = tp->snd_nxt;
2861 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2862 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2863 inet_csk(sk)->icsk_pending = 0;
2868 /* Push out any pending frames which were held back due to
2869 * TCP_CORK or attempt at coalescing tiny packets.
2870 * The socket must be locked by the caller.
2872 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2875 /* If we are closed, the bytes will have to remain here.
2876 * In time closedown will finish, we empty the write queue and
2877 * all will be happy.
2879 if (unlikely(sk->sk_state == TCP_CLOSE))
2882 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2883 sk_gfp_mask(sk, GFP_ATOMIC)))
2884 tcp_check_probe_timer(sk);
2887 /* Send _single_ skb sitting at the send head. This function requires
2888 * true push pending frames to setup probe timer etc.
2890 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2892 struct sk_buff *skb = tcp_send_head(sk);
2894 BUG_ON(!skb || skb->len < mss_now);
2896 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2899 /* This function returns the amount that we can raise the
2900 * usable window based on the following constraints
2902 * 1. The window can never be shrunk once it is offered (RFC 793)
2903 * 2. We limit memory per socket
2906 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2907 * RECV.NEXT + RCV.WIN fixed until:
2908 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2910 * i.e. don't raise the right edge of the window until you can raise
2911 * it at least MSS bytes.
2913 * Unfortunately, the recommended algorithm breaks header prediction,
2914 * since header prediction assumes th->window stays fixed.
2916 * Strictly speaking, keeping th->window fixed violates the receiver
2917 * side SWS prevention criteria. The problem is that under this rule
2918 * a stream of single byte packets will cause the right side of the
2919 * window to always advance by a single byte.
2921 * Of course, if the sender implements sender side SWS prevention
2922 * then this will not be a problem.
2924 * BSD seems to make the following compromise:
2926 * If the free space is less than the 1/4 of the maximum
2927 * space available and the free space is less than 1/2 mss,
2928 * then set the window to 0.
2929 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2930 * Otherwise, just prevent the window from shrinking
2931 * and from being larger than the largest representable value.
2933 * This prevents incremental opening of the window in the regime
2934 * where TCP is limited by the speed of the reader side taking
2935 * data out of the TCP receive queue. It does nothing about
2936 * those cases where the window is constrained on the sender side
2937 * because the pipeline is full.
2939 * BSD also seems to "accidentally" limit itself to windows that are a
2940 * multiple of MSS, at least until the free space gets quite small.
2941 * This would appear to be a side effect of the mbuf implementation.
2942 * Combining these two algorithms results in the observed behavior
2943 * of having a fixed window size at almost all times.
2945 * Below we obtain similar behavior by forcing the offered window to
2946 * a multiple of the mss when it is feasible to do so.
2948 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2949 * Regular options like TIMESTAMP are taken into account.
2951 u32 __tcp_select_window(struct sock *sk)
2953 struct inet_connection_sock *icsk = inet_csk(sk);
2954 struct tcp_sock *tp = tcp_sk(sk);
2955 /* MSS for the peer's data. Previous versions used mss_clamp
2956 * here. I don't know if the value based on our guesses
2957 * of peer's MSS is better for the performance. It's more correct
2958 * but may be worse for the performance because of rcv_mss
2959 * fluctuations. --SAW 1998/11/1
2961 int mss = icsk->icsk_ack.rcv_mss;
2962 int free_space = tcp_space(sk);
2963 int allowed_space = tcp_full_space(sk);
2964 int full_space, window;
2966 if (sk_is_mptcp(sk))
2967 mptcp_space(sk, &free_space, &allowed_space);
2969 full_space = min_t(int, tp->window_clamp, allowed_space);
2971 if (unlikely(mss > full_space)) {
2976 if (free_space < (full_space >> 1)) {
2977 icsk->icsk_ack.quick = 0;
2979 if (tcp_under_memory_pressure(sk))
2980 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2983 /* free_space might become our new window, make sure we don't
2984 * increase it due to wscale.
2986 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2988 /* if free space is less than mss estimate, or is below 1/16th
2989 * of the maximum allowed, try to move to zero-window, else
2990 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2991 * new incoming data is dropped due to memory limits.
2992 * With large window, mss test triggers way too late in order
2993 * to announce zero window in time before rmem limit kicks in.
2995 if (free_space < (allowed_space >> 4) || free_space < mss)
2999 if (free_space > tp->rcv_ssthresh)
3000 free_space = tp->rcv_ssthresh;
3002 /* Don't do rounding if we are using window scaling, since the
3003 * scaled window will not line up with the MSS boundary anyway.
3005 if (tp->rx_opt.rcv_wscale) {
3006 window = free_space;
3008 /* Advertise enough space so that it won't get scaled away.
3009 * Import case: prevent zero window announcement if
3010 * 1<<rcv_wscale > mss.
3012 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
3014 window = tp->rcv_wnd;
3015 /* Get the largest window that is a nice multiple of mss.
3016 * Window clamp already applied above.
3017 * If our current window offering is within 1 mss of the
3018 * free space we just keep it. This prevents the divide
3019 * and multiply from happening most of the time.
3020 * We also don't do any window rounding when the free space
3023 if (window <= free_space - mss || window > free_space)
3024 window = rounddown(free_space, mss);
3025 else if (mss == full_space &&
3026 free_space > window + (full_space >> 1))
3027 window = free_space;
3033 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3034 const struct sk_buff *next_skb)
3036 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3037 const struct skb_shared_info *next_shinfo =
3038 skb_shinfo(next_skb);
3039 struct skb_shared_info *shinfo = skb_shinfo(skb);
3041 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3042 shinfo->tskey = next_shinfo->tskey;
3043 TCP_SKB_CB(skb)->txstamp_ack |=
3044 TCP_SKB_CB(next_skb)->txstamp_ack;
3048 /* Collapses two adjacent SKB's during retransmission. */
3049 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3051 struct tcp_sock *tp = tcp_sk(sk);
3052 struct sk_buff *next_skb = skb_rb_next(skb);
3055 next_skb_size = next_skb->len;
3057 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3059 if (next_skb_size) {
3060 if (next_skb_size <= skb_availroom(skb))
3061 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
3063 else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3066 tcp_highest_sack_replace(sk, next_skb, skb);
3068 /* Update sequence range on original skb. */
3069 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3071 /* Merge over control information. This moves PSH/FIN etc. over */
3072 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3074 /* All done, get rid of second SKB and account for it so
3075 * packet counting does not break.
3077 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3078 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3080 /* changed transmit queue under us so clear hints */
3081 tcp_clear_retrans_hints_partial(tp);
3082 if (next_skb == tp->retransmit_skb_hint)
3083 tp->retransmit_skb_hint = skb;
3085 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3087 tcp_skb_collapse_tstamp(skb, next_skb);
3089 tcp_rtx_queue_unlink_and_free(next_skb, sk);
3093 /* Check if coalescing SKBs is legal. */
3094 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3096 if (tcp_skb_pcount(skb) > 1)
3098 if (skb_cloned(skb))
3100 /* Some heuristics for collapsing over SACK'd could be invented */
3101 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3107 /* Collapse packets in the retransmit queue to make to create
3108 * less packets on the wire. This is only done on retransmission.
3110 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3113 struct tcp_sock *tp = tcp_sk(sk);
3114 struct sk_buff *skb = to, *tmp;
3117 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
3119 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3122 skb_rbtree_walk_from_safe(skb, tmp) {
3123 if (!tcp_can_collapse(sk, skb))
3126 if (!tcp_skb_can_collapse(to, skb))
3139 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3142 if (!tcp_collapse_retrans(sk, to))
3147 /* This retransmits one SKB. Policy decisions and retransmit queue
3148 * state updates are done by the caller. Returns non-zero if an
3149 * error occurred which prevented the send.
3151 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3153 struct inet_connection_sock *icsk = inet_csk(sk);
3154 struct tcp_sock *tp = tcp_sk(sk);
3155 unsigned int cur_mss;
3159 /* Inconclusive MTU probe */
3160 if (icsk->icsk_mtup.probe_size)
3161 icsk->icsk_mtup.probe_size = 0;
3163 /* Do not sent more than we queued. 1/4 is reserved for possible
3164 * copying overhead: fragmentation, tunneling, mangling etc.
3166 if (refcount_read(&sk->sk_wmem_alloc) >
3167 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
3171 if (skb_still_in_host_queue(sk, skb))
3175 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
3176 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3177 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
3178 TCP_SKB_CB(skb)->seq++;
3181 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3185 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3189 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3190 return -EHOSTUNREACH; /* Routing failure or similar. */
3192 cur_mss = tcp_current_mss(sk);
3193 avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3195 /* If receiver has shrunk his window, and skb is out of
3196 * new window, do not retransmit it. The exception is the
3197 * case, when window is shrunk to zero. In this case
3198 * our retransmit of one segment serves as a zero window probe.
3200 if (avail_wnd <= 0) {
3201 if (TCP_SKB_CB(skb)->seq != tp->snd_una)
3203 avail_wnd = cur_mss;
3206 len = cur_mss * segs;
3207 if (len > avail_wnd) {
3208 len = rounddown(avail_wnd, cur_mss);
3212 if (skb->len > len) {
3213 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3214 cur_mss, GFP_ATOMIC))
3215 return -ENOMEM; /* We'll try again later. */
3217 if (skb_unclone(skb, GFP_ATOMIC))
3220 diff = tcp_skb_pcount(skb);
3221 tcp_set_skb_tso_segs(skb, cur_mss);
3222 diff -= tcp_skb_pcount(skb);
3224 tcp_adjust_pcount(sk, skb, diff);
3225 avail_wnd = min_t(int, avail_wnd, cur_mss);
3226 if (skb->len < avail_wnd)
3227 tcp_retrans_try_collapse(sk, skb, avail_wnd);
3230 /* RFC3168, section 6.1.1.1. ECN fallback */
3231 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3232 tcp_ecn_clear_syn(sk, skb);
3234 /* Update global and local TCP statistics. */
3235 segs = tcp_skb_pcount(skb);
3236 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3237 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3238 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3239 tp->total_retrans += segs;
3240 tp->bytes_retrans += skb->len;
3242 /* make sure skb->data is aligned on arches that require it
3243 * and check if ack-trimming & collapsing extended the headroom
3244 * beyond what csum_start can cover.
3246 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3247 skb_headroom(skb) >= 0xFFFF)) {
3248 struct sk_buff *nskb;
3250 tcp_skb_tsorted_save(skb) {
3251 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3254 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3258 } tcp_skb_tsorted_restore(skb);
3261 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3262 tcp_rate_skb_sent(sk, skb);
3265 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3268 /* To avoid taking spuriously low RTT samples based on a timestamp
3269 * for a transmit that never happened, always mark EVER_RETRANS
3271 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3273 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3274 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3275 TCP_SKB_CB(skb)->seq, segs, err);
3278 trace_tcp_retransmit_skb(sk, skb);
3279 } else if (err != -EBUSY) {
3280 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3285 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3287 struct tcp_sock *tp = tcp_sk(sk);
3288 int err = __tcp_retransmit_skb(sk, skb, segs);
3291 #if FASTRETRANS_DEBUG > 0
3292 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3293 net_dbg_ratelimited("retrans_out leaked\n");
3296 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3297 tp->retrans_out += tcp_skb_pcount(skb);
3300 /* Save stamp of the first (attempted) retransmit. */
3301 if (!tp->retrans_stamp)
3302 tp->retrans_stamp = tcp_skb_timestamp(skb);
3304 if (tp->undo_retrans < 0)
3305 tp->undo_retrans = 0;
3306 tp->undo_retrans += tcp_skb_pcount(skb);
3310 /* This gets called after a retransmit timeout, and the initially
3311 * retransmitted data is acknowledged. It tries to continue
3312 * resending the rest of the retransmit queue, until either
3313 * we've sent it all or the congestion window limit is reached.
3315 void tcp_xmit_retransmit_queue(struct sock *sk)
3317 const struct inet_connection_sock *icsk = inet_csk(sk);
3318 struct sk_buff *skb, *rtx_head, *hole = NULL;
3319 struct tcp_sock *tp = tcp_sk(sk);
3320 bool rearm_timer = false;
3324 if (!tp->packets_out)
3327 rtx_head = tcp_rtx_queue_head(sk);
3328 skb = tp->retransmit_skb_hint ?: rtx_head;
3329 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3330 skb_rbtree_walk_from(skb) {
3334 if (tcp_pacing_check(sk))
3337 /* we could do better than to assign each time */
3339 tp->retransmit_skb_hint = skb;
3341 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
3344 sacked = TCP_SKB_CB(skb)->sacked;
3345 /* In case tcp_shift_skb_data() have aggregated large skbs,
3346 * we need to make sure not sending too bigs TSO packets
3348 segs = min_t(int, segs, max_segs);
3350 if (tp->retrans_out >= tp->lost_out) {
3352 } else if (!(sacked & TCPCB_LOST)) {
3353 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3358 if (icsk->icsk_ca_state != TCP_CA_Loss)
3359 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3361 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3364 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3367 if (tcp_small_queue_check(sk, skb, 1))
3370 if (tcp_retransmit_skb(sk, skb, segs))
3373 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3375 if (tcp_in_cwnd_reduction(sk))
3376 tp->prr_out += tcp_skb_pcount(skb);
3378 if (skb == rtx_head &&
3379 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3384 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3385 inet_csk(sk)->icsk_rto,
3389 /* We allow to exceed memory limits for FIN packets to expedite
3390 * connection tear down and (memory) recovery.
3391 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3392 * or even be forced to close flow without any FIN.
3393 * In general, we want to allow one skb per socket to avoid hangs
3394 * with edge trigger epoll()
3396 void sk_forced_mem_schedule(struct sock *sk, int size)
3400 delta = size - sk->sk_forward_alloc;
3403 amt = sk_mem_pages(delta);
3404 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3405 sk_memory_allocated_add(sk, amt);
3407 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3408 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3411 /* Send a FIN. The caller locks the socket for us.
3412 * We should try to send a FIN packet really hard, but eventually give up.
3414 void tcp_send_fin(struct sock *sk)
3416 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3417 struct tcp_sock *tp = tcp_sk(sk);
3419 /* Optimization, tack on the FIN if we have one skb in write queue and
3420 * this skb was not yet sent, or we are under memory pressure.
3421 * Note: in the latter case, FIN packet will be sent after a timeout,
3422 * as TCP stack thinks it has already been transmitted.
3425 if (!tskb && tcp_under_memory_pressure(sk))
3426 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3429 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3430 TCP_SKB_CB(tskb)->end_seq++;
3433 /* This means tskb was already sent.
3434 * Pretend we included the FIN on previous transmit.
3435 * We need to set tp->snd_nxt to the value it would have
3436 * if FIN had been sent. This is because retransmit path
3437 * does not change tp->snd_nxt.
3439 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3443 skb = alloc_skb_fclone(MAX_TCP_HEADER,
3444 sk_gfp_mask(sk, GFP_ATOMIC |
3449 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3450 skb_reserve(skb, MAX_TCP_HEADER);
3451 sk_forced_mem_schedule(sk, skb->truesize);
3452 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3453 tcp_init_nondata_skb(skb, tp->write_seq,
3454 TCPHDR_ACK | TCPHDR_FIN);
3455 tcp_queue_skb(sk, skb);
3457 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3460 /* We get here when a process closes a file descriptor (either due to
3461 * an explicit close() or as a byproduct of exit()'ing) and there
3462 * was unread data in the receive queue. This behavior is recommended
3463 * by RFC 2525, section 2.17. -DaveM
3465 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3467 struct sk_buff *skb;
3469 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3471 /* NOTE: No TCP options attached and we never retransmit this. */
3472 skb = alloc_skb(MAX_TCP_HEADER, priority);
3474 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3478 /* Reserve space for headers and prepare control bits. */
3479 skb_reserve(skb, MAX_TCP_HEADER);
3480 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3481 TCPHDR_ACK | TCPHDR_RST);
3482 tcp_mstamp_refresh(tcp_sk(sk));
3484 if (tcp_transmit_skb(sk, skb, 0, priority))
3485 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3487 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3488 * skb here is different to the troublesome skb, so use NULL
3490 trace_tcp_send_reset(sk, NULL);
3493 /* Send a crossed SYN-ACK during socket establishment.
3494 * WARNING: This routine must only be called when we have already sent
3495 * a SYN packet that crossed the incoming SYN that caused this routine
3496 * to get called. If this assumption fails then the initial rcv_wnd
3497 * and rcv_wscale values will not be correct.
3499 int tcp_send_synack(struct sock *sk)
3501 struct sk_buff *skb;
3503 skb = tcp_rtx_queue_head(sk);
3504 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3505 pr_err("%s: wrong queue state\n", __func__);
3508 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3509 if (skb_cloned(skb)) {
3510 struct sk_buff *nskb;
3512 tcp_skb_tsorted_save(skb) {
3513 nskb = skb_copy(skb, GFP_ATOMIC);
3514 } tcp_skb_tsorted_restore(skb);
3517 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3518 tcp_highest_sack_replace(sk, skb, nskb);
3519 tcp_rtx_queue_unlink_and_free(skb, sk);
3520 __skb_header_release(nskb);
3521 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3522 sk_wmem_queued_add(sk, nskb->truesize);
3523 sk_mem_charge(sk, nskb->truesize);
3527 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3528 tcp_ecn_send_synack(sk, skb);
3530 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3534 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3535 * @sk: listener socket
3536 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3537 * should not use it again.
3538 * @req: request_sock pointer
3539 * @foc: cookie for tcp fast open
3540 * @synack_type: Type of synack to prepare
3541 * @syn_skb: SYN packet just received. It could be NULL for rtx case.
3543 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3544 struct request_sock *req,
3545 struct tcp_fastopen_cookie *foc,
3546 enum tcp_synack_type synack_type,
3547 struct sk_buff *syn_skb)
3549 struct inet_request_sock *ireq = inet_rsk(req);
3550 const struct tcp_sock *tp = tcp_sk(sk);
3551 struct tcp_md5sig_key *md5 = NULL;
3552 struct tcp_out_options opts;
3553 struct sk_buff *skb;
3554 int tcp_header_size;
3559 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3560 if (unlikely(!skb)) {
3564 /* Reserve space for headers. */
3565 skb_reserve(skb, MAX_TCP_HEADER);
3567 switch (synack_type) {
3568 case TCP_SYNACK_NORMAL:
3569 skb_set_owner_w(skb, req_to_sk(req));
3571 case TCP_SYNACK_COOKIE:
3572 /* Under synflood, we do not attach skb to a socket,
3573 * to avoid false sharing.
3576 case TCP_SYNACK_FASTOPEN:
3577 /* sk is a const pointer, because we want to express multiple
3578 * cpu might call us concurrently.
3579 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3581 skb_set_owner_w(skb, (struct sock *)sk);
3584 skb_dst_set(skb, dst);
3586 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3588 memset(&opts, 0, sizeof(opts));
3589 now = tcp_clock_ns();
3590 #ifdef CONFIG_SYN_COOKIES
3591 if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3592 skb->skb_mstamp_ns = cookie_init_timestamp(req, now);
3596 skb->skb_mstamp_ns = now;
3597 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3598 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3601 #ifdef CONFIG_TCP_MD5SIG
3603 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3605 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3606 /* bpf program will be interested in the tcp_flags */
3607 TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3608 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3610 syn_skb) + sizeof(*th);
3612 skb_push(skb, tcp_header_size);
3613 skb_reset_transport_header(skb);
3615 th = (struct tcphdr *)skb->data;
3616 memset(th, 0, sizeof(struct tcphdr));
3619 tcp_ecn_make_synack(req, th);
3620 th->source = htons(ireq->ir_num);
3621 th->dest = ireq->ir_rmt_port;
3622 skb->mark = ireq->ir_mark;
3623 skb->ip_summed = CHECKSUM_PARTIAL;
3624 th->seq = htonl(tcp_rsk(req)->snt_isn);
3625 /* XXX data is queued and acked as is. No buffer/window check */
3626 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3628 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3629 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3630 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3631 th->doff = (tcp_header_size >> 2);
3632 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3634 #ifdef CONFIG_TCP_MD5SIG
3635 /* Okay, we have all we need - do the md5 hash if needed */
3637 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3638 md5, req_to_sk(req), skb);
3642 bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
3643 synack_type, &opts);
3645 skb->skb_mstamp_ns = now;
3646 tcp_add_tx_delay(skb, tp);
3650 EXPORT_SYMBOL(tcp_make_synack);
3652 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3654 struct inet_connection_sock *icsk = inet_csk(sk);
3655 const struct tcp_congestion_ops *ca;
3656 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3658 if (ca_key == TCP_CA_UNSPEC)
3662 ca = tcp_ca_find_key(ca_key);
3663 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3664 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3665 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3666 icsk->icsk_ca_ops = ca;
3671 /* Do all connect socket setups that can be done AF independent. */
3672 static void tcp_connect_init(struct sock *sk)
3674 const struct dst_entry *dst = __sk_dst_get(sk);
3675 struct tcp_sock *tp = tcp_sk(sk);
3679 /* We'll fix this up when we get a response from the other end.
3680 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3682 tp->tcp_header_len = sizeof(struct tcphdr);
3683 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps))
3684 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3686 #ifdef CONFIG_TCP_MD5SIG
3687 if (tp->af_specific->md5_lookup(sk, sk))
3688 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3691 /* If user gave his TCP_MAXSEG, record it to clamp */
3692 if (tp->rx_opt.user_mss)
3693 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3696 tcp_sync_mss(sk, dst_mtu(dst));
3698 tcp_ca_dst_init(sk, dst);
3700 if (!tp->window_clamp)
3701 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3702 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3704 tcp_initialize_rcv_mss(sk);
3706 /* limit the window selection if the user enforce a smaller rx buffer */
3707 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3708 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3709 tp->window_clamp = tcp_full_space(sk);
3711 rcv_wnd = tcp_rwnd_init_bpf(sk);
3713 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3715 tcp_select_initial_window(sk, tcp_full_space(sk),
3716 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3719 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling),
3723 tp->rx_opt.rcv_wscale = rcv_wscale;
3724 tp->rcv_ssthresh = tp->rcv_wnd;
3727 sock_reset_flag(sk, SOCK_DONE);
3730 tcp_write_queue_purge(sk);
3731 tp->snd_una = tp->write_seq;
3732 tp->snd_sml = tp->write_seq;
3733 tp->snd_up = tp->write_seq;
3734 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3736 if (likely(!tp->repair))
3739 tp->rcv_tstamp = tcp_jiffies32;
3740 tp->rcv_wup = tp->rcv_nxt;
3741 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3743 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3744 inet_csk(sk)->icsk_retransmits = 0;
3745 tcp_clear_retrans(tp);
3748 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3750 struct tcp_sock *tp = tcp_sk(sk);
3751 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3753 tcb->end_seq += skb->len;
3754 __skb_header_release(skb);
3755 sk_wmem_queued_add(sk, skb->truesize);
3756 sk_mem_charge(sk, skb->truesize);
3757 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3758 tp->packets_out += tcp_skb_pcount(skb);
3761 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3762 * queue a data-only packet after the regular SYN, such that regular SYNs
3763 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3764 * only the SYN sequence, the data are retransmitted in the first ACK.
3765 * If cookie is not cached or other error occurs, falls back to send a
3766 * regular SYN with Fast Open cookie request option.
3768 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3770 struct inet_connection_sock *icsk = inet_csk(sk);
3771 struct tcp_sock *tp = tcp_sk(sk);
3772 struct tcp_fastopen_request *fo = tp->fastopen_req;
3774 struct sk_buff *syn_data;
3776 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3777 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3780 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3781 * user-MSS. Reserve maximum option space for middleboxes that add
3782 * private TCP options. The cost is reduced data space in SYN :(
3784 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3785 /* Sync mss_cache after updating the mss_clamp */
3786 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3788 space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3789 MAX_TCP_OPTION_SPACE;
3791 space = min_t(size_t, space, fo->size);
3793 /* limit to order-0 allocations */
3794 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3796 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3799 syn_data->ip_summed = CHECKSUM_PARTIAL;
3800 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3802 int copied = copy_from_iter(skb_put(syn_data, space), space,
3803 &fo->data->msg_iter);
3804 if (unlikely(!copied)) {
3805 tcp_skb_tsorted_anchor_cleanup(syn_data);
3806 kfree_skb(syn_data);
3809 if (copied != space) {
3810 skb_trim(syn_data, copied);
3813 skb_zcopy_set(syn_data, fo->uarg, NULL);
3815 /* No more data pending in inet_wait_for_connect() */
3816 if (space == fo->size)
3820 tcp_connect_queue_skb(sk, syn_data);
3822 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3824 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3826 syn->skb_mstamp_ns = syn_data->skb_mstamp_ns;
3828 /* Now full SYN+DATA was cloned and sent (or not),
3829 * remove the SYN from the original skb (syn_data)
3830 * we keep in write queue in case of a retransmit, as we
3831 * also have the SYN packet (with no data) in the same queue.
3833 TCP_SKB_CB(syn_data)->seq++;
3834 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3836 tp->syn_data = (fo->copied > 0);
3837 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3838 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3842 /* data was not sent, put it in write_queue */
3843 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3844 tp->packets_out -= tcp_skb_pcount(syn_data);
3847 /* Send a regular SYN with Fast Open cookie request option */
3848 if (fo->cookie.len > 0)
3850 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3852 tp->syn_fastopen = 0;
3854 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3858 /* Build a SYN and send it off. */
3859 int tcp_connect(struct sock *sk)
3861 struct tcp_sock *tp = tcp_sk(sk);
3862 struct sk_buff *buff;
3865 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3867 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3868 return -EHOSTUNREACH; /* Routing failure or similar. */
3870 tcp_connect_init(sk);
3872 if (unlikely(tp->repair)) {
3873 tcp_finish_connect(sk, NULL);
3877 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3878 if (unlikely(!buff))
3881 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3882 tcp_mstamp_refresh(tp);
3883 tp->retrans_stamp = tcp_time_stamp(tp);
3884 tcp_connect_queue_skb(sk, buff);
3885 tcp_ecn_send_syn(sk, buff);
3886 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3888 /* Send off SYN; include data in Fast Open. */
3889 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3890 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3891 if (err == -ECONNREFUSED)
3894 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3895 * in order to make this packet get counted in tcpOutSegs.
3897 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3898 tp->pushed_seq = tp->write_seq;
3899 buff = tcp_send_head(sk);
3900 if (unlikely(buff)) {
3901 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3902 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3904 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3906 /* Timer for repeating the SYN until an answer. */
3907 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3908 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3911 EXPORT_SYMBOL(tcp_connect);
3913 /* Send out a delayed ack, the caller does the policy checking
3914 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3917 void tcp_send_delayed_ack(struct sock *sk)
3919 struct inet_connection_sock *icsk = inet_csk(sk);
3920 int ato = icsk->icsk_ack.ato;
3921 unsigned long timeout;
3923 if (ato > TCP_DELACK_MIN) {
3924 const struct tcp_sock *tp = tcp_sk(sk);
3925 int max_ato = HZ / 2;
3927 if (inet_csk_in_pingpong_mode(sk) ||
3928 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3929 max_ato = TCP_DELACK_MAX;
3931 /* Slow path, intersegment interval is "high". */
3933 /* If some rtt estimate is known, use it to bound delayed ack.
3934 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3938 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3945 ato = min(ato, max_ato);
3948 ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max);
3950 /* Stay within the limit we were given */
3951 timeout = jiffies + ato;
3953 /* Use new timeout only if there wasn't a older one earlier. */
3954 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3955 /* If delack timer is about to expire, send ACK now. */
3956 if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3961 if (!time_before(timeout, icsk->icsk_ack.timeout))
3962 timeout = icsk->icsk_ack.timeout;
3964 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3965 icsk->icsk_ack.timeout = timeout;
3966 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3969 /* This routine sends an ack and also updates the window. */
3970 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3972 struct sk_buff *buff;
3974 /* If we have been reset, we may not send again. */
3975 if (sk->sk_state == TCP_CLOSE)
3978 /* We are not putting this on the write queue, so
3979 * tcp_transmit_skb() will set the ownership to this
3982 buff = alloc_skb(MAX_TCP_HEADER,
3983 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3984 if (unlikely(!buff)) {
3985 struct inet_connection_sock *icsk = inet_csk(sk);
3986 unsigned long delay;
3988 delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
3989 if (delay < TCP_RTO_MAX)
3990 icsk->icsk_ack.retry++;
3991 inet_csk_schedule_ack(sk);
3992 icsk->icsk_ack.ato = TCP_ATO_MIN;
3993 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
3997 /* Reserve space for headers and prepare control bits. */
3998 skb_reserve(buff, MAX_TCP_HEADER);
3999 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
4001 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
4003 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
4005 skb_set_tcp_pure_ack(buff);
4007 /* Send it off, this clears delayed acks for us. */
4008 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
4010 EXPORT_SYMBOL_GPL(__tcp_send_ack);
4012 void tcp_send_ack(struct sock *sk)
4014 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
4017 /* This routine sends a packet with an out of date sequence
4018 * number. It assumes the other end will try to ack it.
4020 * Question: what should we make while urgent mode?
4021 * 4.4BSD forces sending single byte of data. We cannot send
4022 * out of window data, because we have SND.NXT==SND.MAX...
4024 * Current solution: to send TWO zero-length segments in urgent mode:
4025 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
4026 * out-of-date with SND.UNA-1 to probe window.
4028 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
4030 struct tcp_sock *tp = tcp_sk(sk);
4031 struct sk_buff *skb;
4033 /* We don't queue it, tcp_transmit_skb() sets ownership. */
4034 skb = alloc_skb(MAX_TCP_HEADER,
4035 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4039 /* Reserve space for headers and set control bits. */
4040 skb_reserve(skb, MAX_TCP_HEADER);
4041 /* Use a previous sequence. This should cause the other
4042 * end to send an ack. Don't queue or clone SKB, just
4045 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
4046 NET_INC_STATS(sock_net(sk), mib);
4047 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4050 /* Called from setsockopt( ... TCP_REPAIR ) */
4051 void tcp_send_window_probe(struct sock *sk)
4053 if (sk->sk_state == TCP_ESTABLISHED) {
4054 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4055 tcp_mstamp_refresh(tcp_sk(sk));
4056 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4060 /* Initiate keepalive or window probe from timer. */
4061 int tcp_write_wakeup(struct sock *sk, int mib)
4063 struct tcp_sock *tp = tcp_sk(sk);
4064 struct sk_buff *skb;
4066 if (sk->sk_state == TCP_CLOSE)
4069 skb = tcp_send_head(sk);
4070 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4072 unsigned int mss = tcp_current_mss(sk);
4073 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4075 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4076 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4078 /* We are probing the opening of a window
4079 * but the window size is != 0
4080 * must have been a result SWS avoidance ( sender )
4082 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4084 seg_size = min(seg_size, mss);
4085 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4086 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4087 skb, seg_size, mss, GFP_ATOMIC))
4089 } else if (!tcp_skb_pcount(skb))
4090 tcp_set_skb_tso_segs(skb, mss);
4092 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4093 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4095 tcp_event_new_data_sent(sk, skb);
4098 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4099 tcp_xmit_probe_skb(sk, 1, mib);
4100 return tcp_xmit_probe_skb(sk, 0, mib);
4104 /* A window probe timeout has occurred. If window is not closed send
4105 * a partial packet else a zero probe.
4107 void tcp_send_probe0(struct sock *sk)
4109 struct inet_connection_sock *icsk = inet_csk(sk);
4110 struct tcp_sock *tp = tcp_sk(sk);
4111 struct net *net = sock_net(sk);
4112 unsigned long timeout;
4115 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4117 if (tp->packets_out || tcp_write_queue_empty(sk)) {
4118 /* Cancel probe timer, if it is not required. */
4119 icsk->icsk_probes_out = 0;
4120 icsk->icsk_backoff = 0;
4121 icsk->icsk_probes_tstamp = 0;
4125 icsk->icsk_probes_out++;
4127 if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
4128 icsk->icsk_backoff++;
4129 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
4131 /* If packet was not sent due to local congestion,
4132 * Let senders fight for local resources conservatively.
4134 timeout = TCP_RESOURCE_PROBE_INTERVAL;
4137 timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout);
4138 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
4141 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4143 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4147 tcp_rsk(req)->txhash = net_tx_rndhash();
4148 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4151 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4152 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4153 if (unlikely(tcp_passive_fastopen(sk)))
4154 tcp_sk(sk)->total_retrans++;
4155 trace_tcp_retransmit_synack(sk, req);
4159 EXPORT_SYMBOL(tcp_rtx_synack);