1 // SPDX-License-Identifier: GPL-2.0-only
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
38 #define pr_fmt(fmt) "TCP: " fmt
41 #include <net/mptcp.h>
43 #include <linux/compiler.h>
44 #include <linux/gfp.h>
45 #include <linux/module.h>
46 #include <linux/static_key.h>
48 #include <trace/events/tcp.h>
50 /* Refresh clocks of a TCP socket,
51 * ensuring monotically increasing values.
53 void tcp_mstamp_refresh(struct tcp_sock *tp)
55 u64 val = tcp_clock_ns();
57 tp->tcp_clock_cache = val;
58 tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
61 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
62 int push_one, gfp_t gfp);
64 /* Account for new data that has been sent to the network. */
65 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
67 struct inet_connection_sock *icsk = inet_csk(sk);
68 struct tcp_sock *tp = tcp_sk(sk);
69 unsigned int prior_packets = tp->packets_out;
71 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
73 __skb_unlink(skb, &sk->sk_write_queue);
74 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
76 if (tp->highest_sack == NULL)
77 tp->highest_sack = skb;
79 tp->packets_out += tcp_skb_pcount(skb);
80 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
83 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
88 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
89 * window scaling factor due to loss of precision.
90 * If window has been shrunk, what should we make? It is not clear at all.
91 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
92 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
93 * invalid. OK, let's make this for now:
95 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
97 const struct tcp_sock *tp = tcp_sk(sk);
99 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
100 (tp->rx_opt.wscale_ok &&
101 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
104 return tcp_wnd_end(tp);
107 /* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
121 static __u16 tcp_advertise_mss(struct sock *sk)
123 struct tcp_sock *tp = tcp_sk(sk);
124 const struct dst_entry *dst = __sk_dst_get(sk);
125 int mss = tp->advmss;
128 unsigned int metric = dst_metric_advmss(dst);
139 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism.
142 void tcp_cwnd_restart(struct sock *sk, s32 delta)
144 struct tcp_sock *tp = tcp_sk(sk);
145 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
146 u32 cwnd = tcp_snd_cwnd(tp);
148 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
150 tp->snd_ssthresh = tcp_current_ssthresh(sk);
151 restart_cwnd = min(restart_cwnd, cwnd);
153 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
155 tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd));
156 tp->snd_cwnd_stamp = tcp_jiffies32;
157 tp->snd_cwnd_used = 0;
160 /* Congestion state accounting after a packet has been sent. */
161 static void tcp_event_data_sent(struct tcp_sock *tp,
164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const u32 now = tcp_jiffies32;
167 if (tcp_packets_in_flight(tp) == 0)
168 tcp_ca_event(sk, CA_EVENT_TX_START);
170 /* If this is the first data packet sent in response to the
171 * previous received data,
172 * and it is a reply for ato after last received packet,
173 * increase pingpong count.
175 if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) &&
176 (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
177 inet_csk_inc_pingpong_cnt(sk);
182 /* Account for an ACK we sent. */
183 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
186 struct tcp_sock *tp = tcp_sk(sk);
188 if (unlikely(tp->compressed_ack)) {
189 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
191 tp->compressed_ack = 0;
192 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
196 if (unlikely(rcv_nxt != tp->rcv_nxt))
197 return; /* Special ACK sent by DCTCP to reflect ECN */
198 tcp_dec_quickack_mode(sk, pkts);
199 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
202 /* Determine a window scaling and initial window to offer.
203 * Based on the assumption that the given amount of space
204 * will be offered. Store the results in the tp structure.
205 * NOTE: for smooth operation initial space offering should
206 * be a multiple of mss if possible. We assume here that mss >= 1.
207 * This MUST be enforced by all callers.
209 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
210 __u32 *rcv_wnd, __u32 *window_clamp,
211 int wscale_ok, __u8 *rcv_wscale,
214 unsigned int space = (__space < 0 ? 0 : __space);
216 /* If no clamp set the clamp to the max possible scaled window */
217 if (*window_clamp == 0)
218 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
219 space = min(*window_clamp, space);
221 /* Quantize space offering to a multiple of mss if possible. */
223 space = rounddown(space, mss);
225 /* NOTE: offering an initial window larger than 32767
226 * will break some buggy TCP stacks. If the admin tells us
227 * it is likely we could be speaking with such a buggy stack
228 * we will truncate our initial window offering to 32K-1
229 * unless the remote has sent us a window scaling option,
230 * which we interpret as a sign the remote TCP is not
231 * misinterpreting the window field as a signed quantity.
233 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
234 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
236 (*rcv_wnd) = min_t(u32, space, U16_MAX);
239 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
243 /* Set window scaling on max possible window */
244 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
245 space = max_t(u32, space, sysctl_rmem_max);
246 space = min_t(u32, space, *window_clamp);
247 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
250 /* Set the clamp no higher than max representable value */
251 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
253 EXPORT_SYMBOL(tcp_select_initial_window);
255 /* Chose a new window to advertise, update state in tcp_sock for the
256 * socket, and return result with RFC1323 scaling applied. The return
257 * value can be stuffed directly into th->window for an outgoing
260 static u16 tcp_select_window(struct sock *sk)
262 struct tcp_sock *tp = tcp_sk(sk);
263 u32 old_win = tp->rcv_wnd;
264 u32 cur_win = tcp_receive_window(tp);
265 u32 new_win = __tcp_select_window(sk);
267 /* Never shrink the offered window */
268 if (new_win < cur_win) {
269 /* Danger Will Robinson!
270 * Don't update rcv_wup/rcv_wnd here or else
271 * we will not be able to advertise a zero
272 * window in time. --DaveM
274 * Relax Will Robinson.
277 NET_INC_STATS(sock_net(sk),
278 LINUX_MIB_TCPWANTZEROWINDOWADV);
279 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
281 tp->rcv_wnd = new_win;
282 tp->rcv_wup = tp->rcv_nxt;
284 /* Make sure we do not exceed the maximum possible
287 if (!tp->rx_opt.rcv_wscale &&
288 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
289 new_win = min(new_win, MAX_TCP_WINDOW);
291 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
293 /* RFC1323 scaling applied */
294 new_win >>= tp->rx_opt.rcv_wscale;
296 /* If we advertise zero window, disable fast path. */
300 NET_INC_STATS(sock_net(sk),
301 LINUX_MIB_TCPTOZEROWINDOWADV);
302 } else if (old_win == 0) {
303 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
309 /* Packet ECN state for a SYN-ACK */
310 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
312 const struct tcp_sock *tp = tcp_sk(sk);
314 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
315 if (!(tp->ecn_flags & TCP_ECN_OK))
316 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
317 else if (tcp_ca_needs_ecn(sk) ||
318 tcp_bpf_ca_needs_ecn(sk))
322 /* Packet ECN state for a SYN. */
323 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
325 struct tcp_sock *tp = tcp_sk(sk);
326 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
327 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
328 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
331 const struct dst_entry *dst = __sk_dst_get(sk);
333 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
340 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
341 tp->ecn_flags = TCP_ECN_OK;
342 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
347 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
349 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
350 /* tp->ecn_flags are cleared at a later point in time when
351 * SYN ACK is ultimatively being received.
353 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
357 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
359 if (inet_rsk(req)->ecn_ok)
363 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
366 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
367 struct tcphdr *th, int tcp_header_len)
369 struct tcp_sock *tp = tcp_sk(sk);
371 if (tp->ecn_flags & TCP_ECN_OK) {
372 /* Not-retransmitted data segment: set ECT and inject CWR. */
373 if (skb->len != tcp_header_len &&
374 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
376 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
377 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
379 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
381 } else if (!tcp_ca_needs_ecn(sk)) {
382 /* ACK or retransmitted segment: clear ECT|CE */
383 INET_ECN_dontxmit(sk);
385 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
390 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
391 * auto increment end seqno.
393 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
395 skb->ip_summed = CHECKSUM_PARTIAL;
397 TCP_SKB_CB(skb)->tcp_flags = flags;
398 TCP_SKB_CB(skb)->sacked = 0;
400 tcp_skb_pcount_set(skb, 1);
402 TCP_SKB_CB(skb)->seq = seq;
403 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
405 TCP_SKB_CB(skb)->end_seq = seq;
408 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
410 return tp->snd_una != tp->snd_up;
413 #define OPTION_SACK_ADVERTISE (1 << 0)
414 #define OPTION_TS (1 << 1)
415 #define OPTION_MD5 (1 << 2)
416 #define OPTION_WSCALE (1 << 3)
417 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
418 #define OPTION_SMC (1 << 9)
419 #define OPTION_MPTCP (1 << 10)
421 static void smc_options_write(__be32 *ptr, u16 *options)
423 #if IS_ENABLED(CONFIG_SMC)
424 if (static_branch_unlikely(&tcp_have_smc)) {
425 if (unlikely(OPTION_SMC & *options)) {
426 *ptr++ = htonl((TCPOPT_NOP << 24) |
429 (TCPOLEN_EXP_SMC_BASE));
430 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
436 struct tcp_out_options {
437 u16 options; /* bit field of OPTION_* */
438 u16 mss; /* 0 to disable */
439 u8 ws; /* window scale, 0 to disable */
440 u8 num_sack_blocks; /* number of SACK blocks to include */
441 u8 hash_size; /* bytes in hash_location */
442 u8 bpf_opt_len; /* length of BPF hdr option */
443 __u8 *hash_location; /* temporary pointer, overloaded */
444 __u32 tsval, tsecr; /* need to include OPTION_TS */
445 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
446 struct mptcp_out_options mptcp;
449 static void mptcp_options_write(__be32 *ptr, const struct tcp_sock *tp,
450 struct tcp_out_options *opts)
452 #if IS_ENABLED(CONFIG_MPTCP)
453 if (unlikely(OPTION_MPTCP & opts->options))
454 mptcp_write_options(ptr, tp, &opts->mptcp);
458 #ifdef CONFIG_CGROUP_BPF
459 static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
460 enum tcp_synack_type synack_type)
463 return BPF_WRITE_HDR_TCP_CURRENT_MSS;
465 if (unlikely(synack_type == TCP_SYNACK_COOKIE))
466 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
471 /* req, syn_skb and synack_type are used when writing synack */
472 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
473 struct request_sock *req,
474 struct sk_buff *syn_skb,
475 enum tcp_synack_type synack_type,
476 struct tcp_out_options *opts,
477 unsigned int *remaining)
479 struct bpf_sock_ops_kern sock_ops;
482 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
483 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
487 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
490 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
492 sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
495 /* The listen "sk" cannot be passed here because
496 * it is not locked. It would not make too much
497 * sense to do bpf_setsockopt(listen_sk) based
498 * on individual connection request also.
500 * Thus, "req" is passed here and the cgroup-bpf-progs
501 * of the listen "sk" will be run.
503 * "req" is also used here for fastopen even the "sk" here is
504 * a fullsock "child" sk. It is to keep the behavior
505 * consistent between fastopen and non-fastopen on
506 * the bpf programming side.
508 sock_ops.sk = (struct sock *)req;
509 sock_ops.syn_skb = syn_skb;
511 sock_owned_by_me(sk);
513 sock_ops.is_fullsock = 1;
517 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
518 sock_ops.remaining_opt_len = *remaining;
519 /* tcp_current_mss() does not pass a skb */
521 bpf_skops_init_skb(&sock_ops, skb, 0);
523 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
525 if (err || sock_ops.remaining_opt_len == *remaining)
528 opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
529 /* round up to 4 bytes */
530 opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
532 *remaining -= opts->bpf_opt_len;
535 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
536 struct request_sock *req,
537 struct sk_buff *syn_skb,
538 enum tcp_synack_type synack_type,
539 struct tcp_out_options *opts)
541 u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
542 struct bpf_sock_ops_kern sock_ops;
545 if (likely(!max_opt_len))
548 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
550 sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
553 sock_ops.sk = (struct sock *)req;
554 sock_ops.syn_skb = syn_skb;
556 sock_owned_by_me(sk);
558 sock_ops.is_fullsock = 1;
562 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
563 sock_ops.remaining_opt_len = max_opt_len;
564 first_opt_off = tcp_hdrlen(skb) - max_opt_len;
565 bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
567 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
572 nr_written = max_opt_len - sock_ops.remaining_opt_len;
574 if (nr_written < max_opt_len)
575 memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
576 max_opt_len - nr_written);
579 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
580 struct request_sock *req,
581 struct sk_buff *syn_skb,
582 enum tcp_synack_type synack_type,
583 struct tcp_out_options *opts,
584 unsigned int *remaining)
588 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
589 struct request_sock *req,
590 struct sk_buff *syn_skb,
591 enum tcp_synack_type synack_type,
592 struct tcp_out_options *opts)
597 /* Write previously computed TCP options to the packet.
599 * Beware: Something in the Internet is very sensitive to the ordering of
600 * TCP options, we learned this through the hard way, so be careful here.
601 * Luckily we can at least blame others for their non-compliance but from
602 * inter-operability perspective it seems that we're somewhat stuck with
603 * the ordering which we have been using if we want to keep working with
604 * those broken things (not that it currently hurts anybody as there isn't
605 * particular reason why the ordering would need to be changed).
607 * At least SACK_PERM as the first option is known to lead to a disaster
608 * (but it may well be that other scenarios fail similarly).
610 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
611 struct tcp_out_options *opts)
613 u16 options = opts->options; /* mungable copy */
615 if (unlikely(OPTION_MD5 & options)) {
616 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
617 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
618 /* overload cookie hash location */
619 opts->hash_location = (__u8 *)ptr;
623 if (unlikely(opts->mss)) {
624 *ptr++ = htonl((TCPOPT_MSS << 24) |
625 (TCPOLEN_MSS << 16) |
629 if (likely(OPTION_TS & options)) {
630 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
631 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
632 (TCPOLEN_SACK_PERM << 16) |
633 (TCPOPT_TIMESTAMP << 8) |
635 options &= ~OPTION_SACK_ADVERTISE;
637 *ptr++ = htonl((TCPOPT_NOP << 24) |
639 (TCPOPT_TIMESTAMP << 8) |
642 *ptr++ = htonl(opts->tsval);
643 *ptr++ = htonl(opts->tsecr);
646 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
647 *ptr++ = htonl((TCPOPT_NOP << 24) |
649 (TCPOPT_SACK_PERM << 8) |
653 if (unlikely(OPTION_WSCALE & options)) {
654 *ptr++ = htonl((TCPOPT_NOP << 24) |
655 (TCPOPT_WINDOW << 16) |
656 (TCPOLEN_WINDOW << 8) |
660 if (unlikely(opts->num_sack_blocks)) {
661 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
662 tp->duplicate_sack : tp->selective_acks;
665 *ptr++ = htonl((TCPOPT_NOP << 24) |
668 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
669 TCPOLEN_SACK_PERBLOCK)));
671 for (this_sack = 0; this_sack < opts->num_sack_blocks;
673 *ptr++ = htonl(sp[this_sack].start_seq);
674 *ptr++ = htonl(sp[this_sack].end_seq);
677 tp->rx_opt.dsack = 0;
680 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
681 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
683 u32 len; /* Fast Open option length */
686 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
687 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
688 TCPOPT_FASTOPEN_MAGIC);
689 p += TCPOLEN_EXP_FASTOPEN_BASE;
691 len = TCPOLEN_FASTOPEN_BASE + foc->len;
692 *p++ = TCPOPT_FASTOPEN;
696 memcpy(p, foc->val, foc->len);
697 if ((len & 3) == 2) {
698 p[foc->len] = TCPOPT_NOP;
699 p[foc->len + 1] = TCPOPT_NOP;
701 ptr += (len + 3) >> 2;
704 smc_options_write(ptr, &options);
706 mptcp_options_write(ptr, tp, opts);
709 static void smc_set_option(const struct tcp_sock *tp,
710 struct tcp_out_options *opts,
711 unsigned int *remaining)
713 #if IS_ENABLED(CONFIG_SMC)
714 if (static_branch_unlikely(&tcp_have_smc)) {
716 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
717 opts->options |= OPTION_SMC;
718 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
725 static void smc_set_option_cond(const struct tcp_sock *tp,
726 const struct inet_request_sock *ireq,
727 struct tcp_out_options *opts,
728 unsigned int *remaining)
730 #if IS_ENABLED(CONFIG_SMC)
731 if (static_branch_unlikely(&tcp_have_smc)) {
732 if (tp->syn_smc && ireq->smc_ok) {
733 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
734 opts->options |= OPTION_SMC;
735 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
742 static void mptcp_set_option_cond(const struct request_sock *req,
743 struct tcp_out_options *opts,
744 unsigned int *remaining)
746 if (rsk_is_mptcp(req)) {
749 if (mptcp_synack_options(req, &size, &opts->mptcp)) {
750 if (*remaining >= size) {
751 opts->options |= OPTION_MPTCP;
758 /* Compute TCP options for SYN packets. This is not the final
759 * network wire format yet.
761 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
762 struct tcp_out_options *opts,
763 struct tcp_md5sig_key **md5)
765 struct tcp_sock *tp = tcp_sk(sk);
766 unsigned int remaining = MAX_TCP_OPTION_SPACE;
767 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
770 #ifdef CONFIG_TCP_MD5SIG
771 if (static_branch_unlikely(&tcp_md5_needed) &&
772 rcu_access_pointer(tp->md5sig_info)) {
773 *md5 = tp->af_specific->md5_lookup(sk, sk);
775 opts->options |= OPTION_MD5;
776 remaining -= TCPOLEN_MD5SIG_ALIGNED;
781 /* We always get an MSS option. The option bytes which will be seen in
782 * normal data packets should timestamps be used, must be in the MSS
783 * advertised. But we subtract them from tp->mss_cache so that
784 * calculations in tcp_sendmsg are simpler etc. So account for this
785 * fact here if necessary. If we don't do this correctly, as a
786 * receiver we won't recognize data packets as being full sized when we
787 * should, and thus we won't abide by the delayed ACK rules correctly.
788 * SACKs don't matter, we never delay an ACK when we have any of those
790 opts->mss = tcp_advertise_mss(sk);
791 remaining -= TCPOLEN_MSS_ALIGNED;
793 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
794 opts->options |= OPTION_TS;
795 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
796 opts->tsecr = tp->rx_opt.ts_recent;
797 remaining -= TCPOLEN_TSTAMP_ALIGNED;
799 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
800 opts->ws = tp->rx_opt.rcv_wscale;
801 opts->options |= OPTION_WSCALE;
802 remaining -= TCPOLEN_WSCALE_ALIGNED;
804 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
805 opts->options |= OPTION_SACK_ADVERTISE;
806 if (unlikely(!(OPTION_TS & opts->options)))
807 remaining -= TCPOLEN_SACKPERM_ALIGNED;
810 if (fastopen && fastopen->cookie.len >= 0) {
811 u32 need = fastopen->cookie.len;
813 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
814 TCPOLEN_FASTOPEN_BASE;
815 need = (need + 3) & ~3U; /* Align to 32 bits */
816 if (remaining >= need) {
817 opts->options |= OPTION_FAST_OPEN_COOKIE;
818 opts->fastopen_cookie = &fastopen->cookie;
820 tp->syn_fastopen = 1;
821 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
825 smc_set_option(tp, opts, &remaining);
827 if (sk_is_mptcp(sk)) {
830 if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
831 opts->options |= OPTION_MPTCP;
836 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
838 return MAX_TCP_OPTION_SPACE - remaining;
841 /* Set up TCP options for SYN-ACKs. */
842 static unsigned int tcp_synack_options(const struct sock *sk,
843 struct request_sock *req,
844 unsigned int mss, struct sk_buff *skb,
845 struct tcp_out_options *opts,
846 const struct tcp_md5sig_key *md5,
847 struct tcp_fastopen_cookie *foc,
848 enum tcp_synack_type synack_type,
849 struct sk_buff *syn_skb)
851 struct inet_request_sock *ireq = inet_rsk(req);
852 unsigned int remaining = MAX_TCP_OPTION_SPACE;
854 #ifdef CONFIG_TCP_MD5SIG
856 opts->options |= OPTION_MD5;
857 remaining -= TCPOLEN_MD5SIG_ALIGNED;
859 /* We can't fit any SACK blocks in a packet with MD5 + TS
860 * options. There was discussion about disabling SACK
861 * rather than TS in order to fit in better with old,
862 * buggy kernels, but that was deemed to be unnecessary.
864 if (synack_type != TCP_SYNACK_COOKIE)
865 ireq->tstamp_ok &= !ireq->sack_ok;
869 /* We always send an MSS option. */
871 remaining -= TCPOLEN_MSS_ALIGNED;
873 if (likely(ireq->wscale_ok)) {
874 opts->ws = ireq->rcv_wscale;
875 opts->options |= OPTION_WSCALE;
876 remaining -= TCPOLEN_WSCALE_ALIGNED;
878 if (likely(ireq->tstamp_ok)) {
879 opts->options |= OPTION_TS;
880 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
881 opts->tsecr = req->ts_recent;
882 remaining -= TCPOLEN_TSTAMP_ALIGNED;
884 if (likely(ireq->sack_ok)) {
885 opts->options |= OPTION_SACK_ADVERTISE;
886 if (unlikely(!ireq->tstamp_ok))
887 remaining -= TCPOLEN_SACKPERM_ALIGNED;
889 if (foc != NULL && foc->len >= 0) {
892 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
893 TCPOLEN_FASTOPEN_BASE;
894 need = (need + 3) & ~3U; /* Align to 32 bits */
895 if (remaining >= need) {
896 opts->options |= OPTION_FAST_OPEN_COOKIE;
897 opts->fastopen_cookie = foc;
902 mptcp_set_option_cond(req, opts, &remaining);
904 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
906 bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb,
907 synack_type, opts, &remaining);
909 return MAX_TCP_OPTION_SPACE - remaining;
912 /* Compute TCP options for ESTABLISHED sockets. This is not the
913 * final wire format yet.
915 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
916 struct tcp_out_options *opts,
917 struct tcp_md5sig_key **md5)
919 struct tcp_sock *tp = tcp_sk(sk);
920 unsigned int size = 0;
921 unsigned int eff_sacks;
926 #ifdef CONFIG_TCP_MD5SIG
927 if (static_branch_unlikely(&tcp_md5_needed) &&
928 rcu_access_pointer(tp->md5sig_info)) {
929 *md5 = tp->af_specific->md5_lookup(sk, sk);
931 opts->options |= OPTION_MD5;
932 size += TCPOLEN_MD5SIG_ALIGNED;
937 if (likely(tp->rx_opt.tstamp_ok)) {
938 opts->options |= OPTION_TS;
939 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
940 opts->tsecr = tp->rx_opt.ts_recent;
941 size += TCPOLEN_TSTAMP_ALIGNED;
944 /* MPTCP options have precedence over SACK for the limited TCP
945 * option space because a MPTCP connection would be forced to
946 * fall back to regular TCP if a required multipath option is
947 * missing. SACK still gets a chance to use whatever space is
950 if (sk_is_mptcp(sk)) {
951 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
952 unsigned int opt_size = 0;
954 if (mptcp_established_options(sk, skb, &opt_size, remaining,
956 opts->options |= OPTION_MPTCP;
961 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
962 if (unlikely(eff_sacks)) {
963 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
964 if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
965 TCPOLEN_SACK_PERBLOCK))
968 opts->num_sack_blocks =
969 min_t(unsigned int, eff_sacks,
970 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
971 TCPOLEN_SACK_PERBLOCK);
973 size += TCPOLEN_SACK_BASE_ALIGNED +
974 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
977 if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
978 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
979 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
981 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
983 size = MAX_TCP_OPTION_SPACE - remaining;
990 /* TCP SMALL QUEUES (TSQ)
992 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
993 * to reduce RTT and bufferbloat.
994 * We do this using a special skb destructor (tcp_wfree).
996 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
997 * needs to be reallocated in a driver.
998 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
1000 * Since transmit from skb destructor is forbidden, we use a tasklet
1001 * to process all sockets that eventually need to send more skbs.
1002 * We use one tasklet per cpu, with its own queue of sockets.
1004 struct tsq_tasklet {
1005 struct tasklet_struct tasklet;
1006 struct list_head head; /* queue of tcp sockets */
1008 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
1010 static void tcp_tsq_write(struct sock *sk)
1012 if ((1 << sk->sk_state) &
1013 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1014 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
1015 struct tcp_sock *tp = tcp_sk(sk);
1017 if (tp->lost_out > tp->retrans_out &&
1018 tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) {
1019 tcp_mstamp_refresh(tp);
1020 tcp_xmit_retransmit_queue(sk);
1023 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
1028 static void tcp_tsq_handler(struct sock *sk)
1031 if (!sock_owned_by_user(sk))
1033 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
1038 * One tasklet per cpu tries to send more skbs.
1039 * We run in tasklet context but need to disable irqs when
1040 * transferring tsq->head because tcp_wfree() might
1041 * interrupt us (non NAPI drivers)
1043 static void tcp_tasklet_func(struct tasklet_struct *t)
1045 struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet);
1047 unsigned long flags;
1048 struct list_head *q, *n;
1049 struct tcp_sock *tp;
1052 local_irq_save(flags);
1053 list_splice_init(&tsq->head, &list);
1054 local_irq_restore(flags);
1056 list_for_each_safe(q, n, &list) {
1057 tp = list_entry(q, struct tcp_sock, tsq_node);
1058 list_del(&tp->tsq_node);
1060 sk = (struct sock *)tp;
1061 smp_mb__before_atomic();
1062 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
1064 tcp_tsq_handler(sk);
1069 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
1070 TCPF_WRITE_TIMER_DEFERRED | \
1071 TCPF_DELACK_TIMER_DEFERRED | \
1072 TCPF_MTU_REDUCED_DEFERRED)
1074 * tcp_release_cb - tcp release_sock() callback
1077 * called from release_sock() to perform protocol dependent
1078 * actions before socket release.
1080 void tcp_release_cb(struct sock *sk)
1082 unsigned long flags, nflags;
1084 /* perform an atomic operation only if at least one flag is set */
1086 flags = sk->sk_tsq_flags;
1087 if (!(flags & TCP_DEFERRED_ALL))
1089 nflags = flags & ~TCP_DEFERRED_ALL;
1090 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
1092 if (flags & TCPF_TSQ_DEFERRED) {
1096 /* Here begins the tricky part :
1097 * We are called from release_sock() with :
1099 * 2) sk_lock.slock spinlock held
1100 * 3) socket owned by us (sk->sk_lock.owned == 1)
1102 * But following code is meant to be called from BH handlers,
1103 * so we should keep BH disabled, but early release socket ownership
1105 sock_release_ownership(sk);
1107 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1108 tcp_write_timer_handler(sk);
1111 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1112 tcp_delack_timer_handler(sk);
1115 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1116 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1120 EXPORT_SYMBOL(tcp_release_cb);
1122 void __init tcp_tasklet_init(void)
1126 for_each_possible_cpu(i) {
1127 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
1129 INIT_LIST_HEAD(&tsq->head);
1130 tasklet_setup(&tsq->tasklet, tcp_tasklet_func);
1135 * Write buffer destructor automatically called from kfree_skb.
1136 * We can't xmit new skbs from this context, as we might already
1139 void tcp_wfree(struct sk_buff *skb)
1141 struct sock *sk = skb->sk;
1142 struct tcp_sock *tp = tcp_sk(sk);
1143 unsigned long flags, nval, oval;
1145 /* Keep one reference on sk_wmem_alloc.
1146 * Will be released by sk_free() from here or tcp_tasklet_func()
1148 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1150 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
1151 * Wait until our queues (qdisc + devices) are drained.
1153 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1154 * - chance for incoming ACK (processed by another cpu maybe)
1155 * to migrate this flow (skb->ooo_okay will be eventually set)
1157 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1160 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
1161 struct tsq_tasklet *tsq;
1164 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1167 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1168 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
1172 /* queue this socket to tasklet queue */
1173 local_irq_save(flags);
1174 tsq = this_cpu_ptr(&tsq_tasklet);
1175 empty = list_empty(&tsq->head);
1176 list_add(&tp->tsq_node, &tsq->head);
1178 tasklet_schedule(&tsq->tasklet);
1179 local_irq_restore(flags);
1186 /* Note: Called under soft irq.
1187 * We can call TCP stack right away, unless socket is owned by user.
1189 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1191 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1192 struct sock *sk = (struct sock *)tp;
1194 tcp_tsq_handler(sk);
1197 return HRTIMER_NORESTART;
1200 static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1203 struct tcp_sock *tp = tcp_sk(sk);
1205 if (sk->sk_pacing_status != SK_PACING_NONE) {
1206 unsigned long rate = sk->sk_pacing_rate;
1208 /* Original sch_fq does not pace first 10 MSS
1209 * Note that tp->data_segs_out overflows after 2^32 packets,
1210 * this is a minor annoyance.
1212 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1213 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1214 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1216 /* take into account OS jitter */
1217 len_ns -= min_t(u64, len_ns / 2, credit);
1218 tp->tcp_wstamp_ns += len_ns;
1221 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1224 INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1225 INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1226 INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1228 /* This routine actually transmits TCP packets queued in by
1229 * tcp_do_sendmsg(). This is used by both the initial
1230 * transmission and possible later retransmissions.
1231 * All SKB's seen here are completely headerless. It is our
1232 * job to build the TCP header, and pass the packet down to
1233 * IP so it can do the same plus pass the packet off to the
1236 * We are working here with either a clone of the original
1237 * SKB, or a fresh unique copy made by the retransmit engine.
1239 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1240 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1242 const struct inet_connection_sock *icsk = inet_csk(sk);
1243 struct inet_sock *inet;
1244 struct tcp_sock *tp;
1245 struct tcp_skb_cb *tcb;
1246 struct tcp_out_options opts;
1247 unsigned int tcp_options_size, tcp_header_size;
1248 struct sk_buff *oskb = NULL;
1249 struct tcp_md5sig_key *md5;
1254 BUG_ON(!skb || !tcp_skb_pcount(skb));
1256 prior_wstamp = tp->tcp_wstamp_ns;
1257 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1258 skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
1260 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1264 tcp_skb_tsorted_save(oskb) {
1265 if (unlikely(skb_cloned(oskb)))
1266 skb = pskb_copy(oskb, gfp_mask);
1268 skb = skb_clone(oskb, gfp_mask);
1269 } tcp_skb_tsorted_restore(oskb);
1273 /* retransmit skbs might have a non zero value in skb->dev
1274 * because skb->dev is aliased with skb->rbnode.rb_left
1280 tcb = TCP_SKB_CB(skb);
1281 memset(&opts, 0, sizeof(opts));
1283 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1284 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1286 tcp_options_size = tcp_established_options(sk, skb, &opts,
1288 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1289 * at receiver : This slightly improve GRO performance.
1290 * Note that we do not force the PSH flag for non GSO packets,
1291 * because they might be sent under high congestion events,
1292 * and in this case it is better to delay the delivery of 1-MSS
1293 * packets and thus the corresponding ACK packet that would
1294 * release the following packet.
1296 if (tcp_skb_pcount(skb) > 1)
1297 tcb->tcp_flags |= TCPHDR_PSH;
1299 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1301 /* if no packet is in qdisc/device queue, then allow XPS to select
1302 * another queue. We can be called from tcp_tsq_handler()
1303 * which holds one reference to sk.
1305 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1306 * One way to get this would be to set skb->truesize = 2 on them.
1308 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1310 /* If we had to use memory reserve to allocate this skb,
1311 * this might cause drops if packet is looped back :
1312 * Other socket might not have SOCK_MEMALLOC.
1313 * Packets not looped back do not care about pfmemalloc.
1315 skb->pfmemalloc = 0;
1317 skb_push(skb, tcp_header_size);
1318 skb_reset_transport_header(skb);
1322 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1323 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1325 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1327 /* Build TCP header and checksum it. */
1328 th = (struct tcphdr *)skb->data;
1329 th->source = inet->inet_sport;
1330 th->dest = inet->inet_dport;
1331 th->seq = htonl(tcb->seq);
1332 th->ack_seq = htonl(rcv_nxt);
1333 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1339 /* The urg_mode check is necessary during a below snd_una win probe */
1340 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1341 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1342 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1344 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1345 th->urg_ptr = htons(0xFFFF);
1350 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1351 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1352 th->window = htons(tcp_select_window(sk));
1353 tcp_ecn_send(sk, skb, th, tcp_header_size);
1355 /* RFC1323: The window in SYN & SYN/ACK segments
1358 th->window = htons(min(tp->rcv_wnd, 65535U));
1361 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1363 #ifdef CONFIG_TCP_MD5SIG
1364 /* Calculate the MD5 hash, as we have all we need now */
1366 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1367 tp->af_specific->calc_md5_hash(opts.hash_location,
1372 /* BPF prog is the last one writing header option */
1373 bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1375 INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1376 tcp_v6_send_check, tcp_v4_send_check,
1379 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1380 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1382 if (skb->len != tcp_header_size) {
1383 tcp_event_data_sent(tp, sk);
1384 tp->data_segs_out += tcp_skb_pcount(skb);
1385 tp->bytes_sent += skb->len - tcp_header_size;
1388 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1389 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1390 tcp_skb_pcount(skb));
1392 tp->segs_out += tcp_skb_pcount(skb);
1393 skb_set_hash_from_sk(skb, sk);
1394 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1395 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1396 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1398 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1400 /* Cleanup our debris for IP stacks */
1401 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1402 sizeof(struct inet6_skb_parm)));
1404 tcp_add_tx_delay(skb, tp);
1406 err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1407 inet6_csk_xmit, ip_queue_xmit,
1408 sk, skb, &inet->cork.fl);
1410 if (unlikely(err > 0)) {
1412 err = net_xmit_eval(err);
1415 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1416 tcp_rate_skb_sent(sk, oskb);
1421 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1424 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1425 tcp_sk(sk)->rcv_nxt);
1428 /* This routine just queues the buffer for sending.
1430 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1431 * otherwise socket can stall.
1433 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1435 struct tcp_sock *tp = tcp_sk(sk);
1437 /* Advance write_seq and place onto the write_queue. */
1438 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1439 __skb_header_release(skb);
1440 tcp_add_write_queue_tail(sk, skb);
1441 sk_wmem_queued_add(sk, skb->truesize);
1442 sk_mem_charge(sk, skb->truesize);
1445 /* Initialize TSO segments for a packet. */
1446 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1448 if (skb->len <= mss_now) {
1449 /* Avoid the costly divide in the normal
1452 tcp_skb_pcount_set(skb, 1);
1453 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1455 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1456 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1460 /* Pcount in the middle of the write queue got changed, we need to do various
1461 * tweaks to fix counters
1463 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1465 struct tcp_sock *tp = tcp_sk(sk);
1467 tp->packets_out -= decr;
1469 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1470 tp->sacked_out -= decr;
1471 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1472 tp->retrans_out -= decr;
1473 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1474 tp->lost_out -= decr;
1476 /* Reno case is special. Sigh... */
1477 if (tcp_is_reno(tp) && decr > 0)
1478 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1480 if (tp->lost_skb_hint &&
1481 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1482 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1483 tp->lost_cnt_hint -= decr;
1485 tcp_verify_left_out(tp);
1488 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1490 return TCP_SKB_CB(skb)->txstamp_ack ||
1491 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1494 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1496 struct skb_shared_info *shinfo = skb_shinfo(skb);
1498 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1499 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1500 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1501 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1503 shinfo->tx_flags &= ~tsflags;
1504 shinfo2->tx_flags |= tsflags;
1505 swap(shinfo->tskey, shinfo2->tskey);
1506 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1507 TCP_SKB_CB(skb)->txstamp_ack = 0;
1511 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1513 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1514 TCP_SKB_CB(skb)->eor = 0;
1517 /* Insert buff after skb on the write or rtx queue of sk. */
1518 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1519 struct sk_buff *buff,
1521 enum tcp_queue tcp_queue)
1523 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1524 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1526 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1529 /* Function to create two new TCP segments. Shrinks the given segment
1530 * to the specified size and appends a new segment with the rest of the
1531 * packet to the list. This won't be called frequently, I hope.
1532 * Remember, these are still headerless SKBs at this point.
1534 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1535 struct sk_buff *skb, u32 len,
1536 unsigned int mss_now, gfp_t gfp)
1538 struct tcp_sock *tp = tcp_sk(sk);
1539 struct sk_buff *buff;
1540 int nsize, old_factor;
1545 if (WARN_ON(len > skb->len))
1548 nsize = skb_headlen(skb) - len;
1552 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1553 * We need some allowance to not penalize applications setting small
1555 * Also allow first and last skb in retransmit queue to be split.
1557 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1558 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1559 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1560 skb != tcp_rtx_queue_head(sk) &&
1561 skb != tcp_rtx_queue_tail(sk))) {
1562 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1566 if (skb_unclone_keeptruesize(skb, gfp))
1569 /* Get a new skb... force flag on. */
1570 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1572 return -ENOMEM; /* We'll just try again later. */
1573 skb_copy_decrypted(buff, skb);
1574 mptcp_skb_ext_copy(buff, skb);
1576 sk_wmem_queued_add(sk, buff->truesize);
1577 sk_mem_charge(sk, buff->truesize);
1578 nlen = skb->len - len - nsize;
1579 buff->truesize += nlen;
1580 skb->truesize -= nlen;
1582 /* Correct the sequence numbers. */
1583 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1584 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1585 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1587 /* PSH and FIN should only be set in the second packet. */
1588 flags = TCP_SKB_CB(skb)->tcp_flags;
1589 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1590 TCP_SKB_CB(buff)->tcp_flags = flags;
1591 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1592 tcp_skb_fragment_eor(skb, buff);
1594 skb_split(skb, buff, len);
1596 buff->ip_summed = CHECKSUM_PARTIAL;
1598 buff->tstamp = skb->tstamp;
1599 tcp_fragment_tstamp(skb, buff);
1601 old_factor = tcp_skb_pcount(skb);
1603 /* Fix up tso_factor for both original and new SKB. */
1604 tcp_set_skb_tso_segs(skb, mss_now);
1605 tcp_set_skb_tso_segs(buff, mss_now);
1607 /* Update delivered info for the new segment */
1608 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1610 /* If this packet has been sent out already, we must
1611 * adjust the various packet counters.
1613 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1614 int diff = old_factor - tcp_skb_pcount(skb) -
1615 tcp_skb_pcount(buff);
1618 tcp_adjust_pcount(sk, skb, diff);
1621 /* Link BUFF into the send queue. */
1622 __skb_header_release(buff);
1623 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1624 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1625 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1630 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1631 * data is not copied, but immediately discarded.
1633 static int __pskb_trim_head(struct sk_buff *skb, int len)
1635 struct skb_shared_info *shinfo;
1638 eat = min_t(int, len, skb_headlen(skb));
1640 __skb_pull(skb, eat);
1647 shinfo = skb_shinfo(skb);
1648 for (i = 0; i < shinfo->nr_frags; i++) {
1649 int size = skb_frag_size(&shinfo->frags[i]);
1652 skb_frag_unref(skb, i);
1655 shinfo->frags[k] = shinfo->frags[i];
1657 skb_frag_off_add(&shinfo->frags[k], eat);
1658 skb_frag_size_sub(&shinfo->frags[k], eat);
1664 shinfo->nr_frags = k;
1666 skb->data_len -= len;
1667 skb->len = skb->data_len;
1671 /* Remove acked data from a packet in the transmit queue. */
1672 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1676 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
1679 delta_truesize = __pskb_trim_head(skb, len);
1681 TCP_SKB_CB(skb)->seq += len;
1682 skb->ip_summed = CHECKSUM_PARTIAL;
1684 if (delta_truesize) {
1685 skb->truesize -= delta_truesize;
1686 sk_wmem_queued_add(sk, -delta_truesize);
1687 sk_mem_uncharge(sk, delta_truesize);
1690 /* Any change of skb->len requires recalculation of tso factor. */
1691 if (tcp_skb_pcount(skb) > 1)
1692 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1697 /* Calculate MSS not accounting any TCP options. */
1698 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1700 const struct tcp_sock *tp = tcp_sk(sk);
1701 const struct inet_connection_sock *icsk = inet_csk(sk);
1704 /* Calculate base mss without TCP options:
1705 It is MMS_S - sizeof(tcphdr) of rfc1122
1707 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1709 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1710 if (icsk->icsk_af_ops->net_frag_header_len) {
1711 const struct dst_entry *dst = __sk_dst_get(sk);
1713 if (dst && dst_allfrag(dst))
1714 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1717 /* Clamp it (mss_clamp does not include tcp options) */
1718 if (mss_now > tp->rx_opt.mss_clamp)
1719 mss_now = tp->rx_opt.mss_clamp;
1721 /* Now subtract optional transport overhead */
1722 mss_now -= icsk->icsk_ext_hdr_len;
1724 /* Then reserve room for full set of TCP options and 8 bytes of data */
1725 mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1729 /* Calculate MSS. Not accounting for SACKs here. */
1730 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1732 /* Subtract TCP options size, not including SACKs */
1733 return __tcp_mtu_to_mss(sk, pmtu) -
1734 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1736 EXPORT_SYMBOL(tcp_mtu_to_mss);
1738 /* Inverse of above */
1739 int tcp_mss_to_mtu(struct sock *sk, int mss)
1741 const struct tcp_sock *tp = tcp_sk(sk);
1742 const struct inet_connection_sock *icsk = inet_csk(sk);
1746 tp->tcp_header_len +
1747 icsk->icsk_ext_hdr_len +
1748 icsk->icsk_af_ops->net_header_len;
1750 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1751 if (icsk->icsk_af_ops->net_frag_header_len) {
1752 const struct dst_entry *dst = __sk_dst_get(sk);
1754 if (dst && dst_allfrag(dst))
1755 mtu += icsk->icsk_af_ops->net_frag_header_len;
1759 EXPORT_SYMBOL(tcp_mss_to_mtu);
1761 /* MTU probing init per socket */
1762 void tcp_mtup_init(struct sock *sk)
1764 struct tcp_sock *tp = tcp_sk(sk);
1765 struct inet_connection_sock *icsk = inet_csk(sk);
1766 struct net *net = sock_net(sk);
1768 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1769 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1770 icsk->icsk_af_ops->net_header_len;
1771 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1772 icsk->icsk_mtup.probe_size = 0;
1773 if (icsk->icsk_mtup.enabled)
1774 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1776 EXPORT_SYMBOL(tcp_mtup_init);
1778 /* This function synchronize snd mss to current pmtu/exthdr set.
1780 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1781 for TCP options, but includes only bare TCP header.
1783 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1784 It is minimum of user_mss and mss received with SYN.
1785 It also does not include TCP options.
1787 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1789 tp->mss_cache is current effective sending mss, including
1790 all tcp options except for SACKs. It is evaluated,
1791 taking into account current pmtu, but never exceeds
1792 tp->rx_opt.mss_clamp.
1794 NOTE1. rfc1122 clearly states that advertised MSS
1795 DOES NOT include either tcp or ip options.
1797 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1798 are READ ONLY outside this function. --ANK (980731)
1800 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1802 struct tcp_sock *tp = tcp_sk(sk);
1803 struct inet_connection_sock *icsk = inet_csk(sk);
1806 if (icsk->icsk_mtup.search_high > pmtu)
1807 icsk->icsk_mtup.search_high = pmtu;
1809 mss_now = tcp_mtu_to_mss(sk, pmtu);
1810 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1812 /* And store cached results */
1813 icsk->icsk_pmtu_cookie = pmtu;
1814 if (icsk->icsk_mtup.enabled)
1815 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1816 tp->mss_cache = mss_now;
1820 EXPORT_SYMBOL(tcp_sync_mss);
1822 /* Compute the current effective MSS, taking SACKs and IP options,
1823 * and even PMTU discovery events into account.
1825 unsigned int tcp_current_mss(struct sock *sk)
1827 const struct tcp_sock *tp = tcp_sk(sk);
1828 const struct dst_entry *dst = __sk_dst_get(sk);
1830 unsigned int header_len;
1831 struct tcp_out_options opts;
1832 struct tcp_md5sig_key *md5;
1834 mss_now = tp->mss_cache;
1837 u32 mtu = dst_mtu(dst);
1838 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1839 mss_now = tcp_sync_mss(sk, mtu);
1842 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1843 sizeof(struct tcphdr);
1844 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1845 * some common options. If this is an odd packet (because we have SACK
1846 * blocks etc) then our calculated header_len will be different, and
1847 * we have to adjust mss_now correspondingly */
1848 if (header_len != tp->tcp_header_len) {
1849 int delta = (int) header_len - tp->tcp_header_len;
1856 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1857 * As additional protections, we do not touch cwnd in retransmission phases,
1858 * and if application hit its sndbuf limit recently.
1860 static void tcp_cwnd_application_limited(struct sock *sk)
1862 struct tcp_sock *tp = tcp_sk(sk);
1864 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1865 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1866 /* Limited by application or receiver window. */
1867 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1868 u32 win_used = max(tp->snd_cwnd_used, init_win);
1869 if (win_used < tcp_snd_cwnd(tp)) {
1870 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1871 tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1);
1873 tp->snd_cwnd_used = 0;
1875 tp->snd_cwnd_stamp = tcp_jiffies32;
1878 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1880 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1881 struct tcp_sock *tp = tcp_sk(sk);
1883 /* Track the maximum number of outstanding packets in each
1884 * window, and remember whether we were cwnd-limited then.
1886 if (!before(tp->snd_una, tp->max_packets_seq) ||
1887 tp->packets_out > tp->max_packets_out ||
1889 tp->max_packets_out = tp->packets_out;
1890 tp->max_packets_seq = tp->snd_nxt;
1891 tp->is_cwnd_limited = is_cwnd_limited;
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 (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 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 = tcp_snd_cwnd(tp);
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);
2131 mptcp_skb_ext_copy(buff, skb);
2133 sk_wmem_queued_add(sk, buff->truesize);
2134 sk_mem_charge(sk, buff->truesize);
2135 buff->truesize += nlen;
2136 skb->truesize -= nlen;
2138 /* Correct the sequence numbers. */
2139 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2140 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2141 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2143 /* PSH and FIN should only be set in the second packet. */
2144 flags = TCP_SKB_CB(skb)->tcp_flags;
2145 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2146 TCP_SKB_CB(buff)->tcp_flags = flags;
2148 /* This packet was never sent out yet, so no SACK bits. */
2149 TCP_SKB_CB(buff)->sacked = 0;
2151 tcp_skb_fragment_eor(skb, buff);
2153 buff->ip_summed = CHECKSUM_PARTIAL;
2154 skb_split(skb, buff, len);
2155 tcp_fragment_tstamp(skb, buff);
2157 /* Fix up tso_factor for both original and new SKB. */
2158 tcp_set_skb_tso_segs(skb, mss_now);
2159 tcp_set_skb_tso_segs(buff, mss_now);
2161 /* Link BUFF into the send queue. */
2162 __skb_header_release(buff);
2163 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2168 /* Try to defer sending, if possible, in order to minimize the amount
2169 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2171 * This algorithm is from John Heffner.
2173 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2174 bool *is_cwnd_limited,
2175 bool *is_rwnd_limited,
2178 const struct inet_connection_sock *icsk = inet_csk(sk);
2179 u32 send_win, cong_win, limit, in_flight;
2180 struct tcp_sock *tp = tcp_sk(sk);
2181 struct sk_buff *head;
2185 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2188 /* Avoid bursty behavior by allowing defer
2189 * only if the last write was recent (1 ms).
2190 * Note that tp->tcp_wstamp_ns can be in the future if we have
2191 * packets waiting in a qdisc or device for EDT delivery.
2193 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2197 in_flight = tcp_packets_in_flight(tp);
2199 BUG_ON(tcp_skb_pcount(skb) <= 1);
2200 BUG_ON(tcp_snd_cwnd(tp) <= in_flight);
2202 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2204 /* From in_flight test above, we know that cwnd > in_flight. */
2205 cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache;
2207 limit = min(send_win, cong_win);
2209 /* If a full-sized TSO skb can be sent, do it. */
2210 if (limit >= max_segs * tp->mss_cache)
2213 /* Middle in queue won't get any more data, full sendable already? */
2214 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2217 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2219 u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache);
2221 /* If at least some fraction of a window is available,
2224 chunk /= win_divisor;
2228 /* Different approach, try not to defer past a single
2229 * ACK. Receiver should ACK every other full sized
2230 * frame, so if we have space for more than 3 frames
2233 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2237 /* TODO : use tsorted_sent_queue ? */
2238 head = tcp_rtx_queue_head(sk);
2241 delta = tp->tcp_clock_cache - head->tstamp;
2242 /* If next ACK is likely to come too late (half srtt), do not defer */
2243 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2246 /* Ok, it looks like it is advisable to defer.
2247 * Three cases are tracked :
2248 * 1) We are cwnd-limited
2249 * 2) We are rwnd-limited
2250 * 3) We are application limited.
2252 if (cong_win < send_win) {
2253 if (cong_win <= skb->len) {
2254 *is_cwnd_limited = true;
2258 if (send_win <= skb->len) {
2259 *is_rwnd_limited = true;
2264 /* If this packet won't get more data, do not wait. */
2265 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2266 TCP_SKB_CB(skb)->eor)
2275 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2277 struct inet_connection_sock *icsk = inet_csk(sk);
2278 struct tcp_sock *tp = tcp_sk(sk);
2279 struct net *net = sock_net(sk);
2283 interval = net->ipv4.sysctl_tcp_probe_interval;
2284 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2285 if (unlikely(delta >= interval * HZ)) {
2286 int mss = tcp_current_mss(sk);
2288 /* Update current search range */
2289 icsk->icsk_mtup.probe_size = 0;
2290 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2291 sizeof(struct tcphdr) +
2292 icsk->icsk_af_ops->net_header_len;
2293 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2295 /* Update probe time stamp */
2296 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2300 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2302 struct sk_buff *skb, *next;
2304 skb = tcp_send_head(sk);
2305 tcp_for_write_queue_from_safe(skb, next, sk) {
2306 if (len <= skb->len)
2309 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2318 /* Create a new MTU probe if we are ready.
2319 * MTU probe is regularly attempting to increase the path MTU by
2320 * deliberately sending larger packets. This discovers routing
2321 * changes resulting in larger path MTUs.
2323 * Returns 0 if we should wait to probe (no cwnd available),
2324 * 1 if a probe was sent,
2327 static int tcp_mtu_probe(struct sock *sk)
2329 struct inet_connection_sock *icsk = inet_csk(sk);
2330 struct tcp_sock *tp = tcp_sk(sk);
2331 struct sk_buff *skb, *nskb, *next;
2332 struct net *net = sock_net(sk);
2339 /* Not currently probing/verifying,
2341 * have enough cwnd, and
2342 * not SACKing (the variable headers throw things off)
2344 if (likely(!icsk->icsk_mtup.enabled ||
2345 icsk->icsk_mtup.probe_size ||
2346 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2347 tcp_snd_cwnd(tp) < 11 ||
2348 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2351 /* Use binary search for probe_size between tcp_mss_base,
2352 * and current mss_clamp. if (search_high - search_low)
2353 * smaller than a threshold, backoff from probing.
2355 mss_now = tcp_current_mss(sk);
2356 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2357 icsk->icsk_mtup.search_low) >> 1);
2358 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2359 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2360 /* When misfortune happens, we are reprobing actively,
2361 * and then reprobe timer has expired. We stick with current
2362 * probing process by not resetting search range to its orignal.
2364 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2365 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2366 /* Check whether enough time has elaplased for
2367 * another round of probing.
2369 tcp_mtu_check_reprobe(sk);
2373 /* Have enough data in the send queue to probe? */
2374 if (tp->write_seq - tp->snd_nxt < size_needed)
2377 if (tp->snd_wnd < size_needed)
2379 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2382 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2383 if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) {
2384 if (!tcp_packets_in_flight(tp))
2390 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2393 /* We're allowed to probe. Build it now. */
2394 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2397 sk_wmem_queued_add(sk, nskb->truesize);
2398 sk_mem_charge(sk, nskb->truesize);
2400 skb = tcp_send_head(sk);
2401 skb_copy_decrypted(nskb, skb);
2402 mptcp_skb_ext_copy(nskb, skb);
2404 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2405 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2406 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2407 TCP_SKB_CB(nskb)->sacked = 0;
2409 nskb->ip_summed = CHECKSUM_PARTIAL;
2411 tcp_insert_write_queue_before(nskb, skb, sk);
2412 tcp_highest_sack_replace(sk, skb, nskb);
2415 tcp_for_write_queue_from_safe(skb, next, sk) {
2416 copy = min_t(int, skb->len, probe_size - len);
2417 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2419 if (skb->len <= copy) {
2420 /* We've eaten all the data from this skb.
2422 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2423 /* If this is the last SKB we copy and eor is set
2424 * we need to propagate it to the new skb.
2426 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2427 tcp_skb_collapse_tstamp(nskb, skb);
2428 tcp_unlink_write_queue(skb, sk);
2429 sk_wmem_free_skb(sk, skb);
2431 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2432 ~(TCPHDR_FIN|TCPHDR_PSH);
2433 if (!skb_shinfo(skb)->nr_frags) {
2434 skb_pull(skb, copy);
2436 __pskb_trim_head(skb, copy);
2437 tcp_set_skb_tso_segs(skb, mss_now);
2439 TCP_SKB_CB(skb)->seq += copy;
2444 if (len >= probe_size)
2447 tcp_init_tso_segs(nskb, nskb->len);
2449 /* We're ready to send. If this fails, the probe will
2450 * be resegmented into mss-sized pieces by tcp_write_xmit().
2452 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2453 /* Decrement cwnd here because we are sending
2454 * effectively two packets. */
2455 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1);
2456 tcp_event_new_data_sent(sk, nskb);
2458 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2459 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2460 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2468 static bool tcp_pacing_check(struct sock *sk)
2470 struct tcp_sock *tp = tcp_sk(sk);
2472 if (!tcp_needs_internal_pacing(sk))
2475 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2478 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2479 hrtimer_start(&tp->pacing_timer,
2480 ns_to_ktime(tp->tcp_wstamp_ns),
2481 HRTIMER_MODE_ABS_PINNED_SOFT);
2487 /* TCP Small Queues :
2488 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2489 * (These limits are doubled for retransmits)
2491 * - better RTT estimation and ACK scheduling
2494 * Alas, some drivers / subsystems require a fair amount
2495 * of queued bytes to ensure line rate.
2496 * One example is wifi aggregation (802.11 AMPDU)
2498 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2499 unsigned int factor)
2501 unsigned long limit;
2503 limit = max_t(unsigned long,
2505 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2506 if (sk->sk_pacing_status == SK_PACING_NONE)
2507 limit = min_t(unsigned long, limit,
2508 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2511 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2512 tcp_sk(sk)->tcp_tx_delay) {
2513 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2515 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2516 * approximate our needs assuming an ~100% skb->truesize overhead.
2517 * USEC_PER_SEC is approximated by 2^20.
2518 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2520 extra_bytes >>= (20 - 1);
2521 limit += extra_bytes;
2523 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2524 /* Always send skb if rtx queue is empty.
2525 * No need to wait for TX completion to call us back,
2526 * after softirq/tasklet schedule.
2527 * This helps when TX completions are delayed too much.
2529 if (tcp_rtx_queue_empty(sk))
2532 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2533 /* It is possible TX completion already happened
2534 * before we set TSQ_THROTTLED, so we must
2535 * test again the condition.
2537 smp_mb__after_atomic();
2538 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2544 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2546 const u32 now = tcp_jiffies32;
2547 enum tcp_chrono old = tp->chrono_type;
2549 if (old > TCP_CHRONO_UNSPEC)
2550 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2551 tp->chrono_start = now;
2552 tp->chrono_type = new;
2555 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2557 struct tcp_sock *tp = tcp_sk(sk);
2559 /* If there are multiple conditions worthy of tracking in a
2560 * chronograph then the highest priority enum takes precedence
2561 * over the other conditions. So that if something "more interesting"
2562 * starts happening, stop the previous chrono and start a new one.
2564 if (type > tp->chrono_type)
2565 tcp_chrono_set(tp, type);
2568 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2570 struct tcp_sock *tp = tcp_sk(sk);
2573 /* There are multiple conditions worthy of tracking in a
2574 * chronograph, so that the highest priority enum takes
2575 * precedence over the other conditions (see tcp_chrono_start).
2576 * If a condition stops, we only stop chrono tracking if
2577 * it's the "most interesting" or current chrono we are
2578 * tracking and starts busy chrono if we have pending data.
2580 if (tcp_rtx_and_write_queues_empty(sk))
2581 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2582 else if (type == tp->chrono_type)
2583 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2586 /* This routine writes packets to the network. It advances the
2587 * send_head. This happens as incoming acks open up the remote
2590 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2591 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2592 * account rare use of URG, this is not a big flaw.
2594 * Send at most one packet when push_one > 0. Temporarily ignore
2595 * cwnd limit to force at most one packet out when push_one == 2.
2597 * Returns true, if no segments are in flight and we have queued segments,
2598 * but cannot send anything now because of SWS or another problem.
2600 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2601 int push_one, gfp_t gfp)
2603 struct tcp_sock *tp = tcp_sk(sk);
2604 struct sk_buff *skb;
2605 unsigned int tso_segs, sent_pkts;
2608 bool is_cwnd_limited = false, is_rwnd_limited = false;
2613 tcp_mstamp_refresh(tp);
2615 /* Do MTU probing. */
2616 result = tcp_mtu_probe(sk);
2619 } else if (result > 0) {
2624 max_segs = tcp_tso_segs(sk, mss_now);
2625 while ((skb = tcp_send_head(sk))) {
2628 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2629 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2630 skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2631 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2632 tcp_init_tso_segs(skb, mss_now);
2633 goto repair; /* Skip network transmission */
2636 if (tcp_pacing_check(sk))
2639 tso_segs = tcp_init_tso_segs(skb, mss_now);
2642 cwnd_quota = tcp_cwnd_test(tp, skb);
2645 /* Force out a loss probe pkt. */
2651 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2652 is_rwnd_limited = true;
2656 if (tso_segs == 1) {
2657 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2658 (tcp_skb_is_last(sk, skb) ?
2659 nonagle : TCP_NAGLE_PUSH))))
2663 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2664 &is_rwnd_limited, max_segs))
2669 if (tso_segs > 1 && !tcp_urg_mode(tp))
2670 limit = tcp_mss_split_point(sk, skb, mss_now,
2676 if (skb->len > limit &&
2677 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2680 if (tcp_small_queue_check(sk, skb, 0))
2683 /* Argh, we hit an empty skb(), presumably a thread
2684 * is sleeping in sendmsg()/sk_stream_wait_memory().
2685 * We do not want to send a pure-ack packet and have
2686 * a strange looking rtx queue with empty packet(s).
2688 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2691 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2695 /* Advance the send_head. This one is sent out.
2696 * This call will increment packets_out.
2698 tcp_event_new_data_sent(sk, skb);
2700 tcp_minshall_update(tp, mss_now, skb);
2701 sent_pkts += tcp_skb_pcount(skb);
2707 if (is_rwnd_limited)
2708 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2710 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2712 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp));
2713 if (likely(sent_pkts || is_cwnd_limited))
2714 tcp_cwnd_validate(sk, is_cwnd_limited);
2716 if (likely(sent_pkts)) {
2717 if (tcp_in_cwnd_reduction(sk))
2718 tp->prr_out += sent_pkts;
2720 /* Send one loss probe per tail loss episode. */
2722 tcp_schedule_loss_probe(sk, false);
2725 return !tp->packets_out && !tcp_write_queue_empty(sk);
2728 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2730 struct inet_connection_sock *icsk = inet_csk(sk);
2731 struct tcp_sock *tp = tcp_sk(sk);
2732 u32 timeout, rto_delta_us;
2735 /* Don't do any loss probe on a Fast Open connection before 3WHS
2738 if (rcu_access_pointer(tp->fastopen_rsk))
2741 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2742 /* Schedule a loss probe in 2*RTT for SACK capable connections
2743 * not in loss recovery, that are either limited by cwnd or application.
2745 if ((early_retrans != 3 && early_retrans != 4) ||
2746 !tp->packets_out || !tcp_is_sack(tp) ||
2747 (icsk->icsk_ca_state != TCP_CA_Open &&
2748 icsk->icsk_ca_state != TCP_CA_CWR))
2751 /* Probe timeout is 2*rtt. Add minimum RTO to account
2752 * for delayed ack when there's one outstanding packet. If no RTT
2753 * sample is available then probe after TCP_TIMEOUT_INIT.
2756 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2757 if (tp->packets_out == 1)
2758 timeout += TCP_RTO_MIN;
2760 timeout += TCP_TIMEOUT_MIN;
2762 timeout = TCP_TIMEOUT_INIT;
2765 /* If the RTO formula yields an earlier time, then use that time. */
2766 rto_delta_us = advancing_rto ?
2767 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2768 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2769 if (rto_delta_us > 0)
2770 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2772 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2776 /* Thanks to skb fast clones, we can detect if a prior transmit of
2777 * a packet is still in a qdisc or driver queue.
2778 * In this case, there is very little point doing a retransmit !
2780 static bool skb_still_in_host_queue(struct sock *sk,
2781 const struct sk_buff *skb)
2783 if (unlikely(skb_fclone_busy(sk, skb))) {
2784 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2785 smp_mb__after_atomic();
2786 if (skb_fclone_busy(sk, skb)) {
2787 NET_INC_STATS(sock_net(sk),
2788 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2795 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2796 * retransmit the last segment.
2798 void tcp_send_loss_probe(struct sock *sk)
2800 struct tcp_sock *tp = tcp_sk(sk);
2801 struct sk_buff *skb;
2803 int mss = tcp_current_mss(sk);
2805 /* At most one outstanding TLP */
2806 if (tp->tlp_high_seq)
2809 tp->tlp_retrans = 0;
2810 skb = tcp_send_head(sk);
2811 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2812 pcount = tp->packets_out;
2813 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2814 if (tp->packets_out > pcount)
2818 skb = skb_rb_last(&sk->tcp_rtx_queue);
2819 if (unlikely(!skb)) {
2820 WARN_ONCE(tp->packets_out,
2821 "invalid inflight: %u state %u cwnd %u mss %d\n",
2822 tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss);
2823 inet_csk(sk)->icsk_pending = 0;
2827 if (skb_still_in_host_queue(sk, skb))
2830 pcount = tcp_skb_pcount(skb);
2831 if (WARN_ON(!pcount))
2834 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2835 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2836 (pcount - 1) * mss, mss,
2839 skb = skb_rb_next(skb);
2842 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2845 if (__tcp_retransmit_skb(sk, skb, 1))
2848 tp->tlp_retrans = 1;
2851 /* Record snd_nxt for loss detection. */
2852 tp->tlp_high_seq = tp->snd_nxt;
2854 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2855 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2856 inet_csk(sk)->icsk_pending = 0;
2861 /* Push out any pending frames which were held back due to
2862 * TCP_CORK or attempt at coalescing tiny packets.
2863 * The socket must be locked by the caller.
2865 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2868 /* If we are closed, the bytes will have to remain here.
2869 * In time closedown will finish, we empty the write queue and
2870 * all will be happy.
2872 if (unlikely(sk->sk_state == TCP_CLOSE))
2875 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2876 sk_gfp_mask(sk, GFP_ATOMIC)))
2877 tcp_check_probe_timer(sk);
2880 /* Send _single_ skb sitting at the send head. This function requires
2881 * true push pending frames to setup probe timer etc.
2883 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2885 struct sk_buff *skb = tcp_send_head(sk);
2887 BUG_ON(!skb || skb->len < mss_now);
2889 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2892 /* This function returns the amount that we can raise the
2893 * usable window based on the following constraints
2895 * 1. The window can never be shrunk once it is offered (RFC 793)
2896 * 2. We limit memory per socket
2899 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2900 * RECV.NEXT + RCV.WIN fixed until:
2901 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2903 * i.e. don't raise the right edge of the window until you can raise
2904 * it at least MSS bytes.
2906 * Unfortunately, the recommended algorithm breaks header prediction,
2907 * since header prediction assumes th->window stays fixed.
2909 * Strictly speaking, keeping th->window fixed violates the receiver
2910 * side SWS prevention criteria. The problem is that under this rule
2911 * a stream of single byte packets will cause the right side of the
2912 * window to always advance by a single byte.
2914 * Of course, if the sender implements sender side SWS prevention
2915 * then this will not be a problem.
2917 * BSD seems to make the following compromise:
2919 * If the free space is less than the 1/4 of the maximum
2920 * space available and the free space is less than 1/2 mss,
2921 * then set the window to 0.
2922 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2923 * Otherwise, just prevent the window from shrinking
2924 * and from being larger than the largest representable value.
2926 * This prevents incremental opening of the window in the regime
2927 * where TCP is limited by the speed of the reader side taking
2928 * data out of the TCP receive queue. It does nothing about
2929 * those cases where the window is constrained on the sender side
2930 * because the pipeline is full.
2932 * BSD also seems to "accidentally" limit itself to windows that are a
2933 * multiple of MSS, at least until the free space gets quite small.
2934 * This would appear to be a side effect of the mbuf implementation.
2935 * Combining these two algorithms results in the observed behavior
2936 * of having a fixed window size at almost all times.
2938 * Below we obtain similar behavior by forcing the offered window to
2939 * a multiple of the mss when it is feasible to do so.
2941 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2942 * Regular options like TIMESTAMP are taken into account.
2944 u32 __tcp_select_window(struct sock *sk)
2946 struct inet_connection_sock *icsk = inet_csk(sk);
2947 struct tcp_sock *tp = tcp_sk(sk);
2948 /* MSS for the peer's data. Previous versions used mss_clamp
2949 * here. I don't know if the value based on our guesses
2950 * of peer's MSS is better for the performance. It's more correct
2951 * but may be worse for the performance because of rcv_mss
2952 * fluctuations. --SAW 1998/11/1
2954 int mss = icsk->icsk_ack.rcv_mss;
2955 int free_space = tcp_space(sk);
2956 int allowed_space = tcp_full_space(sk);
2957 int full_space, window;
2959 if (sk_is_mptcp(sk))
2960 mptcp_space(sk, &free_space, &allowed_space);
2962 full_space = min_t(int, tp->window_clamp, allowed_space);
2964 if (unlikely(mss > full_space)) {
2969 if (free_space < (full_space >> 1)) {
2970 icsk->icsk_ack.quick = 0;
2972 if (tcp_under_memory_pressure(sk))
2973 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2976 /* free_space might become our new window, make sure we don't
2977 * increase it due to wscale.
2979 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2981 /* if free space is less than mss estimate, or is below 1/16th
2982 * of the maximum allowed, try to move to zero-window, else
2983 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2984 * new incoming data is dropped due to memory limits.
2985 * With large window, mss test triggers way too late in order
2986 * to announce zero window in time before rmem limit kicks in.
2988 if (free_space < (allowed_space >> 4) || free_space < mss)
2992 if (free_space > tp->rcv_ssthresh)
2993 free_space = tp->rcv_ssthresh;
2995 /* Don't do rounding if we are using window scaling, since the
2996 * scaled window will not line up with the MSS boundary anyway.
2998 if (tp->rx_opt.rcv_wscale) {
2999 window = free_space;
3001 /* Advertise enough space so that it won't get scaled away.
3002 * Import case: prevent zero window announcement if
3003 * 1<<rcv_wscale > mss.
3005 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
3007 window = tp->rcv_wnd;
3008 /* Get the largest window that is a nice multiple of mss.
3009 * Window clamp already applied above.
3010 * If our current window offering is within 1 mss of the
3011 * free space we just keep it. This prevents the divide
3012 * and multiply from happening most of the time.
3013 * We also don't do any window rounding when the free space
3016 if (window <= free_space - mss || window > free_space)
3017 window = rounddown(free_space, mss);
3018 else if (mss == full_space &&
3019 free_space > window + (full_space >> 1))
3020 window = free_space;
3026 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3027 const struct sk_buff *next_skb)
3029 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3030 const struct skb_shared_info *next_shinfo =
3031 skb_shinfo(next_skb);
3032 struct skb_shared_info *shinfo = skb_shinfo(skb);
3034 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3035 shinfo->tskey = next_shinfo->tskey;
3036 TCP_SKB_CB(skb)->txstamp_ack |=
3037 TCP_SKB_CB(next_skb)->txstamp_ack;
3041 /* Collapses two adjacent SKB's during retransmission. */
3042 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3044 struct tcp_sock *tp = tcp_sk(sk);
3045 struct sk_buff *next_skb = skb_rb_next(skb);
3048 next_skb_size = next_skb->len;
3050 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3052 if (next_skb_size) {
3053 if (next_skb_size <= skb_availroom(skb))
3054 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
3056 else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3059 tcp_highest_sack_replace(sk, next_skb, skb);
3061 /* Update sequence range on original skb. */
3062 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3064 /* Merge over control information. This moves PSH/FIN etc. over */
3065 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3067 /* All done, get rid of second SKB and account for it so
3068 * packet counting does not break.
3070 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3071 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3073 /* changed transmit queue under us so clear hints */
3074 tcp_clear_retrans_hints_partial(tp);
3075 if (next_skb == tp->retransmit_skb_hint)
3076 tp->retransmit_skb_hint = skb;
3078 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3080 tcp_skb_collapse_tstamp(skb, next_skb);
3082 tcp_rtx_queue_unlink_and_free(next_skb, sk);
3086 /* Check if coalescing SKBs is legal. */
3087 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3089 if (tcp_skb_pcount(skb) > 1)
3091 if (skb_cloned(skb))
3093 /* Some heuristics for collapsing over SACK'd could be invented */
3094 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3100 /* Collapse packets in the retransmit queue to make to create
3101 * less packets on the wire. This is only done on retransmission.
3103 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3106 struct tcp_sock *tp = tcp_sk(sk);
3107 struct sk_buff *skb = to, *tmp;
3110 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
3112 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3115 skb_rbtree_walk_from_safe(skb, tmp) {
3116 if (!tcp_can_collapse(sk, skb))
3119 if (!tcp_skb_can_collapse(to, skb))
3132 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3135 if (!tcp_collapse_retrans(sk, to))
3140 /* This retransmits one SKB. Policy decisions and retransmit queue
3141 * state updates are done by the caller. Returns non-zero if an
3142 * error occurred which prevented the send.
3144 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3146 struct inet_connection_sock *icsk = inet_csk(sk);
3147 struct tcp_sock *tp = tcp_sk(sk);
3148 unsigned int cur_mss;
3152 /* Inconclusive MTU probe */
3153 if (icsk->icsk_mtup.probe_size)
3154 icsk->icsk_mtup.probe_size = 0;
3156 if (skb_still_in_host_queue(sk, skb))
3159 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
3160 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3164 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3168 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3169 return -EHOSTUNREACH; /* Routing failure or similar. */
3171 cur_mss = tcp_current_mss(sk);
3173 /* If receiver has shrunk his window, and skb is out of
3174 * new window, do not retransmit it. The exception is the
3175 * case, when window is shrunk to zero. In this case
3176 * our retransmit serves as a zero window probe.
3178 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
3179 TCP_SKB_CB(skb)->seq != tp->snd_una)
3182 len = cur_mss * segs;
3183 if (skb->len > len) {
3184 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3185 cur_mss, GFP_ATOMIC))
3186 return -ENOMEM; /* We'll try again later. */
3188 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
3191 diff = tcp_skb_pcount(skb);
3192 tcp_set_skb_tso_segs(skb, cur_mss);
3193 diff -= tcp_skb_pcount(skb);
3195 tcp_adjust_pcount(sk, skb, diff);
3196 if (skb->len < cur_mss)
3197 tcp_retrans_try_collapse(sk, skb, cur_mss);
3200 /* RFC3168, section 6.1.1.1. ECN fallback */
3201 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3202 tcp_ecn_clear_syn(sk, skb);
3204 /* Update global and local TCP statistics. */
3205 segs = tcp_skb_pcount(skb);
3206 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3207 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3208 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3209 tp->total_retrans += segs;
3210 tp->bytes_retrans += skb->len;
3212 /* make sure skb->data is aligned on arches that require it
3213 * and check if ack-trimming & collapsing extended the headroom
3214 * beyond what csum_start can cover.
3216 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3217 skb_headroom(skb) >= 0xFFFF)) {
3218 struct sk_buff *nskb;
3220 tcp_skb_tsorted_save(skb) {
3221 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3224 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3228 } tcp_skb_tsorted_restore(skb);
3231 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3232 tcp_rate_skb_sent(sk, skb);
3235 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3238 /* To avoid taking spuriously low RTT samples based on a timestamp
3239 * for a transmit that never happened, always mark EVER_RETRANS
3241 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3243 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3244 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3245 TCP_SKB_CB(skb)->seq, segs, err);
3248 trace_tcp_retransmit_skb(sk, skb);
3249 } else if (err != -EBUSY) {
3250 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3255 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3257 struct tcp_sock *tp = tcp_sk(sk);
3258 int err = __tcp_retransmit_skb(sk, skb, segs);
3261 #if FASTRETRANS_DEBUG > 0
3262 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3263 net_dbg_ratelimited("retrans_out leaked\n");
3266 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3267 tp->retrans_out += tcp_skb_pcount(skb);
3270 /* Save stamp of the first (attempted) retransmit. */
3271 if (!tp->retrans_stamp)
3272 tp->retrans_stamp = tcp_skb_timestamp(skb);
3274 if (tp->undo_retrans < 0)
3275 tp->undo_retrans = 0;
3276 tp->undo_retrans += tcp_skb_pcount(skb);
3280 /* This gets called after a retransmit timeout, and the initially
3281 * retransmitted data is acknowledged. It tries to continue
3282 * resending the rest of the retransmit queue, until either
3283 * we've sent it all or the congestion window limit is reached.
3285 void tcp_xmit_retransmit_queue(struct sock *sk)
3287 const struct inet_connection_sock *icsk = inet_csk(sk);
3288 struct sk_buff *skb, *rtx_head, *hole = NULL;
3289 struct tcp_sock *tp = tcp_sk(sk);
3290 bool rearm_timer = false;
3294 if (!tp->packets_out)
3297 rtx_head = tcp_rtx_queue_head(sk);
3298 skb = tp->retransmit_skb_hint ?: rtx_head;
3299 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3300 skb_rbtree_walk_from(skb) {
3304 if (tcp_pacing_check(sk))
3307 /* we could do better than to assign each time */
3309 tp->retransmit_skb_hint = skb;
3311 segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp);
3314 sacked = TCP_SKB_CB(skb)->sacked;
3315 /* In case tcp_shift_skb_data() have aggregated large skbs,
3316 * we need to make sure not sending too bigs TSO packets
3318 segs = min_t(int, segs, max_segs);
3320 if (tp->retrans_out >= tp->lost_out) {
3322 } else if (!(sacked & TCPCB_LOST)) {
3323 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3328 if (icsk->icsk_ca_state != TCP_CA_Loss)
3329 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3331 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3334 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3337 if (tcp_small_queue_check(sk, skb, 1))
3340 if (tcp_retransmit_skb(sk, skb, segs))
3343 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3345 if (tcp_in_cwnd_reduction(sk))
3346 tp->prr_out += tcp_skb_pcount(skb);
3348 if (skb == rtx_head &&
3349 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3354 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3355 inet_csk(sk)->icsk_rto,
3359 /* We allow to exceed memory limits for FIN packets to expedite
3360 * connection tear down and (memory) recovery.
3361 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3362 * or even be forced to close flow without any FIN.
3363 * In general, we want to allow one skb per socket to avoid hangs
3364 * with edge trigger epoll()
3366 void sk_forced_mem_schedule(struct sock *sk, int size)
3370 if (size <= sk->sk_forward_alloc)
3372 amt = sk_mem_pages(size);
3373 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3374 sk_memory_allocated_add(sk, amt);
3376 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3377 mem_cgroup_charge_skmem(sk->sk_memcg, amt,
3378 gfp_memcg_charge() | __GFP_NOFAIL);
3381 /* Send a FIN. The caller locks the socket for us.
3382 * We should try to send a FIN packet really hard, but eventually give up.
3384 void tcp_send_fin(struct sock *sk)
3386 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3387 struct tcp_sock *tp = tcp_sk(sk);
3389 /* Optimization, tack on the FIN if we have one skb in write queue and
3390 * this skb was not yet sent, or we are under memory pressure.
3391 * Note: in the latter case, FIN packet will be sent after a timeout,
3392 * as TCP stack thinks it has already been transmitted.
3395 if (!tskb && tcp_under_memory_pressure(sk))
3396 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3399 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3400 TCP_SKB_CB(tskb)->end_seq++;
3403 /* This means tskb was already sent.
3404 * Pretend we included the FIN on previous transmit.
3405 * We need to set tp->snd_nxt to the value it would have
3406 * if FIN had been sent. This is because retransmit path
3407 * does not change tp->snd_nxt.
3409 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3413 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3417 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3418 skb_reserve(skb, MAX_TCP_HEADER);
3419 sk_forced_mem_schedule(sk, skb->truesize);
3420 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3421 tcp_init_nondata_skb(skb, tp->write_seq,
3422 TCPHDR_ACK | TCPHDR_FIN);
3423 tcp_queue_skb(sk, skb);
3425 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3428 /* We get here when a process closes a file descriptor (either due to
3429 * an explicit close() or as a byproduct of exit()'ing) and there
3430 * was unread data in the receive queue. This behavior is recommended
3431 * by RFC 2525, section 2.17. -DaveM
3433 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3435 struct sk_buff *skb;
3437 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3439 /* NOTE: No TCP options attached and we never retransmit this. */
3440 skb = alloc_skb(MAX_TCP_HEADER, priority);
3442 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3446 /* Reserve space for headers and prepare control bits. */
3447 skb_reserve(skb, MAX_TCP_HEADER);
3448 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3449 TCPHDR_ACK | TCPHDR_RST);
3450 tcp_mstamp_refresh(tcp_sk(sk));
3452 if (tcp_transmit_skb(sk, skb, 0, priority))
3453 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3455 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3456 * skb here is different to the troublesome skb, so use NULL
3458 trace_tcp_send_reset(sk, NULL);
3461 /* Send a crossed SYN-ACK during socket establishment.
3462 * WARNING: This routine must only be called when we have already sent
3463 * a SYN packet that crossed the incoming SYN that caused this routine
3464 * to get called. If this assumption fails then the initial rcv_wnd
3465 * and rcv_wscale values will not be correct.
3467 int tcp_send_synack(struct sock *sk)
3469 struct sk_buff *skb;
3471 skb = tcp_rtx_queue_head(sk);
3472 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3473 pr_err("%s: wrong queue state\n", __func__);
3476 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3477 if (skb_cloned(skb)) {
3478 struct sk_buff *nskb;
3480 tcp_skb_tsorted_save(skb) {
3481 nskb = skb_copy(skb, GFP_ATOMIC);
3482 } tcp_skb_tsorted_restore(skb);
3485 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3486 tcp_highest_sack_replace(sk, skb, nskb);
3487 tcp_rtx_queue_unlink_and_free(skb, sk);
3488 __skb_header_release(nskb);
3489 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3490 sk_wmem_queued_add(sk, nskb->truesize);
3491 sk_mem_charge(sk, nskb->truesize);
3495 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3496 tcp_ecn_send_synack(sk, skb);
3498 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3502 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3503 * @sk: listener socket
3504 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3505 * should not use it again.
3506 * @req: request_sock pointer
3507 * @foc: cookie for tcp fast open
3508 * @synack_type: Type of synack to prepare
3509 * @syn_skb: SYN packet just received. It could be NULL for rtx case.
3511 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3512 struct request_sock *req,
3513 struct tcp_fastopen_cookie *foc,
3514 enum tcp_synack_type synack_type,
3515 struct sk_buff *syn_skb)
3517 struct inet_request_sock *ireq = inet_rsk(req);
3518 const struct tcp_sock *tp = tcp_sk(sk);
3519 struct tcp_md5sig_key *md5 = NULL;
3520 struct tcp_out_options opts;
3521 struct sk_buff *skb;
3522 int tcp_header_size;
3527 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3528 if (unlikely(!skb)) {
3532 /* Reserve space for headers. */
3533 skb_reserve(skb, MAX_TCP_HEADER);
3535 switch (synack_type) {
3536 case TCP_SYNACK_NORMAL:
3537 skb_set_owner_w(skb, req_to_sk(req));
3539 case TCP_SYNACK_COOKIE:
3540 /* Under synflood, we do not attach skb to a socket,
3541 * to avoid false sharing.
3544 case TCP_SYNACK_FASTOPEN:
3545 /* sk is a const pointer, because we want to express multiple
3546 * cpu might call us concurrently.
3547 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3549 skb_set_owner_w(skb, (struct sock *)sk);
3552 skb_dst_set(skb, dst);
3554 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3556 memset(&opts, 0, sizeof(opts));
3557 now = tcp_clock_ns();
3558 #ifdef CONFIG_SYN_COOKIES
3559 if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3560 skb->skb_mstamp_ns = cookie_init_timestamp(req, now);
3564 skb->skb_mstamp_ns = now;
3565 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3566 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3569 #ifdef CONFIG_TCP_MD5SIG
3571 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3573 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3574 /* bpf program will be interested in the tcp_flags */
3575 TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3576 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3578 syn_skb) + sizeof(*th);
3580 skb_push(skb, tcp_header_size);
3581 skb_reset_transport_header(skb);
3583 th = (struct tcphdr *)skb->data;
3584 memset(th, 0, sizeof(struct tcphdr));
3587 tcp_ecn_make_synack(req, th);
3588 th->source = htons(ireq->ir_num);
3589 th->dest = ireq->ir_rmt_port;
3590 skb->mark = ireq->ir_mark;
3591 skb->ip_summed = CHECKSUM_PARTIAL;
3592 th->seq = htonl(tcp_rsk(req)->snt_isn);
3593 /* XXX data is queued and acked as is. No buffer/window check */
3594 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3596 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3597 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3598 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3599 th->doff = (tcp_header_size >> 2);
3600 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3602 #ifdef CONFIG_TCP_MD5SIG
3603 /* Okay, we have all we need - do the md5 hash if needed */
3605 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3606 md5, req_to_sk(req), skb);
3610 bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
3611 synack_type, &opts);
3613 skb->skb_mstamp_ns = now;
3614 tcp_add_tx_delay(skb, tp);
3618 EXPORT_SYMBOL(tcp_make_synack);
3620 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3622 struct inet_connection_sock *icsk = inet_csk(sk);
3623 const struct tcp_congestion_ops *ca;
3624 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3626 if (ca_key == TCP_CA_UNSPEC)
3630 ca = tcp_ca_find_key(ca_key);
3631 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3632 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3633 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3634 icsk->icsk_ca_ops = ca;
3639 /* Do all connect socket setups that can be done AF independent. */
3640 static void tcp_connect_init(struct sock *sk)
3642 const struct dst_entry *dst = __sk_dst_get(sk);
3643 struct tcp_sock *tp = tcp_sk(sk);
3647 /* We'll fix this up when we get a response from the other end.
3648 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3650 tp->tcp_header_len = sizeof(struct tcphdr);
3651 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3652 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3654 #ifdef CONFIG_TCP_MD5SIG
3655 if (tp->af_specific->md5_lookup(sk, sk))
3656 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3659 /* If user gave his TCP_MAXSEG, record it to clamp */
3660 if (tp->rx_opt.user_mss)
3661 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3664 tcp_sync_mss(sk, dst_mtu(dst));
3666 tcp_ca_dst_init(sk, dst);
3668 if (!tp->window_clamp)
3669 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3670 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3672 tcp_initialize_rcv_mss(sk);
3674 /* limit the window selection if the user enforce a smaller rx buffer */
3675 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3676 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3677 tp->window_clamp = tcp_full_space(sk);
3679 rcv_wnd = tcp_rwnd_init_bpf(sk);
3681 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3683 tcp_select_initial_window(sk, tcp_full_space(sk),
3684 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3687 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3691 tp->rx_opt.rcv_wscale = rcv_wscale;
3692 tp->rcv_ssthresh = tp->rcv_wnd;
3695 sock_reset_flag(sk, SOCK_DONE);
3698 tcp_write_queue_purge(sk);
3699 tp->snd_una = tp->write_seq;
3700 tp->snd_sml = tp->write_seq;
3701 tp->snd_up = tp->write_seq;
3702 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3704 if (likely(!tp->repair))
3707 tp->rcv_tstamp = tcp_jiffies32;
3708 tp->rcv_wup = tp->rcv_nxt;
3709 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3711 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3712 inet_csk(sk)->icsk_retransmits = 0;
3713 tcp_clear_retrans(tp);
3716 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3718 struct tcp_sock *tp = tcp_sk(sk);
3719 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3721 tcb->end_seq += skb->len;
3722 __skb_header_release(skb);
3723 sk_wmem_queued_add(sk, skb->truesize);
3724 sk_mem_charge(sk, skb->truesize);
3725 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3726 tp->packets_out += tcp_skb_pcount(skb);
3729 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3730 * queue a data-only packet after the regular SYN, such that regular SYNs
3731 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3732 * only the SYN sequence, the data are retransmitted in the first ACK.
3733 * If cookie is not cached or other error occurs, falls back to send a
3734 * regular SYN with Fast Open cookie request option.
3736 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3738 struct inet_connection_sock *icsk = inet_csk(sk);
3739 struct tcp_sock *tp = tcp_sk(sk);
3740 struct tcp_fastopen_request *fo = tp->fastopen_req;
3742 struct sk_buff *syn_data;
3744 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3745 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3748 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3749 * user-MSS. Reserve maximum option space for middleboxes that add
3750 * private TCP options. The cost is reduced data space in SYN :(
3752 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3753 /* Sync mss_cache after updating the mss_clamp */
3754 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3756 space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3757 MAX_TCP_OPTION_SPACE;
3759 space = min_t(size_t, space, fo->size);
3761 /* limit to order-0 allocations */
3762 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3764 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3767 syn_data->ip_summed = CHECKSUM_PARTIAL;
3768 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3770 int copied = copy_from_iter(skb_put(syn_data, space), space,
3771 &fo->data->msg_iter);
3772 if (unlikely(!copied)) {
3773 tcp_skb_tsorted_anchor_cleanup(syn_data);
3774 kfree_skb(syn_data);
3777 if (copied != space) {
3778 skb_trim(syn_data, copied);
3781 skb_zcopy_set(syn_data, fo->uarg, NULL);
3783 /* No more data pending in inet_wait_for_connect() */
3784 if (space == fo->size)
3788 tcp_connect_queue_skb(sk, syn_data);
3790 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3792 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3794 syn->skb_mstamp_ns = syn_data->skb_mstamp_ns;
3796 /* Now full SYN+DATA was cloned and sent (or not),
3797 * remove the SYN from the original skb (syn_data)
3798 * we keep in write queue in case of a retransmit, as we
3799 * also have the SYN packet (with no data) in the same queue.
3801 TCP_SKB_CB(syn_data)->seq++;
3802 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3804 tp->syn_data = (fo->copied > 0);
3805 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3806 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3810 /* data was not sent, put it in write_queue */
3811 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3812 tp->packets_out -= tcp_skb_pcount(syn_data);
3815 /* Send a regular SYN with Fast Open cookie request option */
3816 if (fo->cookie.len > 0)
3818 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3820 tp->syn_fastopen = 0;
3822 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3826 /* Build a SYN and send it off. */
3827 int tcp_connect(struct sock *sk)
3829 struct tcp_sock *tp = tcp_sk(sk);
3830 struct sk_buff *buff;
3833 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3835 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3836 return -EHOSTUNREACH; /* Routing failure or similar. */
3838 tcp_connect_init(sk);
3840 if (unlikely(tp->repair)) {
3841 tcp_finish_connect(sk, NULL);
3845 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3846 if (unlikely(!buff))
3849 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3850 tcp_mstamp_refresh(tp);
3851 tp->retrans_stamp = tcp_time_stamp(tp);
3852 tcp_connect_queue_skb(sk, buff);
3853 tcp_ecn_send_syn(sk, buff);
3854 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3856 /* Send off SYN; include data in Fast Open. */
3857 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3858 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3859 if (err == -ECONNREFUSED)
3862 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3863 * in order to make this packet get counted in tcpOutSegs.
3865 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3866 tp->pushed_seq = tp->write_seq;
3867 buff = tcp_send_head(sk);
3868 if (unlikely(buff)) {
3869 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3870 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3872 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3874 /* Timer for repeating the SYN until an answer. */
3875 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3876 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3879 EXPORT_SYMBOL(tcp_connect);
3881 /* Send out a delayed ack, the caller does the policy checking
3882 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3885 void tcp_send_delayed_ack(struct sock *sk)
3887 struct inet_connection_sock *icsk = inet_csk(sk);
3888 int ato = icsk->icsk_ack.ato;
3889 unsigned long timeout;
3891 if (ato > TCP_DELACK_MIN) {
3892 const struct tcp_sock *tp = tcp_sk(sk);
3893 int max_ato = HZ / 2;
3895 if (inet_csk_in_pingpong_mode(sk) ||
3896 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3897 max_ato = TCP_DELACK_MAX;
3899 /* Slow path, intersegment interval is "high". */
3901 /* If some rtt estimate is known, use it to bound delayed ack.
3902 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3906 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3913 ato = min(ato, max_ato);
3916 ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max);
3918 /* Stay within the limit we were given */
3919 timeout = jiffies + ato;
3921 /* Use new timeout only if there wasn't a older one earlier. */
3922 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3923 /* If delack timer is about to expire, send ACK now. */
3924 if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3929 if (!time_before(timeout, icsk->icsk_ack.timeout))
3930 timeout = icsk->icsk_ack.timeout;
3932 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3933 icsk->icsk_ack.timeout = timeout;
3934 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3937 /* This routine sends an ack and also updates the window. */
3938 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3940 struct sk_buff *buff;
3942 /* If we have been reset, we may not send again. */
3943 if (sk->sk_state == TCP_CLOSE)
3946 /* We are not putting this on the write queue, so
3947 * tcp_transmit_skb() will set the ownership to this
3950 buff = alloc_skb(MAX_TCP_HEADER,
3951 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3952 if (unlikely(!buff)) {
3953 struct inet_connection_sock *icsk = inet_csk(sk);
3954 unsigned long delay;
3956 delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
3957 if (delay < TCP_RTO_MAX)
3958 icsk->icsk_ack.retry++;
3959 inet_csk_schedule_ack(sk);
3960 icsk->icsk_ack.ato = TCP_ATO_MIN;
3961 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
3965 /* Reserve space for headers and prepare control bits. */
3966 skb_reserve(buff, MAX_TCP_HEADER);
3967 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3969 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3971 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3973 skb_set_tcp_pure_ack(buff);
3975 /* Send it off, this clears delayed acks for us. */
3976 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3978 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3980 void tcp_send_ack(struct sock *sk)
3982 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3985 /* This routine sends a packet with an out of date sequence
3986 * number. It assumes the other end will try to ack it.
3988 * Question: what should we make while urgent mode?
3989 * 4.4BSD forces sending single byte of data. We cannot send
3990 * out of window data, because we have SND.NXT==SND.MAX...
3992 * Current solution: to send TWO zero-length segments in urgent mode:
3993 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3994 * out-of-date with SND.UNA-1 to probe window.
3996 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3998 struct tcp_sock *tp = tcp_sk(sk);
3999 struct sk_buff *skb;
4001 /* We don't queue it, tcp_transmit_skb() sets ownership. */
4002 skb = alloc_skb(MAX_TCP_HEADER,
4003 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4007 /* Reserve space for headers and set control bits. */
4008 skb_reserve(skb, MAX_TCP_HEADER);
4009 /* Use a previous sequence. This should cause the other
4010 * end to send an ack. Don't queue or clone SKB, just
4013 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
4014 NET_INC_STATS(sock_net(sk), mib);
4015 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4018 /* Called from setsockopt( ... TCP_REPAIR ) */
4019 void tcp_send_window_probe(struct sock *sk)
4021 if (sk->sk_state == TCP_ESTABLISHED) {
4022 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4023 tcp_mstamp_refresh(tcp_sk(sk));
4024 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4028 /* Initiate keepalive or window probe from timer. */
4029 int tcp_write_wakeup(struct sock *sk, int mib)
4031 struct tcp_sock *tp = tcp_sk(sk);
4032 struct sk_buff *skb;
4034 if (sk->sk_state == TCP_CLOSE)
4037 skb = tcp_send_head(sk);
4038 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4040 unsigned int mss = tcp_current_mss(sk);
4041 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4043 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4044 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4046 /* We are probing the opening of a window
4047 * but the window size is != 0
4048 * must have been a result SWS avoidance ( sender )
4050 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4052 seg_size = min(seg_size, mss);
4053 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4054 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4055 skb, seg_size, mss, GFP_ATOMIC))
4057 } else if (!tcp_skb_pcount(skb))
4058 tcp_set_skb_tso_segs(skb, mss);
4060 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4061 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4063 tcp_event_new_data_sent(sk, skb);
4066 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4067 tcp_xmit_probe_skb(sk, 1, mib);
4068 return tcp_xmit_probe_skb(sk, 0, mib);
4072 /* A window probe timeout has occurred. If window is not closed send
4073 * a partial packet else a zero probe.
4075 void tcp_send_probe0(struct sock *sk)
4077 struct inet_connection_sock *icsk = inet_csk(sk);
4078 struct tcp_sock *tp = tcp_sk(sk);
4079 struct net *net = sock_net(sk);
4080 unsigned long timeout;
4083 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4085 if (tp->packets_out || tcp_write_queue_empty(sk)) {
4086 /* Cancel probe timer, if it is not required. */
4087 icsk->icsk_probes_out = 0;
4088 icsk->icsk_backoff = 0;
4089 icsk->icsk_probes_tstamp = 0;
4093 icsk->icsk_probes_out++;
4095 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
4096 icsk->icsk_backoff++;
4097 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
4099 /* If packet was not sent due to local congestion,
4100 * Let senders fight for local resources conservatively.
4102 timeout = TCP_RESOURCE_PROBE_INTERVAL;
4105 timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout);
4106 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
4109 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4111 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4115 tcp_rsk(req)->txhash = net_tx_rndhash();
4116 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4119 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4120 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4121 if (unlikely(tcp_passive_fastopen(sk)))
4122 tcp_sk(sk)->total_retrans++;
4123 trace_tcp_retransmit_synack(sk, req);
4127 EXPORT_SYMBOL(tcp_rtx_synack);