1 // SPDX-License-Identifier: GPL-2.0-only
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
38 #define pr_fmt(fmt) "TCP: " fmt
41 #include <net/mptcp.h>
43 #include <linux/compiler.h>
44 #include <linux/gfp.h>
45 #include <linux/module.h>
46 #include <linux/static_key.h>
48 #include <trace/events/tcp.h>
50 /* Refresh clocks of a TCP socket,
51 * ensuring monotically increasing values.
53 void tcp_mstamp_refresh(struct tcp_sock *tp)
55 u64 val = tcp_clock_ns();
57 tp->tcp_clock_cache = val;
58 tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
61 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
62 int push_one, gfp_t gfp);
64 /* Account for new data that has been sent to the network. */
65 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
67 struct inet_connection_sock *icsk = inet_csk(sk);
68 struct tcp_sock *tp = tcp_sk(sk);
69 unsigned int prior_packets = tp->packets_out;
71 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
73 __skb_unlink(skb, &sk->sk_write_queue);
74 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
76 if (tp->highest_sack == NULL)
77 tp->highest_sack = skb;
79 tp->packets_out += tcp_skb_pcount(skb);
80 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
83 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
87 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
88 * window scaling factor due to loss of precision.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
94 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
96 const struct tcp_sock *tp = tcp_sk(sk);
98 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
99 (tp->rx_opt.wscale_ok &&
100 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
103 return tcp_wnd_end(tp);
106 /* Calculate mss to advertise in SYN segment.
107 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
109 * 1. It is independent of path mtu.
110 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
111 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
112 * attached devices, because some buggy hosts are confused by
114 * 4. We do not make 3, we advertise MSS, calculated from first
115 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
116 * This may be overridden via information stored in routing table.
117 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
118 * probably even Jumbo".
120 static __u16 tcp_advertise_mss(struct sock *sk)
122 struct tcp_sock *tp = tcp_sk(sk);
123 const struct dst_entry *dst = __sk_dst_get(sk);
124 int mss = tp->advmss;
127 unsigned int metric = dst_metric_advmss(dst);
138 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
139 * This is the first part of cwnd validation mechanism.
141 void tcp_cwnd_restart(struct sock *sk, s32 delta)
143 struct tcp_sock *tp = tcp_sk(sk);
144 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
145 u32 cwnd = tp->snd_cwnd;
147 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
149 tp->snd_ssthresh = tcp_current_ssthresh(sk);
150 restart_cwnd = min(restart_cwnd, cwnd);
152 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
154 tp->snd_cwnd = max(cwnd, restart_cwnd);
155 tp->snd_cwnd_stamp = tcp_jiffies32;
156 tp->snd_cwnd_used = 0;
159 /* Congestion state accounting after a packet has been sent. */
160 static void tcp_event_data_sent(struct tcp_sock *tp,
163 struct inet_connection_sock *icsk = inet_csk(sk);
164 const u32 now = tcp_jiffies32;
166 if (tcp_packets_in_flight(tp) == 0)
167 tcp_ca_event(sk, CA_EVENT_TX_START);
169 /* If this is the first data packet sent in response to the
170 * previous received data,
171 * and it is a reply for ato after last received packet,
172 * increase pingpong count.
174 if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) &&
175 (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
176 inet_csk_inc_pingpong_cnt(sk);
181 /* Account for an ACK we sent. */
182 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
185 struct tcp_sock *tp = tcp_sk(sk);
187 if (unlikely(tp->compressed_ack > TCP_FASTRETRANS_THRESH)) {
188 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
189 tp->compressed_ack - TCP_FASTRETRANS_THRESH);
190 tp->compressed_ack = TCP_FASTRETRANS_THRESH;
191 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
195 if (unlikely(rcv_nxt != tp->rcv_nxt))
196 return; /* Special ACK sent by DCTCP to reflect ECN */
197 tcp_dec_quickack_mode(sk, pkts);
198 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
201 /* Determine a window scaling and initial window to offer.
202 * Based on the assumption that the given amount of space
203 * will be offered. Store the results in the tp structure.
204 * NOTE: for smooth operation initial space offering should
205 * be a multiple of mss if possible. We assume here that mss >= 1.
206 * This MUST be enforced by all callers.
208 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
209 __u32 *rcv_wnd, __u32 *window_clamp,
210 int wscale_ok, __u8 *rcv_wscale,
213 unsigned int space = (__space < 0 ? 0 : __space);
215 /* If no clamp set the clamp to the max possible scaled window */
216 if (*window_clamp == 0)
217 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
218 space = min(*window_clamp, space);
220 /* Quantize space offering to a multiple of mss if possible. */
222 space = rounddown(space, mss);
224 /* NOTE: offering an initial window larger than 32767
225 * will break some buggy TCP stacks. If the admin tells us
226 * it is likely we could be speaking with such a buggy stack
227 * we will truncate our initial window offering to 32K-1
228 * unless the remote has sent us a window scaling option,
229 * which we interpret as a sign the remote TCP is not
230 * misinterpreting the window field as a signed quantity.
232 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
233 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
235 (*rcv_wnd) = min_t(u32, space, U16_MAX);
238 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
242 /* Set window scaling on max possible window */
243 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
244 space = max_t(u32, space, sysctl_rmem_max);
245 space = min_t(u32, space, *window_clamp);
246 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
249 /* Set the clamp no higher than max representable value */
250 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
252 EXPORT_SYMBOL(tcp_select_initial_window);
254 /* Chose a new window to advertise, update state in tcp_sock for the
255 * socket, and return result with RFC1323 scaling applied. The return
256 * value can be stuffed directly into th->window for an outgoing
259 static u16 tcp_select_window(struct sock *sk)
261 struct tcp_sock *tp = tcp_sk(sk);
262 u32 old_win = tp->rcv_wnd;
263 u32 cur_win = tcp_receive_window(tp);
264 u32 new_win = __tcp_select_window(sk);
266 /* Never shrink the offered window */
267 if (new_win < cur_win) {
268 /* Danger Will Robinson!
269 * Don't update rcv_wup/rcv_wnd here or else
270 * we will not be able to advertise a zero
271 * window in time. --DaveM
273 * Relax Will Robinson.
276 NET_INC_STATS(sock_net(sk),
277 LINUX_MIB_TCPWANTZEROWINDOWADV);
278 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
280 tp->rcv_wnd = new_win;
281 tp->rcv_wup = tp->rcv_nxt;
283 /* Make sure we do not exceed the maximum possible
286 if (!tp->rx_opt.rcv_wscale &&
287 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
288 new_win = min(new_win, MAX_TCP_WINDOW);
290 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
292 /* RFC1323 scaling applied */
293 new_win >>= tp->rx_opt.rcv_wscale;
295 /* If we advertise zero window, disable fast path. */
299 NET_INC_STATS(sock_net(sk),
300 LINUX_MIB_TCPTOZEROWINDOWADV);
301 } else if (old_win == 0) {
302 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
308 /* Packet ECN state for a SYN-ACK */
309 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
311 const struct tcp_sock *tp = tcp_sk(sk);
313 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
314 if (!(tp->ecn_flags & TCP_ECN_OK))
315 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
316 else if (tcp_ca_needs_ecn(sk) ||
317 tcp_bpf_ca_needs_ecn(sk))
321 /* Packet ECN state for a SYN. */
322 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
324 struct tcp_sock *tp = tcp_sk(sk);
325 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
326 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
327 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
330 const struct dst_entry *dst = __sk_dst_get(sk);
332 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
339 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
340 tp->ecn_flags = TCP_ECN_OK;
341 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
346 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
348 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
349 /* tp->ecn_flags are cleared at a later point in time when
350 * SYN ACK is ultimatively being received.
352 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
356 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
358 if (inet_rsk(req)->ecn_ok)
362 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
365 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
366 struct tcphdr *th, int tcp_header_len)
368 struct tcp_sock *tp = tcp_sk(sk);
370 if (tp->ecn_flags & TCP_ECN_OK) {
371 /* Not-retransmitted data segment: set ECT and inject CWR. */
372 if (skb->len != tcp_header_len &&
373 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
375 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
376 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
378 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
380 } else if (!tcp_ca_needs_ecn(sk)) {
381 /* ACK or retransmitted segment: clear ECT|CE */
382 INET_ECN_dontxmit(sk);
384 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
389 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
390 * auto increment end seqno.
392 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
394 skb->ip_summed = CHECKSUM_PARTIAL;
396 TCP_SKB_CB(skb)->tcp_flags = flags;
397 TCP_SKB_CB(skb)->sacked = 0;
399 tcp_skb_pcount_set(skb, 1);
401 TCP_SKB_CB(skb)->seq = seq;
402 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
404 TCP_SKB_CB(skb)->end_seq = seq;
407 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
409 return tp->snd_una != tp->snd_up;
412 #define OPTION_SACK_ADVERTISE (1 << 0)
413 #define OPTION_TS (1 << 1)
414 #define OPTION_MD5 (1 << 2)
415 #define OPTION_WSCALE (1 << 3)
416 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
417 #define OPTION_SMC (1 << 9)
418 #define OPTION_MPTCP (1 << 10)
420 static void smc_options_write(__be32 *ptr, u16 *options)
422 #if IS_ENABLED(CONFIG_SMC)
423 if (static_branch_unlikely(&tcp_have_smc)) {
424 if (unlikely(OPTION_SMC & *options)) {
425 *ptr++ = htonl((TCPOPT_NOP << 24) |
428 (TCPOLEN_EXP_SMC_BASE));
429 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
435 struct tcp_out_options {
436 u16 options; /* bit field of OPTION_* */
437 u16 mss; /* 0 to disable */
438 u8 ws; /* window scale, 0 to disable */
439 u8 num_sack_blocks; /* number of SACK blocks to include */
440 u8 hash_size; /* bytes in hash_location */
441 __u8 *hash_location; /* temporary pointer, overloaded */
442 __u32 tsval, tsecr; /* need to include OPTION_TS */
443 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
444 struct mptcp_out_options mptcp;
447 static void mptcp_options_write(__be32 *ptr, struct tcp_out_options *opts)
449 #if IS_ENABLED(CONFIG_MPTCP)
450 if (unlikely(OPTION_MPTCP & opts->options))
451 mptcp_write_options(ptr, &opts->mptcp);
455 /* Write previously computed TCP options to the packet.
457 * Beware: Something in the Internet is very sensitive to the ordering of
458 * TCP options, we learned this through the hard way, so be careful here.
459 * Luckily we can at least blame others for their non-compliance but from
460 * inter-operability perspective it seems that we're somewhat stuck with
461 * the ordering which we have been using if we want to keep working with
462 * those broken things (not that it currently hurts anybody as there isn't
463 * particular reason why the ordering would need to be changed).
465 * At least SACK_PERM as the first option is known to lead to a disaster
466 * (but it may well be that other scenarios fail similarly).
468 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
469 struct tcp_out_options *opts)
471 u16 options = opts->options; /* mungable copy */
473 if (unlikely(OPTION_MD5 & options)) {
474 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
475 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
476 /* overload cookie hash location */
477 opts->hash_location = (__u8 *)ptr;
481 if (unlikely(opts->mss)) {
482 *ptr++ = htonl((TCPOPT_MSS << 24) |
483 (TCPOLEN_MSS << 16) |
487 if (likely(OPTION_TS & options)) {
488 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
489 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
490 (TCPOLEN_SACK_PERM << 16) |
491 (TCPOPT_TIMESTAMP << 8) |
493 options &= ~OPTION_SACK_ADVERTISE;
495 *ptr++ = htonl((TCPOPT_NOP << 24) |
497 (TCPOPT_TIMESTAMP << 8) |
500 *ptr++ = htonl(opts->tsval);
501 *ptr++ = htonl(opts->tsecr);
504 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
505 *ptr++ = htonl((TCPOPT_NOP << 24) |
507 (TCPOPT_SACK_PERM << 8) |
511 if (unlikely(OPTION_WSCALE & options)) {
512 *ptr++ = htonl((TCPOPT_NOP << 24) |
513 (TCPOPT_WINDOW << 16) |
514 (TCPOLEN_WINDOW << 8) |
518 if (unlikely(opts->num_sack_blocks)) {
519 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
520 tp->duplicate_sack : tp->selective_acks;
523 *ptr++ = htonl((TCPOPT_NOP << 24) |
526 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
527 TCPOLEN_SACK_PERBLOCK)));
529 for (this_sack = 0; this_sack < opts->num_sack_blocks;
531 *ptr++ = htonl(sp[this_sack].start_seq);
532 *ptr++ = htonl(sp[this_sack].end_seq);
535 tp->rx_opt.dsack = 0;
538 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
539 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
541 u32 len; /* Fast Open option length */
544 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
545 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
546 TCPOPT_FASTOPEN_MAGIC);
547 p += TCPOLEN_EXP_FASTOPEN_BASE;
549 len = TCPOLEN_FASTOPEN_BASE + foc->len;
550 *p++ = TCPOPT_FASTOPEN;
554 memcpy(p, foc->val, foc->len);
555 if ((len & 3) == 2) {
556 p[foc->len] = TCPOPT_NOP;
557 p[foc->len + 1] = TCPOPT_NOP;
559 ptr += (len + 3) >> 2;
562 smc_options_write(ptr, &options);
564 mptcp_options_write(ptr, opts);
567 static void smc_set_option(const struct tcp_sock *tp,
568 struct tcp_out_options *opts,
569 unsigned int *remaining)
571 #if IS_ENABLED(CONFIG_SMC)
572 if (static_branch_unlikely(&tcp_have_smc)) {
574 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
575 opts->options |= OPTION_SMC;
576 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
583 static void smc_set_option_cond(const struct tcp_sock *tp,
584 const struct inet_request_sock *ireq,
585 struct tcp_out_options *opts,
586 unsigned int *remaining)
588 #if IS_ENABLED(CONFIG_SMC)
589 if (static_branch_unlikely(&tcp_have_smc)) {
590 if (tp->syn_smc && ireq->smc_ok) {
591 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
592 opts->options |= OPTION_SMC;
593 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
600 static void mptcp_set_option_cond(const struct request_sock *req,
601 struct tcp_out_options *opts,
602 unsigned int *remaining)
604 if (rsk_is_mptcp(req)) {
607 if (mptcp_synack_options(req, &size, &opts->mptcp)) {
608 if (*remaining >= size) {
609 opts->options |= OPTION_MPTCP;
616 /* Compute TCP options for SYN packets. This is not the final
617 * network wire format yet.
619 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
620 struct tcp_out_options *opts,
621 struct tcp_md5sig_key **md5)
623 struct tcp_sock *tp = tcp_sk(sk);
624 unsigned int remaining = MAX_TCP_OPTION_SPACE;
625 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
628 #ifdef CONFIG_TCP_MD5SIG
629 if (static_branch_unlikely(&tcp_md5_needed) &&
630 rcu_access_pointer(tp->md5sig_info)) {
631 *md5 = tp->af_specific->md5_lookup(sk, sk);
633 opts->options |= OPTION_MD5;
634 remaining -= TCPOLEN_MD5SIG_ALIGNED;
639 /* We always get an MSS option. The option bytes which will be seen in
640 * normal data packets should timestamps be used, must be in the MSS
641 * advertised. But we subtract them from tp->mss_cache so that
642 * calculations in tcp_sendmsg are simpler etc. So account for this
643 * fact here if necessary. If we don't do this correctly, as a
644 * receiver we won't recognize data packets as being full sized when we
645 * should, and thus we won't abide by the delayed ACK rules correctly.
646 * SACKs don't matter, we never delay an ACK when we have any of those
648 opts->mss = tcp_advertise_mss(sk);
649 remaining -= TCPOLEN_MSS_ALIGNED;
651 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
652 opts->options |= OPTION_TS;
653 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
654 opts->tsecr = tp->rx_opt.ts_recent;
655 remaining -= TCPOLEN_TSTAMP_ALIGNED;
657 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
658 opts->ws = tp->rx_opt.rcv_wscale;
659 opts->options |= OPTION_WSCALE;
660 remaining -= TCPOLEN_WSCALE_ALIGNED;
662 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
663 opts->options |= OPTION_SACK_ADVERTISE;
664 if (unlikely(!(OPTION_TS & opts->options)))
665 remaining -= TCPOLEN_SACKPERM_ALIGNED;
668 if (fastopen && fastopen->cookie.len >= 0) {
669 u32 need = fastopen->cookie.len;
671 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
672 TCPOLEN_FASTOPEN_BASE;
673 need = (need + 3) & ~3U; /* Align to 32 bits */
674 if (remaining >= need) {
675 opts->options |= OPTION_FAST_OPEN_COOKIE;
676 opts->fastopen_cookie = &fastopen->cookie;
678 tp->syn_fastopen = 1;
679 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
683 smc_set_option(tp, opts, &remaining);
685 if (sk_is_mptcp(sk)) {
688 if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
689 opts->options |= OPTION_MPTCP;
694 return MAX_TCP_OPTION_SPACE - remaining;
697 /* Set up TCP options for SYN-ACKs. */
698 static unsigned int tcp_synack_options(const struct sock *sk,
699 struct request_sock *req,
700 unsigned int mss, struct sk_buff *skb,
701 struct tcp_out_options *opts,
702 const struct tcp_md5sig_key *md5,
703 struct tcp_fastopen_cookie *foc)
705 struct inet_request_sock *ireq = inet_rsk(req);
706 unsigned int remaining = MAX_TCP_OPTION_SPACE;
708 #ifdef CONFIG_TCP_MD5SIG
710 opts->options |= OPTION_MD5;
711 remaining -= TCPOLEN_MD5SIG_ALIGNED;
713 /* We can't fit any SACK blocks in a packet with MD5 + TS
714 * options. There was discussion about disabling SACK
715 * rather than TS in order to fit in better with old,
716 * buggy kernels, but that was deemed to be unnecessary.
718 ireq->tstamp_ok &= !ireq->sack_ok;
722 /* We always send an MSS option. */
724 remaining -= TCPOLEN_MSS_ALIGNED;
726 if (likely(ireq->wscale_ok)) {
727 opts->ws = ireq->rcv_wscale;
728 opts->options |= OPTION_WSCALE;
729 remaining -= TCPOLEN_WSCALE_ALIGNED;
731 if (likely(ireq->tstamp_ok)) {
732 opts->options |= OPTION_TS;
733 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
734 opts->tsecr = req->ts_recent;
735 remaining -= TCPOLEN_TSTAMP_ALIGNED;
737 if (likely(ireq->sack_ok)) {
738 opts->options |= OPTION_SACK_ADVERTISE;
739 if (unlikely(!ireq->tstamp_ok))
740 remaining -= TCPOLEN_SACKPERM_ALIGNED;
742 if (foc != NULL && foc->len >= 0) {
745 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
746 TCPOLEN_FASTOPEN_BASE;
747 need = (need + 3) & ~3U; /* Align to 32 bits */
748 if (remaining >= need) {
749 opts->options |= OPTION_FAST_OPEN_COOKIE;
750 opts->fastopen_cookie = foc;
755 mptcp_set_option_cond(req, opts, &remaining);
757 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
759 return MAX_TCP_OPTION_SPACE - remaining;
762 /* Compute TCP options for ESTABLISHED sockets. This is not the
763 * final wire format yet.
765 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
766 struct tcp_out_options *opts,
767 struct tcp_md5sig_key **md5)
769 struct tcp_sock *tp = tcp_sk(sk);
770 unsigned int size = 0;
771 unsigned int eff_sacks;
776 #ifdef CONFIG_TCP_MD5SIG
777 if (static_branch_unlikely(&tcp_md5_needed) &&
778 rcu_access_pointer(tp->md5sig_info)) {
779 *md5 = tp->af_specific->md5_lookup(sk, sk);
781 opts->options |= OPTION_MD5;
782 size += TCPOLEN_MD5SIG_ALIGNED;
787 if (likely(tp->rx_opt.tstamp_ok)) {
788 opts->options |= OPTION_TS;
789 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
790 opts->tsecr = tp->rx_opt.ts_recent;
791 size += TCPOLEN_TSTAMP_ALIGNED;
794 /* MPTCP options have precedence over SACK for the limited TCP
795 * option space because a MPTCP connection would be forced to
796 * fall back to regular TCP if a required multipath option is
797 * missing. SACK still gets a chance to use whatever space is
800 if (sk_is_mptcp(sk)) {
801 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
802 unsigned int opt_size = 0;
804 if (mptcp_established_options(sk, skb, &opt_size, remaining,
806 opts->options |= OPTION_MPTCP;
811 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
812 if (unlikely(eff_sacks)) {
813 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
814 if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
815 TCPOLEN_SACK_PERBLOCK))
818 opts->num_sack_blocks =
819 min_t(unsigned int, eff_sacks,
820 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
821 TCPOLEN_SACK_PERBLOCK);
823 size += TCPOLEN_SACK_BASE_ALIGNED +
824 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
831 /* TCP SMALL QUEUES (TSQ)
833 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
834 * to reduce RTT and bufferbloat.
835 * We do this using a special skb destructor (tcp_wfree).
837 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
838 * needs to be reallocated in a driver.
839 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
841 * Since transmit from skb destructor is forbidden, we use a tasklet
842 * to process all sockets that eventually need to send more skbs.
843 * We use one tasklet per cpu, with its own queue of sockets.
846 struct tasklet_struct tasklet;
847 struct list_head head; /* queue of tcp sockets */
849 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
851 static void tcp_tsq_write(struct sock *sk)
853 if ((1 << sk->sk_state) &
854 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
855 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
856 struct tcp_sock *tp = tcp_sk(sk);
858 if (tp->lost_out > tp->retrans_out &&
859 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
860 tcp_mstamp_refresh(tp);
861 tcp_xmit_retransmit_queue(sk);
864 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
869 static void tcp_tsq_handler(struct sock *sk)
872 if (!sock_owned_by_user(sk))
874 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
879 * One tasklet per cpu tries to send more skbs.
880 * We run in tasklet context but need to disable irqs when
881 * transferring tsq->head because tcp_wfree() might
882 * interrupt us (non NAPI drivers)
884 static void tcp_tasklet_func(unsigned long data)
886 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
889 struct list_head *q, *n;
893 local_irq_save(flags);
894 list_splice_init(&tsq->head, &list);
895 local_irq_restore(flags);
897 list_for_each_safe(q, n, &list) {
898 tp = list_entry(q, struct tcp_sock, tsq_node);
899 list_del(&tp->tsq_node);
901 sk = (struct sock *)tp;
902 smp_mb__before_atomic();
903 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
910 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
911 TCPF_WRITE_TIMER_DEFERRED | \
912 TCPF_DELACK_TIMER_DEFERRED | \
913 TCPF_MTU_REDUCED_DEFERRED)
915 * tcp_release_cb - tcp release_sock() callback
918 * called from release_sock() to perform protocol dependent
919 * actions before socket release.
921 void tcp_release_cb(struct sock *sk)
923 unsigned long flags, nflags;
925 /* perform an atomic operation only if at least one flag is set */
927 flags = sk->sk_tsq_flags;
928 if (!(flags & TCP_DEFERRED_ALL))
930 nflags = flags & ~TCP_DEFERRED_ALL;
931 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
933 if (flags & TCPF_TSQ_DEFERRED) {
937 /* Here begins the tricky part :
938 * We are called from release_sock() with :
940 * 2) sk_lock.slock spinlock held
941 * 3) socket owned by us (sk->sk_lock.owned == 1)
943 * But following code is meant to be called from BH handlers,
944 * so we should keep BH disabled, but early release socket ownership
946 sock_release_ownership(sk);
948 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
949 tcp_write_timer_handler(sk);
952 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
953 tcp_delack_timer_handler(sk);
956 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
957 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
961 EXPORT_SYMBOL(tcp_release_cb);
963 void __init tcp_tasklet_init(void)
967 for_each_possible_cpu(i) {
968 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
970 INIT_LIST_HEAD(&tsq->head);
971 tasklet_init(&tsq->tasklet,
978 * Write buffer destructor automatically called from kfree_skb.
979 * We can't xmit new skbs from this context, as we might already
982 void tcp_wfree(struct sk_buff *skb)
984 struct sock *sk = skb->sk;
985 struct tcp_sock *tp = tcp_sk(sk);
986 unsigned long flags, nval, oval;
988 /* Keep one reference on sk_wmem_alloc.
989 * Will be released by sk_free() from here or tcp_tasklet_func()
991 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
993 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
994 * Wait until our queues (qdisc + devices) are drained.
996 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
997 * - chance for incoming ACK (processed by another cpu maybe)
998 * to migrate this flow (skb->ooo_okay will be eventually set)
1000 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1003 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
1004 struct tsq_tasklet *tsq;
1007 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1010 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1011 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
1015 /* queue this socket to tasklet queue */
1016 local_irq_save(flags);
1017 tsq = this_cpu_ptr(&tsq_tasklet);
1018 empty = list_empty(&tsq->head);
1019 list_add(&tp->tsq_node, &tsq->head);
1021 tasklet_schedule(&tsq->tasklet);
1022 local_irq_restore(flags);
1029 /* Note: Called under soft irq.
1030 * We can call TCP stack right away, unless socket is owned by user.
1032 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1034 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1035 struct sock *sk = (struct sock *)tp;
1037 tcp_tsq_handler(sk);
1040 return HRTIMER_NORESTART;
1043 static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1046 struct tcp_sock *tp = tcp_sk(sk);
1048 if (sk->sk_pacing_status != SK_PACING_NONE) {
1049 unsigned long rate = sk->sk_pacing_rate;
1051 /* Original sch_fq does not pace first 10 MSS
1052 * Note that tp->data_segs_out overflows after 2^32 packets,
1053 * this is a minor annoyance.
1055 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1056 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1057 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1059 /* take into account OS jitter */
1060 len_ns -= min_t(u64, len_ns / 2, credit);
1061 tp->tcp_wstamp_ns += len_ns;
1064 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1067 /* This routine actually transmits TCP packets queued in by
1068 * tcp_do_sendmsg(). This is used by both the initial
1069 * transmission and possible later retransmissions.
1070 * All SKB's seen here are completely headerless. It is our
1071 * job to build the TCP header, and pass the packet down to
1072 * IP so it can do the same plus pass the packet off to the
1075 * We are working here with either a clone of the original
1076 * SKB, or a fresh unique copy made by the retransmit engine.
1078 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1079 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1081 const struct inet_connection_sock *icsk = inet_csk(sk);
1082 struct inet_sock *inet;
1083 struct tcp_sock *tp;
1084 struct tcp_skb_cb *tcb;
1085 struct tcp_out_options opts;
1086 unsigned int tcp_options_size, tcp_header_size;
1087 struct sk_buff *oskb = NULL;
1088 struct tcp_md5sig_key *md5;
1093 BUG_ON(!skb || !tcp_skb_pcount(skb));
1095 prior_wstamp = tp->tcp_wstamp_ns;
1096 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1097 skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
1099 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1103 tcp_skb_tsorted_save(oskb) {
1104 if (unlikely(skb_cloned(oskb)))
1105 skb = pskb_copy(oskb, gfp_mask);
1107 skb = skb_clone(oskb, gfp_mask);
1108 } tcp_skb_tsorted_restore(oskb);
1112 /* retransmit skbs might have a non zero value in skb->dev
1113 * because skb->dev is aliased with skb->rbnode.rb_left
1119 tcb = TCP_SKB_CB(skb);
1120 memset(&opts, 0, sizeof(opts));
1122 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1123 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1125 tcp_options_size = tcp_established_options(sk, skb, &opts,
1127 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1128 * at receiver : This slightly improve GRO performance.
1129 * Note that we do not force the PSH flag for non GSO packets,
1130 * because they might be sent under high congestion events,
1131 * and in this case it is better to delay the delivery of 1-MSS
1132 * packets and thus the corresponding ACK packet that would
1133 * release the following packet.
1135 if (tcp_skb_pcount(skb) > 1)
1136 tcb->tcp_flags |= TCPHDR_PSH;
1138 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1140 /* if no packet is in qdisc/device queue, then allow XPS to select
1141 * another queue. We can be called from tcp_tsq_handler()
1142 * which holds one reference to sk.
1144 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1145 * One way to get this would be to set skb->truesize = 2 on them.
1147 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1149 /* If we had to use memory reserve to allocate this skb,
1150 * this might cause drops if packet is looped back :
1151 * Other socket might not have SOCK_MEMALLOC.
1152 * Packets not looped back do not care about pfmemalloc.
1154 skb->pfmemalloc = 0;
1156 skb_push(skb, tcp_header_size);
1157 skb_reset_transport_header(skb);
1161 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1162 skb_set_hash_from_sk(skb, sk);
1163 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1165 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1167 /* Build TCP header and checksum it. */
1168 th = (struct tcphdr *)skb->data;
1169 th->source = inet->inet_sport;
1170 th->dest = inet->inet_dport;
1171 th->seq = htonl(tcb->seq);
1172 th->ack_seq = htonl(rcv_nxt);
1173 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1179 /* The urg_mode check is necessary during a below snd_una win probe */
1180 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1181 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1182 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1184 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1185 th->urg_ptr = htons(0xFFFF);
1190 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1191 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1192 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1193 th->window = htons(tcp_select_window(sk));
1194 tcp_ecn_send(sk, skb, th, tcp_header_size);
1196 /* RFC1323: The window in SYN & SYN/ACK segments
1199 th->window = htons(min(tp->rcv_wnd, 65535U));
1201 #ifdef CONFIG_TCP_MD5SIG
1202 /* Calculate the MD5 hash, as we have all we need now */
1204 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1205 tp->af_specific->calc_md5_hash(opts.hash_location,
1210 icsk->icsk_af_ops->send_check(sk, skb);
1212 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1213 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1215 if (skb->len != tcp_header_size) {
1216 tcp_event_data_sent(tp, sk);
1217 tp->data_segs_out += tcp_skb_pcount(skb);
1218 tp->bytes_sent += skb->len - tcp_header_size;
1221 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1222 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1223 tcp_skb_pcount(skb));
1225 tp->segs_out += tcp_skb_pcount(skb);
1226 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1227 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1228 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1230 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1232 /* Cleanup our debris for IP stacks */
1233 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1234 sizeof(struct inet6_skb_parm)));
1236 tcp_add_tx_delay(skb, tp);
1238 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1240 if (unlikely(err > 0)) {
1242 err = net_xmit_eval(err);
1245 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1246 tcp_rate_skb_sent(sk, oskb);
1251 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1254 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1255 tcp_sk(sk)->rcv_nxt);
1258 /* This routine just queues the buffer for sending.
1260 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1261 * otherwise socket can stall.
1263 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1265 struct tcp_sock *tp = tcp_sk(sk);
1267 /* Advance write_seq and place onto the write_queue. */
1268 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1269 __skb_header_release(skb);
1270 tcp_add_write_queue_tail(sk, skb);
1271 sk_wmem_queued_add(sk, skb->truesize);
1272 sk_mem_charge(sk, skb->truesize);
1275 /* Initialize TSO segments for a packet. */
1276 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1278 if (skb->len <= mss_now) {
1279 /* Avoid the costly divide in the normal
1282 tcp_skb_pcount_set(skb, 1);
1283 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1285 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1286 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1290 /* Pcount in the middle of the write queue got changed, we need to do various
1291 * tweaks to fix counters
1293 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1295 struct tcp_sock *tp = tcp_sk(sk);
1297 tp->packets_out -= decr;
1299 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1300 tp->sacked_out -= decr;
1301 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1302 tp->retrans_out -= decr;
1303 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1304 tp->lost_out -= decr;
1306 /* Reno case is special. Sigh... */
1307 if (tcp_is_reno(tp) && decr > 0)
1308 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1310 if (tp->lost_skb_hint &&
1311 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1312 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1313 tp->lost_cnt_hint -= decr;
1315 tcp_verify_left_out(tp);
1318 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1320 return TCP_SKB_CB(skb)->txstamp_ack ||
1321 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1324 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1326 struct skb_shared_info *shinfo = skb_shinfo(skb);
1328 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1329 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1330 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1331 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1333 shinfo->tx_flags &= ~tsflags;
1334 shinfo2->tx_flags |= tsflags;
1335 swap(shinfo->tskey, shinfo2->tskey);
1336 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1337 TCP_SKB_CB(skb)->txstamp_ack = 0;
1341 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1343 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1344 TCP_SKB_CB(skb)->eor = 0;
1347 /* Insert buff after skb on the write or rtx queue of sk. */
1348 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1349 struct sk_buff *buff,
1351 enum tcp_queue tcp_queue)
1353 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1354 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1356 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1359 /* Function to create two new TCP segments. Shrinks the given segment
1360 * to the specified size and appends a new segment with the rest of the
1361 * packet to the list. This won't be called frequently, I hope.
1362 * Remember, these are still headerless SKBs at this point.
1364 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1365 struct sk_buff *skb, u32 len,
1366 unsigned int mss_now, gfp_t gfp)
1368 struct tcp_sock *tp = tcp_sk(sk);
1369 struct sk_buff *buff;
1370 int nsize, old_factor;
1375 if (WARN_ON(len > skb->len))
1378 nsize = skb_headlen(skb) - len;
1382 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1383 * We need some allowance to not penalize applications setting small
1385 * Also allow first and last skb in retransmit queue to be split.
1387 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1388 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1389 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1390 skb != tcp_rtx_queue_head(sk) &&
1391 skb != tcp_rtx_queue_tail(sk))) {
1392 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1396 if (skb_unclone(skb, gfp))
1399 /* Get a new skb... force flag on. */
1400 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1402 return -ENOMEM; /* We'll just try again later. */
1403 skb_copy_decrypted(buff, skb);
1405 sk_wmem_queued_add(sk, buff->truesize);
1406 sk_mem_charge(sk, buff->truesize);
1407 nlen = skb->len - len - nsize;
1408 buff->truesize += nlen;
1409 skb->truesize -= nlen;
1411 /* Correct the sequence numbers. */
1412 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1413 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1414 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1416 /* PSH and FIN should only be set in the second packet. */
1417 flags = TCP_SKB_CB(skb)->tcp_flags;
1418 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1419 TCP_SKB_CB(buff)->tcp_flags = flags;
1420 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1421 tcp_skb_fragment_eor(skb, buff);
1423 skb_split(skb, buff, len);
1425 buff->ip_summed = CHECKSUM_PARTIAL;
1427 buff->tstamp = skb->tstamp;
1428 tcp_fragment_tstamp(skb, buff);
1430 old_factor = tcp_skb_pcount(skb);
1432 /* Fix up tso_factor for both original and new SKB. */
1433 tcp_set_skb_tso_segs(skb, mss_now);
1434 tcp_set_skb_tso_segs(buff, mss_now);
1436 /* Update delivered info for the new segment */
1437 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1439 /* If this packet has been sent out already, we must
1440 * adjust the various packet counters.
1442 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1443 int diff = old_factor - tcp_skb_pcount(skb) -
1444 tcp_skb_pcount(buff);
1447 tcp_adjust_pcount(sk, skb, diff);
1450 /* Link BUFF into the send queue. */
1451 __skb_header_release(buff);
1452 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1453 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1454 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1459 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1460 * data is not copied, but immediately discarded.
1462 static int __pskb_trim_head(struct sk_buff *skb, int len)
1464 struct skb_shared_info *shinfo;
1467 eat = min_t(int, len, skb_headlen(skb));
1469 __skb_pull(skb, eat);
1476 shinfo = skb_shinfo(skb);
1477 for (i = 0; i < shinfo->nr_frags; i++) {
1478 int size = skb_frag_size(&shinfo->frags[i]);
1481 skb_frag_unref(skb, i);
1484 shinfo->frags[k] = shinfo->frags[i];
1486 skb_frag_off_add(&shinfo->frags[k], eat);
1487 skb_frag_size_sub(&shinfo->frags[k], eat);
1493 shinfo->nr_frags = k;
1495 skb->data_len -= len;
1496 skb->len = skb->data_len;
1500 /* Remove acked data from a packet in the transmit queue. */
1501 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1505 if (skb_unclone(skb, GFP_ATOMIC))
1508 delta_truesize = __pskb_trim_head(skb, len);
1510 TCP_SKB_CB(skb)->seq += len;
1511 skb->ip_summed = CHECKSUM_PARTIAL;
1513 if (delta_truesize) {
1514 skb->truesize -= delta_truesize;
1515 sk_wmem_queued_add(sk, -delta_truesize);
1516 sk_mem_uncharge(sk, delta_truesize);
1517 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1520 /* Any change of skb->len requires recalculation of tso factor. */
1521 if (tcp_skb_pcount(skb) > 1)
1522 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1527 /* Calculate MSS not accounting any TCP options. */
1528 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1530 const struct tcp_sock *tp = tcp_sk(sk);
1531 const struct inet_connection_sock *icsk = inet_csk(sk);
1534 /* Calculate base mss without TCP options:
1535 It is MMS_S - sizeof(tcphdr) of rfc1122
1537 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1539 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1540 if (icsk->icsk_af_ops->net_frag_header_len) {
1541 const struct dst_entry *dst = __sk_dst_get(sk);
1543 if (dst && dst_allfrag(dst))
1544 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1547 /* Clamp it (mss_clamp does not include tcp options) */
1548 if (mss_now > tp->rx_opt.mss_clamp)
1549 mss_now = tp->rx_opt.mss_clamp;
1551 /* Now subtract optional transport overhead */
1552 mss_now -= icsk->icsk_ext_hdr_len;
1554 /* Then reserve room for full set of TCP options and 8 bytes of data */
1555 mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1559 /* Calculate MSS. Not accounting for SACKs here. */
1560 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1562 /* Subtract TCP options size, not including SACKs */
1563 return __tcp_mtu_to_mss(sk, pmtu) -
1564 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1567 /* Inverse of above */
1568 int tcp_mss_to_mtu(struct sock *sk, int mss)
1570 const struct tcp_sock *tp = tcp_sk(sk);
1571 const struct inet_connection_sock *icsk = inet_csk(sk);
1575 tp->tcp_header_len +
1576 icsk->icsk_ext_hdr_len +
1577 icsk->icsk_af_ops->net_header_len;
1579 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1580 if (icsk->icsk_af_ops->net_frag_header_len) {
1581 const struct dst_entry *dst = __sk_dst_get(sk);
1583 if (dst && dst_allfrag(dst))
1584 mtu += icsk->icsk_af_ops->net_frag_header_len;
1588 EXPORT_SYMBOL(tcp_mss_to_mtu);
1590 /* MTU probing init per socket */
1591 void tcp_mtup_init(struct sock *sk)
1593 struct tcp_sock *tp = tcp_sk(sk);
1594 struct inet_connection_sock *icsk = inet_csk(sk);
1595 struct net *net = sock_net(sk);
1597 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1598 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1599 icsk->icsk_af_ops->net_header_len;
1600 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1601 icsk->icsk_mtup.probe_size = 0;
1602 if (icsk->icsk_mtup.enabled)
1603 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1605 EXPORT_SYMBOL(tcp_mtup_init);
1607 /* This function synchronize snd mss to current pmtu/exthdr set.
1609 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1610 for TCP options, but includes only bare TCP header.
1612 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1613 It is minimum of user_mss and mss received with SYN.
1614 It also does not include TCP options.
1616 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1618 tp->mss_cache is current effective sending mss, including
1619 all tcp options except for SACKs. It is evaluated,
1620 taking into account current pmtu, but never exceeds
1621 tp->rx_opt.mss_clamp.
1623 NOTE1. rfc1122 clearly states that advertised MSS
1624 DOES NOT include either tcp or ip options.
1626 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1627 are READ ONLY outside this function. --ANK (980731)
1629 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1631 struct tcp_sock *tp = tcp_sk(sk);
1632 struct inet_connection_sock *icsk = inet_csk(sk);
1635 if (icsk->icsk_mtup.search_high > pmtu)
1636 icsk->icsk_mtup.search_high = pmtu;
1638 mss_now = tcp_mtu_to_mss(sk, pmtu);
1639 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1641 /* And store cached results */
1642 icsk->icsk_pmtu_cookie = pmtu;
1643 if (icsk->icsk_mtup.enabled)
1644 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1645 tp->mss_cache = mss_now;
1649 EXPORT_SYMBOL(tcp_sync_mss);
1651 /* Compute the current effective MSS, taking SACKs and IP options,
1652 * and even PMTU discovery events into account.
1654 unsigned int tcp_current_mss(struct sock *sk)
1656 const struct tcp_sock *tp = tcp_sk(sk);
1657 const struct dst_entry *dst = __sk_dst_get(sk);
1659 unsigned int header_len;
1660 struct tcp_out_options opts;
1661 struct tcp_md5sig_key *md5;
1663 mss_now = tp->mss_cache;
1666 u32 mtu = dst_mtu(dst);
1667 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1668 mss_now = tcp_sync_mss(sk, mtu);
1671 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1672 sizeof(struct tcphdr);
1673 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1674 * some common options. If this is an odd packet (because we have SACK
1675 * blocks etc) then our calculated header_len will be different, and
1676 * we have to adjust mss_now correspondingly */
1677 if (header_len != tp->tcp_header_len) {
1678 int delta = (int) header_len - tp->tcp_header_len;
1685 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1686 * As additional protections, we do not touch cwnd in retransmission phases,
1687 * and if application hit its sndbuf limit recently.
1689 static void tcp_cwnd_application_limited(struct sock *sk)
1691 struct tcp_sock *tp = tcp_sk(sk);
1693 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1694 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1695 /* Limited by application or receiver window. */
1696 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1697 u32 win_used = max(tp->snd_cwnd_used, init_win);
1698 if (win_used < tp->snd_cwnd) {
1699 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1700 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1702 tp->snd_cwnd_used = 0;
1704 tp->snd_cwnd_stamp = tcp_jiffies32;
1707 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1709 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1710 struct tcp_sock *tp = tcp_sk(sk);
1712 /* Track the maximum number of outstanding packets in each
1713 * window, and remember whether we were cwnd-limited then.
1715 if (!before(tp->snd_una, tp->max_packets_seq) ||
1716 tp->packets_out > tp->max_packets_out) {
1717 tp->max_packets_out = tp->packets_out;
1718 tp->max_packets_seq = tp->snd_nxt;
1719 tp->is_cwnd_limited = is_cwnd_limited;
1722 if (tcp_is_cwnd_limited(sk)) {
1723 /* Network is feed fully. */
1724 tp->snd_cwnd_used = 0;
1725 tp->snd_cwnd_stamp = tcp_jiffies32;
1727 /* Network starves. */
1728 if (tp->packets_out > tp->snd_cwnd_used)
1729 tp->snd_cwnd_used = tp->packets_out;
1731 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1732 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1733 !ca_ops->cong_control)
1734 tcp_cwnd_application_limited(sk);
1736 /* The following conditions together indicate the starvation
1737 * is caused by insufficient sender buffer:
1738 * 1) just sent some data (see tcp_write_xmit)
1739 * 2) not cwnd limited (this else condition)
1740 * 3) no more data to send (tcp_write_queue_empty())
1741 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1743 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1744 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1745 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1746 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1750 /* Minshall's variant of the Nagle send check. */
1751 static bool tcp_minshall_check(const struct tcp_sock *tp)
1753 return after(tp->snd_sml, tp->snd_una) &&
1754 !after(tp->snd_sml, tp->snd_nxt);
1757 /* Update snd_sml if this skb is under mss
1758 * Note that a TSO packet might end with a sub-mss segment
1759 * The test is really :
1760 * if ((skb->len % mss) != 0)
1761 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1762 * But we can avoid doing the divide again given we already have
1763 * skb_pcount = skb->len / mss_now
1765 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1766 const struct sk_buff *skb)
1768 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1769 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1772 /* Return false, if packet can be sent now without violation Nagle's rules:
1773 * 1. It is full sized. (provided by caller in %partial bool)
1774 * 2. Or it contains FIN. (already checked by caller)
1775 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1776 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1777 * With Minshall's modification: all sent small packets are ACKed.
1779 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1783 ((nonagle & TCP_NAGLE_CORK) ||
1784 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1787 /* Return how many segs we'd like on a TSO packet,
1788 * to send one TSO packet per ms
1790 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1795 bytes = min_t(unsigned long,
1796 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift),
1797 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1799 /* Goal is to send at least one packet per ms,
1800 * not one big TSO packet every 100 ms.
1801 * This preserves ACK clocking and is consistent
1802 * with tcp_tso_should_defer() heuristic.
1804 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1809 /* Return the number of segments we want in the skb we are transmitting.
1810 * See if congestion control module wants to decide; otherwise, autosize.
1812 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1814 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1815 u32 min_tso, tso_segs;
1817 min_tso = ca_ops->min_tso_segs ?
1818 ca_ops->min_tso_segs(sk) :
1819 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1821 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1822 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1825 /* Returns the portion of skb which can be sent right away */
1826 static unsigned int tcp_mss_split_point(const struct sock *sk,
1827 const struct sk_buff *skb,
1828 unsigned int mss_now,
1829 unsigned int max_segs,
1832 const struct tcp_sock *tp = tcp_sk(sk);
1833 u32 partial, needed, window, max_len;
1835 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1836 max_len = mss_now * max_segs;
1838 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1841 needed = min(skb->len, window);
1843 if (max_len <= needed)
1846 partial = needed % mss_now;
1847 /* If last segment is not a full MSS, check if Nagle rules allow us
1848 * to include this last segment in this skb.
1849 * Otherwise, we'll split the skb at last MSS boundary
1851 if (tcp_nagle_check(partial != 0, tp, nonagle))
1852 return needed - partial;
1857 /* Can at least one segment of SKB be sent right now, according to the
1858 * congestion window rules? If so, return how many segments are allowed.
1860 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1861 const struct sk_buff *skb)
1863 u32 in_flight, cwnd, halfcwnd;
1865 /* Don't be strict about the congestion window for the final FIN. */
1866 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1867 tcp_skb_pcount(skb) == 1)
1870 in_flight = tcp_packets_in_flight(tp);
1871 cwnd = tp->snd_cwnd;
1872 if (in_flight >= cwnd)
1875 /* For better scheduling, ensure we have at least
1876 * 2 GSO packets in flight.
1878 halfcwnd = max(cwnd >> 1, 1U);
1879 return min(halfcwnd, cwnd - in_flight);
1882 /* Initialize TSO state of a skb.
1883 * This must be invoked the first time we consider transmitting
1884 * SKB onto the wire.
1886 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1888 int tso_segs = tcp_skb_pcount(skb);
1890 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1891 tcp_set_skb_tso_segs(skb, mss_now);
1892 tso_segs = tcp_skb_pcount(skb);
1898 /* Return true if the Nagle test allows this packet to be
1901 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1902 unsigned int cur_mss, int nonagle)
1904 /* Nagle rule does not apply to frames, which sit in the middle of the
1905 * write_queue (they have no chances to get new data).
1907 * This is implemented in the callers, where they modify the 'nonagle'
1908 * argument based upon the location of SKB in the send queue.
1910 if (nonagle & TCP_NAGLE_PUSH)
1913 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1914 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1917 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1923 /* Does at least the first segment of SKB fit into the send window? */
1924 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1925 const struct sk_buff *skb,
1926 unsigned int cur_mss)
1928 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1930 if (skb->len > cur_mss)
1931 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1933 return !after(end_seq, tcp_wnd_end(tp));
1936 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1937 * which is put after SKB on the list. It is very much like
1938 * tcp_fragment() except that it may make several kinds of assumptions
1939 * in order to speed up the splitting operation. In particular, we
1940 * know that all the data is in scatter-gather pages, and that the
1941 * packet has never been sent out before (and thus is not cloned).
1943 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1944 unsigned int mss_now, gfp_t gfp)
1946 int nlen = skb->len - len;
1947 struct sk_buff *buff;
1950 /* All of a TSO frame must be composed of paged data. */
1951 if (skb->len != skb->data_len)
1952 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
1953 skb, len, mss_now, gfp);
1955 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1956 if (unlikely(!buff))
1958 skb_copy_decrypted(buff, skb);
1960 sk_wmem_queued_add(sk, buff->truesize);
1961 sk_mem_charge(sk, buff->truesize);
1962 buff->truesize += nlen;
1963 skb->truesize -= nlen;
1965 /* Correct the sequence numbers. */
1966 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1967 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1968 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1970 /* PSH and FIN should only be set in the second packet. */
1971 flags = TCP_SKB_CB(skb)->tcp_flags;
1972 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1973 TCP_SKB_CB(buff)->tcp_flags = flags;
1975 /* This packet was never sent out yet, so no SACK bits. */
1976 TCP_SKB_CB(buff)->sacked = 0;
1978 tcp_skb_fragment_eor(skb, buff);
1980 buff->ip_summed = CHECKSUM_PARTIAL;
1981 skb_split(skb, buff, len);
1982 tcp_fragment_tstamp(skb, buff);
1984 /* Fix up tso_factor for both original and new SKB. */
1985 tcp_set_skb_tso_segs(skb, mss_now);
1986 tcp_set_skb_tso_segs(buff, mss_now);
1988 /* Link BUFF into the send queue. */
1989 __skb_header_release(buff);
1990 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
1995 /* Try to defer sending, if possible, in order to minimize the amount
1996 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1998 * This algorithm is from John Heffner.
2000 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2001 bool *is_cwnd_limited,
2002 bool *is_rwnd_limited,
2005 const struct inet_connection_sock *icsk = inet_csk(sk);
2006 u32 send_win, cong_win, limit, in_flight;
2007 struct tcp_sock *tp = tcp_sk(sk);
2008 struct sk_buff *head;
2012 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2015 /* Avoid bursty behavior by allowing defer
2016 * only if the last write was recent (1 ms).
2017 * Note that tp->tcp_wstamp_ns can be in the future if we have
2018 * packets waiting in a qdisc or device for EDT delivery.
2020 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2024 in_flight = tcp_packets_in_flight(tp);
2026 BUG_ON(tcp_skb_pcount(skb) <= 1);
2027 BUG_ON(tp->snd_cwnd <= in_flight);
2029 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2031 /* From in_flight test above, we know that cwnd > in_flight. */
2032 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
2034 limit = min(send_win, cong_win);
2036 /* If a full-sized TSO skb can be sent, do it. */
2037 if (limit >= max_segs * tp->mss_cache)
2040 /* Middle in queue won't get any more data, full sendable already? */
2041 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2044 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2046 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
2048 /* If at least some fraction of a window is available,
2051 chunk /= win_divisor;
2055 /* Different approach, try not to defer past a single
2056 * ACK. Receiver should ACK every other full sized
2057 * frame, so if we have space for more than 3 frames
2060 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2064 /* TODO : use tsorted_sent_queue ? */
2065 head = tcp_rtx_queue_head(sk);
2068 delta = tp->tcp_clock_cache - head->tstamp;
2069 /* If next ACK is likely to come too late (half srtt), do not defer */
2070 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2073 /* Ok, it looks like it is advisable to defer.
2074 * Three cases are tracked :
2075 * 1) We are cwnd-limited
2076 * 2) We are rwnd-limited
2077 * 3) We are application limited.
2079 if (cong_win < send_win) {
2080 if (cong_win <= skb->len) {
2081 *is_cwnd_limited = true;
2085 if (send_win <= skb->len) {
2086 *is_rwnd_limited = true;
2091 /* If this packet won't get more data, do not wait. */
2092 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2093 TCP_SKB_CB(skb)->eor)
2102 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2104 struct inet_connection_sock *icsk = inet_csk(sk);
2105 struct tcp_sock *tp = tcp_sk(sk);
2106 struct net *net = sock_net(sk);
2110 interval = net->ipv4.sysctl_tcp_probe_interval;
2111 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2112 if (unlikely(delta >= interval * HZ)) {
2113 int mss = tcp_current_mss(sk);
2115 /* Update current search range */
2116 icsk->icsk_mtup.probe_size = 0;
2117 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2118 sizeof(struct tcphdr) +
2119 icsk->icsk_af_ops->net_header_len;
2120 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2122 /* Update probe time stamp */
2123 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2127 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2129 struct sk_buff *skb, *next;
2131 skb = tcp_send_head(sk);
2132 tcp_for_write_queue_from_safe(skb, next, sk) {
2133 if (len <= skb->len)
2136 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2145 /* Create a new MTU probe if we are ready.
2146 * MTU probe is regularly attempting to increase the path MTU by
2147 * deliberately sending larger packets. This discovers routing
2148 * changes resulting in larger path MTUs.
2150 * Returns 0 if we should wait to probe (no cwnd available),
2151 * 1 if a probe was sent,
2154 static int tcp_mtu_probe(struct sock *sk)
2156 struct inet_connection_sock *icsk = inet_csk(sk);
2157 struct tcp_sock *tp = tcp_sk(sk);
2158 struct sk_buff *skb, *nskb, *next;
2159 struct net *net = sock_net(sk);
2166 /* Not currently probing/verifying,
2168 * have enough cwnd, and
2169 * not SACKing (the variable headers throw things off)
2171 if (likely(!icsk->icsk_mtup.enabled ||
2172 icsk->icsk_mtup.probe_size ||
2173 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2174 tp->snd_cwnd < 11 ||
2175 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2178 /* Use binary search for probe_size between tcp_mss_base,
2179 * and current mss_clamp. if (search_high - search_low)
2180 * smaller than a threshold, backoff from probing.
2182 mss_now = tcp_current_mss(sk);
2183 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2184 icsk->icsk_mtup.search_low) >> 1);
2185 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2186 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2187 /* When misfortune happens, we are reprobing actively,
2188 * and then reprobe timer has expired. We stick with current
2189 * probing process by not resetting search range to its orignal.
2191 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2192 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2193 /* Check whether enough time has elaplased for
2194 * another round of probing.
2196 tcp_mtu_check_reprobe(sk);
2200 /* Have enough data in the send queue to probe? */
2201 if (tp->write_seq - tp->snd_nxt < size_needed)
2204 if (tp->snd_wnd < size_needed)
2206 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2209 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2210 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2211 if (!tcp_packets_in_flight(tp))
2217 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2220 /* We're allowed to probe. Build it now. */
2221 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2224 sk_wmem_queued_add(sk, nskb->truesize);
2225 sk_mem_charge(sk, nskb->truesize);
2227 skb = tcp_send_head(sk);
2228 skb_copy_decrypted(nskb, skb);
2230 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2231 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2232 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2233 TCP_SKB_CB(nskb)->sacked = 0;
2235 nskb->ip_summed = CHECKSUM_PARTIAL;
2237 tcp_insert_write_queue_before(nskb, skb, sk);
2238 tcp_highest_sack_replace(sk, skb, nskb);
2241 tcp_for_write_queue_from_safe(skb, next, sk) {
2242 copy = min_t(int, skb->len, probe_size - len);
2243 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2245 if (skb->len <= copy) {
2246 /* We've eaten all the data from this skb.
2248 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2249 /* If this is the last SKB we copy and eor is set
2250 * we need to propagate it to the new skb.
2252 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2253 tcp_skb_collapse_tstamp(nskb, skb);
2254 tcp_unlink_write_queue(skb, sk);
2255 sk_wmem_free_skb(sk, skb);
2257 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2258 ~(TCPHDR_FIN|TCPHDR_PSH);
2259 if (!skb_shinfo(skb)->nr_frags) {
2260 skb_pull(skb, copy);
2262 __pskb_trim_head(skb, copy);
2263 tcp_set_skb_tso_segs(skb, mss_now);
2265 TCP_SKB_CB(skb)->seq += copy;
2270 if (len >= probe_size)
2273 tcp_init_tso_segs(nskb, nskb->len);
2275 /* We're ready to send. If this fails, the probe will
2276 * be resegmented into mss-sized pieces by tcp_write_xmit().
2278 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2279 /* Decrement cwnd here because we are sending
2280 * effectively two packets. */
2282 tcp_event_new_data_sent(sk, nskb);
2284 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2285 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2286 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2294 static bool tcp_pacing_check(struct sock *sk)
2296 struct tcp_sock *tp = tcp_sk(sk);
2298 if (!tcp_needs_internal_pacing(sk))
2301 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2304 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2305 hrtimer_start(&tp->pacing_timer,
2306 ns_to_ktime(tp->tcp_wstamp_ns),
2307 HRTIMER_MODE_ABS_PINNED_SOFT);
2313 /* TCP Small Queues :
2314 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2315 * (These limits are doubled for retransmits)
2317 * - better RTT estimation and ACK scheduling
2320 * Alas, some drivers / subsystems require a fair amount
2321 * of queued bytes to ensure line rate.
2322 * One example is wifi aggregation (802.11 AMPDU)
2324 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2325 unsigned int factor)
2327 unsigned long limit;
2329 limit = max_t(unsigned long,
2331 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2332 if (sk->sk_pacing_status == SK_PACING_NONE)
2333 limit = min_t(unsigned long, limit,
2334 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2337 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2338 tcp_sk(sk)->tcp_tx_delay) {
2339 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2341 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2342 * approximate our needs assuming an ~100% skb->truesize overhead.
2343 * USEC_PER_SEC is approximated by 2^20.
2344 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2346 extra_bytes >>= (20 - 1);
2347 limit += extra_bytes;
2349 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2350 /* Always send skb if rtx queue is empty.
2351 * No need to wait for TX completion to call us back,
2352 * after softirq/tasklet schedule.
2353 * This helps when TX completions are delayed too much.
2355 if (tcp_rtx_queue_empty(sk))
2358 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2359 /* It is possible TX completion already happened
2360 * before we set TSQ_THROTTLED, so we must
2361 * test again the condition.
2363 smp_mb__after_atomic();
2364 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2370 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2372 const u32 now = tcp_jiffies32;
2373 enum tcp_chrono old = tp->chrono_type;
2375 if (old > TCP_CHRONO_UNSPEC)
2376 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2377 tp->chrono_start = now;
2378 tp->chrono_type = new;
2381 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2383 struct tcp_sock *tp = tcp_sk(sk);
2385 /* If there are multiple conditions worthy of tracking in a
2386 * chronograph then the highest priority enum takes precedence
2387 * over the other conditions. So that if something "more interesting"
2388 * starts happening, stop the previous chrono and start a new one.
2390 if (type > tp->chrono_type)
2391 tcp_chrono_set(tp, type);
2394 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2396 struct tcp_sock *tp = tcp_sk(sk);
2399 /* There are multiple conditions worthy of tracking in a
2400 * chronograph, so that the highest priority enum takes
2401 * precedence over the other conditions (see tcp_chrono_start).
2402 * If a condition stops, we only stop chrono tracking if
2403 * it's the "most interesting" or current chrono we are
2404 * tracking and starts busy chrono if we have pending data.
2406 if (tcp_rtx_and_write_queues_empty(sk))
2407 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2408 else if (type == tp->chrono_type)
2409 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2412 /* This routine writes packets to the network. It advances the
2413 * send_head. This happens as incoming acks open up the remote
2416 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2417 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2418 * account rare use of URG, this is not a big flaw.
2420 * Send at most one packet when push_one > 0. Temporarily ignore
2421 * cwnd limit to force at most one packet out when push_one == 2.
2423 * Returns true, if no segments are in flight and we have queued segments,
2424 * but cannot send anything now because of SWS or another problem.
2426 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2427 int push_one, gfp_t gfp)
2429 struct tcp_sock *tp = tcp_sk(sk);
2430 struct sk_buff *skb;
2431 unsigned int tso_segs, sent_pkts;
2434 bool is_cwnd_limited = false, is_rwnd_limited = false;
2439 tcp_mstamp_refresh(tp);
2441 /* Do MTU probing. */
2442 result = tcp_mtu_probe(sk);
2445 } else if (result > 0) {
2450 max_segs = tcp_tso_segs(sk, mss_now);
2451 while ((skb = tcp_send_head(sk))) {
2454 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2455 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2456 skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2457 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2458 tcp_init_tso_segs(skb, mss_now);
2459 goto repair; /* Skip network transmission */
2462 if (tcp_pacing_check(sk))
2465 tso_segs = tcp_init_tso_segs(skb, mss_now);
2468 cwnd_quota = tcp_cwnd_test(tp, skb);
2471 /* Force out a loss probe pkt. */
2477 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2478 is_rwnd_limited = true;
2482 if (tso_segs == 1) {
2483 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2484 (tcp_skb_is_last(sk, skb) ?
2485 nonagle : TCP_NAGLE_PUSH))))
2489 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2490 &is_rwnd_limited, max_segs))
2495 if (tso_segs > 1 && !tcp_urg_mode(tp))
2496 limit = tcp_mss_split_point(sk, skb, mss_now,
2502 if (skb->len > limit &&
2503 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2506 if (tcp_small_queue_check(sk, skb, 0))
2509 /* Argh, we hit an empty skb(), presumably a thread
2510 * is sleeping in sendmsg()/sk_stream_wait_memory().
2511 * We do not want to send a pure-ack packet and have
2512 * a strange looking rtx queue with empty packet(s).
2514 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2517 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2521 /* Advance the send_head. This one is sent out.
2522 * This call will increment packets_out.
2524 tcp_event_new_data_sent(sk, skb);
2526 tcp_minshall_update(tp, mss_now, skb);
2527 sent_pkts += tcp_skb_pcount(skb);
2533 if (is_rwnd_limited)
2534 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2536 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2538 if (likely(sent_pkts)) {
2539 if (tcp_in_cwnd_reduction(sk))
2540 tp->prr_out += sent_pkts;
2542 /* Send one loss probe per tail loss episode. */
2544 tcp_schedule_loss_probe(sk, false);
2545 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2546 tcp_cwnd_validate(sk, is_cwnd_limited);
2549 return !tp->packets_out && !tcp_write_queue_empty(sk);
2552 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2554 struct inet_connection_sock *icsk = inet_csk(sk);
2555 struct tcp_sock *tp = tcp_sk(sk);
2556 u32 timeout, rto_delta_us;
2559 /* Don't do any loss probe on a Fast Open connection before 3WHS
2562 if (rcu_access_pointer(tp->fastopen_rsk))
2565 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2566 /* Schedule a loss probe in 2*RTT for SACK capable connections
2567 * not in loss recovery, that are either limited by cwnd or application.
2569 if ((early_retrans != 3 && early_retrans != 4) ||
2570 !tp->packets_out || !tcp_is_sack(tp) ||
2571 (icsk->icsk_ca_state != TCP_CA_Open &&
2572 icsk->icsk_ca_state != TCP_CA_CWR))
2575 /* Probe timeout is 2*rtt. Add minimum RTO to account
2576 * for delayed ack when there's one outstanding packet. If no RTT
2577 * sample is available then probe after TCP_TIMEOUT_INIT.
2580 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2581 if (tp->packets_out == 1)
2582 timeout += TCP_RTO_MIN;
2584 timeout += TCP_TIMEOUT_MIN;
2586 timeout = TCP_TIMEOUT_INIT;
2589 /* If the RTO formula yields an earlier time, then use that time. */
2590 rto_delta_us = advancing_rto ?
2591 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2592 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2593 if (rto_delta_us > 0)
2594 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2596 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2601 /* Thanks to skb fast clones, we can detect if a prior transmit of
2602 * a packet is still in a qdisc or driver queue.
2603 * In this case, there is very little point doing a retransmit !
2605 static bool skb_still_in_host_queue(const struct sock *sk,
2606 const struct sk_buff *skb)
2608 if (unlikely(skb_fclone_busy(sk, skb))) {
2609 NET_INC_STATS(sock_net(sk),
2610 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2616 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2617 * retransmit the last segment.
2619 void tcp_send_loss_probe(struct sock *sk)
2621 struct tcp_sock *tp = tcp_sk(sk);
2622 struct sk_buff *skb;
2624 int mss = tcp_current_mss(sk);
2626 skb = tcp_send_head(sk);
2627 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2628 pcount = tp->packets_out;
2629 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2630 if (tp->packets_out > pcount)
2634 skb = skb_rb_last(&sk->tcp_rtx_queue);
2635 if (unlikely(!skb)) {
2636 WARN_ONCE(tp->packets_out,
2637 "invalid inflight: %u state %u cwnd %u mss %d\n",
2638 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2639 inet_csk(sk)->icsk_pending = 0;
2643 /* At most one outstanding TLP retransmission. */
2644 if (tp->tlp_high_seq)
2647 if (skb_still_in_host_queue(sk, skb))
2650 pcount = tcp_skb_pcount(skb);
2651 if (WARN_ON(!pcount))
2654 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2655 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2656 (pcount - 1) * mss, mss,
2659 skb = skb_rb_next(skb);
2662 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2665 if (__tcp_retransmit_skb(sk, skb, 1))
2668 /* Record snd_nxt for loss detection. */
2669 tp->tlp_high_seq = tp->snd_nxt;
2672 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2673 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2674 inet_csk(sk)->icsk_pending = 0;
2679 /* Push out any pending frames which were held back due to
2680 * TCP_CORK or attempt at coalescing tiny packets.
2681 * The socket must be locked by the caller.
2683 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2686 /* If we are closed, the bytes will have to remain here.
2687 * In time closedown will finish, we empty the write queue and
2688 * all will be happy.
2690 if (unlikely(sk->sk_state == TCP_CLOSE))
2693 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2694 sk_gfp_mask(sk, GFP_ATOMIC)))
2695 tcp_check_probe_timer(sk);
2698 /* Send _single_ skb sitting at the send head. This function requires
2699 * true push pending frames to setup probe timer etc.
2701 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2703 struct sk_buff *skb = tcp_send_head(sk);
2705 BUG_ON(!skb || skb->len < mss_now);
2707 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2710 /* This function returns the amount that we can raise the
2711 * usable window based on the following constraints
2713 * 1. The window can never be shrunk once it is offered (RFC 793)
2714 * 2. We limit memory per socket
2717 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2718 * RECV.NEXT + RCV.WIN fixed until:
2719 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2721 * i.e. don't raise the right edge of the window until you can raise
2722 * it at least MSS bytes.
2724 * Unfortunately, the recommended algorithm breaks header prediction,
2725 * since header prediction assumes th->window stays fixed.
2727 * Strictly speaking, keeping th->window fixed violates the receiver
2728 * side SWS prevention criteria. The problem is that under this rule
2729 * a stream of single byte packets will cause the right side of the
2730 * window to always advance by a single byte.
2732 * Of course, if the sender implements sender side SWS prevention
2733 * then this will not be a problem.
2735 * BSD seems to make the following compromise:
2737 * If the free space is less than the 1/4 of the maximum
2738 * space available and the free space is less than 1/2 mss,
2739 * then set the window to 0.
2740 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2741 * Otherwise, just prevent the window from shrinking
2742 * and from being larger than the largest representable value.
2744 * This prevents incremental opening of the window in the regime
2745 * where TCP is limited by the speed of the reader side taking
2746 * data out of the TCP receive queue. It does nothing about
2747 * those cases where the window is constrained on the sender side
2748 * because the pipeline is full.
2750 * BSD also seems to "accidentally" limit itself to windows that are a
2751 * multiple of MSS, at least until the free space gets quite small.
2752 * This would appear to be a side effect of the mbuf implementation.
2753 * Combining these two algorithms results in the observed behavior
2754 * of having a fixed window size at almost all times.
2756 * Below we obtain similar behavior by forcing the offered window to
2757 * a multiple of the mss when it is feasible to do so.
2759 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2760 * Regular options like TIMESTAMP are taken into account.
2762 u32 __tcp_select_window(struct sock *sk)
2764 struct inet_connection_sock *icsk = inet_csk(sk);
2765 struct tcp_sock *tp = tcp_sk(sk);
2766 /* MSS for the peer's data. Previous versions used mss_clamp
2767 * here. I don't know if the value based on our guesses
2768 * of peer's MSS is better for the performance. It's more correct
2769 * but may be worse for the performance because of rcv_mss
2770 * fluctuations. --SAW 1998/11/1
2772 int mss = icsk->icsk_ack.rcv_mss;
2773 int free_space = tcp_space(sk);
2774 int allowed_space = tcp_full_space(sk);
2775 int full_space = min_t(int, tp->window_clamp, allowed_space);
2778 if (unlikely(mss > full_space)) {
2783 if (free_space < (full_space >> 1)) {
2784 icsk->icsk_ack.quick = 0;
2786 if (tcp_under_memory_pressure(sk))
2787 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2790 /* free_space might become our new window, make sure we don't
2791 * increase it due to wscale.
2793 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2795 /* if free space is less than mss estimate, or is below 1/16th
2796 * of the maximum allowed, try to move to zero-window, else
2797 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2798 * new incoming data is dropped due to memory limits.
2799 * With large window, mss test triggers way too late in order
2800 * to announce zero window in time before rmem limit kicks in.
2802 if (free_space < (allowed_space >> 4) || free_space < mss)
2806 if (free_space > tp->rcv_ssthresh)
2807 free_space = tp->rcv_ssthresh;
2809 /* Don't do rounding if we are using window scaling, since the
2810 * scaled window will not line up with the MSS boundary anyway.
2812 if (tp->rx_opt.rcv_wscale) {
2813 window = free_space;
2815 /* Advertise enough space so that it won't get scaled away.
2816 * Import case: prevent zero window announcement if
2817 * 1<<rcv_wscale > mss.
2819 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2821 window = tp->rcv_wnd;
2822 /* Get the largest window that is a nice multiple of mss.
2823 * Window clamp already applied above.
2824 * If our current window offering is within 1 mss of the
2825 * free space we just keep it. This prevents the divide
2826 * and multiply from happening most of the time.
2827 * We also don't do any window rounding when the free space
2830 if (window <= free_space - mss || window > free_space)
2831 window = rounddown(free_space, mss);
2832 else if (mss == full_space &&
2833 free_space > window + (full_space >> 1))
2834 window = free_space;
2840 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2841 const struct sk_buff *next_skb)
2843 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2844 const struct skb_shared_info *next_shinfo =
2845 skb_shinfo(next_skb);
2846 struct skb_shared_info *shinfo = skb_shinfo(skb);
2848 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2849 shinfo->tskey = next_shinfo->tskey;
2850 TCP_SKB_CB(skb)->txstamp_ack |=
2851 TCP_SKB_CB(next_skb)->txstamp_ack;
2855 /* Collapses two adjacent SKB's during retransmission. */
2856 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2858 struct tcp_sock *tp = tcp_sk(sk);
2859 struct sk_buff *next_skb = skb_rb_next(skb);
2862 next_skb_size = next_skb->len;
2864 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2866 if (next_skb_size) {
2867 if (next_skb_size <= skb_availroom(skb))
2868 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2870 else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
2873 tcp_highest_sack_replace(sk, next_skb, skb);
2875 /* Update sequence range on original skb. */
2876 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2878 /* Merge over control information. This moves PSH/FIN etc. over */
2879 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2881 /* All done, get rid of second SKB and account for it so
2882 * packet counting does not break.
2884 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2885 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2887 /* changed transmit queue under us so clear hints */
2888 tcp_clear_retrans_hints_partial(tp);
2889 if (next_skb == tp->retransmit_skb_hint)
2890 tp->retransmit_skb_hint = skb;
2892 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2894 tcp_skb_collapse_tstamp(skb, next_skb);
2896 tcp_rtx_queue_unlink_and_free(next_skb, sk);
2900 /* Check if coalescing SKBs is legal. */
2901 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2903 if (tcp_skb_pcount(skb) > 1)
2905 if (skb_cloned(skb))
2907 /* Some heuristics for collapsing over SACK'd could be invented */
2908 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2914 /* Collapse packets in the retransmit queue to make to create
2915 * less packets on the wire. This is only done on retransmission.
2917 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2920 struct tcp_sock *tp = tcp_sk(sk);
2921 struct sk_buff *skb = to, *tmp;
2924 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2926 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2929 skb_rbtree_walk_from_safe(skb, tmp) {
2930 if (!tcp_can_collapse(sk, skb))
2933 if (!tcp_skb_can_collapse(to, skb))
2946 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2949 if (!tcp_collapse_retrans(sk, to))
2954 /* This retransmits one SKB. Policy decisions and retransmit queue
2955 * state updates are done by the caller. Returns non-zero if an
2956 * error occurred which prevented the send.
2958 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2960 struct inet_connection_sock *icsk = inet_csk(sk);
2961 struct tcp_sock *tp = tcp_sk(sk);
2962 unsigned int cur_mss;
2966 /* Inconclusive MTU probe */
2967 if (icsk->icsk_mtup.probe_size)
2968 icsk->icsk_mtup.probe_size = 0;
2970 /* Do not sent more than we queued. 1/4 is reserved for possible
2971 * copying overhead: fragmentation, tunneling, mangling etc.
2973 if (refcount_read(&sk->sk_wmem_alloc) >
2974 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2978 if (skb_still_in_host_queue(sk, skb))
2981 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2982 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2986 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2990 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2991 return -EHOSTUNREACH; /* Routing failure or similar. */
2993 cur_mss = tcp_current_mss(sk);
2995 /* If receiver has shrunk his window, and skb is out of
2996 * new window, do not retransmit it. The exception is the
2997 * case, when window is shrunk to zero. In this case
2998 * our retransmit serves as a zero window probe.
3000 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
3001 TCP_SKB_CB(skb)->seq != tp->snd_una)
3004 len = cur_mss * segs;
3005 if (skb->len > len) {
3006 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3007 cur_mss, GFP_ATOMIC))
3008 return -ENOMEM; /* We'll try again later. */
3010 if (skb_unclone(skb, GFP_ATOMIC))
3013 diff = tcp_skb_pcount(skb);
3014 tcp_set_skb_tso_segs(skb, cur_mss);
3015 diff -= tcp_skb_pcount(skb);
3017 tcp_adjust_pcount(sk, skb, diff);
3018 if (skb->len < cur_mss)
3019 tcp_retrans_try_collapse(sk, skb, cur_mss);
3022 /* RFC3168, section 6.1.1.1. ECN fallback */
3023 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3024 tcp_ecn_clear_syn(sk, skb);
3026 /* Update global and local TCP statistics. */
3027 segs = tcp_skb_pcount(skb);
3028 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3029 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3030 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3031 tp->total_retrans += segs;
3032 tp->bytes_retrans += skb->len;
3034 /* make sure skb->data is aligned on arches that require it
3035 * and check if ack-trimming & collapsing extended the headroom
3036 * beyond what csum_start can cover.
3038 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3039 skb_headroom(skb) >= 0xFFFF)) {
3040 struct sk_buff *nskb;
3042 tcp_skb_tsorted_save(skb) {
3043 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3046 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3050 } tcp_skb_tsorted_restore(skb);
3053 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3054 tcp_rate_skb_sent(sk, skb);
3057 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3060 /* To avoid taking spuriously low RTT samples based on a timestamp
3061 * for a transmit that never happened, always mark EVER_RETRANS
3063 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3065 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3066 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3067 TCP_SKB_CB(skb)->seq, segs, err);
3070 trace_tcp_retransmit_skb(sk, skb);
3071 } else if (err != -EBUSY) {
3072 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3077 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3079 struct tcp_sock *tp = tcp_sk(sk);
3080 int err = __tcp_retransmit_skb(sk, skb, segs);
3083 #if FASTRETRANS_DEBUG > 0
3084 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3085 net_dbg_ratelimited("retrans_out leaked\n");
3088 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3089 tp->retrans_out += tcp_skb_pcount(skb);
3092 /* Save stamp of the first (attempted) retransmit. */
3093 if (!tp->retrans_stamp)
3094 tp->retrans_stamp = tcp_skb_timestamp(skb);
3096 if (tp->undo_retrans < 0)
3097 tp->undo_retrans = 0;
3098 tp->undo_retrans += tcp_skb_pcount(skb);
3102 /* This gets called after a retransmit timeout, and the initially
3103 * retransmitted data is acknowledged. It tries to continue
3104 * resending the rest of the retransmit queue, until either
3105 * we've sent it all or the congestion window limit is reached.
3107 void tcp_xmit_retransmit_queue(struct sock *sk)
3109 const struct inet_connection_sock *icsk = inet_csk(sk);
3110 struct sk_buff *skb, *rtx_head, *hole = NULL;
3111 struct tcp_sock *tp = tcp_sk(sk);
3115 if (!tp->packets_out)
3118 rtx_head = tcp_rtx_queue_head(sk);
3119 skb = tp->retransmit_skb_hint ?: rtx_head;
3120 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3121 skb_rbtree_walk_from(skb) {
3125 if (tcp_pacing_check(sk))
3128 /* we could do better than to assign each time */
3130 tp->retransmit_skb_hint = skb;
3132 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
3135 sacked = TCP_SKB_CB(skb)->sacked;
3136 /* In case tcp_shift_skb_data() have aggregated large skbs,
3137 * we need to make sure not sending too bigs TSO packets
3139 segs = min_t(int, segs, max_segs);
3141 if (tp->retrans_out >= tp->lost_out) {
3143 } else if (!(sacked & TCPCB_LOST)) {
3144 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3149 if (icsk->icsk_ca_state != TCP_CA_Loss)
3150 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3152 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3155 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3158 if (tcp_small_queue_check(sk, skb, 1))
3161 if (tcp_retransmit_skb(sk, skb, segs))
3164 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3166 if (tcp_in_cwnd_reduction(sk))
3167 tp->prr_out += tcp_skb_pcount(skb);
3169 if (skb == rtx_head &&
3170 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3171 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3172 inet_csk(sk)->icsk_rto,
3178 /* We allow to exceed memory limits for FIN packets to expedite
3179 * connection tear down and (memory) recovery.
3180 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3181 * or even be forced to close flow without any FIN.
3182 * In general, we want to allow one skb per socket to avoid hangs
3183 * with edge trigger epoll()
3185 void sk_forced_mem_schedule(struct sock *sk, int size)
3189 if (size <= sk->sk_forward_alloc)
3191 amt = sk_mem_pages(size);
3192 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3193 sk_memory_allocated_add(sk, amt);
3195 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3196 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3199 /* Send a FIN. The caller locks the socket for us.
3200 * We should try to send a FIN packet really hard, but eventually give up.
3202 void tcp_send_fin(struct sock *sk)
3204 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3205 struct tcp_sock *tp = tcp_sk(sk);
3207 /* Optimization, tack on the FIN if we have one skb in write queue and
3208 * this skb was not yet sent, or we are under memory pressure.
3209 * Note: in the latter case, FIN packet will be sent after a timeout,
3210 * as TCP stack thinks it has already been transmitted.
3213 if (!tskb && tcp_under_memory_pressure(sk))
3214 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3217 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3218 TCP_SKB_CB(tskb)->end_seq++;
3221 /* This means tskb was already sent.
3222 * Pretend we included the FIN on previous transmit.
3223 * We need to set tp->snd_nxt to the value it would have
3224 * if FIN had been sent. This is because retransmit path
3225 * does not change tp->snd_nxt.
3227 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3231 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3235 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3236 skb_reserve(skb, MAX_TCP_HEADER);
3237 sk_forced_mem_schedule(sk, skb->truesize);
3238 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3239 tcp_init_nondata_skb(skb, tp->write_seq,
3240 TCPHDR_ACK | TCPHDR_FIN);
3241 tcp_queue_skb(sk, skb);
3243 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3246 /* We get here when a process closes a file descriptor (either due to
3247 * an explicit close() or as a byproduct of exit()'ing) and there
3248 * was unread data in the receive queue. This behavior is recommended
3249 * by RFC 2525, section 2.17. -DaveM
3251 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3253 struct sk_buff *skb;
3255 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3257 /* NOTE: No TCP options attached and we never retransmit this. */
3258 skb = alloc_skb(MAX_TCP_HEADER, priority);
3260 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3264 /* Reserve space for headers and prepare control bits. */
3265 skb_reserve(skb, MAX_TCP_HEADER);
3266 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3267 TCPHDR_ACK | TCPHDR_RST);
3268 tcp_mstamp_refresh(tcp_sk(sk));
3270 if (tcp_transmit_skb(sk, skb, 0, priority))
3271 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3273 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3274 * skb here is different to the troublesome skb, so use NULL
3276 trace_tcp_send_reset(sk, NULL);
3279 /* Send a crossed SYN-ACK during socket establishment.
3280 * WARNING: This routine must only be called when we have already sent
3281 * a SYN packet that crossed the incoming SYN that caused this routine
3282 * to get called. If this assumption fails then the initial rcv_wnd
3283 * and rcv_wscale values will not be correct.
3285 int tcp_send_synack(struct sock *sk)
3287 struct sk_buff *skb;
3289 skb = tcp_rtx_queue_head(sk);
3290 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3291 pr_err("%s: wrong queue state\n", __func__);
3294 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3295 if (skb_cloned(skb)) {
3296 struct sk_buff *nskb;
3298 tcp_skb_tsorted_save(skb) {
3299 nskb = skb_copy(skb, GFP_ATOMIC);
3300 } tcp_skb_tsorted_restore(skb);
3303 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3304 tcp_highest_sack_replace(sk, skb, nskb);
3305 tcp_rtx_queue_unlink_and_free(skb, sk);
3306 __skb_header_release(nskb);
3307 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3308 sk_wmem_queued_add(sk, nskb->truesize);
3309 sk_mem_charge(sk, nskb->truesize);
3313 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3314 tcp_ecn_send_synack(sk, skb);
3316 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3320 * tcp_make_synack - Prepare a SYN-ACK.
3321 * sk: listener socket
3322 * dst: dst entry attached to the SYNACK
3323 * req: request_sock pointer
3325 * Allocate one skb and build a SYNACK packet.
3326 * @dst is consumed : Caller should not use it again.
3328 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3329 struct request_sock *req,
3330 struct tcp_fastopen_cookie *foc,
3331 enum tcp_synack_type synack_type)
3333 struct inet_request_sock *ireq = inet_rsk(req);
3334 const struct tcp_sock *tp = tcp_sk(sk);
3335 struct tcp_md5sig_key *md5 = NULL;
3336 struct tcp_out_options opts;
3337 struct sk_buff *skb;
3338 int tcp_header_size;
3343 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3344 if (unlikely(!skb)) {
3348 /* Reserve space for headers. */
3349 skb_reserve(skb, MAX_TCP_HEADER);
3351 switch (synack_type) {
3352 case TCP_SYNACK_NORMAL:
3353 skb_set_owner_w(skb, req_to_sk(req));
3355 case TCP_SYNACK_COOKIE:
3356 /* Under synflood, we do not attach skb to a socket,
3357 * to avoid false sharing.
3360 case TCP_SYNACK_FASTOPEN:
3361 /* sk is a const pointer, because we want to express multiple
3362 * cpu might call us concurrently.
3363 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3365 skb_set_owner_w(skb, (struct sock *)sk);
3368 skb_dst_set(skb, dst);
3370 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3372 memset(&opts, 0, sizeof(opts));
3373 now = tcp_clock_ns();
3374 #ifdef CONFIG_SYN_COOKIES
3375 if (unlikely(req->cookie_ts))
3376 skb->skb_mstamp_ns = cookie_init_timestamp(req, now);
3380 skb->skb_mstamp_ns = now;
3381 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3382 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3385 #ifdef CONFIG_TCP_MD5SIG
3387 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3389 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3390 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3393 skb_push(skb, tcp_header_size);
3394 skb_reset_transport_header(skb);
3396 th = (struct tcphdr *)skb->data;
3397 memset(th, 0, sizeof(struct tcphdr));
3400 tcp_ecn_make_synack(req, th);
3401 th->source = htons(ireq->ir_num);
3402 th->dest = ireq->ir_rmt_port;
3403 skb->mark = ireq->ir_mark;
3404 skb->ip_summed = CHECKSUM_PARTIAL;
3405 th->seq = htonl(tcp_rsk(req)->snt_isn);
3406 /* XXX data is queued and acked as is. No buffer/window check */
3407 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3409 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3410 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3411 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3412 th->doff = (tcp_header_size >> 2);
3413 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3415 #ifdef CONFIG_TCP_MD5SIG
3416 /* Okay, we have all we need - do the md5 hash if needed */
3418 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3419 md5, req_to_sk(req), skb);
3423 skb->skb_mstamp_ns = now;
3424 tcp_add_tx_delay(skb, tp);
3428 EXPORT_SYMBOL(tcp_make_synack);
3430 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3432 struct inet_connection_sock *icsk = inet_csk(sk);
3433 const struct tcp_congestion_ops *ca;
3434 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3436 if (ca_key == TCP_CA_UNSPEC)
3440 ca = tcp_ca_find_key(ca_key);
3441 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3442 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3443 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3444 icsk->icsk_ca_ops = ca;
3449 /* Do all connect socket setups that can be done AF independent. */
3450 static void tcp_connect_init(struct sock *sk)
3452 const struct dst_entry *dst = __sk_dst_get(sk);
3453 struct tcp_sock *tp = tcp_sk(sk);
3457 /* We'll fix this up when we get a response from the other end.
3458 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3460 tp->tcp_header_len = sizeof(struct tcphdr);
3461 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3462 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3464 #ifdef CONFIG_TCP_MD5SIG
3465 if (tp->af_specific->md5_lookup(sk, sk))
3466 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3469 /* If user gave his TCP_MAXSEG, record it to clamp */
3470 if (tp->rx_opt.user_mss)
3471 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3474 tcp_sync_mss(sk, dst_mtu(dst));
3476 tcp_ca_dst_init(sk, dst);
3478 if (!tp->window_clamp)
3479 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3480 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3482 tcp_initialize_rcv_mss(sk);
3484 /* limit the window selection if the user enforce a smaller rx buffer */
3485 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3486 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3487 tp->window_clamp = tcp_full_space(sk);
3489 rcv_wnd = tcp_rwnd_init_bpf(sk);
3491 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3493 tcp_select_initial_window(sk, tcp_full_space(sk),
3494 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3497 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3501 tp->rx_opt.rcv_wscale = rcv_wscale;
3502 tp->rcv_ssthresh = tp->rcv_wnd;
3505 sock_reset_flag(sk, SOCK_DONE);
3508 tcp_write_queue_purge(sk);
3509 tp->snd_una = tp->write_seq;
3510 tp->snd_sml = tp->write_seq;
3511 tp->snd_up = tp->write_seq;
3512 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3514 if (likely(!tp->repair))
3517 tp->rcv_tstamp = tcp_jiffies32;
3518 tp->rcv_wup = tp->rcv_nxt;
3519 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3521 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3522 inet_csk(sk)->icsk_retransmits = 0;
3523 tcp_clear_retrans(tp);
3526 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3528 struct tcp_sock *tp = tcp_sk(sk);
3529 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3531 tcb->end_seq += skb->len;
3532 __skb_header_release(skb);
3533 sk_wmem_queued_add(sk, skb->truesize);
3534 sk_mem_charge(sk, skb->truesize);
3535 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3536 tp->packets_out += tcp_skb_pcount(skb);
3539 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3540 * queue a data-only packet after the regular SYN, such that regular SYNs
3541 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3542 * only the SYN sequence, the data are retransmitted in the first ACK.
3543 * If cookie is not cached or other error occurs, falls back to send a
3544 * regular SYN with Fast Open cookie request option.
3546 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3548 struct tcp_sock *tp = tcp_sk(sk);
3549 struct tcp_fastopen_request *fo = tp->fastopen_req;
3551 struct sk_buff *syn_data;
3553 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3554 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3557 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3558 * user-MSS. Reserve maximum option space for middleboxes that add
3559 * private TCP options. The cost is reduced data space in SYN :(
3561 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3563 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3564 MAX_TCP_OPTION_SPACE;
3566 space = min_t(size_t, space, fo->size);
3568 /* limit to order-0 allocations */
3569 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3571 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3574 syn_data->ip_summed = CHECKSUM_PARTIAL;
3575 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3577 int copied = copy_from_iter(skb_put(syn_data, space), space,
3578 &fo->data->msg_iter);
3579 if (unlikely(!copied)) {
3580 tcp_skb_tsorted_anchor_cleanup(syn_data);
3581 kfree_skb(syn_data);
3584 if (copied != space) {
3585 skb_trim(syn_data, copied);
3588 skb_zcopy_set(syn_data, fo->uarg, NULL);
3590 /* No more data pending in inet_wait_for_connect() */
3591 if (space == fo->size)
3595 tcp_connect_queue_skb(sk, syn_data);
3597 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3599 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3601 syn->skb_mstamp_ns = syn_data->skb_mstamp_ns;
3603 /* Now full SYN+DATA was cloned and sent (or not),
3604 * remove the SYN from the original skb (syn_data)
3605 * we keep in write queue in case of a retransmit, as we
3606 * also have the SYN packet (with no data) in the same queue.
3608 TCP_SKB_CB(syn_data)->seq++;
3609 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3611 tp->syn_data = (fo->copied > 0);
3612 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3613 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3617 /* data was not sent, put it in write_queue */
3618 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3619 tp->packets_out -= tcp_skb_pcount(syn_data);
3622 /* Send a regular SYN with Fast Open cookie request option */
3623 if (fo->cookie.len > 0)
3625 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3627 tp->syn_fastopen = 0;
3629 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3633 /* Build a SYN and send it off. */
3634 int tcp_connect(struct sock *sk)
3636 struct tcp_sock *tp = tcp_sk(sk);
3637 struct sk_buff *buff;
3640 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3642 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3643 return -EHOSTUNREACH; /* Routing failure or similar. */
3645 tcp_connect_init(sk);
3647 if (unlikely(tp->repair)) {
3648 tcp_finish_connect(sk, NULL);
3652 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3653 if (unlikely(!buff))
3656 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3657 tcp_mstamp_refresh(tp);
3658 tp->retrans_stamp = tcp_time_stamp(tp);
3659 tcp_connect_queue_skb(sk, buff);
3660 tcp_ecn_send_syn(sk, buff);
3661 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3663 /* Send off SYN; include data in Fast Open. */
3664 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3665 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3666 if (err == -ECONNREFUSED)
3669 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3670 * in order to make this packet get counted in tcpOutSegs.
3672 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3673 tp->pushed_seq = tp->write_seq;
3674 buff = tcp_send_head(sk);
3675 if (unlikely(buff)) {
3676 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3677 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3679 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3681 /* Timer for repeating the SYN until an answer. */
3682 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3683 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3686 EXPORT_SYMBOL(tcp_connect);
3688 /* Send out a delayed ack, the caller does the policy checking
3689 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3692 void tcp_send_delayed_ack(struct sock *sk)
3694 struct inet_connection_sock *icsk = inet_csk(sk);
3695 int ato = icsk->icsk_ack.ato;
3696 unsigned long timeout;
3698 if (ato > TCP_DELACK_MIN) {
3699 const struct tcp_sock *tp = tcp_sk(sk);
3700 int max_ato = HZ / 2;
3702 if (inet_csk_in_pingpong_mode(sk) ||
3703 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3704 max_ato = TCP_DELACK_MAX;
3706 /* Slow path, intersegment interval is "high". */
3708 /* If some rtt estimate is known, use it to bound delayed ack.
3709 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3713 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3720 ato = min(ato, max_ato);
3723 /* Stay within the limit we were given */
3724 timeout = jiffies + ato;
3726 /* Use new timeout only if there wasn't a older one earlier. */
3727 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3728 /* If delack timer was blocked or is about to expire,
3731 if (icsk->icsk_ack.blocked ||
3732 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3737 if (!time_before(timeout, icsk->icsk_ack.timeout))
3738 timeout = icsk->icsk_ack.timeout;
3740 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3741 icsk->icsk_ack.timeout = timeout;
3742 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3745 /* This routine sends an ack and also updates the window. */
3746 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3748 struct sk_buff *buff;
3750 /* If we have been reset, we may not send again. */
3751 if (sk->sk_state == TCP_CLOSE)
3754 /* We are not putting this on the write queue, so
3755 * tcp_transmit_skb() will set the ownership to this
3758 buff = alloc_skb(MAX_TCP_HEADER,
3759 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3760 if (unlikely(!buff)) {
3761 inet_csk_schedule_ack(sk);
3762 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3763 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3764 TCP_DELACK_MAX, TCP_RTO_MAX);
3768 /* Reserve space for headers and prepare control bits. */
3769 skb_reserve(buff, MAX_TCP_HEADER);
3770 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3772 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3774 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3776 skb_set_tcp_pure_ack(buff);
3778 /* Send it off, this clears delayed acks for us. */
3779 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3781 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3783 void tcp_send_ack(struct sock *sk)
3785 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3788 /* This routine sends a packet with an out of date sequence
3789 * number. It assumes the other end will try to ack it.
3791 * Question: what should we make while urgent mode?
3792 * 4.4BSD forces sending single byte of data. We cannot send
3793 * out of window data, because we have SND.NXT==SND.MAX...
3795 * Current solution: to send TWO zero-length segments in urgent mode:
3796 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3797 * out-of-date with SND.UNA-1 to probe window.
3799 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3801 struct tcp_sock *tp = tcp_sk(sk);
3802 struct sk_buff *skb;
3804 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3805 skb = alloc_skb(MAX_TCP_HEADER,
3806 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3810 /* Reserve space for headers and set control bits. */
3811 skb_reserve(skb, MAX_TCP_HEADER);
3812 /* Use a previous sequence. This should cause the other
3813 * end to send an ack. Don't queue or clone SKB, just
3816 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3817 NET_INC_STATS(sock_net(sk), mib);
3818 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3821 /* Called from setsockopt( ... TCP_REPAIR ) */
3822 void tcp_send_window_probe(struct sock *sk)
3824 if (sk->sk_state == TCP_ESTABLISHED) {
3825 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3826 tcp_mstamp_refresh(tcp_sk(sk));
3827 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3831 /* Initiate keepalive or window probe from timer. */
3832 int tcp_write_wakeup(struct sock *sk, int mib)
3834 struct tcp_sock *tp = tcp_sk(sk);
3835 struct sk_buff *skb;
3837 if (sk->sk_state == TCP_CLOSE)
3840 skb = tcp_send_head(sk);
3841 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3843 unsigned int mss = tcp_current_mss(sk);
3844 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3846 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3847 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3849 /* We are probing the opening of a window
3850 * but the window size is != 0
3851 * must have been a result SWS avoidance ( sender )
3853 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3855 seg_size = min(seg_size, mss);
3856 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3857 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3858 skb, seg_size, mss, GFP_ATOMIC))
3860 } else if (!tcp_skb_pcount(skb))
3861 tcp_set_skb_tso_segs(skb, mss);
3863 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3864 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3866 tcp_event_new_data_sent(sk, skb);
3869 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3870 tcp_xmit_probe_skb(sk, 1, mib);
3871 return tcp_xmit_probe_skb(sk, 0, mib);
3875 /* A window probe timeout has occurred. If window is not closed send
3876 * a partial packet else a zero probe.
3878 void tcp_send_probe0(struct sock *sk)
3880 struct inet_connection_sock *icsk = inet_csk(sk);
3881 struct tcp_sock *tp = tcp_sk(sk);
3882 struct net *net = sock_net(sk);
3883 unsigned long timeout;
3886 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3888 if (tp->packets_out || tcp_write_queue_empty(sk)) {
3889 /* Cancel probe timer, if it is not required. */
3890 icsk->icsk_probes_out = 0;
3891 icsk->icsk_backoff = 0;
3895 icsk->icsk_probes_out++;
3897 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3898 icsk->icsk_backoff++;
3899 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
3901 /* If packet was not sent due to local congestion,
3902 * Let senders fight for local resources conservatively.
3904 timeout = TCP_RESOURCE_PROBE_INTERVAL;
3906 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX, NULL);
3909 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3911 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3915 tcp_rsk(req)->txhash = net_tx_rndhash();
3916 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3918 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3919 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3920 if (unlikely(tcp_passive_fastopen(sk)))
3921 tcp_sk(sk)->total_retrans++;
3922 trace_tcp_retransmit_synack(sk, req);
3926 EXPORT_SYMBOL(tcp_rtx_synack);