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
42 #include <linux/compiler.h>
43 #include <linux/gfp.h>
44 #include <linux/module.h>
45 #include <linux/static_key.h>
47 #include <trace/events/tcp.h>
49 /* Refresh clocks of a TCP socket,
50 * ensuring monotically increasing values.
52 void tcp_mstamp_refresh(struct tcp_sock *tp)
54 u64 val = tcp_clock_ns();
56 tp->tcp_clock_cache = val;
57 tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
60 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
61 int push_one, gfp_t gfp);
63 /* Account for new data that has been sent to the network. */
64 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
66 struct inet_connection_sock *icsk = inet_csk(sk);
67 struct tcp_sock *tp = tcp_sk(sk);
68 unsigned int prior_packets = tp->packets_out;
70 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
72 __skb_unlink(skb, &sk->sk_write_queue);
73 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
75 tp->packets_out += tcp_skb_pcount(skb);
76 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
79 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
83 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
84 * window scaling factor due to loss of precision.
85 * If window has been shrunk, what should we make? It is not clear at all.
86 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
87 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
88 * invalid. OK, let's make this for now:
90 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
92 const struct tcp_sock *tp = tcp_sk(sk);
94 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
95 (tp->rx_opt.wscale_ok &&
96 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
99 return tcp_wnd_end(tp);
102 /* Calculate mss to advertise in SYN segment.
103 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
105 * 1. It is independent of path mtu.
106 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
107 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
108 * attached devices, because some buggy hosts are confused by
110 * 4. We do not make 3, we advertise MSS, calculated from first
111 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
112 * This may be overridden via information stored in routing table.
113 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
114 * probably even Jumbo".
116 static __u16 tcp_advertise_mss(struct sock *sk)
118 struct tcp_sock *tp = tcp_sk(sk);
119 const struct dst_entry *dst = __sk_dst_get(sk);
120 int mss = tp->advmss;
123 unsigned int metric = dst_metric_advmss(dst);
134 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
135 * This is the first part of cwnd validation mechanism.
137 void tcp_cwnd_restart(struct sock *sk, s32 delta)
139 struct tcp_sock *tp = tcp_sk(sk);
140 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
141 u32 cwnd = tp->snd_cwnd;
143 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
145 tp->snd_ssthresh = tcp_current_ssthresh(sk);
146 restart_cwnd = min(restart_cwnd, cwnd);
148 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
150 tp->snd_cwnd = max(cwnd, restart_cwnd);
151 tp->snd_cwnd_stamp = tcp_jiffies32;
152 tp->snd_cwnd_used = 0;
155 /* Congestion state accounting after a packet has been sent. */
156 static void tcp_event_data_sent(struct tcp_sock *tp,
159 struct inet_connection_sock *icsk = inet_csk(sk);
160 const u32 now = tcp_jiffies32;
162 if (tcp_packets_in_flight(tp) == 0)
163 tcp_ca_event(sk, CA_EVENT_TX_START);
165 /* If this is the first data packet sent in response to the
166 * previous received data,
167 * and it is a reply for ato after last received packet,
168 * increase pingpong count.
170 if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) &&
171 (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
172 inet_csk_inc_pingpong_cnt(sk);
177 /* Account for an ACK we sent. */
178 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
181 struct tcp_sock *tp = tcp_sk(sk);
183 if (unlikely(tp->compressed_ack > TCP_FASTRETRANS_THRESH)) {
184 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
185 tp->compressed_ack - TCP_FASTRETRANS_THRESH);
186 tp->compressed_ack = TCP_FASTRETRANS_THRESH;
187 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
191 if (unlikely(rcv_nxt != tp->rcv_nxt))
192 return; /* Special ACK sent by DCTCP to reflect ECN */
193 tcp_dec_quickack_mode(sk, pkts);
194 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
197 /* Determine a window scaling and initial window to offer.
198 * Based on the assumption that the given amount of space
199 * will be offered. Store the results in the tp structure.
200 * NOTE: for smooth operation initial space offering should
201 * be a multiple of mss if possible. We assume here that mss >= 1.
202 * This MUST be enforced by all callers.
204 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
205 __u32 *rcv_wnd, __u32 *window_clamp,
206 int wscale_ok, __u8 *rcv_wscale,
209 unsigned int space = (__space < 0 ? 0 : __space);
211 /* If no clamp set the clamp to the max possible scaled window */
212 if (*window_clamp == 0)
213 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
214 space = min(*window_clamp, space);
216 /* Quantize space offering to a multiple of mss if possible. */
218 space = rounddown(space, mss);
220 /* NOTE: offering an initial window larger than 32767
221 * will break some buggy TCP stacks. If the admin tells us
222 * it is likely we could be speaking with such a buggy stack
223 * we will truncate our initial window offering to 32K-1
224 * unless the remote has sent us a window scaling option,
225 * which we interpret as a sign the remote TCP is not
226 * misinterpreting the window field as a signed quantity.
228 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
229 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
231 (*rcv_wnd) = min_t(u32, space, U16_MAX);
234 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
238 /* Set window scaling on max possible window */
239 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
240 space = max_t(u32, space, sysctl_rmem_max);
241 space = min_t(u32, space, *window_clamp);
242 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
245 /* Set the clamp no higher than max representable value */
246 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
248 EXPORT_SYMBOL(tcp_select_initial_window);
250 /* Chose a new window to advertise, update state in tcp_sock for the
251 * socket, and return result with RFC1323 scaling applied. The return
252 * value can be stuffed directly into th->window for an outgoing
255 static u16 tcp_select_window(struct sock *sk)
257 struct tcp_sock *tp = tcp_sk(sk);
258 u32 old_win = tp->rcv_wnd;
259 u32 cur_win = tcp_receive_window(tp);
260 u32 new_win = __tcp_select_window(sk);
262 /* Never shrink the offered window */
263 if (new_win < cur_win) {
264 /* Danger Will Robinson!
265 * Don't update rcv_wup/rcv_wnd here or else
266 * we will not be able to advertise a zero
267 * window in time. --DaveM
269 * Relax Will Robinson.
272 NET_INC_STATS(sock_net(sk),
273 LINUX_MIB_TCPWANTZEROWINDOWADV);
274 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
276 tp->rcv_wnd = new_win;
277 tp->rcv_wup = tp->rcv_nxt;
279 /* Make sure we do not exceed the maximum possible
282 if (!tp->rx_opt.rcv_wscale &&
283 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
284 new_win = min(new_win, MAX_TCP_WINDOW);
286 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
288 /* RFC1323 scaling applied */
289 new_win >>= tp->rx_opt.rcv_wscale;
291 /* If we advertise zero window, disable fast path. */
295 NET_INC_STATS(sock_net(sk),
296 LINUX_MIB_TCPTOZEROWINDOWADV);
297 } else if (old_win == 0) {
298 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
304 /* Packet ECN state for a SYN-ACK */
305 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
307 const struct tcp_sock *tp = tcp_sk(sk);
309 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
310 if (!(tp->ecn_flags & TCP_ECN_OK))
311 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
312 else if (tcp_ca_needs_ecn(sk) ||
313 tcp_bpf_ca_needs_ecn(sk))
317 /* Packet ECN state for a SYN. */
318 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
320 struct tcp_sock *tp = tcp_sk(sk);
321 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
322 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
323 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
326 const struct dst_entry *dst = __sk_dst_get(sk);
328 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
335 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
336 tp->ecn_flags = TCP_ECN_OK;
337 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
342 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
344 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
345 /* tp->ecn_flags are cleared at a later point in time when
346 * SYN ACK is ultimatively being received.
348 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
352 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
354 if (inet_rsk(req)->ecn_ok)
358 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
361 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
362 struct tcphdr *th, int tcp_header_len)
364 struct tcp_sock *tp = tcp_sk(sk);
366 if (tp->ecn_flags & TCP_ECN_OK) {
367 /* Not-retransmitted data segment: set ECT and inject CWR. */
368 if (skb->len != tcp_header_len &&
369 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
371 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
372 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
374 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
376 } else if (!tcp_ca_needs_ecn(sk)) {
377 /* ACK or retransmitted segment: clear ECT|CE */
378 INET_ECN_dontxmit(sk);
380 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
385 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
386 * auto increment end seqno.
388 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
390 skb->ip_summed = CHECKSUM_PARTIAL;
392 TCP_SKB_CB(skb)->tcp_flags = flags;
393 TCP_SKB_CB(skb)->sacked = 0;
395 tcp_skb_pcount_set(skb, 1);
397 TCP_SKB_CB(skb)->seq = seq;
398 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
400 TCP_SKB_CB(skb)->end_seq = seq;
403 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
405 return tp->snd_una != tp->snd_up;
408 #define OPTION_SACK_ADVERTISE (1 << 0)
409 #define OPTION_TS (1 << 1)
410 #define OPTION_MD5 (1 << 2)
411 #define OPTION_WSCALE (1 << 3)
412 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
413 #define OPTION_SMC (1 << 9)
415 static void smc_options_write(__be32 *ptr, u16 *options)
417 #if IS_ENABLED(CONFIG_SMC)
418 if (static_branch_unlikely(&tcp_have_smc)) {
419 if (unlikely(OPTION_SMC & *options)) {
420 *ptr++ = htonl((TCPOPT_NOP << 24) |
423 (TCPOLEN_EXP_SMC_BASE));
424 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
430 struct tcp_out_options {
431 u16 options; /* bit field of OPTION_* */
432 u16 mss; /* 0 to disable */
433 u8 ws; /* window scale, 0 to disable */
434 u8 num_sack_blocks; /* number of SACK blocks to include */
435 u8 hash_size; /* bytes in hash_location */
436 __u8 *hash_location; /* temporary pointer, overloaded */
437 __u32 tsval, tsecr; /* need to include OPTION_TS */
438 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
441 /* Write previously computed TCP options to the packet.
443 * Beware: Something in the Internet is very sensitive to the ordering of
444 * TCP options, we learned this through the hard way, so be careful here.
445 * Luckily we can at least blame others for their non-compliance but from
446 * inter-operability perspective it seems that we're somewhat stuck with
447 * the ordering which we have been using if we want to keep working with
448 * those broken things (not that it currently hurts anybody as there isn't
449 * particular reason why the ordering would need to be changed).
451 * At least SACK_PERM as the first option is known to lead to a disaster
452 * (but it may well be that other scenarios fail similarly).
454 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
455 struct tcp_out_options *opts)
457 u16 options = opts->options; /* mungable copy */
459 if (unlikely(OPTION_MD5 & options)) {
460 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
461 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
462 /* overload cookie hash location */
463 opts->hash_location = (__u8 *)ptr;
467 if (unlikely(opts->mss)) {
468 *ptr++ = htonl((TCPOPT_MSS << 24) |
469 (TCPOLEN_MSS << 16) |
473 if (likely(OPTION_TS & options)) {
474 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
475 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
476 (TCPOLEN_SACK_PERM << 16) |
477 (TCPOPT_TIMESTAMP << 8) |
479 options &= ~OPTION_SACK_ADVERTISE;
481 *ptr++ = htonl((TCPOPT_NOP << 24) |
483 (TCPOPT_TIMESTAMP << 8) |
486 *ptr++ = htonl(opts->tsval);
487 *ptr++ = htonl(opts->tsecr);
490 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
491 *ptr++ = htonl((TCPOPT_NOP << 24) |
493 (TCPOPT_SACK_PERM << 8) |
497 if (unlikely(OPTION_WSCALE & options)) {
498 *ptr++ = htonl((TCPOPT_NOP << 24) |
499 (TCPOPT_WINDOW << 16) |
500 (TCPOLEN_WINDOW << 8) |
504 if (unlikely(opts->num_sack_blocks)) {
505 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
506 tp->duplicate_sack : tp->selective_acks;
509 *ptr++ = htonl((TCPOPT_NOP << 24) |
512 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
513 TCPOLEN_SACK_PERBLOCK)));
515 for (this_sack = 0; this_sack < opts->num_sack_blocks;
517 *ptr++ = htonl(sp[this_sack].start_seq);
518 *ptr++ = htonl(sp[this_sack].end_seq);
521 tp->rx_opt.dsack = 0;
524 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
525 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
527 u32 len; /* Fast Open option length */
530 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
531 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
532 TCPOPT_FASTOPEN_MAGIC);
533 p += TCPOLEN_EXP_FASTOPEN_BASE;
535 len = TCPOLEN_FASTOPEN_BASE + foc->len;
536 *p++ = TCPOPT_FASTOPEN;
540 memcpy(p, foc->val, foc->len);
541 if ((len & 3) == 2) {
542 p[foc->len] = TCPOPT_NOP;
543 p[foc->len + 1] = TCPOPT_NOP;
545 ptr += (len + 3) >> 2;
548 smc_options_write(ptr, &options);
551 static void smc_set_option(const struct tcp_sock *tp,
552 struct tcp_out_options *opts,
553 unsigned int *remaining)
555 #if IS_ENABLED(CONFIG_SMC)
556 if (static_branch_unlikely(&tcp_have_smc)) {
558 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
559 opts->options |= OPTION_SMC;
560 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
567 static void smc_set_option_cond(const struct tcp_sock *tp,
568 const struct inet_request_sock *ireq,
569 struct tcp_out_options *opts,
570 unsigned int *remaining)
572 #if IS_ENABLED(CONFIG_SMC)
573 if (static_branch_unlikely(&tcp_have_smc)) {
574 if (tp->syn_smc && ireq->smc_ok) {
575 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
576 opts->options |= OPTION_SMC;
577 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
584 /* Compute TCP options for SYN packets. This is not the final
585 * network wire format yet.
587 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
588 struct tcp_out_options *opts,
589 struct tcp_md5sig_key **md5)
591 struct tcp_sock *tp = tcp_sk(sk);
592 unsigned int remaining = MAX_TCP_OPTION_SPACE;
593 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
596 #ifdef CONFIG_TCP_MD5SIG
597 if (static_branch_unlikely(&tcp_md5_needed) &&
598 rcu_access_pointer(tp->md5sig_info)) {
599 *md5 = tp->af_specific->md5_lookup(sk, sk);
601 opts->options |= OPTION_MD5;
602 remaining -= TCPOLEN_MD5SIG_ALIGNED;
607 /* We always get an MSS option. The option bytes which will be seen in
608 * normal data packets should timestamps be used, must be in the MSS
609 * advertised. But we subtract them from tp->mss_cache so that
610 * calculations in tcp_sendmsg are simpler etc. So account for this
611 * fact here if necessary. If we don't do this correctly, as a
612 * receiver we won't recognize data packets as being full sized when we
613 * should, and thus we won't abide by the delayed ACK rules correctly.
614 * SACKs don't matter, we never delay an ACK when we have any of those
616 opts->mss = tcp_advertise_mss(sk);
617 remaining -= TCPOLEN_MSS_ALIGNED;
619 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
620 opts->options |= OPTION_TS;
621 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
622 opts->tsecr = tp->rx_opt.ts_recent;
623 remaining -= TCPOLEN_TSTAMP_ALIGNED;
625 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
626 opts->ws = tp->rx_opt.rcv_wscale;
627 opts->options |= OPTION_WSCALE;
628 remaining -= TCPOLEN_WSCALE_ALIGNED;
630 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
631 opts->options |= OPTION_SACK_ADVERTISE;
632 if (unlikely(!(OPTION_TS & opts->options)))
633 remaining -= TCPOLEN_SACKPERM_ALIGNED;
636 if (fastopen && fastopen->cookie.len >= 0) {
637 u32 need = fastopen->cookie.len;
639 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
640 TCPOLEN_FASTOPEN_BASE;
641 need = (need + 3) & ~3U; /* Align to 32 bits */
642 if (remaining >= need) {
643 opts->options |= OPTION_FAST_OPEN_COOKIE;
644 opts->fastopen_cookie = &fastopen->cookie;
646 tp->syn_fastopen = 1;
647 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
651 smc_set_option(tp, opts, &remaining);
653 return MAX_TCP_OPTION_SPACE - remaining;
656 /* Set up TCP options for SYN-ACKs. */
657 static unsigned int tcp_synack_options(const struct sock *sk,
658 struct request_sock *req,
659 unsigned int mss, struct sk_buff *skb,
660 struct tcp_out_options *opts,
661 const struct tcp_md5sig_key *md5,
662 struct tcp_fastopen_cookie *foc)
664 struct inet_request_sock *ireq = inet_rsk(req);
665 unsigned int remaining = MAX_TCP_OPTION_SPACE;
667 #ifdef CONFIG_TCP_MD5SIG
669 opts->options |= OPTION_MD5;
670 remaining -= TCPOLEN_MD5SIG_ALIGNED;
672 /* We can't fit any SACK blocks in a packet with MD5 + TS
673 * options. There was discussion about disabling SACK
674 * rather than TS in order to fit in better with old,
675 * buggy kernels, but that was deemed to be unnecessary.
677 ireq->tstamp_ok &= !ireq->sack_ok;
681 /* We always send an MSS option. */
683 remaining -= TCPOLEN_MSS_ALIGNED;
685 if (likely(ireq->wscale_ok)) {
686 opts->ws = ireq->rcv_wscale;
687 opts->options |= OPTION_WSCALE;
688 remaining -= TCPOLEN_WSCALE_ALIGNED;
690 if (likely(ireq->tstamp_ok)) {
691 opts->options |= OPTION_TS;
692 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
693 opts->tsecr = req->ts_recent;
694 remaining -= TCPOLEN_TSTAMP_ALIGNED;
696 if (likely(ireq->sack_ok)) {
697 opts->options |= OPTION_SACK_ADVERTISE;
698 if (unlikely(!ireq->tstamp_ok))
699 remaining -= TCPOLEN_SACKPERM_ALIGNED;
701 if (foc != NULL && foc->len >= 0) {
704 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
705 TCPOLEN_FASTOPEN_BASE;
706 need = (need + 3) & ~3U; /* Align to 32 bits */
707 if (remaining >= need) {
708 opts->options |= OPTION_FAST_OPEN_COOKIE;
709 opts->fastopen_cookie = foc;
714 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
716 return MAX_TCP_OPTION_SPACE - remaining;
719 /* Compute TCP options for ESTABLISHED sockets. This is not the
720 * final wire format yet.
722 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
723 struct tcp_out_options *opts,
724 struct tcp_md5sig_key **md5)
726 struct tcp_sock *tp = tcp_sk(sk);
727 unsigned int size = 0;
728 unsigned int eff_sacks;
733 #ifdef CONFIG_TCP_MD5SIG
734 if (static_branch_unlikely(&tcp_md5_needed) &&
735 rcu_access_pointer(tp->md5sig_info)) {
736 *md5 = tp->af_specific->md5_lookup(sk, sk);
738 opts->options |= OPTION_MD5;
739 size += TCPOLEN_MD5SIG_ALIGNED;
744 if (likely(tp->rx_opt.tstamp_ok)) {
745 opts->options |= OPTION_TS;
746 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
747 opts->tsecr = tp->rx_opt.ts_recent;
748 size += TCPOLEN_TSTAMP_ALIGNED;
751 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
752 if (unlikely(eff_sacks)) {
753 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
754 opts->num_sack_blocks =
755 min_t(unsigned int, eff_sacks,
756 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
757 TCPOLEN_SACK_PERBLOCK);
758 size += TCPOLEN_SACK_BASE_ALIGNED +
759 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
766 /* TCP SMALL QUEUES (TSQ)
768 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
769 * to reduce RTT and bufferbloat.
770 * We do this using a special skb destructor (tcp_wfree).
772 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
773 * needs to be reallocated in a driver.
774 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
776 * Since transmit from skb destructor is forbidden, we use a tasklet
777 * to process all sockets that eventually need to send more skbs.
778 * We use one tasklet per cpu, with its own queue of sockets.
781 struct tasklet_struct tasklet;
782 struct list_head head; /* queue of tcp sockets */
784 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
786 static void tcp_tsq_write(struct sock *sk)
788 if ((1 << sk->sk_state) &
789 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
790 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
791 struct tcp_sock *tp = tcp_sk(sk);
793 if (tp->lost_out > tp->retrans_out &&
794 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
795 tcp_mstamp_refresh(tp);
796 tcp_xmit_retransmit_queue(sk);
799 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
804 static void tcp_tsq_handler(struct sock *sk)
807 if (!sock_owned_by_user(sk))
809 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
814 * One tasklet per cpu tries to send more skbs.
815 * We run in tasklet context but need to disable irqs when
816 * transferring tsq->head because tcp_wfree() might
817 * interrupt us (non NAPI drivers)
819 static void tcp_tasklet_func(unsigned long data)
821 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
824 struct list_head *q, *n;
828 local_irq_save(flags);
829 list_splice_init(&tsq->head, &list);
830 local_irq_restore(flags);
832 list_for_each_safe(q, n, &list) {
833 tp = list_entry(q, struct tcp_sock, tsq_node);
834 list_del(&tp->tsq_node);
836 sk = (struct sock *)tp;
837 smp_mb__before_atomic();
838 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
845 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
846 TCPF_WRITE_TIMER_DEFERRED | \
847 TCPF_DELACK_TIMER_DEFERRED | \
848 TCPF_MTU_REDUCED_DEFERRED)
850 * tcp_release_cb - tcp release_sock() callback
853 * called from release_sock() to perform protocol dependent
854 * actions before socket release.
856 void tcp_release_cb(struct sock *sk)
858 unsigned long flags, nflags;
860 /* perform an atomic operation only if at least one flag is set */
862 flags = sk->sk_tsq_flags;
863 if (!(flags & TCP_DEFERRED_ALL))
865 nflags = flags & ~TCP_DEFERRED_ALL;
866 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
868 if (flags & TCPF_TSQ_DEFERRED) {
872 /* Here begins the tricky part :
873 * We are called from release_sock() with :
875 * 2) sk_lock.slock spinlock held
876 * 3) socket owned by us (sk->sk_lock.owned == 1)
878 * But following code is meant to be called from BH handlers,
879 * so we should keep BH disabled, but early release socket ownership
881 sock_release_ownership(sk);
883 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
884 tcp_write_timer_handler(sk);
887 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
888 tcp_delack_timer_handler(sk);
891 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
892 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
896 EXPORT_SYMBOL(tcp_release_cb);
898 void __init tcp_tasklet_init(void)
902 for_each_possible_cpu(i) {
903 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
905 INIT_LIST_HEAD(&tsq->head);
906 tasklet_init(&tsq->tasklet,
913 * Write buffer destructor automatically called from kfree_skb.
914 * We can't xmit new skbs from this context, as we might already
917 void tcp_wfree(struct sk_buff *skb)
919 struct sock *sk = skb->sk;
920 struct tcp_sock *tp = tcp_sk(sk);
921 unsigned long flags, nval, oval;
923 /* Keep one reference on sk_wmem_alloc.
924 * Will be released by sk_free() from here or tcp_tasklet_func()
926 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
928 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
929 * Wait until our queues (qdisc + devices) are drained.
931 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
932 * - chance for incoming ACK (processed by another cpu maybe)
933 * to migrate this flow (skb->ooo_okay will be eventually set)
935 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
938 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
939 struct tsq_tasklet *tsq;
942 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
945 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
946 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
950 /* queue this socket to tasklet queue */
951 local_irq_save(flags);
952 tsq = this_cpu_ptr(&tsq_tasklet);
953 empty = list_empty(&tsq->head);
954 list_add(&tp->tsq_node, &tsq->head);
956 tasklet_schedule(&tsq->tasklet);
957 local_irq_restore(flags);
964 /* Note: Called under soft irq.
965 * We can call TCP stack right away, unless socket is owned by user.
967 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
969 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
970 struct sock *sk = (struct sock *)tp;
975 return HRTIMER_NORESTART;
978 static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
981 struct tcp_sock *tp = tcp_sk(sk);
983 if (sk->sk_pacing_status != SK_PACING_NONE) {
984 unsigned long rate = sk->sk_pacing_rate;
986 /* Original sch_fq does not pace first 10 MSS
987 * Note that tp->data_segs_out overflows after 2^32 packets,
988 * this is a minor annoyance.
990 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
991 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
992 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
994 /* take into account OS jitter */
995 len_ns -= min_t(u64, len_ns / 2, credit);
996 tp->tcp_wstamp_ns += len_ns;
999 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1002 /* This routine actually transmits TCP packets queued in by
1003 * tcp_do_sendmsg(). This is used by both the initial
1004 * transmission and possible later retransmissions.
1005 * All SKB's seen here are completely headerless. It is our
1006 * job to build the TCP header, and pass the packet down to
1007 * IP so it can do the same plus pass the packet off to the
1010 * We are working here with either a clone of the original
1011 * SKB, or a fresh unique copy made by the retransmit engine.
1013 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1014 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1016 const struct inet_connection_sock *icsk = inet_csk(sk);
1017 struct inet_sock *inet;
1018 struct tcp_sock *tp;
1019 struct tcp_skb_cb *tcb;
1020 struct tcp_out_options opts;
1021 unsigned int tcp_options_size, tcp_header_size;
1022 struct sk_buff *oskb = NULL;
1023 struct tcp_md5sig_key *md5;
1028 BUG_ON(!skb || !tcp_skb_pcount(skb));
1030 prior_wstamp = tp->tcp_wstamp_ns;
1031 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1032 skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
1034 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1038 tcp_skb_tsorted_save(oskb) {
1039 if (unlikely(skb_cloned(oskb)))
1040 skb = pskb_copy(oskb, gfp_mask);
1042 skb = skb_clone(oskb, gfp_mask);
1043 } tcp_skb_tsorted_restore(oskb);
1050 tcb = TCP_SKB_CB(skb);
1051 memset(&opts, 0, sizeof(opts));
1053 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1054 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1056 tcp_options_size = tcp_established_options(sk, skb, &opts,
1058 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1060 /* if no packet is in qdisc/device queue, then allow XPS to select
1061 * another queue. We can be called from tcp_tsq_handler()
1062 * which holds one reference to sk.
1064 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1065 * One way to get this would be to set skb->truesize = 2 on them.
1067 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1069 /* If we had to use memory reserve to allocate this skb,
1070 * this might cause drops if packet is looped back :
1071 * Other socket might not have SOCK_MEMALLOC.
1072 * Packets not looped back do not care about pfmemalloc.
1074 skb->pfmemalloc = 0;
1076 skb_push(skb, tcp_header_size);
1077 skb_reset_transport_header(skb);
1081 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1082 skb_set_hash_from_sk(skb, sk);
1083 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1085 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1087 /* Build TCP header and checksum it. */
1088 th = (struct tcphdr *)skb->data;
1089 th->source = inet->inet_sport;
1090 th->dest = inet->inet_dport;
1091 th->seq = htonl(tcb->seq);
1092 th->ack_seq = htonl(rcv_nxt);
1093 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1099 /* The urg_mode check is necessary during a below snd_una win probe */
1100 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1101 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1102 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1104 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1105 th->urg_ptr = htons(0xFFFF);
1110 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1111 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1112 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1113 th->window = htons(tcp_select_window(sk));
1114 tcp_ecn_send(sk, skb, th, tcp_header_size);
1116 /* RFC1323: The window in SYN & SYN/ACK segments
1119 th->window = htons(min(tp->rcv_wnd, 65535U));
1121 #ifdef CONFIG_TCP_MD5SIG
1122 /* Calculate the MD5 hash, as we have all we need now */
1124 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1125 tp->af_specific->calc_md5_hash(opts.hash_location,
1130 icsk->icsk_af_ops->send_check(sk, skb);
1132 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1133 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1135 if (skb->len != tcp_header_size) {
1136 tcp_event_data_sent(tp, sk);
1137 tp->data_segs_out += tcp_skb_pcount(skb);
1138 tp->bytes_sent += skb->len - tcp_header_size;
1141 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1142 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1143 tcp_skb_pcount(skb));
1145 tp->segs_out += tcp_skb_pcount(skb);
1146 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1147 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1148 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1150 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1152 /* Cleanup our debris for IP stacks */
1153 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1154 sizeof(struct inet6_skb_parm)));
1156 tcp_add_tx_delay(skb, tp);
1158 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1160 if (unlikely(err > 0)) {
1162 err = net_xmit_eval(err);
1165 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1166 tcp_rate_skb_sent(sk, oskb);
1171 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1174 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1175 tcp_sk(sk)->rcv_nxt);
1178 /* This routine just queues the buffer for sending.
1180 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1181 * otherwise socket can stall.
1183 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1185 struct tcp_sock *tp = tcp_sk(sk);
1187 /* Advance write_seq and place onto the write_queue. */
1188 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1189 __skb_header_release(skb);
1190 tcp_add_write_queue_tail(sk, skb);
1191 sk->sk_wmem_queued += skb->truesize;
1192 sk_mem_charge(sk, skb->truesize);
1195 /* Initialize TSO segments for a packet. */
1196 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1198 if (skb->len <= mss_now) {
1199 /* Avoid the costly divide in the normal
1202 tcp_skb_pcount_set(skb, 1);
1203 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1205 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1206 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1210 /* Pcount in the middle of the write queue got changed, we need to do various
1211 * tweaks to fix counters
1213 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1215 struct tcp_sock *tp = tcp_sk(sk);
1217 tp->packets_out -= decr;
1219 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1220 tp->sacked_out -= decr;
1221 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1222 tp->retrans_out -= decr;
1223 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1224 tp->lost_out -= decr;
1226 /* Reno case is special. Sigh... */
1227 if (tcp_is_reno(tp) && decr > 0)
1228 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1230 if (tp->lost_skb_hint &&
1231 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1232 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1233 tp->lost_cnt_hint -= decr;
1235 tcp_verify_left_out(tp);
1238 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1240 return TCP_SKB_CB(skb)->txstamp_ack ||
1241 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1244 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1246 struct skb_shared_info *shinfo = skb_shinfo(skb);
1248 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1249 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1250 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1251 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1253 shinfo->tx_flags &= ~tsflags;
1254 shinfo2->tx_flags |= tsflags;
1255 swap(shinfo->tskey, shinfo2->tskey);
1256 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1257 TCP_SKB_CB(skb)->txstamp_ack = 0;
1261 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1263 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1264 TCP_SKB_CB(skb)->eor = 0;
1267 /* Insert buff after skb on the write or rtx queue of sk. */
1268 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1269 struct sk_buff *buff,
1271 enum tcp_queue tcp_queue)
1273 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1274 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1276 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1279 /* Function to create two new TCP segments. Shrinks the given segment
1280 * to the specified size and appends a new segment with the rest of the
1281 * packet to the list. This won't be called frequently, I hope.
1282 * Remember, these are still headerless SKBs at this point.
1284 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1285 struct sk_buff *skb, u32 len,
1286 unsigned int mss_now, gfp_t gfp)
1288 struct tcp_sock *tp = tcp_sk(sk);
1289 struct sk_buff *buff;
1290 int nsize, old_factor;
1295 if (WARN_ON(len > skb->len))
1298 nsize = skb_headlen(skb) - len;
1302 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1303 * We need some allowance to not penalize applications setting small
1305 * Also allow first and last skb in retransmit queue to be split.
1307 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1308 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1309 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1310 skb != tcp_rtx_queue_head(sk) &&
1311 skb != tcp_rtx_queue_tail(sk))) {
1312 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1316 if (skb_unclone(skb, gfp))
1319 /* Get a new skb... force flag on. */
1320 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1322 return -ENOMEM; /* We'll just try again later. */
1324 sk->sk_wmem_queued += buff->truesize;
1325 sk_mem_charge(sk, buff->truesize);
1326 nlen = skb->len - len - nsize;
1327 buff->truesize += nlen;
1328 skb->truesize -= nlen;
1330 /* Correct the sequence numbers. */
1331 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1332 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1333 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1335 /* PSH and FIN should only be set in the second packet. */
1336 flags = TCP_SKB_CB(skb)->tcp_flags;
1337 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1338 TCP_SKB_CB(buff)->tcp_flags = flags;
1339 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1340 tcp_skb_fragment_eor(skb, buff);
1342 skb_split(skb, buff, len);
1344 buff->ip_summed = CHECKSUM_PARTIAL;
1346 buff->tstamp = skb->tstamp;
1347 tcp_fragment_tstamp(skb, buff);
1349 old_factor = tcp_skb_pcount(skb);
1351 /* Fix up tso_factor for both original and new SKB. */
1352 tcp_set_skb_tso_segs(skb, mss_now);
1353 tcp_set_skb_tso_segs(buff, mss_now);
1355 /* Update delivered info for the new segment */
1356 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1358 /* If this packet has been sent out already, we must
1359 * adjust the various packet counters.
1361 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1362 int diff = old_factor - tcp_skb_pcount(skb) -
1363 tcp_skb_pcount(buff);
1366 tcp_adjust_pcount(sk, skb, diff);
1369 /* Link BUFF into the send queue. */
1370 __skb_header_release(buff);
1371 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1372 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1373 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1378 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1379 * data is not copied, but immediately discarded.
1381 static int __pskb_trim_head(struct sk_buff *skb, int len)
1383 struct skb_shared_info *shinfo;
1386 eat = min_t(int, len, skb_headlen(skb));
1388 __skb_pull(skb, eat);
1395 shinfo = skb_shinfo(skb);
1396 for (i = 0; i < shinfo->nr_frags; i++) {
1397 int size = skb_frag_size(&shinfo->frags[i]);
1400 skb_frag_unref(skb, i);
1403 shinfo->frags[k] = shinfo->frags[i];
1405 shinfo->frags[k].page_offset += eat;
1406 skb_frag_size_sub(&shinfo->frags[k], eat);
1412 shinfo->nr_frags = k;
1414 skb->data_len -= len;
1415 skb->len = skb->data_len;
1419 /* Remove acked data from a packet in the transmit queue. */
1420 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1424 if (skb_unclone(skb, GFP_ATOMIC))
1427 delta_truesize = __pskb_trim_head(skb, len);
1429 TCP_SKB_CB(skb)->seq += len;
1430 skb->ip_summed = CHECKSUM_PARTIAL;
1432 if (delta_truesize) {
1433 skb->truesize -= delta_truesize;
1434 sk->sk_wmem_queued -= delta_truesize;
1435 sk_mem_uncharge(sk, delta_truesize);
1436 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1439 /* Any change of skb->len requires recalculation of tso factor. */
1440 if (tcp_skb_pcount(skb) > 1)
1441 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1446 /* Calculate MSS not accounting any TCP options. */
1447 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1449 const struct tcp_sock *tp = tcp_sk(sk);
1450 const struct inet_connection_sock *icsk = inet_csk(sk);
1453 /* Calculate base mss without TCP options:
1454 It is MMS_S - sizeof(tcphdr) of rfc1122
1456 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1458 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1459 if (icsk->icsk_af_ops->net_frag_header_len) {
1460 const struct dst_entry *dst = __sk_dst_get(sk);
1462 if (dst && dst_allfrag(dst))
1463 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1466 /* Clamp it (mss_clamp does not include tcp options) */
1467 if (mss_now > tp->rx_opt.mss_clamp)
1468 mss_now = tp->rx_opt.mss_clamp;
1470 /* Now subtract optional transport overhead */
1471 mss_now -= icsk->icsk_ext_hdr_len;
1473 /* Then reserve room for full set of TCP options and 8 bytes of data */
1474 mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1478 /* Calculate MSS. Not accounting for SACKs here. */
1479 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1481 /* Subtract TCP options size, not including SACKs */
1482 return __tcp_mtu_to_mss(sk, pmtu) -
1483 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1486 /* Inverse of above */
1487 int tcp_mss_to_mtu(struct sock *sk, int mss)
1489 const struct tcp_sock *tp = tcp_sk(sk);
1490 const struct inet_connection_sock *icsk = inet_csk(sk);
1494 tp->tcp_header_len +
1495 icsk->icsk_ext_hdr_len +
1496 icsk->icsk_af_ops->net_header_len;
1498 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1499 if (icsk->icsk_af_ops->net_frag_header_len) {
1500 const struct dst_entry *dst = __sk_dst_get(sk);
1502 if (dst && dst_allfrag(dst))
1503 mtu += icsk->icsk_af_ops->net_frag_header_len;
1507 EXPORT_SYMBOL(tcp_mss_to_mtu);
1509 /* MTU probing init per socket */
1510 void tcp_mtup_init(struct sock *sk)
1512 struct tcp_sock *tp = tcp_sk(sk);
1513 struct inet_connection_sock *icsk = inet_csk(sk);
1514 struct net *net = sock_net(sk);
1516 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1517 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1518 icsk->icsk_af_ops->net_header_len;
1519 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1520 icsk->icsk_mtup.probe_size = 0;
1521 if (icsk->icsk_mtup.enabled)
1522 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1524 EXPORT_SYMBOL(tcp_mtup_init);
1526 /* This function synchronize snd mss to current pmtu/exthdr set.
1528 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1529 for TCP options, but includes only bare TCP header.
1531 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1532 It is minimum of user_mss and mss received with SYN.
1533 It also does not include TCP options.
1535 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1537 tp->mss_cache is current effective sending mss, including
1538 all tcp options except for SACKs. It is evaluated,
1539 taking into account current pmtu, but never exceeds
1540 tp->rx_opt.mss_clamp.
1542 NOTE1. rfc1122 clearly states that advertised MSS
1543 DOES NOT include either tcp or ip options.
1545 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1546 are READ ONLY outside this function. --ANK (980731)
1548 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1550 struct tcp_sock *tp = tcp_sk(sk);
1551 struct inet_connection_sock *icsk = inet_csk(sk);
1554 if (icsk->icsk_mtup.search_high > pmtu)
1555 icsk->icsk_mtup.search_high = pmtu;
1557 mss_now = tcp_mtu_to_mss(sk, pmtu);
1558 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1560 /* And store cached results */
1561 icsk->icsk_pmtu_cookie = pmtu;
1562 if (icsk->icsk_mtup.enabled)
1563 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1564 tp->mss_cache = mss_now;
1568 EXPORT_SYMBOL(tcp_sync_mss);
1570 /* Compute the current effective MSS, taking SACKs and IP options,
1571 * and even PMTU discovery events into account.
1573 unsigned int tcp_current_mss(struct sock *sk)
1575 const struct tcp_sock *tp = tcp_sk(sk);
1576 const struct dst_entry *dst = __sk_dst_get(sk);
1578 unsigned int header_len;
1579 struct tcp_out_options opts;
1580 struct tcp_md5sig_key *md5;
1582 mss_now = tp->mss_cache;
1585 u32 mtu = dst_mtu(dst);
1586 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1587 mss_now = tcp_sync_mss(sk, mtu);
1590 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1591 sizeof(struct tcphdr);
1592 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1593 * some common options. If this is an odd packet (because we have SACK
1594 * blocks etc) then our calculated header_len will be different, and
1595 * we have to adjust mss_now correspondingly */
1596 if (header_len != tp->tcp_header_len) {
1597 int delta = (int) header_len - tp->tcp_header_len;
1604 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1605 * As additional protections, we do not touch cwnd in retransmission phases,
1606 * and if application hit its sndbuf limit recently.
1608 static void tcp_cwnd_application_limited(struct sock *sk)
1610 struct tcp_sock *tp = tcp_sk(sk);
1612 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1613 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1614 /* Limited by application or receiver window. */
1615 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1616 u32 win_used = max(tp->snd_cwnd_used, init_win);
1617 if (win_used < tp->snd_cwnd) {
1618 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1619 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1621 tp->snd_cwnd_used = 0;
1623 tp->snd_cwnd_stamp = tcp_jiffies32;
1626 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1628 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1629 struct tcp_sock *tp = tcp_sk(sk);
1631 /* Track the maximum number of outstanding packets in each
1632 * window, and remember whether we were cwnd-limited then.
1634 if (!before(tp->snd_una, tp->max_packets_seq) ||
1635 tp->packets_out > tp->max_packets_out) {
1636 tp->max_packets_out = tp->packets_out;
1637 tp->max_packets_seq = tp->snd_nxt;
1638 tp->is_cwnd_limited = is_cwnd_limited;
1641 if (tcp_is_cwnd_limited(sk)) {
1642 /* Network is feed fully. */
1643 tp->snd_cwnd_used = 0;
1644 tp->snd_cwnd_stamp = tcp_jiffies32;
1646 /* Network starves. */
1647 if (tp->packets_out > tp->snd_cwnd_used)
1648 tp->snd_cwnd_used = tp->packets_out;
1650 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1651 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1652 !ca_ops->cong_control)
1653 tcp_cwnd_application_limited(sk);
1655 /* The following conditions together indicate the starvation
1656 * is caused by insufficient sender buffer:
1657 * 1) just sent some data (see tcp_write_xmit)
1658 * 2) not cwnd limited (this else condition)
1659 * 3) no more data to send (tcp_write_queue_empty())
1660 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1662 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1663 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1664 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1665 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1669 /* Minshall's variant of the Nagle send check. */
1670 static bool tcp_minshall_check(const struct tcp_sock *tp)
1672 return after(tp->snd_sml, tp->snd_una) &&
1673 !after(tp->snd_sml, tp->snd_nxt);
1676 /* Update snd_sml if this skb is under mss
1677 * Note that a TSO packet might end with a sub-mss segment
1678 * The test is really :
1679 * if ((skb->len % mss) != 0)
1680 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1681 * But we can avoid doing the divide again given we already have
1682 * skb_pcount = skb->len / mss_now
1684 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1685 const struct sk_buff *skb)
1687 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1688 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1691 /* Return false, if packet can be sent now without violation Nagle's rules:
1692 * 1. It is full sized. (provided by caller in %partial bool)
1693 * 2. Or it contains FIN. (already checked by caller)
1694 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1695 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1696 * With Minshall's modification: all sent small packets are ACKed.
1698 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1702 ((nonagle & TCP_NAGLE_CORK) ||
1703 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1706 /* Return how many segs we'd like on a TSO packet,
1707 * to send one TSO packet per ms
1709 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1714 bytes = min_t(unsigned long,
1715 sk->sk_pacing_rate >> sk->sk_pacing_shift,
1716 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1718 /* Goal is to send at least one packet per ms,
1719 * not one big TSO packet every 100 ms.
1720 * This preserves ACK clocking and is consistent
1721 * with tcp_tso_should_defer() heuristic.
1723 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1728 /* Return the number of segments we want in the skb we are transmitting.
1729 * See if congestion control module wants to decide; otherwise, autosize.
1731 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1733 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1734 u32 min_tso, tso_segs;
1736 min_tso = ca_ops->min_tso_segs ?
1737 ca_ops->min_tso_segs(sk) :
1738 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1740 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1741 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1744 /* Returns the portion of skb which can be sent right away */
1745 static unsigned int tcp_mss_split_point(const struct sock *sk,
1746 const struct sk_buff *skb,
1747 unsigned int mss_now,
1748 unsigned int max_segs,
1751 const struct tcp_sock *tp = tcp_sk(sk);
1752 u32 partial, needed, window, max_len;
1754 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1755 max_len = mss_now * max_segs;
1757 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1760 needed = min(skb->len, window);
1762 if (max_len <= needed)
1765 partial = needed % mss_now;
1766 /* If last segment is not a full MSS, check if Nagle rules allow us
1767 * to include this last segment in this skb.
1768 * Otherwise, we'll split the skb at last MSS boundary
1770 if (tcp_nagle_check(partial != 0, tp, nonagle))
1771 return needed - partial;
1776 /* Can at least one segment of SKB be sent right now, according to the
1777 * congestion window rules? If so, return how many segments are allowed.
1779 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1780 const struct sk_buff *skb)
1782 u32 in_flight, cwnd, halfcwnd;
1784 /* Don't be strict about the congestion window for the final FIN. */
1785 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1786 tcp_skb_pcount(skb) == 1)
1789 in_flight = tcp_packets_in_flight(tp);
1790 cwnd = tp->snd_cwnd;
1791 if (in_flight >= cwnd)
1794 /* For better scheduling, ensure we have at least
1795 * 2 GSO packets in flight.
1797 halfcwnd = max(cwnd >> 1, 1U);
1798 return min(halfcwnd, cwnd - in_flight);
1801 /* Initialize TSO state of a skb.
1802 * This must be invoked the first time we consider transmitting
1803 * SKB onto the wire.
1805 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1807 int tso_segs = tcp_skb_pcount(skb);
1809 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1810 tcp_set_skb_tso_segs(skb, mss_now);
1811 tso_segs = tcp_skb_pcount(skb);
1817 /* Return true if the Nagle test allows this packet to be
1820 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1821 unsigned int cur_mss, int nonagle)
1823 /* Nagle rule does not apply to frames, which sit in the middle of the
1824 * write_queue (they have no chances to get new data).
1826 * This is implemented in the callers, where they modify the 'nonagle'
1827 * argument based upon the location of SKB in the send queue.
1829 if (nonagle & TCP_NAGLE_PUSH)
1832 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1833 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1836 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1842 /* Does at least the first segment of SKB fit into the send window? */
1843 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1844 const struct sk_buff *skb,
1845 unsigned int cur_mss)
1847 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1849 if (skb->len > cur_mss)
1850 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1852 return !after(end_seq, tcp_wnd_end(tp));
1855 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1856 * which is put after SKB on the list. It is very much like
1857 * tcp_fragment() except that it may make several kinds of assumptions
1858 * in order to speed up the splitting operation. In particular, we
1859 * know that all the data is in scatter-gather pages, and that the
1860 * packet has never been sent out before (and thus is not cloned).
1862 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1863 unsigned int mss_now, gfp_t gfp)
1865 int nlen = skb->len - len;
1866 struct sk_buff *buff;
1869 /* All of a TSO frame must be composed of paged data. */
1870 if (skb->len != skb->data_len)
1871 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
1872 skb, len, mss_now, gfp);
1874 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1875 if (unlikely(!buff))
1878 sk->sk_wmem_queued += buff->truesize;
1879 sk_mem_charge(sk, buff->truesize);
1880 buff->truesize += nlen;
1881 skb->truesize -= nlen;
1883 /* Correct the sequence numbers. */
1884 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1885 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1886 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1888 /* PSH and FIN should only be set in the second packet. */
1889 flags = TCP_SKB_CB(skb)->tcp_flags;
1890 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1891 TCP_SKB_CB(buff)->tcp_flags = flags;
1893 /* This packet was never sent out yet, so no SACK bits. */
1894 TCP_SKB_CB(buff)->sacked = 0;
1896 tcp_skb_fragment_eor(skb, buff);
1898 buff->ip_summed = CHECKSUM_PARTIAL;
1899 skb_split(skb, buff, len);
1900 tcp_fragment_tstamp(skb, buff);
1902 /* Fix up tso_factor for both original and new SKB. */
1903 tcp_set_skb_tso_segs(skb, mss_now);
1904 tcp_set_skb_tso_segs(buff, mss_now);
1906 /* Link BUFF into the send queue. */
1907 __skb_header_release(buff);
1908 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
1913 /* Try to defer sending, if possible, in order to minimize the amount
1914 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1916 * This algorithm is from John Heffner.
1918 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1919 bool *is_cwnd_limited,
1920 bool *is_rwnd_limited,
1923 const struct inet_connection_sock *icsk = inet_csk(sk);
1924 u32 send_win, cong_win, limit, in_flight;
1925 struct tcp_sock *tp = tcp_sk(sk);
1926 struct sk_buff *head;
1930 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1933 /* Avoid bursty behavior by allowing defer
1934 * only if the last write was recent (1 ms).
1935 * Note that tp->tcp_wstamp_ns can be in the future if we have
1936 * packets waiting in a qdisc or device for EDT delivery.
1938 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
1942 in_flight = tcp_packets_in_flight(tp);
1944 BUG_ON(tcp_skb_pcount(skb) <= 1);
1945 BUG_ON(tp->snd_cwnd <= in_flight);
1947 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1949 /* From in_flight test above, we know that cwnd > in_flight. */
1950 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1952 limit = min(send_win, cong_win);
1954 /* If a full-sized TSO skb can be sent, do it. */
1955 if (limit >= max_segs * tp->mss_cache)
1958 /* Middle in queue won't get any more data, full sendable already? */
1959 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1962 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1964 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1966 /* If at least some fraction of a window is available,
1969 chunk /= win_divisor;
1973 /* Different approach, try not to defer past a single
1974 * ACK. Receiver should ACK every other full sized
1975 * frame, so if we have space for more than 3 frames
1978 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1982 /* TODO : use tsorted_sent_queue ? */
1983 head = tcp_rtx_queue_head(sk);
1986 delta = tp->tcp_clock_cache - head->tstamp;
1987 /* If next ACK is likely to come too late (half srtt), do not defer */
1988 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
1991 /* Ok, it looks like it is advisable to defer.
1992 * Three cases are tracked :
1993 * 1) We are cwnd-limited
1994 * 2) We are rwnd-limited
1995 * 3) We are application limited.
1997 if (cong_win < send_win) {
1998 if (cong_win <= skb->len) {
1999 *is_cwnd_limited = true;
2003 if (send_win <= skb->len) {
2004 *is_rwnd_limited = true;
2009 /* If this packet won't get more data, do not wait. */
2010 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2011 TCP_SKB_CB(skb)->eor)
2020 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2022 struct inet_connection_sock *icsk = inet_csk(sk);
2023 struct tcp_sock *tp = tcp_sk(sk);
2024 struct net *net = sock_net(sk);
2028 interval = net->ipv4.sysctl_tcp_probe_interval;
2029 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2030 if (unlikely(delta >= interval * HZ)) {
2031 int mss = tcp_current_mss(sk);
2033 /* Update current search range */
2034 icsk->icsk_mtup.probe_size = 0;
2035 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2036 sizeof(struct tcphdr) +
2037 icsk->icsk_af_ops->net_header_len;
2038 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2040 /* Update probe time stamp */
2041 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2045 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2047 struct sk_buff *skb, *next;
2049 skb = tcp_send_head(sk);
2050 tcp_for_write_queue_from_safe(skb, next, sk) {
2051 if (len <= skb->len)
2054 if (unlikely(TCP_SKB_CB(skb)->eor))
2063 /* Create a new MTU probe if we are ready.
2064 * MTU probe is regularly attempting to increase the path MTU by
2065 * deliberately sending larger packets. This discovers routing
2066 * changes resulting in larger path MTUs.
2068 * Returns 0 if we should wait to probe (no cwnd available),
2069 * 1 if a probe was sent,
2072 static int tcp_mtu_probe(struct sock *sk)
2074 struct inet_connection_sock *icsk = inet_csk(sk);
2075 struct tcp_sock *tp = tcp_sk(sk);
2076 struct sk_buff *skb, *nskb, *next;
2077 struct net *net = sock_net(sk);
2084 /* Not currently probing/verifying,
2086 * have enough cwnd, and
2087 * not SACKing (the variable headers throw things off)
2089 if (likely(!icsk->icsk_mtup.enabled ||
2090 icsk->icsk_mtup.probe_size ||
2091 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2092 tp->snd_cwnd < 11 ||
2093 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2096 /* Use binary search for probe_size between tcp_mss_base,
2097 * and current mss_clamp. if (search_high - search_low)
2098 * smaller than a threshold, backoff from probing.
2100 mss_now = tcp_current_mss(sk);
2101 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2102 icsk->icsk_mtup.search_low) >> 1);
2103 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2104 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2105 /* When misfortune happens, we are reprobing actively,
2106 * and then reprobe timer has expired. We stick with current
2107 * probing process by not resetting search range to its orignal.
2109 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2110 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2111 /* Check whether enough time has elaplased for
2112 * another round of probing.
2114 tcp_mtu_check_reprobe(sk);
2118 /* Have enough data in the send queue to probe? */
2119 if (tp->write_seq - tp->snd_nxt < size_needed)
2122 if (tp->snd_wnd < size_needed)
2124 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2127 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2128 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2129 if (!tcp_packets_in_flight(tp))
2135 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2138 /* We're allowed to probe. Build it now. */
2139 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2142 sk->sk_wmem_queued += nskb->truesize;
2143 sk_mem_charge(sk, nskb->truesize);
2145 skb = tcp_send_head(sk);
2147 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2148 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2149 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2150 TCP_SKB_CB(nskb)->sacked = 0;
2152 nskb->ip_summed = CHECKSUM_PARTIAL;
2154 tcp_insert_write_queue_before(nskb, skb, sk);
2155 tcp_highest_sack_replace(sk, skb, nskb);
2158 tcp_for_write_queue_from_safe(skb, next, sk) {
2159 copy = min_t(int, skb->len, probe_size - len);
2160 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2162 if (skb->len <= copy) {
2163 /* We've eaten all the data from this skb.
2165 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2166 /* If this is the last SKB we copy and eor is set
2167 * we need to propagate it to the new skb.
2169 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2170 tcp_unlink_write_queue(skb, sk);
2171 sk_wmem_free_skb(sk, skb);
2173 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2174 ~(TCPHDR_FIN|TCPHDR_PSH);
2175 if (!skb_shinfo(skb)->nr_frags) {
2176 skb_pull(skb, copy);
2178 __pskb_trim_head(skb, copy);
2179 tcp_set_skb_tso_segs(skb, mss_now);
2181 TCP_SKB_CB(skb)->seq += copy;
2186 if (len >= probe_size)
2189 tcp_init_tso_segs(nskb, nskb->len);
2191 /* We're ready to send. If this fails, the probe will
2192 * be resegmented into mss-sized pieces by tcp_write_xmit().
2194 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2195 /* Decrement cwnd here because we are sending
2196 * effectively two packets. */
2198 tcp_event_new_data_sent(sk, nskb);
2200 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2201 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2202 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2210 static bool tcp_pacing_check(struct sock *sk)
2212 struct tcp_sock *tp = tcp_sk(sk);
2214 if (!tcp_needs_internal_pacing(sk))
2217 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2220 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2221 hrtimer_start(&tp->pacing_timer,
2222 ns_to_ktime(tp->tcp_wstamp_ns),
2223 HRTIMER_MODE_ABS_PINNED_SOFT);
2229 /* TCP Small Queues :
2230 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2231 * (These limits are doubled for retransmits)
2233 * - better RTT estimation and ACK scheduling
2236 * Alas, some drivers / subsystems require a fair amount
2237 * of queued bytes to ensure line rate.
2238 * One example is wifi aggregation (802.11 AMPDU)
2240 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2241 unsigned int factor)
2243 unsigned long limit;
2245 limit = max_t(unsigned long,
2247 sk->sk_pacing_rate >> sk->sk_pacing_shift);
2248 if (sk->sk_pacing_status == SK_PACING_NONE)
2249 limit = min_t(unsigned long, limit,
2250 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2253 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2254 tcp_sk(sk)->tcp_tx_delay) {
2255 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2257 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2258 * approximate our needs assuming an ~100% skb->truesize overhead.
2259 * USEC_PER_SEC is approximated by 2^20.
2260 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2262 extra_bytes >>= (20 - 1);
2263 limit += extra_bytes;
2265 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2266 /* Always send skb if rtx queue is empty.
2267 * No need to wait for TX completion to call us back,
2268 * after softirq/tasklet schedule.
2269 * This helps when TX completions are delayed too much.
2271 if (tcp_rtx_queue_empty(sk))
2274 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2275 /* It is possible TX completion already happened
2276 * before we set TSQ_THROTTLED, so we must
2277 * test again the condition.
2279 smp_mb__after_atomic();
2280 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2286 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2288 const u32 now = tcp_jiffies32;
2289 enum tcp_chrono old = tp->chrono_type;
2291 if (old > TCP_CHRONO_UNSPEC)
2292 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2293 tp->chrono_start = now;
2294 tp->chrono_type = new;
2297 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2299 struct tcp_sock *tp = tcp_sk(sk);
2301 /* If there are multiple conditions worthy of tracking in a
2302 * chronograph then the highest priority enum takes precedence
2303 * over the other conditions. So that if something "more interesting"
2304 * starts happening, stop the previous chrono and start a new one.
2306 if (type > tp->chrono_type)
2307 tcp_chrono_set(tp, type);
2310 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2312 struct tcp_sock *tp = tcp_sk(sk);
2315 /* There are multiple conditions worthy of tracking in a
2316 * chronograph, so that the highest priority enum takes
2317 * precedence over the other conditions (see tcp_chrono_start).
2318 * If a condition stops, we only stop chrono tracking if
2319 * it's the "most interesting" or current chrono we are
2320 * tracking and starts busy chrono if we have pending data.
2322 if (tcp_rtx_and_write_queues_empty(sk))
2323 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2324 else if (type == tp->chrono_type)
2325 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2328 /* This routine writes packets to the network. It advances the
2329 * send_head. This happens as incoming acks open up the remote
2332 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2333 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2334 * account rare use of URG, this is not a big flaw.
2336 * Send at most one packet when push_one > 0. Temporarily ignore
2337 * cwnd limit to force at most one packet out when push_one == 2.
2339 * Returns true, if no segments are in flight and we have queued segments,
2340 * but cannot send anything now because of SWS or another problem.
2342 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2343 int push_one, gfp_t gfp)
2345 struct tcp_sock *tp = tcp_sk(sk);
2346 struct sk_buff *skb;
2347 unsigned int tso_segs, sent_pkts;
2350 bool is_cwnd_limited = false, is_rwnd_limited = false;
2355 tcp_mstamp_refresh(tp);
2357 /* Do MTU probing. */
2358 result = tcp_mtu_probe(sk);
2361 } else if (result > 0) {
2366 max_segs = tcp_tso_segs(sk, mss_now);
2367 while ((skb = tcp_send_head(sk))) {
2370 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2371 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2372 skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2373 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2374 tcp_init_tso_segs(skb, mss_now);
2375 goto repair; /* Skip network transmission */
2378 if (tcp_pacing_check(sk))
2381 tso_segs = tcp_init_tso_segs(skb, mss_now);
2384 cwnd_quota = tcp_cwnd_test(tp, skb);
2387 /* Force out a loss probe pkt. */
2393 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2394 is_rwnd_limited = true;
2398 if (tso_segs == 1) {
2399 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2400 (tcp_skb_is_last(sk, skb) ?
2401 nonagle : TCP_NAGLE_PUSH))))
2405 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2406 &is_rwnd_limited, max_segs))
2411 if (tso_segs > 1 && !tcp_urg_mode(tp))
2412 limit = tcp_mss_split_point(sk, skb, mss_now,
2418 if (skb->len > limit &&
2419 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2422 if (tcp_small_queue_check(sk, skb, 0))
2425 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2429 /* Advance the send_head. This one is sent out.
2430 * This call will increment packets_out.
2432 tcp_event_new_data_sent(sk, skb);
2434 tcp_minshall_update(tp, mss_now, skb);
2435 sent_pkts += tcp_skb_pcount(skb);
2441 if (is_rwnd_limited)
2442 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2444 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2446 if (likely(sent_pkts)) {
2447 if (tcp_in_cwnd_reduction(sk))
2448 tp->prr_out += sent_pkts;
2450 /* Send one loss probe per tail loss episode. */
2452 tcp_schedule_loss_probe(sk, false);
2453 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2454 tcp_cwnd_validate(sk, is_cwnd_limited);
2457 return !tp->packets_out && !tcp_write_queue_empty(sk);
2460 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2462 struct inet_connection_sock *icsk = inet_csk(sk);
2463 struct tcp_sock *tp = tcp_sk(sk);
2464 u32 timeout, rto_delta_us;
2467 /* Don't do any loss probe on a Fast Open connection before 3WHS
2470 if (tp->fastopen_rsk)
2473 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2474 /* Schedule a loss probe in 2*RTT for SACK capable connections
2475 * not in loss recovery, that are either limited by cwnd or application.
2477 if ((early_retrans != 3 && early_retrans != 4) ||
2478 !tp->packets_out || !tcp_is_sack(tp) ||
2479 (icsk->icsk_ca_state != TCP_CA_Open &&
2480 icsk->icsk_ca_state != TCP_CA_CWR))
2483 /* Probe timeout is 2*rtt. Add minimum RTO to account
2484 * for delayed ack when there's one outstanding packet. If no RTT
2485 * sample is available then probe after TCP_TIMEOUT_INIT.
2488 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2489 if (tp->packets_out == 1)
2490 timeout += TCP_RTO_MIN;
2492 timeout += TCP_TIMEOUT_MIN;
2494 timeout = TCP_TIMEOUT_INIT;
2497 /* If the RTO formula yields an earlier time, then use that time. */
2498 rto_delta_us = advancing_rto ?
2499 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2500 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2501 if (rto_delta_us > 0)
2502 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2504 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2509 /* Thanks to skb fast clones, we can detect if a prior transmit of
2510 * a packet is still in a qdisc or driver queue.
2511 * In this case, there is very little point doing a retransmit !
2513 static bool skb_still_in_host_queue(const struct sock *sk,
2514 const struct sk_buff *skb)
2516 if (unlikely(skb_fclone_busy(sk, skb))) {
2517 NET_INC_STATS(sock_net(sk),
2518 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2524 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2525 * retransmit the last segment.
2527 void tcp_send_loss_probe(struct sock *sk)
2529 struct tcp_sock *tp = tcp_sk(sk);
2530 struct sk_buff *skb;
2532 int mss = tcp_current_mss(sk);
2534 skb = tcp_send_head(sk);
2535 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2536 pcount = tp->packets_out;
2537 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2538 if (tp->packets_out > pcount)
2542 skb = skb_rb_last(&sk->tcp_rtx_queue);
2543 if (unlikely(!skb)) {
2544 WARN_ONCE(tp->packets_out,
2545 "invalid inflight: %u state %u cwnd %u mss %d\n",
2546 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2547 inet_csk(sk)->icsk_pending = 0;
2551 /* At most one outstanding TLP retransmission. */
2552 if (tp->tlp_high_seq)
2555 if (skb_still_in_host_queue(sk, skb))
2558 pcount = tcp_skb_pcount(skb);
2559 if (WARN_ON(!pcount))
2562 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2563 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2564 (pcount - 1) * mss, mss,
2567 skb = skb_rb_next(skb);
2570 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2573 if (__tcp_retransmit_skb(sk, skb, 1))
2576 /* Record snd_nxt for loss detection. */
2577 tp->tlp_high_seq = tp->snd_nxt;
2580 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2581 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2582 inet_csk(sk)->icsk_pending = 0;
2587 /* Push out any pending frames which were held back due to
2588 * TCP_CORK or attempt at coalescing tiny packets.
2589 * The socket must be locked by the caller.
2591 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2594 /* If we are closed, the bytes will have to remain here.
2595 * In time closedown will finish, we empty the write queue and
2596 * all will be happy.
2598 if (unlikely(sk->sk_state == TCP_CLOSE))
2601 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2602 sk_gfp_mask(sk, GFP_ATOMIC)))
2603 tcp_check_probe_timer(sk);
2606 /* Send _single_ skb sitting at the send head. This function requires
2607 * true push pending frames to setup probe timer etc.
2609 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2611 struct sk_buff *skb = tcp_send_head(sk);
2613 BUG_ON(!skb || skb->len < mss_now);
2615 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2618 /* This function returns the amount that we can raise the
2619 * usable window based on the following constraints
2621 * 1. The window can never be shrunk once it is offered (RFC 793)
2622 * 2. We limit memory per socket
2625 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2626 * RECV.NEXT + RCV.WIN fixed until:
2627 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2629 * i.e. don't raise the right edge of the window until you can raise
2630 * it at least MSS bytes.
2632 * Unfortunately, the recommended algorithm breaks header prediction,
2633 * since header prediction assumes th->window stays fixed.
2635 * Strictly speaking, keeping th->window fixed violates the receiver
2636 * side SWS prevention criteria. The problem is that under this rule
2637 * a stream of single byte packets will cause the right side of the
2638 * window to always advance by a single byte.
2640 * Of course, if the sender implements sender side SWS prevention
2641 * then this will not be a problem.
2643 * BSD seems to make the following compromise:
2645 * If the free space is less than the 1/4 of the maximum
2646 * space available and the free space is less than 1/2 mss,
2647 * then set the window to 0.
2648 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2649 * Otherwise, just prevent the window from shrinking
2650 * and from being larger than the largest representable value.
2652 * This prevents incremental opening of the window in the regime
2653 * where TCP is limited by the speed of the reader side taking
2654 * data out of the TCP receive queue. It does nothing about
2655 * those cases where the window is constrained on the sender side
2656 * because the pipeline is full.
2658 * BSD also seems to "accidentally" limit itself to windows that are a
2659 * multiple of MSS, at least until the free space gets quite small.
2660 * This would appear to be a side effect of the mbuf implementation.
2661 * Combining these two algorithms results in the observed behavior
2662 * of having a fixed window size at almost all times.
2664 * Below we obtain similar behavior by forcing the offered window to
2665 * a multiple of the mss when it is feasible to do so.
2667 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2668 * Regular options like TIMESTAMP are taken into account.
2670 u32 __tcp_select_window(struct sock *sk)
2672 struct inet_connection_sock *icsk = inet_csk(sk);
2673 struct tcp_sock *tp = tcp_sk(sk);
2674 /* MSS for the peer's data. Previous versions used mss_clamp
2675 * here. I don't know if the value based on our guesses
2676 * of peer's MSS is better for the performance. It's more correct
2677 * but may be worse for the performance because of rcv_mss
2678 * fluctuations. --SAW 1998/11/1
2680 int mss = icsk->icsk_ack.rcv_mss;
2681 int free_space = tcp_space(sk);
2682 int allowed_space = tcp_full_space(sk);
2683 int full_space = min_t(int, tp->window_clamp, allowed_space);
2686 if (unlikely(mss > full_space)) {
2691 if (free_space < (full_space >> 1)) {
2692 icsk->icsk_ack.quick = 0;
2694 if (tcp_under_memory_pressure(sk))
2695 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2698 /* free_space might become our new window, make sure we don't
2699 * increase it due to wscale.
2701 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2703 /* if free space is less than mss estimate, or is below 1/16th
2704 * of the maximum allowed, try to move to zero-window, else
2705 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2706 * new incoming data is dropped due to memory limits.
2707 * With large window, mss test triggers way too late in order
2708 * to announce zero window in time before rmem limit kicks in.
2710 if (free_space < (allowed_space >> 4) || free_space < mss)
2714 if (free_space > tp->rcv_ssthresh)
2715 free_space = tp->rcv_ssthresh;
2717 /* Don't do rounding if we are using window scaling, since the
2718 * scaled window will not line up with the MSS boundary anyway.
2720 if (tp->rx_opt.rcv_wscale) {
2721 window = free_space;
2723 /* Advertise enough space so that it won't get scaled away.
2724 * Import case: prevent zero window announcement if
2725 * 1<<rcv_wscale > mss.
2727 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2729 window = tp->rcv_wnd;
2730 /* Get the largest window that is a nice multiple of mss.
2731 * Window clamp already applied above.
2732 * If our current window offering is within 1 mss of the
2733 * free space we just keep it. This prevents the divide
2734 * and multiply from happening most of the time.
2735 * We also don't do any window rounding when the free space
2738 if (window <= free_space - mss || window > free_space)
2739 window = rounddown(free_space, mss);
2740 else if (mss == full_space &&
2741 free_space > window + (full_space >> 1))
2742 window = free_space;
2748 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2749 const struct sk_buff *next_skb)
2751 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2752 const struct skb_shared_info *next_shinfo =
2753 skb_shinfo(next_skb);
2754 struct skb_shared_info *shinfo = skb_shinfo(skb);
2756 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2757 shinfo->tskey = next_shinfo->tskey;
2758 TCP_SKB_CB(skb)->txstamp_ack |=
2759 TCP_SKB_CB(next_skb)->txstamp_ack;
2763 /* Collapses two adjacent SKB's during retransmission. */
2764 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2766 struct tcp_sock *tp = tcp_sk(sk);
2767 struct sk_buff *next_skb = skb_rb_next(skb);
2770 next_skb_size = next_skb->len;
2772 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2774 if (next_skb_size) {
2775 if (next_skb_size <= skb_availroom(skb))
2776 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2778 else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
2781 tcp_highest_sack_replace(sk, next_skb, skb);
2783 /* Update sequence range on original skb. */
2784 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2786 /* Merge over control information. This moves PSH/FIN etc. over */
2787 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2789 /* All done, get rid of second SKB and account for it so
2790 * packet counting does not break.
2792 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2793 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2795 /* changed transmit queue under us so clear hints */
2796 tcp_clear_retrans_hints_partial(tp);
2797 if (next_skb == tp->retransmit_skb_hint)
2798 tp->retransmit_skb_hint = skb;
2800 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2802 tcp_skb_collapse_tstamp(skb, next_skb);
2804 tcp_rtx_queue_unlink_and_free(next_skb, sk);
2808 /* Check if coalescing SKBs is legal. */
2809 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2811 if (tcp_skb_pcount(skb) > 1)
2813 if (skb_cloned(skb))
2815 /* Some heuristics for collapsing over SACK'd could be invented */
2816 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2822 /* Collapse packets in the retransmit queue to make to create
2823 * less packets on the wire. This is only done on retransmission.
2825 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2828 struct tcp_sock *tp = tcp_sk(sk);
2829 struct sk_buff *skb = to, *tmp;
2832 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2834 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2837 skb_rbtree_walk_from_safe(skb, tmp) {
2838 if (!tcp_can_collapse(sk, skb))
2841 if (!tcp_skb_can_collapse_to(to))
2854 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2857 if (!tcp_collapse_retrans(sk, to))
2862 /* This retransmits one SKB. Policy decisions and retransmit queue
2863 * state updates are done by the caller. Returns non-zero if an
2864 * error occurred which prevented the send.
2866 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2868 struct inet_connection_sock *icsk = inet_csk(sk);
2869 struct tcp_sock *tp = tcp_sk(sk);
2870 unsigned int cur_mss;
2874 /* Inconclusive MTU probe */
2875 if (icsk->icsk_mtup.probe_size)
2876 icsk->icsk_mtup.probe_size = 0;
2878 /* Do not sent more than we queued. 1/4 is reserved for possible
2879 * copying overhead: fragmentation, tunneling, mangling etc.
2881 if (refcount_read(&sk->sk_wmem_alloc) >
2882 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2886 if (skb_still_in_host_queue(sk, skb))
2889 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2890 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2894 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2898 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2899 return -EHOSTUNREACH; /* Routing failure or similar. */
2901 cur_mss = tcp_current_mss(sk);
2903 /* If receiver has shrunk his window, and skb is out of
2904 * new window, do not retransmit it. The exception is the
2905 * case, when window is shrunk to zero. In this case
2906 * our retransmit serves as a zero window probe.
2908 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2909 TCP_SKB_CB(skb)->seq != tp->snd_una)
2912 len = cur_mss * segs;
2913 if (skb->len > len) {
2914 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2915 cur_mss, GFP_ATOMIC))
2916 return -ENOMEM; /* We'll try again later. */
2918 if (skb_unclone(skb, GFP_ATOMIC))
2921 diff = tcp_skb_pcount(skb);
2922 tcp_set_skb_tso_segs(skb, cur_mss);
2923 diff -= tcp_skb_pcount(skb);
2925 tcp_adjust_pcount(sk, skb, diff);
2926 if (skb->len < cur_mss)
2927 tcp_retrans_try_collapse(sk, skb, cur_mss);
2930 /* RFC3168, section 6.1.1.1. ECN fallback */
2931 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2932 tcp_ecn_clear_syn(sk, skb);
2934 /* Update global and local TCP statistics. */
2935 segs = tcp_skb_pcount(skb);
2936 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2937 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2938 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2939 tp->total_retrans += segs;
2940 tp->bytes_retrans += skb->len;
2942 /* make sure skb->data is aligned on arches that require it
2943 * and check if ack-trimming & collapsing extended the headroom
2944 * beyond what csum_start can cover.
2946 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2947 skb_headroom(skb) >= 0xFFFF)) {
2948 struct sk_buff *nskb;
2950 tcp_skb_tsorted_save(skb) {
2951 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2952 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2954 } tcp_skb_tsorted_restore(skb);
2957 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
2958 tcp_rate_skb_sent(sk, skb);
2961 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2964 /* To avoid taking spuriously low RTT samples based on a timestamp
2965 * for a transmit that never happened, always mark EVER_RETRANS
2967 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2969 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
2970 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
2971 TCP_SKB_CB(skb)->seq, segs, err);
2974 trace_tcp_retransmit_skb(sk, skb);
2975 } else if (err != -EBUSY) {
2976 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
2981 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2983 struct tcp_sock *tp = tcp_sk(sk);
2984 int err = __tcp_retransmit_skb(sk, skb, segs);
2987 #if FASTRETRANS_DEBUG > 0
2988 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2989 net_dbg_ratelimited("retrans_out leaked\n");
2992 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2993 tp->retrans_out += tcp_skb_pcount(skb);
2996 /* Save stamp of the first (attempted) retransmit. */
2997 if (!tp->retrans_stamp)
2998 tp->retrans_stamp = tcp_skb_timestamp(skb);
3000 if (tp->undo_retrans < 0)
3001 tp->undo_retrans = 0;
3002 tp->undo_retrans += tcp_skb_pcount(skb);
3006 /* This gets called after a retransmit timeout, and the initially
3007 * retransmitted data is acknowledged. It tries to continue
3008 * resending the rest of the retransmit queue, until either
3009 * we've sent it all or the congestion window limit is reached.
3011 void tcp_xmit_retransmit_queue(struct sock *sk)
3013 const struct inet_connection_sock *icsk = inet_csk(sk);
3014 struct sk_buff *skb, *rtx_head, *hole = NULL;
3015 struct tcp_sock *tp = tcp_sk(sk);
3019 if (!tp->packets_out)
3022 rtx_head = tcp_rtx_queue_head(sk);
3023 skb = tp->retransmit_skb_hint ?: rtx_head;
3024 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3025 skb_rbtree_walk_from(skb) {
3029 if (tcp_pacing_check(sk))
3032 /* we could do better than to assign each time */
3034 tp->retransmit_skb_hint = skb;
3036 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
3039 sacked = TCP_SKB_CB(skb)->sacked;
3040 /* In case tcp_shift_skb_data() have aggregated large skbs,
3041 * we need to make sure not sending too bigs TSO packets
3043 segs = min_t(int, segs, max_segs);
3045 if (tp->retrans_out >= tp->lost_out) {
3047 } else if (!(sacked & TCPCB_LOST)) {
3048 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3053 if (icsk->icsk_ca_state != TCP_CA_Loss)
3054 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3056 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3059 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3062 if (tcp_small_queue_check(sk, skb, 1))
3065 if (tcp_retransmit_skb(sk, skb, segs))
3068 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3070 if (tcp_in_cwnd_reduction(sk))
3071 tp->prr_out += tcp_skb_pcount(skb);
3073 if (skb == rtx_head &&
3074 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3075 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3076 inet_csk(sk)->icsk_rto,
3082 /* We allow to exceed memory limits for FIN packets to expedite
3083 * connection tear down and (memory) recovery.
3084 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3085 * or even be forced to close flow without any FIN.
3086 * In general, we want to allow one skb per socket to avoid hangs
3087 * with edge trigger epoll()
3089 void sk_forced_mem_schedule(struct sock *sk, int size)
3093 if (size <= sk->sk_forward_alloc)
3095 amt = sk_mem_pages(size);
3096 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3097 sk_memory_allocated_add(sk, amt);
3099 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3100 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3103 /* Send a FIN. The caller locks the socket for us.
3104 * We should try to send a FIN packet really hard, but eventually give up.
3106 void tcp_send_fin(struct sock *sk)
3108 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3109 struct tcp_sock *tp = tcp_sk(sk);
3111 /* Optimization, tack on the FIN if we have one skb in write queue and
3112 * this skb was not yet sent, or we are under memory pressure.
3113 * Note: in the latter case, FIN packet will be sent after a timeout,
3114 * as TCP stack thinks it has already been transmitted.
3116 if (!tskb && tcp_under_memory_pressure(sk))
3117 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3120 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3121 TCP_SKB_CB(tskb)->end_seq++;
3123 if (tcp_write_queue_empty(sk)) {
3124 /* This means tskb was already sent.
3125 * Pretend we included the FIN on previous transmit.
3126 * We need to set tp->snd_nxt to the value it would have
3127 * if FIN had been sent. This is because retransmit path
3128 * does not change tp->snd_nxt.
3134 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3138 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3139 skb_reserve(skb, MAX_TCP_HEADER);
3140 sk_forced_mem_schedule(sk, skb->truesize);
3141 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3142 tcp_init_nondata_skb(skb, tp->write_seq,
3143 TCPHDR_ACK | TCPHDR_FIN);
3144 tcp_queue_skb(sk, skb);
3146 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3149 /* We get here when a process closes a file descriptor (either due to
3150 * an explicit close() or as a byproduct of exit()'ing) and there
3151 * was unread data in the receive queue. This behavior is recommended
3152 * by RFC 2525, section 2.17. -DaveM
3154 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3156 struct sk_buff *skb;
3158 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3160 /* NOTE: No TCP options attached and we never retransmit this. */
3161 skb = alloc_skb(MAX_TCP_HEADER, priority);
3163 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3167 /* Reserve space for headers and prepare control bits. */
3168 skb_reserve(skb, MAX_TCP_HEADER);
3169 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3170 TCPHDR_ACK | TCPHDR_RST);
3171 tcp_mstamp_refresh(tcp_sk(sk));
3173 if (tcp_transmit_skb(sk, skb, 0, priority))
3174 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3176 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3177 * skb here is different to the troublesome skb, so use NULL
3179 trace_tcp_send_reset(sk, NULL);
3182 /* Send a crossed SYN-ACK during socket establishment.
3183 * WARNING: This routine must only be called when we have already sent
3184 * a SYN packet that crossed the incoming SYN that caused this routine
3185 * to get called. If this assumption fails then the initial rcv_wnd
3186 * and rcv_wscale values will not be correct.
3188 int tcp_send_synack(struct sock *sk)
3190 struct sk_buff *skb;
3192 skb = tcp_rtx_queue_head(sk);
3193 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3194 pr_err("%s: wrong queue state\n", __func__);
3197 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3198 if (skb_cloned(skb)) {
3199 struct sk_buff *nskb;
3201 tcp_skb_tsorted_save(skb) {
3202 nskb = skb_copy(skb, GFP_ATOMIC);
3203 } tcp_skb_tsorted_restore(skb);
3206 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3207 tcp_rtx_queue_unlink_and_free(skb, sk);
3208 __skb_header_release(nskb);
3209 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3210 sk->sk_wmem_queued += nskb->truesize;
3211 sk_mem_charge(sk, nskb->truesize);
3215 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3216 tcp_ecn_send_synack(sk, skb);
3218 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3222 * tcp_make_synack - Prepare a SYN-ACK.
3223 * sk: listener socket
3224 * dst: dst entry attached to the SYNACK
3225 * req: request_sock pointer
3227 * Allocate one skb and build a SYNACK packet.
3228 * @dst is consumed : Caller should not use it again.
3230 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3231 struct request_sock *req,
3232 struct tcp_fastopen_cookie *foc,
3233 enum tcp_synack_type synack_type)
3235 struct inet_request_sock *ireq = inet_rsk(req);
3236 const struct tcp_sock *tp = tcp_sk(sk);
3237 struct tcp_md5sig_key *md5 = NULL;
3238 struct tcp_out_options opts;
3239 struct sk_buff *skb;
3240 int tcp_header_size;
3245 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3246 if (unlikely(!skb)) {
3250 /* Reserve space for headers. */
3251 skb_reserve(skb, MAX_TCP_HEADER);
3253 switch (synack_type) {
3254 case TCP_SYNACK_NORMAL:
3255 skb_set_owner_w(skb, req_to_sk(req));
3257 case TCP_SYNACK_COOKIE:
3258 /* Under synflood, we do not attach skb to a socket,
3259 * to avoid false sharing.
3262 case TCP_SYNACK_FASTOPEN:
3263 /* sk is a const pointer, because we want to express multiple
3264 * cpu might call us concurrently.
3265 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3267 skb_set_owner_w(skb, (struct sock *)sk);
3270 skb_dst_set(skb, dst);
3272 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3274 memset(&opts, 0, sizeof(opts));
3275 now = tcp_clock_ns();
3276 #ifdef CONFIG_SYN_COOKIES
3277 if (unlikely(req->cookie_ts))
3278 skb->skb_mstamp_ns = cookie_init_timestamp(req);
3282 skb->skb_mstamp_ns = now;
3283 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3284 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3287 #ifdef CONFIG_TCP_MD5SIG
3289 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3291 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3292 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3295 skb_push(skb, tcp_header_size);
3296 skb_reset_transport_header(skb);
3298 th = (struct tcphdr *)skb->data;
3299 memset(th, 0, sizeof(struct tcphdr));
3302 tcp_ecn_make_synack(req, th);
3303 th->source = htons(ireq->ir_num);
3304 th->dest = ireq->ir_rmt_port;
3305 skb->mark = ireq->ir_mark;
3306 skb->ip_summed = CHECKSUM_PARTIAL;
3307 th->seq = htonl(tcp_rsk(req)->snt_isn);
3308 /* XXX data is queued and acked as is. No buffer/window check */
3309 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3311 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3312 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3313 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3314 th->doff = (tcp_header_size >> 2);
3315 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3317 #ifdef CONFIG_TCP_MD5SIG
3318 /* Okay, we have all we need - do the md5 hash if needed */
3320 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3321 md5, req_to_sk(req), skb);
3325 skb->skb_mstamp_ns = now;
3326 tcp_add_tx_delay(skb, tp);
3330 EXPORT_SYMBOL(tcp_make_synack);
3332 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3334 struct inet_connection_sock *icsk = inet_csk(sk);
3335 const struct tcp_congestion_ops *ca;
3336 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3338 if (ca_key == TCP_CA_UNSPEC)
3342 ca = tcp_ca_find_key(ca_key);
3343 if (likely(ca && try_module_get(ca->owner))) {
3344 module_put(icsk->icsk_ca_ops->owner);
3345 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3346 icsk->icsk_ca_ops = ca;
3351 /* Do all connect socket setups that can be done AF independent. */
3352 static void tcp_connect_init(struct sock *sk)
3354 const struct dst_entry *dst = __sk_dst_get(sk);
3355 struct tcp_sock *tp = tcp_sk(sk);
3359 /* We'll fix this up when we get a response from the other end.
3360 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3362 tp->tcp_header_len = sizeof(struct tcphdr);
3363 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3364 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3366 #ifdef CONFIG_TCP_MD5SIG
3367 if (tp->af_specific->md5_lookup(sk, sk))
3368 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3371 /* If user gave his TCP_MAXSEG, record it to clamp */
3372 if (tp->rx_opt.user_mss)
3373 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3376 tcp_sync_mss(sk, dst_mtu(dst));
3378 tcp_ca_dst_init(sk, dst);
3380 if (!tp->window_clamp)
3381 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3382 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3384 tcp_initialize_rcv_mss(sk);
3386 /* limit the window selection if the user enforce a smaller rx buffer */
3387 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3388 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3389 tp->window_clamp = tcp_full_space(sk);
3391 rcv_wnd = tcp_rwnd_init_bpf(sk);
3393 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3395 tcp_select_initial_window(sk, tcp_full_space(sk),
3396 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3399 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3403 tp->rx_opt.rcv_wscale = rcv_wscale;
3404 tp->rcv_ssthresh = tp->rcv_wnd;
3407 sock_reset_flag(sk, SOCK_DONE);
3410 tcp_write_queue_purge(sk);
3411 tp->snd_una = tp->write_seq;
3412 tp->snd_sml = tp->write_seq;
3413 tp->snd_up = tp->write_seq;
3414 tp->snd_nxt = tp->write_seq;
3416 if (likely(!tp->repair))
3419 tp->rcv_tstamp = tcp_jiffies32;
3420 tp->rcv_wup = tp->rcv_nxt;
3421 tp->copied_seq = tp->rcv_nxt;
3423 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3424 inet_csk(sk)->icsk_retransmits = 0;
3425 tcp_clear_retrans(tp);
3428 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3430 struct tcp_sock *tp = tcp_sk(sk);
3431 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3433 tcb->end_seq += skb->len;
3434 __skb_header_release(skb);
3435 sk->sk_wmem_queued += skb->truesize;
3436 sk_mem_charge(sk, skb->truesize);
3437 tp->write_seq = tcb->end_seq;
3438 tp->packets_out += tcp_skb_pcount(skb);
3441 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3442 * queue a data-only packet after the regular SYN, such that regular SYNs
3443 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3444 * only the SYN sequence, the data are retransmitted in the first ACK.
3445 * If cookie is not cached or other error occurs, falls back to send a
3446 * regular SYN with Fast Open cookie request option.
3448 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3450 struct tcp_sock *tp = tcp_sk(sk);
3451 struct tcp_fastopen_request *fo = tp->fastopen_req;
3453 struct sk_buff *syn_data;
3455 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3456 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3459 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3460 * user-MSS. Reserve maximum option space for middleboxes that add
3461 * private TCP options. The cost is reduced data space in SYN :(
3463 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3465 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3466 MAX_TCP_OPTION_SPACE;
3468 space = min_t(size_t, space, fo->size);
3470 /* limit to order-0 allocations */
3471 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3473 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3476 syn_data->ip_summed = CHECKSUM_PARTIAL;
3477 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3479 int copied = copy_from_iter(skb_put(syn_data, space), space,
3480 &fo->data->msg_iter);
3481 if (unlikely(!copied)) {
3482 tcp_skb_tsorted_anchor_cleanup(syn_data);
3483 kfree_skb(syn_data);
3486 if (copied != space) {
3487 skb_trim(syn_data, copied);
3490 skb_zcopy_set(syn_data, fo->uarg, NULL);
3492 /* No more data pending in inet_wait_for_connect() */
3493 if (space == fo->size)
3497 tcp_connect_queue_skb(sk, syn_data);
3499 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3501 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3503 syn->skb_mstamp_ns = syn_data->skb_mstamp_ns;
3505 /* Now full SYN+DATA was cloned and sent (or not),
3506 * remove the SYN from the original skb (syn_data)
3507 * we keep in write queue in case of a retransmit, as we
3508 * also have the SYN packet (with no data) in the same queue.
3510 TCP_SKB_CB(syn_data)->seq++;
3511 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3513 tp->syn_data = (fo->copied > 0);
3514 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3515 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3519 /* data was not sent, put it in write_queue */
3520 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3521 tp->packets_out -= tcp_skb_pcount(syn_data);
3524 /* Send a regular SYN with Fast Open cookie request option */
3525 if (fo->cookie.len > 0)
3527 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3529 tp->syn_fastopen = 0;
3531 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3535 /* Build a SYN and send it off. */
3536 int tcp_connect(struct sock *sk)
3538 struct tcp_sock *tp = tcp_sk(sk);
3539 struct sk_buff *buff;
3542 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3544 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3545 return -EHOSTUNREACH; /* Routing failure or similar. */
3547 tcp_connect_init(sk);
3549 if (unlikely(tp->repair)) {
3550 tcp_finish_connect(sk, NULL);
3554 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3555 if (unlikely(!buff))
3558 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3559 tcp_mstamp_refresh(tp);
3560 tp->retrans_stamp = tcp_time_stamp(tp);
3561 tcp_connect_queue_skb(sk, buff);
3562 tcp_ecn_send_syn(sk, buff);
3563 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3565 /* Send off SYN; include data in Fast Open. */
3566 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3567 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3568 if (err == -ECONNREFUSED)
3571 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3572 * in order to make this packet get counted in tcpOutSegs.
3574 tp->snd_nxt = tp->write_seq;
3575 tp->pushed_seq = tp->write_seq;
3576 buff = tcp_send_head(sk);
3577 if (unlikely(buff)) {
3578 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3579 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3581 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3583 /* Timer for repeating the SYN until an answer. */
3584 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3585 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3588 EXPORT_SYMBOL(tcp_connect);
3590 /* Send out a delayed ack, the caller does the policy checking
3591 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3594 void tcp_send_delayed_ack(struct sock *sk)
3596 struct inet_connection_sock *icsk = inet_csk(sk);
3597 int ato = icsk->icsk_ack.ato;
3598 unsigned long timeout;
3600 if (ato > TCP_DELACK_MIN) {
3601 const struct tcp_sock *tp = tcp_sk(sk);
3602 int max_ato = HZ / 2;
3604 if (inet_csk_in_pingpong_mode(sk) ||
3605 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3606 max_ato = TCP_DELACK_MAX;
3608 /* Slow path, intersegment interval is "high". */
3610 /* If some rtt estimate is known, use it to bound delayed ack.
3611 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3615 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3622 ato = min(ato, max_ato);
3625 /* Stay within the limit we were given */
3626 timeout = jiffies + ato;
3628 /* Use new timeout only if there wasn't a older one earlier. */
3629 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3630 /* If delack timer was blocked or is about to expire,
3633 if (icsk->icsk_ack.blocked ||
3634 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3639 if (!time_before(timeout, icsk->icsk_ack.timeout))
3640 timeout = icsk->icsk_ack.timeout;
3642 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3643 icsk->icsk_ack.timeout = timeout;
3644 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3647 /* This routine sends an ack and also updates the window. */
3648 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3650 struct sk_buff *buff;
3652 /* If we have been reset, we may not send again. */
3653 if (sk->sk_state == TCP_CLOSE)
3656 /* We are not putting this on the write queue, so
3657 * tcp_transmit_skb() will set the ownership to this
3660 buff = alloc_skb(MAX_TCP_HEADER,
3661 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3662 if (unlikely(!buff)) {
3663 inet_csk_schedule_ack(sk);
3664 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3665 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3666 TCP_DELACK_MAX, TCP_RTO_MAX);
3670 /* Reserve space for headers and prepare control bits. */
3671 skb_reserve(buff, MAX_TCP_HEADER);
3672 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3674 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3676 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3678 skb_set_tcp_pure_ack(buff);
3680 /* Send it off, this clears delayed acks for us. */
3681 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3683 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3685 void tcp_send_ack(struct sock *sk)
3687 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3690 /* This routine sends a packet with an out of date sequence
3691 * number. It assumes the other end will try to ack it.
3693 * Question: what should we make while urgent mode?
3694 * 4.4BSD forces sending single byte of data. We cannot send
3695 * out of window data, because we have SND.NXT==SND.MAX...
3697 * Current solution: to send TWO zero-length segments in urgent mode:
3698 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3699 * out-of-date with SND.UNA-1 to probe window.
3701 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3703 struct tcp_sock *tp = tcp_sk(sk);
3704 struct sk_buff *skb;
3706 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3707 skb = alloc_skb(MAX_TCP_HEADER,
3708 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3712 /* Reserve space for headers and set control bits. */
3713 skb_reserve(skb, MAX_TCP_HEADER);
3714 /* Use a previous sequence. This should cause the other
3715 * end to send an ack. Don't queue or clone SKB, just
3718 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3719 NET_INC_STATS(sock_net(sk), mib);
3720 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3723 /* Called from setsockopt( ... TCP_REPAIR ) */
3724 void tcp_send_window_probe(struct sock *sk)
3726 if (sk->sk_state == TCP_ESTABLISHED) {
3727 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3728 tcp_mstamp_refresh(tcp_sk(sk));
3729 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3733 /* Initiate keepalive or window probe from timer. */
3734 int tcp_write_wakeup(struct sock *sk, int mib)
3736 struct tcp_sock *tp = tcp_sk(sk);
3737 struct sk_buff *skb;
3739 if (sk->sk_state == TCP_CLOSE)
3742 skb = tcp_send_head(sk);
3743 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3745 unsigned int mss = tcp_current_mss(sk);
3746 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3748 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3749 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3751 /* We are probing the opening of a window
3752 * but the window size is != 0
3753 * must have been a result SWS avoidance ( sender )
3755 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3757 seg_size = min(seg_size, mss);
3758 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3759 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3760 skb, seg_size, mss, GFP_ATOMIC))
3762 } else if (!tcp_skb_pcount(skb))
3763 tcp_set_skb_tso_segs(skb, mss);
3765 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3766 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3768 tcp_event_new_data_sent(sk, skb);
3771 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3772 tcp_xmit_probe_skb(sk, 1, mib);
3773 return tcp_xmit_probe_skb(sk, 0, mib);
3777 /* A window probe timeout has occurred. If window is not closed send
3778 * a partial packet else a zero probe.
3780 void tcp_send_probe0(struct sock *sk)
3782 struct inet_connection_sock *icsk = inet_csk(sk);
3783 struct tcp_sock *tp = tcp_sk(sk);
3784 struct net *net = sock_net(sk);
3785 unsigned long timeout;
3788 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3790 if (tp->packets_out || tcp_write_queue_empty(sk)) {
3791 /* Cancel probe timer, if it is not required. */
3792 icsk->icsk_probes_out = 0;
3793 icsk->icsk_backoff = 0;
3797 icsk->icsk_probes_out++;
3799 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3800 icsk->icsk_backoff++;
3801 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
3803 /* If packet was not sent due to local congestion,
3804 * Let senders fight for local resources conservatively.
3806 timeout = TCP_RESOURCE_PROBE_INTERVAL;
3808 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX, NULL);
3811 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3813 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3817 tcp_rsk(req)->txhash = net_tx_rndhash();
3818 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3820 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3821 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3822 if (unlikely(tcp_passive_fastopen(sk)))
3823 tcp_sk(sk)->total_retrans++;
3824 trace_tcp_retransmit_synack(sk, req);
3828 EXPORT_SYMBOL(tcp_rtx_synack);