2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
44 #include <linux/static_key.h>
46 #include <trace/events/tcp.h>
48 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
49 int push_one, gfp_t gfp);
51 /* Account for new data that has been sent to the network. */
52 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
54 struct inet_connection_sock *icsk = inet_csk(sk);
55 struct tcp_sock *tp = tcp_sk(sk);
56 unsigned int prior_packets = tp->packets_out;
58 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
60 __skb_unlink(skb, &sk->sk_write_queue);
61 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
63 tp->packets_out += tcp_skb_pcount(skb);
64 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
67 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
71 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
72 * window scaling factor due to loss of precision.
73 * If window has been shrunk, what should we make? It is not clear at all.
74 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
75 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
76 * invalid. OK, let's make this for now:
78 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
80 const struct tcp_sock *tp = tcp_sk(sk);
82 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
83 (tp->rx_opt.wscale_ok &&
84 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
87 return tcp_wnd_end(tp);
90 /* Calculate mss to advertise in SYN segment.
91 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
93 * 1. It is independent of path mtu.
94 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
95 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
96 * attached devices, because some buggy hosts are confused by
98 * 4. We do not make 3, we advertise MSS, calculated from first
99 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
100 * This may be overridden via information stored in routing table.
101 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
102 * probably even Jumbo".
104 static __u16 tcp_advertise_mss(struct sock *sk)
106 struct tcp_sock *tp = tcp_sk(sk);
107 const struct dst_entry *dst = __sk_dst_get(sk);
108 int mss = tp->advmss;
111 unsigned int metric = dst_metric_advmss(dst);
122 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
123 * This is the first part of cwnd validation mechanism.
125 void tcp_cwnd_restart(struct sock *sk, s32 delta)
127 struct tcp_sock *tp = tcp_sk(sk);
128 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
129 u32 cwnd = tp->snd_cwnd;
131 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
133 tp->snd_ssthresh = tcp_current_ssthresh(sk);
134 restart_cwnd = min(restart_cwnd, cwnd);
136 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
138 tp->snd_cwnd = max(cwnd, restart_cwnd);
139 tp->snd_cwnd_stamp = tcp_jiffies32;
140 tp->snd_cwnd_used = 0;
143 /* Congestion state accounting after a packet has been sent. */
144 static void tcp_event_data_sent(struct tcp_sock *tp,
147 struct inet_connection_sock *icsk = inet_csk(sk);
148 const u32 now = tcp_jiffies32;
150 if (tcp_packets_in_flight(tp) == 0)
151 tcp_ca_event(sk, CA_EVENT_TX_START);
155 /* If it is a reply for ato after last received
156 * packet, enter pingpong mode.
158 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
159 icsk->icsk_ack.pingpong = 1;
162 /* Account for an ACK we sent. */
163 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
165 tcp_dec_quickack_mode(sk, pkts);
166 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
170 u32 tcp_default_init_rwnd(u32 mss)
172 /* Initial receive window should be twice of TCP_INIT_CWND to
173 * enable proper sending of new unsent data during fast recovery
174 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
175 * limit when mss is larger than 1460.
177 u32 init_rwnd = TCP_INIT_CWND * 2;
180 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
184 /* Determine a window scaling and initial window to offer.
185 * Based on the assumption that the given amount of space
186 * will be offered. Store the results in the tp structure.
187 * NOTE: for smooth operation initial space offering should
188 * be a multiple of mss if possible. We assume here that mss >= 1.
189 * This MUST be enforced by all callers.
191 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
192 __u32 *rcv_wnd, __u32 *window_clamp,
193 int wscale_ok, __u8 *rcv_wscale,
196 unsigned int space = (__space < 0 ? 0 : __space);
198 /* If no clamp set the clamp to the max possible scaled window */
199 if (*window_clamp == 0)
200 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
201 space = min(*window_clamp, space);
203 /* Quantize space offering to a multiple of mss if possible. */
205 space = rounddown(space, mss);
207 /* NOTE: offering an initial window larger than 32767
208 * will break some buggy TCP stacks. If the admin tells us
209 * it is likely we could be speaking with such a buggy stack
210 * we will truncate our initial window offering to 32K-1
211 * unless the remote has sent us a window scaling option,
212 * which we interpret as a sign the remote TCP is not
213 * misinterpreting the window field as a signed quantity.
215 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
216 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
222 /* Set window scaling on max possible window */
223 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
224 space = max_t(u32, space, sysctl_rmem_max);
225 space = min_t(u32, space, *window_clamp);
226 while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
232 if (mss > (1 << *rcv_wscale)) {
233 if (!init_rcv_wnd) /* Use default unless specified otherwise */
234 init_rcv_wnd = tcp_default_init_rwnd(mss);
235 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
238 /* Set the clamp no higher than max representable value */
239 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
241 EXPORT_SYMBOL(tcp_select_initial_window);
243 /* Chose a new window to advertise, update state in tcp_sock for the
244 * socket, and return result with RFC1323 scaling applied. The return
245 * value can be stuffed directly into th->window for an outgoing
248 static u16 tcp_select_window(struct sock *sk)
250 struct tcp_sock *tp = tcp_sk(sk);
251 u32 old_win = tp->rcv_wnd;
252 u32 cur_win = tcp_receive_window(tp);
253 u32 new_win = __tcp_select_window(sk);
255 /* Never shrink the offered window */
256 if (new_win < cur_win) {
257 /* Danger Will Robinson!
258 * Don't update rcv_wup/rcv_wnd here or else
259 * we will not be able to advertise a zero
260 * window in time. --DaveM
262 * Relax Will Robinson.
265 NET_INC_STATS(sock_net(sk),
266 LINUX_MIB_TCPWANTZEROWINDOWADV);
267 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
269 tp->rcv_wnd = new_win;
270 tp->rcv_wup = tp->rcv_nxt;
272 /* Make sure we do not exceed the maximum possible
275 if (!tp->rx_opt.rcv_wscale &&
276 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
277 new_win = min(new_win, MAX_TCP_WINDOW);
279 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
281 /* RFC1323 scaling applied */
282 new_win >>= tp->rx_opt.rcv_wscale;
284 /* If we advertise zero window, disable fast path. */
288 NET_INC_STATS(sock_net(sk),
289 LINUX_MIB_TCPTOZEROWINDOWADV);
290 } else if (old_win == 0) {
291 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
297 /* Packet ECN state for a SYN-ACK */
298 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
300 const struct tcp_sock *tp = tcp_sk(sk);
302 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
303 if (!(tp->ecn_flags & TCP_ECN_OK))
304 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
305 else if (tcp_ca_needs_ecn(sk) ||
306 tcp_bpf_ca_needs_ecn(sk))
310 /* Packet ECN state for a SYN. */
311 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
313 struct tcp_sock *tp = tcp_sk(sk);
314 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
315 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
316 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
319 const struct dst_entry *dst = __sk_dst_get(sk);
321 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
328 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
329 tp->ecn_flags = TCP_ECN_OK;
330 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
335 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
337 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
338 /* tp->ecn_flags are cleared at a later point in time when
339 * SYN ACK is ultimatively being received.
341 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
345 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
347 if (inet_rsk(req)->ecn_ok)
351 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
354 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
355 struct tcphdr *th, int tcp_header_len)
357 struct tcp_sock *tp = tcp_sk(sk);
359 if (tp->ecn_flags & TCP_ECN_OK) {
360 /* Not-retransmitted data segment: set ECT and inject CWR. */
361 if (skb->len != tcp_header_len &&
362 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
364 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
365 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
367 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
369 } else if (!tcp_ca_needs_ecn(sk)) {
370 /* ACK or retransmitted segment: clear ECT|CE */
371 INET_ECN_dontxmit(sk);
373 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
378 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
379 * auto increment end seqno.
381 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
383 skb->ip_summed = CHECKSUM_PARTIAL;
385 TCP_SKB_CB(skb)->tcp_flags = flags;
386 TCP_SKB_CB(skb)->sacked = 0;
388 tcp_skb_pcount_set(skb, 1);
390 TCP_SKB_CB(skb)->seq = seq;
391 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
393 TCP_SKB_CB(skb)->end_seq = seq;
396 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
398 return tp->snd_una != tp->snd_up;
401 #define OPTION_SACK_ADVERTISE (1 << 0)
402 #define OPTION_TS (1 << 1)
403 #define OPTION_MD5 (1 << 2)
404 #define OPTION_WSCALE (1 << 3)
405 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
406 #define OPTION_SMC (1 << 9)
408 static void smc_options_write(__be32 *ptr, u16 *options)
410 #if IS_ENABLED(CONFIG_SMC)
411 if (static_branch_unlikely(&tcp_have_smc)) {
412 if (unlikely(OPTION_SMC & *options)) {
413 *ptr++ = htonl((TCPOPT_NOP << 24) |
416 (TCPOLEN_EXP_SMC_BASE));
417 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
423 struct tcp_out_options {
424 u16 options; /* bit field of OPTION_* */
425 u16 mss; /* 0 to disable */
426 u8 ws; /* window scale, 0 to disable */
427 u8 num_sack_blocks; /* number of SACK blocks to include */
428 u8 hash_size; /* bytes in hash_location */
429 __u8 *hash_location; /* temporary pointer, overloaded */
430 __u32 tsval, tsecr; /* need to include OPTION_TS */
431 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
434 /* Write previously computed TCP options to the packet.
436 * Beware: Something in the Internet is very sensitive to the ordering of
437 * TCP options, we learned this through the hard way, so be careful here.
438 * Luckily we can at least blame others for their non-compliance but from
439 * inter-operability perspective it seems that we're somewhat stuck with
440 * the ordering which we have been using if we want to keep working with
441 * those broken things (not that it currently hurts anybody as there isn't
442 * particular reason why the ordering would need to be changed).
444 * At least SACK_PERM as the first option is known to lead to a disaster
445 * (but it may well be that other scenarios fail similarly).
447 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
448 struct tcp_out_options *opts)
450 u16 options = opts->options; /* mungable copy */
452 if (unlikely(OPTION_MD5 & options)) {
453 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
454 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
455 /* overload cookie hash location */
456 opts->hash_location = (__u8 *)ptr;
460 if (unlikely(opts->mss)) {
461 *ptr++ = htonl((TCPOPT_MSS << 24) |
462 (TCPOLEN_MSS << 16) |
466 if (likely(OPTION_TS & options)) {
467 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
468 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
469 (TCPOLEN_SACK_PERM << 16) |
470 (TCPOPT_TIMESTAMP << 8) |
472 options &= ~OPTION_SACK_ADVERTISE;
474 *ptr++ = htonl((TCPOPT_NOP << 24) |
476 (TCPOPT_TIMESTAMP << 8) |
479 *ptr++ = htonl(opts->tsval);
480 *ptr++ = htonl(opts->tsecr);
483 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
484 *ptr++ = htonl((TCPOPT_NOP << 24) |
486 (TCPOPT_SACK_PERM << 8) |
490 if (unlikely(OPTION_WSCALE & options)) {
491 *ptr++ = htonl((TCPOPT_NOP << 24) |
492 (TCPOPT_WINDOW << 16) |
493 (TCPOLEN_WINDOW << 8) |
497 if (unlikely(opts->num_sack_blocks)) {
498 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
499 tp->duplicate_sack : tp->selective_acks;
502 *ptr++ = htonl((TCPOPT_NOP << 24) |
505 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
506 TCPOLEN_SACK_PERBLOCK)));
508 for (this_sack = 0; this_sack < opts->num_sack_blocks;
510 *ptr++ = htonl(sp[this_sack].start_seq);
511 *ptr++ = htonl(sp[this_sack].end_seq);
514 tp->rx_opt.dsack = 0;
517 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
518 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
520 u32 len; /* Fast Open option length */
523 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
524 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
525 TCPOPT_FASTOPEN_MAGIC);
526 p += TCPOLEN_EXP_FASTOPEN_BASE;
528 len = TCPOLEN_FASTOPEN_BASE + foc->len;
529 *p++ = TCPOPT_FASTOPEN;
533 memcpy(p, foc->val, foc->len);
534 if ((len & 3) == 2) {
535 p[foc->len] = TCPOPT_NOP;
536 p[foc->len + 1] = TCPOPT_NOP;
538 ptr += (len + 3) >> 2;
541 smc_options_write(ptr, &options);
544 static void smc_set_option(const struct tcp_sock *tp,
545 struct tcp_out_options *opts,
546 unsigned int *remaining)
548 #if IS_ENABLED(CONFIG_SMC)
549 if (static_branch_unlikely(&tcp_have_smc)) {
551 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
552 opts->options |= OPTION_SMC;
553 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
560 static void smc_set_option_cond(const struct tcp_sock *tp,
561 const struct inet_request_sock *ireq,
562 struct tcp_out_options *opts,
563 unsigned int *remaining)
565 #if IS_ENABLED(CONFIG_SMC)
566 if (static_branch_unlikely(&tcp_have_smc)) {
567 if (tp->syn_smc && ireq->smc_ok) {
568 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
569 opts->options |= OPTION_SMC;
570 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
577 /* Compute TCP options for SYN packets. This is not the final
578 * network wire format yet.
580 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
581 struct tcp_out_options *opts,
582 struct tcp_md5sig_key **md5)
584 struct tcp_sock *tp = tcp_sk(sk);
585 unsigned int remaining = MAX_TCP_OPTION_SPACE;
586 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
588 #ifdef CONFIG_TCP_MD5SIG
589 *md5 = tp->af_specific->md5_lookup(sk, sk);
591 opts->options |= OPTION_MD5;
592 remaining -= TCPOLEN_MD5SIG_ALIGNED;
598 /* We always get an MSS option. The option bytes which will be seen in
599 * normal data packets should timestamps be used, must be in the MSS
600 * advertised. But we subtract them from tp->mss_cache so that
601 * calculations in tcp_sendmsg are simpler etc. So account for this
602 * fact here if necessary. If we don't do this correctly, as a
603 * receiver we won't recognize data packets as being full sized when we
604 * should, and thus we won't abide by the delayed ACK rules correctly.
605 * SACKs don't matter, we never delay an ACK when we have any of those
607 opts->mss = tcp_advertise_mss(sk);
608 remaining -= TCPOLEN_MSS_ALIGNED;
610 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
611 opts->options |= OPTION_TS;
612 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
613 opts->tsecr = tp->rx_opt.ts_recent;
614 remaining -= TCPOLEN_TSTAMP_ALIGNED;
616 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
617 opts->ws = tp->rx_opt.rcv_wscale;
618 opts->options |= OPTION_WSCALE;
619 remaining -= TCPOLEN_WSCALE_ALIGNED;
621 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
622 opts->options |= OPTION_SACK_ADVERTISE;
623 if (unlikely(!(OPTION_TS & opts->options)))
624 remaining -= TCPOLEN_SACKPERM_ALIGNED;
627 if (fastopen && fastopen->cookie.len >= 0) {
628 u32 need = fastopen->cookie.len;
630 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
631 TCPOLEN_FASTOPEN_BASE;
632 need = (need + 3) & ~3U; /* Align to 32 bits */
633 if (remaining >= need) {
634 opts->options |= OPTION_FAST_OPEN_COOKIE;
635 opts->fastopen_cookie = &fastopen->cookie;
637 tp->syn_fastopen = 1;
638 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
642 smc_set_option(tp, opts, &remaining);
644 return MAX_TCP_OPTION_SPACE - remaining;
647 /* Set up TCP options for SYN-ACKs. */
648 static unsigned int tcp_synack_options(const struct sock *sk,
649 struct request_sock *req,
650 unsigned int mss, struct sk_buff *skb,
651 struct tcp_out_options *opts,
652 const struct tcp_md5sig_key *md5,
653 struct tcp_fastopen_cookie *foc)
655 struct inet_request_sock *ireq = inet_rsk(req);
656 unsigned int remaining = MAX_TCP_OPTION_SPACE;
658 #ifdef CONFIG_TCP_MD5SIG
660 opts->options |= OPTION_MD5;
661 remaining -= TCPOLEN_MD5SIG_ALIGNED;
663 /* We can't fit any SACK blocks in a packet with MD5 + TS
664 * options. There was discussion about disabling SACK
665 * rather than TS in order to fit in better with old,
666 * buggy kernels, but that was deemed to be unnecessary.
668 ireq->tstamp_ok &= !ireq->sack_ok;
672 /* We always send an MSS option. */
674 remaining -= TCPOLEN_MSS_ALIGNED;
676 if (likely(ireq->wscale_ok)) {
677 opts->ws = ireq->rcv_wscale;
678 opts->options |= OPTION_WSCALE;
679 remaining -= TCPOLEN_WSCALE_ALIGNED;
681 if (likely(ireq->tstamp_ok)) {
682 opts->options |= OPTION_TS;
683 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
684 opts->tsecr = req->ts_recent;
685 remaining -= TCPOLEN_TSTAMP_ALIGNED;
687 if (likely(ireq->sack_ok)) {
688 opts->options |= OPTION_SACK_ADVERTISE;
689 if (unlikely(!ireq->tstamp_ok))
690 remaining -= TCPOLEN_SACKPERM_ALIGNED;
692 if (foc != NULL && foc->len >= 0) {
695 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
696 TCPOLEN_FASTOPEN_BASE;
697 need = (need + 3) & ~3U; /* Align to 32 bits */
698 if (remaining >= need) {
699 opts->options |= OPTION_FAST_OPEN_COOKIE;
700 opts->fastopen_cookie = foc;
705 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
707 return MAX_TCP_OPTION_SPACE - remaining;
710 /* Compute TCP options for ESTABLISHED sockets. This is not the
711 * final wire format yet.
713 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
714 struct tcp_out_options *opts,
715 struct tcp_md5sig_key **md5)
717 struct tcp_sock *tp = tcp_sk(sk);
718 unsigned int size = 0;
719 unsigned int eff_sacks;
723 #ifdef CONFIG_TCP_MD5SIG
724 *md5 = tp->af_specific->md5_lookup(sk, sk);
725 if (unlikely(*md5)) {
726 opts->options |= OPTION_MD5;
727 size += TCPOLEN_MD5SIG_ALIGNED;
733 if (likely(tp->rx_opt.tstamp_ok)) {
734 opts->options |= OPTION_TS;
735 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
736 opts->tsecr = tp->rx_opt.ts_recent;
737 size += TCPOLEN_TSTAMP_ALIGNED;
740 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
741 if (unlikely(eff_sacks)) {
742 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
743 opts->num_sack_blocks =
744 min_t(unsigned int, eff_sacks,
745 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
746 TCPOLEN_SACK_PERBLOCK);
747 size += TCPOLEN_SACK_BASE_ALIGNED +
748 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
755 /* TCP SMALL QUEUES (TSQ)
757 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
758 * to reduce RTT and bufferbloat.
759 * We do this using a special skb destructor (tcp_wfree).
761 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
762 * needs to be reallocated in a driver.
763 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
765 * Since transmit from skb destructor is forbidden, we use a tasklet
766 * to process all sockets that eventually need to send more skbs.
767 * We use one tasklet per cpu, with its own queue of sockets.
770 struct tasklet_struct tasklet;
771 struct list_head head; /* queue of tcp sockets */
773 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
775 static void tcp_tsq_handler(struct sock *sk)
777 if ((1 << sk->sk_state) &
778 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
779 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
780 struct tcp_sock *tp = tcp_sk(sk);
782 if (tp->lost_out > tp->retrans_out &&
783 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
784 tcp_mstamp_refresh(tp);
785 tcp_xmit_retransmit_queue(sk);
788 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
793 * One tasklet per cpu tries to send more skbs.
794 * We run in tasklet context but need to disable irqs when
795 * transferring tsq->head because tcp_wfree() might
796 * interrupt us (non NAPI drivers)
798 static void tcp_tasklet_func(unsigned long data)
800 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
803 struct list_head *q, *n;
807 local_irq_save(flags);
808 list_splice_init(&tsq->head, &list);
809 local_irq_restore(flags);
811 list_for_each_safe(q, n, &list) {
812 tp = list_entry(q, struct tcp_sock, tsq_node);
813 list_del(&tp->tsq_node);
815 sk = (struct sock *)tp;
816 smp_mb__before_atomic();
817 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
819 if (!sk->sk_lock.owned &&
820 test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) {
822 if (!sock_owned_by_user(sk)) {
823 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
833 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
834 TCPF_WRITE_TIMER_DEFERRED | \
835 TCPF_DELACK_TIMER_DEFERRED | \
836 TCPF_MTU_REDUCED_DEFERRED)
838 * tcp_release_cb - tcp release_sock() callback
841 * called from release_sock() to perform protocol dependent
842 * actions before socket release.
844 void tcp_release_cb(struct sock *sk)
846 unsigned long flags, nflags;
848 /* perform an atomic operation only if at least one flag is set */
850 flags = sk->sk_tsq_flags;
851 if (!(flags & TCP_DEFERRED_ALL))
853 nflags = flags & ~TCP_DEFERRED_ALL;
854 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
856 if (flags & TCPF_TSQ_DEFERRED)
859 /* Here begins the tricky part :
860 * We are called from release_sock() with :
862 * 2) sk_lock.slock spinlock held
863 * 3) socket owned by us (sk->sk_lock.owned == 1)
865 * But following code is meant to be called from BH handlers,
866 * so we should keep BH disabled, but early release socket ownership
868 sock_release_ownership(sk);
870 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
871 tcp_write_timer_handler(sk);
874 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
875 tcp_delack_timer_handler(sk);
878 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
879 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
883 EXPORT_SYMBOL(tcp_release_cb);
885 void __init tcp_tasklet_init(void)
889 for_each_possible_cpu(i) {
890 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
892 INIT_LIST_HEAD(&tsq->head);
893 tasklet_init(&tsq->tasklet,
900 * Write buffer destructor automatically called from kfree_skb.
901 * We can't xmit new skbs from this context, as we might already
904 void tcp_wfree(struct sk_buff *skb)
906 struct sock *sk = skb->sk;
907 struct tcp_sock *tp = tcp_sk(sk);
908 unsigned long flags, nval, oval;
910 /* Keep one reference on sk_wmem_alloc.
911 * Will be released by sk_free() from here or tcp_tasklet_func()
913 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
915 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
916 * Wait until our queues (qdisc + devices) are drained.
918 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
919 * - chance for incoming ACK (processed by another cpu maybe)
920 * to migrate this flow (skb->ooo_okay will be eventually set)
922 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
925 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
926 struct tsq_tasklet *tsq;
929 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
932 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
933 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
937 /* queue this socket to tasklet queue */
938 local_irq_save(flags);
939 tsq = this_cpu_ptr(&tsq_tasklet);
940 empty = list_empty(&tsq->head);
941 list_add(&tp->tsq_node, &tsq->head);
943 tasklet_schedule(&tsq->tasklet);
944 local_irq_restore(flags);
951 /* Note: Called under hard irq.
952 * We can not call TCP stack right away.
954 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
956 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
957 struct sock *sk = (struct sock *)tp;
958 unsigned long nval, oval;
960 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
961 struct tsq_tasklet *tsq;
964 if (oval & TSQF_QUEUED)
967 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
968 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
972 if (!refcount_inc_not_zero(&sk->sk_wmem_alloc))
974 /* queue this socket to tasklet queue */
975 tsq = this_cpu_ptr(&tsq_tasklet);
976 empty = list_empty(&tsq->head);
977 list_add(&tp->tsq_node, &tsq->head);
979 tasklet_schedule(&tsq->tasklet);
982 return HRTIMER_NORESTART;
985 /* BBR congestion control needs pacing.
986 * Same remark for SO_MAX_PACING_RATE.
987 * sch_fq packet scheduler is efficiently handling pacing,
988 * but is not always installed/used.
989 * Return true if TCP stack should pace packets itself.
991 static bool tcp_needs_internal_pacing(const struct sock *sk)
993 return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
996 static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
1001 if (!tcp_needs_internal_pacing(sk))
1003 rate = sk->sk_pacing_rate;
1004 if (!rate || rate == ~0U)
1007 /* Should account for header sizes as sch_fq does,
1008 * but lets make things simple.
1010 len_ns = (u64)skb->len * NSEC_PER_SEC;
1011 do_div(len_ns, rate);
1012 hrtimer_start(&tcp_sk(sk)->pacing_timer,
1013 ktime_add_ns(ktime_get(), len_ns),
1014 HRTIMER_MODE_ABS_PINNED);
1017 static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb)
1019 skb->skb_mstamp = tp->tcp_mstamp;
1020 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1023 /* This routine actually transmits TCP packets queued in by
1024 * tcp_do_sendmsg(). This is used by both the initial
1025 * transmission and possible later retransmissions.
1026 * All SKB's seen here are completely headerless. It is our
1027 * job to build the TCP header, and pass the packet down to
1028 * IP so it can do the same plus pass the packet off to the
1031 * We are working here with either a clone of the original
1032 * SKB, or a fresh unique copy made by the retransmit engine.
1034 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1037 const struct inet_connection_sock *icsk = inet_csk(sk);
1038 struct inet_sock *inet;
1039 struct tcp_sock *tp;
1040 struct tcp_skb_cb *tcb;
1041 struct tcp_out_options opts;
1042 unsigned int tcp_options_size, tcp_header_size;
1043 struct sk_buff *oskb = NULL;
1044 struct tcp_md5sig_key *md5;
1048 BUG_ON(!skb || !tcp_skb_pcount(skb));
1052 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1056 tcp_skb_tsorted_save(oskb) {
1057 if (unlikely(skb_cloned(oskb)))
1058 skb = pskb_copy(oskb, gfp_mask);
1060 skb = skb_clone(oskb, gfp_mask);
1061 } tcp_skb_tsorted_restore(oskb);
1066 skb->skb_mstamp = tp->tcp_mstamp;
1069 tcb = TCP_SKB_CB(skb);
1070 memset(&opts, 0, sizeof(opts));
1072 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1073 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1075 tcp_options_size = tcp_established_options(sk, skb, &opts,
1077 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1079 /* if no packet is in qdisc/device queue, then allow XPS to select
1080 * another queue. We can be called from tcp_tsq_handler()
1081 * which holds one reference to sk_wmem_alloc.
1083 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1084 * One way to get this would be to set skb->truesize = 2 on them.
1086 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1088 /* If we had to use memory reserve to allocate this skb,
1089 * this might cause drops if packet is looped back :
1090 * Other socket might not have SOCK_MEMALLOC.
1091 * Packets not looped back do not care about pfmemalloc.
1093 skb->pfmemalloc = 0;
1095 skb_push(skb, tcp_header_size);
1096 skb_reset_transport_header(skb);
1100 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1101 skb_set_hash_from_sk(skb, sk);
1102 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1104 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1106 /* Build TCP header and checksum it. */
1107 th = (struct tcphdr *)skb->data;
1108 th->source = inet->inet_sport;
1109 th->dest = inet->inet_dport;
1110 th->seq = htonl(tcb->seq);
1111 th->ack_seq = htonl(tp->rcv_nxt);
1112 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1118 /* The urg_mode check is necessary during a below snd_una win probe */
1119 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1120 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1121 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1123 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1124 th->urg_ptr = htons(0xFFFF);
1129 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1130 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1131 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1132 th->window = htons(tcp_select_window(sk));
1133 tcp_ecn_send(sk, skb, th, tcp_header_size);
1135 /* RFC1323: The window in SYN & SYN/ACK segments
1138 th->window = htons(min(tp->rcv_wnd, 65535U));
1140 #ifdef CONFIG_TCP_MD5SIG
1141 /* Calculate the MD5 hash, as we have all we need now */
1143 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1144 tp->af_specific->calc_md5_hash(opts.hash_location,
1149 icsk->icsk_af_ops->send_check(sk, skb);
1151 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1152 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1154 if (skb->len != tcp_header_size) {
1155 tcp_event_data_sent(tp, sk);
1156 tp->data_segs_out += tcp_skb_pcount(skb);
1157 tcp_internal_pacing(sk, skb);
1160 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1161 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1162 tcp_skb_pcount(skb));
1164 tp->segs_out += tcp_skb_pcount(skb);
1165 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1166 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1167 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1169 /* Our usage of tstamp should remain private */
1172 /* Cleanup our debris for IP stacks */
1173 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1174 sizeof(struct inet6_skb_parm)));
1176 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1178 if (unlikely(err > 0)) {
1180 err = net_xmit_eval(err);
1183 tcp_update_skb_after_send(tp, oskb);
1184 tcp_rate_skb_sent(sk, oskb);
1189 /* This routine just queues the buffer for sending.
1191 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1192 * otherwise socket can stall.
1194 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1196 struct tcp_sock *tp = tcp_sk(sk);
1198 /* Advance write_seq and place onto the write_queue. */
1199 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1200 __skb_header_release(skb);
1201 tcp_add_write_queue_tail(sk, skb);
1202 sk->sk_wmem_queued += skb->truesize;
1203 sk_mem_charge(sk, skb->truesize);
1206 /* Initialize TSO segments for a packet. */
1207 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1209 if (skb->len <= mss_now) {
1210 /* Avoid the costly divide in the normal
1213 tcp_skb_pcount_set(skb, 1);
1214 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1216 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1217 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1221 /* Pcount in the middle of the write queue got changed, we need to do various
1222 * tweaks to fix counters
1224 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1226 struct tcp_sock *tp = tcp_sk(sk);
1228 tp->packets_out -= decr;
1230 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1231 tp->sacked_out -= decr;
1232 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1233 tp->retrans_out -= decr;
1234 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1235 tp->lost_out -= decr;
1237 /* Reno case is special. Sigh... */
1238 if (tcp_is_reno(tp) && decr > 0)
1239 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1241 if (tp->lost_skb_hint &&
1242 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1243 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1244 tp->lost_cnt_hint -= decr;
1246 tcp_verify_left_out(tp);
1249 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1251 return TCP_SKB_CB(skb)->txstamp_ack ||
1252 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1255 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1257 struct skb_shared_info *shinfo = skb_shinfo(skb);
1259 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1260 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1261 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1262 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1264 shinfo->tx_flags &= ~tsflags;
1265 shinfo2->tx_flags |= tsflags;
1266 swap(shinfo->tskey, shinfo2->tskey);
1267 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1268 TCP_SKB_CB(skb)->txstamp_ack = 0;
1272 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1274 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1275 TCP_SKB_CB(skb)->eor = 0;
1278 /* Insert buff after skb on the write or rtx queue of sk. */
1279 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1280 struct sk_buff *buff,
1282 enum tcp_queue tcp_queue)
1284 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1285 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1287 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1290 /* Function to create two new TCP segments. Shrinks the given segment
1291 * to the specified size and appends a new segment with the rest of the
1292 * packet to the list. This won't be called frequently, I hope.
1293 * Remember, these are still headerless SKBs at this point.
1295 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1296 struct sk_buff *skb, u32 len,
1297 unsigned int mss_now, gfp_t gfp)
1299 struct tcp_sock *tp = tcp_sk(sk);
1300 struct sk_buff *buff;
1301 int nsize, old_factor;
1305 if (WARN_ON(len > skb->len))
1308 nsize = skb_headlen(skb) - len;
1312 if (skb_unclone(skb, gfp))
1315 /* Get a new skb... force flag on. */
1316 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1318 return -ENOMEM; /* We'll just try again later. */
1320 sk->sk_wmem_queued += buff->truesize;
1321 sk_mem_charge(sk, buff->truesize);
1322 nlen = skb->len - len - nsize;
1323 buff->truesize += nlen;
1324 skb->truesize -= nlen;
1326 /* Correct the sequence numbers. */
1327 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1328 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1329 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1331 /* PSH and FIN should only be set in the second packet. */
1332 flags = TCP_SKB_CB(skb)->tcp_flags;
1333 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1334 TCP_SKB_CB(buff)->tcp_flags = flags;
1335 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1336 tcp_skb_fragment_eor(skb, buff);
1338 skb_split(skb, buff, len);
1340 buff->ip_summed = CHECKSUM_PARTIAL;
1342 buff->tstamp = skb->tstamp;
1343 tcp_fragment_tstamp(skb, buff);
1345 old_factor = tcp_skb_pcount(skb);
1347 /* Fix up tso_factor for both original and new SKB. */
1348 tcp_set_skb_tso_segs(skb, mss_now);
1349 tcp_set_skb_tso_segs(buff, mss_now);
1351 /* Update delivered info for the new segment */
1352 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1354 /* If this packet has been sent out already, we must
1355 * adjust the various packet counters.
1357 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1358 int diff = old_factor - tcp_skb_pcount(skb) -
1359 tcp_skb_pcount(buff);
1362 tcp_adjust_pcount(sk, skb, diff);
1365 /* Link BUFF into the send queue. */
1366 __skb_header_release(buff);
1367 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1368 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1369 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1374 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1375 * data is not copied, but immediately discarded.
1377 static int __pskb_trim_head(struct sk_buff *skb, int len)
1379 struct skb_shared_info *shinfo;
1382 eat = min_t(int, len, skb_headlen(skb));
1384 __skb_pull(skb, eat);
1391 shinfo = skb_shinfo(skb);
1392 for (i = 0; i < shinfo->nr_frags; i++) {
1393 int size = skb_frag_size(&shinfo->frags[i]);
1396 skb_frag_unref(skb, i);
1399 shinfo->frags[k] = shinfo->frags[i];
1401 shinfo->frags[k].page_offset += eat;
1402 skb_frag_size_sub(&shinfo->frags[k], eat);
1408 shinfo->nr_frags = k;
1410 skb->data_len -= len;
1411 skb->len = skb->data_len;
1415 /* Remove acked data from a packet in the transmit queue. */
1416 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1420 if (skb_unclone(skb, GFP_ATOMIC))
1423 delta_truesize = __pskb_trim_head(skb, len);
1425 TCP_SKB_CB(skb)->seq += len;
1426 skb->ip_summed = CHECKSUM_PARTIAL;
1428 if (delta_truesize) {
1429 skb->truesize -= delta_truesize;
1430 sk->sk_wmem_queued -= delta_truesize;
1431 sk_mem_uncharge(sk, delta_truesize);
1432 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1435 /* Any change of skb->len requires recalculation of tso factor. */
1436 if (tcp_skb_pcount(skb) > 1)
1437 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1442 /* Calculate MSS not accounting any TCP options. */
1443 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1445 const struct tcp_sock *tp = tcp_sk(sk);
1446 const struct inet_connection_sock *icsk = inet_csk(sk);
1449 /* Calculate base mss without TCP options:
1450 It is MMS_S - sizeof(tcphdr) of rfc1122
1452 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1454 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1455 if (icsk->icsk_af_ops->net_frag_header_len) {
1456 const struct dst_entry *dst = __sk_dst_get(sk);
1458 if (dst && dst_allfrag(dst))
1459 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1462 /* Clamp it (mss_clamp does not include tcp options) */
1463 if (mss_now > tp->rx_opt.mss_clamp)
1464 mss_now = tp->rx_opt.mss_clamp;
1466 /* Now subtract optional transport overhead */
1467 mss_now -= icsk->icsk_ext_hdr_len;
1469 /* Then reserve room for full set of TCP options and 8 bytes of data */
1475 /* Calculate MSS. Not accounting for SACKs here. */
1476 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1478 /* Subtract TCP options size, not including SACKs */
1479 return __tcp_mtu_to_mss(sk, pmtu) -
1480 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1483 /* Inverse of above */
1484 int tcp_mss_to_mtu(struct sock *sk, int mss)
1486 const struct tcp_sock *tp = tcp_sk(sk);
1487 const struct inet_connection_sock *icsk = inet_csk(sk);
1491 tp->tcp_header_len +
1492 icsk->icsk_ext_hdr_len +
1493 icsk->icsk_af_ops->net_header_len;
1495 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1496 if (icsk->icsk_af_ops->net_frag_header_len) {
1497 const struct dst_entry *dst = __sk_dst_get(sk);
1499 if (dst && dst_allfrag(dst))
1500 mtu += icsk->icsk_af_ops->net_frag_header_len;
1504 EXPORT_SYMBOL(tcp_mss_to_mtu);
1506 /* MTU probing init per socket */
1507 void tcp_mtup_init(struct sock *sk)
1509 struct tcp_sock *tp = tcp_sk(sk);
1510 struct inet_connection_sock *icsk = inet_csk(sk);
1511 struct net *net = sock_net(sk);
1513 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1514 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1515 icsk->icsk_af_ops->net_header_len;
1516 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1517 icsk->icsk_mtup.probe_size = 0;
1518 if (icsk->icsk_mtup.enabled)
1519 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1521 EXPORT_SYMBOL(tcp_mtup_init);
1523 /* This function synchronize snd mss to current pmtu/exthdr set.
1525 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1526 for TCP options, but includes only bare TCP header.
1528 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1529 It is minimum of user_mss and mss received with SYN.
1530 It also does not include TCP options.
1532 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1534 tp->mss_cache is current effective sending mss, including
1535 all tcp options except for SACKs. It is evaluated,
1536 taking into account current pmtu, but never exceeds
1537 tp->rx_opt.mss_clamp.
1539 NOTE1. rfc1122 clearly states that advertised MSS
1540 DOES NOT include either tcp or ip options.
1542 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1543 are READ ONLY outside this function. --ANK (980731)
1545 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1547 struct tcp_sock *tp = tcp_sk(sk);
1548 struct inet_connection_sock *icsk = inet_csk(sk);
1551 if (icsk->icsk_mtup.search_high > pmtu)
1552 icsk->icsk_mtup.search_high = pmtu;
1554 mss_now = tcp_mtu_to_mss(sk, pmtu);
1555 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1557 /* And store cached results */
1558 icsk->icsk_pmtu_cookie = pmtu;
1559 if (icsk->icsk_mtup.enabled)
1560 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1561 tp->mss_cache = mss_now;
1565 EXPORT_SYMBOL(tcp_sync_mss);
1567 /* Compute the current effective MSS, taking SACKs and IP options,
1568 * and even PMTU discovery events into account.
1570 unsigned int tcp_current_mss(struct sock *sk)
1572 const struct tcp_sock *tp = tcp_sk(sk);
1573 const struct dst_entry *dst = __sk_dst_get(sk);
1575 unsigned int header_len;
1576 struct tcp_out_options opts;
1577 struct tcp_md5sig_key *md5;
1579 mss_now = tp->mss_cache;
1582 u32 mtu = dst_mtu(dst);
1583 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1584 mss_now = tcp_sync_mss(sk, mtu);
1587 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1588 sizeof(struct tcphdr);
1589 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1590 * some common options. If this is an odd packet (because we have SACK
1591 * blocks etc) then our calculated header_len will be different, and
1592 * we have to adjust mss_now correspondingly */
1593 if (header_len != tp->tcp_header_len) {
1594 int delta = (int) header_len - tp->tcp_header_len;
1601 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1602 * As additional protections, we do not touch cwnd in retransmission phases,
1603 * and if application hit its sndbuf limit recently.
1605 static void tcp_cwnd_application_limited(struct sock *sk)
1607 struct tcp_sock *tp = tcp_sk(sk);
1609 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1610 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1611 /* Limited by application or receiver window. */
1612 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1613 u32 win_used = max(tp->snd_cwnd_used, init_win);
1614 if (win_used < tp->snd_cwnd) {
1615 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1616 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1618 tp->snd_cwnd_used = 0;
1620 tp->snd_cwnd_stamp = tcp_jiffies32;
1623 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1625 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1626 struct tcp_sock *tp = tcp_sk(sk);
1628 /* Track the maximum number of outstanding packets in each
1629 * window, and remember whether we were cwnd-limited then.
1631 if (!before(tp->snd_una, tp->max_packets_seq) ||
1632 tp->packets_out > tp->max_packets_out) {
1633 tp->max_packets_out = tp->packets_out;
1634 tp->max_packets_seq = tp->snd_nxt;
1635 tp->is_cwnd_limited = is_cwnd_limited;
1638 if (tcp_is_cwnd_limited(sk)) {
1639 /* Network is feed fully. */
1640 tp->snd_cwnd_used = 0;
1641 tp->snd_cwnd_stamp = tcp_jiffies32;
1643 /* Network starves. */
1644 if (tp->packets_out > tp->snd_cwnd_used)
1645 tp->snd_cwnd_used = tp->packets_out;
1647 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1648 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1649 !ca_ops->cong_control)
1650 tcp_cwnd_application_limited(sk);
1652 /* The following conditions together indicate the starvation
1653 * is caused by insufficient sender buffer:
1654 * 1) just sent some data (see tcp_write_xmit)
1655 * 2) not cwnd limited (this else condition)
1656 * 3) no more data to send (tcp_write_queue_empty())
1657 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1659 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1660 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1661 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1662 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1666 /* Minshall's variant of the Nagle send check. */
1667 static bool tcp_minshall_check(const struct tcp_sock *tp)
1669 return after(tp->snd_sml, tp->snd_una) &&
1670 !after(tp->snd_sml, tp->snd_nxt);
1673 /* Update snd_sml if this skb is under mss
1674 * Note that a TSO packet might end with a sub-mss segment
1675 * The test is really :
1676 * if ((skb->len % mss) != 0)
1677 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1678 * But we can avoid doing the divide again given we already have
1679 * skb_pcount = skb->len / mss_now
1681 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1682 const struct sk_buff *skb)
1684 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1685 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1688 /* Return false, if packet can be sent now without violation Nagle's rules:
1689 * 1. It is full sized. (provided by caller in %partial bool)
1690 * 2. Or it contains FIN. (already checked by caller)
1691 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1692 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1693 * With Minshall's modification: all sent small packets are ACKed.
1695 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1699 ((nonagle & TCP_NAGLE_CORK) ||
1700 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1703 /* Return how many segs we'd like on a TSO packet,
1704 * to send one TSO packet per ms
1706 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1711 bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift,
1712 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1714 /* Goal is to send at least one packet per ms,
1715 * not one big TSO packet every 100 ms.
1716 * This preserves ACK clocking and is consistent
1717 * with tcp_tso_should_defer() heuristic.
1719 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1724 /* Return the number of segments we want in the skb we are transmitting.
1725 * See if congestion control module wants to decide; otherwise, autosize.
1727 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1729 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1730 u32 min_tso, tso_segs;
1732 min_tso = ca_ops->min_tso_segs ?
1733 ca_ops->min_tso_segs(sk) :
1734 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1736 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1737 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1740 /* Returns the portion of skb which can be sent right away */
1741 static unsigned int tcp_mss_split_point(const struct sock *sk,
1742 const struct sk_buff *skb,
1743 unsigned int mss_now,
1744 unsigned int max_segs,
1747 const struct tcp_sock *tp = tcp_sk(sk);
1748 u32 partial, needed, window, max_len;
1750 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1751 max_len = mss_now * max_segs;
1753 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1756 needed = min(skb->len, window);
1758 if (max_len <= needed)
1761 partial = needed % mss_now;
1762 /* If last segment is not a full MSS, check if Nagle rules allow us
1763 * to include this last segment in this skb.
1764 * Otherwise, we'll split the skb at last MSS boundary
1766 if (tcp_nagle_check(partial != 0, tp, nonagle))
1767 return needed - partial;
1772 /* Can at least one segment of SKB be sent right now, according to the
1773 * congestion window rules? If so, return how many segments are allowed.
1775 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1776 const struct sk_buff *skb)
1778 u32 in_flight, cwnd, halfcwnd;
1780 /* Don't be strict about the congestion window for the final FIN. */
1781 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1782 tcp_skb_pcount(skb) == 1)
1785 in_flight = tcp_packets_in_flight(tp);
1786 cwnd = tp->snd_cwnd;
1787 if (in_flight >= cwnd)
1790 /* For better scheduling, ensure we have at least
1791 * 2 GSO packets in flight.
1793 halfcwnd = max(cwnd >> 1, 1U);
1794 return min(halfcwnd, cwnd - in_flight);
1797 /* Initialize TSO state of a skb.
1798 * This must be invoked the first time we consider transmitting
1799 * SKB onto the wire.
1801 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1803 int tso_segs = tcp_skb_pcount(skb);
1805 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1806 tcp_set_skb_tso_segs(skb, mss_now);
1807 tso_segs = tcp_skb_pcount(skb);
1813 /* Return true if the Nagle test allows this packet to be
1816 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1817 unsigned int cur_mss, int nonagle)
1819 /* Nagle rule does not apply to frames, which sit in the middle of the
1820 * write_queue (they have no chances to get new data).
1822 * This is implemented in the callers, where they modify the 'nonagle'
1823 * argument based upon the location of SKB in the send queue.
1825 if (nonagle & TCP_NAGLE_PUSH)
1828 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1829 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1832 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1838 /* Does at least the first segment of SKB fit into the send window? */
1839 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1840 const struct sk_buff *skb,
1841 unsigned int cur_mss)
1843 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1845 if (skb->len > cur_mss)
1846 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1848 return !after(end_seq, tcp_wnd_end(tp));
1851 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1852 * which is put after SKB on the list. It is very much like
1853 * tcp_fragment() except that it may make several kinds of assumptions
1854 * in order to speed up the splitting operation. In particular, we
1855 * know that all the data is in scatter-gather pages, and that the
1856 * packet has never been sent out before (and thus is not cloned).
1858 static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1859 struct sk_buff *skb, unsigned int len,
1860 unsigned int mss_now, gfp_t gfp)
1862 struct sk_buff *buff;
1863 int nlen = skb->len - len;
1866 /* All of a TSO frame must be composed of paged data. */
1867 if (skb->len != skb->data_len)
1868 return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp);
1870 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1871 if (unlikely(!buff))
1874 sk->sk_wmem_queued += buff->truesize;
1875 sk_mem_charge(sk, buff->truesize);
1876 buff->truesize += nlen;
1877 skb->truesize -= nlen;
1879 /* Correct the sequence numbers. */
1880 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1881 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1882 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1884 /* PSH and FIN should only be set in the second packet. */
1885 flags = TCP_SKB_CB(skb)->tcp_flags;
1886 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1887 TCP_SKB_CB(buff)->tcp_flags = flags;
1889 /* This packet was never sent out yet, so no SACK bits. */
1890 TCP_SKB_CB(buff)->sacked = 0;
1892 tcp_skb_fragment_eor(skb, buff);
1894 buff->ip_summed = CHECKSUM_PARTIAL;
1895 skb_split(skb, buff, len);
1896 tcp_fragment_tstamp(skb, buff);
1898 /* Fix up tso_factor for both original and new SKB. */
1899 tcp_set_skb_tso_segs(skb, mss_now);
1900 tcp_set_skb_tso_segs(buff, mss_now);
1902 /* Link BUFF into the send queue. */
1903 __skb_header_release(buff);
1904 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1909 /* Try to defer sending, if possible, in order to minimize the amount
1910 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1912 * This algorithm is from John Heffner.
1914 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1915 bool *is_cwnd_limited, u32 max_segs)
1917 const struct inet_connection_sock *icsk = inet_csk(sk);
1918 u32 age, send_win, cong_win, limit, in_flight;
1919 struct tcp_sock *tp = tcp_sk(sk);
1920 struct sk_buff *head;
1923 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1926 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1929 /* Avoid bursty behavior by allowing defer
1930 * only if the last write was recent.
1932 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
1935 in_flight = tcp_packets_in_flight(tp);
1937 BUG_ON(tcp_skb_pcount(skb) <= 1);
1938 BUG_ON(tp->snd_cwnd <= in_flight);
1940 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1942 /* From in_flight test above, we know that cwnd > in_flight. */
1943 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1945 limit = min(send_win, cong_win);
1947 /* If a full-sized TSO skb can be sent, do it. */
1948 if (limit >= max_segs * tp->mss_cache)
1951 /* Middle in queue won't get any more data, full sendable already? */
1952 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1955 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1957 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1959 /* If at least some fraction of a window is available,
1962 chunk /= win_divisor;
1966 /* Different approach, try not to defer past a single
1967 * ACK. Receiver should ACK every other full sized
1968 * frame, so if we have space for more than 3 frames
1971 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1975 /* TODO : use tsorted_sent_queue ? */
1976 head = tcp_rtx_queue_head(sk);
1979 age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
1980 /* If next ACK is likely to come too late (half srtt), do not defer */
1981 if (age < (tp->srtt_us >> 4))
1984 /* Ok, it looks like it is advisable to defer. */
1986 if (cong_win < send_win && cong_win <= skb->len)
1987 *is_cwnd_limited = true;
1995 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1997 struct inet_connection_sock *icsk = inet_csk(sk);
1998 struct tcp_sock *tp = tcp_sk(sk);
1999 struct net *net = sock_net(sk);
2003 interval = net->ipv4.sysctl_tcp_probe_interval;
2004 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2005 if (unlikely(delta >= interval * HZ)) {
2006 int mss = tcp_current_mss(sk);
2008 /* Update current search range */
2009 icsk->icsk_mtup.probe_size = 0;
2010 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2011 sizeof(struct tcphdr) +
2012 icsk->icsk_af_ops->net_header_len;
2013 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2015 /* Update probe time stamp */
2016 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2020 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2022 struct sk_buff *skb, *next;
2024 skb = tcp_send_head(sk);
2025 tcp_for_write_queue_from_safe(skb, next, sk) {
2026 if (len <= skb->len)
2029 if (unlikely(TCP_SKB_CB(skb)->eor))
2038 /* Create a new MTU probe if we are ready.
2039 * MTU probe is regularly attempting to increase the path MTU by
2040 * deliberately sending larger packets. This discovers routing
2041 * changes resulting in larger path MTUs.
2043 * Returns 0 if we should wait to probe (no cwnd available),
2044 * 1 if a probe was sent,
2047 static int tcp_mtu_probe(struct sock *sk)
2049 struct inet_connection_sock *icsk = inet_csk(sk);
2050 struct tcp_sock *tp = tcp_sk(sk);
2051 struct sk_buff *skb, *nskb, *next;
2052 struct net *net = sock_net(sk);
2059 /* Not currently probing/verifying,
2061 * have enough cwnd, and
2062 * not SACKing (the variable headers throw things off)
2064 if (likely(!icsk->icsk_mtup.enabled ||
2065 icsk->icsk_mtup.probe_size ||
2066 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2067 tp->snd_cwnd < 11 ||
2068 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2071 /* Use binary search for probe_size between tcp_mss_base,
2072 * and current mss_clamp. if (search_high - search_low)
2073 * smaller than a threshold, backoff from probing.
2075 mss_now = tcp_current_mss(sk);
2076 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2077 icsk->icsk_mtup.search_low) >> 1);
2078 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2079 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2080 /* When misfortune happens, we are reprobing actively,
2081 * and then reprobe timer has expired. We stick with current
2082 * probing process by not resetting search range to its orignal.
2084 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2085 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2086 /* Check whether enough time has elaplased for
2087 * another round of probing.
2089 tcp_mtu_check_reprobe(sk);
2093 /* Have enough data in the send queue to probe? */
2094 if (tp->write_seq - tp->snd_nxt < size_needed)
2097 if (tp->snd_wnd < size_needed)
2099 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2102 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2103 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2104 if (!tcp_packets_in_flight(tp))
2110 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2113 /* We're allowed to probe. Build it now. */
2114 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2117 sk->sk_wmem_queued += nskb->truesize;
2118 sk_mem_charge(sk, nskb->truesize);
2120 skb = tcp_send_head(sk);
2122 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2123 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2124 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2125 TCP_SKB_CB(nskb)->sacked = 0;
2127 nskb->ip_summed = CHECKSUM_PARTIAL;
2129 tcp_insert_write_queue_before(nskb, skb, sk);
2130 tcp_highest_sack_replace(sk, skb, nskb);
2133 tcp_for_write_queue_from_safe(skb, next, sk) {
2134 copy = min_t(int, skb->len, probe_size - len);
2135 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2137 if (skb->len <= copy) {
2138 /* We've eaten all the data from this skb.
2140 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2141 /* If this is the last SKB we copy and eor is set
2142 * we need to propagate it to the new skb.
2144 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2145 tcp_unlink_write_queue(skb, sk);
2146 sk_wmem_free_skb(sk, skb);
2148 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2149 ~(TCPHDR_FIN|TCPHDR_PSH);
2150 if (!skb_shinfo(skb)->nr_frags) {
2151 skb_pull(skb, copy);
2153 __pskb_trim_head(skb, copy);
2154 tcp_set_skb_tso_segs(skb, mss_now);
2156 TCP_SKB_CB(skb)->seq += copy;
2161 if (len >= probe_size)
2164 tcp_init_tso_segs(nskb, nskb->len);
2166 /* We're ready to send. If this fails, the probe will
2167 * be resegmented into mss-sized pieces by tcp_write_xmit().
2169 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2170 /* Decrement cwnd here because we are sending
2171 * effectively two packets. */
2173 tcp_event_new_data_sent(sk, nskb);
2175 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2176 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2177 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2185 static bool tcp_pacing_check(const struct sock *sk)
2187 return tcp_needs_internal_pacing(sk) &&
2188 hrtimer_active(&tcp_sk(sk)->pacing_timer);
2191 /* TCP Small Queues :
2192 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2193 * (These limits are doubled for retransmits)
2195 * - better RTT estimation and ACK scheduling
2198 * Alas, some drivers / subsystems require a fair amount
2199 * of queued bytes to ensure line rate.
2200 * One example is wifi aggregation (802.11 AMPDU)
2202 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2203 unsigned int factor)
2207 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift);
2208 limit = min_t(u32, limit,
2209 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2212 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2213 /* Always send skb if rtx queue is empty.
2214 * No need to wait for TX completion to call us back,
2215 * after softirq/tasklet schedule.
2216 * This helps when TX completions are delayed too much.
2218 if (tcp_rtx_queue_empty(sk))
2221 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2222 /* It is possible TX completion already happened
2223 * before we set TSQ_THROTTLED, so we must
2224 * test again the condition.
2226 smp_mb__after_atomic();
2227 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2233 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2235 const u32 now = tcp_jiffies32;
2236 enum tcp_chrono old = tp->chrono_type;
2238 if (old > TCP_CHRONO_UNSPEC)
2239 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2240 tp->chrono_start = now;
2241 tp->chrono_type = new;
2244 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2246 struct tcp_sock *tp = tcp_sk(sk);
2248 /* If there are multiple conditions worthy of tracking in a
2249 * chronograph then the highest priority enum takes precedence
2250 * over the other conditions. So that if something "more interesting"
2251 * starts happening, stop the previous chrono and start a new one.
2253 if (type > tp->chrono_type)
2254 tcp_chrono_set(tp, type);
2257 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2259 struct tcp_sock *tp = tcp_sk(sk);
2262 /* There are multiple conditions worthy of tracking in a
2263 * chronograph, so that the highest priority enum takes
2264 * precedence over the other conditions (see tcp_chrono_start).
2265 * If a condition stops, we only stop chrono tracking if
2266 * it's the "most interesting" or current chrono we are
2267 * tracking and starts busy chrono if we have pending data.
2269 if (tcp_rtx_and_write_queues_empty(sk))
2270 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2271 else if (type == tp->chrono_type)
2272 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2275 /* This routine writes packets to the network. It advances the
2276 * send_head. This happens as incoming acks open up the remote
2279 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2280 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2281 * account rare use of URG, this is not a big flaw.
2283 * Send at most one packet when push_one > 0. Temporarily ignore
2284 * cwnd limit to force at most one packet out when push_one == 2.
2286 * Returns true, if no segments are in flight and we have queued segments,
2287 * but cannot send anything now because of SWS or another problem.
2289 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2290 int push_one, gfp_t gfp)
2292 struct tcp_sock *tp = tcp_sk(sk);
2293 struct sk_buff *skb;
2294 unsigned int tso_segs, sent_pkts;
2297 bool is_cwnd_limited = false, is_rwnd_limited = false;
2302 tcp_mstamp_refresh(tp);
2304 /* Do MTU probing. */
2305 result = tcp_mtu_probe(sk);
2308 } else if (result > 0) {
2313 max_segs = tcp_tso_segs(sk, mss_now);
2314 while ((skb = tcp_send_head(sk))) {
2317 if (tcp_pacing_check(sk))
2320 tso_segs = tcp_init_tso_segs(skb, mss_now);
2323 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2324 /* "skb_mstamp" is used as a start point for the retransmit timer */
2325 tcp_update_skb_after_send(tp, skb);
2326 goto repair; /* Skip network transmission */
2329 cwnd_quota = tcp_cwnd_test(tp, skb);
2332 /* Force out a loss probe pkt. */
2338 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2339 is_rwnd_limited = true;
2343 if (tso_segs == 1) {
2344 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2345 (tcp_skb_is_last(sk, skb) ?
2346 nonagle : TCP_NAGLE_PUSH))))
2350 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2356 if (tso_segs > 1 && !tcp_urg_mode(tp))
2357 limit = tcp_mss_split_point(sk, skb, mss_now,
2363 if (skb->len > limit &&
2364 unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2365 skb, limit, mss_now, gfp)))
2368 if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
2369 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
2370 if (tcp_small_queue_check(sk, skb, 0))
2373 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2377 /* Advance the send_head. This one is sent out.
2378 * This call will increment packets_out.
2380 tcp_event_new_data_sent(sk, skb);
2382 tcp_minshall_update(tp, mss_now, skb);
2383 sent_pkts += tcp_skb_pcount(skb);
2389 if (is_rwnd_limited)
2390 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2392 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2394 if (likely(sent_pkts)) {
2395 if (tcp_in_cwnd_reduction(sk))
2396 tp->prr_out += sent_pkts;
2398 /* Send one loss probe per tail loss episode. */
2400 tcp_schedule_loss_probe(sk, false);
2401 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2402 tcp_cwnd_validate(sk, is_cwnd_limited);
2405 return !tp->packets_out && !tcp_write_queue_empty(sk);
2408 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2410 struct inet_connection_sock *icsk = inet_csk(sk);
2411 struct tcp_sock *tp = tcp_sk(sk);
2412 u32 timeout, rto_delta_us;
2415 /* Don't do any loss probe on a Fast Open connection before 3WHS
2418 if (tp->fastopen_rsk)
2421 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2422 /* Schedule a loss probe in 2*RTT for SACK capable connections
2423 * not in loss recovery, that are either limited by cwnd or application.
2425 if ((early_retrans != 3 && early_retrans != 4) ||
2426 !tp->packets_out || !tcp_is_sack(tp) ||
2427 (icsk->icsk_ca_state != TCP_CA_Open &&
2428 icsk->icsk_ca_state != TCP_CA_CWR))
2431 /* Probe timeout is 2*rtt. Add minimum RTO to account
2432 * for delayed ack when there's one outstanding packet. If no RTT
2433 * sample is available then probe after TCP_TIMEOUT_INIT.
2436 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2437 if (tp->packets_out == 1)
2438 timeout += TCP_RTO_MIN;
2440 timeout += TCP_TIMEOUT_MIN;
2442 timeout = TCP_TIMEOUT_INIT;
2445 /* If the RTO formula yields an earlier time, then use that time. */
2446 rto_delta_us = advancing_rto ?
2447 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2448 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2449 if (rto_delta_us > 0)
2450 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2452 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2457 /* Thanks to skb fast clones, we can detect if a prior transmit of
2458 * a packet is still in a qdisc or driver queue.
2459 * In this case, there is very little point doing a retransmit !
2461 static bool skb_still_in_host_queue(const struct sock *sk,
2462 const struct sk_buff *skb)
2464 if (unlikely(skb_fclone_busy(sk, skb))) {
2465 NET_INC_STATS(sock_net(sk),
2466 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2472 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2473 * retransmit the last segment.
2475 void tcp_send_loss_probe(struct sock *sk)
2477 struct tcp_sock *tp = tcp_sk(sk);
2478 struct sk_buff *skb;
2480 int mss = tcp_current_mss(sk);
2482 skb = tcp_send_head(sk);
2483 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2484 pcount = tp->packets_out;
2485 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2486 if (tp->packets_out > pcount)
2490 skb = skb_rb_last(&sk->tcp_rtx_queue);
2492 /* At most one outstanding TLP retransmission. */
2493 if (tp->tlp_high_seq)
2496 /* Retransmit last segment. */
2500 if (skb_still_in_host_queue(sk, skb))
2503 pcount = tcp_skb_pcount(skb);
2504 if (WARN_ON(!pcount))
2507 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2508 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2509 (pcount - 1) * mss, mss,
2512 skb = skb_rb_next(skb);
2515 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2518 if (__tcp_retransmit_skb(sk, skb, 1))
2521 /* Record snd_nxt for loss detection. */
2522 tp->tlp_high_seq = tp->snd_nxt;
2525 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2526 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2527 inet_csk(sk)->icsk_pending = 0;
2532 /* Push out any pending frames which were held back due to
2533 * TCP_CORK or attempt at coalescing tiny packets.
2534 * The socket must be locked by the caller.
2536 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2539 /* If we are closed, the bytes will have to remain here.
2540 * In time closedown will finish, we empty the write queue and
2541 * all will be happy.
2543 if (unlikely(sk->sk_state == TCP_CLOSE))
2546 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2547 sk_gfp_mask(sk, GFP_ATOMIC)))
2548 tcp_check_probe_timer(sk);
2551 /* Send _single_ skb sitting at the send head. This function requires
2552 * true push pending frames to setup probe timer etc.
2554 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2556 struct sk_buff *skb = tcp_send_head(sk);
2558 BUG_ON(!skb || skb->len < mss_now);
2560 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2563 /* This function returns the amount that we can raise the
2564 * usable window based on the following constraints
2566 * 1. The window can never be shrunk once it is offered (RFC 793)
2567 * 2. We limit memory per socket
2570 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2571 * RECV.NEXT + RCV.WIN fixed until:
2572 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2574 * i.e. don't raise the right edge of the window until you can raise
2575 * it at least MSS bytes.
2577 * Unfortunately, the recommended algorithm breaks header prediction,
2578 * since header prediction assumes th->window stays fixed.
2580 * Strictly speaking, keeping th->window fixed violates the receiver
2581 * side SWS prevention criteria. The problem is that under this rule
2582 * a stream of single byte packets will cause the right side of the
2583 * window to always advance by a single byte.
2585 * Of course, if the sender implements sender side SWS prevention
2586 * then this will not be a problem.
2588 * BSD seems to make the following compromise:
2590 * If the free space is less than the 1/4 of the maximum
2591 * space available and the free space is less than 1/2 mss,
2592 * then set the window to 0.
2593 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2594 * Otherwise, just prevent the window from shrinking
2595 * and from being larger than the largest representable value.
2597 * This prevents incremental opening of the window in the regime
2598 * where TCP is limited by the speed of the reader side taking
2599 * data out of the TCP receive queue. It does nothing about
2600 * those cases where the window is constrained on the sender side
2601 * because the pipeline is full.
2603 * BSD also seems to "accidentally" limit itself to windows that are a
2604 * multiple of MSS, at least until the free space gets quite small.
2605 * This would appear to be a side effect of the mbuf implementation.
2606 * Combining these two algorithms results in the observed behavior
2607 * of having a fixed window size at almost all times.
2609 * Below we obtain similar behavior by forcing the offered window to
2610 * a multiple of the mss when it is feasible to do so.
2612 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2613 * Regular options like TIMESTAMP are taken into account.
2615 u32 __tcp_select_window(struct sock *sk)
2617 struct inet_connection_sock *icsk = inet_csk(sk);
2618 struct tcp_sock *tp = tcp_sk(sk);
2619 /* MSS for the peer's data. Previous versions used mss_clamp
2620 * here. I don't know if the value based on our guesses
2621 * of peer's MSS is better for the performance. It's more correct
2622 * but may be worse for the performance because of rcv_mss
2623 * fluctuations. --SAW 1998/11/1
2625 int mss = icsk->icsk_ack.rcv_mss;
2626 int free_space = tcp_space(sk);
2627 int allowed_space = tcp_full_space(sk);
2628 int full_space = min_t(int, tp->window_clamp, allowed_space);
2631 if (unlikely(mss > full_space)) {
2636 if (free_space < (full_space >> 1)) {
2637 icsk->icsk_ack.quick = 0;
2639 if (tcp_under_memory_pressure(sk))
2640 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2643 /* free_space might become our new window, make sure we don't
2644 * increase it due to wscale.
2646 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2648 /* if free space is less than mss estimate, or is below 1/16th
2649 * of the maximum allowed, try to move to zero-window, else
2650 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2651 * new incoming data is dropped due to memory limits.
2652 * With large window, mss test triggers way too late in order
2653 * to announce zero window in time before rmem limit kicks in.
2655 if (free_space < (allowed_space >> 4) || free_space < mss)
2659 if (free_space > tp->rcv_ssthresh)
2660 free_space = tp->rcv_ssthresh;
2662 /* Don't do rounding if we are using window scaling, since the
2663 * scaled window will not line up with the MSS boundary anyway.
2665 if (tp->rx_opt.rcv_wscale) {
2666 window = free_space;
2668 /* Advertise enough space so that it won't get scaled away.
2669 * Import case: prevent zero window announcement if
2670 * 1<<rcv_wscale > mss.
2672 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2674 window = tp->rcv_wnd;
2675 /* Get the largest window that is a nice multiple of mss.
2676 * Window clamp already applied above.
2677 * If our current window offering is within 1 mss of the
2678 * free space we just keep it. This prevents the divide
2679 * and multiply from happening most of the time.
2680 * We also don't do any window rounding when the free space
2683 if (window <= free_space - mss || window > free_space)
2684 window = rounddown(free_space, mss);
2685 else if (mss == full_space &&
2686 free_space > window + (full_space >> 1))
2687 window = free_space;
2693 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2694 const struct sk_buff *next_skb)
2696 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2697 const struct skb_shared_info *next_shinfo =
2698 skb_shinfo(next_skb);
2699 struct skb_shared_info *shinfo = skb_shinfo(skb);
2701 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2702 shinfo->tskey = next_shinfo->tskey;
2703 TCP_SKB_CB(skb)->txstamp_ack |=
2704 TCP_SKB_CB(next_skb)->txstamp_ack;
2708 /* Collapses two adjacent SKB's during retransmission. */
2709 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2711 struct tcp_sock *tp = tcp_sk(sk);
2712 struct sk_buff *next_skb = skb_rb_next(skb);
2713 int skb_size, next_skb_size;
2715 skb_size = skb->len;
2716 next_skb_size = next_skb->len;
2718 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2720 if (next_skb_size) {
2721 if (next_skb_size <= skb_availroom(skb))
2722 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2724 else if (!skb_shift(skb, next_skb, next_skb_size))
2727 tcp_highest_sack_replace(sk, next_skb, skb);
2729 /* Update sequence range on original skb. */
2730 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2732 /* Merge over control information. This moves PSH/FIN etc. over */
2733 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2735 /* All done, get rid of second SKB and account for it so
2736 * packet counting does not break.
2738 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2739 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2741 /* changed transmit queue under us so clear hints */
2742 tcp_clear_retrans_hints_partial(tp);
2743 if (next_skb == tp->retransmit_skb_hint)
2744 tp->retransmit_skb_hint = skb;
2746 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2748 tcp_skb_collapse_tstamp(skb, next_skb);
2750 tcp_rtx_queue_unlink_and_free(next_skb, sk);
2754 /* Check if coalescing SKBs is legal. */
2755 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2757 if (tcp_skb_pcount(skb) > 1)
2759 if (skb_cloned(skb))
2761 /* Some heuristics for collapsing over SACK'd could be invented */
2762 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2768 /* Collapse packets in the retransmit queue to make to create
2769 * less packets on the wire. This is only done on retransmission.
2771 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2774 struct tcp_sock *tp = tcp_sk(sk);
2775 struct sk_buff *skb = to, *tmp;
2778 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2780 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2783 skb_rbtree_walk_from_safe(skb, tmp) {
2784 if (!tcp_can_collapse(sk, skb))
2787 if (!tcp_skb_can_collapse_to(to))
2800 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2803 if (!tcp_collapse_retrans(sk, to))
2808 /* This retransmits one SKB. Policy decisions and retransmit queue
2809 * state updates are done by the caller. Returns non-zero if an
2810 * error occurred which prevented the send.
2812 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2814 struct inet_connection_sock *icsk = inet_csk(sk);
2815 struct tcp_sock *tp = tcp_sk(sk);
2816 unsigned int cur_mss;
2820 /* Inconclusive MTU probe */
2821 if (icsk->icsk_mtup.probe_size)
2822 icsk->icsk_mtup.probe_size = 0;
2824 /* Do not sent more than we queued. 1/4 is reserved for possible
2825 * copying overhead: fragmentation, tunneling, mangling etc.
2827 if (refcount_read(&sk->sk_wmem_alloc) >
2828 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2832 if (skb_still_in_host_queue(sk, skb))
2835 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2836 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2838 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2842 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2843 return -EHOSTUNREACH; /* Routing failure or similar. */
2845 cur_mss = tcp_current_mss(sk);
2847 /* If receiver has shrunk his window, and skb is out of
2848 * new window, do not retransmit it. The exception is the
2849 * case, when window is shrunk to zero. In this case
2850 * our retransmit serves as a zero window probe.
2852 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2853 TCP_SKB_CB(skb)->seq != tp->snd_una)
2856 len = cur_mss * segs;
2857 if (skb->len > len) {
2858 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2859 cur_mss, GFP_ATOMIC))
2860 return -ENOMEM; /* We'll try again later. */
2862 if (skb_unclone(skb, GFP_ATOMIC))
2865 diff = tcp_skb_pcount(skb);
2866 tcp_set_skb_tso_segs(skb, cur_mss);
2867 diff -= tcp_skb_pcount(skb);
2869 tcp_adjust_pcount(sk, skb, diff);
2870 if (skb->len < cur_mss)
2871 tcp_retrans_try_collapse(sk, skb, cur_mss);
2874 /* RFC3168, section 6.1.1.1. ECN fallback */
2875 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2876 tcp_ecn_clear_syn(sk, skb);
2878 /* Update global and local TCP statistics. */
2879 segs = tcp_skb_pcount(skb);
2880 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2881 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2882 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2883 tp->total_retrans += segs;
2885 /* make sure skb->data is aligned on arches that require it
2886 * and check if ack-trimming & collapsing extended the headroom
2887 * beyond what csum_start can cover.
2889 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2890 skb_headroom(skb) >= 0xFFFF)) {
2891 struct sk_buff *nskb;
2893 tcp_skb_tsorted_save(skb) {
2894 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2895 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2897 } tcp_skb_tsorted_restore(skb);
2900 tcp_update_skb_after_send(tp, skb);
2901 tcp_rate_skb_sent(sk, skb);
2904 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2907 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
2908 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
2909 TCP_SKB_CB(skb)->seq, segs, err);
2912 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2913 trace_tcp_retransmit_skb(sk, skb);
2914 } else if (err != -EBUSY) {
2915 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2920 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2922 struct tcp_sock *tp = tcp_sk(sk);
2923 int err = __tcp_retransmit_skb(sk, skb, segs);
2926 #if FASTRETRANS_DEBUG > 0
2927 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2928 net_dbg_ratelimited("retrans_out leaked\n");
2931 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2932 tp->retrans_out += tcp_skb_pcount(skb);
2934 /* Save stamp of the first retransmit. */
2935 if (!tp->retrans_stamp)
2936 tp->retrans_stamp = tcp_skb_timestamp(skb);
2940 if (tp->undo_retrans < 0)
2941 tp->undo_retrans = 0;
2942 tp->undo_retrans += tcp_skb_pcount(skb);
2946 /* This gets called after a retransmit timeout, and the initially
2947 * retransmitted data is acknowledged. It tries to continue
2948 * resending the rest of the retransmit queue, until either
2949 * we've sent it all or the congestion window limit is reached.
2951 void tcp_xmit_retransmit_queue(struct sock *sk)
2953 const struct inet_connection_sock *icsk = inet_csk(sk);
2954 struct sk_buff *skb, *rtx_head, *hole = NULL;
2955 struct tcp_sock *tp = tcp_sk(sk);
2959 if (!tp->packets_out)
2962 rtx_head = tcp_rtx_queue_head(sk);
2963 skb = tp->retransmit_skb_hint ?: rtx_head;
2964 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2965 skb_rbtree_walk_from(skb) {
2969 if (tcp_pacing_check(sk))
2972 /* we could do better than to assign each time */
2974 tp->retransmit_skb_hint = skb;
2976 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2979 sacked = TCP_SKB_CB(skb)->sacked;
2980 /* In case tcp_shift_skb_data() have aggregated large skbs,
2981 * we need to make sure not sending too bigs TSO packets
2983 segs = min_t(int, segs, max_segs);
2985 if (tp->retrans_out >= tp->lost_out) {
2987 } else if (!(sacked & TCPCB_LOST)) {
2988 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2993 if (icsk->icsk_ca_state != TCP_CA_Loss)
2994 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2996 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2999 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3002 if (tcp_small_queue_check(sk, skb, 1))
3005 if (tcp_retransmit_skb(sk, skb, segs))
3008 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3010 if (tcp_in_cwnd_reduction(sk))
3011 tp->prr_out += tcp_skb_pcount(skb);
3013 if (skb == rtx_head &&
3014 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3015 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3016 inet_csk(sk)->icsk_rto,
3021 /* We allow to exceed memory limits for FIN packets to expedite
3022 * connection tear down and (memory) recovery.
3023 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3024 * or even be forced to close flow without any FIN.
3025 * In general, we want to allow one skb per socket to avoid hangs
3026 * with edge trigger epoll()
3028 void sk_forced_mem_schedule(struct sock *sk, int size)
3032 if (size <= sk->sk_forward_alloc)
3034 amt = sk_mem_pages(size);
3035 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3036 sk_memory_allocated_add(sk, amt);
3038 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3039 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3042 /* Send a FIN. The caller locks the socket for us.
3043 * We should try to send a FIN packet really hard, but eventually give up.
3045 void tcp_send_fin(struct sock *sk)
3047 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3048 struct tcp_sock *tp = tcp_sk(sk);
3050 /* Optimization, tack on the FIN if we have one skb in write queue and
3051 * this skb was not yet sent, or we are under memory pressure.
3052 * Note: in the latter case, FIN packet will be sent after a timeout,
3053 * as TCP stack thinks it has already been transmitted.
3055 if (!tskb && tcp_under_memory_pressure(sk))
3056 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3060 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3061 TCP_SKB_CB(tskb)->end_seq++;
3063 if (tcp_write_queue_empty(sk)) {
3064 /* This means tskb was already sent.
3065 * Pretend we included the FIN on previous transmit.
3066 * We need to set tp->snd_nxt to the value it would have
3067 * if FIN had been sent. This is because retransmit path
3068 * does not change tp->snd_nxt.
3074 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3075 if (unlikely(!skb)) {
3080 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3081 skb_reserve(skb, MAX_TCP_HEADER);
3082 sk_forced_mem_schedule(sk, skb->truesize);
3083 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3084 tcp_init_nondata_skb(skb, tp->write_seq,
3085 TCPHDR_ACK | TCPHDR_FIN);
3086 tcp_queue_skb(sk, skb);
3088 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3091 /* We get here when a process closes a file descriptor (either due to
3092 * an explicit close() or as a byproduct of exit()'ing) and there
3093 * was unread data in the receive queue. This behavior is recommended
3094 * by RFC 2525, section 2.17. -DaveM
3096 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3098 struct sk_buff *skb;
3100 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3102 /* NOTE: No TCP options attached and we never retransmit this. */
3103 skb = alloc_skb(MAX_TCP_HEADER, priority);
3105 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3109 /* Reserve space for headers and prepare control bits. */
3110 skb_reserve(skb, MAX_TCP_HEADER);
3111 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3112 TCPHDR_ACK | TCPHDR_RST);
3113 tcp_mstamp_refresh(tcp_sk(sk));
3115 if (tcp_transmit_skb(sk, skb, 0, priority))
3116 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3118 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3119 * skb here is different to the troublesome skb, so use NULL
3121 trace_tcp_send_reset(sk, NULL);
3124 /* Send a crossed SYN-ACK during socket establishment.
3125 * WARNING: This routine must only be called when we have already sent
3126 * a SYN packet that crossed the incoming SYN that caused this routine
3127 * to get called. If this assumption fails then the initial rcv_wnd
3128 * and rcv_wscale values will not be correct.
3130 int tcp_send_synack(struct sock *sk)
3132 struct sk_buff *skb;
3134 skb = tcp_rtx_queue_head(sk);
3135 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3136 pr_err("%s: wrong queue state\n", __func__);
3139 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3140 if (skb_cloned(skb)) {
3141 struct sk_buff *nskb;
3143 tcp_skb_tsorted_save(skb) {
3144 nskb = skb_copy(skb, GFP_ATOMIC);
3145 } tcp_skb_tsorted_restore(skb);
3148 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3149 tcp_rtx_queue_unlink_and_free(skb, sk);
3150 __skb_header_release(nskb);
3151 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3152 sk->sk_wmem_queued += nskb->truesize;
3153 sk_mem_charge(sk, nskb->truesize);
3157 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3158 tcp_ecn_send_synack(sk, skb);
3160 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3164 * tcp_make_synack - Prepare a SYN-ACK.
3165 * sk: listener socket
3166 * dst: dst entry attached to the SYNACK
3167 * req: request_sock pointer
3169 * Allocate one skb and build a SYNACK packet.
3170 * @dst is consumed : Caller should not use it again.
3172 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3173 struct request_sock *req,
3174 struct tcp_fastopen_cookie *foc,
3175 enum tcp_synack_type synack_type)
3177 struct inet_request_sock *ireq = inet_rsk(req);
3178 const struct tcp_sock *tp = tcp_sk(sk);
3179 struct tcp_md5sig_key *md5 = NULL;
3180 struct tcp_out_options opts;
3181 struct sk_buff *skb;
3182 int tcp_header_size;
3186 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3187 if (unlikely(!skb)) {
3191 /* Reserve space for headers. */
3192 skb_reserve(skb, MAX_TCP_HEADER);
3194 switch (synack_type) {
3195 case TCP_SYNACK_NORMAL:
3196 skb_set_owner_w(skb, req_to_sk(req));
3198 case TCP_SYNACK_COOKIE:
3199 /* Under synflood, we do not attach skb to a socket,
3200 * to avoid false sharing.
3203 case TCP_SYNACK_FASTOPEN:
3204 /* sk is a const pointer, because we want to express multiple
3205 * cpu might call us concurrently.
3206 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3208 skb_set_owner_w(skb, (struct sock *)sk);
3211 skb_dst_set(skb, dst);
3213 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3215 memset(&opts, 0, sizeof(opts));
3216 #ifdef CONFIG_SYN_COOKIES
3217 if (unlikely(req->cookie_ts))
3218 skb->skb_mstamp = cookie_init_timestamp(req);
3221 skb->skb_mstamp = tcp_clock_us();
3223 #ifdef CONFIG_TCP_MD5SIG
3225 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3227 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3228 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3231 skb_push(skb, tcp_header_size);
3232 skb_reset_transport_header(skb);
3234 th = (struct tcphdr *)skb->data;
3235 memset(th, 0, sizeof(struct tcphdr));
3238 tcp_ecn_make_synack(req, th);
3239 th->source = htons(ireq->ir_num);
3240 th->dest = ireq->ir_rmt_port;
3241 skb->mark = ireq->ir_mark;
3242 skb->ip_summed = CHECKSUM_PARTIAL;
3243 th->seq = htonl(tcp_rsk(req)->snt_isn);
3244 /* XXX data is queued and acked as is. No buffer/window check */
3245 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3247 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3248 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3249 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3250 th->doff = (tcp_header_size >> 2);
3251 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3253 #ifdef CONFIG_TCP_MD5SIG
3254 /* Okay, we have all we need - do the md5 hash if needed */
3256 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3257 md5, req_to_sk(req), skb);
3261 /* Do not fool tcpdump (if any), clean our debris */
3265 EXPORT_SYMBOL(tcp_make_synack);
3267 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3269 struct inet_connection_sock *icsk = inet_csk(sk);
3270 const struct tcp_congestion_ops *ca;
3271 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3273 if (ca_key == TCP_CA_UNSPEC)
3277 ca = tcp_ca_find_key(ca_key);
3278 if (likely(ca && try_module_get(ca->owner))) {
3279 module_put(icsk->icsk_ca_ops->owner);
3280 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3281 icsk->icsk_ca_ops = ca;
3286 /* Do all connect socket setups that can be done AF independent. */
3287 static void tcp_connect_init(struct sock *sk)
3289 const struct dst_entry *dst = __sk_dst_get(sk);
3290 struct tcp_sock *tp = tcp_sk(sk);
3294 /* We'll fix this up when we get a response from the other end.
3295 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3297 tp->tcp_header_len = sizeof(struct tcphdr);
3298 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3299 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3301 #ifdef CONFIG_TCP_MD5SIG
3302 if (tp->af_specific->md5_lookup(sk, sk))
3303 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3306 /* If user gave his TCP_MAXSEG, record it to clamp */
3307 if (tp->rx_opt.user_mss)
3308 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3311 tcp_sync_mss(sk, dst_mtu(dst));
3313 tcp_ca_dst_init(sk, dst);
3315 if (!tp->window_clamp)
3316 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3317 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3319 tcp_initialize_rcv_mss(sk);
3321 /* limit the window selection if the user enforce a smaller rx buffer */
3322 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3323 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3324 tp->window_clamp = tcp_full_space(sk);
3326 rcv_wnd = tcp_rwnd_init_bpf(sk);
3328 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3330 tcp_select_initial_window(sk, tcp_full_space(sk),
3331 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3334 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3338 tp->rx_opt.rcv_wscale = rcv_wscale;
3339 tp->rcv_ssthresh = tp->rcv_wnd;
3342 sock_reset_flag(sk, SOCK_DONE);
3345 tp->snd_una = tp->write_seq;
3346 tp->snd_sml = tp->write_seq;
3347 tp->snd_up = tp->write_seq;
3348 tp->snd_nxt = tp->write_seq;
3350 if (likely(!tp->repair))
3353 tp->rcv_tstamp = tcp_jiffies32;
3354 tp->rcv_wup = tp->rcv_nxt;
3355 tp->copied_seq = tp->rcv_nxt;
3357 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3358 inet_csk(sk)->icsk_retransmits = 0;
3359 tcp_clear_retrans(tp);
3362 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3364 struct tcp_sock *tp = tcp_sk(sk);
3365 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3367 tcb->end_seq += skb->len;
3368 __skb_header_release(skb);
3369 sk->sk_wmem_queued += skb->truesize;
3370 sk_mem_charge(sk, skb->truesize);
3371 tp->write_seq = tcb->end_seq;
3372 tp->packets_out += tcp_skb_pcount(skb);
3375 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3376 * queue a data-only packet after the regular SYN, such that regular SYNs
3377 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3378 * only the SYN sequence, the data are retransmitted in the first ACK.
3379 * If cookie is not cached or other error occurs, falls back to send a
3380 * regular SYN with Fast Open cookie request option.
3382 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3384 struct tcp_sock *tp = tcp_sk(sk);
3385 struct tcp_fastopen_request *fo = tp->fastopen_req;
3387 struct sk_buff *syn_data;
3389 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3390 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3393 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3394 * user-MSS. Reserve maximum option space for middleboxes that add
3395 * private TCP options. The cost is reduced data space in SYN :(
3397 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3399 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3400 MAX_TCP_OPTION_SPACE;
3402 space = min_t(size_t, space, fo->size);
3404 /* limit to order-0 allocations */
3405 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3407 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3410 syn_data->ip_summed = CHECKSUM_PARTIAL;
3411 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3413 int copied = copy_from_iter(skb_put(syn_data, space), space,
3414 &fo->data->msg_iter);
3415 if (unlikely(!copied)) {
3416 tcp_skb_tsorted_anchor_cleanup(syn_data);
3417 kfree_skb(syn_data);
3420 if (copied != space) {
3421 skb_trim(syn_data, copied);
3425 /* No more data pending in inet_wait_for_connect() */
3426 if (space == fo->size)
3430 tcp_connect_queue_skb(sk, syn_data);
3432 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3434 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3436 syn->skb_mstamp = syn_data->skb_mstamp;
3438 /* Now full SYN+DATA was cloned and sent (or not),
3439 * remove the SYN from the original skb (syn_data)
3440 * we keep in write queue in case of a retransmit, as we
3441 * also have the SYN packet (with no data) in the same queue.
3443 TCP_SKB_CB(syn_data)->seq++;
3444 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3446 tp->syn_data = (fo->copied > 0);
3447 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3448 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3452 /* data was not sent, put it in write_queue */
3453 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3454 tp->packets_out -= tcp_skb_pcount(syn_data);
3457 /* Send a regular SYN with Fast Open cookie request option */
3458 if (fo->cookie.len > 0)
3460 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3462 tp->syn_fastopen = 0;
3464 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3468 /* Build a SYN and send it off. */
3469 int tcp_connect(struct sock *sk)
3471 struct tcp_sock *tp = tcp_sk(sk);
3472 struct sk_buff *buff;
3475 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3477 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3478 return -EHOSTUNREACH; /* Routing failure or similar. */
3480 tcp_connect_init(sk);
3482 if (unlikely(tp->repair)) {
3483 tcp_finish_connect(sk, NULL);
3487 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3488 if (unlikely(!buff))
3491 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3492 tcp_mstamp_refresh(tp);
3493 tp->retrans_stamp = tcp_time_stamp(tp);
3494 tcp_connect_queue_skb(sk, buff);
3495 tcp_ecn_send_syn(sk, buff);
3496 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3498 /* Send off SYN; include data in Fast Open. */
3499 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3500 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3501 if (err == -ECONNREFUSED)
3504 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3505 * in order to make this packet get counted in tcpOutSegs.
3507 tp->snd_nxt = tp->write_seq;
3508 tp->pushed_seq = tp->write_seq;
3509 buff = tcp_send_head(sk);
3510 if (unlikely(buff)) {
3511 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3512 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3514 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3516 /* Timer for repeating the SYN until an answer. */
3517 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3518 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3521 EXPORT_SYMBOL(tcp_connect);
3523 /* Send out a delayed ack, the caller does the policy checking
3524 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3527 void tcp_send_delayed_ack(struct sock *sk)
3529 struct inet_connection_sock *icsk = inet_csk(sk);
3530 int ato = icsk->icsk_ack.ato;
3531 unsigned long timeout;
3533 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3535 if (ato > TCP_DELACK_MIN) {
3536 const struct tcp_sock *tp = tcp_sk(sk);
3537 int max_ato = HZ / 2;
3539 if (icsk->icsk_ack.pingpong ||
3540 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3541 max_ato = TCP_DELACK_MAX;
3543 /* Slow path, intersegment interval is "high". */
3545 /* If some rtt estimate is known, use it to bound delayed ack.
3546 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3550 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3557 ato = min(ato, max_ato);
3560 /* Stay within the limit we were given */
3561 timeout = jiffies + ato;
3563 /* Use new timeout only if there wasn't a older one earlier. */
3564 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3565 /* If delack timer was blocked or is about to expire,
3568 if (icsk->icsk_ack.blocked ||
3569 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3574 if (!time_before(timeout, icsk->icsk_ack.timeout))
3575 timeout = icsk->icsk_ack.timeout;
3577 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3578 icsk->icsk_ack.timeout = timeout;
3579 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3582 /* This routine sends an ack and also updates the window. */
3583 void tcp_send_ack(struct sock *sk)
3585 struct sk_buff *buff;
3587 /* If we have been reset, we may not send again. */
3588 if (sk->sk_state == TCP_CLOSE)
3591 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3593 /* We are not putting this on the write queue, so
3594 * tcp_transmit_skb() will set the ownership to this
3597 buff = alloc_skb(MAX_TCP_HEADER,
3598 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3599 if (unlikely(!buff)) {
3600 inet_csk_schedule_ack(sk);
3601 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3602 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3603 TCP_DELACK_MAX, TCP_RTO_MAX);
3607 /* Reserve space for headers and prepare control bits. */
3608 skb_reserve(buff, MAX_TCP_HEADER);
3609 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3611 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3613 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3615 skb_set_tcp_pure_ack(buff);
3617 /* Send it off, this clears delayed acks for us. */
3618 tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3620 EXPORT_SYMBOL_GPL(tcp_send_ack);
3622 /* This routine sends a packet with an out of date sequence
3623 * number. It assumes the other end will try to ack it.
3625 * Question: what should we make while urgent mode?
3626 * 4.4BSD forces sending single byte of data. We cannot send
3627 * out of window data, because we have SND.NXT==SND.MAX...
3629 * Current solution: to send TWO zero-length segments in urgent mode:
3630 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3631 * out-of-date with SND.UNA-1 to probe window.
3633 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3635 struct tcp_sock *tp = tcp_sk(sk);
3636 struct sk_buff *skb;
3638 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3639 skb = alloc_skb(MAX_TCP_HEADER,
3640 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3644 /* Reserve space for headers and set control bits. */
3645 skb_reserve(skb, MAX_TCP_HEADER);
3646 /* Use a previous sequence. This should cause the other
3647 * end to send an ack. Don't queue or clone SKB, just
3650 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3651 NET_INC_STATS(sock_net(sk), mib);
3652 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3655 /* Called from setsockopt( ... TCP_REPAIR ) */
3656 void tcp_send_window_probe(struct sock *sk)
3658 if (sk->sk_state == TCP_ESTABLISHED) {
3659 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3660 tcp_mstamp_refresh(tcp_sk(sk));
3661 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3665 /* Initiate keepalive or window probe from timer. */
3666 int tcp_write_wakeup(struct sock *sk, int mib)
3668 struct tcp_sock *tp = tcp_sk(sk);
3669 struct sk_buff *skb;
3671 if (sk->sk_state == TCP_CLOSE)
3674 skb = tcp_send_head(sk);
3675 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3677 unsigned int mss = tcp_current_mss(sk);
3678 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3680 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3681 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3683 /* We are probing the opening of a window
3684 * but the window size is != 0
3685 * must have been a result SWS avoidance ( sender )
3687 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3689 seg_size = min(seg_size, mss);
3690 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3691 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3692 skb, seg_size, mss, GFP_ATOMIC))
3694 } else if (!tcp_skb_pcount(skb))
3695 tcp_set_skb_tso_segs(skb, mss);
3697 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3698 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3700 tcp_event_new_data_sent(sk, skb);
3703 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3704 tcp_xmit_probe_skb(sk, 1, mib);
3705 return tcp_xmit_probe_skb(sk, 0, mib);
3709 /* A window probe timeout has occurred. If window is not closed send
3710 * a partial packet else a zero probe.
3712 void tcp_send_probe0(struct sock *sk)
3714 struct inet_connection_sock *icsk = inet_csk(sk);
3715 struct tcp_sock *tp = tcp_sk(sk);
3716 struct net *net = sock_net(sk);
3717 unsigned long probe_max;
3720 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3722 if (tp->packets_out || tcp_write_queue_empty(sk)) {
3723 /* Cancel probe timer, if it is not required. */
3724 icsk->icsk_probes_out = 0;
3725 icsk->icsk_backoff = 0;
3730 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3731 icsk->icsk_backoff++;
3732 icsk->icsk_probes_out++;
3733 probe_max = TCP_RTO_MAX;
3735 /* If packet was not sent due to local congestion,
3736 * do not backoff and do not remember icsk_probes_out.
3737 * Let local senders to fight for local resources.
3739 * Use accumulated backoff yet.
3741 if (!icsk->icsk_probes_out)
3742 icsk->icsk_probes_out = 1;
3743 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3745 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3746 tcp_probe0_when(sk, probe_max),
3750 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3752 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3756 tcp_rsk(req)->txhash = net_tx_rndhash();
3757 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3759 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3760 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3761 if (unlikely(tcp_passive_fastopen(sk)))
3762 tcp_sk(sk)->total_retrans++;
3763 trace_tcp_retransmit_synack(sk, req);
3767 EXPORT_SYMBOL(tcp_rtx_synack);