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,
166 struct tcp_sock *tp = tcp_sk(sk);
168 if (unlikely(tp->compressed_ack)) {
169 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
171 tp->compressed_ack = 0;
172 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
176 if (unlikely(rcv_nxt != tp->rcv_nxt))
177 return; /* Special ACK sent by DCTCP to reflect ECN */
178 tcp_dec_quickack_mode(sk, pkts);
179 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
183 u32 tcp_default_init_rwnd(u32 mss)
185 /* Initial receive window should be twice of TCP_INIT_CWND to
186 * enable proper sending of new unsent data during fast recovery
187 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
188 * limit when mss is larger than 1460.
190 u32 init_rwnd = TCP_INIT_CWND * 2;
193 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
197 /* Determine a window scaling and initial window to offer.
198 * Based on the assumption that the given amount of space
199 * will be offered. Store the results in the tp structure.
200 * NOTE: for smooth operation initial space offering should
201 * be a multiple of mss if possible. We assume here that mss >= 1.
202 * This MUST be enforced by all callers.
204 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
205 __u32 *rcv_wnd, __u32 *window_clamp,
206 int wscale_ok, __u8 *rcv_wscale,
209 unsigned int space = (__space < 0 ? 0 : __space);
211 /* If no clamp set the clamp to the max possible scaled window */
212 if (*window_clamp == 0)
213 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
214 space = min(*window_clamp, space);
216 /* Quantize space offering to a multiple of mss if possible. */
218 space = rounddown(space, mss);
220 /* NOTE: offering an initial window larger than 32767
221 * will break some buggy TCP stacks. If the admin tells us
222 * it is likely we could be speaking with such a buggy stack
223 * we will truncate our initial window offering to 32K-1
224 * unless the remote has sent us a window scaling option,
225 * which we interpret as a sign the remote TCP is not
226 * misinterpreting the window field as a signed quantity.
228 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
229 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
235 /* Set window scaling on max possible window */
236 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
237 space = max_t(u32, space, sysctl_rmem_max);
238 space = min_t(u32, space, *window_clamp);
239 while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
245 if (!init_rcv_wnd) /* Use default unless specified otherwise */
246 init_rcv_wnd = tcp_default_init_rwnd(mss);
247 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
249 /* Set the clamp no higher than max representable value */
250 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
252 EXPORT_SYMBOL(tcp_select_initial_window);
254 /* Chose a new window to advertise, update state in tcp_sock for the
255 * socket, and return result with RFC1323 scaling applied. The return
256 * value can be stuffed directly into th->window for an outgoing
259 static u16 tcp_select_window(struct sock *sk)
261 struct tcp_sock *tp = tcp_sk(sk);
262 u32 old_win = tp->rcv_wnd;
263 u32 cur_win = tcp_receive_window(tp);
264 u32 new_win = __tcp_select_window(sk);
266 /* Never shrink the offered window */
267 if (new_win < cur_win) {
268 /* Danger Will Robinson!
269 * Don't update rcv_wup/rcv_wnd here or else
270 * we will not be able to advertise a zero
271 * window in time. --DaveM
273 * Relax Will Robinson.
276 NET_INC_STATS(sock_net(sk),
277 LINUX_MIB_TCPWANTZEROWINDOWADV);
278 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
280 tp->rcv_wnd = new_win;
281 tp->rcv_wup = tp->rcv_nxt;
283 /* Make sure we do not exceed the maximum possible
286 if (!tp->rx_opt.rcv_wscale &&
287 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
288 new_win = min(new_win, MAX_TCP_WINDOW);
290 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
292 /* RFC1323 scaling applied */
293 new_win >>= tp->rx_opt.rcv_wscale;
295 /* If we advertise zero window, disable fast path. */
299 NET_INC_STATS(sock_net(sk),
300 LINUX_MIB_TCPTOZEROWINDOWADV);
301 } else if (old_win == 0) {
302 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
308 /* Packet ECN state for a SYN-ACK */
309 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
311 const struct tcp_sock *tp = tcp_sk(sk);
313 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
314 if (!(tp->ecn_flags & TCP_ECN_OK))
315 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
316 else if (tcp_ca_needs_ecn(sk) ||
317 tcp_bpf_ca_needs_ecn(sk))
321 /* Packet ECN state for a SYN. */
322 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
324 struct tcp_sock *tp = tcp_sk(sk);
325 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
326 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
327 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
330 const struct dst_entry *dst = __sk_dst_get(sk);
332 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
339 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
340 tp->ecn_flags = TCP_ECN_OK;
341 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
346 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
348 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
349 /* tp->ecn_flags are cleared at a later point in time when
350 * SYN ACK is ultimatively being received.
352 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
356 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
358 if (inet_rsk(req)->ecn_ok)
362 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
365 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
366 struct tcphdr *th, int tcp_header_len)
368 struct tcp_sock *tp = tcp_sk(sk);
370 if (tp->ecn_flags & TCP_ECN_OK) {
371 /* Not-retransmitted data segment: set ECT and inject CWR. */
372 if (skb->len != tcp_header_len &&
373 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
375 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
376 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
378 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
380 } else if (!tcp_ca_needs_ecn(sk)) {
381 /* ACK or retransmitted segment: clear ECT|CE */
382 INET_ECN_dontxmit(sk);
384 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
389 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
390 * auto increment end seqno.
392 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
394 skb->ip_summed = CHECKSUM_PARTIAL;
396 TCP_SKB_CB(skb)->tcp_flags = flags;
397 TCP_SKB_CB(skb)->sacked = 0;
399 tcp_skb_pcount_set(skb, 1);
401 TCP_SKB_CB(skb)->seq = seq;
402 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
404 TCP_SKB_CB(skb)->end_seq = seq;
407 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
409 return tp->snd_una != tp->snd_up;
412 #define OPTION_SACK_ADVERTISE (1 << 0)
413 #define OPTION_TS (1 << 1)
414 #define OPTION_MD5 (1 << 2)
415 #define OPTION_WSCALE (1 << 3)
416 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
417 #define OPTION_SMC (1 << 9)
419 static void smc_options_write(__be32 *ptr, u16 *options)
421 #if IS_ENABLED(CONFIG_SMC)
422 if (static_branch_unlikely(&tcp_have_smc)) {
423 if (unlikely(OPTION_SMC & *options)) {
424 *ptr++ = htonl((TCPOPT_NOP << 24) |
427 (TCPOLEN_EXP_SMC_BASE));
428 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
434 struct tcp_out_options {
435 u16 options; /* bit field of OPTION_* */
436 u16 mss; /* 0 to disable */
437 u8 ws; /* window scale, 0 to disable */
438 u8 num_sack_blocks; /* number of SACK blocks to include */
439 u8 hash_size; /* bytes in hash_location */
440 __u8 *hash_location; /* temporary pointer, overloaded */
441 __u32 tsval, tsecr; /* need to include OPTION_TS */
442 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
445 /* Write previously computed TCP options to the packet.
447 * Beware: Something in the Internet is very sensitive to the ordering of
448 * TCP options, we learned this through the hard way, so be careful here.
449 * Luckily we can at least blame others for their non-compliance but from
450 * inter-operability perspective it seems that we're somewhat stuck with
451 * the ordering which we have been using if we want to keep working with
452 * those broken things (not that it currently hurts anybody as there isn't
453 * particular reason why the ordering would need to be changed).
455 * At least SACK_PERM as the first option is known to lead to a disaster
456 * (but it may well be that other scenarios fail similarly).
458 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
459 struct tcp_out_options *opts)
461 u16 options = opts->options; /* mungable copy */
463 if (unlikely(OPTION_MD5 & options)) {
464 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
465 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
466 /* overload cookie hash location */
467 opts->hash_location = (__u8 *)ptr;
471 if (unlikely(opts->mss)) {
472 *ptr++ = htonl((TCPOPT_MSS << 24) |
473 (TCPOLEN_MSS << 16) |
477 if (likely(OPTION_TS & options)) {
478 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
479 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
480 (TCPOLEN_SACK_PERM << 16) |
481 (TCPOPT_TIMESTAMP << 8) |
483 options &= ~OPTION_SACK_ADVERTISE;
485 *ptr++ = htonl((TCPOPT_NOP << 24) |
487 (TCPOPT_TIMESTAMP << 8) |
490 *ptr++ = htonl(opts->tsval);
491 *ptr++ = htonl(opts->tsecr);
494 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
495 *ptr++ = htonl((TCPOPT_NOP << 24) |
497 (TCPOPT_SACK_PERM << 8) |
501 if (unlikely(OPTION_WSCALE & options)) {
502 *ptr++ = htonl((TCPOPT_NOP << 24) |
503 (TCPOPT_WINDOW << 16) |
504 (TCPOLEN_WINDOW << 8) |
508 if (unlikely(opts->num_sack_blocks)) {
509 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
510 tp->duplicate_sack : tp->selective_acks;
513 *ptr++ = htonl((TCPOPT_NOP << 24) |
516 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
517 TCPOLEN_SACK_PERBLOCK)));
519 for (this_sack = 0; this_sack < opts->num_sack_blocks;
521 *ptr++ = htonl(sp[this_sack].start_seq);
522 *ptr++ = htonl(sp[this_sack].end_seq);
525 tp->rx_opt.dsack = 0;
528 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
529 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
531 u32 len; /* Fast Open option length */
534 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
535 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
536 TCPOPT_FASTOPEN_MAGIC);
537 p += TCPOLEN_EXP_FASTOPEN_BASE;
539 len = TCPOLEN_FASTOPEN_BASE + foc->len;
540 *p++ = TCPOPT_FASTOPEN;
544 memcpy(p, foc->val, foc->len);
545 if ((len & 3) == 2) {
546 p[foc->len] = TCPOPT_NOP;
547 p[foc->len + 1] = TCPOPT_NOP;
549 ptr += (len + 3) >> 2;
552 smc_options_write(ptr, &options);
555 static void smc_set_option(const struct tcp_sock *tp,
556 struct tcp_out_options *opts,
557 unsigned int *remaining)
559 #if IS_ENABLED(CONFIG_SMC)
560 if (static_branch_unlikely(&tcp_have_smc)) {
562 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
563 opts->options |= OPTION_SMC;
564 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
571 static void smc_set_option_cond(const struct tcp_sock *tp,
572 const struct inet_request_sock *ireq,
573 struct tcp_out_options *opts,
574 unsigned int *remaining)
576 #if IS_ENABLED(CONFIG_SMC)
577 if (static_branch_unlikely(&tcp_have_smc)) {
578 if (tp->syn_smc && ireq->smc_ok) {
579 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
580 opts->options |= OPTION_SMC;
581 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
588 /* Compute TCP options for SYN packets. This is not the final
589 * network wire format yet.
591 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
592 struct tcp_out_options *opts,
593 struct tcp_md5sig_key **md5)
595 struct tcp_sock *tp = tcp_sk(sk);
596 unsigned int remaining = MAX_TCP_OPTION_SPACE;
597 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
600 #ifdef CONFIG_TCP_MD5SIG
601 if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
602 *md5 = tp->af_specific->md5_lookup(sk, sk);
604 opts->options |= OPTION_MD5;
605 remaining -= TCPOLEN_MD5SIG_ALIGNED;
610 /* We always get an MSS option. The option bytes which will be seen in
611 * normal data packets should timestamps be used, must be in the MSS
612 * advertised. But we subtract them from tp->mss_cache so that
613 * calculations in tcp_sendmsg are simpler etc. So account for this
614 * fact here if necessary. If we don't do this correctly, as a
615 * receiver we won't recognize data packets as being full sized when we
616 * should, and thus we won't abide by the delayed ACK rules correctly.
617 * SACKs don't matter, we never delay an ACK when we have any of those
619 opts->mss = tcp_advertise_mss(sk);
620 remaining -= TCPOLEN_MSS_ALIGNED;
622 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
623 opts->options |= OPTION_TS;
624 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
625 opts->tsecr = tp->rx_opt.ts_recent;
626 remaining -= TCPOLEN_TSTAMP_ALIGNED;
628 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
629 opts->ws = tp->rx_opt.rcv_wscale;
630 opts->options |= OPTION_WSCALE;
631 remaining -= TCPOLEN_WSCALE_ALIGNED;
633 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
634 opts->options |= OPTION_SACK_ADVERTISE;
635 if (unlikely(!(OPTION_TS & opts->options)))
636 remaining -= TCPOLEN_SACKPERM_ALIGNED;
639 if (fastopen && fastopen->cookie.len >= 0) {
640 u32 need = fastopen->cookie.len;
642 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
643 TCPOLEN_FASTOPEN_BASE;
644 need = (need + 3) & ~3U; /* Align to 32 bits */
645 if (remaining >= need) {
646 opts->options |= OPTION_FAST_OPEN_COOKIE;
647 opts->fastopen_cookie = &fastopen->cookie;
649 tp->syn_fastopen = 1;
650 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
654 smc_set_option(tp, opts, &remaining);
656 return MAX_TCP_OPTION_SPACE - remaining;
659 /* Set up TCP options for SYN-ACKs. */
660 static unsigned int tcp_synack_options(const struct sock *sk,
661 struct request_sock *req,
662 unsigned int mss, struct sk_buff *skb,
663 struct tcp_out_options *opts,
664 const struct tcp_md5sig_key *md5,
665 struct tcp_fastopen_cookie *foc)
667 struct inet_request_sock *ireq = inet_rsk(req);
668 unsigned int remaining = MAX_TCP_OPTION_SPACE;
670 #ifdef CONFIG_TCP_MD5SIG
672 opts->options |= OPTION_MD5;
673 remaining -= TCPOLEN_MD5SIG_ALIGNED;
675 /* We can't fit any SACK blocks in a packet with MD5 + TS
676 * options. There was discussion about disabling SACK
677 * rather than TS in order to fit in better with old,
678 * buggy kernels, but that was deemed to be unnecessary.
680 ireq->tstamp_ok &= !ireq->sack_ok;
684 /* We always send an MSS option. */
686 remaining -= TCPOLEN_MSS_ALIGNED;
688 if (likely(ireq->wscale_ok)) {
689 opts->ws = ireq->rcv_wscale;
690 opts->options |= OPTION_WSCALE;
691 remaining -= TCPOLEN_WSCALE_ALIGNED;
693 if (likely(ireq->tstamp_ok)) {
694 opts->options |= OPTION_TS;
695 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
696 opts->tsecr = req->ts_recent;
697 remaining -= TCPOLEN_TSTAMP_ALIGNED;
699 if (likely(ireq->sack_ok)) {
700 opts->options |= OPTION_SACK_ADVERTISE;
701 if (unlikely(!ireq->tstamp_ok))
702 remaining -= TCPOLEN_SACKPERM_ALIGNED;
704 if (foc != NULL && foc->len >= 0) {
707 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
708 TCPOLEN_FASTOPEN_BASE;
709 need = (need + 3) & ~3U; /* Align to 32 bits */
710 if (remaining >= need) {
711 opts->options |= OPTION_FAST_OPEN_COOKIE;
712 opts->fastopen_cookie = foc;
717 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
719 return MAX_TCP_OPTION_SPACE - remaining;
722 /* Compute TCP options for ESTABLISHED sockets. This is not the
723 * final wire format yet.
725 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
726 struct tcp_out_options *opts,
727 struct tcp_md5sig_key **md5)
729 struct tcp_sock *tp = tcp_sk(sk);
730 unsigned int size = 0;
731 unsigned int eff_sacks;
736 #ifdef CONFIG_TCP_MD5SIG
737 if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
738 *md5 = tp->af_specific->md5_lookup(sk, sk);
740 opts->options |= OPTION_MD5;
741 size += TCPOLEN_MD5SIG_ALIGNED;
746 if (likely(tp->rx_opt.tstamp_ok)) {
747 opts->options |= OPTION_TS;
748 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
749 opts->tsecr = tp->rx_opt.ts_recent;
750 size += TCPOLEN_TSTAMP_ALIGNED;
753 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
754 if (unlikely(eff_sacks)) {
755 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
756 opts->num_sack_blocks =
757 min_t(unsigned int, eff_sacks,
758 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
759 TCPOLEN_SACK_PERBLOCK);
760 size += TCPOLEN_SACK_BASE_ALIGNED +
761 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
768 /* TCP SMALL QUEUES (TSQ)
770 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
771 * to reduce RTT and bufferbloat.
772 * We do this using a special skb destructor (tcp_wfree).
774 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
775 * needs to be reallocated in a driver.
776 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
778 * Since transmit from skb destructor is forbidden, we use a tasklet
779 * to process all sockets that eventually need to send more skbs.
780 * We use one tasklet per cpu, with its own queue of sockets.
783 struct tasklet_struct tasklet;
784 struct list_head head; /* queue of tcp sockets */
786 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
788 static void tcp_tsq_write(struct sock *sk)
790 if ((1 << sk->sk_state) &
791 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
792 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
793 struct tcp_sock *tp = tcp_sk(sk);
795 if (tp->lost_out > tp->retrans_out &&
796 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
797 tcp_mstamp_refresh(tp);
798 tcp_xmit_retransmit_queue(sk);
801 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
806 static void tcp_tsq_handler(struct sock *sk)
809 if (!sock_owned_by_user(sk))
811 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
816 * One tasklet per cpu tries to send more skbs.
817 * We run in tasklet context but need to disable irqs when
818 * transferring tsq->head because tcp_wfree() might
819 * interrupt us (non NAPI drivers)
821 static void tcp_tasklet_func(unsigned long data)
823 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
826 struct list_head *q, *n;
830 local_irq_save(flags);
831 list_splice_init(&tsq->head, &list);
832 local_irq_restore(flags);
834 list_for_each_safe(q, n, &list) {
835 tp = list_entry(q, struct tcp_sock, tsq_node);
836 list_del(&tp->tsq_node);
838 sk = (struct sock *)tp;
839 smp_mb__before_atomic();
840 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
847 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
848 TCPF_WRITE_TIMER_DEFERRED | \
849 TCPF_DELACK_TIMER_DEFERRED | \
850 TCPF_MTU_REDUCED_DEFERRED)
852 * tcp_release_cb - tcp release_sock() callback
855 * called from release_sock() to perform protocol dependent
856 * actions before socket release.
858 void tcp_release_cb(struct sock *sk)
860 unsigned long flags, nflags;
862 /* perform an atomic operation only if at least one flag is set */
864 flags = sk->sk_tsq_flags;
865 if (!(flags & TCP_DEFERRED_ALL))
867 nflags = flags & ~TCP_DEFERRED_ALL;
868 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
870 if (flags & TCPF_TSQ_DEFERRED) {
874 /* Here begins the tricky part :
875 * We are called from release_sock() with :
877 * 2) sk_lock.slock spinlock held
878 * 3) socket owned by us (sk->sk_lock.owned == 1)
880 * But following code is meant to be called from BH handlers,
881 * so we should keep BH disabled, but early release socket ownership
883 sock_release_ownership(sk);
885 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
886 tcp_write_timer_handler(sk);
889 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
890 tcp_delack_timer_handler(sk);
893 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
894 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
898 EXPORT_SYMBOL(tcp_release_cb);
900 void __init tcp_tasklet_init(void)
904 for_each_possible_cpu(i) {
905 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
907 INIT_LIST_HEAD(&tsq->head);
908 tasklet_init(&tsq->tasklet,
915 * Write buffer destructor automatically called from kfree_skb.
916 * We can't xmit new skbs from this context, as we might already
919 void tcp_wfree(struct sk_buff *skb)
921 struct sock *sk = skb->sk;
922 struct tcp_sock *tp = tcp_sk(sk);
923 unsigned long flags, nval, oval;
925 /* Keep one reference on sk_wmem_alloc.
926 * Will be released by sk_free() from here or tcp_tasklet_func()
928 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
930 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
931 * Wait until our queues (qdisc + devices) are drained.
933 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
934 * - chance for incoming ACK (processed by another cpu maybe)
935 * to migrate this flow (skb->ooo_okay will be eventually set)
937 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
940 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
941 struct tsq_tasklet *tsq;
944 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
947 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
948 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
952 /* queue this socket to tasklet queue */
953 local_irq_save(flags);
954 tsq = this_cpu_ptr(&tsq_tasklet);
955 empty = list_empty(&tsq->head);
956 list_add(&tp->tsq_node, &tsq->head);
958 tasklet_schedule(&tsq->tasklet);
959 local_irq_restore(flags);
966 /* Note: Called under soft irq.
967 * We can call TCP stack right away, unless socket is owned by user.
969 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
971 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
972 struct sock *sk = (struct sock *)tp;
977 return HRTIMER_NORESTART;
980 /* BBR congestion control needs pacing.
981 * Same remark for SO_MAX_PACING_RATE.
982 * sch_fq packet scheduler is efficiently handling pacing,
983 * but is not always installed/used.
984 * Return true if TCP stack should pace packets itself.
986 static bool tcp_needs_internal_pacing(const struct sock *sk)
988 return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
991 static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
996 if (!tcp_needs_internal_pacing(sk))
998 rate = sk->sk_pacing_rate;
999 if (!rate || rate == ~0U)
1002 /* Should account for header sizes as sch_fq does,
1003 * but lets make things simple.
1005 len_ns = (u64)skb->len * NSEC_PER_SEC;
1006 do_div(len_ns, rate);
1007 hrtimer_start(&tcp_sk(sk)->pacing_timer,
1008 ktime_add_ns(ktime_get(), len_ns),
1009 HRTIMER_MODE_ABS_PINNED_SOFT);
1013 static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb)
1015 skb->skb_mstamp = tp->tcp_mstamp;
1016 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1019 /* This routine actually transmits TCP packets queued in by
1020 * tcp_do_sendmsg(). This is used by both the initial
1021 * transmission and possible later retransmissions.
1022 * All SKB's seen here are completely headerless. It is our
1023 * job to build the TCP header, and pass the packet down to
1024 * IP so it can do the same plus pass the packet off to the
1027 * We are working here with either a clone of the original
1028 * SKB, or a fresh unique copy made by the retransmit engine.
1030 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1031 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1033 const struct inet_connection_sock *icsk = inet_csk(sk);
1034 struct inet_sock *inet;
1035 struct tcp_sock *tp;
1036 struct tcp_skb_cb *tcb;
1037 struct tcp_out_options opts;
1038 unsigned int tcp_options_size, tcp_header_size;
1039 struct sk_buff *oskb = NULL;
1040 struct tcp_md5sig_key *md5;
1044 BUG_ON(!skb || !tcp_skb_pcount(skb));
1048 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1052 tcp_skb_tsorted_save(oskb) {
1053 if (unlikely(skb_cloned(oskb)))
1054 skb = pskb_copy(oskb, gfp_mask);
1056 skb = skb_clone(oskb, gfp_mask);
1057 } tcp_skb_tsorted_restore(oskb);
1062 skb->skb_mstamp = tp->tcp_mstamp;
1065 tcb = TCP_SKB_CB(skb);
1066 memset(&opts, 0, sizeof(opts));
1068 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1069 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1071 tcp_options_size = tcp_established_options(sk, skb, &opts,
1073 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1075 /* if no packet is in qdisc/device queue, then allow XPS to select
1076 * another queue. We can be called from tcp_tsq_handler()
1077 * which holds one reference to sk.
1079 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1080 * One way to get this would be to set skb->truesize = 2 on them.
1082 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1084 /* If we had to use memory reserve to allocate this skb,
1085 * this might cause drops if packet is looped back :
1086 * Other socket might not have SOCK_MEMALLOC.
1087 * Packets not looped back do not care about pfmemalloc.
1089 skb->pfmemalloc = 0;
1091 skb_push(skb, tcp_header_size);
1092 skb_reset_transport_header(skb);
1096 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1097 skb_set_hash_from_sk(skb, sk);
1098 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1100 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1102 /* Build TCP header and checksum it. */
1103 th = (struct tcphdr *)skb->data;
1104 th->source = inet->inet_sport;
1105 th->dest = inet->inet_dport;
1106 th->seq = htonl(tcb->seq);
1107 th->ack_seq = htonl(rcv_nxt);
1108 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1114 /* The urg_mode check is necessary during a below snd_una win probe */
1115 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1116 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1117 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1119 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1120 th->urg_ptr = htons(0xFFFF);
1125 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1126 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1127 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1128 th->window = htons(tcp_select_window(sk));
1129 tcp_ecn_send(sk, skb, th, tcp_header_size);
1131 /* RFC1323: The window in SYN & SYN/ACK segments
1134 th->window = htons(min(tp->rcv_wnd, 65535U));
1136 #ifdef CONFIG_TCP_MD5SIG
1137 /* Calculate the MD5 hash, as we have all we need now */
1139 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1140 tp->af_specific->calc_md5_hash(opts.hash_location,
1145 icsk->icsk_af_ops->send_check(sk, skb);
1147 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1148 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1150 if (skb->len != tcp_header_size) {
1151 tcp_event_data_sent(tp, sk);
1152 tp->data_segs_out += tcp_skb_pcount(skb);
1153 tcp_internal_pacing(sk, skb);
1156 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1157 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1158 tcp_skb_pcount(skb));
1160 tp->segs_out += tcp_skb_pcount(skb);
1161 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1162 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1163 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1165 /* Our usage of tstamp should remain private */
1168 /* Cleanup our debris for IP stacks */
1169 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1170 sizeof(struct inet6_skb_parm)));
1172 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1174 if (unlikely(err > 0)) {
1176 err = net_xmit_eval(err);
1179 tcp_update_skb_after_send(tp, oskb);
1180 tcp_rate_skb_sent(sk, oskb);
1185 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1188 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1189 tcp_sk(sk)->rcv_nxt);
1192 /* This routine just queues the buffer for sending.
1194 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1195 * otherwise socket can stall.
1197 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1199 struct tcp_sock *tp = tcp_sk(sk);
1201 /* Advance write_seq and place onto the write_queue. */
1202 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1203 __skb_header_release(skb);
1204 tcp_add_write_queue_tail(sk, skb);
1205 sk->sk_wmem_queued += skb->truesize;
1206 sk_mem_charge(sk, skb->truesize);
1209 /* Initialize TSO segments for a packet. */
1210 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1212 if (skb->len <= mss_now) {
1213 /* Avoid the costly divide in the normal
1216 tcp_skb_pcount_set(skb, 1);
1217 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1219 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1220 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1224 /* Pcount in the middle of the write queue got changed, we need to do various
1225 * tweaks to fix counters
1227 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1229 struct tcp_sock *tp = tcp_sk(sk);
1231 tp->packets_out -= decr;
1233 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1234 tp->sacked_out -= decr;
1235 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1236 tp->retrans_out -= decr;
1237 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1238 tp->lost_out -= decr;
1240 /* Reno case is special. Sigh... */
1241 if (tcp_is_reno(tp) && decr > 0)
1242 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1244 if (tp->lost_skb_hint &&
1245 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1246 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1247 tp->lost_cnt_hint -= decr;
1249 tcp_verify_left_out(tp);
1252 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1254 return TCP_SKB_CB(skb)->txstamp_ack ||
1255 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1258 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1260 struct skb_shared_info *shinfo = skb_shinfo(skb);
1262 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1263 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1264 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1265 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1267 shinfo->tx_flags &= ~tsflags;
1268 shinfo2->tx_flags |= tsflags;
1269 swap(shinfo->tskey, shinfo2->tskey);
1270 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1271 TCP_SKB_CB(skb)->txstamp_ack = 0;
1275 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1277 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1278 TCP_SKB_CB(skb)->eor = 0;
1281 /* Insert buff after skb on the write or rtx queue of sk. */
1282 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1283 struct sk_buff *buff,
1285 enum tcp_queue tcp_queue)
1287 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1288 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1290 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1293 /* Function to create two new TCP segments. Shrinks the given segment
1294 * to the specified size and appends a new segment with the rest of the
1295 * packet to the list. This won't be called frequently, I hope.
1296 * Remember, these are still headerless SKBs at this point.
1298 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1299 struct sk_buff *skb, u32 len,
1300 unsigned int mss_now, gfp_t gfp)
1302 struct tcp_sock *tp = tcp_sk(sk);
1303 struct sk_buff *buff;
1304 int nsize, old_factor;
1308 if (WARN_ON(len > skb->len))
1311 nsize = skb_headlen(skb) - len;
1315 if (skb_unclone(skb, gfp))
1318 /* Get a new skb... force flag on. */
1319 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1321 return -ENOMEM; /* We'll just try again later. */
1323 sk->sk_wmem_queued += buff->truesize;
1324 sk_mem_charge(sk, buff->truesize);
1325 nlen = skb->len - len - nsize;
1326 buff->truesize += nlen;
1327 skb->truesize -= nlen;
1329 /* Correct the sequence numbers. */
1330 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1331 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1332 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1334 /* PSH and FIN should only be set in the second packet. */
1335 flags = TCP_SKB_CB(skb)->tcp_flags;
1336 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1337 TCP_SKB_CB(buff)->tcp_flags = flags;
1338 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1339 tcp_skb_fragment_eor(skb, buff);
1341 skb_split(skb, buff, len);
1343 buff->ip_summed = CHECKSUM_PARTIAL;
1345 buff->tstamp = skb->tstamp;
1346 tcp_fragment_tstamp(skb, buff);
1348 old_factor = tcp_skb_pcount(skb);
1350 /* Fix up tso_factor for both original and new SKB. */
1351 tcp_set_skb_tso_segs(skb, mss_now);
1352 tcp_set_skb_tso_segs(buff, mss_now);
1354 /* Update delivered info for the new segment */
1355 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1357 /* If this packet has been sent out already, we must
1358 * adjust the various packet counters.
1360 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1361 int diff = old_factor - tcp_skb_pcount(skb) -
1362 tcp_skb_pcount(buff);
1365 tcp_adjust_pcount(sk, skb, diff);
1368 /* Link BUFF into the send queue. */
1369 __skb_header_release(buff);
1370 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1371 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1372 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1377 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1378 * data is not copied, but immediately discarded.
1380 static int __pskb_trim_head(struct sk_buff *skb, int len)
1382 struct skb_shared_info *shinfo;
1385 eat = min_t(int, len, skb_headlen(skb));
1387 __skb_pull(skb, eat);
1394 shinfo = skb_shinfo(skb);
1395 for (i = 0; i < shinfo->nr_frags; i++) {
1396 int size = skb_frag_size(&shinfo->frags[i]);
1399 skb_frag_unref(skb, i);
1402 shinfo->frags[k] = shinfo->frags[i];
1404 shinfo->frags[k].page_offset += eat;
1405 skb_frag_size_sub(&shinfo->frags[k], eat);
1411 shinfo->nr_frags = k;
1413 skb->data_len -= len;
1414 skb->len = skb->data_len;
1418 /* Remove acked data from a packet in the transmit queue. */
1419 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1423 if (skb_unclone(skb, GFP_ATOMIC))
1426 delta_truesize = __pskb_trim_head(skb, len);
1428 TCP_SKB_CB(skb)->seq += len;
1429 skb->ip_summed = CHECKSUM_PARTIAL;
1431 if (delta_truesize) {
1432 skb->truesize -= delta_truesize;
1433 sk->sk_wmem_queued -= delta_truesize;
1434 sk_mem_uncharge(sk, delta_truesize);
1435 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1438 /* Any change of skb->len requires recalculation of tso factor. */
1439 if (tcp_skb_pcount(skb) > 1)
1440 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1445 /* Calculate MSS not accounting any TCP options. */
1446 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1448 const struct tcp_sock *tp = tcp_sk(sk);
1449 const struct inet_connection_sock *icsk = inet_csk(sk);
1452 /* Calculate base mss without TCP options:
1453 It is MMS_S - sizeof(tcphdr) of rfc1122
1455 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1457 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1458 if (icsk->icsk_af_ops->net_frag_header_len) {
1459 const struct dst_entry *dst = __sk_dst_get(sk);
1461 if (dst && dst_allfrag(dst))
1462 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1465 /* Clamp it (mss_clamp does not include tcp options) */
1466 if (mss_now > tp->rx_opt.mss_clamp)
1467 mss_now = tp->rx_opt.mss_clamp;
1469 /* Now subtract optional transport overhead */
1470 mss_now -= icsk->icsk_ext_hdr_len;
1472 /* Then reserve room for full set of TCP options and 8 bytes of data */
1478 /* Calculate MSS. Not accounting for SACKs here. */
1479 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1481 /* Subtract TCP options size, not including SACKs */
1482 return __tcp_mtu_to_mss(sk, pmtu) -
1483 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1486 /* Inverse of above */
1487 int tcp_mss_to_mtu(struct sock *sk, int mss)
1489 const struct tcp_sock *tp = tcp_sk(sk);
1490 const struct inet_connection_sock *icsk = inet_csk(sk);
1494 tp->tcp_header_len +
1495 icsk->icsk_ext_hdr_len +
1496 icsk->icsk_af_ops->net_header_len;
1498 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1499 if (icsk->icsk_af_ops->net_frag_header_len) {
1500 const struct dst_entry *dst = __sk_dst_get(sk);
1502 if (dst && dst_allfrag(dst))
1503 mtu += icsk->icsk_af_ops->net_frag_header_len;
1507 EXPORT_SYMBOL(tcp_mss_to_mtu);
1509 /* MTU probing init per socket */
1510 void tcp_mtup_init(struct sock *sk)
1512 struct tcp_sock *tp = tcp_sk(sk);
1513 struct inet_connection_sock *icsk = inet_csk(sk);
1514 struct net *net = sock_net(sk);
1516 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1517 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1518 icsk->icsk_af_ops->net_header_len;
1519 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1520 icsk->icsk_mtup.probe_size = 0;
1521 if (icsk->icsk_mtup.enabled)
1522 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1524 EXPORT_SYMBOL(tcp_mtup_init);
1526 /* This function synchronize snd mss to current pmtu/exthdr set.
1528 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1529 for TCP options, but includes only bare TCP header.
1531 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1532 It is minimum of user_mss and mss received with SYN.
1533 It also does not include TCP options.
1535 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1537 tp->mss_cache is current effective sending mss, including
1538 all tcp options except for SACKs. It is evaluated,
1539 taking into account current pmtu, but never exceeds
1540 tp->rx_opt.mss_clamp.
1542 NOTE1. rfc1122 clearly states that advertised MSS
1543 DOES NOT include either tcp or ip options.
1545 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1546 are READ ONLY outside this function. --ANK (980731)
1548 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1550 struct tcp_sock *tp = tcp_sk(sk);
1551 struct inet_connection_sock *icsk = inet_csk(sk);
1554 if (icsk->icsk_mtup.search_high > pmtu)
1555 icsk->icsk_mtup.search_high = pmtu;
1557 mss_now = tcp_mtu_to_mss(sk, pmtu);
1558 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1560 /* And store cached results */
1561 icsk->icsk_pmtu_cookie = pmtu;
1562 if (icsk->icsk_mtup.enabled)
1563 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1564 tp->mss_cache = mss_now;
1568 EXPORT_SYMBOL(tcp_sync_mss);
1570 /* Compute the current effective MSS, taking SACKs and IP options,
1571 * and even PMTU discovery events into account.
1573 unsigned int tcp_current_mss(struct sock *sk)
1575 const struct tcp_sock *tp = tcp_sk(sk);
1576 const struct dst_entry *dst = __sk_dst_get(sk);
1578 unsigned int header_len;
1579 struct tcp_out_options opts;
1580 struct tcp_md5sig_key *md5;
1582 mss_now = tp->mss_cache;
1585 u32 mtu = dst_mtu(dst);
1586 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1587 mss_now = tcp_sync_mss(sk, mtu);
1590 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1591 sizeof(struct tcphdr);
1592 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1593 * some common options. If this is an odd packet (because we have SACK
1594 * blocks etc) then our calculated header_len will be different, and
1595 * we have to adjust mss_now correspondingly */
1596 if (header_len != tp->tcp_header_len) {
1597 int delta = (int) header_len - tp->tcp_header_len;
1604 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1605 * As additional protections, we do not touch cwnd in retransmission phases,
1606 * and if application hit its sndbuf limit recently.
1608 static void tcp_cwnd_application_limited(struct sock *sk)
1610 struct tcp_sock *tp = tcp_sk(sk);
1612 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1613 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1614 /* Limited by application or receiver window. */
1615 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1616 u32 win_used = max(tp->snd_cwnd_used, init_win);
1617 if (win_used < tp->snd_cwnd) {
1618 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1619 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1621 tp->snd_cwnd_used = 0;
1623 tp->snd_cwnd_stamp = tcp_jiffies32;
1626 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1628 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1629 struct tcp_sock *tp = tcp_sk(sk);
1631 /* Track the maximum number of outstanding packets in each
1632 * window, and remember whether we were cwnd-limited then.
1634 if (!before(tp->snd_una, tp->max_packets_seq) ||
1635 tp->packets_out > tp->max_packets_out) {
1636 tp->max_packets_out = tp->packets_out;
1637 tp->max_packets_seq = tp->snd_nxt;
1638 tp->is_cwnd_limited = is_cwnd_limited;
1641 if (tcp_is_cwnd_limited(sk)) {
1642 /* Network is feed fully. */
1643 tp->snd_cwnd_used = 0;
1644 tp->snd_cwnd_stamp = tcp_jiffies32;
1646 /* Network starves. */
1647 if (tp->packets_out > tp->snd_cwnd_used)
1648 tp->snd_cwnd_used = tp->packets_out;
1650 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1651 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1652 !ca_ops->cong_control)
1653 tcp_cwnd_application_limited(sk);
1655 /* The following conditions together indicate the starvation
1656 * is caused by insufficient sender buffer:
1657 * 1) just sent some data (see tcp_write_xmit)
1658 * 2) not cwnd limited (this else condition)
1659 * 3) no more data to send (tcp_write_queue_empty())
1660 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1662 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1663 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1664 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1665 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1669 /* Minshall's variant of the Nagle send check. */
1670 static bool tcp_minshall_check(const struct tcp_sock *tp)
1672 return after(tp->snd_sml, tp->snd_una) &&
1673 !after(tp->snd_sml, tp->snd_nxt);
1676 /* Update snd_sml if this skb is under mss
1677 * Note that a TSO packet might end with a sub-mss segment
1678 * The test is really :
1679 * if ((skb->len % mss) != 0)
1680 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1681 * But we can avoid doing the divide again given we already have
1682 * skb_pcount = skb->len / mss_now
1684 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1685 const struct sk_buff *skb)
1687 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1688 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1691 /* Return false, if packet can be sent now without violation Nagle's rules:
1692 * 1. It is full sized. (provided by caller in %partial bool)
1693 * 2. Or it contains FIN. (already checked by caller)
1694 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1695 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1696 * With Minshall's modification: all sent small packets are ACKed.
1698 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1702 ((nonagle & TCP_NAGLE_CORK) ||
1703 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1706 /* Return how many segs we'd like on a TSO packet,
1707 * to send one TSO packet per ms
1709 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1714 bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift,
1715 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1717 /* Goal is to send at least one packet per ms,
1718 * not one big TSO packet every 100 ms.
1719 * This preserves ACK clocking and is consistent
1720 * with tcp_tso_should_defer() heuristic.
1722 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1727 /* Return the number of segments we want in the skb we are transmitting.
1728 * See if congestion control module wants to decide; otherwise, autosize.
1730 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1732 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1733 u32 min_tso, tso_segs;
1735 min_tso = ca_ops->min_tso_segs ?
1736 ca_ops->min_tso_segs(sk) :
1737 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1739 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1740 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1743 /* Returns the portion of skb which can be sent right away */
1744 static unsigned int tcp_mss_split_point(const struct sock *sk,
1745 const struct sk_buff *skb,
1746 unsigned int mss_now,
1747 unsigned int max_segs,
1750 const struct tcp_sock *tp = tcp_sk(sk);
1751 u32 partial, needed, window, max_len;
1753 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1754 max_len = mss_now * max_segs;
1756 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1759 needed = min(skb->len, window);
1761 if (max_len <= needed)
1764 partial = needed % mss_now;
1765 /* If last segment is not a full MSS, check if Nagle rules allow us
1766 * to include this last segment in this skb.
1767 * Otherwise, we'll split the skb at last MSS boundary
1769 if (tcp_nagle_check(partial != 0, tp, nonagle))
1770 return needed - partial;
1775 /* Can at least one segment of SKB be sent right now, according to the
1776 * congestion window rules? If so, return how many segments are allowed.
1778 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1779 const struct sk_buff *skb)
1781 u32 in_flight, cwnd, halfcwnd;
1783 /* Don't be strict about the congestion window for the final FIN. */
1784 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1785 tcp_skb_pcount(skb) == 1)
1788 in_flight = tcp_packets_in_flight(tp);
1789 cwnd = tp->snd_cwnd;
1790 if (in_flight >= cwnd)
1793 /* For better scheduling, ensure we have at least
1794 * 2 GSO packets in flight.
1796 halfcwnd = max(cwnd >> 1, 1U);
1797 return min(halfcwnd, cwnd - in_flight);
1800 /* Initialize TSO state of a skb.
1801 * This must be invoked the first time we consider transmitting
1802 * SKB onto the wire.
1804 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1806 int tso_segs = tcp_skb_pcount(skb);
1808 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1809 tcp_set_skb_tso_segs(skb, mss_now);
1810 tso_segs = tcp_skb_pcount(skb);
1816 /* Return true if the Nagle test allows this packet to be
1819 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1820 unsigned int cur_mss, int nonagle)
1822 /* Nagle rule does not apply to frames, which sit in the middle of the
1823 * write_queue (they have no chances to get new data).
1825 * This is implemented in the callers, where they modify the 'nonagle'
1826 * argument based upon the location of SKB in the send queue.
1828 if (nonagle & TCP_NAGLE_PUSH)
1831 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1832 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1835 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1841 /* Does at least the first segment of SKB fit into the send window? */
1842 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1843 const struct sk_buff *skb,
1844 unsigned int cur_mss)
1846 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1848 if (skb->len > cur_mss)
1849 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1851 return !after(end_seq, tcp_wnd_end(tp));
1854 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1855 * which is put after SKB on the list. It is very much like
1856 * tcp_fragment() except that it may make several kinds of assumptions
1857 * in order to speed up the splitting operation. In particular, we
1858 * know that all the data is in scatter-gather pages, and that the
1859 * packet has never been sent out before (and thus is not cloned).
1861 static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1862 struct sk_buff *skb, unsigned int len,
1863 unsigned int mss_now, gfp_t gfp)
1865 struct sk_buff *buff;
1866 int nlen = skb->len - len;
1869 /* All of a TSO frame must be composed of paged data. */
1870 if (skb->len != skb->data_len)
1871 return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp);
1873 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1874 if (unlikely(!buff))
1877 sk->sk_wmem_queued += buff->truesize;
1878 sk_mem_charge(sk, buff->truesize);
1879 buff->truesize += nlen;
1880 skb->truesize -= nlen;
1882 /* Correct the sequence numbers. */
1883 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1884 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1885 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1887 /* PSH and FIN should only be set in the second packet. */
1888 flags = TCP_SKB_CB(skb)->tcp_flags;
1889 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1890 TCP_SKB_CB(buff)->tcp_flags = flags;
1892 /* This packet was never sent out yet, so no SACK bits. */
1893 TCP_SKB_CB(buff)->sacked = 0;
1895 tcp_skb_fragment_eor(skb, buff);
1897 buff->ip_summed = CHECKSUM_PARTIAL;
1898 skb_split(skb, buff, len);
1899 tcp_fragment_tstamp(skb, buff);
1901 /* Fix up tso_factor for both original and new SKB. */
1902 tcp_set_skb_tso_segs(skb, mss_now);
1903 tcp_set_skb_tso_segs(buff, mss_now);
1905 /* Link BUFF into the send queue. */
1906 __skb_header_release(buff);
1907 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1912 /* Try to defer sending, if possible, in order to minimize the amount
1913 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1915 * This algorithm is from John Heffner.
1917 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1918 bool *is_cwnd_limited, u32 max_segs)
1920 const struct inet_connection_sock *icsk = inet_csk(sk);
1921 u32 age, send_win, cong_win, limit, in_flight;
1922 struct tcp_sock *tp = tcp_sk(sk);
1923 struct sk_buff *head;
1926 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1929 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1932 /* Avoid bursty behavior by allowing defer
1933 * only if the last write was recent.
1935 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
1938 in_flight = tcp_packets_in_flight(tp);
1940 BUG_ON(tcp_skb_pcount(skb) <= 1);
1941 BUG_ON(tp->snd_cwnd <= in_flight);
1943 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1945 /* From in_flight test above, we know that cwnd > in_flight. */
1946 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1948 limit = min(send_win, cong_win);
1950 /* If a full-sized TSO skb can be sent, do it. */
1951 if (limit >= max_segs * tp->mss_cache)
1954 /* Middle in queue won't get any more data, full sendable already? */
1955 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1958 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1960 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1962 /* If at least some fraction of a window is available,
1965 chunk /= win_divisor;
1969 /* Different approach, try not to defer past a single
1970 * ACK. Receiver should ACK every other full sized
1971 * frame, so if we have space for more than 3 frames
1974 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1978 /* TODO : use tsorted_sent_queue ? */
1979 head = tcp_rtx_queue_head(sk);
1982 age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
1983 /* If next ACK is likely to come too late (half srtt), do not defer */
1984 if (age < (tp->srtt_us >> 4))
1987 /* Ok, it looks like it is advisable to defer. */
1989 if (cong_win < send_win && cong_win <= skb->len)
1990 *is_cwnd_limited = true;
1998 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2000 struct inet_connection_sock *icsk = inet_csk(sk);
2001 struct tcp_sock *tp = tcp_sk(sk);
2002 struct net *net = sock_net(sk);
2006 interval = net->ipv4.sysctl_tcp_probe_interval;
2007 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2008 if (unlikely(delta >= interval * HZ)) {
2009 int mss = tcp_current_mss(sk);
2011 /* Update current search range */
2012 icsk->icsk_mtup.probe_size = 0;
2013 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2014 sizeof(struct tcphdr) +
2015 icsk->icsk_af_ops->net_header_len;
2016 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2018 /* Update probe time stamp */
2019 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2023 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2025 struct sk_buff *skb, *next;
2027 skb = tcp_send_head(sk);
2028 tcp_for_write_queue_from_safe(skb, next, sk) {
2029 if (len <= skb->len)
2032 if (unlikely(TCP_SKB_CB(skb)->eor))
2041 /* Create a new MTU probe if we are ready.
2042 * MTU probe is regularly attempting to increase the path MTU by
2043 * deliberately sending larger packets. This discovers routing
2044 * changes resulting in larger path MTUs.
2046 * Returns 0 if we should wait to probe (no cwnd available),
2047 * 1 if a probe was sent,
2050 static int tcp_mtu_probe(struct sock *sk)
2052 struct inet_connection_sock *icsk = inet_csk(sk);
2053 struct tcp_sock *tp = tcp_sk(sk);
2054 struct sk_buff *skb, *nskb, *next;
2055 struct net *net = sock_net(sk);
2062 /* Not currently probing/verifying,
2064 * have enough cwnd, and
2065 * not SACKing (the variable headers throw things off)
2067 if (likely(!icsk->icsk_mtup.enabled ||
2068 icsk->icsk_mtup.probe_size ||
2069 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2070 tp->snd_cwnd < 11 ||
2071 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2074 /* Use binary search for probe_size between tcp_mss_base,
2075 * and current mss_clamp. if (search_high - search_low)
2076 * smaller than a threshold, backoff from probing.
2078 mss_now = tcp_current_mss(sk);
2079 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2080 icsk->icsk_mtup.search_low) >> 1);
2081 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2082 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2083 /* When misfortune happens, we are reprobing actively,
2084 * and then reprobe timer has expired. We stick with current
2085 * probing process by not resetting search range to its orignal.
2087 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2088 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2089 /* Check whether enough time has elaplased for
2090 * another round of probing.
2092 tcp_mtu_check_reprobe(sk);
2096 /* Have enough data in the send queue to probe? */
2097 if (tp->write_seq - tp->snd_nxt < size_needed)
2100 if (tp->snd_wnd < size_needed)
2102 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2105 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2106 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2107 if (!tcp_packets_in_flight(tp))
2113 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2116 /* We're allowed to probe. Build it now. */
2117 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2120 sk->sk_wmem_queued += nskb->truesize;
2121 sk_mem_charge(sk, nskb->truesize);
2123 skb = tcp_send_head(sk);
2125 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2126 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2127 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2128 TCP_SKB_CB(nskb)->sacked = 0;
2130 nskb->ip_summed = CHECKSUM_PARTIAL;
2132 tcp_insert_write_queue_before(nskb, skb, sk);
2133 tcp_highest_sack_replace(sk, skb, nskb);
2136 tcp_for_write_queue_from_safe(skb, next, sk) {
2137 copy = min_t(int, skb->len, probe_size - len);
2138 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2140 if (skb->len <= copy) {
2141 /* We've eaten all the data from this skb.
2143 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2144 /* If this is the last SKB we copy and eor is set
2145 * we need to propagate it to the new skb.
2147 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2148 tcp_unlink_write_queue(skb, sk);
2149 sk_wmem_free_skb(sk, skb);
2151 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2152 ~(TCPHDR_FIN|TCPHDR_PSH);
2153 if (!skb_shinfo(skb)->nr_frags) {
2154 skb_pull(skb, copy);
2156 __pskb_trim_head(skb, copy);
2157 tcp_set_skb_tso_segs(skb, mss_now);
2159 TCP_SKB_CB(skb)->seq += copy;
2164 if (len >= probe_size)
2167 tcp_init_tso_segs(nskb, nskb->len);
2169 /* We're ready to send. If this fails, the probe will
2170 * be resegmented into mss-sized pieces by tcp_write_xmit().
2172 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2173 /* Decrement cwnd here because we are sending
2174 * effectively two packets. */
2176 tcp_event_new_data_sent(sk, nskb);
2178 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2179 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2180 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2188 static bool tcp_pacing_check(const struct sock *sk)
2190 return tcp_needs_internal_pacing(sk) &&
2191 hrtimer_is_queued(&tcp_sk(sk)->pacing_timer);
2194 /* TCP Small Queues :
2195 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2196 * (These limits are doubled for retransmits)
2198 * - better RTT estimation and ACK scheduling
2201 * Alas, some drivers / subsystems require a fair amount
2202 * of queued bytes to ensure line rate.
2203 * One example is wifi aggregation (802.11 AMPDU)
2205 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2206 unsigned int factor)
2210 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift);
2211 limit = min_t(u32, limit,
2212 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2215 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2216 /* Always send skb if rtx queue is empty.
2217 * No need to wait for TX completion to call us back,
2218 * after softirq/tasklet schedule.
2219 * This helps when TX completions are delayed too much.
2221 if (tcp_rtx_queue_empty(sk))
2224 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2225 /* It is possible TX completion already happened
2226 * before we set TSQ_THROTTLED, so we must
2227 * test again the condition.
2229 smp_mb__after_atomic();
2230 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2236 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2238 const u32 now = tcp_jiffies32;
2239 enum tcp_chrono old = tp->chrono_type;
2241 if (old > TCP_CHRONO_UNSPEC)
2242 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2243 tp->chrono_start = now;
2244 tp->chrono_type = new;
2247 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2249 struct tcp_sock *tp = tcp_sk(sk);
2251 /* If there are multiple conditions worthy of tracking in a
2252 * chronograph then the highest priority enum takes precedence
2253 * over the other conditions. So that if something "more interesting"
2254 * starts happening, stop the previous chrono and start a new one.
2256 if (type > tp->chrono_type)
2257 tcp_chrono_set(tp, type);
2260 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2262 struct tcp_sock *tp = tcp_sk(sk);
2265 /* There are multiple conditions worthy of tracking in a
2266 * chronograph, so that the highest priority enum takes
2267 * precedence over the other conditions (see tcp_chrono_start).
2268 * If a condition stops, we only stop chrono tracking if
2269 * it's the "most interesting" or current chrono we are
2270 * tracking and starts busy chrono if we have pending data.
2272 if (tcp_rtx_and_write_queues_empty(sk))
2273 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2274 else if (type == tp->chrono_type)
2275 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2278 /* This routine writes packets to the network. It advances the
2279 * send_head. This happens as incoming acks open up the remote
2282 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2283 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2284 * account rare use of URG, this is not a big flaw.
2286 * Send at most one packet when push_one > 0. Temporarily ignore
2287 * cwnd limit to force at most one packet out when push_one == 2.
2289 * Returns true, if no segments are in flight and we have queued segments,
2290 * but cannot send anything now because of SWS or another problem.
2292 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2293 int push_one, gfp_t gfp)
2295 struct tcp_sock *tp = tcp_sk(sk);
2296 struct sk_buff *skb;
2297 unsigned int tso_segs, sent_pkts;
2300 bool is_cwnd_limited = false, is_rwnd_limited = false;
2305 tcp_mstamp_refresh(tp);
2307 /* Do MTU probing. */
2308 result = tcp_mtu_probe(sk);
2311 } else if (result > 0) {
2316 max_segs = tcp_tso_segs(sk, mss_now);
2317 while ((skb = tcp_send_head(sk))) {
2320 if (tcp_pacing_check(sk))
2323 tso_segs = tcp_init_tso_segs(skb, mss_now);
2326 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2327 /* "skb_mstamp" is used as a start point for the retransmit timer */
2328 tcp_update_skb_after_send(tp, skb);
2329 goto repair; /* Skip network transmission */
2332 cwnd_quota = tcp_cwnd_test(tp, skb);
2335 /* Force out a loss probe pkt. */
2341 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2342 is_rwnd_limited = true;
2346 if (tso_segs == 1) {
2347 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2348 (tcp_skb_is_last(sk, skb) ?
2349 nonagle : TCP_NAGLE_PUSH))))
2353 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2359 if (tso_segs > 1 && !tcp_urg_mode(tp))
2360 limit = tcp_mss_split_point(sk, skb, mss_now,
2366 if (skb->len > limit &&
2367 unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2368 skb, limit, mss_now, gfp)))
2371 if (tcp_small_queue_check(sk, skb, 0))
2374 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2378 /* Advance the send_head. This one is sent out.
2379 * This call will increment packets_out.
2381 tcp_event_new_data_sent(sk, skb);
2383 tcp_minshall_update(tp, mss_now, skb);
2384 sent_pkts += tcp_skb_pcount(skb);
2390 if (is_rwnd_limited)
2391 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2393 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2395 if (likely(sent_pkts)) {
2396 if (tcp_in_cwnd_reduction(sk))
2397 tp->prr_out += sent_pkts;
2399 /* Send one loss probe per tail loss episode. */
2401 tcp_schedule_loss_probe(sk, false);
2402 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2403 tcp_cwnd_validate(sk, is_cwnd_limited);
2406 return !tp->packets_out && !tcp_write_queue_empty(sk);
2409 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2411 struct inet_connection_sock *icsk = inet_csk(sk);
2412 struct tcp_sock *tp = tcp_sk(sk);
2413 u32 timeout, rto_delta_us;
2416 /* Don't do any loss probe on a Fast Open connection before 3WHS
2419 if (tp->fastopen_rsk)
2422 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2423 /* Schedule a loss probe in 2*RTT for SACK capable connections
2424 * not in loss recovery, that are either limited by cwnd or application.
2426 if ((early_retrans != 3 && early_retrans != 4) ||
2427 !tp->packets_out || !tcp_is_sack(tp) ||
2428 (icsk->icsk_ca_state != TCP_CA_Open &&
2429 icsk->icsk_ca_state != TCP_CA_CWR))
2432 /* Probe timeout is 2*rtt. Add minimum RTO to account
2433 * for delayed ack when there's one outstanding packet. If no RTT
2434 * sample is available then probe after TCP_TIMEOUT_INIT.
2437 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2438 if (tp->packets_out == 1)
2439 timeout += TCP_RTO_MIN;
2441 timeout += TCP_TIMEOUT_MIN;
2443 timeout = TCP_TIMEOUT_INIT;
2446 /* If the RTO formula yields an earlier time, then use that time. */
2447 rto_delta_us = advancing_rto ?
2448 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2449 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2450 if (rto_delta_us > 0)
2451 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2453 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2458 /* Thanks to skb fast clones, we can detect if a prior transmit of
2459 * a packet is still in a qdisc or driver queue.
2460 * In this case, there is very little point doing a retransmit !
2462 static bool skb_still_in_host_queue(const struct sock *sk,
2463 const struct sk_buff *skb)
2465 if (unlikely(skb_fclone_busy(sk, skb))) {
2466 NET_INC_STATS(sock_net(sk),
2467 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2473 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2474 * retransmit the last segment.
2476 void tcp_send_loss_probe(struct sock *sk)
2478 struct tcp_sock *tp = tcp_sk(sk);
2479 struct sk_buff *skb;
2481 int mss = tcp_current_mss(sk);
2483 skb = tcp_send_head(sk);
2484 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2485 pcount = tp->packets_out;
2486 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2487 if (tp->packets_out > pcount)
2491 skb = skb_rb_last(&sk->tcp_rtx_queue);
2493 /* At most one outstanding TLP retransmission. */
2494 if (tp->tlp_high_seq)
2497 /* Retransmit last segment. */
2501 if (skb_still_in_host_queue(sk, skb))
2504 pcount = tcp_skb_pcount(skb);
2505 if (WARN_ON(!pcount))
2508 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2509 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2510 (pcount - 1) * mss, mss,
2513 skb = skb_rb_next(skb);
2516 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2519 if (__tcp_retransmit_skb(sk, skb, 1))
2522 /* Record snd_nxt for loss detection. */
2523 tp->tlp_high_seq = tp->snd_nxt;
2526 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2527 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2528 inet_csk(sk)->icsk_pending = 0;
2533 /* Push out any pending frames which were held back due to
2534 * TCP_CORK or attempt at coalescing tiny packets.
2535 * The socket must be locked by the caller.
2537 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2540 /* If we are closed, the bytes will have to remain here.
2541 * In time closedown will finish, we empty the write queue and
2542 * all will be happy.
2544 if (unlikely(sk->sk_state == TCP_CLOSE))
2547 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2548 sk_gfp_mask(sk, GFP_ATOMIC)))
2549 tcp_check_probe_timer(sk);
2552 /* Send _single_ skb sitting at the send head. This function requires
2553 * true push pending frames to setup probe timer etc.
2555 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2557 struct sk_buff *skb = tcp_send_head(sk);
2559 BUG_ON(!skb || skb->len < mss_now);
2561 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2564 /* This function returns the amount that we can raise the
2565 * usable window based on the following constraints
2567 * 1. The window can never be shrunk once it is offered (RFC 793)
2568 * 2. We limit memory per socket
2571 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2572 * RECV.NEXT + RCV.WIN fixed until:
2573 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2575 * i.e. don't raise the right edge of the window until you can raise
2576 * it at least MSS bytes.
2578 * Unfortunately, the recommended algorithm breaks header prediction,
2579 * since header prediction assumes th->window stays fixed.
2581 * Strictly speaking, keeping th->window fixed violates the receiver
2582 * side SWS prevention criteria. The problem is that under this rule
2583 * a stream of single byte packets will cause the right side of the
2584 * window to always advance by a single byte.
2586 * Of course, if the sender implements sender side SWS prevention
2587 * then this will not be a problem.
2589 * BSD seems to make the following compromise:
2591 * If the free space is less than the 1/4 of the maximum
2592 * space available and the free space is less than 1/2 mss,
2593 * then set the window to 0.
2594 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2595 * Otherwise, just prevent the window from shrinking
2596 * and from being larger than the largest representable value.
2598 * This prevents incremental opening of the window in the regime
2599 * where TCP is limited by the speed of the reader side taking
2600 * data out of the TCP receive queue. It does nothing about
2601 * those cases where the window is constrained on the sender side
2602 * because the pipeline is full.
2604 * BSD also seems to "accidentally" limit itself to windows that are a
2605 * multiple of MSS, at least until the free space gets quite small.
2606 * This would appear to be a side effect of the mbuf implementation.
2607 * Combining these two algorithms results in the observed behavior
2608 * of having a fixed window size at almost all times.
2610 * Below we obtain similar behavior by forcing the offered window to
2611 * a multiple of the mss when it is feasible to do so.
2613 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2614 * Regular options like TIMESTAMP are taken into account.
2616 u32 __tcp_select_window(struct sock *sk)
2618 struct inet_connection_sock *icsk = inet_csk(sk);
2619 struct tcp_sock *tp = tcp_sk(sk);
2620 /* MSS for the peer's data. Previous versions used mss_clamp
2621 * here. I don't know if the value based on our guesses
2622 * of peer's MSS is better for the performance. It's more correct
2623 * but may be worse for the performance because of rcv_mss
2624 * fluctuations. --SAW 1998/11/1
2626 int mss = icsk->icsk_ack.rcv_mss;
2627 int free_space = tcp_space(sk);
2628 int allowed_space = tcp_full_space(sk);
2629 int full_space = min_t(int, tp->window_clamp, allowed_space);
2632 if (unlikely(mss > full_space)) {
2637 if (free_space < (full_space >> 1)) {
2638 icsk->icsk_ack.quick = 0;
2640 if (tcp_under_memory_pressure(sk))
2641 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2644 /* free_space might become our new window, make sure we don't
2645 * increase it due to wscale.
2647 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2649 /* if free space is less than mss estimate, or is below 1/16th
2650 * of the maximum allowed, try to move to zero-window, else
2651 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2652 * new incoming data is dropped due to memory limits.
2653 * With large window, mss test triggers way too late in order
2654 * to announce zero window in time before rmem limit kicks in.
2656 if (free_space < (allowed_space >> 4) || free_space < mss)
2660 if (free_space > tp->rcv_ssthresh)
2661 free_space = tp->rcv_ssthresh;
2663 /* Don't do rounding if we are using window scaling, since the
2664 * scaled window will not line up with the MSS boundary anyway.
2666 if (tp->rx_opt.rcv_wscale) {
2667 window = free_space;
2669 /* Advertise enough space so that it won't get scaled away.
2670 * Import case: prevent zero window announcement if
2671 * 1<<rcv_wscale > mss.
2673 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2675 window = tp->rcv_wnd;
2676 /* Get the largest window that is a nice multiple of mss.
2677 * Window clamp already applied above.
2678 * If our current window offering is within 1 mss of the
2679 * free space we just keep it. This prevents the divide
2680 * and multiply from happening most of the time.
2681 * We also don't do any window rounding when the free space
2684 if (window <= free_space - mss || window > free_space)
2685 window = rounddown(free_space, mss);
2686 else if (mss == full_space &&
2687 free_space > window + (full_space >> 1))
2688 window = free_space;
2694 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2695 const struct sk_buff *next_skb)
2697 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2698 const struct skb_shared_info *next_shinfo =
2699 skb_shinfo(next_skb);
2700 struct skb_shared_info *shinfo = skb_shinfo(skb);
2702 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2703 shinfo->tskey = next_shinfo->tskey;
2704 TCP_SKB_CB(skb)->txstamp_ack |=
2705 TCP_SKB_CB(next_skb)->txstamp_ack;
2709 /* Collapses two adjacent SKB's during retransmission. */
2710 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2712 struct tcp_sock *tp = tcp_sk(sk);
2713 struct sk_buff *next_skb = skb_rb_next(skb);
2714 int skb_size, next_skb_size;
2716 skb_size = skb->len;
2717 next_skb_size = next_skb->len;
2719 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2721 if (next_skb_size) {
2722 if (next_skb_size <= skb_availroom(skb))
2723 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2725 else if (!skb_shift(skb, next_skb, next_skb_size))
2728 tcp_highest_sack_replace(sk, next_skb, skb);
2730 /* Update sequence range on original skb. */
2731 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2733 /* Merge over control information. This moves PSH/FIN etc. over */
2734 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2736 /* All done, get rid of second SKB and account for it so
2737 * packet counting does not break.
2739 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2740 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2742 /* changed transmit queue under us so clear hints */
2743 tcp_clear_retrans_hints_partial(tp);
2744 if (next_skb == tp->retransmit_skb_hint)
2745 tp->retransmit_skb_hint = skb;
2747 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2749 tcp_skb_collapse_tstamp(skb, next_skb);
2751 tcp_rtx_queue_unlink_and_free(next_skb, sk);
2755 /* Check if coalescing SKBs is legal. */
2756 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2758 if (tcp_skb_pcount(skb) > 1)
2760 if (skb_cloned(skb))
2762 /* Some heuristics for collapsing over SACK'd could be invented */
2763 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2769 /* Collapse packets in the retransmit queue to make to create
2770 * less packets on the wire. This is only done on retransmission.
2772 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2775 struct tcp_sock *tp = tcp_sk(sk);
2776 struct sk_buff *skb = to, *tmp;
2779 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2781 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2784 skb_rbtree_walk_from_safe(skb, tmp) {
2785 if (!tcp_can_collapse(sk, skb))
2788 if (!tcp_skb_can_collapse_to(to))
2801 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2804 if (!tcp_collapse_retrans(sk, to))
2809 /* This retransmits one SKB. Policy decisions and retransmit queue
2810 * state updates are done by the caller. Returns non-zero if an
2811 * error occurred which prevented the send.
2813 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2815 struct inet_connection_sock *icsk = inet_csk(sk);
2816 struct tcp_sock *tp = tcp_sk(sk);
2817 unsigned int cur_mss;
2821 /* Inconclusive MTU probe */
2822 if (icsk->icsk_mtup.probe_size)
2823 icsk->icsk_mtup.probe_size = 0;
2825 /* Do not sent more than we queued. 1/4 is reserved for possible
2826 * copying overhead: fragmentation, tunneling, mangling etc.
2828 if (refcount_read(&sk->sk_wmem_alloc) >
2829 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2833 if (skb_still_in_host_queue(sk, skb))
2836 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2837 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2841 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2845 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2846 return -EHOSTUNREACH; /* Routing failure or similar. */
2848 cur_mss = tcp_current_mss(sk);
2850 /* If receiver has shrunk his window, and skb is out of
2851 * new window, do not retransmit it. The exception is the
2852 * case, when window is shrunk to zero. In this case
2853 * our retransmit serves as a zero window probe.
2855 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2856 TCP_SKB_CB(skb)->seq != tp->snd_una)
2859 len = cur_mss * segs;
2860 if (skb->len > len) {
2861 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2862 cur_mss, GFP_ATOMIC))
2863 return -ENOMEM; /* We'll try again later. */
2865 if (skb_unclone(skb, GFP_ATOMIC))
2868 diff = tcp_skb_pcount(skb);
2869 tcp_set_skb_tso_segs(skb, cur_mss);
2870 diff -= tcp_skb_pcount(skb);
2872 tcp_adjust_pcount(sk, skb, diff);
2873 if (skb->len < cur_mss)
2874 tcp_retrans_try_collapse(sk, skb, cur_mss);
2877 /* RFC3168, section 6.1.1.1. ECN fallback */
2878 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2879 tcp_ecn_clear_syn(sk, skb);
2881 /* Update global and local TCP statistics. */
2882 segs = tcp_skb_pcount(skb);
2883 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2884 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2885 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2886 tp->total_retrans += segs;
2888 /* make sure skb->data is aligned on arches that require it
2889 * and check if ack-trimming & collapsing extended the headroom
2890 * beyond what csum_start can cover.
2892 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2893 skb_headroom(skb) >= 0xFFFF)) {
2894 struct sk_buff *nskb;
2896 tcp_skb_tsorted_save(skb) {
2897 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2898 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2900 } tcp_skb_tsorted_restore(skb);
2903 tcp_update_skb_after_send(tp, skb);
2904 tcp_rate_skb_sent(sk, skb);
2907 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2910 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
2911 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
2912 TCP_SKB_CB(skb)->seq, segs, err);
2915 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2916 trace_tcp_retransmit_skb(sk, skb);
2917 } else if (err != -EBUSY) {
2918 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2923 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2925 struct tcp_sock *tp = tcp_sk(sk);
2926 int err = __tcp_retransmit_skb(sk, skb, segs);
2929 #if FASTRETRANS_DEBUG > 0
2930 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2931 net_dbg_ratelimited("retrans_out leaked\n");
2934 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2935 tp->retrans_out += tcp_skb_pcount(skb);
2937 /* Save stamp of the first retransmit. */
2938 if (!tp->retrans_stamp)
2939 tp->retrans_stamp = tcp_skb_timestamp(skb);
2943 if (tp->undo_retrans < 0)
2944 tp->undo_retrans = 0;
2945 tp->undo_retrans += tcp_skb_pcount(skb);
2949 /* This gets called after a retransmit timeout, and the initially
2950 * retransmitted data is acknowledged. It tries to continue
2951 * resending the rest of the retransmit queue, until either
2952 * we've sent it all or the congestion window limit is reached.
2954 void tcp_xmit_retransmit_queue(struct sock *sk)
2956 const struct inet_connection_sock *icsk = inet_csk(sk);
2957 struct sk_buff *skb, *rtx_head, *hole = NULL;
2958 struct tcp_sock *tp = tcp_sk(sk);
2962 if (!tp->packets_out)
2965 rtx_head = tcp_rtx_queue_head(sk);
2966 skb = tp->retransmit_skb_hint ?: rtx_head;
2967 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2968 skb_rbtree_walk_from(skb) {
2972 if (tcp_pacing_check(sk))
2975 /* we could do better than to assign each time */
2977 tp->retransmit_skb_hint = skb;
2979 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2982 sacked = TCP_SKB_CB(skb)->sacked;
2983 /* In case tcp_shift_skb_data() have aggregated large skbs,
2984 * we need to make sure not sending too bigs TSO packets
2986 segs = min_t(int, segs, max_segs);
2988 if (tp->retrans_out >= tp->lost_out) {
2990 } else if (!(sacked & TCPCB_LOST)) {
2991 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2996 if (icsk->icsk_ca_state != TCP_CA_Loss)
2997 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2999 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3002 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3005 if (tcp_small_queue_check(sk, skb, 1))
3008 if (tcp_retransmit_skb(sk, skb, segs))
3011 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3013 if (tcp_in_cwnd_reduction(sk))
3014 tp->prr_out += tcp_skb_pcount(skb);
3016 if (skb == rtx_head &&
3017 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3018 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3019 inet_csk(sk)->icsk_rto,
3024 /* We allow to exceed memory limits for FIN packets to expedite
3025 * connection tear down and (memory) recovery.
3026 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3027 * or even be forced to close flow without any FIN.
3028 * In general, we want to allow one skb per socket to avoid hangs
3029 * with edge trigger epoll()
3031 void sk_forced_mem_schedule(struct sock *sk, int size)
3035 if (size <= sk->sk_forward_alloc)
3037 amt = sk_mem_pages(size);
3038 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3039 sk_memory_allocated_add(sk, amt);
3041 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3042 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3045 /* Send a FIN. The caller locks the socket for us.
3046 * We should try to send a FIN packet really hard, but eventually give up.
3048 void tcp_send_fin(struct sock *sk)
3050 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3051 struct tcp_sock *tp = tcp_sk(sk);
3053 /* Optimization, tack on the FIN if we have one skb in write queue and
3054 * this skb was not yet sent, or we are under memory pressure.
3055 * Note: in the latter case, FIN packet will be sent after a timeout,
3056 * as TCP stack thinks it has already been transmitted.
3058 if (!tskb && tcp_under_memory_pressure(sk))
3059 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3063 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3064 TCP_SKB_CB(tskb)->end_seq++;
3066 if (tcp_write_queue_empty(sk)) {
3067 /* This means tskb was already sent.
3068 * Pretend we included the FIN on previous transmit.
3069 * We need to set tp->snd_nxt to the value it would have
3070 * if FIN had been sent. This is because retransmit path
3071 * does not change tp->snd_nxt.
3077 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3078 if (unlikely(!skb)) {
3083 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3084 skb_reserve(skb, MAX_TCP_HEADER);
3085 sk_forced_mem_schedule(sk, skb->truesize);
3086 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3087 tcp_init_nondata_skb(skb, tp->write_seq,
3088 TCPHDR_ACK | TCPHDR_FIN);
3089 tcp_queue_skb(sk, skb);
3091 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3094 /* We get here when a process closes a file descriptor (either due to
3095 * an explicit close() or as a byproduct of exit()'ing) and there
3096 * was unread data in the receive queue. This behavior is recommended
3097 * by RFC 2525, section 2.17. -DaveM
3099 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3101 struct sk_buff *skb;
3103 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3105 /* NOTE: No TCP options attached and we never retransmit this. */
3106 skb = alloc_skb(MAX_TCP_HEADER, priority);
3108 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3112 /* Reserve space for headers and prepare control bits. */
3113 skb_reserve(skb, MAX_TCP_HEADER);
3114 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3115 TCPHDR_ACK | TCPHDR_RST);
3116 tcp_mstamp_refresh(tcp_sk(sk));
3118 if (tcp_transmit_skb(sk, skb, 0, priority))
3119 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3121 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3122 * skb here is different to the troublesome skb, so use NULL
3124 trace_tcp_send_reset(sk, NULL);
3127 /* Send a crossed SYN-ACK during socket establishment.
3128 * WARNING: This routine must only be called when we have already sent
3129 * a SYN packet that crossed the incoming SYN that caused this routine
3130 * to get called. If this assumption fails then the initial rcv_wnd
3131 * and rcv_wscale values will not be correct.
3133 int tcp_send_synack(struct sock *sk)
3135 struct sk_buff *skb;
3137 skb = tcp_rtx_queue_head(sk);
3138 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3139 pr_err("%s: wrong queue state\n", __func__);
3142 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3143 if (skb_cloned(skb)) {
3144 struct sk_buff *nskb;
3146 tcp_skb_tsorted_save(skb) {
3147 nskb = skb_copy(skb, GFP_ATOMIC);
3148 } tcp_skb_tsorted_restore(skb);
3151 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3152 tcp_rtx_queue_unlink_and_free(skb, sk);
3153 __skb_header_release(nskb);
3154 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3155 sk->sk_wmem_queued += nskb->truesize;
3156 sk_mem_charge(sk, nskb->truesize);
3160 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3161 tcp_ecn_send_synack(sk, skb);
3163 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3167 * tcp_make_synack - Prepare a SYN-ACK.
3168 * sk: listener socket
3169 * dst: dst entry attached to the SYNACK
3170 * req: request_sock pointer
3172 * Allocate one skb and build a SYNACK packet.
3173 * @dst is consumed : Caller should not use it again.
3175 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3176 struct request_sock *req,
3177 struct tcp_fastopen_cookie *foc,
3178 enum tcp_synack_type synack_type)
3180 struct inet_request_sock *ireq = inet_rsk(req);
3181 const struct tcp_sock *tp = tcp_sk(sk);
3182 struct tcp_md5sig_key *md5 = NULL;
3183 struct tcp_out_options opts;
3184 struct sk_buff *skb;
3185 int tcp_header_size;
3189 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3190 if (unlikely(!skb)) {
3194 /* Reserve space for headers. */
3195 skb_reserve(skb, MAX_TCP_HEADER);
3197 switch (synack_type) {
3198 case TCP_SYNACK_NORMAL:
3199 skb_set_owner_w(skb, req_to_sk(req));
3201 case TCP_SYNACK_COOKIE:
3202 /* Under synflood, we do not attach skb to a socket,
3203 * to avoid false sharing.
3206 case TCP_SYNACK_FASTOPEN:
3207 /* sk is a const pointer, because we want to express multiple
3208 * cpu might call us concurrently.
3209 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3211 skb_set_owner_w(skb, (struct sock *)sk);
3214 skb_dst_set(skb, dst);
3216 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3218 memset(&opts, 0, sizeof(opts));
3219 #ifdef CONFIG_SYN_COOKIES
3220 if (unlikely(req->cookie_ts))
3221 skb->skb_mstamp = cookie_init_timestamp(req);
3224 skb->skb_mstamp = tcp_clock_us();
3226 #ifdef CONFIG_TCP_MD5SIG
3228 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3230 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3231 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3234 skb_push(skb, tcp_header_size);
3235 skb_reset_transport_header(skb);
3237 th = (struct tcphdr *)skb->data;
3238 memset(th, 0, sizeof(struct tcphdr));
3241 tcp_ecn_make_synack(req, th);
3242 th->source = htons(ireq->ir_num);
3243 th->dest = ireq->ir_rmt_port;
3244 skb->mark = ireq->ir_mark;
3245 skb->ip_summed = CHECKSUM_PARTIAL;
3246 th->seq = htonl(tcp_rsk(req)->snt_isn);
3247 /* XXX data is queued and acked as is. No buffer/window check */
3248 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3250 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3251 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3252 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3253 th->doff = (tcp_header_size >> 2);
3254 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3256 #ifdef CONFIG_TCP_MD5SIG
3257 /* Okay, we have all we need - do the md5 hash if needed */
3259 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3260 md5, req_to_sk(req), skb);
3264 /* Do not fool tcpdump (if any), clean our debris */
3268 EXPORT_SYMBOL(tcp_make_synack);
3270 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3272 struct inet_connection_sock *icsk = inet_csk(sk);
3273 const struct tcp_congestion_ops *ca;
3274 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3276 if (ca_key == TCP_CA_UNSPEC)
3280 ca = tcp_ca_find_key(ca_key);
3281 if (likely(ca && try_module_get(ca->owner))) {
3282 module_put(icsk->icsk_ca_ops->owner);
3283 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3284 icsk->icsk_ca_ops = ca;
3289 /* Do all connect socket setups that can be done AF independent. */
3290 static void tcp_connect_init(struct sock *sk)
3292 const struct dst_entry *dst = __sk_dst_get(sk);
3293 struct tcp_sock *tp = tcp_sk(sk);
3297 /* We'll fix this up when we get a response from the other end.
3298 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3300 tp->tcp_header_len = sizeof(struct tcphdr);
3301 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3302 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3304 #ifdef CONFIG_TCP_MD5SIG
3305 if (tp->af_specific->md5_lookup(sk, sk))
3306 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3309 /* If user gave his TCP_MAXSEG, record it to clamp */
3310 if (tp->rx_opt.user_mss)
3311 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3314 tcp_sync_mss(sk, dst_mtu(dst));
3316 tcp_ca_dst_init(sk, dst);
3318 if (!tp->window_clamp)
3319 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3320 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3322 tcp_initialize_rcv_mss(sk);
3324 /* limit the window selection if the user enforce a smaller rx buffer */
3325 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3326 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3327 tp->window_clamp = tcp_full_space(sk);
3329 rcv_wnd = tcp_rwnd_init_bpf(sk);
3331 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3333 tcp_select_initial_window(sk, tcp_full_space(sk),
3334 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3337 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3341 tp->rx_opt.rcv_wscale = rcv_wscale;
3342 tp->rcv_ssthresh = tp->rcv_wnd;
3345 sock_reset_flag(sk, SOCK_DONE);
3348 tcp_write_queue_purge(sk);
3349 tp->snd_una = tp->write_seq;
3350 tp->snd_sml = tp->write_seq;
3351 tp->snd_up = tp->write_seq;
3352 tp->snd_nxt = tp->write_seq;
3354 if (likely(!tp->repair))
3357 tp->rcv_tstamp = tcp_jiffies32;
3358 tp->rcv_wup = tp->rcv_nxt;
3359 tp->copied_seq = tp->rcv_nxt;
3361 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3362 inet_csk(sk)->icsk_retransmits = 0;
3363 tcp_clear_retrans(tp);
3366 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3368 struct tcp_sock *tp = tcp_sk(sk);
3369 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3371 tcb->end_seq += skb->len;
3372 __skb_header_release(skb);
3373 sk->sk_wmem_queued += skb->truesize;
3374 sk_mem_charge(sk, skb->truesize);
3375 tp->write_seq = tcb->end_seq;
3376 tp->packets_out += tcp_skb_pcount(skb);
3379 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3380 * queue a data-only packet after the regular SYN, such that regular SYNs
3381 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3382 * only the SYN sequence, the data are retransmitted in the first ACK.
3383 * If cookie is not cached or other error occurs, falls back to send a
3384 * regular SYN with Fast Open cookie request option.
3386 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3388 struct tcp_sock *tp = tcp_sk(sk);
3389 struct tcp_fastopen_request *fo = tp->fastopen_req;
3391 struct sk_buff *syn_data;
3393 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3394 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3397 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3398 * user-MSS. Reserve maximum option space for middleboxes that add
3399 * private TCP options. The cost is reduced data space in SYN :(
3401 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3403 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3404 MAX_TCP_OPTION_SPACE;
3406 space = min_t(size_t, space, fo->size);
3408 /* limit to order-0 allocations */
3409 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3411 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3414 syn_data->ip_summed = CHECKSUM_PARTIAL;
3415 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3417 int copied = copy_from_iter(skb_put(syn_data, space), space,
3418 &fo->data->msg_iter);
3419 if (unlikely(!copied)) {
3420 tcp_skb_tsorted_anchor_cleanup(syn_data);
3421 kfree_skb(syn_data);
3424 if (copied != space) {
3425 skb_trim(syn_data, copied);
3429 /* No more data pending in inet_wait_for_connect() */
3430 if (space == fo->size)
3434 tcp_connect_queue_skb(sk, syn_data);
3436 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3438 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3440 syn->skb_mstamp = syn_data->skb_mstamp;
3442 /* Now full SYN+DATA was cloned and sent (or not),
3443 * remove the SYN from the original skb (syn_data)
3444 * we keep in write queue in case of a retransmit, as we
3445 * also have the SYN packet (with no data) in the same queue.
3447 TCP_SKB_CB(syn_data)->seq++;
3448 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3450 tp->syn_data = (fo->copied > 0);
3451 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3452 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3456 /* data was not sent, put it in write_queue */
3457 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3458 tp->packets_out -= tcp_skb_pcount(syn_data);
3461 /* Send a regular SYN with Fast Open cookie request option */
3462 if (fo->cookie.len > 0)
3464 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3466 tp->syn_fastopen = 0;
3468 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3472 /* Build a SYN and send it off. */
3473 int tcp_connect(struct sock *sk)
3475 struct tcp_sock *tp = tcp_sk(sk);
3476 struct sk_buff *buff;
3479 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3481 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3482 return -EHOSTUNREACH; /* Routing failure or similar. */
3484 tcp_connect_init(sk);
3486 if (unlikely(tp->repair)) {
3487 tcp_finish_connect(sk, NULL);
3491 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3492 if (unlikely(!buff))
3495 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3496 tcp_mstamp_refresh(tp);
3497 tp->retrans_stamp = tcp_time_stamp(tp);
3498 tcp_connect_queue_skb(sk, buff);
3499 tcp_ecn_send_syn(sk, buff);
3500 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3502 /* Send off SYN; include data in Fast Open. */
3503 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3504 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3505 if (err == -ECONNREFUSED)
3508 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3509 * in order to make this packet get counted in tcpOutSegs.
3511 tp->snd_nxt = tp->write_seq;
3512 tp->pushed_seq = tp->write_seq;
3513 buff = tcp_send_head(sk);
3514 if (unlikely(buff)) {
3515 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3516 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3518 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3520 /* Timer for repeating the SYN until an answer. */
3521 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3522 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3525 EXPORT_SYMBOL(tcp_connect);
3527 /* Send out a delayed ack, the caller does the policy checking
3528 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3531 void tcp_send_delayed_ack(struct sock *sk)
3533 struct inet_connection_sock *icsk = inet_csk(sk);
3534 int ato = icsk->icsk_ack.ato;
3535 unsigned long timeout;
3537 if (ato > TCP_DELACK_MIN) {
3538 const struct tcp_sock *tp = tcp_sk(sk);
3539 int max_ato = HZ / 2;
3541 if (icsk->icsk_ack.pingpong ||
3542 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3543 max_ato = TCP_DELACK_MAX;
3545 /* Slow path, intersegment interval is "high". */
3547 /* If some rtt estimate is known, use it to bound delayed ack.
3548 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3552 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3559 ato = min(ato, max_ato);
3562 /* Stay within the limit we were given */
3563 timeout = jiffies + ato;
3565 /* Use new timeout only if there wasn't a older one earlier. */
3566 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3567 /* If delack timer was blocked or is about to expire,
3570 if (icsk->icsk_ack.blocked ||
3571 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3576 if (!time_before(timeout, icsk->icsk_ack.timeout))
3577 timeout = icsk->icsk_ack.timeout;
3579 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3580 icsk->icsk_ack.timeout = timeout;
3581 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3584 /* This routine sends an ack and also updates the window. */
3585 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3587 struct sk_buff *buff;
3589 /* If we have been reset, we may not send again. */
3590 if (sk->sk_state == TCP_CLOSE)
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, rcv_nxt);
3620 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3622 void tcp_send_ack(struct sock *sk)
3624 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3627 /* This routine sends a packet with an out of date sequence
3628 * number. It assumes the other end will try to ack it.
3630 * Question: what should we make while urgent mode?
3631 * 4.4BSD forces sending single byte of data. We cannot send
3632 * out of window data, because we have SND.NXT==SND.MAX...
3634 * Current solution: to send TWO zero-length segments in urgent mode:
3635 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3636 * out-of-date with SND.UNA-1 to probe window.
3638 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3640 struct tcp_sock *tp = tcp_sk(sk);
3641 struct sk_buff *skb;
3643 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3644 skb = alloc_skb(MAX_TCP_HEADER,
3645 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3649 /* Reserve space for headers and set control bits. */
3650 skb_reserve(skb, MAX_TCP_HEADER);
3651 /* Use a previous sequence. This should cause the other
3652 * end to send an ack. Don't queue or clone SKB, just
3655 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3656 NET_INC_STATS(sock_net(sk), mib);
3657 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3660 /* Called from setsockopt( ... TCP_REPAIR ) */
3661 void tcp_send_window_probe(struct sock *sk)
3663 if (sk->sk_state == TCP_ESTABLISHED) {
3664 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3665 tcp_mstamp_refresh(tcp_sk(sk));
3666 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3670 /* Initiate keepalive or window probe from timer. */
3671 int tcp_write_wakeup(struct sock *sk, int mib)
3673 struct tcp_sock *tp = tcp_sk(sk);
3674 struct sk_buff *skb;
3676 if (sk->sk_state == TCP_CLOSE)
3679 skb = tcp_send_head(sk);
3680 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3682 unsigned int mss = tcp_current_mss(sk);
3683 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3685 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3686 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3688 /* We are probing the opening of a window
3689 * but the window size is != 0
3690 * must have been a result SWS avoidance ( sender )
3692 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3694 seg_size = min(seg_size, mss);
3695 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3696 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3697 skb, seg_size, mss, GFP_ATOMIC))
3699 } else if (!tcp_skb_pcount(skb))
3700 tcp_set_skb_tso_segs(skb, mss);
3702 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3703 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3705 tcp_event_new_data_sent(sk, skb);
3708 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3709 tcp_xmit_probe_skb(sk, 1, mib);
3710 return tcp_xmit_probe_skb(sk, 0, mib);
3714 /* A window probe timeout has occurred. If window is not closed send
3715 * a partial packet else a zero probe.
3717 void tcp_send_probe0(struct sock *sk)
3719 struct inet_connection_sock *icsk = inet_csk(sk);
3720 struct tcp_sock *tp = tcp_sk(sk);
3721 struct net *net = sock_net(sk);
3722 unsigned long probe_max;
3725 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3727 if (tp->packets_out || tcp_write_queue_empty(sk)) {
3728 /* Cancel probe timer, if it is not required. */
3729 icsk->icsk_probes_out = 0;
3730 icsk->icsk_backoff = 0;
3735 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3736 icsk->icsk_backoff++;
3737 icsk->icsk_probes_out++;
3738 probe_max = TCP_RTO_MAX;
3740 /* If packet was not sent due to local congestion,
3741 * do not backoff and do not remember icsk_probes_out.
3742 * Let local senders to fight for local resources.
3744 * Use accumulated backoff yet.
3746 if (!icsk->icsk_probes_out)
3747 icsk->icsk_probes_out = 1;
3748 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3750 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3751 tcp_probe0_when(sk, probe_max),
3755 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3757 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3761 tcp_rsk(req)->txhash = net_tx_rndhash();
3762 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3764 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3765 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3766 if (unlikely(tcp_passive_fastopen(sk)))
3767 tcp_sk(sk)->total_retrans++;
3768 trace_tcp_retransmit_synack(sk, req);
3772 EXPORT_SYMBOL(tcp_rtx_synack);