2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
44 #include <linux/static_key.h>
46 #include <trace/events/tcp.h>
48 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
49 int push_one, gfp_t gfp);
51 /* Account for new data that has been sent to the network. */
52 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
54 struct inet_connection_sock *icsk = inet_csk(sk);
55 struct tcp_sock *tp = tcp_sk(sk);
56 unsigned int prior_packets = tp->packets_out;
58 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
60 __skb_unlink(skb, &sk->sk_write_queue);
61 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
63 tp->packets_out += tcp_skb_pcount(skb);
64 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
67 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
71 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
72 * window scaling factor due to loss of precision.
73 * If window has been shrunk, what should we make? It is not clear at all.
74 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
75 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
76 * invalid. OK, let's make this for now:
78 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
80 const struct tcp_sock *tp = tcp_sk(sk);
82 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
83 (tp->rx_opt.wscale_ok &&
84 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
87 return tcp_wnd_end(tp);
90 /* Calculate mss to advertise in SYN segment.
91 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
93 * 1. It is independent of path mtu.
94 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
95 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
96 * attached devices, because some buggy hosts are confused by
98 * 4. We do not make 3, we advertise MSS, calculated from first
99 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
100 * This may be overridden via information stored in routing table.
101 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
102 * probably even Jumbo".
104 static __u16 tcp_advertise_mss(struct sock *sk)
106 struct tcp_sock *tp = tcp_sk(sk);
107 const struct dst_entry *dst = __sk_dst_get(sk);
108 int mss = tp->advmss;
111 unsigned int metric = dst_metric_advmss(dst);
122 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
123 * This is the first part of cwnd validation mechanism.
125 void tcp_cwnd_restart(struct sock *sk, s32 delta)
127 struct tcp_sock *tp = tcp_sk(sk);
128 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
129 u32 cwnd = tp->snd_cwnd;
131 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
133 tp->snd_ssthresh = tcp_current_ssthresh(sk);
134 restart_cwnd = min(restart_cwnd, cwnd);
136 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
138 tp->snd_cwnd = max(cwnd, restart_cwnd);
139 tp->snd_cwnd_stamp = tcp_jiffies32;
140 tp->snd_cwnd_used = 0;
143 /* Congestion state accounting after a packet has been sent. */
144 static void tcp_event_data_sent(struct tcp_sock *tp,
147 struct inet_connection_sock *icsk = inet_csk(sk);
148 const u32 now = tcp_jiffies32;
150 if (tcp_packets_in_flight(tp) == 0)
151 tcp_ca_event(sk, CA_EVENT_TX_START);
155 /* If it is a reply for ato after last received
156 * packet, enter pingpong mode.
158 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
159 icsk->icsk_ack.pingpong = 1;
162 /* Account for an ACK we sent. */
163 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
165 struct tcp_sock *tp = tcp_sk(sk);
167 if (unlikely(tp->compressed_ack)) {
168 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
170 tp->compressed_ack = 0;
171 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
174 tcp_dec_quickack_mode(sk, pkts);
175 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
179 u32 tcp_default_init_rwnd(u32 mss)
181 /* Initial receive window should be twice of TCP_INIT_CWND to
182 * enable proper sending of new unsent data during fast recovery
183 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
184 * limit when mss is larger than 1460.
186 u32 init_rwnd = TCP_INIT_CWND * 2;
189 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
193 /* Determine a window scaling and initial window to offer.
194 * Based on the assumption that the given amount of space
195 * will be offered. Store the results in the tp structure.
196 * NOTE: for smooth operation initial space offering should
197 * be a multiple of mss if possible. We assume here that mss >= 1.
198 * This MUST be enforced by all callers.
200 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
201 __u32 *rcv_wnd, __u32 *window_clamp,
202 int wscale_ok, __u8 *rcv_wscale,
205 unsigned int space = (__space < 0 ? 0 : __space);
207 /* If no clamp set the clamp to the max possible scaled window */
208 if (*window_clamp == 0)
209 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
210 space = min(*window_clamp, space);
212 /* Quantize space offering to a multiple of mss if possible. */
214 space = rounddown(space, mss);
216 /* NOTE: offering an initial window larger than 32767
217 * will break some buggy TCP stacks. If the admin tells us
218 * it is likely we could be speaking with such a buggy stack
219 * we will truncate our initial window offering to 32K-1
220 * unless the remote has sent us a window scaling option,
221 * which we interpret as a sign the remote TCP is not
222 * misinterpreting the window field as a signed quantity.
224 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
225 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
231 /* Set window scaling on max possible window */
232 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
233 space = max_t(u32, space, sysctl_rmem_max);
234 space = min_t(u32, space, *window_clamp);
235 while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
241 if (!init_rcv_wnd) /* Use default unless specified otherwise */
242 init_rcv_wnd = tcp_default_init_rwnd(mss);
243 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
245 /* Set the clamp no higher than max representable value */
246 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
248 EXPORT_SYMBOL(tcp_select_initial_window);
250 /* Chose a new window to advertise, update state in tcp_sock for the
251 * socket, and return result with RFC1323 scaling applied. The return
252 * value can be stuffed directly into th->window for an outgoing
255 static u16 tcp_select_window(struct sock *sk)
257 struct tcp_sock *tp = tcp_sk(sk);
258 u32 old_win = tp->rcv_wnd;
259 u32 cur_win = tcp_receive_window(tp);
260 u32 new_win = __tcp_select_window(sk);
262 /* Never shrink the offered window */
263 if (new_win < cur_win) {
264 /* Danger Will Robinson!
265 * Don't update rcv_wup/rcv_wnd here or else
266 * we will not be able to advertise a zero
267 * window in time. --DaveM
269 * Relax Will Robinson.
272 NET_INC_STATS(sock_net(sk),
273 LINUX_MIB_TCPWANTZEROWINDOWADV);
274 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
276 tp->rcv_wnd = new_win;
277 tp->rcv_wup = tp->rcv_nxt;
279 /* Make sure we do not exceed the maximum possible
282 if (!tp->rx_opt.rcv_wscale &&
283 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
284 new_win = min(new_win, MAX_TCP_WINDOW);
286 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
288 /* RFC1323 scaling applied */
289 new_win >>= tp->rx_opt.rcv_wscale;
291 /* If we advertise zero window, disable fast path. */
295 NET_INC_STATS(sock_net(sk),
296 LINUX_MIB_TCPTOZEROWINDOWADV);
297 } else if (old_win == 0) {
298 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
304 /* Packet ECN state for a SYN-ACK */
305 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
307 const struct tcp_sock *tp = tcp_sk(sk);
309 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
310 if (!(tp->ecn_flags & TCP_ECN_OK))
311 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
312 else if (tcp_ca_needs_ecn(sk) ||
313 tcp_bpf_ca_needs_ecn(sk))
317 /* Packet ECN state for a SYN. */
318 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
320 struct tcp_sock *tp = tcp_sk(sk);
321 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
322 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
323 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
326 const struct dst_entry *dst = __sk_dst_get(sk);
328 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
335 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
336 tp->ecn_flags = TCP_ECN_OK;
337 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
342 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
344 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
345 /* tp->ecn_flags are cleared at a later point in time when
346 * SYN ACK is ultimatively being received.
348 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
352 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
354 if (inet_rsk(req)->ecn_ok)
358 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
361 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
362 struct tcphdr *th, int tcp_header_len)
364 struct tcp_sock *tp = tcp_sk(sk);
366 if (tp->ecn_flags & TCP_ECN_OK) {
367 /* Not-retransmitted data segment: set ECT and inject CWR. */
368 if (skb->len != tcp_header_len &&
369 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
371 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
372 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
374 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
376 } else if (!tcp_ca_needs_ecn(sk)) {
377 /* ACK or retransmitted segment: clear ECT|CE */
378 INET_ECN_dontxmit(sk);
380 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
385 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
386 * auto increment end seqno.
388 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
390 skb->ip_summed = CHECKSUM_PARTIAL;
392 TCP_SKB_CB(skb)->tcp_flags = flags;
393 TCP_SKB_CB(skb)->sacked = 0;
395 tcp_skb_pcount_set(skb, 1);
397 TCP_SKB_CB(skb)->seq = seq;
398 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
400 TCP_SKB_CB(skb)->end_seq = seq;
403 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
405 return tp->snd_una != tp->snd_up;
408 #define OPTION_SACK_ADVERTISE (1 << 0)
409 #define OPTION_TS (1 << 1)
410 #define OPTION_MD5 (1 << 2)
411 #define OPTION_WSCALE (1 << 3)
412 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
413 #define OPTION_SMC (1 << 9)
415 static void smc_options_write(__be32 *ptr, u16 *options)
417 #if IS_ENABLED(CONFIG_SMC)
418 if (static_branch_unlikely(&tcp_have_smc)) {
419 if (unlikely(OPTION_SMC & *options)) {
420 *ptr++ = htonl((TCPOPT_NOP << 24) |
423 (TCPOLEN_EXP_SMC_BASE));
424 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
430 struct tcp_out_options {
431 u16 options; /* bit field of OPTION_* */
432 u16 mss; /* 0 to disable */
433 u8 ws; /* window scale, 0 to disable */
434 u8 num_sack_blocks; /* number of SACK blocks to include */
435 u8 hash_size; /* bytes in hash_location */
436 __u8 *hash_location; /* temporary pointer, overloaded */
437 __u32 tsval, tsecr; /* need to include OPTION_TS */
438 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
441 /* Write previously computed TCP options to the packet.
443 * Beware: Something in the Internet is very sensitive to the ordering of
444 * TCP options, we learned this through the hard way, so be careful here.
445 * Luckily we can at least blame others for their non-compliance but from
446 * inter-operability perspective it seems that we're somewhat stuck with
447 * the ordering which we have been using if we want to keep working with
448 * those broken things (not that it currently hurts anybody as there isn't
449 * particular reason why the ordering would need to be changed).
451 * At least SACK_PERM as the first option is known to lead to a disaster
452 * (but it may well be that other scenarios fail similarly).
454 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
455 struct tcp_out_options *opts)
457 u16 options = opts->options; /* mungable copy */
459 if (unlikely(OPTION_MD5 & options)) {
460 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
461 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
462 /* overload cookie hash location */
463 opts->hash_location = (__u8 *)ptr;
467 if (unlikely(opts->mss)) {
468 *ptr++ = htonl((TCPOPT_MSS << 24) |
469 (TCPOLEN_MSS << 16) |
473 if (likely(OPTION_TS & options)) {
474 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
475 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
476 (TCPOLEN_SACK_PERM << 16) |
477 (TCPOPT_TIMESTAMP << 8) |
479 options &= ~OPTION_SACK_ADVERTISE;
481 *ptr++ = htonl((TCPOPT_NOP << 24) |
483 (TCPOPT_TIMESTAMP << 8) |
486 *ptr++ = htonl(opts->tsval);
487 *ptr++ = htonl(opts->tsecr);
490 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
491 *ptr++ = htonl((TCPOPT_NOP << 24) |
493 (TCPOPT_SACK_PERM << 8) |
497 if (unlikely(OPTION_WSCALE & options)) {
498 *ptr++ = htonl((TCPOPT_NOP << 24) |
499 (TCPOPT_WINDOW << 16) |
500 (TCPOLEN_WINDOW << 8) |
504 if (unlikely(opts->num_sack_blocks)) {
505 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
506 tp->duplicate_sack : tp->selective_acks;
509 *ptr++ = htonl((TCPOPT_NOP << 24) |
512 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
513 TCPOLEN_SACK_PERBLOCK)));
515 for (this_sack = 0; this_sack < opts->num_sack_blocks;
517 *ptr++ = htonl(sp[this_sack].start_seq);
518 *ptr++ = htonl(sp[this_sack].end_seq);
521 tp->rx_opt.dsack = 0;
524 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
525 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
527 u32 len; /* Fast Open option length */
530 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
531 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
532 TCPOPT_FASTOPEN_MAGIC);
533 p += TCPOLEN_EXP_FASTOPEN_BASE;
535 len = TCPOLEN_FASTOPEN_BASE + foc->len;
536 *p++ = TCPOPT_FASTOPEN;
540 memcpy(p, foc->val, foc->len);
541 if ((len & 3) == 2) {
542 p[foc->len] = TCPOPT_NOP;
543 p[foc->len + 1] = TCPOPT_NOP;
545 ptr += (len + 3) >> 2;
548 smc_options_write(ptr, &options);
551 static void smc_set_option(const struct tcp_sock *tp,
552 struct tcp_out_options *opts,
553 unsigned int *remaining)
555 #if IS_ENABLED(CONFIG_SMC)
556 if (static_branch_unlikely(&tcp_have_smc)) {
558 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
559 opts->options |= OPTION_SMC;
560 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
567 static void smc_set_option_cond(const struct tcp_sock *tp,
568 const struct inet_request_sock *ireq,
569 struct tcp_out_options *opts,
570 unsigned int *remaining)
572 #if IS_ENABLED(CONFIG_SMC)
573 if (static_branch_unlikely(&tcp_have_smc)) {
574 if (tp->syn_smc && ireq->smc_ok) {
575 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
576 opts->options |= OPTION_SMC;
577 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
584 /* Compute TCP options for SYN packets. This is not the final
585 * network wire format yet.
587 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
588 struct tcp_out_options *opts,
589 struct tcp_md5sig_key **md5)
591 struct tcp_sock *tp = tcp_sk(sk);
592 unsigned int remaining = MAX_TCP_OPTION_SPACE;
593 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
596 #ifdef CONFIG_TCP_MD5SIG
597 if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
598 *md5 = tp->af_specific->md5_lookup(sk, sk);
600 opts->options |= OPTION_MD5;
601 remaining -= TCPOLEN_MD5SIG_ALIGNED;
606 /* We always get an MSS option. The option bytes which will be seen in
607 * normal data packets should timestamps be used, must be in the MSS
608 * advertised. But we subtract them from tp->mss_cache so that
609 * calculations in tcp_sendmsg are simpler etc. So account for this
610 * fact here if necessary. If we don't do this correctly, as a
611 * receiver we won't recognize data packets as being full sized when we
612 * should, and thus we won't abide by the delayed ACK rules correctly.
613 * SACKs don't matter, we never delay an ACK when we have any of those
615 opts->mss = tcp_advertise_mss(sk);
616 remaining -= TCPOLEN_MSS_ALIGNED;
618 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
619 opts->options |= OPTION_TS;
620 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
621 opts->tsecr = tp->rx_opt.ts_recent;
622 remaining -= TCPOLEN_TSTAMP_ALIGNED;
624 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
625 opts->ws = tp->rx_opt.rcv_wscale;
626 opts->options |= OPTION_WSCALE;
627 remaining -= TCPOLEN_WSCALE_ALIGNED;
629 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
630 opts->options |= OPTION_SACK_ADVERTISE;
631 if (unlikely(!(OPTION_TS & opts->options)))
632 remaining -= TCPOLEN_SACKPERM_ALIGNED;
635 if (fastopen && fastopen->cookie.len >= 0) {
636 u32 need = fastopen->cookie.len;
638 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
639 TCPOLEN_FASTOPEN_BASE;
640 need = (need + 3) & ~3U; /* Align to 32 bits */
641 if (remaining >= need) {
642 opts->options |= OPTION_FAST_OPEN_COOKIE;
643 opts->fastopen_cookie = &fastopen->cookie;
645 tp->syn_fastopen = 1;
646 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
650 smc_set_option(tp, opts, &remaining);
652 return MAX_TCP_OPTION_SPACE - remaining;
655 /* Set up TCP options for SYN-ACKs. */
656 static unsigned int tcp_synack_options(const struct sock *sk,
657 struct request_sock *req,
658 unsigned int mss, struct sk_buff *skb,
659 struct tcp_out_options *opts,
660 const struct tcp_md5sig_key *md5,
661 struct tcp_fastopen_cookie *foc)
663 struct inet_request_sock *ireq = inet_rsk(req);
664 unsigned int remaining = MAX_TCP_OPTION_SPACE;
666 #ifdef CONFIG_TCP_MD5SIG
668 opts->options |= OPTION_MD5;
669 remaining -= TCPOLEN_MD5SIG_ALIGNED;
671 /* We can't fit any SACK blocks in a packet with MD5 + TS
672 * options. There was discussion about disabling SACK
673 * rather than TS in order to fit in better with old,
674 * buggy kernels, but that was deemed to be unnecessary.
676 ireq->tstamp_ok &= !ireq->sack_ok;
680 /* We always send an MSS option. */
682 remaining -= TCPOLEN_MSS_ALIGNED;
684 if (likely(ireq->wscale_ok)) {
685 opts->ws = ireq->rcv_wscale;
686 opts->options |= OPTION_WSCALE;
687 remaining -= TCPOLEN_WSCALE_ALIGNED;
689 if (likely(ireq->tstamp_ok)) {
690 opts->options |= OPTION_TS;
691 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
692 opts->tsecr = req->ts_recent;
693 remaining -= TCPOLEN_TSTAMP_ALIGNED;
695 if (likely(ireq->sack_ok)) {
696 opts->options |= OPTION_SACK_ADVERTISE;
697 if (unlikely(!ireq->tstamp_ok))
698 remaining -= TCPOLEN_SACKPERM_ALIGNED;
700 if (foc != NULL && foc->len >= 0) {
703 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
704 TCPOLEN_FASTOPEN_BASE;
705 need = (need + 3) & ~3U; /* Align to 32 bits */
706 if (remaining >= need) {
707 opts->options |= OPTION_FAST_OPEN_COOKIE;
708 opts->fastopen_cookie = foc;
713 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
715 return MAX_TCP_OPTION_SPACE - remaining;
718 /* Compute TCP options for ESTABLISHED sockets. This is not the
719 * final wire format yet.
721 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
722 struct tcp_out_options *opts,
723 struct tcp_md5sig_key **md5)
725 struct tcp_sock *tp = tcp_sk(sk);
726 unsigned int size = 0;
727 unsigned int eff_sacks;
732 #ifdef CONFIG_TCP_MD5SIG
733 if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
734 *md5 = tp->af_specific->md5_lookup(sk, sk);
736 opts->options |= OPTION_MD5;
737 size += TCPOLEN_MD5SIG_ALIGNED;
742 if (likely(tp->rx_opt.tstamp_ok)) {
743 opts->options |= OPTION_TS;
744 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
745 opts->tsecr = tp->rx_opt.ts_recent;
746 size += TCPOLEN_TSTAMP_ALIGNED;
749 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
750 if (unlikely(eff_sacks)) {
751 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
752 opts->num_sack_blocks =
753 min_t(unsigned int, eff_sacks,
754 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
755 TCPOLEN_SACK_PERBLOCK);
756 size += TCPOLEN_SACK_BASE_ALIGNED +
757 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
764 /* TCP SMALL QUEUES (TSQ)
766 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
767 * to reduce RTT and bufferbloat.
768 * We do this using a special skb destructor (tcp_wfree).
770 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
771 * needs to be reallocated in a driver.
772 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
774 * Since transmit from skb destructor is forbidden, we use a tasklet
775 * to process all sockets that eventually need to send more skbs.
776 * We use one tasklet per cpu, with its own queue of sockets.
779 struct tasklet_struct tasklet;
780 struct list_head head; /* queue of tcp sockets */
782 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
784 static void tcp_tsq_write(struct sock *sk)
786 if ((1 << sk->sk_state) &
787 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
788 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
789 struct tcp_sock *tp = tcp_sk(sk);
791 if (tp->lost_out > tp->retrans_out &&
792 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
793 tcp_mstamp_refresh(tp);
794 tcp_xmit_retransmit_queue(sk);
797 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
802 static void tcp_tsq_handler(struct sock *sk)
805 if (!sock_owned_by_user(sk))
807 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
812 * One tasklet per cpu tries to send more skbs.
813 * We run in tasklet context but need to disable irqs when
814 * transferring tsq->head because tcp_wfree() might
815 * interrupt us (non NAPI drivers)
817 static void tcp_tasklet_func(unsigned long data)
819 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
822 struct list_head *q, *n;
826 local_irq_save(flags);
827 list_splice_init(&tsq->head, &list);
828 local_irq_restore(flags);
830 list_for_each_safe(q, n, &list) {
831 tp = list_entry(q, struct tcp_sock, tsq_node);
832 list_del(&tp->tsq_node);
834 sk = (struct sock *)tp;
835 smp_mb__before_atomic();
836 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
843 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
844 TCPF_WRITE_TIMER_DEFERRED | \
845 TCPF_DELACK_TIMER_DEFERRED | \
846 TCPF_MTU_REDUCED_DEFERRED)
848 * tcp_release_cb - tcp release_sock() callback
851 * called from release_sock() to perform protocol dependent
852 * actions before socket release.
854 void tcp_release_cb(struct sock *sk)
856 unsigned long flags, nflags;
858 /* perform an atomic operation only if at least one flag is set */
860 flags = sk->sk_tsq_flags;
861 if (!(flags & TCP_DEFERRED_ALL))
863 nflags = flags & ~TCP_DEFERRED_ALL;
864 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
866 if (flags & TCPF_TSQ_DEFERRED) {
870 /* Here begins the tricky part :
871 * We are called from release_sock() with :
873 * 2) sk_lock.slock spinlock held
874 * 3) socket owned by us (sk->sk_lock.owned == 1)
876 * But following code is meant to be called from BH handlers,
877 * so we should keep BH disabled, but early release socket ownership
879 sock_release_ownership(sk);
881 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
882 tcp_write_timer_handler(sk);
885 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
886 tcp_delack_timer_handler(sk);
889 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
890 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
894 EXPORT_SYMBOL(tcp_release_cb);
896 void __init tcp_tasklet_init(void)
900 for_each_possible_cpu(i) {
901 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
903 INIT_LIST_HEAD(&tsq->head);
904 tasklet_init(&tsq->tasklet,
911 * Write buffer destructor automatically called from kfree_skb.
912 * We can't xmit new skbs from this context, as we might already
915 void tcp_wfree(struct sk_buff *skb)
917 struct sock *sk = skb->sk;
918 struct tcp_sock *tp = tcp_sk(sk);
919 unsigned long flags, nval, oval;
921 /* Keep one reference on sk_wmem_alloc.
922 * Will be released by sk_free() from here or tcp_tasklet_func()
924 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
926 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
927 * Wait until our queues (qdisc + devices) are drained.
929 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
930 * - chance for incoming ACK (processed by another cpu maybe)
931 * to migrate this flow (skb->ooo_okay will be eventually set)
933 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
936 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
937 struct tsq_tasklet *tsq;
940 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
943 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
944 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
948 /* queue this socket to tasklet queue */
949 local_irq_save(flags);
950 tsq = this_cpu_ptr(&tsq_tasklet);
951 empty = list_empty(&tsq->head);
952 list_add(&tp->tsq_node, &tsq->head);
954 tasklet_schedule(&tsq->tasklet);
955 local_irq_restore(flags);
962 /* Note: Called under soft irq.
963 * We can call TCP stack right away, unless socket is owned by user.
965 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
967 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
968 struct sock *sk = (struct sock *)tp;
973 return HRTIMER_NORESTART;
976 /* BBR congestion control needs pacing.
977 * Same remark for SO_MAX_PACING_RATE.
978 * sch_fq packet scheduler is efficiently handling pacing,
979 * but is not always installed/used.
980 * Return true if TCP stack should pace packets itself.
982 static bool tcp_needs_internal_pacing(const struct sock *sk)
984 return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
987 static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
992 if (!tcp_needs_internal_pacing(sk))
994 rate = sk->sk_pacing_rate;
995 if (!rate || rate == ~0U)
998 /* Should account for header sizes as sch_fq does,
999 * but lets make things simple.
1001 len_ns = (u64)skb->len * NSEC_PER_SEC;
1002 do_div(len_ns, rate);
1003 hrtimer_start(&tcp_sk(sk)->pacing_timer,
1004 ktime_add_ns(ktime_get(), len_ns),
1005 HRTIMER_MODE_ABS_PINNED_SOFT);
1009 static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb)
1011 skb->skb_mstamp = tp->tcp_mstamp;
1012 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1015 /* This routine actually transmits TCP packets queued in by
1016 * tcp_do_sendmsg(). This is used by both the initial
1017 * transmission and possible later retransmissions.
1018 * All SKB's seen here are completely headerless. It is our
1019 * job to build the TCP header, and pass the packet down to
1020 * IP so it can do the same plus pass the packet off to the
1023 * We are working here with either a clone of the original
1024 * SKB, or a fresh unique copy made by the retransmit engine.
1026 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1029 const struct inet_connection_sock *icsk = inet_csk(sk);
1030 struct inet_sock *inet;
1031 struct tcp_sock *tp;
1032 struct tcp_skb_cb *tcb;
1033 struct tcp_out_options opts;
1034 unsigned int tcp_options_size, tcp_header_size;
1035 struct sk_buff *oskb = NULL;
1036 struct tcp_md5sig_key *md5;
1040 BUG_ON(!skb || !tcp_skb_pcount(skb));
1044 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1048 tcp_skb_tsorted_save(oskb) {
1049 if (unlikely(skb_cloned(oskb)))
1050 skb = pskb_copy(oskb, gfp_mask);
1052 skb = skb_clone(oskb, gfp_mask);
1053 } tcp_skb_tsorted_restore(oskb);
1058 skb->skb_mstamp = tp->tcp_mstamp;
1061 tcb = TCP_SKB_CB(skb);
1062 memset(&opts, 0, sizeof(opts));
1064 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1065 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1067 tcp_options_size = tcp_established_options(sk, skb, &opts,
1069 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1071 /* if no packet is in qdisc/device queue, then allow XPS to select
1072 * another queue. We can be called from tcp_tsq_handler()
1073 * which holds one reference to sk.
1075 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1076 * One way to get this would be to set skb->truesize = 2 on them.
1078 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1080 /* If we had to use memory reserve to allocate this skb,
1081 * this might cause drops if packet is looped back :
1082 * Other socket might not have SOCK_MEMALLOC.
1083 * Packets not looped back do not care about pfmemalloc.
1085 skb->pfmemalloc = 0;
1087 skb_push(skb, tcp_header_size);
1088 skb_reset_transport_header(skb);
1092 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1093 skb_set_hash_from_sk(skb, sk);
1094 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1096 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1098 /* Build TCP header and checksum it. */
1099 th = (struct tcphdr *)skb->data;
1100 th->source = inet->inet_sport;
1101 th->dest = inet->inet_dport;
1102 th->seq = htonl(tcb->seq);
1103 th->ack_seq = htonl(tp->rcv_nxt);
1104 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1110 /* The urg_mode check is necessary during a below snd_una win probe */
1111 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1112 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1113 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1115 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1116 th->urg_ptr = htons(0xFFFF);
1121 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1122 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1123 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1124 th->window = htons(tcp_select_window(sk));
1125 tcp_ecn_send(sk, skb, th, tcp_header_size);
1127 /* RFC1323: The window in SYN & SYN/ACK segments
1130 th->window = htons(min(tp->rcv_wnd, 65535U));
1132 #ifdef CONFIG_TCP_MD5SIG
1133 /* Calculate the MD5 hash, as we have all we need now */
1135 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1136 tp->af_specific->calc_md5_hash(opts.hash_location,
1141 icsk->icsk_af_ops->send_check(sk, skb);
1143 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1144 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1146 if (skb->len != tcp_header_size) {
1147 tcp_event_data_sent(tp, sk);
1148 tp->data_segs_out += tcp_skb_pcount(skb);
1149 tcp_internal_pacing(sk, skb);
1152 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1153 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1154 tcp_skb_pcount(skb));
1156 tp->segs_out += tcp_skb_pcount(skb);
1157 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1158 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1159 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1161 /* Our usage of tstamp should remain private */
1164 /* Cleanup our debris for IP stacks */
1165 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1166 sizeof(struct inet6_skb_parm)));
1168 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1170 if (unlikely(err > 0)) {
1172 err = net_xmit_eval(err);
1175 tcp_update_skb_after_send(tp, oskb);
1176 tcp_rate_skb_sent(sk, oskb);
1181 /* This routine just queues the buffer for sending.
1183 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1184 * otherwise socket can stall.
1186 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1188 struct tcp_sock *tp = tcp_sk(sk);
1190 /* Advance write_seq and place onto the write_queue. */
1191 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1192 __skb_header_release(skb);
1193 tcp_add_write_queue_tail(sk, skb);
1194 sk->sk_wmem_queued += skb->truesize;
1195 sk_mem_charge(sk, skb->truesize);
1198 /* Initialize TSO segments for a packet. */
1199 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1201 if (skb->len <= mss_now) {
1202 /* Avoid the costly divide in the normal
1205 tcp_skb_pcount_set(skb, 1);
1206 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1208 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1209 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1213 /* Pcount in the middle of the write queue got changed, we need to do various
1214 * tweaks to fix counters
1216 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1218 struct tcp_sock *tp = tcp_sk(sk);
1220 tp->packets_out -= decr;
1222 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1223 tp->sacked_out -= decr;
1224 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1225 tp->retrans_out -= decr;
1226 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1227 tp->lost_out -= decr;
1229 /* Reno case is special. Sigh... */
1230 if (tcp_is_reno(tp) && decr > 0)
1231 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1233 if (tp->lost_skb_hint &&
1234 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1235 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1236 tp->lost_cnt_hint -= decr;
1238 tcp_verify_left_out(tp);
1241 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1243 return TCP_SKB_CB(skb)->txstamp_ack ||
1244 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1247 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1249 struct skb_shared_info *shinfo = skb_shinfo(skb);
1251 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1252 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1253 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1254 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1256 shinfo->tx_flags &= ~tsflags;
1257 shinfo2->tx_flags |= tsflags;
1258 swap(shinfo->tskey, shinfo2->tskey);
1259 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1260 TCP_SKB_CB(skb)->txstamp_ack = 0;
1264 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1266 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1267 TCP_SKB_CB(skb)->eor = 0;
1270 /* Insert buff after skb on the write or rtx queue of sk. */
1271 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1272 struct sk_buff *buff,
1274 enum tcp_queue tcp_queue)
1276 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1277 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1279 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1282 /* Function to create two new TCP segments. Shrinks the given segment
1283 * to the specified size and appends a new segment with the rest of the
1284 * packet to the list. This won't be called frequently, I hope.
1285 * Remember, these are still headerless SKBs at this point.
1287 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1288 struct sk_buff *skb, u32 len,
1289 unsigned int mss_now, gfp_t gfp)
1291 struct tcp_sock *tp = tcp_sk(sk);
1292 struct sk_buff *buff;
1293 int nsize, old_factor;
1297 if (WARN_ON(len > skb->len))
1300 nsize = skb_headlen(skb) - len;
1304 if (skb_unclone(skb, gfp))
1307 /* Get a new skb... force flag on. */
1308 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1310 return -ENOMEM; /* We'll just try again later. */
1312 sk->sk_wmem_queued += buff->truesize;
1313 sk_mem_charge(sk, buff->truesize);
1314 nlen = skb->len - len - nsize;
1315 buff->truesize += nlen;
1316 skb->truesize -= nlen;
1318 /* Correct the sequence numbers. */
1319 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1320 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1321 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1323 /* PSH and FIN should only be set in the second packet. */
1324 flags = TCP_SKB_CB(skb)->tcp_flags;
1325 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1326 TCP_SKB_CB(buff)->tcp_flags = flags;
1327 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1328 tcp_skb_fragment_eor(skb, buff);
1330 skb_split(skb, buff, len);
1332 buff->ip_summed = CHECKSUM_PARTIAL;
1334 buff->tstamp = skb->tstamp;
1335 tcp_fragment_tstamp(skb, buff);
1337 old_factor = tcp_skb_pcount(skb);
1339 /* Fix up tso_factor for both original and new SKB. */
1340 tcp_set_skb_tso_segs(skb, mss_now);
1341 tcp_set_skb_tso_segs(buff, mss_now);
1343 /* Update delivered info for the new segment */
1344 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1346 /* If this packet has been sent out already, we must
1347 * adjust the various packet counters.
1349 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1350 int diff = old_factor - tcp_skb_pcount(skb) -
1351 tcp_skb_pcount(buff);
1354 tcp_adjust_pcount(sk, skb, diff);
1357 /* Link BUFF into the send queue. */
1358 __skb_header_release(buff);
1359 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1360 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1361 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1366 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1367 * data is not copied, but immediately discarded.
1369 static int __pskb_trim_head(struct sk_buff *skb, int len)
1371 struct skb_shared_info *shinfo;
1374 eat = min_t(int, len, skb_headlen(skb));
1376 __skb_pull(skb, eat);
1383 shinfo = skb_shinfo(skb);
1384 for (i = 0; i < shinfo->nr_frags; i++) {
1385 int size = skb_frag_size(&shinfo->frags[i]);
1388 skb_frag_unref(skb, i);
1391 shinfo->frags[k] = shinfo->frags[i];
1393 shinfo->frags[k].page_offset += eat;
1394 skb_frag_size_sub(&shinfo->frags[k], eat);
1400 shinfo->nr_frags = k;
1402 skb->data_len -= len;
1403 skb->len = skb->data_len;
1407 /* Remove acked data from a packet in the transmit queue. */
1408 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1412 if (skb_unclone(skb, GFP_ATOMIC))
1415 delta_truesize = __pskb_trim_head(skb, len);
1417 TCP_SKB_CB(skb)->seq += len;
1418 skb->ip_summed = CHECKSUM_PARTIAL;
1420 if (delta_truesize) {
1421 skb->truesize -= delta_truesize;
1422 sk->sk_wmem_queued -= delta_truesize;
1423 sk_mem_uncharge(sk, delta_truesize);
1424 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1427 /* Any change of skb->len requires recalculation of tso factor. */
1428 if (tcp_skb_pcount(skb) > 1)
1429 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1434 /* Calculate MSS not accounting any TCP options. */
1435 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1437 const struct tcp_sock *tp = tcp_sk(sk);
1438 const struct inet_connection_sock *icsk = inet_csk(sk);
1441 /* Calculate base mss without TCP options:
1442 It is MMS_S - sizeof(tcphdr) of rfc1122
1444 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1446 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1447 if (icsk->icsk_af_ops->net_frag_header_len) {
1448 const struct dst_entry *dst = __sk_dst_get(sk);
1450 if (dst && dst_allfrag(dst))
1451 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1454 /* Clamp it (mss_clamp does not include tcp options) */
1455 if (mss_now > tp->rx_opt.mss_clamp)
1456 mss_now = tp->rx_opt.mss_clamp;
1458 /* Now subtract optional transport overhead */
1459 mss_now -= icsk->icsk_ext_hdr_len;
1461 /* Then reserve room for full set of TCP options and 8 bytes of data */
1467 /* Calculate MSS. Not accounting for SACKs here. */
1468 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1470 /* Subtract TCP options size, not including SACKs */
1471 return __tcp_mtu_to_mss(sk, pmtu) -
1472 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1475 /* Inverse of above */
1476 int tcp_mss_to_mtu(struct sock *sk, int mss)
1478 const struct tcp_sock *tp = tcp_sk(sk);
1479 const struct inet_connection_sock *icsk = inet_csk(sk);
1483 tp->tcp_header_len +
1484 icsk->icsk_ext_hdr_len +
1485 icsk->icsk_af_ops->net_header_len;
1487 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1488 if (icsk->icsk_af_ops->net_frag_header_len) {
1489 const struct dst_entry *dst = __sk_dst_get(sk);
1491 if (dst && dst_allfrag(dst))
1492 mtu += icsk->icsk_af_ops->net_frag_header_len;
1496 EXPORT_SYMBOL(tcp_mss_to_mtu);
1498 /* MTU probing init per socket */
1499 void tcp_mtup_init(struct sock *sk)
1501 struct tcp_sock *tp = tcp_sk(sk);
1502 struct inet_connection_sock *icsk = inet_csk(sk);
1503 struct net *net = sock_net(sk);
1505 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1506 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1507 icsk->icsk_af_ops->net_header_len;
1508 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1509 icsk->icsk_mtup.probe_size = 0;
1510 if (icsk->icsk_mtup.enabled)
1511 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1513 EXPORT_SYMBOL(tcp_mtup_init);
1515 /* This function synchronize snd mss to current pmtu/exthdr set.
1517 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1518 for TCP options, but includes only bare TCP header.
1520 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1521 It is minimum of user_mss and mss received with SYN.
1522 It also does not include TCP options.
1524 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1526 tp->mss_cache is current effective sending mss, including
1527 all tcp options except for SACKs. It is evaluated,
1528 taking into account current pmtu, but never exceeds
1529 tp->rx_opt.mss_clamp.
1531 NOTE1. rfc1122 clearly states that advertised MSS
1532 DOES NOT include either tcp or ip options.
1534 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1535 are READ ONLY outside this function. --ANK (980731)
1537 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1539 struct tcp_sock *tp = tcp_sk(sk);
1540 struct inet_connection_sock *icsk = inet_csk(sk);
1543 if (icsk->icsk_mtup.search_high > pmtu)
1544 icsk->icsk_mtup.search_high = pmtu;
1546 mss_now = tcp_mtu_to_mss(sk, pmtu);
1547 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1549 /* And store cached results */
1550 icsk->icsk_pmtu_cookie = pmtu;
1551 if (icsk->icsk_mtup.enabled)
1552 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1553 tp->mss_cache = mss_now;
1557 EXPORT_SYMBOL(tcp_sync_mss);
1559 /* Compute the current effective MSS, taking SACKs and IP options,
1560 * and even PMTU discovery events into account.
1562 unsigned int tcp_current_mss(struct sock *sk)
1564 const struct tcp_sock *tp = tcp_sk(sk);
1565 const struct dst_entry *dst = __sk_dst_get(sk);
1567 unsigned int header_len;
1568 struct tcp_out_options opts;
1569 struct tcp_md5sig_key *md5;
1571 mss_now = tp->mss_cache;
1574 u32 mtu = dst_mtu(dst);
1575 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1576 mss_now = tcp_sync_mss(sk, mtu);
1579 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1580 sizeof(struct tcphdr);
1581 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1582 * some common options. If this is an odd packet (because we have SACK
1583 * blocks etc) then our calculated header_len will be different, and
1584 * we have to adjust mss_now correspondingly */
1585 if (header_len != tp->tcp_header_len) {
1586 int delta = (int) header_len - tp->tcp_header_len;
1593 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1594 * As additional protections, we do not touch cwnd in retransmission phases,
1595 * and if application hit its sndbuf limit recently.
1597 static void tcp_cwnd_application_limited(struct sock *sk)
1599 struct tcp_sock *tp = tcp_sk(sk);
1601 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1602 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1603 /* Limited by application or receiver window. */
1604 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1605 u32 win_used = max(tp->snd_cwnd_used, init_win);
1606 if (win_used < tp->snd_cwnd) {
1607 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1608 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1610 tp->snd_cwnd_used = 0;
1612 tp->snd_cwnd_stamp = tcp_jiffies32;
1615 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1617 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1618 struct tcp_sock *tp = tcp_sk(sk);
1620 /* Track the maximum number of outstanding packets in each
1621 * window, and remember whether we were cwnd-limited then.
1623 if (!before(tp->snd_una, tp->max_packets_seq) ||
1624 tp->packets_out > tp->max_packets_out) {
1625 tp->max_packets_out = tp->packets_out;
1626 tp->max_packets_seq = tp->snd_nxt;
1627 tp->is_cwnd_limited = is_cwnd_limited;
1630 if (tcp_is_cwnd_limited(sk)) {
1631 /* Network is feed fully. */
1632 tp->snd_cwnd_used = 0;
1633 tp->snd_cwnd_stamp = tcp_jiffies32;
1635 /* Network starves. */
1636 if (tp->packets_out > tp->snd_cwnd_used)
1637 tp->snd_cwnd_used = tp->packets_out;
1639 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1640 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1641 !ca_ops->cong_control)
1642 tcp_cwnd_application_limited(sk);
1644 /* The following conditions together indicate the starvation
1645 * is caused by insufficient sender buffer:
1646 * 1) just sent some data (see tcp_write_xmit)
1647 * 2) not cwnd limited (this else condition)
1648 * 3) no more data to send (tcp_write_queue_empty())
1649 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1651 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1652 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1653 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1654 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1658 /* Minshall's variant of the Nagle send check. */
1659 static bool tcp_minshall_check(const struct tcp_sock *tp)
1661 return after(tp->snd_sml, tp->snd_una) &&
1662 !after(tp->snd_sml, tp->snd_nxt);
1665 /* Update snd_sml if this skb is under mss
1666 * Note that a TSO packet might end with a sub-mss segment
1667 * The test is really :
1668 * if ((skb->len % mss) != 0)
1669 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1670 * But we can avoid doing the divide again given we already have
1671 * skb_pcount = skb->len / mss_now
1673 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1674 const struct sk_buff *skb)
1676 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1677 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1680 /* Return false, if packet can be sent now without violation Nagle's rules:
1681 * 1. It is full sized. (provided by caller in %partial bool)
1682 * 2. Or it contains FIN. (already checked by caller)
1683 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1684 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1685 * With Minshall's modification: all sent small packets are ACKed.
1687 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1691 ((nonagle & TCP_NAGLE_CORK) ||
1692 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1695 /* Return how many segs we'd like on a TSO packet,
1696 * to send one TSO packet per ms
1698 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1703 bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift,
1704 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1706 /* Goal is to send at least one packet per ms,
1707 * not one big TSO packet every 100 ms.
1708 * This preserves ACK clocking and is consistent
1709 * with tcp_tso_should_defer() heuristic.
1711 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1716 /* Return the number of segments we want in the skb we are transmitting.
1717 * See if congestion control module wants to decide; otherwise, autosize.
1719 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1721 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1722 u32 min_tso, tso_segs;
1724 min_tso = ca_ops->min_tso_segs ?
1725 ca_ops->min_tso_segs(sk) :
1726 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1728 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1729 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1732 /* Returns the portion of skb which can be sent right away */
1733 static unsigned int tcp_mss_split_point(const struct sock *sk,
1734 const struct sk_buff *skb,
1735 unsigned int mss_now,
1736 unsigned int max_segs,
1739 const struct tcp_sock *tp = tcp_sk(sk);
1740 u32 partial, needed, window, max_len;
1742 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1743 max_len = mss_now * max_segs;
1745 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1748 needed = min(skb->len, window);
1750 if (max_len <= needed)
1753 partial = needed % mss_now;
1754 /* If last segment is not a full MSS, check if Nagle rules allow us
1755 * to include this last segment in this skb.
1756 * Otherwise, we'll split the skb at last MSS boundary
1758 if (tcp_nagle_check(partial != 0, tp, nonagle))
1759 return needed - partial;
1764 /* Can at least one segment of SKB be sent right now, according to the
1765 * congestion window rules? If so, return how many segments are allowed.
1767 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1768 const struct sk_buff *skb)
1770 u32 in_flight, cwnd, halfcwnd;
1772 /* Don't be strict about the congestion window for the final FIN. */
1773 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1774 tcp_skb_pcount(skb) == 1)
1777 in_flight = tcp_packets_in_flight(tp);
1778 cwnd = tp->snd_cwnd;
1779 if (in_flight >= cwnd)
1782 /* For better scheduling, ensure we have at least
1783 * 2 GSO packets in flight.
1785 halfcwnd = max(cwnd >> 1, 1U);
1786 return min(halfcwnd, cwnd - in_flight);
1789 /* Initialize TSO state of a skb.
1790 * This must be invoked the first time we consider transmitting
1791 * SKB onto the wire.
1793 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1795 int tso_segs = tcp_skb_pcount(skb);
1797 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1798 tcp_set_skb_tso_segs(skb, mss_now);
1799 tso_segs = tcp_skb_pcount(skb);
1805 /* Return true if the Nagle test allows this packet to be
1808 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1809 unsigned int cur_mss, int nonagle)
1811 /* Nagle rule does not apply to frames, which sit in the middle of the
1812 * write_queue (they have no chances to get new data).
1814 * This is implemented in the callers, where they modify the 'nonagle'
1815 * argument based upon the location of SKB in the send queue.
1817 if (nonagle & TCP_NAGLE_PUSH)
1820 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1821 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1824 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1830 /* Does at least the first segment of SKB fit into the send window? */
1831 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1832 const struct sk_buff *skb,
1833 unsigned int cur_mss)
1835 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1837 if (skb->len > cur_mss)
1838 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1840 return !after(end_seq, tcp_wnd_end(tp));
1843 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1844 * which is put after SKB on the list. It is very much like
1845 * tcp_fragment() except that it may make several kinds of assumptions
1846 * in order to speed up the splitting operation. In particular, we
1847 * know that all the data is in scatter-gather pages, and that the
1848 * packet has never been sent out before (and thus is not cloned).
1850 static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1851 struct sk_buff *skb, unsigned int len,
1852 unsigned int mss_now, gfp_t gfp)
1854 struct sk_buff *buff;
1855 int nlen = skb->len - len;
1858 /* All of a TSO frame must be composed of paged data. */
1859 if (skb->len != skb->data_len)
1860 return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp);
1862 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1863 if (unlikely(!buff))
1866 sk->sk_wmem_queued += buff->truesize;
1867 sk_mem_charge(sk, buff->truesize);
1868 buff->truesize += nlen;
1869 skb->truesize -= nlen;
1871 /* Correct the sequence numbers. */
1872 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1873 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1874 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1876 /* PSH and FIN should only be set in the second packet. */
1877 flags = TCP_SKB_CB(skb)->tcp_flags;
1878 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1879 TCP_SKB_CB(buff)->tcp_flags = flags;
1881 /* This packet was never sent out yet, so no SACK bits. */
1882 TCP_SKB_CB(buff)->sacked = 0;
1884 tcp_skb_fragment_eor(skb, buff);
1886 buff->ip_summed = CHECKSUM_PARTIAL;
1887 skb_split(skb, buff, len);
1888 tcp_fragment_tstamp(skb, buff);
1890 /* Fix up tso_factor for both original and new SKB. */
1891 tcp_set_skb_tso_segs(skb, mss_now);
1892 tcp_set_skb_tso_segs(buff, mss_now);
1894 /* Link BUFF into the send queue. */
1895 __skb_header_release(buff);
1896 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1901 /* Try to defer sending, if possible, in order to minimize the amount
1902 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1904 * This algorithm is from John Heffner.
1906 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1907 bool *is_cwnd_limited, u32 max_segs)
1909 const struct inet_connection_sock *icsk = inet_csk(sk);
1910 u32 age, send_win, cong_win, limit, in_flight;
1911 struct tcp_sock *tp = tcp_sk(sk);
1912 struct sk_buff *head;
1915 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1918 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1921 /* Avoid bursty behavior by allowing defer
1922 * only if the last write was recent.
1924 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
1927 in_flight = tcp_packets_in_flight(tp);
1929 BUG_ON(tcp_skb_pcount(skb) <= 1);
1930 BUG_ON(tp->snd_cwnd <= in_flight);
1932 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1934 /* From in_flight test above, we know that cwnd > in_flight. */
1935 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1937 limit = min(send_win, cong_win);
1939 /* If a full-sized TSO skb can be sent, do it. */
1940 if (limit >= max_segs * tp->mss_cache)
1943 /* Middle in queue won't get any more data, full sendable already? */
1944 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1947 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1949 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1951 /* If at least some fraction of a window is available,
1954 chunk /= win_divisor;
1958 /* Different approach, try not to defer past a single
1959 * ACK. Receiver should ACK every other full sized
1960 * frame, so if we have space for more than 3 frames
1963 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1967 /* TODO : use tsorted_sent_queue ? */
1968 head = tcp_rtx_queue_head(sk);
1971 age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
1972 /* If next ACK is likely to come too late (half srtt), do not defer */
1973 if (age < (tp->srtt_us >> 4))
1976 /* Ok, it looks like it is advisable to defer. */
1978 if (cong_win < send_win && cong_win <= skb->len)
1979 *is_cwnd_limited = true;
1987 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1989 struct inet_connection_sock *icsk = inet_csk(sk);
1990 struct tcp_sock *tp = tcp_sk(sk);
1991 struct net *net = sock_net(sk);
1995 interval = net->ipv4.sysctl_tcp_probe_interval;
1996 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
1997 if (unlikely(delta >= interval * HZ)) {
1998 int mss = tcp_current_mss(sk);
2000 /* Update current search range */
2001 icsk->icsk_mtup.probe_size = 0;
2002 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2003 sizeof(struct tcphdr) +
2004 icsk->icsk_af_ops->net_header_len;
2005 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2007 /* Update probe time stamp */
2008 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2012 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2014 struct sk_buff *skb, *next;
2016 skb = tcp_send_head(sk);
2017 tcp_for_write_queue_from_safe(skb, next, sk) {
2018 if (len <= skb->len)
2021 if (unlikely(TCP_SKB_CB(skb)->eor))
2030 /* Create a new MTU probe if we are ready.
2031 * MTU probe is regularly attempting to increase the path MTU by
2032 * deliberately sending larger packets. This discovers routing
2033 * changes resulting in larger path MTUs.
2035 * Returns 0 if we should wait to probe (no cwnd available),
2036 * 1 if a probe was sent,
2039 static int tcp_mtu_probe(struct sock *sk)
2041 struct inet_connection_sock *icsk = inet_csk(sk);
2042 struct tcp_sock *tp = tcp_sk(sk);
2043 struct sk_buff *skb, *nskb, *next;
2044 struct net *net = sock_net(sk);
2051 /* Not currently probing/verifying,
2053 * have enough cwnd, and
2054 * not SACKing (the variable headers throw things off)
2056 if (likely(!icsk->icsk_mtup.enabled ||
2057 icsk->icsk_mtup.probe_size ||
2058 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2059 tp->snd_cwnd < 11 ||
2060 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2063 /* Use binary search for probe_size between tcp_mss_base,
2064 * and current mss_clamp. if (search_high - search_low)
2065 * smaller than a threshold, backoff from probing.
2067 mss_now = tcp_current_mss(sk);
2068 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2069 icsk->icsk_mtup.search_low) >> 1);
2070 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2071 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2072 /* When misfortune happens, we are reprobing actively,
2073 * and then reprobe timer has expired. We stick with current
2074 * probing process by not resetting search range to its orignal.
2076 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2077 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2078 /* Check whether enough time has elaplased for
2079 * another round of probing.
2081 tcp_mtu_check_reprobe(sk);
2085 /* Have enough data in the send queue to probe? */
2086 if (tp->write_seq - tp->snd_nxt < size_needed)
2089 if (tp->snd_wnd < size_needed)
2091 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2094 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2095 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2096 if (!tcp_packets_in_flight(tp))
2102 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2105 /* We're allowed to probe. Build it now. */
2106 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2109 sk->sk_wmem_queued += nskb->truesize;
2110 sk_mem_charge(sk, nskb->truesize);
2112 skb = tcp_send_head(sk);
2114 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2115 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2116 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2117 TCP_SKB_CB(nskb)->sacked = 0;
2119 nskb->ip_summed = CHECKSUM_PARTIAL;
2121 tcp_insert_write_queue_before(nskb, skb, sk);
2122 tcp_highest_sack_replace(sk, skb, nskb);
2125 tcp_for_write_queue_from_safe(skb, next, sk) {
2126 copy = min_t(int, skb->len, probe_size - len);
2127 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2129 if (skb->len <= copy) {
2130 /* We've eaten all the data from this skb.
2132 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2133 /* If this is the last SKB we copy and eor is set
2134 * we need to propagate it to the new skb.
2136 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2137 tcp_unlink_write_queue(skb, sk);
2138 sk_wmem_free_skb(sk, skb);
2140 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2141 ~(TCPHDR_FIN|TCPHDR_PSH);
2142 if (!skb_shinfo(skb)->nr_frags) {
2143 skb_pull(skb, copy);
2145 __pskb_trim_head(skb, copy);
2146 tcp_set_skb_tso_segs(skb, mss_now);
2148 TCP_SKB_CB(skb)->seq += copy;
2153 if (len >= probe_size)
2156 tcp_init_tso_segs(nskb, nskb->len);
2158 /* We're ready to send. If this fails, the probe will
2159 * be resegmented into mss-sized pieces by tcp_write_xmit().
2161 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2162 /* Decrement cwnd here because we are sending
2163 * effectively two packets. */
2165 tcp_event_new_data_sent(sk, nskb);
2167 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2168 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2169 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2177 static bool tcp_pacing_check(const struct sock *sk)
2179 return tcp_needs_internal_pacing(sk) &&
2180 hrtimer_is_queued(&tcp_sk(sk)->pacing_timer);
2183 /* TCP Small Queues :
2184 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2185 * (These limits are doubled for retransmits)
2187 * - better RTT estimation and ACK scheduling
2190 * Alas, some drivers / subsystems require a fair amount
2191 * of queued bytes to ensure line rate.
2192 * One example is wifi aggregation (802.11 AMPDU)
2194 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2195 unsigned int factor)
2199 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift);
2200 limit = min_t(u32, limit,
2201 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2204 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2205 /* Always send skb if rtx queue is empty.
2206 * No need to wait for TX completion to call us back,
2207 * after softirq/tasklet schedule.
2208 * This helps when TX completions are delayed too much.
2210 if (tcp_rtx_queue_empty(sk))
2213 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2214 /* It is possible TX completion already happened
2215 * before we set TSQ_THROTTLED, so we must
2216 * test again the condition.
2218 smp_mb__after_atomic();
2219 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2225 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2227 const u32 now = tcp_jiffies32;
2228 enum tcp_chrono old = tp->chrono_type;
2230 if (old > TCP_CHRONO_UNSPEC)
2231 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2232 tp->chrono_start = now;
2233 tp->chrono_type = new;
2236 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2238 struct tcp_sock *tp = tcp_sk(sk);
2240 /* If there are multiple conditions worthy of tracking in a
2241 * chronograph then the highest priority enum takes precedence
2242 * over the other conditions. So that if something "more interesting"
2243 * starts happening, stop the previous chrono and start a new one.
2245 if (type > tp->chrono_type)
2246 tcp_chrono_set(tp, type);
2249 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2251 struct tcp_sock *tp = tcp_sk(sk);
2254 /* There are multiple conditions worthy of tracking in a
2255 * chronograph, so that the highest priority enum takes
2256 * precedence over the other conditions (see tcp_chrono_start).
2257 * If a condition stops, we only stop chrono tracking if
2258 * it's the "most interesting" or current chrono we are
2259 * tracking and starts busy chrono if we have pending data.
2261 if (tcp_rtx_and_write_queues_empty(sk))
2262 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2263 else if (type == tp->chrono_type)
2264 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2267 /* This routine writes packets to the network. It advances the
2268 * send_head. This happens as incoming acks open up the remote
2271 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2272 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2273 * account rare use of URG, this is not a big flaw.
2275 * Send at most one packet when push_one > 0. Temporarily ignore
2276 * cwnd limit to force at most one packet out when push_one == 2.
2278 * Returns true, if no segments are in flight and we have queued segments,
2279 * but cannot send anything now because of SWS or another problem.
2281 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2282 int push_one, gfp_t gfp)
2284 struct tcp_sock *tp = tcp_sk(sk);
2285 struct sk_buff *skb;
2286 unsigned int tso_segs, sent_pkts;
2289 bool is_cwnd_limited = false, is_rwnd_limited = false;
2294 tcp_mstamp_refresh(tp);
2296 /* Do MTU probing. */
2297 result = tcp_mtu_probe(sk);
2300 } else if (result > 0) {
2305 max_segs = tcp_tso_segs(sk, mss_now);
2306 while ((skb = tcp_send_head(sk))) {
2309 if (tcp_pacing_check(sk))
2312 tso_segs = tcp_init_tso_segs(skb, mss_now);
2315 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2316 /* "skb_mstamp" is used as a start point for the retransmit timer */
2317 tcp_update_skb_after_send(tp, skb);
2318 goto repair; /* Skip network transmission */
2321 cwnd_quota = tcp_cwnd_test(tp, skb);
2324 /* Force out a loss probe pkt. */
2330 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2331 is_rwnd_limited = true;
2335 if (tso_segs == 1) {
2336 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2337 (tcp_skb_is_last(sk, skb) ?
2338 nonagle : TCP_NAGLE_PUSH))))
2342 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2348 if (tso_segs > 1 && !tcp_urg_mode(tp))
2349 limit = tcp_mss_split_point(sk, skb, mss_now,
2355 if (skb->len > limit &&
2356 unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2357 skb, limit, mss_now, gfp)))
2360 if (tcp_small_queue_check(sk, skb, 0))
2363 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2367 /* Advance the send_head. This one is sent out.
2368 * This call will increment packets_out.
2370 tcp_event_new_data_sent(sk, skb);
2372 tcp_minshall_update(tp, mss_now, skb);
2373 sent_pkts += tcp_skb_pcount(skb);
2379 if (is_rwnd_limited)
2380 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2382 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2384 if (likely(sent_pkts)) {
2385 if (tcp_in_cwnd_reduction(sk))
2386 tp->prr_out += sent_pkts;
2388 /* Send one loss probe per tail loss episode. */
2390 tcp_schedule_loss_probe(sk, false);
2391 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2392 tcp_cwnd_validate(sk, is_cwnd_limited);
2395 return !tp->packets_out && !tcp_write_queue_empty(sk);
2398 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2400 struct inet_connection_sock *icsk = inet_csk(sk);
2401 struct tcp_sock *tp = tcp_sk(sk);
2402 u32 timeout, rto_delta_us;
2405 /* Don't do any loss probe on a Fast Open connection before 3WHS
2408 if (tp->fastopen_rsk)
2411 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2412 /* Schedule a loss probe in 2*RTT for SACK capable connections
2413 * not in loss recovery, that are either limited by cwnd or application.
2415 if ((early_retrans != 3 && early_retrans != 4) ||
2416 !tp->packets_out || !tcp_is_sack(tp) ||
2417 (icsk->icsk_ca_state != TCP_CA_Open &&
2418 icsk->icsk_ca_state != TCP_CA_CWR))
2421 /* Probe timeout is 2*rtt. Add minimum RTO to account
2422 * for delayed ack when there's one outstanding packet. If no RTT
2423 * sample is available then probe after TCP_TIMEOUT_INIT.
2426 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2427 if (tp->packets_out == 1)
2428 timeout += TCP_RTO_MIN;
2430 timeout += TCP_TIMEOUT_MIN;
2432 timeout = TCP_TIMEOUT_INIT;
2435 /* If the RTO formula yields an earlier time, then use that time. */
2436 rto_delta_us = advancing_rto ?
2437 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2438 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2439 if (rto_delta_us > 0)
2440 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2442 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2447 /* Thanks to skb fast clones, we can detect if a prior transmit of
2448 * a packet is still in a qdisc or driver queue.
2449 * In this case, there is very little point doing a retransmit !
2451 static bool skb_still_in_host_queue(const struct sock *sk,
2452 const struct sk_buff *skb)
2454 if (unlikely(skb_fclone_busy(sk, skb))) {
2455 NET_INC_STATS(sock_net(sk),
2456 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2462 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2463 * retransmit the last segment.
2465 void tcp_send_loss_probe(struct sock *sk)
2467 struct tcp_sock *tp = tcp_sk(sk);
2468 struct sk_buff *skb;
2470 int mss = tcp_current_mss(sk);
2472 skb = tcp_send_head(sk);
2473 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2474 pcount = tp->packets_out;
2475 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2476 if (tp->packets_out > pcount)
2480 skb = skb_rb_last(&sk->tcp_rtx_queue);
2482 /* At most one outstanding TLP retransmission. */
2483 if (tp->tlp_high_seq)
2486 /* Retransmit last segment. */
2490 if (skb_still_in_host_queue(sk, skb))
2493 pcount = tcp_skb_pcount(skb);
2494 if (WARN_ON(!pcount))
2497 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2498 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2499 (pcount - 1) * mss, mss,
2502 skb = skb_rb_next(skb);
2505 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2508 if (__tcp_retransmit_skb(sk, skb, 1))
2511 /* Record snd_nxt for loss detection. */
2512 tp->tlp_high_seq = tp->snd_nxt;
2515 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2516 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2517 inet_csk(sk)->icsk_pending = 0;
2522 /* Push out any pending frames which were held back due to
2523 * TCP_CORK or attempt at coalescing tiny packets.
2524 * The socket must be locked by the caller.
2526 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2529 /* If we are closed, the bytes will have to remain here.
2530 * In time closedown will finish, we empty the write queue and
2531 * all will be happy.
2533 if (unlikely(sk->sk_state == TCP_CLOSE))
2536 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2537 sk_gfp_mask(sk, GFP_ATOMIC)))
2538 tcp_check_probe_timer(sk);
2541 /* Send _single_ skb sitting at the send head. This function requires
2542 * true push pending frames to setup probe timer etc.
2544 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2546 struct sk_buff *skb = tcp_send_head(sk);
2548 BUG_ON(!skb || skb->len < mss_now);
2550 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2553 /* This function returns the amount that we can raise the
2554 * usable window based on the following constraints
2556 * 1. The window can never be shrunk once it is offered (RFC 793)
2557 * 2. We limit memory per socket
2560 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2561 * RECV.NEXT + RCV.WIN fixed until:
2562 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2564 * i.e. don't raise the right edge of the window until you can raise
2565 * it at least MSS bytes.
2567 * Unfortunately, the recommended algorithm breaks header prediction,
2568 * since header prediction assumes th->window stays fixed.
2570 * Strictly speaking, keeping th->window fixed violates the receiver
2571 * side SWS prevention criteria. The problem is that under this rule
2572 * a stream of single byte packets will cause the right side of the
2573 * window to always advance by a single byte.
2575 * Of course, if the sender implements sender side SWS prevention
2576 * then this will not be a problem.
2578 * BSD seems to make the following compromise:
2580 * If the free space is less than the 1/4 of the maximum
2581 * space available and the free space is less than 1/2 mss,
2582 * then set the window to 0.
2583 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2584 * Otherwise, just prevent the window from shrinking
2585 * and from being larger than the largest representable value.
2587 * This prevents incremental opening of the window in the regime
2588 * where TCP is limited by the speed of the reader side taking
2589 * data out of the TCP receive queue. It does nothing about
2590 * those cases where the window is constrained on the sender side
2591 * because the pipeline is full.
2593 * BSD also seems to "accidentally" limit itself to windows that are a
2594 * multiple of MSS, at least until the free space gets quite small.
2595 * This would appear to be a side effect of the mbuf implementation.
2596 * Combining these two algorithms results in the observed behavior
2597 * of having a fixed window size at almost all times.
2599 * Below we obtain similar behavior by forcing the offered window to
2600 * a multiple of the mss when it is feasible to do so.
2602 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2603 * Regular options like TIMESTAMP are taken into account.
2605 u32 __tcp_select_window(struct sock *sk)
2607 struct inet_connection_sock *icsk = inet_csk(sk);
2608 struct tcp_sock *tp = tcp_sk(sk);
2609 /* MSS for the peer's data. Previous versions used mss_clamp
2610 * here. I don't know if the value based on our guesses
2611 * of peer's MSS is better for the performance. It's more correct
2612 * but may be worse for the performance because of rcv_mss
2613 * fluctuations. --SAW 1998/11/1
2615 int mss = icsk->icsk_ack.rcv_mss;
2616 int free_space = tcp_space(sk);
2617 int allowed_space = tcp_full_space(sk);
2618 int full_space = min_t(int, tp->window_clamp, allowed_space);
2621 if (unlikely(mss > full_space)) {
2626 if (free_space < (full_space >> 1)) {
2627 icsk->icsk_ack.quick = 0;
2629 if (tcp_under_memory_pressure(sk))
2630 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2633 /* free_space might become our new window, make sure we don't
2634 * increase it due to wscale.
2636 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2638 /* if free space is less than mss estimate, or is below 1/16th
2639 * of the maximum allowed, try to move to zero-window, else
2640 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2641 * new incoming data is dropped due to memory limits.
2642 * With large window, mss test triggers way too late in order
2643 * to announce zero window in time before rmem limit kicks in.
2645 if (free_space < (allowed_space >> 4) || free_space < mss)
2649 if (free_space > tp->rcv_ssthresh)
2650 free_space = tp->rcv_ssthresh;
2652 /* Don't do rounding if we are using window scaling, since the
2653 * scaled window will not line up with the MSS boundary anyway.
2655 if (tp->rx_opt.rcv_wscale) {
2656 window = free_space;
2658 /* Advertise enough space so that it won't get scaled away.
2659 * Import case: prevent zero window announcement if
2660 * 1<<rcv_wscale > mss.
2662 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2664 window = tp->rcv_wnd;
2665 /* Get the largest window that is a nice multiple of mss.
2666 * Window clamp already applied above.
2667 * If our current window offering is within 1 mss of the
2668 * free space we just keep it. This prevents the divide
2669 * and multiply from happening most of the time.
2670 * We also don't do any window rounding when the free space
2673 if (window <= free_space - mss || window > free_space)
2674 window = rounddown(free_space, mss);
2675 else if (mss == full_space &&
2676 free_space > window + (full_space >> 1))
2677 window = free_space;
2683 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2684 const struct sk_buff *next_skb)
2686 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2687 const struct skb_shared_info *next_shinfo =
2688 skb_shinfo(next_skb);
2689 struct skb_shared_info *shinfo = skb_shinfo(skb);
2691 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2692 shinfo->tskey = next_shinfo->tskey;
2693 TCP_SKB_CB(skb)->txstamp_ack |=
2694 TCP_SKB_CB(next_skb)->txstamp_ack;
2698 /* Collapses two adjacent SKB's during retransmission. */
2699 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2701 struct tcp_sock *tp = tcp_sk(sk);
2702 struct sk_buff *next_skb = skb_rb_next(skb);
2703 int skb_size, next_skb_size;
2705 skb_size = skb->len;
2706 next_skb_size = next_skb->len;
2708 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2710 if (next_skb_size) {
2711 if (next_skb_size <= skb_availroom(skb))
2712 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2714 else if (!skb_shift(skb, next_skb, next_skb_size))
2717 tcp_highest_sack_replace(sk, next_skb, skb);
2719 /* Update sequence range on original skb. */
2720 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2722 /* Merge over control information. This moves PSH/FIN etc. over */
2723 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2725 /* All done, get rid of second SKB and account for it so
2726 * packet counting does not break.
2728 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2729 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2731 /* changed transmit queue under us so clear hints */
2732 tcp_clear_retrans_hints_partial(tp);
2733 if (next_skb == tp->retransmit_skb_hint)
2734 tp->retransmit_skb_hint = skb;
2736 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2738 tcp_skb_collapse_tstamp(skb, next_skb);
2740 tcp_rtx_queue_unlink_and_free(next_skb, sk);
2744 /* Check if coalescing SKBs is legal. */
2745 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2747 if (tcp_skb_pcount(skb) > 1)
2749 if (skb_cloned(skb))
2751 /* Some heuristics for collapsing over SACK'd could be invented */
2752 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2758 /* Collapse packets in the retransmit queue to make to create
2759 * less packets on the wire. This is only done on retransmission.
2761 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2764 struct tcp_sock *tp = tcp_sk(sk);
2765 struct sk_buff *skb = to, *tmp;
2768 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2770 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2773 skb_rbtree_walk_from_safe(skb, tmp) {
2774 if (!tcp_can_collapse(sk, skb))
2777 if (!tcp_skb_can_collapse_to(to))
2790 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2793 if (!tcp_collapse_retrans(sk, to))
2798 /* This retransmits one SKB. Policy decisions and retransmit queue
2799 * state updates are done by the caller. Returns non-zero if an
2800 * error occurred which prevented the send.
2802 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2804 struct inet_connection_sock *icsk = inet_csk(sk);
2805 struct tcp_sock *tp = tcp_sk(sk);
2806 unsigned int cur_mss;
2810 /* Inconclusive MTU probe */
2811 if (icsk->icsk_mtup.probe_size)
2812 icsk->icsk_mtup.probe_size = 0;
2814 /* Do not sent more than we queued. 1/4 is reserved for possible
2815 * copying overhead: fragmentation, tunneling, mangling etc.
2817 if (refcount_read(&sk->sk_wmem_alloc) >
2818 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2822 if (skb_still_in_host_queue(sk, skb))
2825 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2826 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2830 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2834 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2835 return -EHOSTUNREACH; /* Routing failure or similar. */
2837 cur_mss = tcp_current_mss(sk);
2839 /* If receiver has shrunk his window, and skb is out of
2840 * new window, do not retransmit it. The exception is the
2841 * case, when window is shrunk to zero. In this case
2842 * our retransmit serves as a zero window probe.
2844 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2845 TCP_SKB_CB(skb)->seq != tp->snd_una)
2848 len = cur_mss * segs;
2849 if (skb->len > len) {
2850 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2851 cur_mss, GFP_ATOMIC))
2852 return -ENOMEM; /* We'll try again later. */
2854 if (skb_unclone(skb, GFP_ATOMIC))
2857 diff = tcp_skb_pcount(skb);
2858 tcp_set_skb_tso_segs(skb, cur_mss);
2859 diff -= tcp_skb_pcount(skb);
2861 tcp_adjust_pcount(sk, skb, diff);
2862 if (skb->len < cur_mss)
2863 tcp_retrans_try_collapse(sk, skb, cur_mss);
2866 /* RFC3168, section 6.1.1.1. ECN fallback */
2867 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2868 tcp_ecn_clear_syn(sk, skb);
2870 /* Update global and local TCP statistics. */
2871 segs = tcp_skb_pcount(skb);
2872 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2873 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2874 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2875 tp->total_retrans += segs;
2877 /* make sure skb->data is aligned on arches that require it
2878 * and check if ack-trimming & collapsing extended the headroom
2879 * beyond what csum_start can cover.
2881 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2882 skb_headroom(skb) >= 0xFFFF)) {
2883 struct sk_buff *nskb;
2885 tcp_skb_tsorted_save(skb) {
2886 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2887 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2889 } tcp_skb_tsorted_restore(skb);
2892 tcp_update_skb_after_send(tp, skb);
2893 tcp_rate_skb_sent(sk, skb);
2896 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2899 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
2900 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
2901 TCP_SKB_CB(skb)->seq, segs, err);
2904 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2905 trace_tcp_retransmit_skb(sk, skb);
2906 } else if (err != -EBUSY) {
2907 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2912 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2914 struct tcp_sock *tp = tcp_sk(sk);
2915 int err = __tcp_retransmit_skb(sk, skb, segs);
2918 #if FASTRETRANS_DEBUG > 0
2919 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2920 net_dbg_ratelimited("retrans_out leaked\n");
2923 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2924 tp->retrans_out += tcp_skb_pcount(skb);
2926 /* Save stamp of the first retransmit. */
2927 if (!tp->retrans_stamp)
2928 tp->retrans_stamp = tcp_skb_timestamp(skb);
2932 if (tp->undo_retrans < 0)
2933 tp->undo_retrans = 0;
2934 tp->undo_retrans += tcp_skb_pcount(skb);
2938 /* This gets called after a retransmit timeout, and the initially
2939 * retransmitted data is acknowledged. It tries to continue
2940 * resending the rest of the retransmit queue, until either
2941 * we've sent it all or the congestion window limit is reached.
2943 void tcp_xmit_retransmit_queue(struct sock *sk)
2945 const struct inet_connection_sock *icsk = inet_csk(sk);
2946 struct sk_buff *skb, *rtx_head, *hole = NULL;
2947 struct tcp_sock *tp = tcp_sk(sk);
2951 if (!tp->packets_out)
2954 rtx_head = tcp_rtx_queue_head(sk);
2955 skb = tp->retransmit_skb_hint ?: rtx_head;
2956 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2957 skb_rbtree_walk_from(skb) {
2961 if (tcp_pacing_check(sk))
2964 /* we could do better than to assign each time */
2966 tp->retransmit_skb_hint = skb;
2968 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2971 sacked = TCP_SKB_CB(skb)->sacked;
2972 /* In case tcp_shift_skb_data() have aggregated large skbs,
2973 * we need to make sure not sending too bigs TSO packets
2975 segs = min_t(int, segs, max_segs);
2977 if (tp->retrans_out >= tp->lost_out) {
2979 } else if (!(sacked & TCPCB_LOST)) {
2980 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2985 if (icsk->icsk_ca_state != TCP_CA_Loss)
2986 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2988 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2991 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2994 if (tcp_small_queue_check(sk, skb, 1))
2997 if (tcp_retransmit_skb(sk, skb, segs))
3000 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3002 if (tcp_in_cwnd_reduction(sk))
3003 tp->prr_out += tcp_skb_pcount(skb);
3005 if (skb == rtx_head &&
3006 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3007 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3008 inet_csk(sk)->icsk_rto,
3013 /* We allow to exceed memory limits for FIN packets to expedite
3014 * connection tear down and (memory) recovery.
3015 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3016 * or even be forced to close flow without any FIN.
3017 * In general, we want to allow one skb per socket to avoid hangs
3018 * with edge trigger epoll()
3020 void sk_forced_mem_schedule(struct sock *sk, int size)
3024 if (size <= sk->sk_forward_alloc)
3026 amt = sk_mem_pages(size);
3027 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3028 sk_memory_allocated_add(sk, amt);
3030 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3031 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3034 /* Send a FIN. The caller locks the socket for us.
3035 * We should try to send a FIN packet really hard, but eventually give up.
3037 void tcp_send_fin(struct sock *sk)
3039 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3040 struct tcp_sock *tp = tcp_sk(sk);
3042 /* Optimization, tack on the FIN if we have one skb in write queue and
3043 * this skb was not yet sent, or we are under memory pressure.
3044 * Note: in the latter case, FIN packet will be sent after a timeout,
3045 * as TCP stack thinks it has already been transmitted.
3047 if (!tskb && tcp_under_memory_pressure(sk))
3048 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3052 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3053 TCP_SKB_CB(tskb)->end_seq++;
3055 if (tcp_write_queue_empty(sk)) {
3056 /* This means tskb was already sent.
3057 * Pretend we included the FIN on previous transmit.
3058 * We need to set tp->snd_nxt to the value it would have
3059 * if FIN had been sent. This is because retransmit path
3060 * does not change tp->snd_nxt.
3066 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3067 if (unlikely(!skb)) {
3072 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3073 skb_reserve(skb, MAX_TCP_HEADER);
3074 sk_forced_mem_schedule(sk, skb->truesize);
3075 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3076 tcp_init_nondata_skb(skb, tp->write_seq,
3077 TCPHDR_ACK | TCPHDR_FIN);
3078 tcp_queue_skb(sk, skb);
3080 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3083 /* We get here when a process closes a file descriptor (either due to
3084 * an explicit close() or as a byproduct of exit()'ing) and there
3085 * was unread data in the receive queue. This behavior is recommended
3086 * by RFC 2525, section 2.17. -DaveM
3088 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3090 struct sk_buff *skb;
3092 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3094 /* NOTE: No TCP options attached and we never retransmit this. */
3095 skb = alloc_skb(MAX_TCP_HEADER, priority);
3097 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3101 /* Reserve space for headers and prepare control bits. */
3102 skb_reserve(skb, MAX_TCP_HEADER);
3103 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3104 TCPHDR_ACK | TCPHDR_RST);
3105 tcp_mstamp_refresh(tcp_sk(sk));
3107 if (tcp_transmit_skb(sk, skb, 0, priority))
3108 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3110 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3111 * skb here is different to the troublesome skb, so use NULL
3113 trace_tcp_send_reset(sk, NULL);
3116 /* Send a crossed SYN-ACK during socket establishment.
3117 * WARNING: This routine must only be called when we have already sent
3118 * a SYN packet that crossed the incoming SYN that caused this routine
3119 * to get called. If this assumption fails then the initial rcv_wnd
3120 * and rcv_wscale values will not be correct.
3122 int tcp_send_synack(struct sock *sk)
3124 struct sk_buff *skb;
3126 skb = tcp_rtx_queue_head(sk);
3127 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3128 pr_err("%s: wrong queue state\n", __func__);
3131 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3132 if (skb_cloned(skb)) {
3133 struct sk_buff *nskb;
3135 tcp_skb_tsorted_save(skb) {
3136 nskb = skb_copy(skb, GFP_ATOMIC);
3137 } tcp_skb_tsorted_restore(skb);
3140 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3141 tcp_rtx_queue_unlink_and_free(skb, sk);
3142 __skb_header_release(nskb);
3143 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3144 sk->sk_wmem_queued += nskb->truesize;
3145 sk_mem_charge(sk, nskb->truesize);
3149 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3150 tcp_ecn_send_synack(sk, skb);
3152 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3156 * tcp_make_synack - Prepare a SYN-ACK.
3157 * sk: listener socket
3158 * dst: dst entry attached to the SYNACK
3159 * req: request_sock pointer
3161 * Allocate one skb and build a SYNACK packet.
3162 * @dst is consumed : Caller should not use it again.
3164 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3165 struct request_sock *req,
3166 struct tcp_fastopen_cookie *foc,
3167 enum tcp_synack_type synack_type)
3169 struct inet_request_sock *ireq = inet_rsk(req);
3170 const struct tcp_sock *tp = tcp_sk(sk);
3171 struct tcp_md5sig_key *md5 = NULL;
3172 struct tcp_out_options opts;
3173 struct sk_buff *skb;
3174 int tcp_header_size;
3178 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3179 if (unlikely(!skb)) {
3183 /* Reserve space for headers. */
3184 skb_reserve(skb, MAX_TCP_HEADER);
3186 switch (synack_type) {
3187 case TCP_SYNACK_NORMAL:
3188 skb_set_owner_w(skb, req_to_sk(req));
3190 case TCP_SYNACK_COOKIE:
3191 /* Under synflood, we do not attach skb to a socket,
3192 * to avoid false sharing.
3195 case TCP_SYNACK_FASTOPEN:
3196 /* sk is a const pointer, because we want to express multiple
3197 * cpu might call us concurrently.
3198 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3200 skb_set_owner_w(skb, (struct sock *)sk);
3203 skb_dst_set(skb, dst);
3205 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3207 memset(&opts, 0, sizeof(opts));
3208 #ifdef CONFIG_SYN_COOKIES
3209 if (unlikely(req->cookie_ts))
3210 skb->skb_mstamp = cookie_init_timestamp(req);
3213 skb->skb_mstamp = tcp_clock_us();
3215 #ifdef CONFIG_TCP_MD5SIG
3217 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3219 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3220 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3223 skb_push(skb, tcp_header_size);
3224 skb_reset_transport_header(skb);
3226 th = (struct tcphdr *)skb->data;
3227 memset(th, 0, sizeof(struct tcphdr));
3230 tcp_ecn_make_synack(req, th);
3231 th->source = htons(ireq->ir_num);
3232 th->dest = ireq->ir_rmt_port;
3233 skb->mark = ireq->ir_mark;
3234 skb->ip_summed = CHECKSUM_PARTIAL;
3235 th->seq = htonl(tcp_rsk(req)->snt_isn);
3236 /* XXX data is queued and acked as is. No buffer/window check */
3237 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3239 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3240 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3241 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3242 th->doff = (tcp_header_size >> 2);
3243 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3245 #ifdef CONFIG_TCP_MD5SIG
3246 /* Okay, we have all we need - do the md5 hash if needed */
3248 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3249 md5, req_to_sk(req), skb);
3253 /* Do not fool tcpdump (if any), clean our debris */
3257 EXPORT_SYMBOL(tcp_make_synack);
3259 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3261 struct inet_connection_sock *icsk = inet_csk(sk);
3262 const struct tcp_congestion_ops *ca;
3263 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3265 if (ca_key == TCP_CA_UNSPEC)
3269 ca = tcp_ca_find_key(ca_key);
3270 if (likely(ca && try_module_get(ca->owner))) {
3271 module_put(icsk->icsk_ca_ops->owner);
3272 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3273 icsk->icsk_ca_ops = ca;
3278 /* Do all connect socket setups that can be done AF independent. */
3279 static void tcp_connect_init(struct sock *sk)
3281 const struct dst_entry *dst = __sk_dst_get(sk);
3282 struct tcp_sock *tp = tcp_sk(sk);
3286 /* We'll fix this up when we get a response from the other end.
3287 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3289 tp->tcp_header_len = sizeof(struct tcphdr);
3290 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3291 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3293 #ifdef CONFIG_TCP_MD5SIG
3294 if (tp->af_specific->md5_lookup(sk, sk))
3295 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3298 /* If user gave his TCP_MAXSEG, record it to clamp */
3299 if (tp->rx_opt.user_mss)
3300 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3303 tcp_sync_mss(sk, dst_mtu(dst));
3305 tcp_ca_dst_init(sk, dst);
3307 if (!tp->window_clamp)
3308 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3309 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3311 tcp_initialize_rcv_mss(sk);
3313 /* limit the window selection if the user enforce a smaller rx buffer */
3314 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3315 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3316 tp->window_clamp = tcp_full_space(sk);
3318 rcv_wnd = tcp_rwnd_init_bpf(sk);
3320 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3322 tcp_select_initial_window(sk, tcp_full_space(sk),
3323 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3326 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3330 tp->rx_opt.rcv_wscale = rcv_wscale;
3331 tp->rcv_ssthresh = tp->rcv_wnd;
3334 sock_reset_flag(sk, SOCK_DONE);
3337 tcp_write_queue_purge(sk);
3338 tp->snd_una = tp->write_seq;
3339 tp->snd_sml = tp->write_seq;
3340 tp->snd_up = tp->write_seq;
3341 tp->snd_nxt = tp->write_seq;
3343 if (likely(!tp->repair))
3346 tp->rcv_tstamp = tcp_jiffies32;
3347 tp->rcv_wup = tp->rcv_nxt;
3348 tp->copied_seq = tp->rcv_nxt;
3350 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3351 inet_csk(sk)->icsk_retransmits = 0;
3352 tcp_clear_retrans(tp);
3355 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3357 struct tcp_sock *tp = tcp_sk(sk);
3358 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3360 tcb->end_seq += skb->len;
3361 __skb_header_release(skb);
3362 sk->sk_wmem_queued += skb->truesize;
3363 sk_mem_charge(sk, skb->truesize);
3364 tp->write_seq = tcb->end_seq;
3365 tp->packets_out += tcp_skb_pcount(skb);
3368 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3369 * queue a data-only packet after the regular SYN, such that regular SYNs
3370 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3371 * only the SYN sequence, the data are retransmitted in the first ACK.
3372 * If cookie is not cached or other error occurs, falls back to send a
3373 * regular SYN with Fast Open cookie request option.
3375 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3377 struct tcp_sock *tp = tcp_sk(sk);
3378 struct tcp_fastopen_request *fo = tp->fastopen_req;
3380 struct sk_buff *syn_data;
3382 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3383 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3386 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3387 * user-MSS. Reserve maximum option space for middleboxes that add
3388 * private TCP options. The cost is reduced data space in SYN :(
3390 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3392 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3393 MAX_TCP_OPTION_SPACE;
3395 space = min_t(size_t, space, fo->size);
3397 /* limit to order-0 allocations */
3398 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3400 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3403 syn_data->ip_summed = CHECKSUM_PARTIAL;
3404 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3406 int copied = copy_from_iter(skb_put(syn_data, space), space,
3407 &fo->data->msg_iter);
3408 if (unlikely(!copied)) {
3409 tcp_skb_tsorted_anchor_cleanup(syn_data);
3410 kfree_skb(syn_data);
3413 if (copied != space) {
3414 skb_trim(syn_data, copied);
3418 /* No more data pending in inet_wait_for_connect() */
3419 if (space == fo->size)
3423 tcp_connect_queue_skb(sk, syn_data);
3425 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3427 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3429 syn->skb_mstamp = syn_data->skb_mstamp;
3431 /* Now full SYN+DATA was cloned and sent (or not),
3432 * remove the SYN from the original skb (syn_data)
3433 * we keep in write queue in case of a retransmit, as we
3434 * also have the SYN packet (with no data) in the same queue.
3436 TCP_SKB_CB(syn_data)->seq++;
3437 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3439 tp->syn_data = (fo->copied > 0);
3440 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3441 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3445 /* data was not sent, put it in write_queue */
3446 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3447 tp->packets_out -= tcp_skb_pcount(syn_data);
3450 /* Send a regular SYN with Fast Open cookie request option */
3451 if (fo->cookie.len > 0)
3453 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3455 tp->syn_fastopen = 0;
3457 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3461 /* Build a SYN and send it off. */
3462 int tcp_connect(struct sock *sk)
3464 struct tcp_sock *tp = tcp_sk(sk);
3465 struct sk_buff *buff;
3468 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3470 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3471 return -EHOSTUNREACH; /* Routing failure or similar. */
3473 tcp_connect_init(sk);
3475 if (unlikely(tp->repair)) {
3476 tcp_finish_connect(sk, NULL);
3480 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3481 if (unlikely(!buff))
3484 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3485 tcp_mstamp_refresh(tp);
3486 tp->retrans_stamp = tcp_time_stamp(tp);
3487 tcp_connect_queue_skb(sk, buff);
3488 tcp_ecn_send_syn(sk, buff);
3489 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3491 /* Send off SYN; include data in Fast Open. */
3492 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3493 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3494 if (err == -ECONNREFUSED)
3497 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3498 * in order to make this packet get counted in tcpOutSegs.
3500 tp->snd_nxt = tp->write_seq;
3501 tp->pushed_seq = tp->write_seq;
3502 buff = tcp_send_head(sk);
3503 if (unlikely(buff)) {
3504 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3505 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3507 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3509 /* Timer for repeating the SYN until an answer. */
3510 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3511 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3514 EXPORT_SYMBOL(tcp_connect);
3516 /* Send out a delayed ack, the caller does the policy checking
3517 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3520 void tcp_send_delayed_ack(struct sock *sk)
3522 struct inet_connection_sock *icsk = inet_csk(sk);
3523 int ato = icsk->icsk_ack.ato;
3524 unsigned long timeout;
3526 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3528 if (ato > TCP_DELACK_MIN) {
3529 const struct tcp_sock *tp = tcp_sk(sk);
3530 int max_ato = HZ / 2;
3532 if (icsk->icsk_ack.pingpong ||
3533 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3534 max_ato = TCP_DELACK_MAX;
3536 /* Slow path, intersegment interval is "high". */
3538 /* If some rtt estimate is known, use it to bound delayed ack.
3539 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3543 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3550 ato = min(ato, max_ato);
3553 /* Stay within the limit we were given */
3554 timeout = jiffies + ato;
3556 /* Use new timeout only if there wasn't a older one earlier. */
3557 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3558 /* If delack timer was blocked or is about to expire,
3561 if (icsk->icsk_ack.blocked ||
3562 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3567 if (!time_before(timeout, icsk->icsk_ack.timeout))
3568 timeout = icsk->icsk_ack.timeout;
3570 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3571 icsk->icsk_ack.timeout = timeout;
3572 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3575 /* This routine sends an ack and also updates the window. */
3576 void tcp_send_ack(struct sock *sk)
3578 struct sk_buff *buff;
3580 /* If we have been reset, we may not send again. */
3581 if (sk->sk_state == TCP_CLOSE)
3584 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3586 /* We are not putting this on the write queue, so
3587 * tcp_transmit_skb() will set the ownership to this
3590 buff = alloc_skb(MAX_TCP_HEADER,
3591 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3592 if (unlikely(!buff)) {
3593 inet_csk_schedule_ack(sk);
3594 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3595 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3596 TCP_DELACK_MAX, TCP_RTO_MAX);
3600 /* Reserve space for headers and prepare control bits. */
3601 skb_reserve(buff, MAX_TCP_HEADER);
3602 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3604 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3606 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3608 skb_set_tcp_pure_ack(buff);
3610 /* Send it off, this clears delayed acks for us. */
3611 tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3613 EXPORT_SYMBOL_GPL(tcp_send_ack);
3615 /* This routine sends a packet with an out of date sequence
3616 * number. It assumes the other end will try to ack it.
3618 * Question: what should we make while urgent mode?
3619 * 4.4BSD forces sending single byte of data. We cannot send
3620 * out of window data, because we have SND.NXT==SND.MAX...
3622 * Current solution: to send TWO zero-length segments in urgent mode:
3623 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3624 * out-of-date with SND.UNA-1 to probe window.
3626 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3628 struct tcp_sock *tp = tcp_sk(sk);
3629 struct sk_buff *skb;
3631 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3632 skb = alloc_skb(MAX_TCP_HEADER,
3633 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3637 /* Reserve space for headers and set control bits. */
3638 skb_reserve(skb, MAX_TCP_HEADER);
3639 /* Use a previous sequence. This should cause the other
3640 * end to send an ack. Don't queue or clone SKB, just
3643 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3644 NET_INC_STATS(sock_net(sk), mib);
3645 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3648 /* Called from setsockopt( ... TCP_REPAIR ) */
3649 void tcp_send_window_probe(struct sock *sk)
3651 if (sk->sk_state == TCP_ESTABLISHED) {
3652 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3653 tcp_mstamp_refresh(tcp_sk(sk));
3654 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3658 /* Initiate keepalive or window probe from timer. */
3659 int tcp_write_wakeup(struct sock *sk, int mib)
3661 struct tcp_sock *tp = tcp_sk(sk);
3662 struct sk_buff *skb;
3664 if (sk->sk_state == TCP_CLOSE)
3667 skb = tcp_send_head(sk);
3668 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3670 unsigned int mss = tcp_current_mss(sk);
3671 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3673 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3674 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3676 /* We are probing the opening of a window
3677 * but the window size is != 0
3678 * must have been a result SWS avoidance ( sender )
3680 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3682 seg_size = min(seg_size, mss);
3683 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3684 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3685 skb, seg_size, mss, GFP_ATOMIC))
3687 } else if (!tcp_skb_pcount(skb))
3688 tcp_set_skb_tso_segs(skb, mss);
3690 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3691 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3693 tcp_event_new_data_sent(sk, skb);
3696 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3697 tcp_xmit_probe_skb(sk, 1, mib);
3698 return tcp_xmit_probe_skb(sk, 0, mib);
3702 /* A window probe timeout has occurred. If window is not closed send
3703 * a partial packet else a zero probe.
3705 void tcp_send_probe0(struct sock *sk)
3707 struct inet_connection_sock *icsk = inet_csk(sk);
3708 struct tcp_sock *tp = tcp_sk(sk);
3709 struct net *net = sock_net(sk);
3710 unsigned long probe_max;
3713 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3715 if (tp->packets_out || tcp_write_queue_empty(sk)) {
3716 /* Cancel probe timer, if it is not required. */
3717 icsk->icsk_probes_out = 0;
3718 icsk->icsk_backoff = 0;
3723 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3724 icsk->icsk_backoff++;
3725 icsk->icsk_probes_out++;
3726 probe_max = TCP_RTO_MAX;
3728 /* If packet was not sent due to local congestion,
3729 * do not backoff and do not remember icsk_probes_out.
3730 * Let local senders to fight for local resources.
3732 * Use accumulated backoff yet.
3734 if (!icsk->icsk_probes_out)
3735 icsk->icsk_probes_out = 1;
3736 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3738 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3739 tcp_probe0_when(sk, probe_max),
3743 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3745 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3749 tcp_rsk(req)->txhash = net_tx_rndhash();
3750 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3752 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3753 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3754 if (unlikely(tcp_passive_fastopen(sk)))
3755 tcp_sk(sk)->total_retrans++;
3756 trace_tcp_retransmit_synack(sk, req);
3760 EXPORT_SYMBOL(tcp_rtx_synack);