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,
1027 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
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(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 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1184 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1185 tcp_sk(sk)->rcv_nxt);
1188 /* This routine just queues the buffer for sending.
1190 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1191 * otherwise socket can stall.
1193 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1195 struct tcp_sock *tp = tcp_sk(sk);
1197 /* Advance write_seq and place onto the write_queue. */
1198 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1199 __skb_header_release(skb);
1200 tcp_add_write_queue_tail(sk, skb);
1201 sk->sk_wmem_queued += skb->truesize;
1202 sk_mem_charge(sk, skb->truesize);
1205 /* Initialize TSO segments for a packet. */
1206 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1208 if (skb->len <= mss_now) {
1209 /* Avoid the costly divide in the normal
1212 tcp_skb_pcount_set(skb, 1);
1213 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1215 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1216 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1220 /* Pcount in the middle of the write queue got changed, we need to do various
1221 * tweaks to fix counters
1223 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1225 struct tcp_sock *tp = tcp_sk(sk);
1227 tp->packets_out -= decr;
1229 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1230 tp->sacked_out -= decr;
1231 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1232 tp->retrans_out -= decr;
1233 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1234 tp->lost_out -= decr;
1236 /* Reno case is special. Sigh... */
1237 if (tcp_is_reno(tp) && decr > 0)
1238 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1240 if (tp->lost_skb_hint &&
1241 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1242 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1243 tp->lost_cnt_hint -= decr;
1245 tcp_verify_left_out(tp);
1248 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1250 return TCP_SKB_CB(skb)->txstamp_ack ||
1251 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1254 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1256 struct skb_shared_info *shinfo = skb_shinfo(skb);
1258 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1259 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1260 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1261 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1263 shinfo->tx_flags &= ~tsflags;
1264 shinfo2->tx_flags |= tsflags;
1265 swap(shinfo->tskey, shinfo2->tskey);
1266 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1267 TCP_SKB_CB(skb)->txstamp_ack = 0;
1271 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1273 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1274 TCP_SKB_CB(skb)->eor = 0;
1277 /* Insert buff after skb on the write or rtx queue of sk. */
1278 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1279 struct sk_buff *buff,
1281 enum tcp_queue tcp_queue)
1283 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1284 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1286 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1289 /* Function to create two new TCP segments. Shrinks the given segment
1290 * to the specified size and appends a new segment with the rest of the
1291 * packet to the list. This won't be called frequently, I hope.
1292 * Remember, these are still headerless SKBs at this point.
1294 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1295 struct sk_buff *skb, u32 len,
1296 unsigned int mss_now, gfp_t gfp)
1298 struct tcp_sock *tp = tcp_sk(sk);
1299 struct sk_buff *buff;
1300 int nsize, old_factor;
1304 if (WARN_ON(len > skb->len))
1307 nsize = skb_headlen(skb) - len;
1311 if (skb_unclone(skb, gfp))
1314 /* Get a new skb... force flag on. */
1315 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1317 return -ENOMEM; /* We'll just try again later. */
1319 sk->sk_wmem_queued += buff->truesize;
1320 sk_mem_charge(sk, buff->truesize);
1321 nlen = skb->len - len - nsize;
1322 buff->truesize += nlen;
1323 skb->truesize -= nlen;
1325 /* Correct the sequence numbers. */
1326 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1327 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1328 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1330 /* PSH and FIN should only be set in the second packet. */
1331 flags = TCP_SKB_CB(skb)->tcp_flags;
1332 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1333 TCP_SKB_CB(buff)->tcp_flags = flags;
1334 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1335 tcp_skb_fragment_eor(skb, buff);
1337 skb_split(skb, buff, len);
1339 buff->ip_summed = CHECKSUM_PARTIAL;
1341 buff->tstamp = skb->tstamp;
1342 tcp_fragment_tstamp(skb, buff);
1344 old_factor = tcp_skb_pcount(skb);
1346 /* Fix up tso_factor for both original and new SKB. */
1347 tcp_set_skb_tso_segs(skb, mss_now);
1348 tcp_set_skb_tso_segs(buff, mss_now);
1350 /* Update delivered info for the new segment */
1351 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1353 /* If this packet has been sent out already, we must
1354 * adjust the various packet counters.
1356 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1357 int diff = old_factor - tcp_skb_pcount(skb) -
1358 tcp_skb_pcount(buff);
1361 tcp_adjust_pcount(sk, skb, diff);
1364 /* Link BUFF into the send queue. */
1365 __skb_header_release(buff);
1366 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1367 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1368 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1373 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1374 * data is not copied, but immediately discarded.
1376 static int __pskb_trim_head(struct sk_buff *skb, int len)
1378 struct skb_shared_info *shinfo;
1381 eat = min_t(int, len, skb_headlen(skb));
1383 __skb_pull(skb, eat);
1390 shinfo = skb_shinfo(skb);
1391 for (i = 0; i < shinfo->nr_frags; i++) {
1392 int size = skb_frag_size(&shinfo->frags[i]);
1395 skb_frag_unref(skb, i);
1398 shinfo->frags[k] = shinfo->frags[i];
1400 shinfo->frags[k].page_offset += eat;
1401 skb_frag_size_sub(&shinfo->frags[k], eat);
1407 shinfo->nr_frags = k;
1409 skb->data_len -= len;
1410 skb->len = skb->data_len;
1414 /* Remove acked data from a packet in the transmit queue. */
1415 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1419 if (skb_unclone(skb, GFP_ATOMIC))
1422 delta_truesize = __pskb_trim_head(skb, len);
1424 TCP_SKB_CB(skb)->seq += len;
1425 skb->ip_summed = CHECKSUM_PARTIAL;
1427 if (delta_truesize) {
1428 skb->truesize -= delta_truesize;
1429 sk->sk_wmem_queued -= delta_truesize;
1430 sk_mem_uncharge(sk, delta_truesize);
1431 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1434 /* Any change of skb->len requires recalculation of tso factor. */
1435 if (tcp_skb_pcount(skb) > 1)
1436 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1441 /* Calculate MSS not accounting any TCP options. */
1442 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1444 const struct tcp_sock *tp = tcp_sk(sk);
1445 const struct inet_connection_sock *icsk = inet_csk(sk);
1448 /* Calculate base mss without TCP options:
1449 It is MMS_S - sizeof(tcphdr) of rfc1122
1451 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1453 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1454 if (icsk->icsk_af_ops->net_frag_header_len) {
1455 const struct dst_entry *dst = __sk_dst_get(sk);
1457 if (dst && dst_allfrag(dst))
1458 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1461 /* Clamp it (mss_clamp does not include tcp options) */
1462 if (mss_now > tp->rx_opt.mss_clamp)
1463 mss_now = tp->rx_opt.mss_clamp;
1465 /* Now subtract optional transport overhead */
1466 mss_now -= icsk->icsk_ext_hdr_len;
1468 /* Then reserve room for full set of TCP options and 8 bytes of data */
1474 /* Calculate MSS. Not accounting for SACKs here. */
1475 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1477 /* Subtract TCP options size, not including SACKs */
1478 return __tcp_mtu_to_mss(sk, pmtu) -
1479 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1482 /* Inverse of above */
1483 int tcp_mss_to_mtu(struct sock *sk, int mss)
1485 const struct tcp_sock *tp = tcp_sk(sk);
1486 const struct inet_connection_sock *icsk = inet_csk(sk);
1490 tp->tcp_header_len +
1491 icsk->icsk_ext_hdr_len +
1492 icsk->icsk_af_ops->net_header_len;
1494 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1495 if (icsk->icsk_af_ops->net_frag_header_len) {
1496 const struct dst_entry *dst = __sk_dst_get(sk);
1498 if (dst && dst_allfrag(dst))
1499 mtu += icsk->icsk_af_ops->net_frag_header_len;
1503 EXPORT_SYMBOL(tcp_mss_to_mtu);
1505 /* MTU probing init per socket */
1506 void tcp_mtup_init(struct sock *sk)
1508 struct tcp_sock *tp = tcp_sk(sk);
1509 struct inet_connection_sock *icsk = inet_csk(sk);
1510 struct net *net = sock_net(sk);
1512 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1513 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1514 icsk->icsk_af_ops->net_header_len;
1515 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1516 icsk->icsk_mtup.probe_size = 0;
1517 if (icsk->icsk_mtup.enabled)
1518 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1520 EXPORT_SYMBOL(tcp_mtup_init);
1522 /* This function synchronize snd mss to current pmtu/exthdr set.
1524 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1525 for TCP options, but includes only bare TCP header.
1527 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1528 It is minimum of user_mss and mss received with SYN.
1529 It also does not include TCP options.
1531 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1533 tp->mss_cache is current effective sending mss, including
1534 all tcp options except for SACKs. It is evaluated,
1535 taking into account current pmtu, but never exceeds
1536 tp->rx_opt.mss_clamp.
1538 NOTE1. rfc1122 clearly states that advertised MSS
1539 DOES NOT include either tcp or ip options.
1541 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1542 are READ ONLY outside this function. --ANK (980731)
1544 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1546 struct tcp_sock *tp = tcp_sk(sk);
1547 struct inet_connection_sock *icsk = inet_csk(sk);
1550 if (icsk->icsk_mtup.search_high > pmtu)
1551 icsk->icsk_mtup.search_high = pmtu;
1553 mss_now = tcp_mtu_to_mss(sk, pmtu);
1554 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1556 /* And store cached results */
1557 icsk->icsk_pmtu_cookie = pmtu;
1558 if (icsk->icsk_mtup.enabled)
1559 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1560 tp->mss_cache = mss_now;
1564 EXPORT_SYMBOL(tcp_sync_mss);
1566 /* Compute the current effective MSS, taking SACKs and IP options,
1567 * and even PMTU discovery events into account.
1569 unsigned int tcp_current_mss(struct sock *sk)
1571 const struct tcp_sock *tp = tcp_sk(sk);
1572 const struct dst_entry *dst = __sk_dst_get(sk);
1574 unsigned int header_len;
1575 struct tcp_out_options opts;
1576 struct tcp_md5sig_key *md5;
1578 mss_now = tp->mss_cache;
1581 u32 mtu = dst_mtu(dst);
1582 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1583 mss_now = tcp_sync_mss(sk, mtu);
1586 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1587 sizeof(struct tcphdr);
1588 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1589 * some common options. If this is an odd packet (because we have SACK
1590 * blocks etc) then our calculated header_len will be different, and
1591 * we have to adjust mss_now correspondingly */
1592 if (header_len != tp->tcp_header_len) {
1593 int delta = (int) header_len - tp->tcp_header_len;
1600 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1601 * As additional protections, we do not touch cwnd in retransmission phases,
1602 * and if application hit its sndbuf limit recently.
1604 static void tcp_cwnd_application_limited(struct sock *sk)
1606 struct tcp_sock *tp = tcp_sk(sk);
1608 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1609 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1610 /* Limited by application or receiver window. */
1611 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1612 u32 win_used = max(tp->snd_cwnd_used, init_win);
1613 if (win_used < tp->snd_cwnd) {
1614 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1615 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1617 tp->snd_cwnd_used = 0;
1619 tp->snd_cwnd_stamp = tcp_jiffies32;
1622 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1624 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1625 struct tcp_sock *tp = tcp_sk(sk);
1627 /* Track the maximum number of outstanding packets in each
1628 * window, and remember whether we were cwnd-limited then.
1630 if (!before(tp->snd_una, tp->max_packets_seq) ||
1631 tp->packets_out > tp->max_packets_out) {
1632 tp->max_packets_out = tp->packets_out;
1633 tp->max_packets_seq = tp->snd_nxt;
1634 tp->is_cwnd_limited = is_cwnd_limited;
1637 if (tcp_is_cwnd_limited(sk)) {
1638 /* Network is feed fully. */
1639 tp->snd_cwnd_used = 0;
1640 tp->snd_cwnd_stamp = tcp_jiffies32;
1642 /* Network starves. */
1643 if (tp->packets_out > tp->snd_cwnd_used)
1644 tp->snd_cwnd_used = tp->packets_out;
1646 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1647 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1648 !ca_ops->cong_control)
1649 tcp_cwnd_application_limited(sk);
1651 /* The following conditions together indicate the starvation
1652 * is caused by insufficient sender buffer:
1653 * 1) just sent some data (see tcp_write_xmit)
1654 * 2) not cwnd limited (this else condition)
1655 * 3) no more data to send (tcp_write_queue_empty())
1656 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1658 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1659 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1660 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1661 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1665 /* Minshall's variant of the Nagle send check. */
1666 static bool tcp_minshall_check(const struct tcp_sock *tp)
1668 return after(tp->snd_sml, tp->snd_una) &&
1669 !after(tp->snd_sml, tp->snd_nxt);
1672 /* Update snd_sml if this skb is under mss
1673 * Note that a TSO packet might end with a sub-mss segment
1674 * The test is really :
1675 * if ((skb->len % mss) != 0)
1676 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1677 * But we can avoid doing the divide again given we already have
1678 * skb_pcount = skb->len / mss_now
1680 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1681 const struct sk_buff *skb)
1683 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1684 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1687 /* Return false, if packet can be sent now without violation Nagle's rules:
1688 * 1. It is full sized. (provided by caller in %partial bool)
1689 * 2. Or it contains FIN. (already checked by caller)
1690 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1691 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1692 * With Minshall's modification: all sent small packets are ACKed.
1694 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1698 ((nonagle & TCP_NAGLE_CORK) ||
1699 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1702 /* Return how many segs we'd like on a TSO packet,
1703 * to send one TSO packet per ms
1705 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1710 bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift,
1711 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1713 /* Goal is to send at least one packet per ms,
1714 * not one big TSO packet every 100 ms.
1715 * This preserves ACK clocking and is consistent
1716 * with tcp_tso_should_defer() heuristic.
1718 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1723 /* Return the number of segments we want in the skb we are transmitting.
1724 * See if congestion control module wants to decide; otherwise, autosize.
1726 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1728 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1729 u32 min_tso, tso_segs;
1731 min_tso = ca_ops->min_tso_segs ?
1732 ca_ops->min_tso_segs(sk) :
1733 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1735 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1736 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1739 /* Returns the portion of skb which can be sent right away */
1740 static unsigned int tcp_mss_split_point(const struct sock *sk,
1741 const struct sk_buff *skb,
1742 unsigned int mss_now,
1743 unsigned int max_segs,
1746 const struct tcp_sock *tp = tcp_sk(sk);
1747 u32 partial, needed, window, max_len;
1749 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1750 max_len = mss_now * max_segs;
1752 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1755 needed = min(skb->len, window);
1757 if (max_len <= needed)
1760 partial = needed % mss_now;
1761 /* If last segment is not a full MSS, check if Nagle rules allow us
1762 * to include this last segment in this skb.
1763 * Otherwise, we'll split the skb at last MSS boundary
1765 if (tcp_nagle_check(partial != 0, tp, nonagle))
1766 return needed - partial;
1771 /* Can at least one segment of SKB be sent right now, according to the
1772 * congestion window rules? If so, return how many segments are allowed.
1774 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1775 const struct sk_buff *skb)
1777 u32 in_flight, cwnd, halfcwnd;
1779 /* Don't be strict about the congestion window for the final FIN. */
1780 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1781 tcp_skb_pcount(skb) == 1)
1784 in_flight = tcp_packets_in_flight(tp);
1785 cwnd = tp->snd_cwnd;
1786 if (in_flight >= cwnd)
1789 /* For better scheduling, ensure we have at least
1790 * 2 GSO packets in flight.
1792 halfcwnd = max(cwnd >> 1, 1U);
1793 return min(halfcwnd, cwnd - in_flight);
1796 /* Initialize TSO state of a skb.
1797 * This must be invoked the first time we consider transmitting
1798 * SKB onto the wire.
1800 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1802 int tso_segs = tcp_skb_pcount(skb);
1804 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1805 tcp_set_skb_tso_segs(skb, mss_now);
1806 tso_segs = tcp_skb_pcount(skb);
1812 /* Return true if the Nagle test allows this packet to be
1815 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1816 unsigned int cur_mss, int nonagle)
1818 /* Nagle rule does not apply to frames, which sit in the middle of the
1819 * write_queue (they have no chances to get new data).
1821 * This is implemented in the callers, where they modify the 'nonagle'
1822 * argument based upon the location of SKB in the send queue.
1824 if (nonagle & TCP_NAGLE_PUSH)
1827 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1828 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1831 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1837 /* Does at least the first segment of SKB fit into the send window? */
1838 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1839 const struct sk_buff *skb,
1840 unsigned int cur_mss)
1842 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1844 if (skb->len > cur_mss)
1845 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1847 return !after(end_seq, tcp_wnd_end(tp));
1850 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1851 * which is put after SKB on the list. It is very much like
1852 * tcp_fragment() except that it may make several kinds of assumptions
1853 * in order to speed up the splitting operation. In particular, we
1854 * know that all the data is in scatter-gather pages, and that the
1855 * packet has never been sent out before (and thus is not cloned).
1857 static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1858 struct sk_buff *skb, unsigned int len,
1859 unsigned int mss_now, gfp_t gfp)
1861 struct sk_buff *buff;
1862 int nlen = skb->len - len;
1865 /* All of a TSO frame must be composed of paged data. */
1866 if (skb->len != skb->data_len)
1867 return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp);
1869 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1870 if (unlikely(!buff))
1873 sk->sk_wmem_queued += buff->truesize;
1874 sk_mem_charge(sk, buff->truesize);
1875 buff->truesize += nlen;
1876 skb->truesize -= nlen;
1878 /* Correct the sequence numbers. */
1879 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1880 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1881 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1883 /* PSH and FIN should only be set in the second packet. */
1884 flags = TCP_SKB_CB(skb)->tcp_flags;
1885 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1886 TCP_SKB_CB(buff)->tcp_flags = flags;
1888 /* This packet was never sent out yet, so no SACK bits. */
1889 TCP_SKB_CB(buff)->sacked = 0;
1891 tcp_skb_fragment_eor(skb, buff);
1893 buff->ip_summed = CHECKSUM_PARTIAL;
1894 skb_split(skb, buff, len);
1895 tcp_fragment_tstamp(skb, buff);
1897 /* Fix up tso_factor for both original and new SKB. */
1898 tcp_set_skb_tso_segs(skb, mss_now);
1899 tcp_set_skb_tso_segs(buff, mss_now);
1901 /* Link BUFF into the send queue. */
1902 __skb_header_release(buff);
1903 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1908 /* Try to defer sending, if possible, in order to minimize the amount
1909 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1911 * This algorithm is from John Heffner.
1913 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1914 bool *is_cwnd_limited, u32 max_segs)
1916 const struct inet_connection_sock *icsk = inet_csk(sk);
1917 u32 age, send_win, cong_win, limit, in_flight;
1918 struct tcp_sock *tp = tcp_sk(sk);
1919 struct sk_buff *head;
1922 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1925 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1928 /* Avoid bursty behavior by allowing defer
1929 * only if the last write was recent.
1931 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
1934 in_flight = tcp_packets_in_flight(tp);
1936 BUG_ON(tcp_skb_pcount(skb) <= 1);
1937 BUG_ON(tp->snd_cwnd <= in_flight);
1939 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1941 /* From in_flight test above, we know that cwnd > in_flight. */
1942 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1944 limit = min(send_win, cong_win);
1946 /* If a full-sized TSO skb can be sent, do it. */
1947 if (limit >= max_segs * tp->mss_cache)
1950 /* Middle in queue won't get any more data, full sendable already? */
1951 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1954 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1956 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1958 /* If at least some fraction of a window is available,
1961 chunk /= win_divisor;
1965 /* Different approach, try not to defer past a single
1966 * ACK. Receiver should ACK every other full sized
1967 * frame, so if we have space for more than 3 frames
1970 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1974 /* TODO : use tsorted_sent_queue ? */
1975 head = tcp_rtx_queue_head(sk);
1978 age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
1979 /* If next ACK is likely to come too late (half srtt), do not defer */
1980 if (age < (tp->srtt_us >> 4))
1983 /* Ok, it looks like it is advisable to defer. */
1985 if (cong_win < send_win && cong_win <= skb->len)
1986 *is_cwnd_limited = true;
1994 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1996 struct inet_connection_sock *icsk = inet_csk(sk);
1997 struct tcp_sock *tp = tcp_sk(sk);
1998 struct net *net = sock_net(sk);
2002 interval = net->ipv4.sysctl_tcp_probe_interval;
2003 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2004 if (unlikely(delta >= interval * HZ)) {
2005 int mss = tcp_current_mss(sk);
2007 /* Update current search range */
2008 icsk->icsk_mtup.probe_size = 0;
2009 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2010 sizeof(struct tcphdr) +
2011 icsk->icsk_af_ops->net_header_len;
2012 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2014 /* Update probe time stamp */
2015 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2019 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2021 struct sk_buff *skb, *next;
2023 skb = tcp_send_head(sk);
2024 tcp_for_write_queue_from_safe(skb, next, sk) {
2025 if (len <= skb->len)
2028 if (unlikely(TCP_SKB_CB(skb)->eor))
2037 /* Create a new MTU probe if we are ready.
2038 * MTU probe is regularly attempting to increase the path MTU by
2039 * deliberately sending larger packets. This discovers routing
2040 * changes resulting in larger path MTUs.
2042 * Returns 0 if we should wait to probe (no cwnd available),
2043 * 1 if a probe was sent,
2046 static int tcp_mtu_probe(struct sock *sk)
2048 struct inet_connection_sock *icsk = inet_csk(sk);
2049 struct tcp_sock *tp = tcp_sk(sk);
2050 struct sk_buff *skb, *nskb, *next;
2051 struct net *net = sock_net(sk);
2058 /* Not currently probing/verifying,
2060 * have enough cwnd, and
2061 * not SACKing (the variable headers throw things off)
2063 if (likely(!icsk->icsk_mtup.enabled ||
2064 icsk->icsk_mtup.probe_size ||
2065 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2066 tp->snd_cwnd < 11 ||
2067 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2070 /* Use binary search for probe_size between tcp_mss_base,
2071 * and current mss_clamp. if (search_high - search_low)
2072 * smaller than a threshold, backoff from probing.
2074 mss_now = tcp_current_mss(sk);
2075 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2076 icsk->icsk_mtup.search_low) >> 1);
2077 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2078 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2079 /* When misfortune happens, we are reprobing actively,
2080 * and then reprobe timer has expired. We stick with current
2081 * probing process by not resetting search range to its orignal.
2083 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2084 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2085 /* Check whether enough time has elaplased for
2086 * another round of probing.
2088 tcp_mtu_check_reprobe(sk);
2092 /* Have enough data in the send queue to probe? */
2093 if (tp->write_seq - tp->snd_nxt < size_needed)
2096 if (tp->snd_wnd < size_needed)
2098 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2101 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2102 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2103 if (!tcp_packets_in_flight(tp))
2109 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2112 /* We're allowed to probe. Build it now. */
2113 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2116 sk->sk_wmem_queued += nskb->truesize;
2117 sk_mem_charge(sk, nskb->truesize);
2119 skb = tcp_send_head(sk);
2121 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2122 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2123 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2124 TCP_SKB_CB(nskb)->sacked = 0;
2126 nskb->ip_summed = CHECKSUM_PARTIAL;
2128 tcp_insert_write_queue_before(nskb, skb, sk);
2129 tcp_highest_sack_replace(sk, skb, nskb);
2132 tcp_for_write_queue_from_safe(skb, next, sk) {
2133 copy = min_t(int, skb->len, probe_size - len);
2134 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2136 if (skb->len <= copy) {
2137 /* We've eaten all the data from this skb.
2139 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2140 /* If this is the last SKB we copy and eor is set
2141 * we need to propagate it to the new skb.
2143 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2144 tcp_unlink_write_queue(skb, sk);
2145 sk_wmem_free_skb(sk, skb);
2147 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2148 ~(TCPHDR_FIN|TCPHDR_PSH);
2149 if (!skb_shinfo(skb)->nr_frags) {
2150 skb_pull(skb, copy);
2152 __pskb_trim_head(skb, copy);
2153 tcp_set_skb_tso_segs(skb, mss_now);
2155 TCP_SKB_CB(skb)->seq += copy;
2160 if (len >= probe_size)
2163 tcp_init_tso_segs(nskb, nskb->len);
2165 /* We're ready to send. If this fails, the probe will
2166 * be resegmented into mss-sized pieces by tcp_write_xmit().
2168 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2169 /* Decrement cwnd here because we are sending
2170 * effectively two packets. */
2172 tcp_event_new_data_sent(sk, nskb);
2174 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2175 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2176 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2184 static bool tcp_pacing_check(const struct sock *sk)
2186 return tcp_needs_internal_pacing(sk) &&
2187 hrtimer_is_queued(&tcp_sk(sk)->pacing_timer);
2190 /* TCP Small Queues :
2191 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2192 * (These limits are doubled for retransmits)
2194 * - better RTT estimation and ACK scheduling
2197 * Alas, some drivers / subsystems require a fair amount
2198 * of queued bytes to ensure line rate.
2199 * One example is wifi aggregation (802.11 AMPDU)
2201 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2202 unsigned int factor)
2206 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift);
2207 limit = min_t(u32, limit,
2208 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2211 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2212 /* Always send skb if rtx queue is empty.
2213 * No need to wait for TX completion to call us back,
2214 * after softirq/tasklet schedule.
2215 * This helps when TX completions are delayed too much.
2217 if (tcp_rtx_queue_empty(sk))
2220 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2221 /* It is possible TX completion already happened
2222 * before we set TSQ_THROTTLED, so we must
2223 * test again the condition.
2225 smp_mb__after_atomic();
2226 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2232 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2234 const u32 now = tcp_jiffies32;
2235 enum tcp_chrono old = tp->chrono_type;
2237 if (old > TCP_CHRONO_UNSPEC)
2238 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2239 tp->chrono_start = now;
2240 tp->chrono_type = new;
2243 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2245 struct tcp_sock *tp = tcp_sk(sk);
2247 /* If there are multiple conditions worthy of tracking in a
2248 * chronograph then the highest priority enum takes precedence
2249 * over the other conditions. So that if something "more interesting"
2250 * starts happening, stop the previous chrono and start a new one.
2252 if (type > tp->chrono_type)
2253 tcp_chrono_set(tp, type);
2256 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2258 struct tcp_sock *tp = tcp_sk(sk);
2261 /* There are multiple conditions worthy of tracking in a
2262 * chronograph, so that the highest priority enum takes
2263 * precedence over the other conditions (see tcp_chrono_start).
2264 * If a condition stops, we only stop chrono tracking if
2265 * it's the "most interesting" or current chrono we are
2266 * tracking and starts busy chrono if we have pending data.
2268 if (tcp_rtx_and_write_queues_empty(sk))
2269 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2270 else if (type == tp->chrono_type)
2271 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2274 /* This routine writes packets to the network. It advances the
2275 * send_head. This happens as incoming acks open up the remote
2278 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2279 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2280 * account rare use of URG, this is not a big flaw.
2282 * Send at most one packet when push_one > 0. Temporarily ignore
2283 * cwnd limit to force at most one packet out when push_one == 2.
2285 * Returns true, if no segments are in flight and we have queued segments,
2286 * but cannot send anything now because of SWS or another problem.
2288 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2289 int push_one, gfp_t gfp)
2291 struct tcp_sock *tp = tcp_sk(sk);
2292 struct sk_buff *skb;
2293 unsigned int tso_segs, sent_pkts;
2296 bool is_cwnd_limited = false, is_rwnd_limited = false;
2301 tcp_mstamp_refresh(tp);
2303 /* Do MTU probing. */
2304 result = tcp_mtu_probe(sk);
2307 } else if (result > 0) {
2312 max_segs = tcp_tso_segs(sk, mss_now);
2313 while ((skb = tcp_send_head(sk))) {
2316 if (tcp_pacing_check(sk))
2319 tso_segs = tcp_init_tso_segs(skb, mss_now);
2322 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2323 /* "skb_mstamp" is used as a start point for the retransmit timer */
2324 tcp_update_skb_after_send(tp, skb);
2325 goto repair; /* Skip network transmission */
2328 cwnd_quota = tcp_cwnd_test(tp, skb);
2331 /* Force out a loss probe pkt. */
2337 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2338 is_rwnd_limited = true;
2342 if (tso_segs == 1) {
2343 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2344 (tcp_skb_is_last(sk, skb) ?
2345 nonagle : TCP_NAGLE_PUSH))))
2349 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2355 if (tso_segs > 1 && !tcp_urg_mode(tp))
2356 limit = tcp_mss_split_point(sk, skb, mss_now,
2362 if (skb->len > limit &&
2363 unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2364 skb, limit, mss_now, gfp)))
2367 if (tcp_small_queue_check(sk, skb, 0))
2370 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2374 /* Advance the send_head. This one is sent out.
2375 * This call will increment packets_out.
2377 tcp_event_new_data_sent(sk, skb);
2379 tcp_minshall_update(tp, mss_now, skb);
2380 sent_pkts += tcp_skb_pcount(skb);
2386 if (is_rwnd_limited)
2387 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2389 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2391 if (likely(sent_pkts)) {
2392 if (tcp_in_cwnd_reduction(sk))
2393 tp->prr_out += sent_pkts;
2395 /* Send one loss probe per tail loss episode. */
2397 tcp_schedule_loss_probe(sk, false);
2398 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2399 tcp_cwnd_validate(sk, is_cwnd_limited);
2402 return !tp->packets_out && !tcp_write_queue_empty(sk);
2405 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2407 struct inet_connection_sock *icsk = inet_csk(sk);
2408 struct tcp_sock *tp = tcp_sk(sk);
2409 u32 timeout, rto_delta_us;
2412 /* Don't do any loss probe on a Fast Open connection before 3WHS
2415 if (tp->fastopen_rsk)
2418 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2419 /* Schedule a loss probe in 2*RTT for SACK capable connections
2420 * not in loss recovery, that are either limited by cwnd or application.
2422 if ((early_retrans != 3 && early_retrans != 4) ||
2423 !tp->packets_out || !tcp_is_sack(tp) ||
2424 (icsk->icsk_ca_state != TCP_CA_Open &&
2425 icsk->icsk_ca_state != TCP_CA_CWR))
2428 /* Probe timeout is 2*rtt. Add minimum RTO to account
2429 * for delayed ack when there's one outstanding packet. If no RTT
2430 * sample is available then probe after TCP_TIMEOUT_INIT.
2433 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2434 if (tp->packets_out == 1)
2435 timeout += TCP_RTO_MIN;
2437 timeout += TCP_TIMEOUT_MIN;
2439 timeout = TCP_TIMEOUT_INIT;
2442 /* If the RTO formula yields an earlier time, then use that time. */
2443 rto_delta_us = advancing_rto ?
2444 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2445 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2446 if (rto_delta_us > 0)
2447 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2449 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2454 /* Thanks to skb fast clones, we can detect if a prior transmit of
2455 * a packet is still in a qdisc or driver queue.
2456 * In this case, there is very little point doing a retransmit !
2458 static bool skb_still_in_host_queue(const struct sock *sk,
2459 const struct sk_buff *skb)
2461 if (unlikely(skb_fclone_busy(sk, skb))) {
2462 NET_INC_STATS(sock_net(sk),
2463 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2469 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2470 * retransmit the last segment.
2472 void tcp_send_loss_probe(struct sock *sk)
2474 struct tcp_sock *tp = tcp_sk(sk);
2475 struct sk_buff *skb;
2477 int mss = tcp_current_mss(sk);
2479 skb = tcp_send_head(sk);
2480 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2481 pcount = tp->packets_out;
2482 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2483 if (tp->packets_out > pcount)
2487 skb = skb_rb_last(&sk->tcp_rtx_queue);
2489 /* At most one outstanding TLP retransmission. */
2490 if (tp->tlp_high_seq)
2493 /* Retransmit last segment. */
2497 if (skb_still_in_host_queue(sk, skb))
2500 pcount = tcp_skb_pcount(skb);
2501 if (WARN_ON(!pcount))
2504 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2505 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2506 (pcount - 1) * mss, mss,
2509 skb = skb_rb_next(skb);
2512 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2515 if (__tcp_retransmit_skb(sk, skb, 1))
2518 /* Record snd_nxt for loss detection. */
2519 tp->tlp_high_seq = tp->snd_nxt;
2522 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2523 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2524 inet_csk(sk)->icsk_pending = 0;
2529 /* Push out any pending frames which were held back due to
2530 * TCP_CORK or attempt at coalescing tiny packets.
2531 * The socket must be locked by the caller.
2533 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2536 /* If we are closed, the bytes will have to remain here.
2537 * In time closedown will finish, we empty the write queue and
2538 * all will be happy.
2540 if (unlikely(sk->sk_state == TCP_CLOSE))
2543 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2544 sk_gfp_mask(sk, GFP_ATOMIC)))
2545 tcp_check_probe_timer(sk);
2548 /* Send _single_ skb sitting at the send head. This function requires
2549 * true push pending frames to setup probe timer etc.
2551 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2553 struct sk_buff *skb = tcp_send_head(sk);
2555 BUG_ON(!skb || skb->len < mss_now);
2557 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2560 /* This function returns the amount that we can raise the
2561 * usable window based on the following constraints
2563 * 1. The window can never be shrunk once it is offered (RFC 793)
2564 * 2. We limit memory per socket
2567 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2568 * RECV.NEXT + RCV.WIN fixed until:
2569 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2571 * i.e. don't raise the right edge of the window until you can raise
2572 * it at least MSS bytes.
2574 * Unfortunately, the recommended algorithm breaks header prediction,
2575 * since header prediction assumes th->window stays fixed.
2577 * Strictly speaking, keeping th->window fixed violates the receiver
2578 * side SWS prevention criteria. The problem is that under this rule
2579 * a stream of single byte packets will cause the right side of the
2580 * window to always advance by a single byte.
2582 * Of course, if the sender implements sender side SWS prevention
2583 * then this will not be a problem.
2585 * BSD seems to make the following compromise:
2587 * If the free space is less than the 1/4 of the maximum
2588 * space available and the free space is less than 1/2 mss,
2589 * then set the window to 0.
2590 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2591 * Otherwise, just prevent the window from shrinking
2592 * and from being larger than the largest representable value.
2594 * This prevents incremental opening of the window in the regime
2595 * where TCP is limited by the speed of the reader side taking
2596 * data out of the TCP receive queue. It does nothing about
2597 * those cases where the window is constrained on the sender side
2598 * because the pipeline is full.
2600 * BSD also seems to "accidentally" limit itself to windows that are a
2601 * multiple of MSS, at least until the free space gets quite small.
2602 * This would appear to be a side effect of the mbuf implementation.
2603 * Combining these two algorithms results in the observed behavior
2604 * of having a fixed window size at almost all times.
2606 * Below we obtain similar behavior by forcing the offered window to
2607 * a multiple of the mss when it is feasible to do so.
2609 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2610 * Regular options like TIMESTAMP are taken into account.
2612 u32 __tcp_select_window(struct sock *sk)
2614 struct inet_connection_sock *icsk = inet_csk(sk);
2615 struct tcp_sock *tp = tcp_sk(sk);
2616 /* MSS for the peer's data. Previous versions used mss_clamp
2617 * here. I don't know if the value based on our guesses
2618 * of peer's MSS is better for the performance. It's more correct
2619 * but may be worse for the performance because of rcv_mss
2620 * fluctuations. --SAW 1998/11/1
2622 int mss = icsk->icsk_ack.rcv_mss;
2623 int free_space = tcp_space(sk);
2624 int allowed_space = tcp_full_space(sk);
2625 int full_space = min_t(int, tp->window_clamp, allowed_space);
2628 if (unlikely(mss > full_space)) {
2633 if (free_space < (full_space >> 1)) {
2634 icsk->icsk_ack.quick = 0;
2636 if (tcp_under_memory_pressure(sk))
2637 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2640 /* free_space might become our new window, make sure we don't
2641 * increase it due to wscale.
2643 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2645 /* if free space is less than mss estimate, or is below 1/16th
2646 * of the maximum allowed, try to move to zero-window, else
2647 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2648 * new incoming data is dropped due to memory limits.
2649 * With large window, mss test triggers way too late in order
2650 * to announce zero window in time before rmem limit kicks in.
2652 if (free_space < (allowed_space >> 4) || free_space < mss)
2656 if (free_space > tp->rcv_ssthresh)
2657 free_space = tp->rcv_ssthresh;
2659 /* Don't do rounding if we are using window scaling, since the
2660 * scaled window will not line up with the MSS boundary anyway.
2662 if (tp->rx_opt.rcv_wscale) {
2663 window = free_space;
2665 /* Advertise enough space so that it won't get scaled away.
2666 * Import case: prevent zero window announcement if
2667 * 1<<rcv_wscale > mss.
2669 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2671 window = tp->rcv_wnd;
2672 /* Get the largest window that is a nice multiple of mss.
2673 * Window clamp already applied above.
2674 * If our current window offering is within 1 mss of the
2675 * free space we just keep it. This prevents the divide
2676 * and multiply from happening most of the time.
2677 * We also don't do any window rounding when the free space
2680 if (window <= free_space - mss || window > free_space)
2681 window = rounddown(free_space, mss);
2682 else if (mss == full_space &&
2683 free_space > window + (full_space >> 1))
2684 window = free_space;
2690 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2691 const struct sk_buff *next_skb)
2693 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2694 const struct skb_shared_info *next_shinfo =
2695 skb_shinfo(next_skb);
2696 struct skb_shared_info *shinfo = skb_shinfo(skb);
2698 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2699 shinfo->tskey = next_shinfo->tskey;
2700 TCP_SKB_CB(skb)->txstamp_ack |=
2701 TCP_SKB_CB(next_skb)->txstamp_ack;
2705 /* Collapses two adjacent SKB's during retransmission. */
2706 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2708 struct tcp_sock *tp = tcp_sk(sk);
2709 struct sk_buff *next_skb = skb_rb_next(skb);
2710 int skb_size, next_skb_size;
2712 skb_size = skb->len;
2713 next_skb_size = next_skb->len;
2715 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2717 if (next_skb_size) {
2718 if (next_skb_size <= skb_availroom(skb))
2719 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2721 else if (!skb_shift(skb, next_skb, next_skb_size))
2724 tcp_highest_sack_replace(sk, next_skb, skb);
2726 /* Update sequence range on original skb. */
2727 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2729 /* Merge over control information. This moves PSH/FIN etc. over */
2730 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2732 /* All done, get rid of second SKB and account for it so
2733 * packet counting does not break.
2735 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2736 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2738 /* changed transmit queue under us so clear hints */
2739 tcp_clear_retrans_hints_partial(tp);
2740 if (next_skb == tp->retransmit_skb_hint)
2741 tp->retransmit_skb_hint = skb;
2743 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2745 tcp_skb_collapse_tstamp(skb, next_skb);
2747 tcp_rtx_queue_unlink_and_free(next_skb, sk);
2751 /* Check if coalescing SKBs is legal. */
2752 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2754 if (tcp_skb_pcount(skb) > 1)
2756 if (skb_cloned(skb))
2758 /* Some heuristics for collapsing over SACK'd could be invented */
2759 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2765 /* Collapse packets in the retransmit queue to make to create
2766 * less packets on the wire. This is only done on retransmission.
2768 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2771 struct tcp_sock *tp = tcp_sk(sk);
2772 struct sk_buff *skb = to, *tmp;
2775 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2777 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2780 skb_rbtree_walk_from_safe(skb, tmp) {
2781 if (!tcp_can_collapse(sk, skb))
2784 if (!tcp_skb_can_collapse_to(to))
2797 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2800 if (!tcp_collapse_retrans(sk, to))
2805 /* This retransmits one SKB. Policy decisions and retransmit queue
2806 * state updates are done by the caller. Returns non-zero if an
2807 * error occurred which prevented the send.
2809 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2811 struct inet_connection_sock *icsk = inet_csk(sk);
2812 struct tcp_sock *tp = tcp_sk(sk);
2813 unsigned int cur_mss;
2817 /* Inconclusive MTU probe */
2818 if (icsk->icsk_mtup.probe_size)
2819 icsk->icsk_mtup.probe_size = 0;
2821 /* Do not sent more than we queued. 1/4 is reserved for possible
2822 * copying overhead: fragmentation, tunneling, mangling etc.
2824 if (refcount_read(&sk->sk_wmem_alloc) >
2825 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2829 if (skb_still_in_host_queue(sk, skb))
2832 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2833 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2837 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2841 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2842 return -EHOSTUNREACH; /* Routing failure or similar. */
2844 cur_mss = tcp_current_mss(sk);
2846 /* If receiver has shrunk his window, and skb is out of
2847 * new window, do not retransmit it. The exception is the
2848 * case, when window is shrunk to zero. In this case
2849 * our retransmit serves as a zero window probe.
2851 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2852 TCP_SKB_CB(skb)->seq != tp->snd_una)
2855 len = cur_mss * segs;
2856 if (skb->len > len) {
2857 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2858 cur_mss, GFP_ATOMIC))
2859 return -ENOMEM; /* We'll try again later. */
2861 if (skb_unclone(skb, GFP_ATOMIC))
2864 diff = tcp_skb_pcount(skb);
2865 tcp_set_skb_tso_segs(skb, cur_mss);
2866 diff -= tcp_skb_pcount(skb);
2868 tcp_adjust_pcount(sk, skb, diff);
2869 if (skb->len < cur_mss)
2870 tcp_retrans_try_collapse(sk, skb, cur_mss);
2873 /* RFC3168, section 6.1.1.1. ECN fallback */
2874 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2875 tcp_ecn_clear_syn(sk, skb);
2877 /* Update global and local TCP statistics. */
2878 segs = tcp_skb_pcount(skb);
2879 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2880 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2881 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2882 tp->total_retrans += segs;
2884 /* make sure skb->data is aligned on arches that require it
2885 * and check if ack-trimming & collapsing extended the headroom
2886 * beyond what csum_start can cover.
2888 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2889 skb_headroom(skb) >= 0xFFFF)) {
2890 struct sk_buff *nskb;
2892 tcp_skb_tsorted_save(skb) {
2893 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2894 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2896 } tcp_skb_tsorted_restore(skb);
2899 tcp_update_skb_after_send(tp, skb);
2900 tcp_rate_skb_sent(sk, skb);
2903 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2906 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
2907 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
2908 TCP_SKB_CB(skb)->seq, segs, err);
2911 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2912 trace_tcp_retransmit_skb(sk, skb);
2913 } else if (err != -EBUSY) {
2914 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2919 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2921 struct tcp_sock *tp = tcp_sk(sk);
2922 int err = __tcp_retransmit_skb(sk, skb, segs);
2925 #if FASTRETRANS_DEBUG > 0
2926 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2927 net_dbg_ratelimited("retrans_out leaked\n");
2930 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2931 tp->retrans_out += tcp_skb_pcount(skb);
2933 /* Save stamp of the first retransmit. */
2934 if (!tp->retrans_stamp)
2935 tp->retrans_stamp = tcp_skb_timestamp(skb);
2939 if (tp->undo_retrans < 0)
2940 tp->undo_retrans = 0;
2941 tp->undo_retrans += tcp_skb_pcount(skb);
2945 /* This gets called after a retransmit timeout, and the initially
2946 * retransmitted data is acknowledged. It tries to continue
2947 * resending the rest of the retransmit queue, until either
2948 * we've sent it all or the congestion window limit is reached.
2950 void tcp_xmit_retransmit_queue(struct sock *sk)
2952 const struct inet_connection_sock *icsk = inet_csk(sk);
2953 struct sk_buff *skb, *rtx_head, *hole = NULL;
2954 struct tcp_sock *tp = tcp_sk(sk);
2958 if (!tp->packets_out)
2961 rtx_head = tcp_rtx_queue_head(sk);
2962 skb = tp->retransmit_skb_hint ?: rtx_head;
2963 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2964 skb_rbtree_walk_from(skb) {
2968 if (tcp_pacing_check(sk))
2971 /* we could do better than to assign each time */
2973 tp->retransmit_skb_hint = skb;
2975 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2978 sacked = TCP_SKB_CB(skb)->sacked;
2979 /* In case tcp_shift_skb_data() have aggregated large skbs,
2980 * we need to make sure not sending too bigs TSO packets
2982 segs = min_t(int, segs, max_segs);
2984 if (tp->retrans_out >= tp->lost_out) {
2986 } else if (!(sacked & TCPCB_LOST)) {
2987 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2992 if (icsk->icsk_ca_state != TCP_CA_Loss)
2993 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2995 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2998 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3001 if (tcp_small_queue_check(sk, skb, 1))
3004 if (tcp_retransmit_skb(sk, skb, segs))
3007 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3009 if (tcp_in_cwnd_reduction(sk))
3010 tp->prr_out += tcp_skb_pcount(skb);
3012 if (skb == rtx_head &&
3013 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3014 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3015 inet_csk(sk)->icsk_rto,
3020 /* We allow to exceed memory limits for FIN packets to expedite
3021 * connection tear down and (memory) recovery.
3022 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3023 * or even be forced to close flow without any FIN.
3024 * In general, we want to allow one skb per socket to avoid hangs
3025 * with edge trigger epoll()
3027 void sk_forced_mem_schedule(struct sock *sk, int size)
3031 if (size <= sk->sk_forward_alloc)
3033 amt = sk_mem_pages(size);
3034 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3035 sk_memory_allocated_add(sk, amt);
3037 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3038 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3041 /* Send a FIN. The caller locks the socket for us.
3042 * We should try to send a FIN packet really hard, but eventually give up.
3044 void tcp_send_fin(struct sock *sk)
3046 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3047 struct tcp_sock *tp = tcp_sk(sk);
3049 /* Optimization, tack on the FIN if we have one skb in write queue and
3050 * this skb was not yet sent, or we are under memory pressure.
3051 * Note: in the latter case, FIN packet will be sent after a timeout,
3052 * as TCP stack thinks it has already been transmitted.
3054 if (!tskb && tcp_under_memory_pressure(sk))
3055 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3059 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3060 TCP_SKB_CB(tskb)->end_seq++;
3062 if (tcp_write_queue_empty(sk)) {
3063 /* This means tskb was already sent.
3064 * Pretend we included the FIN on previous transmit.
3065 * We need to set tp->snd_nxt to the value it would have
3066 * if FIN had been sent. This is because retransmit path
3067 * does not change tp->snd_nxt.
3073 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3074 if (unlikely(!skb)) {
3079 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3080 skb_reserve(skb, MAX_TCP_HEADER);
3081 sk_forced_mem_schedule(sk, skb->truesize);
3082 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3083 tcp_init_nondata_skb(skb, tp->write_seq,
3084 TCPHDR_ACK | TCPHDR_FIN);
3085 tcp_queue_skb(sk, skb);
3087 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3090 /* We get here when a process closes a file descriptor (either due to
3091 * an explicit close() or as a byproduct of exit()'ing) and there
3092 * was unread data in the receive queue. This behavior is recommended
3093 * by RFC 2525, section 2.17. -DaveM
3095 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3097 struct sk_buff *skb;
3099 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3101 /* NOTE: No TCP options attached and we never retransmit this. */
3102 skb = alloc_skb(MAX_TCP_HEADER, priority);
3104 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3108 /* Reserve space for headers and prepare control bits. */
3109 skb_reserve(skb, MAX_TCP_HEADER);
3110 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3111 TCPHDR_ACK | TCPHDR_RST);
3112 tcp_mstamp_refresh(tcp_sk(sk));
3114 if (tcp_transmit_skb(sk, skb, 0, priority))
3115 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3117 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3118 * skb here is different to the troublesome skb, so use NULL
3120 trace_tcp_send_reset(sk, NULL);
3123 /* Send a crossed SYN-ACK during socket establishment.
3124 * WARNING: This routine must only be called when we have already sent
3125 * a SYN packet that crossed the incoming SYN that caused this routine
3126 * to get called. If this assumption fails then the initial rcv_wnd
3127 * and rcv_wscale values will not be correct.
3129 int tcp_send_synack(struct sock *sk)
3131 struct sk_buff *skb;
3133 skb = tcp_rtx_queue_head(sk);
3134 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3135 pr_err("%s: wrong queue state\n", __func__);
3138 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3139 if (skb_cloned(skb)) {
3140 struct sk_buff *nskb;
3142 tcp_skb_tsorted_save(skb) {
3143 nskb = skb_copy(skb, GFP_ATOMIC);
3144 } tcp_skb_tsorted_restore(skb);
3147 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3148 tcp_rtx_queue_unlink_and_free(skb, sk);
3149 __skb_header_release(nskb);
3150 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3151 sk->sk_wmem_queued += nskb->truesize;
3152 sk_mem_charge(sk, nskb->truesize);
3156 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3157 tcp_ecn_send_synack(sk, skb);
3159 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3163 * tcp_make_synack - Prepare a SYN-ACK.
3164 * sk: listener socket
3165 * dst: dst entry attached to the SYNACK
3166 * req: request_sock pointer
3168 * Allocate one skb and build a SYNACK packet.
3169 * @dst is consumed : Caller should not use it again.
3171 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3172 struct request_sock *req,
3173 struct tcp_fastopen_cookie *foc,
3174 enum tcp_synack_type synack_type)
3176 struct inet_request_sock *ireq = inet_rsk(req);
3177 const struct tcp_sock *tp = tcp_sk(sk);
3178 struct tcp_md5sig_key *md5 = NULL;
3179 struct tcp_out_options opts;
3180 struct sk_buff *skb;
3181 int tcp_header_size;
3185 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3186 if (unlikely(!skb)) {
3190 /* Reserve space for headers. */
3191 skb_reserve(skb, MAX_TCP_HEADER);
3193 switch (synack_type) {
3194 case TCP_SYNACK_NORMAL:
3195 skb_set_owner_w(skb, req_to_sk(req));
3197 case TCP_SYNACK_COOKIE:
3198 /* Under synflood, we do not attach skb to a socket,
3199 * to avoid false sharing.
3202 case TCP_SYNACK_FASTOPEN:
3203 /* sk is a const pointer, because we want to express multiple
3204 * cpu might call us concurrently.
3205 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3207 skb_set_owner_w(skb, (struct sock *)sk);
3210 skb_dst_set(skb, dst);
3212 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3214 memset(&opts, 0, sizeof(opts));
3215 #ifdef CONFIG_SYN_COOKIES
3216 if (unlikely(req->cookie_ts))
3217 skb->skb_mstamp = cookie_init_timestamp(req);
3220 skb->skb_mstamp = tcp_clock_us();
3222 #ifdef CONFIG_TCP_MD5SIG
3224 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3226 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3227 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3230 skb_push(skb, tcp_header_size);
3231 skb_reset_transport_header(skb);
3233 th = (struct tcphdr *)skb->data;
3234 memset(th, 0, sizeof(struct tcphdr));
3237 tcp_ecn_make_synack(req, th);
3238 th->source = htons(ireq->ir_num);
3239 th->dest = ireq->ir_rmt_port;
3240 skb->mark = ireq->ir_mark;
3241 skb->ip_summed = CHECKSUM_PARTIAL;
3242 th->seq = htonl(tcp_rsk(req)->snt_isn);
3243 /* XXX data is queued and acked as is. No buffer/window check */
3244 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3246 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3247 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3248 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3249 th->doff = (tcp_header_size >> 2);
3250 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3252 #ifdef CONFIG_TCP_MD5SIG
3253 /* Okay, we have all we need - do the md5 hash if needed */
3255 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3256 md5, req_to_sk(req), skb);
3260 /* Do not fool tcpdump (if any), clean our debris */
3264 EXPORT_SYMBOL(tcp_make_synack);
3266 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3268 struct inet_connection_sock *icsk = inet_csk(sk);
3269 const struct tcp_congestion_ops *ca;
3270 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3272 if (ca_key == TCP_CA_UNSPEC)
3276 ca = tcp_ca_find_key(ca_key);
3277 if (likely(ca && try_module_get(ca->owner))) {
3278 module_put(icsk->icsk_ca_ops->owner);
3279 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3280 icsk->icsk_ca_ops = ca;
3285 /* Do all connect socket setups that can be done AF independent. */
3286 static void tcp_connect_init(struct sock *sk)
3288 const struct dst_entry *dst = __sk_dst_get(sk);
3289 struct tcp_sock *tp = tcp_sk(sk);
3293 /* We'll fix this up when we get a response from the other end.
3294 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3296 tp->tcp_header_len = sizeof(struct tcphdr);
3297 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3298 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3300 #ifdef CONFIG_TCP_MD5SIG
3301 if (tp->af_specific->md5_lookup(sk, sk))
3302 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3305 /* If user gave his TCP_MAXSEG, record it to clamp */
3306 if (tp->rx_opt.user_mss)
3307 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3310 tcp_sync_mss(sk, dst_mtu(dst));
3312 tcp_ca_dst_init(sk, dst);
3314 if (!tp->window_clamp)
3315 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3316 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3318 tcp_initialize_rcv_mss(sk);
3320 /* limit the window selection if the user enforce a smaller rx buffer */
3321 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3322 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3323 tp->window_clamp = tcp_full_space(sk);
3325 rcv_wnd = tcp_rwnd_init_bpf(sk);
3327 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3329 tcp_select_initial_window(sk, tcp_full_space(sk),
3330 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3333 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3337 tp->rx_opt.rcv_wscale = rcv_wscale;
3338 tp->rcv_ssthresh = tp->rcv_wnd;
3341 sock_reset_flag(sk, SOCK_DONE);
3344 tcp_write_queue_purge(sk);
3345 tp->snd_una = tp->write_seq;
3346 tp->snd_sml = tp->write_seq;
3347 tp->snd_up = tp->write_seq;
3348 tp->snd_nxt = tp->write_seq;
3350 if (likely(!tp->repair))
3353 tp->rcv_tstamp = tcp_jiffies32;
3354 tp->rcv_wup = tp->rcv_nxt;
3355 tp->copied_seq = tp->rcv_nxt;
3357 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3358 inet_csk(sk)->icsk_retransmits = 0;
3359 tcp_clear_retrans(tp);
3362 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3364 struct tcp_sock *tp = tcp_sk(sk);
3365 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3367 tcb->end_seq += skb->len;
3368 __skb_header_release(skb);
3369 sk->sk_wmem_queued += skb->truesize;
3370 sk_mem_charge(sk, skb->truesize);
3371 tp->write_seq = tcb->end_seq;
3372 tp->packets_out += tcp_skb_pcount(skb);
3375 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3376 * queue a data-only packet after the regular SYN, such that regular SYNs
3377 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3378 * only the SYN sequence, the data are retransmitted in the first ACK.
3379 * If cookie is not cached or other error occurs, falls back to send a
3380 * regular SYN with Fast Open cookie request option.
3382 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3384 struct tcp_sock *tp = tcp_sk(sk);
3385 struct tcp_fastopen_request *fo = tp->fastopen_req;
3387 struct sk_buff *syn_data;
3389 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3390 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3393 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3394 * user-MSS. Reserve maximum option space for middleboxes that add
3395 * private TCP options. The cost is reduced data space in SYN :(
3397 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3399 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3400 MAX_TCP_OPTION_SPACE;
3402 space = min_t(size_t, space, fo->size);
3404 /* limit to order-0 allocations */
3405 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3407 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3410 syn_data->ip_summed = CHECKSUM_PARTIAL;
3411 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3413 int copied = copy_from_iter(skb_put(syn_data, space), space,
3414 &fo->data->msg_iter);
3415 if (unlikely(!copied)) {
3416 tcp_skb_tsorted_anchor_cleanup(syn_data);
3417 kfree_skb(syn_data);
3420 if (copied != space) {
3421 skb_trim(syn_data, copied);
3425 /* No more data pending in inet_wait_for_connect() */
3426 if (space == fo->size)
3430 tcp_connect_queue_skb(sk, syn_data);
3432 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3434 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3436 syn->skb_mstamp = syn_data->skb_mstamp;
3438 /* Now full SYN+DATA was cloned and sent (or not),
3439 * remove the SYN from the original skb (syn_data)
3440 * we keep in write queue in case of a retransmit, as we
3441 * also have the SYN packet (with no data) in the same queue.
3443 TCP_SKB_CB(syn_data)->seq++;
3444 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3446 tp->syn_data = (fo->copied > 0);
3447 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3448 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3452 /* data was not sent, put it in write_queue */
3453 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3454 tp->packets_out -= tcp_skb_pcount(syn_data);
3457 /* Send a regular SYN with Fast Open cookie request option */
3458 if (fo->cookie.len > 0)
3460 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3462 tp->syn_fastopen = 0;
3464 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3468 /* Build a SYN and send it off. */
3469 int tcp_connect(struct sock *sk)
3471 struct tcp_sock *tp = tcp_sk(sk);
3472 struct sk_buff *buff;
3475 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3477 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3478 return -EHOSTUNREACH; /* Routing failure or similar. */
3480 tcp_connect_init(sk);
3482 if (unlikely(tp->repair)) {
3483 tcp_finish_connect(sk, NULL);
3487 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3488 if (unlikely(!buff))
3491 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3492 tcp_mstamp_refresh(tp);
3493 tp->retrans_stamp = tcp_time_stamp(tp);
3494 tcp_connect_queue_skb(sk, buff);
3495 tcp_ecn_send_syn(sk, buff);
3496 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3498 /* Send off SYN; include data in Fast Open. */
3499 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3500 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3501 if (err == -ECONNREFUSED)
3504 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3505 * in order to make this packet get counted in tcpOutSegs.
3507 tp->snd_nxt = tp->write_seq;
3508 tp->pushed_seq = tp->write_seq;
3509 buff = tcp_send_head(sk);
3510 if (unlikely(buff)) {
3511 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3512 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3514 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3516 /* Timer for repeating the SYN until an answer. */
3517 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3518 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3521 EXPORT_SYMBOL(tcp_connect);
3523 /* Send out a delayed ack, the caller does the policy checking
3524 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3527 void tcp_send_delayed_ack(struct sock *sk)
3529 struct inet_connection_sock *icsk = inet_csk(sk);
3530 int ato = icsk->icsk_ack.ato;
3531 unsigned long timeout;
3533 if (ato > TCP_DELACK_MIN) {
3534 const struct tcp_sock *tp = tcp_sk(sk);
3535 int max_ato = HZ / 2;
3537 if (icsk->icsk_ack.pingpong ||
3538 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3539 max_ato = TCP_DELACK_MAX;
3541 /* Slow path, intersegment interval is "high". */
3543 /* If some rtt estimate is known, use it to bound delayed ack.
3544 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3548 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3555 ato = min(ato, max_ato);
3558 /* Stay within the limit we were given */
3559 timeout = jiffies + ato;
3561 /* Use new timeout only if there wasn't a older one earlier. */
3562 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3563 /* If delack timer was blocked or is about to expire,
3566 if (icsk->icsk_ack.blocked ||
3567 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3572 if (!time_before(timeout, icsk->icsk_ack.timeout))
3573 timeout = icsk->icsk_ack.timeout;
3575 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3576 icsk->icsk_ack.timeout = timeout;
3577 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3580 /* This routine sends an ack and also updates the window. */
3581 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3583 struct sk_buff *buff;
3585 /* If we have been reset, we may not send again. */
3586 if (sk->sk_state == TCP_CLOSE)
3589 /* We are not putting this on the write queue, so
3590 * tcp_transmit_skb() will set the ownership to this
3593 buff = alloc_skb(MAX_TCP_HEADER,
3594 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3595 if (unlikely(!buff)) {
3596 inet_csk_schedule_ack(sk);
3597 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3598 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3599 TCP_DELACK_MAX, TCP_RTO_MAX);
3603 /* Reserve space for headers and prepare control bits. */
3604 skb_reserve(buff, MAX_TCP_HEADER);
3605 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3607 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3609 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3611 skb_set_tcp_pure_ack(buff);
3613 /* Send it off, this clears delayed acks for us. */
3614 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3617 void tcp_send_ack(struct sock *sk)
3619 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3621 EXPORT_SYMBOL_GPL(tcp_send_ack);
3623 /* This routine sends a packet with an out of date sequence
3624 * number. It assumes the other end will try to ack it.
3626 * Question: what should we make while urgent mode?
3627 * 4.4BSD forces sending single byte of data. We cannot send
3628 * out of window data, because we have SND.NXT==SND.MAX...
3630 * Current solution: to send TWO zero-length segments in urgent mode:
3631 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3632 * out-of-date with SND.UNA-1 to probe window.
3634 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3636 struct tcp_sock *tp = tcp_sk(sk);
3637 struct sk_buff *skb;
3639 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3640 skb = alloc_skb(MAX_TCP_HEADER,
3641 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3645 /* Reserve space for headers and set control bits. */
3646 skb_reserve(skb, MAX_TCP_HEADER);
3647 /* Use a previous sequence. This should cause the other
3648 * end to send an ack. Don't queue or clone SKB, just
3651 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3652 NET_INC_STATS(sock_net(sk), mib);
3653 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3656 /* Called from setsockopt( ... TCP_REPAIR ) */
3657 void tcp_send_window_probe(struct sock *sk)
3659 if (sk->sk_state == TCP_ESTABLISHED) {
3660 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3661 tcp_mstamp_refresh(tcp_sk(sk));
3662 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3666 /* Initiate keepalive or window probe from timer. */
3667 int tcp_write_wakeup(struct sock *sk, int mib)
3669 struct tcp_sock *tp = tcp_sk(sk);
3670 struct sk_buff *skb;
3672 if (sk->sk_state == TCP_CLOSE)
3675 skb = tcp_send_head(sk);
3676 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3678 unsigned int mss = tcp_current_mss(sk);
3679 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3681 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3682 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3684 /* We are probing the opening of a window
3685 * but the window size is != 0
3686 * must have been a result SWS avoidance ( sender )
3688 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3690 seg_size = min(seg_size, mss);
3691 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3692 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3693 skb, seg_size, mss, GFP_ATOMIC))
3695 } else if (!tcp_skb_pcount(skb))
3696 tcp_set_skb_tso_segs(skb, mss);
3698 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3699 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3701 tcp_event_new_data_sent(sk, skb);
3704 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3705 tcp_xmit_probe_skb(sk, 1, mib);
3706 return tcp_xmit_probe_skb(sk, 0, mib);
3710 /* A window probe timeout has occurred. If window is not closed send
3711 * a partial packet else a zero probe.
3713 void tcp_send_probe0(struct sock *sk)
3715 struct inet_connection_sock *icsk = inet_csk(sk);
3716 struct tcp_sock *tp = tcp_sk(sk);
3717 struct net *net = sock_net(sk);
3718 unsigned long probe_max;
3721 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3723 if (tp->packets_out || tcp_write_queue_empty(sk)) {
3724 /* Cancel probe timer, if it is not required. */
3725 icsk->icsk_probes_out = 0;
3726 icsk->icsk_backoff = 0;
3731 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3732 icsk->icsk_backoff++;
3733 icsk->icsk_probes_out++;
3734 probe_max = TCP_RTO_MAX;
3736 /* If packet was not sent due to local congestion,
3737 * do not backoff and do not remember icsk_probes_out.
3738 * Let local senders to fight for local resources.
3740 * Use accumulated backoff yet.
3742 if (!icsk->icsk_probes_out)
3743 icsk->icsk_probes_out = 1;
3744 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3746 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3747 tcp_probe0_when(sk, probe_max),
3751 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3753 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3757 tcp_rsk(req)->txhash = net_tx_rndhash();
3758 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3760 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3761 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3762 if (unlikely(tcp_passive_fastopen(sk)))
3763 tcp_sk(sk)->total_retrans++;
3764 trace_tcp_retransmit_synack(sk, req);
3768 EXPORT_SYMBOL(tcp_rtx_synack);