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 tp->compressed_ack = 0;
169 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
172 tcp_dec_quickack_mode(sk, pkts);
173 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
177 u32 tcp_default_init_rwnd(u32 mss)
179 /* Initial receive window should be twice of TCP_INIT_CWND to
180 * enable proper sending of new unsent data during fast recovery
181 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
182 * limit when mss is larger than 1460.
184 u32 init_rwnd = TCP_INIT_CWND * 2;
187 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
191 /* Determine a window scaling and initial window to offer.
192 * Based on the assumption that the given amount of space
193 * will be offered. Store the results in the tp structure.
194 * NOTE: for smooth operation initial space offering should
195 * be a multiple of mss if possible. We assume here that mss >= 1.
196 * This MUST be enforced by all callers.
198 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
199 __u32 *rcv_wnd, __u32 *window_clamp,
200 int wscale_ok, __u8 *rcv_wscale,
203 unsigned int space = (__space < 0 ? 0 : __space);
205 /* If no clamp set the clamp to the max possible scaled window */
206 if (*window_clamp == 0)
207 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
208 space = min(*window_clamp, space);
210 /* Quantize space offering to a multiple of mss if possible. */
212 space = rounddown(space, mss);
214 /* NOTE: offering an initial window larger than 32767
215 * will break some buggy TCP stacks. If the admin tells us
216 * it is likely we could be speaking with such a buggy stack
217 * we will truncate our initial window offering to 32K-1
218 * unless the remote has sent us a window scaling option,
219 * which we interpret as a sign the remote TCP is not
220 * misinterpreting the window field as a signed quantity.
222 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
223 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
229 /* Set window scaling on max possible window */
230 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
231 space = max_t(u32, space, sysctl_rmem_max);
232 space = min_t(u32, space, *window_clamp);
233 while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
239 if (!init_rcv_wnd) /* Use default unless specified otherwise */
240 init_rcv_wnd = tcp_default_init_rwnd(mss);
241 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
243 /* Set the clamp no higher than max representable value */
244 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
246 EXPORT_SYMBOL(tcp_select_initial_window);
248 /* Chose a new window to advertise, update state in tcp_sock for the
249 * socket, and return result with RFC1323 scaling applied. The return
250 * value can be stuffed directly into th->window for an outgoing
253 static u16 tcp_select_window(struct sock *sk)
255 struct tcp_sock *tp = tcp_sk(sk);
256 u32 old_win = tp->rcv_wnd;
257 u32 cur_win = tcp_receive_window(tp);
258 u32 new_win = __tcp_select_window(sk);
260 /* Never shrink the offered window */
261 if (new_win < cur_win) {
262 /* Danger Will Robinson!
263 * Don't update rcv_wup/rcv_wnd here or else
264 * we will not be able to advertise a zero
265 * window in time. --DaveM
267 * Relax Will Robinson.
270 NET_INC_STATS(sock_net(sk),
271 LINUX_MIB_TCPWANTZEROWINDOWADV);
272 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
274 tp->rcv_wnd = new_win;
275 tp->rcv_wup = tp->rcv_nxt;
277 /* Make sure we do not exceed the maximum possible
280 if (!tp->rx_opt.rcv_wscale &&
281 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
282 new_win = min(new_win, MAX_TCP_WINDOW);
284 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
286 /* RFC1323 scaling applied */
287 new_win >>= tp->rx_opt.rcv_wscale;
289 /* If we advertise zero window, disable fast path. */
293 NET_INC_STATS(sock_net(sk),
294 LINUX_MIB_TCPTOZEROWINDOWADV);
295 } else if (old_win == 0) {
296 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
302 /* Packet ECN state for a SYN-ACK */
303 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
305 const struct tcp_sock *tp = tcp_sk(sk);
307 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
308 if (!(tp->ecn_flags & TCP_ECN_OK))
309 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
310 else if (tcp_ca_needs_ecn(sk) ||
311 tcp_bpf_ca_needs_ecn(sk))
315 /* Packet ECN state for a SYN. */
316 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
318 struct tcp_sock *tp = tcp_sk(sk);
319 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
320 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
321 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
324 const struct dst_entry *dst = __sk_dst_get(sk);
326 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
333 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
334 tp->ecn_flags = TCP_ECN_OK;
335 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
340 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
342 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
343 /* tp->ecn_flags are cleared at a later point in time when
344 * SYN ACK is ultimatively being received.
346 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
350 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
352 if (inet_rsk(req)->ecn_ok)
356 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
359 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
360 struct tcphdr *th, int tcp_header_len)
362 struct tcp_sock *tp = tcp_sk(sk);
364 if (tp->ecn_flags & TCP_ECN_OK) {
365 /* Not-retransmitted data segment: set ECT and inject CWR. */
366 if (skb->len != tcp_header_len &&
367 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
369 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
370 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
372 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
374 } else if (!tcp_ca_needs_ecn(sk)) {
375 /* ACK or retransmitted segment: clear ECT|CE */
376 INET_ECN_dontxmit(sk);
378 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
383 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
384 * auto increment end seqno.
386 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
388 skb->ip_summed = CHECKSUM_PARTIAL;
390 TCP_SKB_CB(skb)->tcp_flags = flags;
391 TCP_SKB_CB(skb)->sacked = 0;
393 tcp_skb_pcount_set(skb, 1);
395 TCP_SKB_CB(skb)->seq = seq;
396 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
398 TCP_SKB_CB(skb)->end_seq = seq;
401 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
403 return tp->snd_una != tp->snd_up;
406 #define OPTION_SACK_ADVERTISE (1 << 0)
407 #define OPTION_TS (1 << 1)
408 #define OPTION_MD5 (1 << 2)
409 #define OPTION_WSCALE (1 << 3)
410 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
411 #define OPTION_SMC (1 << 9)
413 static void smc_options_write(__be32 *ptr, u16 *options)
415 #if IS_ENABLED(CONFIG_SMC)
416 if (static_branch_unlikely(&tcp_have_smc)) {
417 if (unlikely(OPTION_SMC & *options)) {
418 *ptr++ = htonl((TCPOPT_NOP << 24) |
421 (TCPOLEN_EXP_SMC_BASE));
422 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
428 struct tcp_out_options {
429 u16 options; /* bit field of OPTION_* */
430 u16 mss; /* 0 to disable */
431 u8 ws; /* window scale, 0 to disable */
432 u8 num_sack_blocks; /* number of SACK blocks to include */
433 u8 hash_size; /* bytes in hash_location */
434 __u8 *hash_location; /* temporary pointer, overloaded */
435 __u32 tsval, tsecr; /* need to include OPTION_TS */
436 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
439 /* Write previously computed TCP options to the packet.
441 * Beware: Something in the Internet is very sensitive to the ordering of
442 * TCP options, we learned this through the hard way, so be careful here.
443 * Luckily we can at least blame others for their non-compliance but from
444 * inter-operability perspective it seems that we're somewhat stuck with
445 * the ordering which we have been using if we want to keep working with
446 * those broken things (not that it currently hurts anybody as there isn't
447 * particular reason why the ordering would need to be changed).
449 * At least SACK_PERM as the first option is known to lead to a disaster
450 * (but it may well be that other scenarios fail similarly).
452 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
453 struct tcp_out_options *opts)
455 u16 options = opts->options; /* mungable copy */
457 if (unlikely(OPTION_MD5 & options)) {
458 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
459 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
460 /* overload cookie hash location */
461 opts->hash_location = (__u8 *)ptr;
465 if (unlikely(opts->mss)) {
466 *ptr++ = htonl((TCPOPT_MSS << 24) |
467 (TCPOLEN_MSS << 16) |
471 if (likely(OPTION_TS & options)) {
472 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
473 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
474 (TCPOLEN_SACK_PERM << 16) |
475 (TCPOPT_TIMESTAMP << 8) |
477 options &= ~OPTION_SACK_ADVERTISE;
479 *ptr++ = htonl((TCPOPT_NOP << 24) |
481 (TCPOPT_TIMESTAMP << 8) |
484 *ptr++ = htonl(opts->tsval);
485 *ptr++ = htonl(opts->tsecr);
488 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
489 *ptr++ = htonl((TCPOPT_NOP << 24) |
491 (TCPOPT_SACK_PERM << 8) |
495 if (unlikely(OPTION_WSCALE & options)) {
496 *ptr++ = htonl((TCPOPT_NOP << 24) |
497 (TCPOPT_WINDOW << 16) |
498 (TCPOLEN_WINDOW << 8) |
502 if (unlikely(opts->num_sack_blocks)) {
503 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
504 tp->duplicate_sack : tp->selective_acks;
507 *ptr++ = htonl((TCPOPT_NOP << 24) |
510 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
511 TCPOLEN_SACK_PERBLOCK)));
513 for (this_sack = 0; this_sack < opts->num_sack_blocks;
515 *ptr++ = htonl(sp[this_sack].start_seq);
516 *ptr++ = htonl(sp[this_sack].end_seq);
519 tp->rx_opt.dsack = 0;
522 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
523 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
525 u32 len; /* Fast Open option length */
528 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
529 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
530 TCPOPT_FASTOPEN_MAGIC);
531 p += TCPOLEN_EXP_FASTOPEN_BASE;
533 len = TCPOLEN_FASTOPEN_BASE + foc->len;
534 *p++ = TCPOPT_FASTOPEN;
538 memcpy(p, foc->val, foc->len);
539 if ((len & 3) == 2) {
540 p[foc->len] = TCPOPT_NOP;
541 p[foc->len + 1] = TCPOPT_NOP;
543 ptr += (len + 3) >> 2;
546 smc_options_write(ptr, &options);
549 static void smc_set_option(const struct tcp_sock *tp,
550 struct tcp_out_options *opts,
551 unsigned int *remaining)
553 #if IS_ENABLED(CONFIG_SMC)
554 if (static_branch_unlikely(&tcp_have_smc)) {
556 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
557 opts->options |= OPTION_SMC;
558 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
565 static void smc_set_option_cond(const struct tcp_sock *tp,
566 const struct inet_request_sock *ireq,
567 struct tcp_out_options *opts,
568 unsigned int *remaining)
570 #if IS_ENABLED(CONFIG_SMC)
571 if (static_branch_unlikely(&tcp_have_smc)) {
572 if (tp->syn_smc && ireq->smc_ok) {
573 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
574 opts->options |= OPTION_SMC;
575 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
582 /* Compute TCP options for SYN packets. This is not the final
583 * network wire format yet.
585 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
586 struct tcp_out_options *opts,
587 struct tcp_md5sig_key **md5)
589 struct tcp_sock *tp = tcp_sk(sk);
590 unsigned int remaining = MAX_TCP_OPTION_SPACE;
591 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
594 #ifdef CONFIG_TCP_MD5SIG
595 if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
596 *md5 = tp->af_specific->md5_lookup(sk, sk);
598 opts->options |= OPTION_MD5;
599 remaining -= TCPOLEN_MD5SIG_ALIGNED;
604 /* We always get an MSS option. The option bytes which will be seen in
605 * normal data packets should timestamps be used, must be in the MSS
606 * advertised. But we subtract them from tp->mss_cache so that
607 * calculations in tcp_sendmsg are simpler etc. So account for this
608 * fact here if necessary. If we don't do this correctly, as a
609 * receiver we won't recognize data packets as being full sized when we
610 * should, and thus we won't abide by the delayed ACK rules correctly.
611 * SACKs don't matter, we never delay an ACK when we have any of those
613 opts->mss = tcp_advertise_mss(sk);
614 remaining -= TCPOLEN_MSS_ALIGNED;
616 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
617 opts->options |= OPTION_TS;
618 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
619 opts->tsecr = tp->rx_opt.ts_recent;
620 remaining -= TCPOLEN_TSTAMP_ALIGNED;
622 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
623 opts->ws = tp->rx_opt.rcv_wscale;
624 opts->options |= OPTION_WSCALE;
625 remaining -= TCPOLEN_WSCALE_ALIGNED;
627 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
628 opts->options |= OPTION_SACK_ADVERTISE;
629 if (unlikely(!(OPTION_TS & opts->options)))
630 remaining -= TCPOLEN_SACKPERM_ALIGNED;
633 if (fastopen && fastopen->cookie.len >= 0) {
634 u32 need = fastopen->cookie.len;
636 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
637 TCPOLEN_FASTOPEN_BASE;
638 need = (need + 3) & ~3U; /* Align to 32 bits */
639 if (remaining >= need) {
640 opts->options |= OPTION_FAST_OPEN_COOKIE;
641 opts->fastopen_cookie = &fastopen->cookie;
643 tp->syn_fastopen = 1;
644 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
648 smc_set_option(tp, opts, &remaining);
650 return MAX_TCP_OPTION_SPACE - remaining;
653 /* Set up TCP options for SYN-ACKs. */
654 static unsigned int tcp_synack_options(const struct sock *sk,
655 struct request_sock *req,
656 unsigned int mss, struct sk_buff *skb,
657 struct tcp_out_options *opts,
658 const struct tcp_md5sig_key *md5,
659 struct tcp_fastopen_cookie *foc)
661 struct inet_request_sock *ireq = inet_rsk(req);
662 unsigned int remaining = MAX_TCP_OPTION_SPACE;
664 #ifdef CONFIG_TCP_MD5SIG
666 opts->options |= OPTION_MD5;
667 remaining -= TCPOLEN_MD5SIG_ALIGNED;
669 /* We can't fit any SACK blocks in a packet with MD5 + TS
670 * options. There was discussion about disabling SACK
671 * rather than TS in order to fit in better with old,
672 * buggy kernels, but that was deemed to be unnecessary.
674 ireq->tstamp_ok &= !ireq->sack_ok;
678 /* We always send an MSS option. */
680 remaining -= TCPOLEN_MSS_ALIGNED;
682 if (likely(ireq->wscale_ok)) {
683 opts->ws = ireq->rcv_wscale;
684 opts->options |= OPTION_WSCALE;
685 remaining -= TCPOLEN_WSCALE_ALIGNED;
687 if (likely(ireq->tstamp_ok)) {
688 opts->options |= OPTION_TS;
689 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
690 opts->tsecr = req->ts_recent;
691 remaining -= TCPOLEN_TSTAMP_ALIGNED;
693 if (likely(ireq->sack_ok)) {
694 opts->options |= OPTION_SACK_ADVERTISE;
695 if (unlikely(!ireq->tstamp_ok))
696 remaining -= TCPOLEN_SACKPERM_ALIGNED;
698 if (foc != NULL && foc->len >= 0) {
701 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
702 TCPOLEN_FASTOPEN_BASE;
703 need = (need + 3) & ~3U; /* Align to 32 bits */
704 if (remaining >= need) {
705 opts->options |= OPTION_FAST_OPEN_COOKIE;
706 opts->fastopen_cookie = foc;
711 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
713 return MAX_TCP_OPTION_SPACE - remaining;
716 /* Compute TCP options for ESTABLISHED sockets. This is not the
717 * final wire format yet.
719 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
720 struct tcp_out_options *opts,
721 struct tcp_md5sig_key **md5)
723 struct tcp_sock *tp = tcp_sk(sk);
724 unsigned int size = 0;
725 unsigned int eff_sacks;
730 #ifdef CONFIG_TCP_MD5SIG
731 if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
732 *md5 = tp->af_specific->md5_lookup(sk, sk);
734 opts->options |= OPTION_MD5;
735 size += TCPOLEN_MD5SIG_ALIGNED;
740 if (likely(tp->rx_opt.tstamp_ok)) {
741 opts->options |= OPTION_TS;
742 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
743 opts->tsecr = tp->rx_opt.ts_recent;
744 size += TCPOLEN_TSTAMP_ALIGNED;
747 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
748 if (unlikely(eff_sacks)) {
749 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
750 opts->num_sack_blocks =
751 min_t(unsigned int, eff_sacks,
752 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
753 TCPOLEN_SACK_PERBLOCK);
754 size += TCPOLEN_SACK_BASE_ALIGNED +
755 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
762 /* TCP SMALL QUEUES (TSQ)
764 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
765 * to reduce RTT and bufferbloat.
766 * We do this using a special skb destructor (tcp_wfree).
768 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
769 * needs to be reallocated in a driver.
770 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
772 * Since transmit from skb destructor is forbidden, we use a tasklet
773 * to process all sockets that eventually need to send more skbs.
774 * We use one tasklet per cpu, with its own queue of sockets.
777 struct tasklet_struct tasklet;
778 struct list_head head; /* queue of tcp sockets */
780 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
782 static void tcp_tsq_write(struct sock *sk)
784 if ((1 << sk->sk_state) &
785 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
786 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
787 struct tcp_sock *tp = tcp_sk(sk);
789 if (tp->lost_out > tp->retrans_out &&
790 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
791 tcp_mstamp_refresh(tp);
792 tcp_xmit_retransmit_queue(sk);
795 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
800 static void tcp_tsq_handler(struct sock *sk)
803 if (!sock_owned_by_user(sk))
805 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
810 * One tasklet per cpu tries to send more skbs.
811 * We run in tasklet context but need to disable irqs when
812 * transferring tsq->head because tcp_wfree() might
813 * interrupt us (non NAPI drivers)
815 static void tcp_tasklet_func(unsigned long data)
817 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
820 struct list_head *q, *n;
824 local_irq_save(flags);
825 list_splice_init(&tsq->head, &list);
826 local_irq_restore(flags);
828 list_for_each_safe(q, n, &list) {
829 tp = list_entry(q, struct tcp_sock, tsq_node);
830 list_del(&tp->tsq_node);
832 sk = (struct sock *)tp;
833 smp_mb__before_atomic();
834 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
841 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
842 TCPF_WRITE_TIMER_DEFERRED | \
843 TCPF_DELACK_TIMER_DEFERRED | \
844 TCPF_MTU_REDUCED_DEFERRED)
846 * tcp_release_cb - tcp release_sock() callback
849 * called from release_sock() to perform protocol dependent
850 * actions before socket release.
852 void tcp_release_cb(struct sock *sk)
854 unsigned long flags, nflags;
856 /* perform an atomic operation only if at least one flag is set */
858 flags = sk->sk_tsq_flags;
859 if (!(flags & TCP_DEFERRED_ALL))
861 nflags = flags & ~TCP_DEFERRED_ALL;
862 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
864 if (flags & TCPF_TSQ_DEFERRED) {
868 /* Here begins the tricky part :
869 * We are called from release_sock() with :
871 * 2) sk_lock.slock spinlock held
872 * 3) socket owned by us (sk->sk_lock.owned == 1)
874 * But following code is meant to be called from BH handlers,
875 * so we should keep BH disabled, but early release socket ownership
877 sock_release_ownership(sk);
879 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
880 tcp_write_timer_handler(sk);
883 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
884 tcp_delack_timer_handler(sk);
887 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
888 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
892 EXPORT_SYMBOL(tcp_release_cb);
894 void __init tcp_tasklet_init(void)
898 for_each_possible_cpu(i) {
899 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
901 INIT_LIST_HEAD(&tsq->head);
902 tasklet_init(&tsq->tasklet,
909 * Write buffer destructor automatically called from kfree_skb.
910 * We can't xmit new skbs from this context, as we might already
913 void tcp_wfree(struct sk_buff *skb)
915 struct sock *sk = skb->sk;
916 struct tcp_sock *tp = tcp_sk(sk);
917 unsigned long flags, nval, oval;
919 /* Keep one reference on sk_wmem_alloc.
920 * Will be released by sk_free() from here or tcp_tasklet_func()
922 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
924 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
925 * Wait until our queues (qdisc + devices) are drained.
927 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
928 * - chance for incoming ACK (processed by another cpu maybe)
929 * to migrate this flow (skb->ooo_okay will be eventually set)
931 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
934 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
935 struct tsq_tasklet *tsq;
938 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
941 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
942 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
946 /* queue this socket to tasklet queue */
947 local_irq_save(flags);
948 tsq = this_cpu_ptr(&tsq_tasklet);
949 empty = list_empty(&tsq->head);
950 list_add(&tp->tsq_node, &tsq->head);
952 tasklet_schedule(&tsq->tasklet);
953 local_irq_restore(flags);
960 /* Note: Called under soft irq.
961 * We can call TCP stack right away, unless socket is owned by user.
963 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
965 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
966 struct sock *sk = (struct sock *)tp;
971 return HRTIMER_NORESTART;
974 /* BBR congestion control needs pacing.
975 * Same remark for SO_MAX_PACING_RATE.
976 * sch_fq packet scheduler is efficiently handling pacing,
977 * but is not always installed/used.
978 * Return true if TCP stack should pace packets itself.
980 static bool tcp_needs_internal_pacing(const struct sock *sk)
982 return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
985 static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
990 if (!tcp_needs_internal_pacing(sk))
992 rate = sk->sk_pacing_rate;
993 if (!rate || rate == ~0U)
996 /* Should account for header sizes as sch_fq does,
997 * but lets make things simple.
999 len_ns = (u64)skb->len * NSEC_PER_SEC;
1000 do_div(len_ns, rate);
1001 hrtimer_start(&tcp_sk(sk)->pacing_timer,
1002 ktime_add_ns(ktime_get(), len_ns),
1003 HRTIMER_MODE_ABS_PINNED_SOFT);
1007 static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb)
1009 skb->skb_mstamp = tp->tcp_mstamp;
1010 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1013 /* This routine actually transmits TCP packets queued in by
1014 * tcp_do_sendmsg(). This is used by both the initial
1015 * transmission and possible later retransmissions.
1016 * All SKB's seen here are completely headerless. It is our
1017 * job to build the TCP header, and pass the packet down to
1018 * IP so it can do the same plus pass the packet off to the
1021 * We are working here with either a clone of the original
1022 * SKB, or a fresh unique copy made by the retransmit engine.
1024 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1027 const struct inet_connection_sock *icsk = inet_csk(sk);
1028 struct inet_sock *inet;
1029 struct tcp_sock *tp;
1030 struct tcp_skb_cb *tcb;
1031 struct tcp_out_options opts;
1032 unsigned int tcp_options_size, tcp_header_size;
1033 struct sk_buff *oskb = NULL;
1034 struct tcp_md5sig_key *md5;
1038 BUG_ON(!skb || !tcp_skb_pcount(skb));
1042 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1046 tcp_skb_tsorted_save(oskb) {
1047 if (unlikely(skb_cloned(oskb)))
1048 skb = pskb_copy(oskb, gfp_mask);
1050 skb = skb_clone(oskb, gfp_mask);
1051 } tcp_skb_tsorted_restore(oskb);
1056 skb->skb_mstamp = tp->tcp_mstamp;
1059 tcb = TCP_SKB_CB(skb);
1060 memset(&opts, 0, sizeof(opts));
1062 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1063 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1065 tcp_options_size = tcp_established_options(sk, skb, &opts,
1067 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1069 /* if no packet is in qdisc/device queue, then allow XPS to select
1070 * another queue. We can be called from tcp_tsq_handler()
1071 * which holds one reference to sk.
1073 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1074 * One way to get this would be to set skb->truesize = 2 on them.
1076 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1078 /* If we had to use memory reserve to allocate this skb,
1079 * this might cause drops if packet is looped back :
1080 * Other socket might not have SOCK_MEMALLOC.
1081 * Packets not looped back do not care about pfmemalloc.
1083 skb->pfmemalloc = 0;
1085 skb_push(skb, tcp_header_size);
1086 skb_reset_transport_header(skb);
1090 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1091 skb_set_hash_from_sk(skb, sk);
1092 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1094 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1096 /* Build TCP header and checksum it. */
1097 th = (struct tcphdr *)skb->data;
1098 th->source = inet->inet_sport;
1099 th->dest = inet->inet_dport;
1100 th->seq = htonl(tcb->seq);
1101 th->ack_seq = htonl(tp->rcv_nxt);
1102 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1108 /* The urg_mode check is necessary during a below snd_una win probe */
1109 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1110 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1111 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1113 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1114 th->urg_ptr = htons(0xFFFF);
1119 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1120 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1121 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1122 th->window = htons(tcp_select_window(sk));
1123 tcp_ecn_send(sk, skb, th, tcp_header_size);
1125 /* RFC1323: The window in SYN & SYN/ACK segments
1128 th->window = htons(min(tp->rcv_wnd, 65535U));
1130 #ifdef CONFIG_TCP_MD5SIG
1131 /* Calculate the MD5 hash, as we have all we need now */
1133 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1134 tp->af_specific->calc_md5_hash(opts.hash_location,
1139 icsk->icsk_af_ops->send_check(sk, skb);
1141 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1142 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1144 if (skb->len != tcp_header_size) {
1145 tcp_event_data_sent(tp, sk);
1146 tp->data_segs_out += tcp_skb_pcount(skb);
1147 tcp_internal_pacing(sk, skb);
1150 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1151 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1152 tcp_skb_pcount(skb));
1154 tp->segs_out += tcp_skb_pcount(skb);
1155 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1156 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1157 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1159 /* Our usage of tstamp should remain private */
1162 /* Cleanup our debris for IP stacks */
1163 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1164 sizeof(struct inet6_skb_parm)));
1166 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1168 if (unlikely(err > 0)) {
1170 err = net_xmit_eval(err);
1173 tcp_update_skb_after_send(tp, oskb);
1174 tcp_rate_skb_sent(sk, oskb);
1179 /* This routine just queues the buffer for sending.
1181 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1182 * otherwise socket can stall.
1184 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1186 struct tcp_sock *tp = tcp_sk(sk);
1188 /* Advance write_seq and place onto the write_queue. */
1189 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1190 __skb_header_release(skb);
1191 tcp_add_write_queue_tail(sk, skb);
1192 sk->sk_wmem_queued += skb->truesize;
1193 sk_mem_charge(sk, skb->truesize);
1196 /* Initialize TSO segments for a packet. */
1197 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1199 if (skb->len <= mss_now) {
1200 /* Avoid the costly divide in the normal
1203 tcp_skb_pcount_set(skb, 1);
1204 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1206 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1207 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1211 /* Pcount in the middle of the write queue got changed, we need to do various
1212 * tweaks to fix counters
1214 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1216 struct tcp_sock *tp = tcp_sk(sk);
1218 tp->packets_out -= decr;
1220 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1221 tp->sacked_out -= decr;
1222 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1223 tp->retrans_out -= decr;
1224 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1225 tp->lost_out -= decr;
1227 /* Reno case is special. Sigh... */
1228 if (tcp_is_reno(tp) && decr > 0)
1229 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1231 if (tp->lost_skb_hint &&
1232 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1233 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1234 tp->lost_cnt_hint -= decr;
1236 tcp_verify_left_out(tp);
1239 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1241 return TCP_SKB_CB(skb)->txstamp_ack ||
1242 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1245 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1247 struct skb_shared_info *shinfo = skb_shinfo(skb);
1249 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1250 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1251 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1252 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1254 shinfo->tx_flags &= ~tsflags;
1255 shinfo2->tx_flags |= tsflags;
1256 swap(shinfo->tskey, shinfo2->tskey);
1257 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1258 TCP_SKB_CB(skb)->txstamp_ack = 0;
1262 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1264 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1265 TCP_SKB_CB(skb)->eor = 0;
1268 /* Insert buff after skb on the write or rtx queue of sk. */
1269 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1270 struct sk_buff *buff,
1272 enum tcp_queue tcp_queue)
1274 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1275 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1277 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1280 /* Function to create two new TCP segments. Shrinks the given segment
1281 * to the specified size and appends a new segment with the rest of the
1282 * packet to the list. This won't be called frequently, I hope.
1283 * Remember, these are still headerless SKBs at this point.
1285 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1286 struct sk_buff *skb, u32 len,
1287 unsigned int mss_now, gfp_t gfp)
1289 struct tcp_sock *tp = tcp_sk(sk);
1290 struct sk_buff *buff;
1291 int nsize, old_factor;
1295 if (WARN_ON(len > skb->len))
1298 nsize = skb_headlen(skb) - len;
1302 if (skb_unclone(skb, gfp))
1305 /* Get a new skb... force flag on. */
1306 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1308 return -ENOMEM; /* We'll just try again later. */
1310 sk->sk_wmem_queued += buff->truesize;
1311 sk_mem_charge(sk, buff->truesize);
1312 nlen = skb->len - len - nsize;
1313 buff->truesize += nlen;
1314 skb->truesize -= nlen;
1316 /* Correct the sequence numbers. */
1317 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1318 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1319 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1321 /* PSH and FIN should only be set in the second packet. */
1322 flags = TCP_SKB_CB(skb)->tcp_flags;
1323 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1324 TCP_SKB_CB(buff)->tcp_flags = flags;
1325 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1326 tcp_skb_fragment_eor(skb, buff);
1328 skb_split(skb, buff, len);
1330 buff->ip_summed = CHECKSUM_PARTIAL;
1332 buff->tstamp = skb->tstamp;
1333 tcp_fragment_tstamp(skb, buff);
1335 old_factor = tcp_skb_pcount(skb);
1337 /* Fix up tso_factor for both original and new SKB. */
1338 tcp_set_skb_tso_segs(skb, mss_now);
1339 tcp_set_skb_tso_segs(buff, mss_now);
1341 /* Update delivered info for the new segment */
1342 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1344 /* If this packet has been sent out already, we must
1345 * adjust the various packet counters.
1347 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1348 int diff = old_factor - tcp_skb_pcount(skb) -
1349 tcp_skb_pcount(buff);
1352 tcp_adjust_pcount(sk, skb, diff);
1355 /* Link BUFF into the send queue. */
1356 __skb_header_release(buff);
1357 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1358 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1359 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1364 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1365 * data is not copied, but immediately discarded.
1367 static int __pskb_trim_head(struct sk_buff *skb, int len)
1369 struct skb_shared_info *shinfo;
1372 eat = min_t(int, len, skb_headlen(skb));
1374 __skb_pull(skb, eat);
1381 shinfo = skb_shinfo(skb);
1382 for (i = 0; i < shinfo->nr_frags; i++) {
1383 int size = skb_frag_size(&shinfo->frags[i]);
1386 skb_frag_unref(skb, i);
1389 shinfo->frags[k] = shinfo->frags[i];
1391 shinfo->frags[k].page_offset += eat;
1392 skb_frag_size_sub(&shinfo->frags[k], eat);
1398 shinfo->nr_frags = k;
1400 skb->data_len -= len;
1401 skb->len = skb->data_len;
1405 /* Remove acked data from a packet in the transmit queue. */
1406 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1410 if (skb_unclone(skb, GFP_ATOMIC))
1413 delta_truesize = __pskb_trim_head(skb, len);
1415 TCP_SKB_CB(skb)->seq += len;
1416 skb->ip_summed = CHECKSUM_PARTIAL;
1418 if (delta_truesize) {
1419 skb->truesize -= delta_truesize;
1420 sk->sk_wmem_queued -= delta_truesize;
1421 sk_mem_uncharge(sk, delta_truesize);
1422 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1425 /* Any change of skb->len requires recalculation of tso factor. */
1426 if (tcp_skb_pcount(skb) > 1)
1427 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1432 /* Calculate MSS not accounting any TCP options. */
1433 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1435 const struct tcp_sock *tp = tcp_sk(sk);
1436 const struct inet_connection_sock *icsk = inet_csk(sk);
1439 /* Calculate base mss without TCP options:
1440 It is MMS_S - sizeof(tcphdr) of rfc1122
1442 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1444 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1445 if (icsk->icsk_af_ops->net_frag_header_len) {
1446 const struct dst_entry *dst = __sk_dst_get(sk);
1448 if (dst && dst_allfrag(dst))
1449 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1452 /* Clamp it (mss_clamp does not include tcp options) */
1453 if (mss_now > tp->rx_opt.mss_clamp)
1454 mss_now = tp->rx_opt.mss_clamp;
1456 /* Now subtract optional transport overhead */
1457 mss_now -= icsk->icsk_ext_hdr_len;
1459 /* Then reserve room for full set of TCP options and 8 bytes of data */
1465 /* Calculate MSS. Not accounting for SACKs here. */
1466 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1468 /* Subtract TCP options size, not including SACKs */
1469 return __tcp_mtu_to_mss(sk, pmtu) -
1470 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1473 /* Inverse of above */
1474 int tcp_mss_to_mtu(struct sock *sk, int mss)
1476 const struct tcp_sock *tp = tcp_sk(sk);
1477 const struct inet_connection_sock *icsk = inet_csk(sk);
1481 tp->tcp_header_len +
1482 icsk->icsk_ext_hdr_len +
1483 icsk->icsk_af_ops->net_header_len;
1485 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1486 if (icsk->icsk_af_ops->net_frag_header_len) {
1487 const struct dst_entry *dst = __sk_dst_get(sk);
1489 if (dst && dst_allfrag(dst))
1490 mtu += icsk->icsk_af_ops->net_frag_header_len;
1494 EXPORT_SYMBOL(tcp_mss_to_mtu);
1496 /* MTU probing init per socket */
1497 void tcp_mtup_init(struct sock *sk)
1499 struct tcp_sock *tp = tcp_sk(sk);
1500 struct inet_connection_sock *icsk = inet_csk(sk);
1501 struct net *net = sock_net(sk);
1503 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1504 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1505 icsk->icsk_af_ops->net_header_len;
1506 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1507 icsk->icsk_mtup.probe_size = 0;
1508 if (icsk->icsk_mtup.enabled)
1509 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1511 EXPORT_SYMBOL(tcp_mtup_init);
1513 /* This function synchronize snd mss to current pmtu/exthdr set.
1515 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1516 for TCP options, but includes only bare TCP header.
1518 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1519 It is minimum of user_mss and mss received with SYN.
1520 It also does not include TCP options.
1522 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1524 tp->mss_cache is current effective sending mss, including
1525 all tcp options except for SACKs. It is evaluated,
1526 taking into account current pmtu, but never exceeds
1527 tp->rx_opt.mss_clamp.
1529 NOTE1. rfc1122 clearly states that advertised MSS
1530 DOES NOT include either tcp or ip options.
1532 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1533 are READ ONLY outside this function. --ANK (980731)
1535 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1537 struct tcp_sock *tp = tcp_sk(sk);
1538 struct inet_connection_sock *icsk = inet_csk(sk);
1541 if (icsk->icsk_mtup.search_high > pmtu)
1542 icsk->icsk_mtup.search_high = pmtu;
1544 mss_now = tcp_mtu_to_mss(sk, pmtu);
1545 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1547 /* And store cached results */
1548 icsk->icsk_pmtu_cookie = pmtu;
1549 if (icsk->icsk_mtup.enabled)
1550 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1551 tp->mss_cache = mss_now;
1555 EXPORT_SYMBOL(tcp_sync_mss);
1557 /* Compute the current effective MSS, taking SACKs and IP options,
1558 * and even PMTU discovery events into account.
1560 unsigned int tcp_current_mss(struct sock *sk)
1562 const struct tcp_sock *tp = tcp_sk(sk);
1563 const struct dst_entry *dst = __sk_dst_get(sk);
1565 unsigned int header_len;
1566 struct tcp_out_options opts;
1567 struct tcp_md5sig_key *md5;
1569 mss_now = tp->mss_cache;
1572 u32 mtu = dst_mtu(dst);
1573 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1574 mss_now = tcp_sync_mss(sk, mtu);
1577 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1578 sizeof(struct tcphdr);
1579 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1580 * some common options. If this is an odd packet (because we have SACK
1581 * blocks etc) then our calculated header_len will be different, and
1582 * we have to adjust mss_now correspondingly */
1583 if (header_len != tp->tcp_header_len) {
1584 int delta = (int) header_len - tp->tcp_header_len;
1591 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1592 * As additional protections, we do not touch cwnd in retransmission phases,
1593 * and if application hit its sndbuf limit recently.
1595 static void tcp_cwnd_application_limited(struct sock *sk)
1597 struct tcp_sock *tp = tcp_sk(sk);
1599 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1600 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1601 /* Limited by application or receiver window. */
1602 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1603 u32 win_used = max(tp->snd_cwnd_used, init_win);
1604 if (win_used < tp->snd_cwnd) {
1605 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1606 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1608 tp->snd_cwnd_used = 0;
1610 tp->snd_cwnd_stamp = tcp_jiffies32;
1613 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1615 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1616 struct tcp_sock *tp = tcp_sk(sk);
1618 /* Track the maximum number of outstanding packets in each
1619 * window, and remember whether we were cwnd-limited then.
1621 if (!before(tp->snd_una, tp->max_packets_seq) ||
1622 tp->packets_out > tp->max_packets_out) {
1623 tp->max_packets_out = tp->packets_out;
1624 tp->max_packets_seq = tp->snd_nxt;
1625 tp->is_cwnd_limited = is_cwnd_limited;
1628 if (tcp_is_cwnd_limited(sk)) {
1629 /* Network is feed fully. */
1630 tp->snd_cwnd_used = 0;
1631 tp->snd_cwnd_stamp = tcp_jiffies32;
1633 /* Network starves. */
1634 if (tp->packets_out > tp->snd_cwnd_used)
1635 tp->snd_cwnd_used = tp->packets_out;
1637 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1638 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1639 !ca_ops->cong_control)
1640 tcp_cwnd_application_limited(sk);
1642 /* The following conditions together indicate the starvation
1643 * is caused by insufficient sender buffer:
1644 * 1) just sent some data (see tcp_write_xmit)
1645 * 2) not cwnd limited (this else condition)
1646 * 3) no more data to send (tcp_write_queue_empty())
1647 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1649 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1650 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1651 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1652 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1656 /* Minshall's variant of the Nagle send check. */
1657 static bool tcp_minshall_check(const struct tcp_sock *tp)
1659 return after(tp->snd_sml, tp->snd_una) &&
1660 !after(tp->snd_sml, tp->snd_nxt);
1663 /* Update snd_sml if this skb is under mss
1664 * Note that a TSO packet might end with a sub-mss segment
1665 * The test is really :
1666 * if ((skb->len % mss) != 0)
1667 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1668 * But we can avoid doing the divide again given we already have
1669 * skb_pcount = skb->len / mss_now
1671 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1672 const struct sk_buff *skb)
1674 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1675 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1678 /* Return false, if packet can be sent now without violation Nagle's rules:
1679 * 1. It is full sized. (provided by caller in %partial bool)
1680 * 2. Or it contains FIN. (already checked by caller)
1681 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1682 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1683 * With Minshall's modification: all sent small packets are ACKed.
1685 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1689 ((nonagle & TCP_NAGLE_CORK) ||
1690 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1693 /* Return how many segs we'd like on a TSO packet,
1694 * to send one TSO packet per ms
1696 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1701 bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift,
1702 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1704 /* Goal is to send at least one packet per ms,
1705 * not one big TSO packet every 100 ms.
1706 * This preserves ACK clocking and is consistent
1707 * with tcp_tso_should_defer() heuristic.
1709 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1714 /* Return the number of segments we want in the skb we are transmitting.
1715 * See if congestion control module wants to decide; otherwise, autosize.
1717 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1719 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1720 u32 min_tso, tso_segs;
1722 min_tso = ca_ops->min_tso_segs ?
1723 ca_ops->min_tso_segs(sk) :
1724 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1726 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1727 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1730 /* Returns the portion of skb which can be sent right away */
1731 static unsigned int tcp_mss_split_point(const struct sock *sk,
1732 const struct sk_buff *skb,
1733 unsigned int mss_now,
1734 unsigned int max_segs,
1737 const struct tcp_sock *tp = tcp_sk(sk);
1738 u32 partial, needed, window, max_len;
1740 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1741 max_len = mss_now * max_segs;
1743 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1746 needed = min(skb->len, window);
1748 if (max_len <= needed)
1751 partial = needed % mss_now;
1752 /* If last segment is not a full MSS, check if Nagle rules allow us
1753 * to include this last segment in this skb.
1754 * Otherwise, we'll split the skb at last MSS boundary
1756 if (tcp_nagle_check(partial != 0, tp, nonagle))
1757 return needed - partial;
1762 /* Can at least one segment of SKB be sent right now, according to the
1763 * congestion window rules? If so, return how many segments are allowed.
1765 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1766 const struct sk_buff *skb)
1768 u32 in_flight, cwnd, halfcwnd;
1770 /* Don't be strict about the congestion window for the final FIN. */
1771 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1772 tcp_skb_pcount(skb) == 1)
1775 in_flight = tcp_packets_in_flight(tp);
1776 cwnd = tp->snd_cwnd;
1777 if (in_flight >= cwnd)
1780 /* For better scheduling, ensure we have at least
1781 * 2 GSO packets in flight.
1783 halfcwnd = max(cwnd >> 1, 1U);
1784 return min(halfcwnd, cwnd - in_flight);
1787 /* Initialize TSO state of a skb.
1788 * This must be invoked the first time we consider transmitting
1789 * SKB onto the wire.
1791 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1793 int tso_segs = tcp_skb_pcount(skb);
1795 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1796 tcp_set_skb_tso_segs(skb, mss_now);
1797 tso_segs = tcp_skb_pcount(skb);
1803 /* Return true if the Nagle test allows this packet to be
1806 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1807 unsigned int cur_mss, int nonagle)
1809 /* Nagle rule does not apply to frames, which sit in the middle of the
1810 * write_queue (they have no chances to get new data).
1812 * This is implemented in the callers, where they modify the 'nonagle'
1813 * argument based upon the location of SKB in the send queue.
1815 if (nonagle & TCP_NAGLE_PUSH)
1818 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1819 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1822 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1828 /* Does at least the first segment of SKB fit into the send window? */
1829 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1830 const struct sk_buff *skb,
1831 unsigned int cur_mss)
1833 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1835 if (skb->len > cur_mss)
1836 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1838 return !after(end_seq, tcp_wnd_end(tp));
1841 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1842 * which is put after SKB on the list. It is very much like
1843 * tcp_fragment() except that it may make several kinds of assumptions
1844 * in order to speed up the splitting operation. In particular, we
1845 * know that all the data is in scatter-gather pages, and that the
1846 * packet has never been sent out before (and thus is not cloned).
1848 static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1849 struct sk_buff *skb, unsigned int len,
1850 unsigned int mss_now, gfp_t gfp)
1852 struct sk_buff *buff;
1853 int nlen = skb->len - len;
1856 /* All of a TSO frame must be composed of paged data. */
1857 if (skb->len != skb->data_len)
1858 return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp);
1860 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1861 if (unlikely(!buff))
1864 sk->sk_wmem_queued += buff->truesize;
1865 sk_mem_charge(sk, buff->truesize);
1866 buff->truesize += nlen;
1867 skb->truesize -= nlen;
1869 /* Correct the sequence numbers. */
1870 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1871 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1872 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1874 /* PSH and FIN should only be set in the second packet. */
1875 flags = TCP_SKB_CB(skb)->tcp_flags;
1876 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1877 TCP_SKB_CB(buff)->tcp_flags = flags;
1879 /* This packet was never sent out yet, so no SACK bits. */
1880 TCP_SKB_CB(buff)->sacked = 0;
1882 tcp_skb_fragment_eor(skb, buff);
1884 buff->ip_summed = CHECKSUM_PARTIAL;
1885 skb_split(skb, buff, len);
1886 tcp_fragment_tstamp(skb, buff);
1888 /* Fix up tso_factor for both original and new SKB. */
1889 tcp_set_skb_tso_segs(skb, mss_now);
1890 tcp_set_skb_tso_segs(buff, mss_now);
1892 /* Link BUFF into the send queue. */
1893 __skb_header_release(buff);
1894 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1899 /* Try to defer sending, if possible, in order to minimize the amount
1900 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1902 * This algorithm is from John Heffner.
1904 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1905 bool *is_cwnd_limited, u32 max_segs)
1907 const struct inet_connection_sock *icsk = inet_csk(sk);
1908 u32 age, send_win, cong_win, limit, in_flight;
1909 struct tcp_sock *tp = tcp_sk(sk);
1910 struct sk_buff *head;
1913 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1916 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1919 /* Avoid bursty behavior by allowing defer
1920 * only if the last write was recent.
1922 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
1925 in_flight = tcp_packets_in_flight(tp);
1927 BUG_ON(tcp_skb_pcount(skb) <= 1);
1928 BUG_ON(tp->snd_cwnd <= in_flight);
1930 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1932 /* From in_flight test above, we know that cwnd > in_flight. */
1933 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1935 limit = min(send_win, cong_win);
1937 /* If a full-sized TSO skb can be sent, do it. */
1938 if (limit >= max_segs * tp->mss_cache)
1941 /* Middle in queue won't get any more data, full sendable already? */
1942 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1945 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1947 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1949 /* If at least some fraction of a window is available,
1952 chunk /= win_divisor;
1956 /* Different approach, try not to defer past a single
1957 * ACK. Receiver should ACK every other full sized
1958 * frame, so if we have space for more than 3 frames
1961 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1965 /* TODO : use tsorted_sent_queue ? */
1966 head = tcp_rtx_queue_head(sk);
1969 age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
1970 /* If next ACK is likely to come too late (half srtt), do not defer */
1971 if (age < (tp->srtt_us >> 4))
1974 /* Ok, it looks like it is advisable to defer. */
1976 if (cong_win < send_win && cong_win <= skb->len)
1977 *is_cwnd_limited = true;
1985 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1987 struct inet_connection_sock *icsk = inet_csk(sk);
1988 struct tcp_sock *tp = tcp_sk(sk);
1989 struct net *net = sock_net(sk);
1993 interval = net->ipv4.sysctl_tcp_probe_interval;
1994 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
1995 if (unlikely(delta >= interval * HZ)) {
1996 int mss = tcp_current_mss(sk);
1998 /* Update current search range */
1999 icsk->icsk_mtup.probe_size = 0;
2000 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2001 sizeof(struct tcphdr) +
2002 icsk->icsk_af_ops->net_header_len;
2003 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2005 /* Update probe time stamp */
2006 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2010 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2012 struct sk_buff *skb, *next;
2014 skb = tcp_send_head(sk);
2015 tcp_for_write_queue_from_safe(skb, next, sk) {
2016 if (len <= skb->len)
2019 if (unlikely(TCP_SKB_CB(skb)->eor))
2028 /* Create a new MTU probe if we are ready.
2029 * MTU probe is regularly attempting to increase the path MTU by
2030 * deliberately sending larger packets. This discovers routing
2031 * changes resulting in larger path MTUs.
2033 * Returns 0 if we should wait to probe (no cwnd available),
2034 * 1 if a probe was sent,
2037 static int tcp_mtu_probe(struct sock *sk)
2039 struct inet_connection_sock *icsk = inet_csk(sk);
2040 struct tcp_sock *tp = tcp_sk(sk);
2041 struct sk_buff *skb, *nskb, *next;
2042 struct net *net = sock_net(sk);
2049 /* Not currently probing/verifying,
2051 * have enough cwnd, and
2052 * not SACKing (the variable headers throw things off)
2054 if (likely(!icsk->icsk_mtup.enabled ||
2055 icsk->icsk_mtup.probe_size ||
2056 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2057 tp->snd_cwnd < 11 ||
2058 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2061 /* Use binary search for probe_size between tcp_mss_base,
2062 * and current mss_clamp. if (search_high - search_low)
2063 * smaller than a threshold, backoff from probing.
2065 mss_now = tcp_current_mss(sk);
2066 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2067 icsk->icsk_mtup.search_low) >> 1);
2068 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2069 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2070 /* When misfortune happens, we are reprobing actively,
2071 * and then reprobe timer has expired. We stick with current
2072 * probing process by not resetting search range to its orignal.
2074 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2075 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2076 /* Check whether enough time has elaplased for
2077 * another round of probing.
2079 tcp_mtu_check_reprobe(sk);
2083 /* Have enough data in the send queue to probe? */
2084 if (tp->write_seq - tp->snd_nxt < size_needed)
2087 if (tp->snd_wnd < size_needed)
2089 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2092 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2093 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2094 if (!tcp_packets_in_flight(tp))
2100 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2103 /* We're allowed to probe. Build it now. */
2104 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2107 sk->sk_wmem_queued += nskb->truesize;
2108 sk_mem_charge(sk, nskb->truesize);
2110 skb = tcp_send_head(sk);
2112 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2113 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2114 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2115 TCP_SKB_CB(nskb)->sacked = 0;
2117 nskb->ip_summed = CHECKSUM_PARTIAL;
2119 tcp_insert_write_queue_before(nskb, skb, sk);
2120 tcp_highest_sack_replace(sk, skb, nskb);
2123 tcp_for_write_queue_from_safe(skb, next, sk) {
2124 copy = min_t(int, skb->len, probe_size - len);
2125 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2127 if (skb->len <= copy) {
2128 /* We've eaten all the data from this skb.
2130 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2131 /* If this is the last SKB we copy and eor is set
2132 * we need to propagate it to the new skb.
2134 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2135 tcp_unlink_write_queue(skb, sk);
2136 sk_wmem_free_skb(sk, skb);
2138 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2139 ~(TCPHDR_FIN|TCPHDR_PSH);
2140 if (!skb_shinfo(skb)->nr_frags) {
2141 skb_pull(skb, copy);
2143 __pskb_trim_head(skb, copy);
2144 tcp_set_skb_tso_segs(skb, mss_now);
2146 TCP_SKB_CB(skb)->seq += copy;
2151 if (len >= probe_size)
2154 tcp_init_tso_segs(nskb, nskb->len);
2156 /* We're ready to send. If this fails, the probe will
2157 * be resegmented into mss-sized pieces by tcp_write_xmit().
2159 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2160 /* Decrement cwnd here because we are sending
2161 * effectively two packets. */
2163 tcp_event_new_data_sent(sk, nskb);
2165 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2166 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2167 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2175 static bool tcp_pacing_check(const struct sock *sk)
2177 return tcp_needs_internal_pacing(sk) &&
2178 hrtimer_is_queued(&tcp_sk(sk)->pacing_timer);
2181 /* TCP Small Queues :
2182 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2183 * (These limits are doubled for retransmits)
2185 * - better RTT estimation and ACK scheduling
2188 * Alas, some drivers / subsystems require a fair amount
2189 * of queued bytes to ensure line rate.
2190 * One example is wifi aggregation (802.11 AMPDU)
2192 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2193 unsigned int factor)
2197 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift);
2198 limit = min_t(u32, limit,
2199 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2202 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2203 /* Always send skb if rtx queue is empty.
2204 * No need to wait for TX completion to call us back,
2205 * after softirq/tasklet schedule.
2206 * This helps when TX completions are delayed too much.
2208 if (tcp_rtx_queue_empty(sk))
2211 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2212 /* It is possible TX completion already happened
2213 * before we set TSQ_THROTTLED, so we must
2214 * test again the condition.
2216 smp_mb__after_atomic();
2217 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2223 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2225 const u32 now = tcp_jiffies32;
2226 enum tcp_chrono old = tp->chrono_type;
2228 if (old > TCP_CHRONO_UNSPEC)
2229 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2230 tp->chrono_start = now;
2231 tp->chrono_type = new;
2234 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2236 struct tcp_sock *tp = tcp_sk(sk);
2238 /* If there are multiple conditions worthy of tracking in a
2239 * chronograph then the highest priority enum takes precedence
2240 * over the other conditions. So that if something "more interesting"
2241 * starts happening, stop the previous chrono and start a new one.
2243 if (type > tp->chrono_type)
2244 tcp_chrono_set(tp, type);
2247 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2249 struct tcp_sock *tp = tcp_sk(sk);
2252 /* There are multiple conditions worthy of tracking in a
2253 * chronograph, so that the highest priority enum takes
2254 * precedence over the other conditions (see tcp_chrono_start).
2255 * If a condition stops, we only stop chrono tracking if
2256 * it's the "most interesting" or current chrono we are
2257 * tracking and starts busy chrono if we have pending data.
2259 if (tcp_rtx_and_write_queues_empty(sk))
2260 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2261 else if (type == tp->chrono_type)
2262 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2265 /* This routine writes packets to the network. It advances the
2266 * send_head. This happens as incoming acks open up the remote
2269 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2270 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2271 * account rare use of URG, this is not a big flaw.
2273 * Send at most one packet when push_one > 0. Temporarily ignore
2274 * cwnd limit to force at most one packet out when push_one == 2.
2276 * Returns true, if no segments are in flight and we have queued segments,
2277 * but cannot send anything now because of SWS or another problem.
2279 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2280 int push_one, gfp_t gfp)
2282 struct tcp_sock *tp = tcp_sk(sk);
2283 struct sk_buff *skb;
2284 unsigned int tso_segs, sent_pkts;
2287 bool is_cwnd_limited = false, is_rwnd_limited = false;
2292 tcp_mstamp_refresh(tp);
2294 /* Do MTU probing. */
2295 result = tcp_mtu_probe(sk);
2298 } else if (result > 0) {
2303 max_segs = tcp_tso_segs(sk, mss_now);
2304 while ((skb = tcp_send_head(sk))) {
2307 if (tcp_pacing_check(sk))
2310 tso_segs = tcp_init_tso_segs(skb, mss_now);
2313 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2314 /* "skb_mstamp" is used as a start point for the retransmit timer */
2315 tcp_update_skb_after_send(tp, skb);
2316 goto repair; /* Skip network transmission */
2319 cwnd_quota = tcp_cwnd_test(tp, skb);
2322 /* Force out a loss probe pkt. */
2328 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2329 is_rwnd_limited = true;
2333 if (tso_segs == 1) {
2334 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2335 (tcp_skb_is_last(sk, skb) ?
2336 nonagle : TCP_NAGLE_PUSH))))
2340 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2346 if (tso_segs > 1 && !tcp_urg_mode(tp))
2347 limit = tcp_mss_split_point(sk, skb, mss_now,
2353 if (skb->len > limit &&
2354 unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2355 skb, limit, mss_now, gfp)))
2358 if (tcp_small_queue_check(sk, skb, 0))
2361 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2365 /* Advance the send_head. This one is sent out.
2366 * This call will increment packets_out.
2368 tcp_event_new_data_sent(sk, skb);
2370 tcp_minshall_update(tp, mss_now, skb);
2371 sent_pkts += tcp_skb_pcount(skb);
2377 if (is_rwnd_limited)
2378 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2380 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2382 if (likely(sent_pkts)) {
2383 if (tcp_in_cwnd_reduction(sk))
2384 tp->prr_out += sent_pkts;
2386 /* Send one loss probe per tail loss episode. */
2388 tcp_schedule_loss_probe(sk, false);
2389 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2390 tcp_cwnd_validate(sk, is_cwnd_limited);
2393 return !tp->packets_out && !tcp_write_queue_empty(sk);
2396 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2398 struct inet_connection_sock *icsk = inet_csk(sk);
2399 struct tcp_sock *tp = tcp_sk(sk);
2400 u32 timeout, rto_delta_us;
2403 /* Don't do any loss probe on a Fast Open connection before 3WHS
2406 if (tp->fastopen_rsk)
2409 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2410 /* Schedule a loss probe in 2*RTT for SACK capable connections
2411 * not in loss recovery, that are either limited by cwnd or application.
2413 if ((early_retrans != 3 && early_retrans != 4) ||
2414 !tp->packets_out || !tcp_is_sack(tp) ||
2415 (icsk->icsk_ca_state != TCP_CA_Open &&
2416 icsk->icsk_ca_state != TCP_CA_CWR))
2419 /* Probe timeout is 2*rtt. Add minimum RTO to account
2420 * for delayed ack when there's one outstanding packet. If no RTT
2421 * sample is available then probe after TCP_TIMEOUT_INIT.
2424 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2425 if (tp->packets_out == 1)
2426 timeout += TCP_RTO_MIN;
2428 timeout += TCP_TIMEOUT_MIN;
2430 timeout = TCP_TIMEOUT_INIT;
2433 /* If the RTO formula yields an earlier time, then use that time. */
2434 rto_delta_us = advancing_rto ?
2435 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2436 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2437 if (rto_delta_us > 0)
2438 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2440 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2445 /* Thanks to skb fast clones, we can detect if a prior transmit of
2446 * a packet is still in a qdisc or driver queue.
2447 * In this case, there is very little point doing a retransmit !
2449 static bool skb_still_in_host_queue(const struct sock *sk,
2450 const struct sk_buff *skb)
2452 if (unlikely(skb_fclone_busy(sk, skb))) {
2453 NET_INC_STATS(sock_net(sk),
2454 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2460 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2461 * retransmit the last segment.
2463 void tcp_send_loss_probe(struct sock *sk)
2465 struct tcp_sock *tp = tcp_sk(sk);
2466 struct sk_buff *skb;
2468 int mss = tcp_current_mss(sk);
2470 skb = tcp_send_head(sk);
2471 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2472 pcount = tp->packets_out;
2473 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2474 if (tp->packets_out > pcount)
2478 skb = skb_rb_last(&sk->tcp_rtx_queue);
2480 /* At most one outstanding TLP retransmission. */
2481 if (tp->tlp_high_seq)
2484 /* Retransmit last segment. */
2488 if (skb_still_in_host_queue(sk, skb))
2491 pcount = tcp_skb_pcount(skb);
2492 if (WARN_ON(!pcount))
2495 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2496 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2497 (pcount - 1) * mss, mss,
2500 skb = skb_rb_next(skb);
2503 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2506 if (__tcp_retransmit_skb(sk, skb, 1))
2509 /* Record snd_nxt for loss detection. */
2510 tp->tlp_high_seq = tp->snd_nxt;
2513 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2514 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2515 inet_csk(sk)->icsk_pending = 0;
2520 /* Push out any pending frames which were held back due to
2521 * TCP_CORK or attempt at coalescing tiny packets.
2522 * The socket must be locked by the caller.
2524 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2527 /* If we are closed, the bytes will have to remain here.
2528 * In time closedown will finish, we empty the write queue and
2529 * all will be happy.
2531 if (unlikely(sk->sk_state == TCP_CLOSE))
2534 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2535 sk_gfp_mask(sk, GFP_ATOMIC)))
2536 tcp_check_probe_timer(sk);
2539 /* Send _single_ skb sitting at the send head. This function requires
2540 * true push pending frames to setup probe timer etc.
2542 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2544 struct sk_buff *skb = tcp_send_head(sk);
2546 BUG_ON(!skb || skb->len < mss_now);
2548 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2551 /* This function returns the amount that we can raise the
2552 * usable window based on the following constraints
2554 * 1. The window can never be shrunk once it is offered (RFC 793)
2555 * 2. We limit memory per socket
2558 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2559 * RECV.NEXT + RCV.WIN fixed until:
2560 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2562 * i.e. don't raise the right edge of the window until you can raise
2563 * it at least MSS bytes.
2565 * Unfortunately, the recommended algorithm breaks header prediction,
2566 * since header prediction assumes th->window stays fixed.
2568 * Strictly speaking, keeping th->window fixed violates the receiver
2569 * side SWS prevention criteria. The problem is that under this rule
2570 * a stream of single byte packets will cause the right side of the
2571 * window to always advance by a single byte.
2573 * Of course, if the sender implements sender side SWS prevention
2574 * then this will not be a problem.
2576 * BSD seems to make the following compromise:
2578 * If the free space is less than the 1/4 of the maximum
2579 * space available and the free space is less than 1/2 mss,
2580 * then set the window to 0.
2581 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2582 * Otherwise, just prevent the window from shrinking
2583 * and from being larger than the largest representable value.
2585 * This prevents incremental opening of the window in the regime
2586 * where TCP is limited by the speed of the reader side taking
2587 * data out of the TCP receive queue. It does nothing about
2588 * those cases where the window is constrained on the sender side
2589 * because the pipeline is full.
2591 * BSD also seems to "accidentally" limit itself to windows that are a
2592 * multiple of MSS, at least until the free space gets quite small.
2593 * This would appear to be a side effect of the mbuf implementation.
2594 * Combining these two algorithms results in the observed behavior
2595 * of having a fixed window size at almost all times.
2597 * Below we obtain similar behavior by forcing the offered window to
2598 * a multiple of the mss when it is feasible to do so.
2600 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2601 * Regular options like TIMESTAMP are taken into account.
2603 u32 __tcp_select_window(struct sock *sk)
2605 struct inet_connection_sock *icsk = inet_csk(sk);
2606 struct tcp_sock *tp = tcp_sk(sk);
2607 /* MSS for the peer's data. Previous versions used mss_clamp
2608 * here. I don't know if the value based on our guesses
2609 * of peer's MSS is better for the performance. It's more correct
2610 * but may be worse for the performance because of rcv_mss
2611 * fluctuations. --SAW 1998/11/1
2613 int mss = icsk->icsk_ack.rcv_mss;
2614 int free_space = tcp_space(sk);
2615 int allowed_space = tcp_full_space(sk);
2616 int full_space = min_t(int, tp->window_clamp, allowed_space);
2619 if (unlikely(mss > full_space)) {
2624 if (free_space < (full_space >> 1)) {
2625 icsk->icsk_ack.quick = 0;
2627 if (tcp_under_memory_pressure(sk))
2628 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2631 /* free_space might become our new window, make sure we don't
2632 * increase it due to wscale.
2634 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2636 /* if free space is less than mss estimate, or is below 1/16th
2637 * of the maximum allowed, try to move to zero-window, else
2638 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2639 * new incoming data is dropped due to memory limits.
2640 * With large window, mss test triggers way too late in order
2641 * to announce zero window in time before rmem limit kicks in.
2643 if (free_space < (allowed_space >> 4) || free_space < mss)
2647 if (free_space > tp->rcv_ssthresh)
2648 free_space = tp->rcv_ssthresh;
2650 /* Don't do rounding if we are using window scaling, since the
2651 * scaled window will not line up with the MSS boundary anyway.
2653 if (tp->rx_opt.rcv_wscale) {
2654 window = free_space;
2656 /* Advertise enough space so that it won't get scaled away.
2657 * Import case: prevent zero window announcement if
2658 * 1<<rcv_wscale > mss.
2660 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2662 window = tp->rcv_wnd;
2663 /* Get the largest window that is a nice multiple of mss.
2664 * Window clamp already applied above.
2665 * If our current window offering is within 1 mss of the
2666 * free space we just keep it. This prevents the divide
2667 * and multiply from happening most of the time.
2668 * We also don't do any window rounding when the free space
2671 if (window <= free_space - mss || window > free_space)
2672 window = rounddown(free_space, mss);
2673 else if (mss == full_space &&
2674 free_space > window + (full_space >> 1))
2675 window = free_space;
2681 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2682 const struct sk_buff *next_skb)
2684 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2685 const struct skb_shared_info *next_shinfo =
2686 skb_shinfo(next_skb);
2687 struct skb_shared_info *shinfo = skb_shinfo(skb);
2689 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2690 shinfo->tskey = next_shinfo->tskey;
2691 TCP_SKB_CB(skb)->txstamp_ack |=
2692 TCP_SKB_CB(next_skb)->txstamp_ack;
2696 /* Collapses two adjacent SKB's during retransmission. */
2697 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2699 struct tcp_sock *tp = tcp_sk(sk);
2700 struct sk_buff *next_skb = skb_rb_next(skb);
2701 int skb_size, next_skb_size;
2703 skb_size = skb->len;
2704 next_skb_size = next_skb->len;
2706 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2708 if (next_skb_size) {
2709 if (next_skb_size <= skb_availroom(skb))
2710 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2712 else if (!skb_shift(skb, next_skb, next_skb_size))
2715 tcp_highest_sack_replace(sk, next_skb, skb);
2717 /* Update sequence range on original skb. */
2718 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2720 /* Merge over control information. This moves PSH/FIN etc. over */
2721 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2723 /* All done, get rid of second SKB and account for it so
2724 * packet counting does not break.
2726 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2727 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2729 /* changed transmit queue under us so clear hints */
2730 tcp_clear_retrans_hints_partial(tp);
2731 if (next_skb == tp->retransmit_skb_hint)
2732 tp->retransmit_skb_hint = skb;
2734 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2736 tcp_skb_collapse_tstamp(skb, next_skb);
2738 tcp_rtx_queue_unlink_and_free(next_skb, sk);
2742 /* Check if coalescing SKBs is legal. */
2743 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2745 if (tcp_skb_pcount(skb) > 1)
2747 if (skb_cloned(skb))
2749 /* Some heuristics for collapsing over SACK'd could be invented */
2750 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2756 /* Collapse packets in the retransmit queue to make to create
2757 * less packets on the wire. This is only done on retransmission.
2759 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2762 struct tcp_sock *tp = tcp_sk(sk);
2763 struct sk_buff *skb = to, *tmp;
2766 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2768 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2771 skb_rbtree_walk_from_safe(skb, tmp) {
2772 if (!tcp_can_collapse(sk, skb))
2775 if (!tcp_skb_can_collapse_to(to))
2788 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2791 if (!tcp_collapse_retrans(sk, to))
2796 /* This retransmits one SKB. Policy decisions and retransmit queue
2797 * state updates are done by the caller. Returns non-zero if an
2798 * error occurred which prevented the send.
2800 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2802 struct inet_connection_sock *icsk = inet_csk(sk);
2803 struct tcp_sock *tp = tcp_sk(sk);
2804 unsigned int cur_mss;
2808 /* Inconclusive MTU probe */
2809 if (icsk->icsk_mtup.probe_size)
2810 icsk->icsk_mtup.probe_size = 0;
2812 /* Do not sent more than we queued. 1/4 is reserved for possible
2813 * copying overhead: fragmentation, tunneling, mangling etc.
2815 if (refcount_read(&sk->sk_wmem_alloc) >
2816 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2820 if (skb_still_in_host_queue(sk, skb))
2823 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2824 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2826 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2830 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2831 return -EHOSTUNREACH; /* Routing failure or similar. */
2833 cur_mss = tcp_current_mss(sk);
2835 /* If receiver has shrunk his window, and skb is out of
2836 * new window, do not retransmit it. The exception is the
2837 * case, when window is shrunk to zero. In this case
2838 * our retransmit serves as a zero window probe.
2840 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2841 TCP_SKB_CB(skb)->seq != tp->snd_una)
2844 len = cur_mss * segs;
2845 if (skb->len > len) {
2846 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2847 cur_mss, GFP_ATOMIC))
2848 return -ENOMEM; /* We'll try again later. */
2850 if (skb_unclone(skb, GFP_ATOMIC))
2853 diff = tcp_skb_pcount(skb);
2854 tcp_set_skb_tso_segs(skb, cur_mss);
2855 diff -= tcp_skb_pcount(skb);
2857 tcp_adjust_pcount(sk, skb, diff);
2858 if (skb->len < cur_mss)
2859 tcp_retrans_try_collapse(sk, skb, cur_mss);
2862 /* RFC3168, section 6.1.1.1. ECN fallback */
2863 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2864 tcp_ecn_clear_syn(sk, skb);
2866 /* Update global and local TCP statistics. */
2867 segs = tcp_skb_pcount(skb);
2868 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2869 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2870 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2871 tp->total_retrans += segs;
2873 /* make sure skb->data is aligned on arches that require it
2874 * and check if ack-trimming & collapsing extended the headroom
2875 * beyond what csum_start can cover.
2877 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2878 skb_headroom(skb) >= 0xFFFF)) {
2879 struct sk_buff *nskb;
2881 tcp_skb_tsorted_save(skb) {
2882 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2883 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2885 } tcp_skb_tsorted_restore(skb);
2888 tcp_update_skb_after_send(tp, skb);
2889 tcp_rate_skb_sent(sk, skb);
2892 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2895 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
2896 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
2897 TCP_SKB_CB(skb)->seq, segs, err);
2900 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2901 trace_tcp_retransmit_skb(sk, skb);
2902 } else if (err != -EBUSY) {
2903 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2908 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2910 struct tcp_sock *tp = tcp_sk(sk);
2911 int err = __tcp_retransmit_skb(sk, skb, segs);
2914 #if FASTRETRANS_DEBUG > 0
2915 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2916 net_dbg_ratelimited("retrans_out leaked\n");
2919 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2920 tp->retrans_out += tcp_skb_pcount(skb);
2922 /* Save stamp of the first retransmit. */
2923 if (!tp->retrans_stamp)
2924 tp->retrans_stamp = tcp_skb_timestamp(skb);
2928 if (tp->undo_retrans < 0)
2929 tp->undo_retrans = 0;
2930 tp->undo_retrans += tcp_skb_pcount(skb);
2934 /* This gets called after a retransmit timeout, and the initially
2935 * retransmitted data is acknowledged. It tries to continue
2936 * resending the rest of the retransmit queue, until either
2937 * we've sent it all or the congestion window limit is reached.
2939 void tcp_xmit_retransmit_queue(struct sock *sk)
2941 const struct inet_connection_sock *icsk = inet_csk(sk);
2942 struct sk_buff *skb, *rtx_head, *hole = NULL;
2943 struct tcp_sock *tp = tcp_sk(sk);
2947 if (!tp->packets_out)
2950 rtx_head = tcp_rtx_queue_head(sk);
2951 skb = tp->retransmit_skb_hint ?: rtx_head;
2952 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2953 skb_rbtree_walk_from(skb) {
2957 if (tcp_pacing_check(sk))
2960 /* we could do better than to assign each time */
2962 tp->retransmit_skb_hint = skb;
2964 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2967 sacked = TCP_SKB_CB(skb)->sacked;
2968 /* In case tcp_shift_skb_data() have aggregated large skbs,
2969 * we need to make sure not sending too bigs TSO packets
2971 segs = min_t(int, segs, max_segs);
2973 if (tp->retrans_out >= tp->lost_out) {
2975 } else if (!(sacked & TCPCB_LOST)) {
2976 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2981 if (icsk->icsk_ca_state != TCP_CA_Loss)
2982 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2984 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2987 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2990 if (tcp_small_queue_check(sk, skb, 1))
2993 if (tcp_retransmit_skb(sk, skb, segs))
2996 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
2998 if (tcp_in_cwnd_reduction(sk))
2999 tp->prr_out += tcp_skb_pcount(skb);
3001 if (skb == rtx_head &&
3002 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3003 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3004 inet_csk(sk)->icsk_rto,
3009 /* We allow to exceed memory limits for FIN packets to expedite
3010 * connection tear down and (memory) recovery.
3011 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3012 * or even be forced to close flow without any FIN.
3013 * In general, we want to allow one skb per socket to avoid hangs
3014 * with edge trigger epoll()
3016 void sk_forced_mem_schedule(struct sock *sk, int size)
3020 if (size <= sk->sk_forward_alloc)
3022 amt = sk_mem_pages(size);
3023 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3024 sk_memory_allocated_add(sk, amt);
3026 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3027 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3030 /* Send a FIN. The caller locks the socket for us.
3031 * We should try to send a FIN packet really hard, but eventually give up.
3033 void tcp_send_fin(struct sock *sk)
3035 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3036 struct tcp_sock *tp = tcp_sk(sk);
3038 /* Optimization, tack on the FIN if we have one skb in write queue and
3039 * this skb was not yet sent, or we are under memory pressure.
3040 * Note: in the latter case, FIN packet will be sent after a timeout,
3041 * as TCP stack thinks it has already been transmitted.
3043 if (!tskb && tcp_under_memory_pressure(sk))
3044 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3048 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3049 TCP_SKB_CB(tskb)->end_seq++;
3051 if (tcp_write_queue_empty(sk)) {
3052 /* This means tskb was already sent.
3053 * Pretend we included the FIN on previous transmit.
3054 * We need to set tp->snd_nxt to the value it would have
3055 * if FIN had been sent. This is because retransmit path
3056 * does not change tp->snd_nxt.
3062 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3063 if (unlikely(!skb)) {
3068 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3069 skb_reserve(skb, MAX_TCP_HEADER);
3070 sk_forced_mem_schedule(sk, skb->truesize);
3071 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3072 tcp_init_nondata_skb(skb, tp->write_seq,
3073 TCPHDR_ACK | TCPHDR_FIN);
3074 tcp_queue_skb(sk, skb);
3076 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3079 /* We get here when a process closes a file descriptor (either due to
3080 * an explicit close() or as a byproduct of exit()'ing) and there
3081 * was unread data in the receive queue. This behavior is recommended
3082 * by RFC 2525, section 2.17. -DaveM
3084 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3086 struct sk_buff *skb;
3088 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3090 /* NOTE: No TCP options attached and we never retransmit this. */
3091 skb = alloc_skb(MAX_TCP_HEADER, priority);
3093 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3097 /* Reserve space for headers and prepare control bits. */
3098 skb_reserve(skb, MAX_TCP_HEADER);
3099 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3100 TCPHDR_ACK | TCPHDR_RST);
3101 tcp_mstamp_refresh(tcp_sk(sk));
3103 if (tcp_transmit_skb(sk, skb, 0, priority))
3104 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3106 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3107 * skb here is different to the troublesome skb, so use NULL
3109 trace_tcp_send_reset(sk, NULL);
3112 /* Send a crossed SYN-ACK during socket establishment.
3113 * WARNING: This routine must only be called when we have already sent
3114 * a SYN packet that crossed the incoming SYN that caused this routine
3115 * to get called. If this assumption fails then the initial rcv_wnd
3116 * and rcv_wscale values will not be correct.
3118 int tcp_send_synack(struct sock *sk)
3120 struct sk_buff *skb;
3122 skb = tcp_rtx_queue_head(sk);
3123 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3124 pr_err("%s: wrong queue state\n", __func__);
3127 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3128 if (skb_cloned(skb)) {
3129 struct sk_buff *nskb;
3131 tcp_skb_tsorted_save(skb) {
3132 nskb = skb_copy(skb, GFP_ATOMIC);
3133 } tcp_skb_tsorted_restore(skb);
3136 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3137 tcp_rtx_queue_unlink_and_free(skb, sk);
3138 __skb_header_release(nskb);
3139 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3140 sk->sk_wmem_queued += nskb->truesize;
3141 sk_mem_charge(sk, nskb->truesize);
3145 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3146 tcp_ecn_send_synack(sk, skb);
3148 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3152 * tcp_make_synack - Prepare a SYN-ACK.
3153 * sk: listener socket
3154 * dst: dst entry attached to the SYNACK
3155 * req: request_sock pointer
3157 * Allocate one skb and build a SYNACK packet.
3158 * @dst is consumed : Caller should not use it again.
3160 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3161 struct request_sock *req,
3162 struct tcp_fastopen_cookie *foc,
3163 enum tcp_synack_type synack_type)
3165 struct inet_request_sock *ireq = inet_rsk(req);
3166 const struct tcp_sock *tp = tcp_sk(sk);
3167 struct tcp_md5sig_key *md5 = NULL;
3168 struct tcp_out_options opts;
3169 struct sk_buff *skb;
3170 int tcp_header_size;
3174 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3175 if (unlikely(!skb)) {
3179 /* Reserve space for headers. */
3180 skb_reserve(skb, MAX_TCP_HEADER);
3182 switch (synack_type) {
3183 case TCP_SYNACK_NORMAL:
3184 skb_set_owner_w(skb, req_to_sk(req));
3186 case TCP_SYNACK_COOKIE:
3187 /* Under synflood, we do not attach skb to a socket,
3188 * to avoid false sharing.
3191 case TCP_SYNACK_FASTOPEN:
3192 /* sk is a const pointer, because we want to express multiple
3193 * cpu might call us concurrently.
3194 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3196 skb_set_owner_w(skb, (struct sock *)sk);
3199 skb_dst_set(skb, dst);
3201 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3203 memset(&opts, 0, sizeof(opts));
3204 #ifdef CONFIG_SYN_COOKIES
3205 if (unlikely(req->cookie_ts))
3206 skb->skb_mstamp = cookie_init_timestamp(req);
3209 skb->skb_mstamp = tcp_clock_us();
3211 #ifdef CONFIG_TCP_MD5SIG
3213 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3215 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3216 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3219 skb_push(skb, tcp_header_size);
3220 skb_reset_transport_header(skb);
3222 th = (struct tcphdr *)skb->data;
3223 memset(th, 0, sizeof(struct tcphdr));
3226 tcp_ecn_make_synack(req, th);
3227 th->source = htons(ireq->ir_num);
3228 th->dest = ireq->ir_rmt_port;
3229 skb->mark = ireq->ir_mark;
3230 skb->ip_summed = CHECKSUM_PARTIAL;
3231 th->seq = htonl(tcp_rsk(req)->snt_isn);
3232 /* XXX data is queued and acked as is. No buffer/window check */
3233 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3235 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3236 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3237 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3238 th->doff = (tcp_header_size >> 2);
3239 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3241 #ifdef CONFIG_TCP_MD5SIG
3242 /* Okay, we have all we need - do the md5 hash if needed */
3244 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3245 md5, req_to_sk(req), skb);
3249 /* Do not fool tcpdump (if any), clean our debris */
3253 EXPORT_SYMBOL(tcp_make_synack);
3255 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3257 struct inet_connection_sock *icsk = inet_csk(sk);
3258 const struct tcp_congestion_ops *ca;
3259 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3261 if (ca_key == TCP_CA_UNSPEC)
3265 ca = tcp_ca_find_key(ca_key);
3266 if (likely(ca && try_module_get(ca->owner))) {
3267 module_put(icsk->icsk_ca_ops->owner);
3268 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3269 icsk->icsk_ca_ops = ca;
3274 /* Do all connect socket setups that can be done AF independent. */
3275 static void tcp_connect_init(struct sock *sk)
3277 const struct dst_entry *dst = __sk_dst_get(sk);
3278 struct tcp_sock *tp = tcp_sk(sk);
3282 /* We'll fix this up when we get a response from the other end.
3283 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3285 tp->tcp_header_len = sizeof(struct tcphdr);
3286 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3287 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3289 #ifdef CONFIG_TCP_MD5SIG
3290 if (tp->af_specific->md5_lookup(sk, sk))
3291 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3294 /* If user gave his TCP_MAXSEG, record it to clamp */
3295 if (tp->rx_opt.user_mss)
3296 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3299 tcp_sync_mss(sk, dst_mtu(dst));
3301 tcp_ca_dst_init(sk, dst);
3303 if (!tp->window_clamp)
3304 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3305 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3307 tcp_initialize_rcv_mss(sk);
3309 /* limit the window selection if the user enforce a smaller rx buffer */
3310 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3311 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3312 tp->window_clamp = tcp_full_space(sk);
3314 rcv_wnd = tcp_rwnd_init_bpf(sk);
3316 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3318 tcp_select_initial_window(sk, tcp_full_space(sk),
3319 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3322 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3326 tp->rx_opt.rcv_wscale = rcv_wscale;
3327 tp->rcv_ssthresh = tp->rcv_wnd;
3330 sock_reset_flag(sk, SOCK_DONE);
3333 tp->snd_una = tp->write_seq;
3334 tp->snd_sml = tp->write_seq;
3335 tp->snd_up = tp->write_seq;
3336 tp->snd_nxt = tp->write_seq;
3338 if (likely(!tp->repair))
3341 tp->rcv_tstamp = tcp_jiffies32;
3342 tp->rcv_wup = tp->rcv_nxt;
3343 tp->copied_seq = tp->rcv_nxt;
3345 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3346 inet_csk(sk)->icsk_retransmits = 0;
3347 tcp_clear_retrans(tp);
3350 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3352 struct tcp_sock *tp = tcp_sk(sk);
3353 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3355 tcb->end_seq += skb->len;
3356 __skb_header_release(skb);
3357 sk->sk_wmem_queued += skb->truesize;
3358 sk_mem_charge(sk, skb->truesize);
3359 tp->write_seq = tcb->end_seq;
3360 tp->packets_out += tcp_skb_pcount(skb);
3363 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3364 * queue a data-only packet after the regular SYN, such that regular SYNs
3365 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3366 * only the SYN sequence, the data are retransmitted in the first ACK.
3367 * If cookie is not cached or other error occurs, falls back to send a
3368 * regular SYN with Fast Open cookie request option.
3370 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3372 struct tcp_sock *tp = tcp_sk(sk);
3373 struct tcp_fastopen_request *fo = tp->fastopen_req;
3375 struct sk_buff *syn_data;
3377 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3378 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3381 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3382 * user-MSS. Reserve maximum option space for middleboxes that add
3383 * private TCP options. The cost is reduced data space in SYN :(
3385 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3387 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3388 MAX_TCP_OPTION_SPACE;
3390 space = min_t(size_t, space, fo->size);
3392 /* limit to order-0 allocations */
3393 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3395 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3398 syn_data->ip_summed = CHECKSUM_PARTIAL;
3399 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3401 int copied = copy_from_iter(skb_put(syn_data, space), space,
3402 &fo->data->msg_iter);
3403 if (unlikely(!copied)) {
3404 tcp_skb_tsorted_anchor_cleanup(syn_data);
3405 kfree_skb(syn_data);
3408 if (copied != space) {
3409 skb_trim(syn_data, copied);
3413 /* No more data pending in inet_wait_for_connect() */
3414 if (space == fo->size)
3418 tcp_connect_queue_skb(sk, syn_data);
3420 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3422 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3424 syn->skb_mstamp = syn_data->skb_mstamp;
3426 /* Now full SYN+DATA was cloned and sent (or not),
3427 * remove the SYN from the original skb (syn_data)
3428 * we keep in write queue in case of a retransmit, as we
3429 * also have the SYN packet (with no data) in the same queue.
3431 TCP_SKB_CB(syn_data)->seq++;
3432 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3434 tp->syn_data = (fo->copied > 0);
3435 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3436 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3440 /* data was not sent, put it in write_queue */
3441 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3442 tp->packets_out -= tcp_skb_pcount(syn_data);
3445 /* Send a regular SYN with Fast Open cookie request option */
3446 if (fo->cookie.len > 0)
3448 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3450 tp->syn_fastopen = 0;
3452 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3456 /* Build a SYN and send it off. */
3457 int tcp_connect(struct sock *sk)
3459 struct tcp_sock *tp = tcp_sk(sk);
3460 struct sk_buff *buff;
3463 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3465 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3466 return -EHOSTUNREACH; /* Routing failure or similar. */
3468 tcp_connect_init(sk);
3470 if (unlikely(tp->repair)) {
3471 tcp_finish_connect(sk, NULL);
3475 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3476 if (unlikely(!buff))
3479 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3480 tcp_mstamp_refresh(tp);
3481 tp->retrans_stamp = tcp_time_stamp(tp);
3482 tcp_connect_queue_skb(sk, buff);
3483 tcp_ecn_send_syn(sk, buff);
3484 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3486 /* Send off SYN; include data in Fast Open. */
3487 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3488 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3489 if (err == -ECONNREFUSED)
3492 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3493 * in order to make this packet get counted in tcpOutSegs.
3495 tp->snd_nxt = tp->write_seq;
3496 tp->pushed_seq = tp->write_seq;
3497 buff = tcp_send_head(sk);
3498 if (unlikely(buff)) {
3499 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3500 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3502 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3504 /* Timer for repeating the SYN until an answer. */
3505 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3506 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3509 EXPORT_SYMBOL(tcp_connect);
3511 /* Send out a delayed ack, the caller does the policy checking
3512 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3515 void tcp_send_delayed_ack(struct sock *sk)
3517 struct inet_connection_sock *icsk = inet_csk(sk);
3518 int ato = icsk->icsk_ack.ato;
3519 unsigned long timeout;
3521 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3523 if (ato > TCP_DELACK_MIN) {
3524 const struct tcp_sock *tp = tcp_sk(sk);
3525 int max_ato = HZ / 2;
3527 if (icsk->icsk_ack.pingpong ||
3528 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3529 max_ato = TCP_DELACK_MAX;
3531 /* Slow path, intersegment interval is "high". */
3533 /* If some rtt estimate is known, use it to bound delayed ack.
3534 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3538 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3545 ato = min(ato, max_ato);
3548 /* Stay within the limit we were given */
3549 timeout = jiffies + ato;
3551 /* Use new timeout only if there wasn't a older one earlier. */
3552 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3553 /* If delack timer was blocked or is about to expire,
3556 if (icsk->icsk_ack.blocked ||
3557 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3562 if (!time_before(timeout, icsk->icsk_ack.timeout))
3563 timeout = icsk->icsk_ack.timeout;
3565 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3566 icsk->icsk_ack.timeout = timeout;
3567 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3570 /* This routine sends an ack and also updates the window. */
3571 void tcp_send_ack(struct sock *sk)
3573 struct sk_buff *buff;
3575 /* If we have been reset, we may not send again. */
3576 if (sk->sk_state == TCP_CLOSE)
3579 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3581 /* We are not putting this on the write queue, so
3582 * tcp_transmit_skb() will set the ownership to this
3585 buff = alloc_skb(MAX_TCP_HEADER,
3586 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3587 if (unlikely(!buff)) {
3588 inet_csk_schedule_ack(sk);
3589 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3590 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3591 TCP_DELACK_MAX, TCP_RTO_MAX);
3595 /* Reserve space for headers and prepare control bits. */
3596 skb_reserve(buff, MAX_TCP_HEADER);
3597 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3599 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3601 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3603 skb_set_tcp_pure_ack(buff);
3605 /* Send it off, this clears delayed acks for us. */
3606 tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3608 EXPORT_SYMBOL_GPL(tcp_send_ack);
3610 /* This routine sends a packet with an out of date sequence
3611 * number. It assumes the other end will try to ack it.
3613 * Question: what should we make while urgent mode?
3614 * 4.4BSD forces sending single byte of data. We cannot send
3615 * out of window data, because we have SND.NXT==SND.MAX...
3617 * Current solution: to send TWO zero-length segments in urgent mode:
3618 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3619 * out-of-date with SND.UNA-1 to probe window.
3621 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3623 struct tcp_sock *tp = tcp_sk(sk);
3624 struct sk_buff *skb;
3626 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3627 skb = alloc_skb(MAX_TCP_HEADER,
3628 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3632 /* Reserve space for headers and set control bits. */
3633 skb_reserve(skb, MAX_TCP_HEADER);
3634 /* Use a previous sequence. This should cause the other
3635 * end to send an ack. Don't queue or clone SKB, just
3638 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3639 NET_INC_STATS(sock_net(sk), mib);
3640 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3643 /* Called from setsockopt( ... TCP_REPAIR ) */
3644 void tcp_send_window_probe(struct sock *sk)
3646 if (sk->sk_state == TCP_ESTABLISHED) {
3647 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3648 tcp_mstamp_refresh(tcp_sk(sk));
3649 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3653 /* Initiate keepalive or window probe from timer. */
3654 int tcp_write_wakeup(struct sock *sk, int mib)
3656 struct tcp_sock *tp = tcp_sk(sk);
3657 struct sk_buff *skb;
3659 if (sk->sk_state == TCP_CLOSE)
3662 skb = tcp_send_head(sk);
3663 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3665 unsigned int mss = tcp_current_mss(sk);
3666 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3668 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3669 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3671 /* We are probing the opening of a window
3672 * but the window size is != 0
3673 * must have been a result SWS avoidance ( sender )
3675 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3677 seg_size = min(seg_size, mss);
3678 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3679 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3680 skb, seg_size, mss, GFP_ATOMIC))
3682 } else if (!tcp_skb_pcount(skb))
3683 tcp_set_skb_tso_segs(skb, mss);
3685 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3686 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3688 tcp_event_new_data_sent(sk, skb);
3691 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3692 tcp_xmit_probe_skb(sk, 1, mib);
3693 return tcp_xmit_probe_skb(sk, 0, mib);
3697 /* A window probe timeout has occurred. If window is not closed send
3698 * a partial packet else a zero probe.
3700 void tcp_send_probe0(struct sock *sk)
3702 struct inet_connection_sock *icsk = inet_csk(sk);
3703 struct tcp_sock *tp = tcp_sk(sk);
3704 struct net *net = sock_net(sk);
3705 unsigned long probe_max;
3708 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3710 if (tp->packets_out || tcp_write_queue_empty(sk)) {
3711 /* Cancel probe timer, if it is not required. */
3712 icsk->icsk_probes_out = 0;
3713 icsk->icsk_backoff = 0;
3718 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3719 icsk->icsk_backoff++;
3720 icsk->icsk_probes_out++;
3721 probe_max = TCP_RTO_MAX;
3723 /* If packet was not sent due to local congestion,
3724 * do not backoff and do not remember icsk_probes_out.
3725 * Let local senders to fight for local resources.
3727 * Use accumulated backoff yet.
3729 if (!icsk->icsk_probes_out)
3730 icsk->icsk_probes_out = 1;
3731 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3733 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3734 tcp_probe0_when(sk, probe_max),
3738 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3740 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3744 tcp_rsk(req)->txhash = net_tx_rndhash();
3745 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3747 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3748 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3749 if (unlikely(tcp_passive_fastopen(sk)))
3750 tcp_sk(sk)->total_retrans++;
3751 trace_tcp_retransmit_synack(sk, req);
3755 EXPORT_SYMBOL(tcp_rtx_synack);