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>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly = 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
53 /* Default TSQ limit of four TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly = 262144;
56 /* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
60 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
65 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
66 int push_one, gfp_t gfp);
68 /* Account for new data that has been sent to the network. */
69 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
71 struct inet_connection_sock *icsk = inet_csk(sk);
72 struct tcp_sock *tp = tcp_sk(sk);
73 unsigned int prior_packets = tp->packets_out;
75 tcp_advance_send_head(sk, skb);
76 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
78 tp->packets_out += tcp_skb_pcount(skb);
79 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
82 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
86 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
87 * window scaling factor due to loss of precision.
88 * If window has been shrunk, what should we make? It is not clear at all.
89 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
90 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
91 * invalid. OK, let's make this for now:
93 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
95 const struct tcp_sock *tp = tcp_sk(sk);
97 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
98 (tp->rx_opt.wscale_ok &&
99 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
102 return tcp_wnd_end(tp);
105 /* Calculate mss to advertise in SYN segment.
106 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
108 * 1. It is independent of path mtu.
109 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
110 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
111 * attached devices, because some buggy hosts are confused by
113 * 4. We do not make 3, we advertise MSS, calculated from first
114 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
115 * This may be overridden via information stored in routing table.
116 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
117 * probably even Jumbo".
119 static __u16 tcp_advertise_mss(struct sock *sk)
121 struct tcp_sock *tp = tcp_sk(sk);
122 const struct dst_entry *dst = __sk_dst_get(sk);
123 int mss = tp->advmss;
126 unsigned int metric = dst_metric_advmss(dst);
137 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
138 * This is the first part of cwnd validation mechanism.
140 void tcp_cwnd_restart(struct sock *sk, s32 delta)
142 struct tcp_sock *tp = tcp_sk(sk);
143 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
144 u32 cwnd = tp->snd_cwnd;
146 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
148 tp->snd_ssthresh = tcp_current_ssthresh(sk);
149 restart_cwnd = min(restart_cwnd, cwnd);
151 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
153 tp->snd_cwnd = max(cwnd, restart_cwnd);
154 tp->snd_cwnd_stamp = tcp_jiffies32;
155 tp->snd_cwnd_used = 0;
158 /* Congestion state accounting after a packet has been sent. */
159 static void tcp_event_data_sent(struct tcp_sock *tp,
162 struct inet_connection_sock *icsk = inet_csk(sk);
163 const u32 now = tcp_jiffies32;
165 if (tcp_packets_in_flight(tp) == 0)
166 tcp_ca_event(sk, CA_EVENT_TX_START);
170 /* If it is a reply for ato after last received
171 * packet, enter pingpong mode.
173 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
174 icsk->icsk_ack.pingpong = 1;
177 /* Account for an ACK we sent. */
178 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
180 tcp_dec_quickack_mode(sk, pkts);
181 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
185 u32 tcp_default_init_rwnd(u32 mss)
187 /* Initial receive window should be twice of TCP_INIT_CWND to
188 * enable proper sending of new unsent data during fast recovery
189 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
190 * limit when mss is larger than 1460.
192 u32 init_rwnd = TCP_INIT_CWND * 2;
195 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
199 /* Determine a window scaling and initial window to offer.
200 * Based on the assumption that the given amount of space
201 * will be offered. Store the results in the tp structure.
202 * NOTE: for smooth operation initial space offering should
203 * be a multiple of mss if possible. We assume here that mss >= 1.
204 * This MUST be enforced by all callers.
206 void tcp_select_initial_window(int __space, __u32 mss,
207 __u32 *rcv_wnd, __u32 *window_clamp,
208 int wscale_ok, __u8 *rcv_wscale,
211 unsigned int space = (__space < 0 ? 0 : __space);
213 /* If no clamp set the clamp to the max possible scaled window */
214 if (*window_clamp == 0)
215 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
216 space = min(*window_clamp, space);
218 /* Quantize space offering to a multiple of mss if possible. */
220 space = rounddown(space, mss);
222 /* NOTE: offering an initial window larger than 32767
223 * will break some buggy TCP stacks. If the admin tells us
224 * it is likely we could be speaking with such a buggy stack
225 * we will truncate our initial window offering to 32K-1
226 * unless the remote has sent us a window scaling option,
227 * which we interpret as a sign the remote TCP is not
228 * misinterpreting the window field as a signed quantity.
230 if (sysctl_tcp_workaround_signed_windows)
231 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
237 /* Set window scaling on max possible window */
238 space = max_t(u32, space, sysctl_tcp_rmem[2]);
239 space = max_t(u32, space, sysctl_rmem_max);
240 space = min_t(u32, space, *window_clamp);
241 while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
247 if (mss > (1 << *rcv_wscale)) {
248 if (!init_rcv_wnd) /* Use default unless specified otherwise */
249 init_rcv_wnd = tcp_default_init_rwnd(mss);
250 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
253 /* Set the clamp no higher than max representable value */
254 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
256 EXPORT_SYMBOL(tcp_select_initial_window);
258 /* Chose a new window to advertise, update state in tcp_sock for the
259 * socket, and return result with RFC1323 scaling applied. The return
260 * value can be stuffed directly into th->window for an outgoing
263 static u16 tcp_select_window(struct sock *sk)
265 struct tcp_sock *tp = tcp_sk(sk);
266 u32 old_win = tp->rcv_wnd;
267 u32 cur_win = tcp_receive_window(tp);
268 u32 new_win = __tcp_select_window(sk);
270 /* Never shrink the offered window */
271 if (new_win < cur_win) {
272 /* Danger Will Robinson!
273 * Don't update rcv_wup/rcv_wnd here or else
274 * we will not be able to advertise a zero
275 * window in time. --DaveM
277 * Relax Will Robinson.
280 NET_INC_STATS(sock_net(sk),
281 LINUX_MIB_TCPWANTZEROWINDOWADV);
282 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
284 tp->rcv_wnd = new_win;
285 tp->rcv_wup = tp->rcv_nxt;
287 /* Make sure we do not exceed the maximum possible
290 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
291 new_win = min(new_win, MAX_TCP_WINDOW);
293 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
295 /* RFC1323 scaling applied */
296 new_win >>= tp->rx_opt.rcv_wscale;
298 /* If we advertise zero window, disable fast path. */
302 NET_INC_STATS(sock_net(sk),
303 LINUX_MIB_TCPTOZEROWINDOWADV);
304 } else if (old_win == 0) {
305 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
311 /* Packet ECN state for a SYN-ACK */
312 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
314 const struct tcp_sock *tp = tcp_sk(sk);
316 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
317 if (!(tp->ecn_flags & TCP_ECN_OK))
318 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
319 else if (tcp_ca_needs_ecn(sk) ||
320 tcp_bpf_ca_needs_ecn(sk))
324 /* Packet ECN state for a SYN. */
325 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
327 struct tcp_sock *tp = tcp_sk(sk);
328 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
329 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
330 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
333 const struct dst_entry *dst = __sk_dst_get(sk);
335 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
342 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
343 tp->ecn_flags = TCP_ECN_OK;
344 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
349 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
351 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
352 /* tp->ecn_flags are cleared at a later point in time when
353 * SYN ACK is ultimatively being received.
355 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
359 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
361 if (inet_rsk(req)->ecn_ok)
365 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
368 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
369 struct tcphdr *th, int tcp_header_len)
371 struct tcp_sock *tp = tcp_sk(sk);
373 if (tp->ecn_flags & TCP_ECN_OK) {
374 /* Not-retransmitted data segment: set ECT and inject CWR. */
375 if (skb->len != tcp_header_len &&
376 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
378 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
379 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
381 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
383 } else if (!tcp_ca_needs_ecn(sk)) {
384 /* ACK or retransmitted segment: clear ECT|CE */
385 INET_ECN_dontxmit(sk);
387 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
392 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
393 * auto increment end seqno.
395 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
397 skb->ip_summed = CHECKSUM_PARTIAL;
400 TCP_SKB_CB(skb)->tcp_flags = flags;
401 TCP_SKB_CB(skb)->sacked = 0;
403 tcp_skb_pcount_set(skb, 1);
405 TCP_SKB_CB(skb)->seq = seq;
406 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
408 TCP_SKB_CB(skb)->end_seq = seq;
411 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
413 return tp->snd_una != tp->snd_up;
416 #define OPTION_SACK_ADVERTISE (1 << 0)
417 #define OPTION_TS (1 << 1)
418 #define OPTION_MD5 (1 << 2)
419 #define OPTION_WSCALE (1 << 3)
420 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
422 struct tcp_out_options {
423 u16 options; /* bit field of OPTION_* */
424 u16 mss; /* 0 to disable */
425 u8 ws; /* window scale, 0 to disable */
426 u8 num_sack_blocks; /* number of SACK blocks to include */
427 u8 hash_size; /* bytes in hash_location */
428 __u8 *hash_location; /* temporary pointer, overloaded */
429 __u32 tsval, tsecr; /* need to include OPTION_TS */
430 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
433 /* Write previously computed TCP options to the packet.
435 * Beware: Something in the Internet is very sensitive to the ordering of
436 * TCP options, we learned this through the hard way, so be careful here.
437 * Luckily we can at least blame others for their non-compliance but from
438 * inter-operability perspective it seems that we're somewhat stuck with
439 * the ordering which we have been using if we want to keep working with
440 * those broken things (not that it currently hurts anybody as there isn't
441 * particular reason why the ordering would need to be changed).
443 * At least SACK_PERM as the first option is known to lead to a disaster
444 * (but it may well be that other scenarios fail similarly).
446 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
447 struct tcp_out_options *opts)
449 u16 options = opts->options; /* mungable copy */
451 if (unlikely(OPTION_MD5 & options)) {
452 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
453 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
454 /* overload cookie hash location */
455 opts->hash_location = (__u8 *)ptr;
459 if (unlikely(opts->mss)) {
460 *ptr++ = htonl((TCPOPT_MSS << 24) |
461 (TCPOLEN_MSS << 16) |
465 if (likely(OPTION_TS & options)) {
466 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
467 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
468 (TCPOLEN_SACK_PERM << 16) |
469 (TCPOPT_TIMESTAMP << 8) |
471 options &= ~OPTION_SACK_ADVERTISE;
473 *ptr++ = htonl((TCPOPT_NOP << 24) |
475 (TCPOPT_TIMESTAMP << 8) |
478 *ptr++ = htonl(opts->tsval);
479 *ptr++ = htonl(opts->tsecr);
482 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
483 *ptr++ = htonl((TCPOPT_NOP << 24) |
485 (TCPOPT_SACK_PERM << 8) |
489 if (unlikely(OPTION_WSCALE & options)) {
490 *ptr++ = htonl((TCPOPT_NOP << 24) |
491 (TCPOPT_WINDOW << 16) |
492 (TCPOLEN_WINDOW << 8) |
496 if (unlikely(opts->num_sack_blocks)) {
497 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
498 tp->duplicate_sack : tp->selective_acks;
501 *ptr++ = htonl((TCPOPT_NOP << 24) |
504 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
505 TCPOLEN_SACK_PERBLOCK)));
507 for (this_sack = 0; this_sack < opts->num_sack_blocks;
509 *ptr++ = htonl(sp[this_sack].start_seq);
510 *ptr++ = htonl(sp[this_sack].end_seq);
513 tp->rx_opt.dsack = 0;
516 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
517 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
519 u32 len; /* Fast Open option length */
522 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
523 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
524 TCPOPT_FASTOPEN_MAGIC);
525 p += TCPOLEN_EXP_FASTOPEN_BASE;
527 len = TCPOLEN_FASTOPEN_BASE + foc->len;
528 *p++ = TCPOPT_FASTOPEN;
532 memcpy(p, foc->val, foc->len);
533 if ((len & 3) == 2) {
534 p[foc->len] = TCPOPT_NOP;
535 p[foc->len + 1] = TCPOPT_NOP;
537 ptr += (len + 3) >> 2;
541 /* Compute TCP options for SYN packets. This is not the final
542 * network wire format yet.
544 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
545 struct tcp_out_options *opts,
546 struct tcp_md5sig_key **md5)
548 struct tcp_sock *tp = tcp_sk(sk);
549 unsigned int remaining = MAX_TCP_OPTION_SPACE;
550 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
552 #ifdef CONFIG_TCP_MD5SIG
553 *md5 = tp->af_specific->md5_lookup(sk, sk);
555 opts->options |= OPTION_MD5;
556 remaining -= TCPOLEN_MD5SIG_ALIGNED;
562 /* We always get an MSS option. The option bytes which will be seen in
563 * normal data packets should timestamps be used, must be in the MSS
564 * advertised. But we subtract them from tp->mss_cache so that
565 * calculations in tcp_sendmsg are simpler etc. So account for this
566 * fact here if necessary. If we don't do this correctly, as a
567 * receiver we won't recognize data packets as being full sized when we
568 * should, and thus we won't abide by the delayed ACK rules correctly.
569 * SACKs don't matter, we never delay an ACK when we have any of those
571 opts->mss = tcp_advertise_mss(sk);
572 remaining -= TCPOLEN_MSS_ALIGNED;
574 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
575 opts->options |= OPTION_TS;
576 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
577 opts->tsecr = tp->rx_opt.ts_recent;
578 remaining -= TCPOLEN_TSTAMP_ALIGNED;
580 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
581 opts->ws = tp->rx_opt.rcv_wscale;
582 opts->options |= OPTION_WSCALE;
583 remaining -= TCPOLEN_WSCALE_ALIGNED;
585 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
586 opts->options |= OPTION_SACK_ADVERTISE;
587 if (unlikely(!(OPTION_TS & opts->options)))
588 remaining -= TCPOLEN_SACKPERM_ALIGNED;
591 if (fastopen && fastopen->cookie.len >= 0) {
592 u32 need = fastopen->cookie.len;
594 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
595 TCPOLEN_FASTOPEN_BASE;
596 need = (need + 3) & ~3U; /* Align to 32 bits */
597 if (remaining >= need) {
598 opts->options |= OPTION_FAST_OPEN_COOKIE;
599 opts->fastopen_cookie = &fastopen->cookie;
601 tp->syn_fastopen = 1;
602 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
606 return MAX_TCP_OPTION_SPACE - remaining;
609 /* Set up TCP options for SYN-ACKs. */
610 static unsigned int tcp_synack_options(struct request_sock *req,
611 unsigned int mss, struct sk_buff *skb,
612 struct tcp_out_options *opts,
613 const struct tcp_md5sig_key *md5,
614 struct tcp_fastopen_cookie *foc)
616 struct inet_request_sock *ireq = inet_rsk(req);
617 unsigned int remaining = MAX_TCP_OPTION_SPACE;
619 #ifdef CONFIG_TCP_MD5SIG
621 opts->options |= OPTION_MD5;
622 remaining -= TCPOLEN_MD5SIG_ALIGNED;
624 /* We can't fit any SACK blocks in a packet with MD5 + TS
625 * options. There was discussion about disabling SACK
626 * rather than TS in order to fit in better with old,
627 * buggy kernels, but that was deemed to be unnecessary.
629 ireq->tstamp_ok &= !ireq->sack_ok;
633 /* We always send an MSS option. */
635 remaining -= TCPOLEN_MSS_ALIGNED;
637 if (likely(ireq->wscale_ok)) {
638 opts->ws = ireq->rcv_wscale;
639 opts->options |= OPTION_WSCALE;
640 remaining -= TCPOLEN_WSCALE_ALIGNED;
642 if (likely(ireq->tstamp_ok)) {
643 opts->options |= OPTION_TS;
644 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
645 opts->tsecr = req->ts_recent;
646 remaining -= TCPOLEN_TSTAMP_ALIGNED;
648 if (likely(ireq->sack_ok)) {
649 opts->options |= OPTION_SACK_ADVERTISE;
650 if (unlikely(!ireq->tstamp_ok))
651 remaining -= TCPOLEN_SACKPERM_ALIGNED;
653 if (foc != NULL && foc->len >= 0) {
656 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
657 TCPOLEN_FASTOPEN_BASE;
658 need = (need + 3) & ~3U; /* Align to 32 bits */
659 if (remaining >= need) {
660 opts->options |= OPTION_FAST_OPEN_COOKIE;
661 opts->fastopen_cookie = foc;
666 return MAX_TCP_OPTION_SPACE - remaining;
669 /* Compute TCP options for ESTABLISHED sockets. This is not the
670 * final wire format yet.
672 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
673 struct tcp_out_options *opts,
674 struct tcp_md5sig_key **md5)
676 struct tcp_sock *tp = tcp_sk(sk);
677 unsigned int size = 0;
678 unsigned int eff_sacks;
682 #ifdef CONFIG_TCP_MD5SIG
683 *md5 = tp->af_specific->md5_lookup(sk, sk);
684 if (unlikely(*md5)) {
685 opts->options |= OPTION_MD5;
686 size += TCPOLEN_MD5SIG_ALIGNED;
692 if (likely(tp->rx_opt.tstamp_ok)) {
693 opts->options |= OPTION_TS;
694 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
695 opts->tsecr = tp->rx_opt.ts_recent;
696 size += TCPOLEN_TSTAMP_ALIGNED;
699 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
700 if (unlikely(eff_sacks)) {
701 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
702 opts->num_sack_blocks =
703 min_t(unsigned int, eff_sacks,
704 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
705 TCPOLEN_SACK_PERBLOCK);
706 size += TCPOLEN_SACK_BASE_ALIGNED +
707 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
714 /* TCP SMALL QUEUES (TSQ)
716 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
717 * to reduce RTT and bufferbloat.
718 * We do this using a special skb destructor (tcp_wfree).
720 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
721 * needs to be reallocated in a driver.
722 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
724 * Since transmit from skb destructor is forbidden, we use a tasklet
725 * to process all sockets that eventually need to send more skbs.
726 * We use one tasklet per cpu, with its own queue of sockets.
729 struct tasklet_struct tasklet;
730 struct list_head head; /* queue of tcp sockets */
732 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
734 static void tcp_tsq_handler(struct sock *sk)
736 if ((1 << sk->sk_state) &
737 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
738 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
739 struct tcp_sock *tp = tcp_sk(sk);
741 if (tp->lost_out > tp->retrans_out &&
742 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
743 tcp_mstamp_refresh(tp);
744 tcp_xmit_retransmit_queue(sk);
747 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
752 * One tasklet per cpu tries to send more skbs.
753 * We run in tasklet context but need to disable irqs when
754 * transferring tsq->head because tcp_wfree() might
755 * interrupt us (non NAPI drivers)
757 static void tcp_tasklet_func(unsigned long data)
759 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
762 struct list_head *q, *n;
766 local_irq_save(flags);
767 list_splice_init(&tsq->head, &list);
768 local_irq_restore(flags);
770 list_for_each_safe(q, n, &list) {
771 tp = list_entry(q, struct tcp_sock, tsq_node);
772 list_del(&tp->tsq_node);
774 sk = (struct sock *)tp;
775 smp_mb__before_atomic();
776 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
778 if (!sk->sk_lock.owned &&
779 test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) {
781 if (!sock_owned_by_user(sk)) {
782 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
792 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
793 TCPF_WRITE_TIMER_DEFERRED | \
794 TCPF_DELACK_TIMER_DEFERRED | \
795 TCPF_MTU_REDUCED_DEFERRED)
797 * tcp_release_cb - tcp release_sock() callback
800 * called from release_sock() to perform protocol dependent
801 * actions before socket release.
803 void tcp_release_cb(struct sock *sk)
805 unsigned long flags, nflags;
807 /* perform an atomic operation only if at least one flag is set */
809 flags = sk->sk_tsq_flags;
810 if (!(flags & TCP_DEFERRED_ALL))
812 nflags = flags & ~TCP_DEFERRED_ALL;
813 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
815 if (flags & TCPF_TSQ_DEFERRED)
818 /* Here begins the tricky part :
819 * We are called from release_sock() with :
821 * 2) sk_lock.slock spinlock held
822 * 3) socket owned by us (sk->sk_lock.owned == 1)
824 * But following code is meant to be called from BH handlers,
825 * so we should keep BH disabled, but early release socket ownership
827 sock_release_ownership(sk);
829 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
830 tcp_write_timer_handler(sk);
833 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
834 tcp_delack_timer_handler(sk);
837 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
838 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
842 EXPORT_SYMBOL(tcp_release_cb);
844 void __init tcp_tasklet_init(void)
848 for_each_possible_cpu(i) {
849 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
851 INIT_LIST_HEAD(&tsq->head);
852 tasklet_init(&tsq->tasklet,
859 * Write buffer destructor automatically called from kfree_skb.
860 * We can't xmit new skbs from this context, as we might already
863 void tcp_wfree(struct sk_buff *skb)
865 struct sock *sk = skb->sk;
866 struct tcp_sock *tp = tcp_sk(sk);
867 unsigned long flags, nval, oval;
869 /* Keep one reference on sk_wmem_alloc.
870 * Will be released by sk_free() from here or tcp_tasklet_func()
872 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
874 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
875 * Wait until our queues (qdisc + devices) are drained.
877 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
878 * - chance for incoming ACK (processed by another cpu maybe)
879 * to migrate this flow (skb->ooo_okay will be eventually set)
881 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
884 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
885 struct tsq_tasklet *tsq;
888 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
891 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
892 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
896 /* queue this socket to tasklet queue */
897 local_irq_save(flags);
898 tsq = this_cpu_ptr(&tsq_tasklet);
899 empty = list_empty(&tsq->head);
900 list_add(&tp->tsq_node, &tsq->head);
902 tasklet_schedule(&tsq->tasklet);
903 local_irq_restore(flags);
910 /* Note: Called under hard irq.
911 * We can not call TCP stack right away.
913 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
915 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
916 struct sock *sk = (struct sock *)tp;
917 unsigned long nval, oval;
919 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
920 struct tsq_tasklet *tsq;
923 if (oval & TSQF_QUEUED)
926 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
927 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
931 if (!refcount_inc_not_zero(&sk->sk_wmem_alloc))
933 /* queue this socket to tasklet queue */
934 tsq = this_cpu_ptr(&tsq_tasklet);
935 empty = list_empty(&tsq->head);
936 list_add(&tp->tsq_node, &tsq->head);
938 tasklet_schedule(&tsq->tasklet);
941 return HRTIMER_NORESTART;
944 /* BBR congestion control needs pacing.
945 * Same remark for SO_MAX_PACING_RATE.
946 * sch_fq packet scheduler is efficiently handling pacing,
947 * but is not always installed/used.
948 * Return true if TCP stack should pace packets itself.
950 static bool tcp_needs_internal_pacing(const struct sock *sk)
952 return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
955 static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
960 if (!tcp_needs_internal_pacing(sk))
962 rate = sk->sk_pacing_rate;
963 if (!rate || rate == ~0U)
966 /* Should account for header sizes as sch_fq does,
967 * but lets make things simple.
969 len_ns = (u64)skb->len * NSEC_PER_SEC;
970 do_div(len_ns, rate);
971 hrtimer_start(&tcp_sk(sk)->pacing_timer,
972 ktime_add_ns(ktime_get(), len_ns),
973 HRTIMER_MODE_ABS_PINNED);
976 /* This routine actually transmits TCP packets queued in by
977 * tcp_do_sendmsg(). This is used by both the initial
978 * transmission and possible later retransmissions.
979 * All SKB's seen here are completely headerless. It is our
980 * job to build the TCP header, and pass the packet down to
981 * IP so it can do the same plus pass the packet off to the
984 * We are working here with either a clone of the original
985 * SKB, or a fresh unique copy made by the retransmit engine.
987 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
990 const struct inet_connection_sock *icsk = inet_csk(sk);
991 struct inet_sock *inet;
993 struct tcp_skb_cb *tcb;
994 struct tcp_out_options opts;
995 unsigned int tcp_options_size, tcp_header_size;
996 struct sk_buff *oskb = NULL;
997 struct tcp_md5sig_key *md5;
1001 BUG_ON(!skb || !tcp_skb_pcount(skb));
1005 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1008 if (unlikely(skb_cloned(skb)))
1009 skb = pskb_copy(skb, gfp_mask);
1011 skb = skb_clone(skb, gfp_mask);
1015 skb->skb_mstamp = tp->tcp_mstamp;
1018 tcb = TCP_SKB_CB(skb);
1019 memset(&opts, 0, sizeof(opts));
1021 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1022 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1024 tcp_options_size = tcp_established_options(sk, skb, &opts,
1026 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1028 /* if no packet is in qdisc/device queue, then allow XPS to select
1029 * another queue. We can be called from tcp_tsq_handler()
1030 * which holds one reference to sk_wmem_alloc.
1032 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1033 * One way to get this would be to set skb->truesize = 2 on them.
1035 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1037 /* If we had to use memory reserve to allocate this skb,
1038 * this might cause drops if packet is looped back :
1039 * Other socket might not have SOCK_MEMALLOC.
1040 * Packets not looped back do not care about pfmemalloc.
1042 skb->pfmemalloc = 0;
1044 skb_push(skb, tcp_header_size);
1045 skb_reset_transport_header(skb);
1049 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1050 skb_set_hash_from_sk(skb, sk);
1051 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1053 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1055 /* Build TCP header and checksum it. */
1056 th = (struct tcphdr *)skb->data;
1057 th->source = inet->inet_sport;
1058 th->dest = inet->inet_dport;
1059 th->seq = htonl(tcb->seq);
1060 th->ack_seq = htonl(tp->rcv_nxt);
1061 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1067 /* The urg_mode check is necessary during a below snd_una win probe */
1068 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1069 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1070 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1072 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1073 th->urg_ptr = htons(0xFFFF);
1078 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1079 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1080 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1081 th->window = htons(tcp_select_window(sk));
1082 tcp_ecn_send(sk, skb, th, tcp_header_size);
1084 /* RFC1323: The window in SYN & SYN/ACK segments
1087 th->window = htons(min(tp->rcv_wnd, 65535U));
1089 #ifdef CONFIG_TCP_MD5SIG
1090 /* Calculate the MD5 hash, as we have all we need now */
1092 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1093 tp->af_specific->calc_md5_hash(opts.hash_location,
1098 icsk->icsk_af_ops->send_check(sk, skb);
1100 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1101 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1103 if (skb->len != tcp_header_size) {
1104 tcp_event_data_sent(tp, sk);
1105 tp->data_segs_out += tcp_skb_pcount(skb);
1106 tcp_internal_pacing(sk, skb);
1109 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1110 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1111 tcp_skb_pcount(skb));
1113 tp->segs_out += tcp_skb_pcount(skb);
1114 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1115 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1116 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1118 /* Our usage of tstamp should remain private */
1121 /* Cleanup our debris for IP stacks */
1122 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1123 sizeof(struct inet6_skb_parm)));
1125 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1127 if (unlikely(err > 0)) {
1129 err = net_xmit_eval(err);
1132 oskb->skb_mstamp = tp->tcp_mstamp;
1133 tcp_rate_skb_sent(sk, oskb);
1138 /* This routine just queues the buffer for sending.
1140 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1141 * otherwise socket can stall.
1143 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1145 struct tcp_sock *tp = tcp_sk(sk);
1147 /* Advance write_seq and place onto the write_queue. */
1148 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1149 __skb_header_release(skb);
1150 tcp_add_write_queue_tail(sk, skb);
1151 sk->sk_wmem_queued += skb->truesize;
1152 sk_mem_charge(sk, skb->truesize);
1155 /* Initialize TSO segments for a packet. */
1156 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1158 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1159 /* Avoid the costly divide in the normal
1162 tcp_skb_pcount_set(skb, 1);
1163 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1165 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1166 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1170 /* When a modification to fackets out becomes necessary, we need to check
1171 * skb is counted to fackets_out or not.
1173 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1176 struct tcp_sock *tp = tcp_sk(sk);
1178 if (!tp->sacked_out || tcp_is_reno(tp))
1181 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1182 tp->fackets_out -= decr;
1185 /* Pcount in the middle of the write queue got changed, we need to do various
1186 * tweaks to fix counters
1188 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1190 struct tcp_sock *tp = tcp_sk(sk);
1192 tp->packets_out -= decr;
1194 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1195 tp->sacked_out -= decr;
1196 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1197 tp->retrans_out -= decr;
1198 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1199 tp->lost_out -= decr;
1201 /* Reno case is special. Sigh... */
1202 if (tcp_is_reno(tp) && decr > 0)
1203 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1205 tcp_adjust_fackets_out(sk, skb, decr);
1207 if (tp->lost_skb_hint &&
1208 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1209 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1210 tp->lost_cnt_hint -= decr;
1212 tcp_verify_left_out(tp);
1215 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1217 return TCP_SKB_CB(skb)->txstamp_ack ||
1218 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1221 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1223 struct skb_shared_info *shinfo = skb_shinfo(skb);
1225 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1226 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1227 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1228 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1230 shinfo->tx_flags &= ~tsflags;
1231 shinfo2->tx_flags |= tsflags;
1232 swap(shinfo->tskey, shinfo2->tskey);
1233 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1234 TCP_SKB_CB(skb)->txstamp_ack = 0;
1238 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1240 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1241 TCP_SKB_CB(skb)->eor = 0;
1244 /* Function to create two new TCP segments. Shrinks the given segment
1245 * to the specified size and appends a new segment with the rest of the
1246 * packet to the list. This won't be called frequently, I hope.
1247 * Remember, these are still headerless SKBs at this point.
1249 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1250 unsigned int mss_now, gfp_t gfp)
1252 struct tcp_sock *tp = tcp_sk(sk);
1253 struct sk_buff *buff;
1254 int nsize, old_factor;
1258 if (WARN_ON(len > skb->len))
1261 nsize = skb_headlen(skb) - len;
1265 if (skb_unclone(skb, gfp))
1268 /* Get a new skb... force flag on. */
1269 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1271 return -ENOMEM; /* We'll just try again later. */
1273 sk->sk_wmem_queued += buff->truesize;
1274 sk_mem_charge(sk, buff->truesize);
1275 nlen = skb->len - len - nsize;
1276 buff->truesize += nlen;
1277 skb->truesize -= nlen;
1279 /* Correct the sequence numbers. */
1280 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1281 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1282 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1284 /* PSH and FIN should only be set in the second packet. */
1285 flags = TCP_SKB_CB(skb)->tcp_flags;
1286 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1287 TCP_SKB_CB(buff)->tcp_flags = flags;
1288 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1289 tcp_skb_fragment_eor(skb, buff);
1291 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1292 /* Copy and checksum data tail into the new buffer. */
1293 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1294 skb_put(buff, nsize),
1299 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1301 skb->ip_summed = CHECKSUM_PARTIAL;
1302 skb_split(skb, buff, len);
1305 buff->ip_summed = skb->ip_summed;
1307 buff->tstamp = skb->tstamp;
1308 tcp_fragment_tstamp(skb, buff);
1310 old_factor = tcp_skb_pcount(skb);
1312 /* Fix up tso_factor for both original and new SKB. */
1313 tcp_set_skb_tso_segs(skb, mss_now);
1314 tcp_set_skb_tso_segs(buff, mss_now);
1316 /* Update delivered info for the new segment */
1317 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1319 /* If this packet has been sent out already, we must
1320 * adjust the various packet counters.
1322 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1323 int diff = old_factor - tcp_skb_pcount(skb) -
1324 tcp_skb_pcount(buff);
1327 tcp_adjust_pcount(sk, skb, diff);
1330 /* Link BUFF into the send queue. */
1331 __skb_header_release(buff);
1332 tcp_insert_write_queue_after(skb, buff, sk);
1337 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1338 * data is not copied, but immediately discarded.
1340 static int __pskb_trim_head(struct sk_buff *skb, int len)
1342 struct skb_shared_info *shinfo;
1345 eat = min_t(int, len, skb_headlen(skb));
1347 __skb_pull(skb, eat);
1354 shinfo = skb_shinfo(skb);
1355 for (i = 0; i < shinfo->nr_frags; i++) {
1356 int size = skb_frag_size(&shinfo->frags[i]);
1359 skb_frag_unref(skb, i);
1362 shinfo->frags[k] = shinfo->frags[i];
1364 shinfo->frags[k].page_offset += eat;
1365 skb_frag_size_sub(&shinfo->frags[k], eat);
1371 shinfo->nr_frags = k;
1373 skb->data_len -= len;
1374 skb->len = skb->data_len;
1378 /* Remove acked data from a packet in the transmit queue. */
1379 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1383 if (skb_unclone(skb, GFP_ATOMIC))
1386 delta_truesize = __pskb_trim_head(skb, len);
1388 TCP_SKB_CB(skb)->seq += len;
1389 skb->ip_summed = CHECKSUM_PARTIAL;
1391 if (delta_truesize) {
1392 skb->truesize -= delta_truesize;
1393 sk->sk_wmem_queued -= delta_truesize;
1394 sk_mem_uncharge(sk, delta_truesize);
1395 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1398 /* Any change of skb->len requires recalculation of tso factor. */
1399 if (tcp_skb_pcount(skb) > 1)
1400 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1405 /* Calculate MSS not accounting any TCP options. */
1406 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1408 const struct tcp_sock *tp = tcp_sk(sk);
1409 const struct inet_connection_sock *icsk = inet_csk(sk);
1412 /* Calculate base mss without TCP options:
1413 It is MMS_S - sizeof(tcphdr) of rfc1122
1415 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1417 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1418 if (icsk->icsk_af_ops->net_frag_header_len) {
1419 const struct dst_entry *dst = __sk_dst_get(sk);
1421 if (dst && dst_allfrag(dst))
1422 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1425 /* Clamp it (mss_clamp does not include tcp options) */
1426 if (mss_now > tp->rx_opt.mss_clamp)
1427 mss_now = tp->rx_opt.mss_clamp;
1429 /* Now subtract optional transport overhead */
1430 mss_now -= icsk->icsk_ext_hdr_len;
1432 /* Then reserve room for full set of TCP options and 8 bytes of data */
1438 /* Calculate MSS. Not accounting for SACKs here. */
1439 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1441 /* Subtract TCP options size, not including SACKs */
1442 return __tcp_mtu_to_mss(sk, pmtu) -
1443 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1446 /* Inverse of above */
1447 int tcp_mss_to_mtu(struct sock *sk, int mss)
1449 const struct tcp_sock *tp = tcp_sk(sk);
1450 const struct inet_connection_sock *icsk = inet_csk(sk);
1454 tp->tcp_header_len +
1455 icsk->icsk_ext_hdr_len +
1456 icsk->icsk_af_ops->net_header_len;
1458 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1459 if (icsk->icsk_af_ops->net_frag_header_len) {
1460 const struct dst_entry *dst = __sk_dst_get(sk);
1462 if (dst && dst_allfrag(dst))
1463 mtu += icsk->icsk_af_ops->net_frag_header_len;
1467 EXPORT_SYMBOL(tcp_mss_to_mtu);
1469 /* MTU probing init per socket */
1470 void tcp_mtup_init(struct sock *sk)
1472 struct tcp_sock *tp = tcp_sk(sk);
1473 struct inet_connection_sock *icsk = inet_csk(sk);
1474 struct net *net = sock_net(sk);
1476 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1477 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1478 icsk->icsk_af_ops->net_header_len;
1479 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1480 icsk->icsk_mtup.probe_size = 0;
1481 if (icsk->icsk_mtup.enabled)
1482 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1484 EXPORT_SYMBOL(tcp_mtup_init);
1486 /* This function synchronize snd mss to current pmtu/exthdr set.
1488 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1489 for TCP options, but includes only bare TCP header.
1491 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1492 It is minimum of user_mss and mss received with SYN.
1493 It also does not include TCP options.
1495 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1497 tp->mss_cache is current effective sending mss, including
1498 all tcp options except for SACKs. It is evaluated,
1499 taking into account current pmtu, but never exceeds
1500 tp->rx_opt.mss_clamp.
1502 NOTE1. rfc1122 clearly states that advertised MSS
1503 DOES NOT include either tcp or ip options.
1505 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1506 are READ ONLY outside this function. --ANK (980731)
1508 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1510 struct tcp_sock *tp = tcp_sk(sk);
1511 struct inet_connection_sock *icsk = inet_csk(sk);
1514 if (icsk->icsk_mtup.search_high > pmtu)
1515 icsk->icsk_mtup.search_high = pmtu;
1517 mss_now = tcp_mtu_to_mss(sk, pmtu);
1518 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1520 /* And store cached results */
1521 icsk->icsk_pmtu_cookie = pmtu;
1522 if (icsk->icsk_mtup.enabled)
1523 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1524 tp->mss_cache = mss_now;
1528 EXPORT_SYMBOL(tcp_sync_mss);
1530 /* Compute the current effective MSS, taking SACKs and IP options,
1531 * and even PMTU discovery events into account.
1533 unsigned int tcp_current_mss(struct sock *sk)
1535 const struct tcp_sock *tp = tcp_sk(sk);
1536 const struct dst_entry *dst = __sk_dst_get(sk);
1538 unsigned int header_len;
1539 struct tcp_out_options opts;
1540 struct tcp_md5sig_key *md5;
1542 mss_now = tp->mss_cache;
1545 u32 mtu = dst_mtu(dst);
1546 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1547 mss_now = tcp_sync_mss(sk, mtu);
1550 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1551 sizeof(struct tcphdr);
1552 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1553 * some common options. If this is an odd packet (because we have SACK
1554 * blocks etc) then our calculated header_len will be different, and
1555 * we have to adjust mss_now correspondingly */
1556 if (header_len != tp->tcp_header_len) {
1557 int delta = (int) header_len - tp->tcp_header_len;
1564 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1565 * As additional protections, we do not touch cwnd in retransmission phases,
1566 * and if application hit its sndbuf limit recently.
1568 static void tcp_cwnd_application_limited(struct sock *sk)
1570 struct tcp_sock *tp = tcp_sk(sk);
1572 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1573 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1574 /* Limited by application or receiver window. */
1575 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1576 u32 win_used = max(tp->snd_cwnd_used, init_win);
1577 if (win_used < tp->snd_cwnd) {
1578 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1579 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1581 tp->snd_cwnd_used = 0;
1583 tp->snd_cwnd_stamp = tcp_jiffies32;
1586 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1588 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1589 struct tcp_sock *tp = tcp_sk(sk);
1591 /* Track the maximum number of outstanding packets in each
1592 * window, and remember whether we were cwnd-limited then.
1594 if (!before(tp->snd_una, tp->max_packets_seq) ||
1595 tp->packets_out > tp->max_packets_out) {
1596 tp->max_packets_out = tp->packets_out;
1597 tp->max_packets_seq = tp->snd_nxt;
1598 tp->is_cwnd_limited = is_cwnd_limited;
1601 if (tcp_is_cwnd_limited(sk)) {
1602 /* Network is feed fully. */
1603 tp->snd_cwnd_used = 0;
1604 tp->snd_cwnd_stamp = tcp_jiffies32;
1606 /* Network starves. */
1607 if (tp->packets_out > tp->snd_cwnd_used)
1608 tp->snd_cwnd_used = tp->packets_out;
1610 if (sysctl_tcp_slow_start_after_idle &&
1611 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1612 !ca_ops->cong_control)
1613 tcp_cwnd_application_limited(sk);
1615 /* The following conditions together indicate the starvation
1616 * is caused by insufficient sender buffer:
1617 * 1) just sent some data (see tcp_write_xmit)
1618 * 2) not cwnd limited (this else condition)
1619 * 3) no more data to send (null tcp_send_head )
1620 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1622 if (!tcp_send_head(sk) && sk->sk_socket &&
1623 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1624 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1625 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1629 /* Minshall's variant of the Nagle send check. */
1630 static bool tcp_minshall_check(const struct tcp_sock *tp)
1632 return after(tp->snd_sml, tp->snd_una) &&
1633 !after(tp->snd_sml, tp->snd_nxt);
1636 /* Update snd_sml if this skb is under mss
1637 * Note that a TSO packet might end with a sub-mss segment
1638 * The test is really :
1639 * if ((skb->len % mss) != 0)
1640 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1641 * But we can avoid doing the divide again given we already have
1642 * skb_pcount = skb->len / mss_now
1644 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1645 const struct sk_buff *skb)
1647 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1648 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1651 /* Return false, if packet can be sent now without violation Nagle's rules:
1652 * 1. It is full sized. (provided by caller in %partial bool)
1653 * 2. Or it contains FIN. (already checked by caller)
1654 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1655 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1656 * With Minshall's modification: all sent small packets are ACKed.
1658 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1662 ((nonagle & TCP_NAGLE_CORK) ||
1663 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1666 /* Return how many segs we'd like on a TSO packet,
1667 * to send one TSO packet per ms
1669 u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1674 bytes = min(sk->sk_pacing_rate >> 10,
1675 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1677 /* Goal is to send at least one packet per ms,
1678 * not one big TSO packet every 100 ms.
1679 * This preserves ACK clocking and is consistent
1680 * with tcp_tso_should_defer() heuristic.
1682 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1684 return min_t(u32, segs, sk->sk_gso_max_segs);
1686 EXPORT_SYMBOL(tcp_tso_autosize);
1688 /* Return the number of segments we want in the skb we are transmitting.
1689 * See if congestion control module wants to decide; otherwise, autosize.
1691 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1693 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1694 u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0;
1697 tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs);
1700 /* Returns the portion of skb which can be sent right away */
1701 static unsigned int tcp_mss_split_point(const struct sock *sk,
1702 const struct sk_buff *skb,
1703 unsigned int mss_now,
1704 unsigned int max_segs,
1707 const struct tcp_sock *tp = tcp_sk(sk);
1708 u32 partial, needed, window, max_len;
1710 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1711 max_len = mss_now * max_segs;
1713 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1716 needed = min(skb->len, window);
1718 if (max_len <= needed)
1721 partial = needed % mss_now;
1722 /* If last segment is not a full MSS, check if Nagle rules allow us
1723 * to include this last segment in this skb.
1724 * Otherwise, we'll split the skb at last MSS boundary
1726 if (tcp_nagle_check(partial != 0, tp, nonagle))
1727 return needed - partial;
1732 /* Can at least one segment of SKB be sent right now, according to the
1733 * congestion window rules? If so, return how many segments are allowed.
1735 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1736 const struct sk_buff *skb)
1738 u32 in_flight, cwnd, halfcwnd;
1740 /* Don't be strict about the congestion window for the final FIN. */
1741 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1742 tcp_skb_pcount(skb) == 1)
1745 in_flight = tcp_packets_in_flight(tp);
1746 cwnd = tp->snd_cwnd;
1747 if (in_flight >= cwnd)
1750 /* For better scheduling, ensure we have at least
1751 * 2 GSO packets in flight.
1753 halfcwnd = max(cwnd >> 1, 1U);
1754 return min(halfcwnd, cwnd - in_flight);
1757 /* Initialize TSO state of a skb.
1758 * This must be invoked the first time we consider transmitting
1759 * SKB onto the wire.
1761 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1763 int tso_segs = tcp_skb_pcount(skb);
1765 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1766 tcp_set_skb_tso_segs(skb, mss_now);
1767 tso_segs = tcp_skb_pcount(skb);
1773 /* Return true if the Nagle test allows this packet to be
1776 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1777 unsigned int cur_mss, int nonagle)
1779 /* Nagle rule does not apply to frames, which sit in the middle of the
1780 * write_queue (they have no chances to get new data).
1782 * This is implemented in the callers, where they modify the 'nonagle'
1783 * argument based upon the location of SKB in the send queue.
1785 if (nonagle & TCP_NAGLE_PUSH)
1788 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1789 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1792 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1798 /* Does at least the first segment of SKB fit into the send window? */
1799 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1800 const struct sk_buff *skb,
1801 unsigned int cur_mss)
1803 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1805 if (skb->len > cur_mss)
1806 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1808 return !after(end_seq, tcp_wnd_end(tp));
1811 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1812 * which is put after SKB on the list. It is very much like
1813 * tcp_fragment() except that it may make several kinds of assumptions
1814 * in order to speed up the splitting operation. In particular, we
1815 * know that all the data is in scatter-gather pages, and that the
1816 * packet has never been sent out before (and thus is not cloned).
1818 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1819 unsigned int mss_now, gfp_t gfp)
1821 struct sk_buff *buff;
1822 int nlen = skb->len - len;
1825 /* All of a TSO frame must be composed of paged data. */
1826 if (skb->len != skb->data_len)
1827 return tcp_fragment(sk, skb, len, mss_now, gfp);
1829 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1830 if (unlikely(!buff))
1833 sk->sk_wmem_queued += buff->truesize;
1834 sk_mem_charge(sk, buff->truesize);
1835 buff->truesize += nlen;
1836 skb->truesize -= nlen;
1838 /* Correct the sequence numbers. */
1839 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1840 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1841 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1843 /* PSH and FIN should only be set in the second packet. */
1844 flags = TCP_SKB_CB(skb)->tcp_flags;
1845 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1846 TCP_SKB_CB(buff)->tcp_flags = flags;
1848 /* This packet was never sent out yet, so no SACK bits. */
1849 TCP_SKB_CB(buff)->sacked = 0;
1851 tcp_skb_fragment_eor(skb, buff);
1853 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1854 skb_split(skb, buff, len);
1855 tcp_fragment_tstamp(skb, buff);
1857 /* Fix up tso_factor for both original and new SKB. */
1858 tcp_set_skb_tso_segs(skb, mss_now);
1859 tcp_set_skb_tso_segs(buff, mss_now);
1861 /* Link BUFF into the send queue. */
1862 __skb_header_release(buff);
1863 tcp_insert_write_queue_after(skb, buff, sk);
1868 /* Try to defer sending, if possible, in order to minimize the amount
1869 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1871 * This algorithm is from John Heffner.
1873 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1874 bool *is_cwnd_limited, u32 max_segs)
1876 const struct inet_connection_sock *icsk = inet_csk(sk);
1877 u32 age, send_win, cong_win, limit, in_flight;
1878 struct tcp_sock *tp = tcp_sk(sk);
1879 struct sk_buff *head;
1882 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1885 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1888 /* Avoid bursty behavior by allowing defer
1889 * only if the last write was recent.
1891 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
1894 in_flight = tcp_packets_in_flight(tp);
1896 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1898 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1900 /* From in_flight test above, we know that cwnd > in_flight. */
1901 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1903 limit = min(send_win, cong_win);
1905 /* If a full-sized TSO skb can be sent, do it. */
1906 if (limit >= max_segs * tp->mss_cache)
1909 /* Middle in queue won't get any more data, full sendable already? */
1910 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1913 win_divisor = READ_ONCE(sysctl_tcp_tso_win_divisor);
1915 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1917 /* If at least some fraction of a window is available,
1920 chunk /= win_divisor;
1924 /* Different approach, try not to defer past a single
1925 * ACK. Receiver should ACK every other full sized
1926 * frame, so if we have space for more than 3 frames
1929 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1933 head = tcp_write_queue_head(sk);
1935 age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
1936 /* If next ACK is likely to come too late (half srtt), do not defer */
1937 if (age < (tp->srtt_us >> 4))
1940 /* Ok, it looks like it is advisable to defer. */
1942 if (cong_win < send_win && cong_win <= skb->len)
1943 *is_cwnd_limited = true;
1951 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1953 struct inet_connection_sock *icsk = inet_csk(sk);
1954 struct tcp_sock *tp = tcp_sk(sk);
1955 struct net *net = sock_net(sk);
1959 interval = net->ipv4.sysctl_tcp_probe_interval;
1960 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
1961 if (unlikely(delta >= interval * HZ)) {
1962 int mss = tcp_current_mss(sk);
1964 /* Update current search range */
1965 icsk->icsk_mtup.probe_size = 0;
1966 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1967 sizeof(struct tcphdr) +
1968 icsk->icsk_af_ops->net_header_len;
1969 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1971 /* Update probe time stamp */
1972 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1976 /* Create a new MTU probe if we are ready.
1977 * MTU probe is regularly attempting to increase the path MTU by
1978 * deliberately sending larger packets. This discovers routing
1979 * changes resulting in larger path MTUs.
1981 * Returns 0 if we should wait to probe (no cwnd available),
1982 * 1 if a probe was sent,
1985 static int tcp_mtu_probe(struct sock *sk)
1987 struct inet_connection_sock *icsk = inet_csk(sk);
1988 struct tcp_sock *tp = tcp_sk(sk);
1989 struct sk_buff *skb, *nskb, *next;
1990 struct net *net = sock_net(sk);
1997 /* Not currently probing/verifying,
1999 * have enough cwnd, and
2000 * not SACKing (the variable headers throw things off)
2002 if (likely(!icsk->icsk_mtup.enabled ||
2003 icsk->icsk_mtup.probe_size ||
2004 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2005 tp->snd_cwnd < 11 ||
2006 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2009 /* Use binary search for probe_size between tcp_mss_base,
2010 * and current mss_clamp. if (search_high - search_low)
2011 * smaller than a threshold, backoff from probing.
2013 mss_now = tcp_current_mss(sk);
2014 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2015 icsk->icsk_mtup.search_low) >> 1);
2016 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2017 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2018 /* When misfortune happens, we are reprobing actively,
2019 * and then reprobe timer has expired. We stick with current
2020 * probing process by not resetting search range to its orignal.
2022 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2023 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2024 /* Check whether enough time has elaplased for
2025 * another round of probing.
2027 tcp_mtu_check_reprobe(sk);
2031 /* Have enough data in the send queue to probe? */
2032 if (tp->write_seq - tp->snd_nxt < size_needed)
2035 if (tp->snd_wnd < size_needed)
2037 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2040 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2041 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2042 if (!tcp_packets_in_flight(tp))
2048 /* We're allowed to probe. Build it now. */
2049 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2052 sk->sk_wmem_queued += nskb->truesize;
2053 sk_mem_charge(sk, nskb->truesize);
2055 skb = tcp_send_head(sk);
2057 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2058 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2059 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2060 TCP_SKB_CB(nskb)->sacked = 0;
2062 nskb->ip_summed = skb->ip_summed;
2064 tcp_insert_write_queue_before(nskb, skb, sk);
2065 tcp_highest_sack_replace(sk, skb, nskb);
2068 tcp_for_write_queue_from_safe(skb, next, sk) {
2069 copy = min_t(int, skb->len, probe_size - len);
2070 if (nskb->ip_summed) {
2071 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2073 __wsum csum = skb_copy_and_csum_bits(skb, 0,
2074 skb_put(nskb, copy),
2076 nskb->csum = csum_block_add(nskb->csum, csum, len);
2079 if (skb->len <= copy) {
2080 /* We've eaten all the data from this skb.
2082 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2083 tcp_unlink_write_queue(skb, sk);
2084 sk_wmem_free_skb(sk, skb);
2086 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2087 ~(TCPHDR_FIN|TCPHDR_PSH);
2088 if (!skb_shinfo(skb)->nr_frags) {
2089 skb_pull(skb, copy);
2090 if (skb->ip_summed != CHECKSUM_PARTIAL)
2091 skb->csum = csum_partial(skb->data,
2094 __pskb_trim_head(skb, copy);
2095 tcp_set_skb_tso_segs(skb, mss_now);
2097 TCP_SKB_CB(skb)->seq += copy;
2102 if (len >= probe_size)
2105 tcp_init_tso_segs(nskb, nskb->len);
2107 /* We're ready to send. If this fails, the probe will
2108 * be resegmented into mss-sized pieces by tcp_write_xmit().
2110 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2111 /* Decrement cwnd here because we are sending
2112 * effectively two packets. */
2114 tcp_event_new_data_sent(sk, nskb);
2116 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2117 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2118 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2126 static bool tcp_pacing_check(const struct sock *sk)
2128 return tcp_needs_internal_pacing(sk) &&
2129 hrtimer_active(&tcp_sk(sk)->pacing_timer);
2132 /* TCP Small Queues :
2133 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2134 * (These limits are doubled for retransmits)
2136 * - better RTT estimation and ACK scheduling
2139 * Alas, some drivers / subsystems require a fair amount
2140 * of queued bytes to ensure line rate.
2141 * One example is wifi aggregation (802.11 AMPDU)
2143 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2144 unsigned int factor)
2148 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2149 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2152 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2153 /* Always send the 1st or 2nd skb in write queue.
2154 * No need to wait for TX completion to call us back,
2155 * after softirq/tasklet schedule.
2156 * This helps when TX completions are delayed too much.
2158 if (skb == sk->sk_write_queue.next ||
2159 skb->prev == sk->sk_write_queue.next)
2162 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2163 /* It is possible TX completion already happened
2164 * before we set TSQ_THROTTLED, so we must
2165 * test again the condition.
2167 smp_mb__after_atomic();
2168 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2174 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2176 const u32 now = tcp_jiffies32;
2177 enum tcp_chrono old = tp->chrono_type;
2179 if (old > TCP_CHRONO_UNSPEC)
2180 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2181 tp->chrono_start = now;
2182 tp->chrono_type = new;
2185 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2187 struct tcp_sock *tp = tcp_sk(sk);
2189 /* If there are multiple conditions worthy of tracking in a
2190 * chronograph then the highest priority enum takes precedence
2191 * over the other conditions. So that if something "more interesting"
2192 * starts happening, stop the previous chrono and start a new one.
2194 if (type > tp->chrono_type)
2195 tcp_chrono_set(tp, type);
2198 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2200 struct tcp_sock *tp = tcp_sk(sk);
2203 /* There are multiple conditions worthy of tracking in a
2204 * chronograph, so that the highest priority enum takes
2205 * precedence over the other conditions (see tcp_chrono_start).
2206 * If a condition stops, we only stop chrono tracking if
2207 * it's the "most interesting" or current chrono we are
2208 * tracking and starts busy chrono if we have pending data.
2210 if (tcp_write_queue_empty(sk))
2211 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2212 else if (type == tp->chrono_type)
2213 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2216 /* This routine writes packets to the network. It advances the
2217 * send_head. This happens as incoming acks open up the remote
2220 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2221 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2222 * account rare use of URG, this is not a big flaw.
2224 * Send at most one packet when push_one > 0. Temporarily ignore
2225 * cwnd limit to force at most one packet out when push_one == 2.
2227 * Returns true, if no segments are in flight and we have queued segments,
2228 * but cannot send anything now because of SWS or another problem.
2230 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2231 int push_one, gfp_t gfp)
2233 struct tcp_sock *tp = tcp_sk(sk);
2234 struct sk_buff *skb;
2235 unsigned int tso_segs, sent_pkts;
2238 bool is_cwnd_limited = false, is_rwnd_limited = false;
2243 tcp_mstamp_refresh(tp);
2245 /* Do MTU probing. */
2246 result = tcp_mtu_probe(sk);
2249 } else if (result > 0) {
2254 max_segs = tcp_tso_segs(sk, mss_now);
2255 while ((skb = tcp_send_head(sk))) {
2258 if (tcp_pacing_check(sk))
2261 tso_segs = tcp_init_tso_segs(skb, mss_now);
2264 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2265 /* "skb_mstamp" is used as a start point for the retransmit timer */
2266 skb->skb_mstamp = tp->tcp_mstamp;
2267 goto repair; /* Skip network transmission */
2270 cwnd_quota = tcp_cwnd_test(tp, skb);
2273 /* Force out a loss probe pkt. */
2279 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2280 is_rwnd_limited = true;
2284 if (tso_segs == 1) {
2285 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2286 (tcp_skb_is_last(sk, skb) ?
2287 nonagle : TCP_NAGLE_PUSH))))
2291 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2297 if (tso_segs > 1 && !tcp_urg_mode(tp))
2298 limit = tcp_mss_split_point(sk, skb, mss_now,
2304 if (skb->len > limit &&
2305 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2308 if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
2309 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
2310 if (tcp_small_queue_check(sk, skb, 0))
2313 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2317 /* Advance the send_head. This one is sent out.
2318 * This call will increment packets_out.
2320 tcp_event_new_data_sent(sk, skb);
2322 tcp_minshall_update(tp, mss_now, skb);
2323 sent_pkts += tcp_skb_pcount(skb);
2329 if (is_rwnd_limited)
2330 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2332 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2334 if (likely(sent_pkts)) {
2335 if (tcp_in_cwnd_reduction(sk))
2336 tp->prr_out += sent_pkts;
2338 /* Send one loss probe per tail loss episode. */
2340 tcp_schedule_loss_probe(sk);
2341 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2342 tcp_cwnd_validate(sk, is_cwnd_limited);
2345 return !tp->packets_out && tcp_send_head(sk);
2348 bool tcp_schedule_loss_probe(struct sock *sk)
2350 struct inet_connection_sock *icsk = inet_csk(sk);
2351 struct tcp_sock *tp = tcp_sk(sk);
2352 u32 timeout, rto_delta_us;
2354 /* Don't do any loss probe on a Fast Open connection before 3WHS
2357 if (tp->fastopen_rsk)
2360 /* Schedule a loss probe in 2*RTT for SACK capable connections
2361 * in Open state, that are either limited by cwnd or application.
2363 if ((sysctl_tcp_early_retrans != 3 && sysctl_tcp_early_retrans != 4) ||
2364 !tp->packets_out || !tcp_is_sack(tp) ||
2365 icsk->icsk_ca_state != TCP_CA_Open)
2368 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2372 /* Probe timeout is 2*rtt. Add minimum RTO to account
2373 * for delayed ack when there's one outstanding packet. If no RTT
2374 * sample is available then probe after TCP_TIMEOUT_INIT.
2377 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2378 if (tp->packets_out == 1)
2379 timeout += TCP_RTO_MIN;
2381 timeout += TCP_TIMEOUT_MIN;
2383 timeout = TCP_TIMEOUT_INIT;
2386 /* If the RTO formula yields an earlier time, then use that time. */
2387 rto_delta_us = tcp_rto_delta_us(sk); /* How far in future is RTO? */
2388 if (rto_delta_us > 0)
2389 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2391 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2396 /* Thanks to skb fast clones, we can detect if a prior transmit of
2397 * a packet is still in a qdisc or driver queue.
2398 * In this case, there is very little point doing a retransmit !
2400 static bool skb_still_in_host_queue(const struct sock *sk,
2401 const struct sk_buff *skb)
2403 if (unlikely(skb_fclone_busy(sk, skb))) {
2404 NET_INC_STATS(sock_net(sk),
2405 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2411 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2412 * retransmit the last segment.
2414 void tcp_send_loss_probe(struct sock *sk)
2416 struct tcp_sock *tp = tcp_sk(sk);
2417 struct sk_buff *skb;
2419 int mss = tcp_current_mss(sk);
2421 skb = tcp_send_head(sk);
2423 if (tcp_snd_wnd_test(tp, skb, mss)) {
2424 pcount = tp->packets_out;
2425 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2426 if (tp->packets_out > pcount)
2430 skb = tcp_write_queue_prev(sk, skb);
2432 skb = tcp_write_queue_tail(sk);
2435 /* At most one outstanding TLP retransmission. */
2436 if (tp->tlp_high_seq)
2439 /* Retransmit last segment. */
2443 if (skb_still_in_host_queue(sk, skb))
2446 pcount = tcp_skb_pcount(skb);
2447 if (WARN_ON(!pcount))
2450 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2451 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2454 skb = tcp_write_queue_next(sk, skb);
2457 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2460 if (__tcp_retransmit_skb(sk, skb, 1))
2463 /* Record snd_nxt for loss detection. */
2464 tp->tlp_high_seq = tp->snd_nxt;
2467 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2468 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2469 inet_csk(sk)->icsk_pending = 0;
2474 /* Push out any pending frames which were held back due to
2475 * TCP_CORK or attempt at coalescing tiny packets.
2476 * The socket must be locked by the caller.
2478 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2481 /* If we are closed, the bytes will have to remain here.
2482 * In time closedown will finish, we empty the write queue and
2483 * all will be happy.
2485 if (unlikely(sk->sk_state == TCP_CLOSE))
2488 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2489 sk_gfp_mask(sk, GFP_ATOMIC)))
2490 tcp_check_probe_timer(sk);
2493 /* Send _single_ skb sitting at the send head. This function requires
2494 * true push pending frames to setup probe timer etc.
2496 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2498 struct sk_buff *skb = tcp_send_head(sk);
2500 BUG_ON(!skb || skb->len < mss_now);
2502 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2505 /* This function returns the amount that we can raise the
2506 * usable window based on the following constraints
2508 * 1. The window can never be shrunk once it is offered (RFC 793)
2509 * 2. We limit memory per socket
2512 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2513 * RECV.NEXT + RCV.WIN fixed until:
2514 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2516 * i.e. don't raise the right edge of the window until you can raise
2517 * it at least MSS bytes.
2519 * Unfortunately, the recommended algorithm breaks header prediction,
2520 * since header prediction assumes th->window stays fixed.
2522 * Strictly speaking, keeping th->window fixed violates the receiver
2523 * side SWS prevention criteria. The problem is that under this rule
2524 * a stream of single byte packets will cause the right side of the
2525 * window to always advance by a single byte.
2527 * Of course, if the sender implements sender side SWS prevention
2528 * then this will not be a problem.
2530 * BSD seems to make the following compromise:
2532 * If the free space is less than the 1/4 of the maximum
2533 * space available and the free space is less than 1/2 mss,
2534 * then set the window to 0.
2535 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2536 * Otherwise, just prevent the window from shrinking
2537 * and from being larger than the largest representable value.
2539 * This prevents incremental opening of the window in the regime
2540 * where TCP is limited by the speed of the reader side taking
2541 * data out of the TCP receive queue. It does nothing about
2542 * those cases where the window is constrained on the sender side
2543 * because the pipeline is full.
2545 * BSD also seems to "accidentally" limit itself to windows that are a
2546 * multiple of MSS, at least until the free space gets quite small.
2547 * This would appear to be a side effect of the mbuf implementation.
2548 * Combining these two algorithms results in the observed behavior
2549 * of having a fixed window size at almost all times.
2551 * Below we obtain similar behavior by forcing the offered window to
2552 * a multiple of the mss when it is feasible to do so.
2554 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2555 * Regular options like TIMESTAMP are taken into account.
2557 u32 __tcp_select_window(struct sock *sk)
2559 struct inet_connection_sock *icsk = inet_csk(sk);
2560 struct tcp_sock *tp = tcp_sk(sk);
2561 /* MSS for the peer's data. Previous versions used mss_clamp
2562 * here. I don't know if the value based on our guesses
2563 * of peer's MSS is better for the performance. It's more correct
2564 * but may be worse for the performance because of rcv_mss
2565 * fluctuations. --SAW 1998/11/1
2567 int mss = icsk->icsk_ack.rcv_mss;
2568 int free_space = tcp_space(sk);
2569 int allowed_space = tcp_full_space(sk);
2570 int full_space = min_t(int, tp->window_clamp, allowed_space);
2573 if (unlikely(mss > full_space)) {
2578 if (free_space < (full_space >> 1)) {
2579 icsk->icsk_ack.quick = 0;
2581 if (tcp_under_memory_pressure(sk))
2582 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2585 /* free_space might become our new window, make sure we don't
2586 * increase it due to wscale.
2588 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2590 /* if free space is less than mss estimate, or is below 1/16th
2591 * of the maximum allowed, try to move to zero-window, else
2592 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2593 * new incoming data is dropped due to memory limits.
2594 * With large window, mss test triggers way too late in order
2595 * to announce zero window in time before rmem limit kicks in.
2597 if (free_space < (allowed_space >> 4) || free_space < mss)
2601 if (free_space > tp->rcv_ssthresh)
2602 free_space = tp->rcv_ssthresh;
2604 /* Don't do rounding if we are using window scaling, since the
2605 * scaled window will not line up with the MSS boundary anyway.
2607 if (tp->rx_opt.rcv_wscale) {
2608 window = free_space;
2610 /* Advertise enough space so that it won't get scaled away.
2611 * Import case: prevent zero window announcement if
2612 * 1<<rcv_wscale > mss.
2614 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2616 window = tp->rcv_wnd;
2617 /* Get the largest window that is a nice multiple of mss.
2618 * Window clamp already applied above.
2619 * If our current window offering is within 1 mss of the
2620 * free space we just keep it. This prevents the divide
2621 * and multiply from happening most of the time.
2622 * We also don't do any window rounding when the free space
2625 if (window <= free_space - mss || window > free_space)
2626 window = rounddown(free_space, mss);
2627 else if (mss == full_space &&
2628 free_space > window + (full_space >> 1))
2629 window = free_space;
2635 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2636 const struct sk_buff *next_skb)
2638 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2639 const struct skb_shared_info *next_shinfo =
2640 skb_shinfo(next_skb);
2641 struct skb_shared_info *shinfo = skb_shinfo(skb);
2643 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2644 shinfo->tskey = next_shinfo->tskey;
2645 TCP_SKB_CB(skb)->txstamp_ack |=
2646 TCP_SKB_CB(next_skb)->txstamp_ack;
2650 /* Collapses two adjacent SKB's during retransmission. */
2651 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2653 struct tcp_sock *tp = tcp_sk(sk);
2654 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2655 int skb_size, next_skb_size;
2657 skb_size = skb->len;
2658 next_skb_size = next_skb->len;
2660 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2662 if (next_skb_size) {
2663 if (next_skb_size <= skb_availroom(skb))
2664 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2666 else if (!skb_shift(skb, next_skb, next_skb_size))
2669 tcp_highest_sack_replace(sk, next_skb, skb);
2671 tcp_unlink_write_queue(next_skb, sk);
2673 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2674 skb->ip_summed = CHECKSUM_PARTIAL;
2676 if (skb->ip_summed != CHECKSUM_PARTIAL)
2677 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2679 /* Update sequence range on original skb. */
2680 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2682 /* Merge over control information. This moves PSH/FIN etc. over */
2683 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2685 /* All done, get rid of second SKB and account for it so
2686 * packet counting does not break.
2688 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2689 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2691 /* changed transmit queue under us so clear hints */
2692 tcp_clear_retrans_hints_partial(tp);
2693 if (next_skb == tp->retransmit_skb_hint)
2694 tp->retransmit_skb_hint = skb;
2696 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2698 tcp_skb_collapse_tstamp(skb, next_skb);
2700 sk_wmem_free_skb(sk, next_skb);
2704 /* Check if coalescing SKBs is legal. */
2705 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2707 if (tcp_skb_pcount(skb) > 1)
2709 if (skb_cloned(skb))
2711 if (skb == tcp_send_head(sk))
2713 /* Some heuristics for collapsing over SACK'd could be invented */
2714 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2720 /* Collapse packets in the retransmit queue to make to create
2721 * less packets on the wire. This is only done on retransmission.
2723 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2726 struct tcp_sock *tp = tcp_sk(sk);
2727 struct sk_buff *skb = to, *tmp;
2730 if (!sysctl_tcp_retrans_collapse)
2732 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2735 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2736 if (!tcp_can_collapse(sk, skb))
2739 if (!tcp_skb_can_collapse_to(to))
2752 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2755 if (!tcp_collapse_retrans(sk, to))
2760 /* This retransmits one SKB. Policy decisions and retransmit queue
2761 * state updates are done by the caller. Returns non-zero if an
2762 * error occurred which prevented the send.
2764 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2766 struct inet_connection_sock *icsk = inet_csk(sk);
2767 struct tcp_sock *tp = tcp_sk(sk);
2768 unsigned int cur_mss;
2772 /* Inconclusive MTU probe */
2773 if (icsk->icsk_mtup.probe_size)
2774 icsk->icsk_mtup.probe_size = 0;
2776 /* Do not sent more than we queued. 1/4 is reserved for possible
2777 * copying overhead: fragmentation, tunneling, mangling etc.
2779 if (refcount_read(&sk->sk_wmem_alloc) >
2780 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2784 if (skb_still_in_host_queue(sk, skb))
2787 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2788 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2790 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2794 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2795 return -EHOSTUNREACH; /* Routing failure or similar. */
2797 cur_mss = tcp_current_mss(sk);
2799 /* If receiver has shrunk his window, and skb is out of
2800 * new window, do not retransmit it. The exception is the
2801 * case, when window is shrunk to zero. In this case
2802 * our retransmit serves as a zero window probe.
2804 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2805 TCP_SKB_CB(skb)->seq != tp->snd_una)
2808 len = cur_mss * segs;
2809 if (skb->len > len) {
2810 if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC))
2811 return -ENOMEM; /* We'll try again later. */
2813 if (skb_unclone(skb, GFP_ATOMIC))
2816 diff = tcp_skb_pcount(skb);
2817 tcp_set_skb_tso_segs(skb, cur_mss);
2818 diff -= tcp_skb_pcount(skb);
2820 tcp_adjust_pcount(sk, skb, diff);
2821 if (skb->len < cur_mss)
2822 tcp_retrans_try_collapse(sk, skb, cur_mss);
2825 /* RFC3168, section 6.1.1.1. ECN fallback */
2826 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2827 tcp_ecn_clear_syn(sk, skb);
2829 /* Update global and local TCP statistics. */
2830 segs = tcp_skb_pcount(skb);
2831 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2832 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2833 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2834 tp->total_retrans += segs;
2836 /* make sure skb->data is aligned on arches that require it
2837 * and check if ack-trimming & collapsing extended the headroom
2838 * beyond what csum_start can cover.
2840 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2841 skb_headroom(skb) >= 0xFFFF)) {
2842 struct sk_buff *nskb;
2844 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2845 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2848 skb->skb_mstamp = tp->tcp_mstamp;
2849 tcp_rate_skb_sent(sk, skb);
2852 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2856 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2857 } else if (err != -EBUSY) {
2858 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2863 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2865 struct tcp_sock *tp = tcp_sk(sk);
2866 int err = __tcp_retransmit_skb(sk, skb, segs);
2869 #if FASTRETRANS_DEBUG > 0
2870 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2871 net_dbg_ratelimited("retrans_out leaked\n");
2874 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2875 tp->retrans_out += tcp_skb_pcount(skb);
2877 /* Save stamp of the first retransmit. */
2878 if (!tp->retrans_stamp)
2879 tp->retrans_stamp = tcp_skb_timestamp(skb);
2883 if (tp->undo_retrans < 0)
2884 tp->undo_retrans = 0;
2885 tp->undo_retrans += tcp_skb_pcount(skb);
2889 /* This gets called after a retransmit timeout, and the initially
2890 * retransmitted data is acknowledged. It tries to continue
2891 * resending the rest of the retransmit queue, until either
2892 * we've sent it all or the congestion window limit is reached.
2893 * If doing SACK, the first ACK which comes back for a timeout
2894 * based retransmit packet might feed us FACK information again.
2895 * If so, we use it to avoid unnecessarily retransmissions.
2897 void tcp_xmit_retransmit_queue(struct sock *sk)
2899 const struct inet_connection_sock *icsk = inet_csk(sk);
2900 struct tcp_sock *tp = tcp_sk(sk);
2901 struct sk_buff *skb;
2902 struct sk_buff *hole = NULL;
2906 if (!tp->packets_out)
2909 if (tp->retransmit_skb_hint) {
2910 skb = tp->retransmit_skb_hint;
2912 skb = tcp_write_queue_head(sk);
2915 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2916 tcp_for_write_queue_from(skb, sk) {
2920 if (skb == tcp_send_head(sk))
2923 if (tcp_pacing_check(sk))
2926 /* we could do better than to assign each time */
2928 tp->retransmit_skb_hint = skb;
2930 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2933 sacked = TCP_SKB_CB(skb)->sacked;
2934 /* In case tcp_shift_skb_data() have aggregated large skbs,
2935 * we need to make sure not sending too bigs TSO packets
2937 segs = min_t(int, segs, max_segs);
2939 if (tp->retrans_out >= tp->lost_out) {
2941 } else if (!(sacked & TCPCB_LOST)) {
2942 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2947 if (icsk->icsk_ca_state != TCP_CA_Loss)
2948 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2950 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2953 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2956 if (tcp_small_queue_check(sk, skb, 1))
2959 if (tcp_retransmit_skb(sk, skb, segs))
2962 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
2964 if (tcp_in_cwnd_reduction(sk))
2965 tp->prr_out += tcp_skb_pcount(skb);
2967 if (skb == tcp_write_queue_head(sk) &&
2968 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
2969 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2970 inet_csk(sk)->icsk_rto,
2975 /* We allow to exceed memory limits for FIN packets to expedite
2976 * connection tear down and (memory) recovery.
2977 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2978 * or even be forced to close flow without any FIN.
2979 * In general, we want to allow one skb per socket to avoid hangs
2980 * with edge trigger epoll()
2982 void sk_forced_mem_schedule(struct sock *sk, int size)
2986 if (size <= sk->sk_forward_alloc)
2988 amt = sk_mem_pages(size);
2989 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2990 sk_memory_allocated_add(sk, amt);
2992 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
2993 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
2996 /* Send a FIN. The caller locks the socket for us.
2997 * We should try to send a FIN packet really hard, but eventually give up.
2999 void tcp_send_fin(struct sock *sk)
3001 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3002 struct tcp_sock *tp = tcp_sk(sk);
3004 /* Optimization, tack on the FIN if we have one skb in write queue and
3005 * this skb was not yet sent, or we are under memory pressure.
3006 * Note: in the latter case, FIN packet will be sent after a timeout,
3007 * as TCP stack thinks it has already been transmitted.
3009 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
3011 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3012 TCP_SKB_CB(tskb)->end_seq++;
3014 if (!tcp_send_head(sk)) {
3015 /* This means tskb was already sent.
3016 * Pretend we included the FIN on previous transmit.
3017 * We need to set tp->snd_nxt to the value it would have
3018 * if FIN had been sent. This is because retransmit path
3019 * does not change tp->snd_nxt.
3025 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3026 if (unlikely(!skb)) {
3031 skb_reserve(skb, MAX_TCP_HEADER);
3032 sk_forced_mem_schedule(sk, skb->truesize);
3033 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3034 tcp_init_nondata_skb(skb, tp->write_seq,
3035 TCPHDR_ACK | TCPHDR_FIN);
3036 tcp_queue_skb(sk, skb);
3038 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3041 /* We get here when a process closes a file descriptor (either due to
3042 * an explicit close() or as a byproduct of exit()'ing) and there
3043 * was unread data in the receive queue. This behavior is recommended
3044 * by RFC 2525, section 2.17. -DaveM
3046 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3048 struct sk_buff *skb;
3050 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3052 /* NOTE: No TCP options attached and we never retransmit this. */
3053 skb = alloc_skb(MAX_TCP_HEADER, priority);
3055 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3059 /* Reserve space for headers and prepare control bits. */
3060 skb_reserve(skb, MAX_TCP_HEADER);
3061 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3062 TCPHDR_ACK | TCPHDR_RST);
3063 tcp_mstamp_refresh(tcp_sk(sk));
3065 if (tcp_transmit_skb(sk, skb, 0, priority))
3066 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3069 /* Send a crossed SYN-ACK during socket establishment.
3070 * WARNING: This routine must only be called when we have already sent
3071 * a SYN packet that crossed the incoming SYN that caused this routine
3072 * to get called. If this assumption fails then the initial rcv_wnd
3073 * and rcv_wscale values will not be correct.
3075 int tcp_send_synack(struct sock *sk)
3077 struct sk_buff *skb;
3079 skb = tcp_write_queue_head(sk);
3080 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3081 pr_debug("%s: wrong queue state\n", __func__);
3084 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3085 if (skb_cloned(skb)) {
3086 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
3089 tcp_unlink_write_queue(skb, sk);
3090 __skb_header_release(nskb);
3091 __tcp_add_write_queue_head(sk, nskb);
3092 sk_wmem_free_skb(sk, skb);
3093 sk->sk_wmem_queued += nskb->truesize;
3094 sk_mem_charge(sk, nskb->truesize);
3098 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3099 tcp_ecn_send_synack(sk, skb);
3101 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3105 * tcp_make_synack - Prepare a SYN-ACK.
3106 * sk: listener socket
3107 * dst: dst entry attached to the SYNACK
3108 * req: request_sock pointer
3110 * Allocate one skb and build a SYNACK packet.
3111 * @dst is consumed : Caller should not use it again.
3113 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3114 struct request_sock *req,
3115 struct tcp_fastopen_cookie *foc,
3116 enum tcp_synack_type synack_type)
3118 struct inet_request_sock *ireq = inet_rsk(req);
3119 const struct tcp_sock *tp = tcp_sk(sk);
3120 struct tcp_md5sig_key *md5 = NULL;
3121 struct tcp_out_options opts;
3122 struct sk_buff *skb;
3123 int tcp_header_size;
3127 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3128 if (unlikely(!skb)) {
3132 /* Reserve space for headers. */
3133 skb_reserve(skb, MAX_TCP_HEADER);
3135 switch (synack_type) {
3136 case TCP_SYNACK_NORMAL:
3137 skb_set_owner_w(skb, req_to_sk(req));
3139 case TCP_SYNACK_COOKIE:
3140 /* Under synflood, we do not attach skb to a socket,
3141 * to avoid false sharing.
3144 case TCP_SYNACK_FASTOPEN:
3145 /* sk is a const pointer, because we want to express multiple
3146 * cpu might call us concurrently.
3147 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3149 skb_set_owner_w(skb, (struct sock *)sk);
3152 skb_dst_set(skb, dst);
3154 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3156 memset(&opts, 0, sizeof(opts));
3157 #ifdef CONFIG_SYN_COOKIES
3158 if (unlikely(req->cookie_ts))
3159 skb->skb_mstamp = cookie_init_timestamp(req);
3162 skb->skb_mstamp = tcp_clock_us();
3164 #ifdef CONFIG_TCP_MD5SIG
3166 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3168 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3169 tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
3172 skb_push(skb, tcp_header_size);
3173 skb_reset_transport_header(skb);
3175 th = (struct tcphdr *)skb->data;
3176 memset(th, 0, sizeof(struct tcphdr));
3179 tcp_ecn_make_synack(req, th);
3180 th->source = htons(ireq->ir_num);
3181 th->dest = ireq->ir_rmt_port;
3182 skb->mark = ireq->ir_mark;
3183 skb->ip_summed = CHECKSUM_PARTIAL;
3184 th->seq = htonl(tcp_rsk(req)->snt_isn);
3185 /* XXX data is queued and acked as is. No buffer/window check */
3186 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3188 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3189 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3190 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3191 th->doff = (tcp_header_size >> 2);
3192 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3194 #ifdef CONFIG_TCP_MD5SIG
3195 /* Okay, we have all we need - do the md5 hash if needed */
3197 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3198 md5, req_to_sk(req), skb);
3202 /* Do not fool tcpdump (if any), clean our debris */
3206 EXPORT_SYMBOL(tcp_make_synack);
3208 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3210 struct inet_connection_sock *icsk = inet_csk(sk);
3211 const struct tcp_congestion_ops *ca;
3212 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3214 if (ca_key == TCP_CA_UNSPEC)
3218 ca = tcp_ca_find_key(ca_key);
3219 if (likely(ca && try_module_get(ca->owner))) {
3220 module_put(icsk->icsk_ca_ops->owner);
3221 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3222 icsk->icsk_ca_ops = ca;
3227 /* Do all connect socket setups that can be done AF independent. */
3228 static void tcp_connect_init(struct sock *sk)
3230 const struct dst_entry *dst = __sk_dst_get(sk);
3231 struct tcp_sock *tp = tcp_sk(sk);
3235 /* We'll fix this up when we get a response from the other end.
3236 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3238 tp->tcp_header_len = sizeof(struct tcphdr);
3239 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3240 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3242 #ifdef CONFIG_TCP_MD5SIG
3243 if (tp->af_specific->md5_lookup(sk, sk))
3244 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3247 /* If user gave his TCP_MAXSEG, record it to clamp */
3248 if (tp->rx_opt.user_mss)
3249 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3252 tcp_sync_mss(sk, dst_mtu(dst));
3254 tcp_ca_dst_init(sk, dst);
3256 if (!tp->window_clamp)
3257 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3258 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3260 tcp_initialize_rcv_mss(sk);
3262 /* limit the window selection if the user enforce a smaller rx buffer */
3263 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3264 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3265 tp->window_clamp = tcp_full_space(sk);
3267 rcv_wnd = tcp_rwnd_init_bpf(sk);
3269 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3271 tcp_select_initial_window(tcp_full_space(sk),
3272 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3275 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3279 tp->rx_opt.rcv_wscale = rcv_wscale;
3280 tp->rcv_ssthresh = tp->rcv_wnd;
3283 sock_reset_flag(sk, SOCK_DONE);
3286 tp->snd_una = tp->write_seq;
3287 tp->snd_sml = tp->write_seq;
3288 tp->snd_up = tp->write_seq;
3289 tp->snd_nxt = tp->write_seq;
3291 if (likely(!tp->repair))
3294 tp->rcv_tstamp = tcp_jiffies32;
3295 tp->rcv_wup = tp->rcv_nxt;
3296 tp->copied_seq = tp->rcv_nxt;
3298 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3299 inet_csk(sk)->icsk_retransmits = 0;
3300 tcp_clear_retrans(tp);
3303 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3305 struct tcp_sock *tp = tcp_sk(sk);
3306 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3308 tcb->end_seq += skb->len;
3309 __skb_header_release(skb);
3310 __tcp_add_write_queue_tail(sk, skb);
3311 sk->sk_wmem_queued += skb->truesize;
3312 sk_mem_charge(sk, skb->truesize);
3313 tp->write_seq = tcb->end_seq;
3314 tp->packets_out += tcp_skb_pcount(skb);
3317 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3318 * queue a data-only packet after the regular SYN, such that regular SYNs
3319 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3320 * only the SYN sequence, the data are retransmitted in the first ACK.
3321 * If cookie is not cached or other error occurs, falls back to send a
3322 * regular SYN with Fast Open cookie request option.
3324 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3326 struct tcp_sock *tp = tcp_sk(sk);
3327 struct tcp_fastopen_request *fo = tp->fastopen_req;
3329 struct sk_buff *syn_data;
3331 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3332 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3335 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3336 * user-MSS. Reserve maximum option space for middleboxes that add
3337 * private TCP options. The cost is reduced data space in SYN :(
3339 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3341 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3342 MAX_TCP_OPTION_SPACE;
3344 space = min_t(size_t, space, fo->size);
3346 /* limit to order-0 allocations */
3347 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3349 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3352 syn_data->ip_summed = CHECKSUM_PARTIAL;
3353 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3355 int copied = copy_from_iter(skb_put(syn_data, space), space,
3356 &fo->data->msg_iter);
3357 if (unlikely(!copied)) {
3358 kfree_skb(syn_data);
3361 if (copied != space) {
3362 skb_trim(syn_data, copied);
3366 /* No more data pending in inet_wait_for_connect() */
3367 if (space == fo->size)
3371 tcp_connect_queue_skb(sk, syn_data);
3373 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3375 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3377 syn->skb_mstamp = syn_data->skb_mstamp;
3379 /* Now full SYN+DATA was cloned and sent (or not),
3380 * remove the SYN from the original skb (syn_data)
3381 * we keep in write queue in case of a retransmit, as we
3382 * also have the SYN packet (with no data) in the same queue.
3384 TCP_SKB_CB(syn_data)->seq++;
3385 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3387 tp->syn_data = (fo->copied > 0);
3388 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3392 /* data was not sent, this is our new send_head */
3393 sk->sk_send_head = syn_data;
3394 tp->packets_out -= tcp_skb_pcount(syn_data);
3397 /* Send a regular SYN with Fast Open cookie request option */
3398 if (fo->cookie.len > 0)
3400 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3402 tp->syn_fastopen = 0;
3404 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3408 /* Build a SYN and send it off. */
3409 int tcp_connect(struct sock *sk)
3411 struct tcp_sock *tp = tcp_sk(sk);
3412 struct sk_buff *buff;
3415 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB);
3417 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3418 return -EHOSTUNREACH; /* Routing failure or similar. */
3420 tcp_connect_init(sk);
3422 if (unlikely(tp->repair)) {
3423 tcp_finish_connect(sk, NULL);
3427 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3428 if (unlikely(!buff))
3431 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3432 tcp_mstamp_refresh(tp);
3433 tp->retrans_stamp = tcp_time_stamp(tp);
3434 tcp_connect_queue_skb(sk, buff);
3435 tcp_ecn_send_syn(sk, buff);
3437 /* Send off SYN; include data in Fast Open. */
3438 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3439 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3440 if (err == -ECONNREFUSED)
3443 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3444 * in order to make this packet get counted in tcpOutSegs.
3446 tp->snd_nxt = tp->write_seq;
3447 tp->pushed_seq = tp->write_seq;
3448 buff = tcp_send_head(sk);
3449 if (unlikely(buff)) {
3450 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3451 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3453 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3455 /* Timer for repeating the SYN until an answer. */
3456 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3457 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3460 EXPORT_SYMBOL(tcp_connect);
3462 /* Send out a delayed ack, the caller does the policy checking
3463 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3466 void tcp_send_delayed_ack(struct sock *sk)
3468 struct inet_connection_sock *icsk = inet_csk(sk);
3469 int ato = icsk->icsk_ack.ato;
3470 unsigned long timeout;
3472 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3474 if (ato > TCP_DELACK_MIN) {
3475 const struct tcp_sock *tp = tcp_sk(sk);
3476 int max_ato = HZ / 2;
3478 if (icsk->icsk_ack.pingpong ||
3479 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3480 max_ato = TCP_DELACK_MAX;
3482 /* Slow path, intersegment interval is "high". */
3484 /* If some rtt estimate is known, use it to bound delayed ack.
3485 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3489 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3496 ato = min(ato, max_ato);
3499 /* Stay within the limit we were given */
3500 timeout = jiffies + ato;
3502 /* Use new timeout only if there wasn't a older one earlier. */
3503 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3504 /* If delack timer was blocked or is about to expire,
3507 if (icsk->icsk_ack.blocked ||
3508 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3513 if (!time_before(timeout, icsk->icsk_ack.timeout))
3514 timeout = icsk->icsk_ack.timeout;
3516 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3517 icsk->icsk_ack.timeout = timeout;
3518 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3521 /* This routine sends an ack and also updates the window. */
3522 void tcp_send_ack(struct sock *sk)
3524 struct sk_buff *buff;
3526 /* If we have been reset, we may not send again. */
3527 if (sk->sk_state == TCP_CLOSE)
3530 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3532 /* We are not putting this on the write queue, so
3533 * tcp_transmit_skb() will set the ownership to this
3536 buff = alloc_skb(MAX_TCP_HEADER,
3537 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3538 if (unlikely(!buff)) {
3539 inet_csk_schedule_ack(sk);
3540 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3541 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3542 TCP_DELACK_MAX, TCP_RTO_MAX);
3546 /* Reserve space for headers and prepare control bits. */
3547 skb_reserve(buff, MAX_TCP_HEADER);
3548 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3550 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3552 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3554 skb_set_tcp_pure_ack(buff);
3556 /* Send it off, this clears delayed acks for us. */
3557 tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3559 EXPORT_SYMBOL_GPL(tcp_send_ack);
3561 /* This routine sends a packet with an out of date sequence
3562 * number. It assumes the other end will try to ack it.
3564 * Question: what should we make while urgent mode?
3565 * 4.4BSD forces sending single byte of data. We cannot send
3566 * out of window data, because we have SND.NXT==SND.MAX...
3568 * Current solution: to send TWO zero-length segments in urgent mode:
3569 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3570 * out-of-date with SND.UNA-1 to probe window.
3572 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3574 struct tcp_sock *tp = tcp_sk(sk);
3575 struct sk_buff *skb;
3577 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3578 skb = alloc_skb(MAX_TCP_HEADER,
3579 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3583 /* Reserve space for headers and set control bits. */
3584 skb_reserve(skb, MAX_TCP_HEADER);
3585 /* Use a previous sequence. This should cause the other
3586 * end to send an ack. Don't queue or clone SKB, just
3589 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3590 NET_INC_STATS(sock_net(sk), mib);
3591 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3594 /* Called from setsockopt( ... TCP_REPAIR ) */
3595 void tcp_send_window_probe(struct sock *sk)
3597 if (sk->sk_state == TCP_ESTABLISHED) {
3598 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3599 tcp_mstamp_refresh(tcp_sk(sk));
3600 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3604 /* Initiate keepalive or window probe from timer. */
3605 int tcp_write_wakeup(struct sock *sk, int mib)
3607 struct tcp_sock *tp = tcp_sk(sk);
3608 struct sk_buff *skb;
3610 if (sk->sk_state == TCP_CLOSE)
3613 skb = tcp_send_head(sk);
3614 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3616 unsigned int mss = tcp_current_mss(sk);
3617 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3619 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3620 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3622 /* We are probing the opening of a window
3623 * but the window size is != 0
3624 * must have been a result SWS avoidance ( sender )
3626 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3628 seg_size = min(seg_size, mss);
3629 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3630 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3632 } else if (!tcp_skb_pcount(skb))
3633 tcp_set_skb_tso_segs(skb, mss);
3635 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3636 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3638 tcp_event_new_data_sent(sk, skb);
3641 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3642 tcp_xmit_probe_skb(sk, 1, mib);
3643 return tcp_xmit_probe_skb(sk, 0, mib);
3647 /* A window probe timeout has occurred. If window is not closed send
3648 * a partial packet else a zero probe.
3650 void tcp_send_probe0(struct sock *sk)
3652 struct inet_connection_sock *icsk = inet_csk(sk);
3653 struct tcp_sock *tp = tcp_sk(sk);
3654 struct net *net = sock_net(sk);
3655 unsigned long probe_max;
3658 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3660 if (tp->packets_out || !tcp_send_head(sk)) {
3661 /* Cancel probe timer, if it is not required. */
3662 icsk->icsk_probes_out = 0;
3663 icsk->icsk_backoff = 0;
3668 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3669 icsk->icsk_backoff++;
3670 icsk->icsk_probes_out++;
3671 probe_max = TCP_RTO_MAX;
3673 /* If packet was not sent due to local congestion,
3674 * do not backoff and do not remember icsk_probes_out.
3675 * Let local senders to fight for local resources.
3677 * Use accumulated backoff yet.
3679 if (!icsk->icsk_probes_out)
3680 icsk->icsk_probes_out = 1;
3681 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3683 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3684 tcp_probe0_when(sk, probe_max),
3688 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3690 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3694 tcp_rsk(req)->txhash = net_tx_rndhash();
3695 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3697 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3698 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3699 if (unlikely(tcp_passive_fastopen(sk)))
3700 tcp_sk(sk)->total_retrans++;
3704 EXPORT_SYMBOL(tcp_rtx_synack);