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_xmit_retransmit_queue(sk);
745 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
750 * One tasklet per cpu tries to send more skbs.
751 * We run in tasklet context but need to disable irqs when
752 * transferring tsq->head because tcp_wfree() might
753 * interrupt us (non NAPI drivers)
755 static void tcp_tasklet_func(unsigned long data)
757 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
760 struct list_head *q, *n;
764 local_irq_save(flags);
765 list_splice_init(&tsq->head, &list);
766 local_irq_restore(flags);
768 list_for_each_safe(q, n, &list) {
769 tp = list_entry(q, struct tcp_sock, tsq_node);
770 list_del(&tp->tsq_node);
772 sk = (struct sock *)tp;
773 smp_mb__before_atomic();
774 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
776 if (!sk->sk_lock.owned &&
777 test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) {
779 if (!sock_owned_by_user(sk)) {
780 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
790 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
791 TCPF_WRITE_TIMER_DEFERRED | \
792 TCPF_DELACK_TIMER_DEFERRED | \
793 TCPF_MTU_REDUCED_DEFERRED)
795 * tcp_release_cb - tcp release_sock() callback
798 * called from release_sock() to perform protocol dependent
799 * actions before socket release.
801 void tcp_release_cb(struct sock *sk)
803 unsigned long flags, nflags;
805 /* perform an atomic operation only if at least one flag is set */
807 flags = sk->sk_tsq_flags;
808 if (!(flags & TCP_DEFERRED_ALL))
810 nflags = flags & ~TCP_DEFERRED_ALL;
811 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
813 if (flags & TCPF_TSQ_DEFERRED)
816 /* Here begins the tricky part :
817 * We are called from release_sock() with :
819 * 2) sk_lock.slock spinlock held
820 * 3) socket owned by us (sk->sk_lock.owned == 1)
822 * But following code is meant to be called from BH handlers,
823 * so we should keep BH disabled, but early release socket ownership
825 sock_release_ownership(sk);
827 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
828 tcp_write_timer_handler(sk);
831 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
832 tcp_delack_timer_handler(sk);
835 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
836 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
840 EXPORT_SYMBOL(tcp_release_cb);
842 void __init tcp_tasklet_init(void)
846 for_each_possible_cpu(i) {
847 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
849 INIT_LIST_HEAD(&tsq->head);
850 tasklet_init(&tsq->tasklet,
857 * Write buffer destructor automatically called from kfree_skb.
858 * We can't xmit new skbs from this context, as we might already
861 void tcp_wfree(struct sk_buff *skb)
863 struct sock *sk = skb->sk;
864 struct tcp_sock *tp = tcp_sk(sk);
865 unsigned long flags, nval, oval;
867 /* Keep one reference on sk_wmem_alloc.
868 * Will be released by sk_free() from here or tcp_tasklet_func()
870 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
872 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
873 * Wait until our queues (qdisc + devices) are drained.
875 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
876 * - chance for incoming ACK (processed by another cpu maybe)
877 * to migrate this flow (skb->ooo_okay will be eventually set)
879 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
882 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
883 struct tsq_tasklet *tsq;
886 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
889 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
890 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
894 /* queue this socket to tasklet queue */
895 local_irq_save(flags);
896 tsq = this_cpu_ptr(&tsq_tasklet);
897 empty = list_empty(&tsq->head);
898 list_add(&tp->tsq_node, &tsq->head);
900 tasklet_schedule(&tsq->tasklet);
901 local_irq_restore(flags);
908 /* Note: Called under hard irq.
909 * We can not call TCP stack right away.
911 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
913 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
914 struct sock *sk = (struct sock *)tp;
915 unsigned long nval, oval;
917 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
918 struct tsq_tasklet *tsq;
921 if (oval & TSQF_QUEUED)
924 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
925 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
929 if (!refcount_inc_not_zero(&sk->sk_wmem_alloc))
931 /* queue this socket to tasklet queue */
932 tsq = this_cpu_ptr(&tsq_tasklet);
933 empty = list_empty(&tsq->head);
934 list_add(&tp->tsq_node, &tsq->head);
936 tasklet_schedule(&tsq->tasklet);
939 return HRTIMER_NORESTART;
942 /* BBR congestion control needs pacing.
943 * Same remark for SO_MAX_PACING_RATE.
944 * sch_fq packet scheduler is efficiently handling pacing,
945 * but is not always installed/used.
946 * Return true if TCP stack should pace packets itself.
948 static bool tcp_needs_internal_pacing(const struct sock *sk)
950 return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
953 static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
958 if (!tcp_needs_internal_pacing(sk))
960 rate = sk->sk_pacing_rate;
961 if (!rate || rate == ~0U)
964 /* Should account for header sizes as sch_fq does,
965 * but lets make things simple.
967 len_ns = (u64)skb->len * NSEC_PER_SEC;
968 do_div(len_ns, rate);
969 hrtimer_start(&tcp_sk(sk)->pacing_timer,
970 ktime_add_ns(ktime_get(), len_ns),
971 HRTIMER_MODE_ABS_PINNED);
974 /* This routine actually transmits TCP packets queued in by
975 * tcp_do_sendmsg(). This is used by both the initial
976 * transmission and possible later retransmissions.
977 * All SKB's seen here are completely headerless. It is our
978 * job to build the TCP header, and pass the packet down to
979 * IP so it can do the same plus pass the packet off to the
982 * We are working here with either a clone of the original
983 * SKB, or a fresh unique copy made by the retransmit engine.
985 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
988 const struct inet_connection_sock *icsk = inet_csk(sk);
989 struct inet_sock *inet;
991 struct tcp_skb_cb *tcb;
992 struct tcp_out_options opts;
993 unsigned int tcp_options_size, tcp_header_size;
994 struct tcp_md5sig_key *md5;
998 BUG_ON(!skb || !tcp_skb_pcount(skb));
1001 skb->skb_mstamp = tp->tcp_mstamp;
1003 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1005 tcp_rate_skb_sent(sk, skb);
1007 if (unlikely(skb_cloned(skb)))
1008 skb = pskb_copy(skb, gfp_mask);
1010 skb = skb_clone(skb, gfp_mask);
1016 tcb = TCP_SKB_CB(skb);
1017 memset(&opts, 0, sizeof(opts));
1019 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1020 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1022 tcp_options_size = tcp_established_options(sk, skb, &opts,
1024 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1026 /* if no packet is in qdisc/device queue, then allow XPS to select
1027 * another queue. We can be called from tcp_tsq_handler()
1028 * which holds one reference to sk_wmem_alloc.
1030 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1031 * One way to get this would be to set skb->truesize = 2 on them.
1033 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1035 /* If we had to use memory reserve to allocate this skb,
1036 * this might cause drops if packet is looped back :
1037 * Other socket might not have SOCK_MEMALLOC.
1038 * Packets not looped back do not care about pfmemalloc.
1040 skb->pfmemalloc = 0;
1042 skb_push(skb, tcp_header_size);
1043 skb_reset_transport_header(skb);
1047 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1048 skb_set_hash_from_sk(skb, sk);
1049 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1051 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1053 /* Build TCP header and checksum it. */
1054 th = (struct tcphdr *)skb->data;
1055 th->source = inet->inet_sport;
1056 th->dest = inet->inet_dport;
1057 th->seq = htonl(tcb->seq);
1058 th->ack_seq = htonl(tp->rcv_nxt);
1059 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1065 /* The urg_mode check is necessary during a below snd_una win probe */
1066 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1067 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1068 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1070 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1071 th->urg_ptr = htons(0xFFFF);
1076 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1077 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1078 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1079 th->window = htons(tcp_select_window(sk));
1080 tcp_ecn_send(sk, skb, th, tcp_header_size);
1082 /* RFC1323: The window in SYN & SYN/ACK segments
1085 th->window = htons(min(tp->rcv_wnd, 65535U));
1087 #ifdef CONFIG_TCP_MD5SIG
1088 /* Calculate the MD5 hash, as we have all we need now */
1090 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1091 tp->af_specific->calc_md5_hash(opts.hash_location,
1096 icsk->icsk_af_ops->send_check(sk, skb);
1098 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1099 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1101 if (skb->len != tcp_header_size) {
1102 tcp_event_data_sent(tp, sk);
1103 tp->data_segs_out += tcp_skb_pcount(skb);
1104 tcp_internal_pacing(sk, skb);
1107 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1108 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1109 tcp_skb_pcount(skb));
1111 tp->segs_out += tcp_skb_pcount(skb);
1112 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1113 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1114 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1116 /* Our usage of tstamp should remain private */
1119 /* Cleanup our debris for IP stacks */
1120 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1121 sizeof(struct inet6_skb_parm)));
1123 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1125 if (likely(err <= 0))
1130 return net_xmit_eval(err);
1133 /* This routine just queues the buffer for sending.
1135 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1136 * otherwise socket can stall.
1138 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1140 struct tcp_sock *tp = tcp_sk(sk);
1142 /* Advance write_seq and place onto the write_queue. */
1143 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1144 __skb_header_release(skb);
1145 tcp_add_write_queue_tail(sk, skb);
1146 sk->sk_wmem_queued += skb->truesize;
1147 sk_mem_charge(sk, skb->truesize);
1150 /* Initialize TSO segments for a packet. */
1151 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1153 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1154 /* Avoid the costly divide in the normal
1157 tcp_skb_pcount_set(skb, 1);
1158 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1160 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1161 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1165 /* When a modification to fackets out becomes necessary, we need to check
1166 * skb is counted to fackets_out or not.
1168 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1171 struct tcp_sock *tp = tcp_sk(sk);
1173 if (!tp->sacked_out || tcp_is_reno(tp))
1176 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1177 tp->fackets_out -= decr;
1180 /* Pcount in the middle of the write queue got changed, we need to do various
1181 * tweaks to fix counters
1183 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1185 struct tcp_sock *tp = tcp_sk(sk);
1187 tp->packets_out -= decr;
1189 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1190 tp->sacked_out -= decr;
1191 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1192 tp->retrans_out -= decr;
1193 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1194 tp->lost_out -= decr;
1196 /* Reno case is special. Sigh... */
1197 if (tcp_is_reno(tp) && decr > 0)
1198 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1200 tcp_adjust_fackets_out(sk, skb, decr);
1202 if (tp->lost_skb_hint &&
1203 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1204 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1205 tp->lost_cnt_hint -= decr;
1207 tcp_verify_left_out(tp);
1210 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1212 return TCP_SKB_CB(skb)->txstamp_ack ||
1213 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1216 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1218 struct skb_shared_info *shinfo = skb_shinfo(skb);
1220 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1221 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1222 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1223 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1225 shinfo->tx_flags &= ~tsflags;
1226 shinfo2->tx_flags |= tsflags;
1227 swap(shinfo->tskey, shinfo2->tskey);
1228 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1229 TCP_SKB_CB(skb)->txstamp_ack = 0;
1233 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1235 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1236 TCP_SKB_CB(skb)->eor = 0;
1239 /* Function to create two new TCP segments. Shrinks the given segment
1240 * to the specified size and appends a new segment with the rest of the
1241 * packet to the list. This won't be called frequently, I hope.
1242 * Remember, these are still headerless SKBs at this point.
1244 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1245 unsigned int mss_now, gfp_t gfp)
1247 struct tcp_sock *tp = tcp_sk(sk);
1248 struct sk_buff *buff;
1249 int nsize, old_factor;
1253 if (WARN_ON(len > skb->len))
1256 nsize = skb_headlen(skb) - len;
1260 if (skb_unclone(skb, gfp))
1263 /* Get a new skb... force flag on. */
1264 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1266 return -ENOMEM; /* We'll just try again later. */
1268 sk->sk_wmem_queued += buff->truesize;
1269 sk_mem_charge(sk, buff->truesize);
1270 nlen = skb->len - len - nsize;
1271 buff->truesize += nlen;
1272 skb->truesize -= nlen;
1274 /* Correct the sequence numbers. */
1275 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1276 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1277 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1279 /* PSH and FIN should only be set in the second packet. */
1280 flags = TCP_SKB_CB(skb)->tcp_flags;
1281 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1282 TCP_SKB_CB(buff)->tcp_flags = flags;
1283 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1284 tcp_skb_fragment_eor(skb, buff);
1286 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1287 /* Copy and checksum data tail into the new buffer. */
1288 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1289 skb_put(buff, nsize),
1294 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1296 skb->ip_summed = CHECKSUM_PARTIAL;
1297 skb_split(skb, buff, len);
1300 buff->ip_summed = skb->ip_summed;
1302 buff->tstamp = skb->tstamp;
1303 tcp_fragment_tstamp(skb, buff);
1305 old_factor = tcp_skb_pcount(skb);
1307 /* Fix up tso_factor for both original and new SKB. */
1308 tcp_set_skb_tso_segs(skb, mss_now);
1309 tcp_set_skb_tso_segs(buff, mss_now);
1311 /* Update delivered info for the new segment */
1312 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1314 /* If this packet has been sent out already, we must
1315 * adjust the various packet counters.
1317 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1318 int diff = old_factor - tcp_skb_pcount(skb) -
1319 tcp_skb_pcount(buff);
1322 tcp_adjust_pcount(sk, skb, diff);
1325 /* Link BUFF into the send queue. */
1326 __skb_header_release(buff);
1327 tcp_insert_write_queue_after(skb, buff, sk);
1332 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1333 * data is not copied, but immediately discarded.
1335 static int __pskb_trim_head(struct sk_buff *skb, int len)
1337 struct skb_shared_info *shinfo;
1340 eat = min_t(int, len, skb_headlen(skb));
1342 __skb_pull(skb, eat);
1349 shinfo = skb_shinfo(skb);
1350 for (i = 0; i < shinfo->nr_frags; i++) {
1351 int size = skb_frag_size(&shinfo->frags[i]);
1354 skb_frag_unref(skb, i);
1357 shinfo->frags[k] = shinfo->frags[i];
1359 shinfo->frags[k].page_offset += eat;
1360 skb_frag_size_sub(&shinfo->frags[k], eat);
1366 shinfo->nr_frags = k;
1368 skb->data_len -= len;
1369 skb->len = skb->data_len;
1373 /* Remove acked data from a packet in the transmit queue. */
1374 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1378 if (skb_unclone(skb, GFP_ATOMIC))
1381 delta_truesize = __pskb_trim_head(skb, len);
1383 TCP_SKB_CB(skb)->seq += len;
1384 skb->ip_summed = CHECKSUM_PARTIAL;
1386 if (delta_truesize) {
1387 skb->truesize -= delta_truesize;
1388 sk->sk_wmem_queued -= delta_truesize;
1389 sk_mem_uncharge(sk, delta_truesize);
1390 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1393 /* Any change of skb->len requires recalculation of tso factor. */
1394 if (tcp_skb_pcount(skb) > 1)
1395 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1400 /* Calculate MSS not accounting any TCP options. */
1401 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1403 const struct tcp_sock *tp = tcp_sk(sk);
1404 const struct inet_connection_sock *icsk = inet_csk(sk);
1407 /* Calculate base mss without TCP options:
1408 It is MMS_S - sizeof(tcphdr) of rfc1122
1410 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1412 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1413 if (icsk->icsk_af_ops->net_frag_header_len) {
1414 const struct dst_entry *dst = __sk_dst_get(sk);
1416 if (dst && dst_allfrag(dst))
1417 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1420 /* Clamp it (mss_clamp does not include tcp options) */
1421 if (mss_now > tp->rx_opt.mss_clamp)
1422 mss_now = tp->rx_opt.mss_clamp;
1424 /* Now subtract optional transport overhead */
1425 mss_now -= icsk->icsk_ext_hdr_len;
1427 /* Then reserve room for full set of TCP options and 8 bytes of data */
1433 /* Calculate MSS. Not accounting for SACKs here. */
1434 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1436 /* Subtract TCP options size, not including SACKs */
1437 return __tcp_mtu_to_mss(sk, pmtu) -
1438 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1441 /* Inverse of above */
1442 int tcp_mss_to_mtu(struct sock *sk, int mss)
1444 const struct tcp_sock *tp = tcp_sk(sk);
1445 const struct inet_connection_sock *icsk = inet_csk(sk);
1449 tp->tcp_header_len +
1450 icsk->icsk_ext_hdr_len +
1451 icsk->icsk_af_ops->net_header_len;
1453 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1454 if (icsk->icsk_af_ops->net_frag_header_len) {
1455 const struct dst_entry *dst = __sk_dst_get(sk);
1457 if (dst && dst_allfrag(dst))
1458 mtu += icsk->icsk_af_ops->net_frag_header_len;
1462 EXPORT_SYMBOL(tcp_mss_to_mtu);
1464 /* MTU probing init per socket */
1465 void tcp_mtup_init(struct sock *sk)
1467 struct tcp_sock *tp = tcp_sk(sk);
1468 struct inet_connection_sock *icsk = inet_csk(sk);
1469 struct net *net = sock_net(sk);
1471 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1472 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1473 icsk->icsk_af_ops->net_header_len;
1474 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1475 icsk->icsk_mtup.probe_size = 0;
1476 if (icsk->icsk_mtup.enabled)
1477 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1479 EXPORT_SYMBOL(tcp_mtup_init);
1481 /* This function synchronize snd mss to current pmtu/exthdr set.
1483 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1484 for TCP options, but includes only bare TCP header.
1486 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1487 It is minimum of user_mss and mss received with SYN.
1488 It also does not include TCP options.
1490 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1492 tp->mss_cache is current effective sending mss, including
1493 all tcp options except for SACKs. It is evaluated,
1494 taking into account current pmtu, but never exceeds
1495 tp->rx_opt.mss_clamp.
1497 NOTE1. rfc1122 clearly states that advertised MSS
1498 DOES NOT include either tcp or ip options.
1500 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1501 are READ ONLY outside this function. --ANK (980731)
1503 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1505 struct tcp_sock *tp = tcp_sk(sk);
1506 struct inet_connection_sock *icsk = inet_csk(sk);
1509 if (icsk->icsk_mtup.search_high > pmtu)
1510 icsk->icsk_mtup.search_high = pmtu;
1512 mss_now = tcp_mtu_to_mss(sk, pmtu);
1513 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1515 /* And store cached results */
1516 icsk->icsk_pmtu_cookie = pmtu;
1517 if (icsk->icsk_mtup.enabled)
1518 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1519 tp->mss_cache = mss_now;
1523 EXPORT_SYMBOL(tcp_sync_mss);
1525 /* Compute the current effective MSS, taking SACKs and IP options,
1526 * and even PMTU discovery events into account.
1528 unsigned int tcp_current_mss(struct sock *sk)
1530 const struct tcp_sock *tp = tcp_sk(sk);
1531 const struct dst_entry *dst = __sk_dst_get(sk);
1533 unsigned int header_len;
1534 struct tcp_out_options opts;
1535 struct tcp_md5sig_key *md5;
1537 mss_now = tp->mss_cache;
1540 u32 mtu = dst_mtu(dst);
1541 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1542 mss_now = tcp_sync_mss(sk, mtu);
1545 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1546 sizeof(struct tcphdr);
1547 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1548 * some common options. If this is an odd packet (because we have SACK
1549 * blocks etc) then our calculated header_len will be different, and
1550 * we have to adjust mss_now correspondingly */
1551 if (header_len != tp->tcp_header_len) {
1552 int delta = (int) header_len - tp->tcp_header_len;
1559 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1560 * As additional protections, we do not touch cwnd in retransmission phases,
1561 * and if application hit its sndbuf limit recently.
1563 static void tcp_cwnd_application_limited(struct sock *sk)
1565 struct tcp_sock *tp = tcp_sk(sk);
1567 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1568 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1569 /* Limited by application or receiver window. */
1570 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1571 u32 win_used = max(tp->snd_cwnd_used, init_win);
1572 if (win_used < tp->snd_cwnd) {
1573 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1574 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1576 tp->snd_cwnd_used = 0;
1578 tp->snd_cwnd_stamp = tcp_jiffies32;
1581 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1583 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1584 struct tcp_sock *tp = tcp_sk(sk);
1586 /* Track the maximum number of outstanding packets in each
1587 * window, and remember whether we were cwnd-limited then.
1589 if (!before(tp->snd_una, tp->max_packets_seq) ||
1590 tp->packets_out > tp->max_packets_out) {
1591 tp->max_packets_out = tp->packets_out;
1592 tp->max_packets_seq = tp->snd_nxt;
1593 tp->is_cwnd_limited = is_cwnd_limited;
1596 if (tcp_is_cwnd_limited(sk)) {
1597 /* Network is feed fully. */
1598 tp->snd_cwnd_used = 0;
1599 tp->snd_cwnd_stamp = tcp_jiffies32;
1601 /* Network starves. */
1602 if (tp->packets_out > tp->snd_cwnd_used)
1603 tp->snd_cwnd_used = tp->packets_out;
1605 if (sysctl_tcp_slow_start_after_idle &&
1606 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1607 !ca_ops->cong_control)
1608 tcp_cwnd_application_limited(sk);
1610 /* The following conditions together indicate the starvation
1611 * is caused by insufficient sender buffer:
1612 * 1) just sent some data (see tcp_write_xmit)
1613 * 2) not cwnd limited (this else condition)
1614 * 3) no more data to send (null tcp_send_head )
1615 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1617 if (!tcp_send_head(sk) && sk->sk_socket &&
1618 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1619 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1620 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1624 /* Minshall's variant of the Nagle send check. */
1625 static bool tcp_minshall_check(const struct tcp_sock *tp)
1627 return after(tp->snd_sml, tp->snd_una) &&
1628 !after(tp->snd_sml, tp->snd_nxt);
1631 /* Update snd_sml if this skb is under mss
1632 * Note that a TSO packet might end with a sub-mss segment
1633 * The test is really :
1634 * if ((skb->len % mss) != 0)
1635 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1636 * But we can avoid doing the divide again given we already have
1637 * skb_pcount = skb->len / mss_now
1639 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1640 const struct sk_buff *skb)
1642 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1643 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1646 /* Return false, if packet can be sent now without violation Nagle's rules:
1647 * 1. It is full sized. (provided by caller in %partial bool)
1648 * 2. Or it contains FIN. (already checked by caller)
1649 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1650 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1651 * With Minshall's modification: all sent small packets are ACKed.
1653 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1657 ((nonagle & TCP_NAGLE_CORK) ||
1658 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1661 /* Return how many segs we'd like on a TSO packet,
1662 * to send one TSO packet per ms
1664 u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1669 bytes = min(sk->sk_pacing_rate >> 10,
1670 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1672 /* Goal is to send at least one packet per ms,
1673 * not one big TSO packet every 100 ms.
1674 * This preserves ACK clocking and is consistent
1675 * with tcp_tso_should_defer() heuristic.
1677 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1679 return min_t(u32, segs, sk->sk_gso_max_segs);
1681 EXPORT_SYMBOL(tcp_tso_autosize);
1683 /* Return the number of segments we want in the skb we are transmitting.
1684 * See if congestion control module wants to decide; otherwise, autosize.
1686 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1688 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1689 u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0;
1692 tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs);
1695 /* Returns the portion of skb which can be sent right away */
1696 static unsigned int tcp_mss_split_point(const struct sock *sk,
1697 const struct sk_buff *skb,
1698 unsigned int mss_now,
1699 unsigned int max_segs,
1702 const struct tcp_sock *tp = tcp_sk(sk);
1703 u32 partial, needed, window, max_len;
1705 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1706 max_len = mss_now * max_segs;
1708 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1711 needed = min(skb->len, window);
1713 if (max_len <= needed)
1716 partial = needed % mss_now;
1717 /* If last segment is not a full MSS, check if Nagle rules allow us
1718 * to include this last segment in this skb.
1719 * Otherwise, we'll split the skb at last MSS boundary
1721 if (tcp_nagle_check(partial != 0, tp, nonagle))
1722 return needed - partial;
1727 /* Can at least one segment of SKB be sent right now, according to the
1728 * congestion window rules? If so, return how many segments are allowed.
1730 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1731 const struct sk_buff *skb)
1733 u32 in_flight, cwnd, halfcwnd;
1735 /* Don't be strict about the congestion window for the final FIN. */
1736 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1737 tcp_skb_pcount(skb) == 1)
1740 in_flight = tcp_packets_in_flight(tp);
1741 cwnd = tp->snd_cwnd;
1742 if (in_flight >= cwnd)
1745 /* For better scheduling, ensure we have at least
1746 * 2 GSO packets in flight.
1748 halfcwnd = max(cwnd >> 1, 1U);
1749 return min(halfcwnd, cwnd - in_flight);
1752 /* Initialize TSO state of a skb.
1753 * This must be invoked the first time we consider transmitting
1754 * SKB onto the wire.
1756 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1758 int tso_segs = tcp_skb_pcount(skb);
1760 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1761 tcp_set_skb_tso_segs(skb, mss_now);
1762 tso_segs = tcp_skb_pcount(skb);
1768 /* Return true if the Nagle test allows this packet to be
1771 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1772 unsigned int cur_mss, int nonagle)
1774 /* Nagle rule does not apply to frames, which sit in the middle of the
1775 * write_queue (they have no chances to get new data).
1777 * This is implemented in the callers, where they modify the 'nonagle'
1778 * argument based upon the location of SKB in the send queue.
1780 if (nonagle & TCP_NAGLE_PUSH)
1783 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1784 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1787 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1793 /* Does at least the first segment of SKB fit into the send window? */
1794 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1795 const struct sk_buff *skb,
1796 unsigned int cur_mss)
1798 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1800 if (skb->len > cur_mss)
1801 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1803 return !after(end_seq, tcp_wnd_end(tp));
1806 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1807 * should be put on the wire right now. If so, it returns the number of
1808 * packets allowed by the congestion window.
1810 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1811 unsigned int cur_mss, int nonagle)
1813 const struct tcp_sock *tp = tcp_sk(sk);
1814 unsigned int cwnd_quota;
1816 tcp_init_tso_segs(skb, cur_mss);
1818 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1821 cwnd_quota = tcp_cwnd_test(tp, skb);
1822 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1828 /* Test if sending is allowed right now. */
1829 bool tcp_may_send_now(struct sock *sk)
1831 const struct tcp_sock *tp = tcp_sk(sk);
1832 struct sk_buff *skb = tcp_send_head(sk);
1835 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1836 (tcp_skb_is_last(sk, skb) ?
1837 tp->nonagle : TCP_NAGLE_PUSH));
1840 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1841 * which is put after SKB on the list. It is very much like
1842 * tcp_fragment() except that it may make several kinds of assumptions
1843 * in order to speed up the splitting operation. In particular, we
1844 * know that all the data is in scatter-gather pages, and that the
1845 * packet has never been sent out before (and thus is not cloned).
1847 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1848 unsigned int mss_now, gfp_t gfp)
1850 struct sk_buff *buff;
1851 int nlen = skb->len - len;
1854 /* All of a TSO frame must be composed of paged data. */
1855 if (skb->len != skb->data_len)
1856 return tcp_fragment(sk, skb, len, mss_now, gfp);
1858 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1859 if (unlikely(!buff))
1862 sk->sk_wmem_queued += buff->truesize;
1863 sk_mem_charge(sk, buff->truesize);
1864 buff->truesize += nlen;
1865 skb->truesize -= nlen;
1867 /* Correct the sequence numbers. */
1868 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1869 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1870 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1872 /* PSH and FIN should only be set in the second packet. */
1873 flags = TCP_SKB_CB(skb)->tcp_flags;
1874 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1875 TCP_SKB_CB(buff)->tcp_flags = flags;
1877 /* This packet was never sent out yet, so no SACK bits. */
1878 TCP_SKB_CB(buff)->sacked = 0;
1880 tcp_skb_fragment_eor(skb, buff);
1882 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1883 skb_split(skb, buff, len);
1884 tcp_fragment_tstamp(skb, buff);
1886 /* Fix up tso_factor for both original and new SKB. */
1887 tcp_set_skb_tso_segs(skb, mss_now);
1888 tcp_set_skb_tso_segs(buff, mss_now);
1890 /* Link BUFF into the send queue. */
1891 __skb_header_release(buff);
1892 tcp_insert_write_queue_after(skb, buff, sk);
1897 /* Try to defer sending, if possible, in order to minimize the amount
1898 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1900 * This algorithm is from John Heffner.
1902 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1903 bool *is_cwnd_limited, u32 max_segs)
1905 const struct inet_connection_sock *icsk = inet_csk(sk);
1906 u32 age, send_win, cong_win, limit, in_flight;
1907 struct tcp_sock *tp = tcp_sk(sk);
1908 struct sk_buff *head;
1911 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1914 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1917 /* Avoid bursty behavior by allowing defer
1918 * only if the last write was recent.
1920 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
1923 in_flight = tcp_packets_in_flight(tp);
1925 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1927 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1929 /* From in_flight test above, we know that cwnd > in_flight. */
1930 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1932 limit = min(send_win, cong_win);
1934 /* If a full-sized TSO skb can be sent, do it. */
1935 if (limit >= max_segs * tp->mss_cache)
1938 /* Middle in queue won't get any more data, full sendable already? */
1939 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1942 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1944 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1946 /* If at least some fraction of a window is available,
1949 chunk /= win_divisor;
1953 /* Different approach, try not to defer past a single
1954 * ACK. Receiver should ACK every other full sized
1955 * frame, so if we have space for more than 3 frames
1958 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1962 head = tcp_write_queue_head(sk);
1964 age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
1965 /* If next ACK is likely to come too late (half srtt), do not defer */
1966 if (age < (tp->srtt_us >> 4))
1969 /* Ok, it looks like it is advisable to defer. */
1971 if (cong_win < send_win && cong_win <= skb->len)
1972 *is_cwnd_limited = true;
1980 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1982 struct inet_connection_sock *icsk = inet_csk(sk);
1983 struct tcp_sock *tp = tcp_sk(sk);
1984 struct net *net = sock_net(sk);
1988 interval = net->ipv4.sysctl_tcp_probe_interval;
1989 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
1990 if (unlikely(delta >= interval * HZ)) {
1991 int mss = tcp_current_mss(sk);
1993 /* Update current search range */
1994 icsk->icsk_mtup.probe_size = 0;
1995 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1996 sizeof(struct tcphdr) +
1997 icsk->icsk_af_ops->net_header_len;
1998 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2000 /* Update probe time stamp */
2001 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2005 /* Create a new MTU probe if we are ready.
2006 * MTU probe is regularly attempting to increase the path MTU by
2007 * deliberately sending larger packets. This discovers routing
2008 * changes resulting in larger path MTUs.
2010 * Returns 0 if we should wait to probe (no cwnd available),
2011 * 1 if a probe was sent,
2014 static int tcp_mtu_probe(struct sock *sk)
2016 struct inet_connection_sock *icsk = inet_csk(sk);
2017 struct tcp_sock *tp = tcp_sk(sk);
2018 struct sk_buff *skb, *nskb, *next;
2019 struct net *net = sock_net(sk);
2026 /* Not currently probing/verifying,
2028 * have enough cwnd, and
2029 * not SACKing (the variable headers throw things off)
2031 if (likely(!icsk->icsk_mtup.enabled ||
2032 icsk->icsk_mtup.probe_size ||
2033 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2034 tp->snd_cwnd < 11 ||
2035 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2038 /* Use binary search for probe_size between tcp_mss_base,
2039 * and current mss_clamp. if (search_high - search_low)
2040 * smaller than a threshold, backoff from probing.
2042 mss_now = tcp_current_mss(sk);
2043 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2044 icsk->icsk_mtup.search_low) >> 1);
2045 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2046 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2047 /* When misfortune happens, we are reprobing actively,
2048 * and then reprobe timer has expired. We stick with current
2049 * probing process by not resetting search range to its orignal.
2051 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2052 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2053 /* Check whether enough time has elaplased for
2054 * another round of probing.
2056 tcp_mtu_check_reprobe(sk);
2060 /* Have enough data in the send queue to probe? */
2061 if (tp->write_seq - tp->snd_nxt < size_needed)
2064 if (tp->snd_wnd < size_needed)
2066 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2069 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2070 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2071 if (!tcp_packets_in_flight(tp))
2077 /* We're allowed to probe. Build it now. */
2078 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2081 sk->sk_wmem_queued += nskb->truesize;
2082 sk_mem_charge(sk, nskb->truesize);
2084 skb = tcp_send_head(sk);
2086 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2087 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2088 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2089 TCP_SKB_CB(nskb)->sacked = 0;
2091 nskb->ip_summed = skb->ip_summed;
2093 tcp_insert_write_queue_before(nskb, skb, sk);
2096 tcp_for_write_queue_from_safe(skb, next, sk) {
2097 copy = min_t(int, skb->len, probe_size - len);
2098 if (nskb->ip_summed) {
2099 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2101 __wsum csum = skb_copy_and_csum_bits(skb, 0,
2102 skb_put(nskb, copy),
2104 nskb->csum = csum_block_add(nskb->csum, csum, len);
2107 if (skb->len <= copy) {
2108 /* We've eaten all the data from this skb.
2110 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2111 tcp_unlink_write_queue(skb, sk);
2112 sk_wmem_free_skb(sk, skb);
2114 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2115 ~(TCPHDR_FIN|TCPHDR_PSH);
2116 if (!skb_shinfo(skb)->nr_frags) {
2117 skb_pull(skb, copy);
2118 if (skb->ip_summed != CHECKSUM_PARTIAL)
2119 skb->csum = csum_partial(skb->data,
2122 __pskb_trim_head(skb, copy);
2123 tcp_set_skb_tso_segs(skb, mss_now);
2125 TCP_SKB_CB(skb)->seq += copy;
2130 if (len >= probe_size)
2133 tcp_init_tso_segs(nskb, nskb->len);
2135 /* We're ready to send. If this fails, the probe will
2136 * be resegmented into mss-sized pieces by tcp_write_xmit().
2138 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2139 /* Decrement cwnd here because we are sending
2140 * effectively two packets. */
2142 tcp_event_new_data_sent(sk, nskb);
2144 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2145 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2146 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2154 static bool tcp_pacing_check(const struct sock *sk)
2156 return tcp_needs_internal_pacing(sk) &&
2157 hrtimer_active(&tcp_sk(sk)->pacing_timer);
2160 /* TCP Small Queues :
2161 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2162 * (These limits are doubled for retransmits)
2164 * - better RTT estimation and ACK scheduling
2167 * Alas, some drivers / subsystems require a fair amount
2168 * of queued bytes to ensure line rate.
2169 * One example is wifi aggregation (802.11 AMPDU)
2171 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2172 unsigned int factor)
2176 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2177 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2180 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2181 /* Always send the 1st or 2nd skb in write queue.
2182 * No need to wait for TX completion to call us back,
2183 * after softirq/tasklet schedule.
2184 * This helps when TX completions are delayed too much.
2186 if (skb == sk->sk_write_queue.next ||
2187 skb->prev == sk->sk_write_queue.next)
2190 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2191 /* It is possible TX completion already happened
2192 * before we set TSQ_THROTTLED, so we must
2193 * test again the condition.
2195 smp_mb__after_atomic();
2196 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2202 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2204 const u32 now = tcp_jiffies32;
2206 if (tp->chrono_type > TCP_CHRONO_UNSPEC)
2207 tp->chrono_stat[tp->chrono_type - 1] += now - tp->chrono_start;
2208 tp->chrono_start = now;
2209 tp->chrono_type = new;
2212 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2214 struct tcp_sock *tp = tcp_sk(sk);
2216 /* If there are multiple conditions worthy of tracking in a
2217 * chronograph then the highest priority enum takes precedence
2218 * over the other conditions. So that if something "more interesting"
2219 * starts happening, stop the previous chrono and start a new one.
2221 if (type > tp->chrono_type)
2222 tcp_chrono_set(tp, type);
2225 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2227 struct tcp_sock *tp = tcp_sk(sk);
2230 /* There are multiple conditions worthy of tracking in a
2231 * chronograph, so that the highest priority enum takes
2232 * precedence over the other conditions (see tcp_chrono_start).
2233 * If a condition stops, we only stop chrono tracking if
2234 * it's the "most interesting" or current chrono we are
2235 * tracking and starts busy chrono if we have pending data.
2237 if (tcp_write_queue_empty(sk))
2238 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2239 else if (type == tp->chrono_type)
2240 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2243 /* This routine writes packets to the network. It advances the
2244 * send_head. This happens as incoming acks open up the remote
2247 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2248 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2249 * account rare use of URG, this is not a big flaw.
2251 * Send at most one packet when push_one > 0. Temporarily ignore
2252 * cwnd limit to force at most one packet out when push_one == 2.
2254 * Returns true, if no segments are in flight and we have queued segments,
2255 * but cannot send anything now because of SWS or another problem.
2257 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2258 int push_one, gfp_t gfp)
2260 struct tcp_sock *tp = tcp_sk(sk);
2261 struct sk_buff *skb;
2262 unsigned int tso_segs, sent_pkts;
2265 bool is_cwnd_limited = false, is_rwnd_limited = false;
2271 /* Do MTU probing. */
2272 result = tcp_mtu_probe(sk);
2275 } else if (result > 0) {
2280 max_segs = tcp_tso_segs(sk, mss_now);
2281 tcp_mstamp_refresh(tp);
2282 while ((skb = tcp_send_head(sk))) {
2285 if (tcp_pacing_check(sk))
2288 tso_segs = tcp_init_tso_segs(skb, mss_now);
2291 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2292 /* "skb_mstamp" is used as a start point for the retransmit timer */
2293 skb->skb_mstamp = tp->tcp_mstamp;
2294 goto repair; /* Skip network transmission */
2297 cwnd_quota = tcp_cwnd_test(tp, skb);
2300 /* Force out a loss probe pkt. */
2306 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2307 is_rwnd_limited = true;
2311 if (tso_segs == 1) {
2312 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2313 (tcp_skb_is_last(sk, skb) ?
2314 nonagle : TCP_NAGLE_PUSH))))
2318 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2324 if (tso_segs > 1 && !tcp_urg_mode(tp))
2325 limit = tcp_mss_split_point(sk, skb, mss_now,
2331 if (skb->len > limit &&
2332 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2335 if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
2336 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
2337 if (tcp_small_queue_check(sk, skb, 0))
2340 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2344 /* Advance the send_head. This one is sent out.
2345 * This call will increment packets_out.
2347 tcp_event_new_data_sent(sk, skb);
2349 tcp_minshall_update(tp, mss_now, skb);
2350 sent_pkts += tcp_skb_pcount(skb);
2356 if (is_rwnd_limited)
2357 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2359 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2361 if (likely(sent_pkts)) {
2362 if (tcp_in_cwnd_reduction(sk))
2363 tp->prr_out += sent_pkts;
2365 /* Send one loss probe per tail loss episode. */
2367 tcp_schedule_loss_probe(sk);
2368 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2369 tcp_cwnd_validate(sk, is_cwnd_limited);
2372 return !tp->packets_out && tcp_send_head(sk);
2375 bool tcp_schedule_loss_probe(struct sock *sk)
2377 struct inet_connection_sock *icsk = inet_csk(sk);
2378 struct tcp_sock *tp = tcp_sk(sk);
2379 u32 timeout, tlp_time_stamp, rto_time_stamp;
2380 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2382 /* No consecutive loss probes. */
2383 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2387 /* Don't do any loss probe on a Fast Open connection before 3WHS
2390 if (tp->fastopen_rsk)
2393 /* TLP is only scheduled when next timer event is RTO. */
2394 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2397 /* Schedule a loss probe in 2*RTT for SACK capable connections
2398 * in Open state, that are either limited by cwnd or application.
2400 if ((sysctl_tcp_early_retrans != 3 && sysctl_tcp_early_retrans != 4) ||
2401 !tp->packets_out || !tcp_is_sack(tp) ||
2402 icsk->icsk_ca_state != TCP_CA_Open)
2405 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2409 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2410 * for delayed ack when there's one outstanding packet. If no RTT
2411 * sample is available then probe after TCP_TIMEOUT_INIT.
2413 timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
2414 if (tp->packets_out == 1)
2415 timeout = max_t(u32, timeout,
2416 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2417 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2419 /* If RTO is shorter, just schedule TLP in its place. */
2420 tlp_time_stamp = tcp_jiffies32 + timeout;
2421 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2422 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2423 s32 delta = rto_time_stamp - tcp_jiffies32;
2428 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2433 /* Thanks to skb fast clones, we can detect if a prior transmit of
2434 * a packet is still in a qdisc or driver queue.
2435 * In this case, there is very little point doing a retransmit !
2437 static bool skb_still_in_host_queue(const struct sock *sk,
2438 const struct sk_buff *skb)
2440 if (unlikely(skb_fclone_busy(sk, skb))) {
2441 NET_INC_STATS(sock_net(sk),
2442 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2448 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2449 * retransmit the last segment.
2451 void tcp_send_loss_probe(struct sock *sk)
2453 struct tcp_sock *tp = tcp_sk(sk);
2454 struct sk_buff *skb;
2456 int mss = tcp_current_mss(sk);
2458 skb = tcp_send_head(sk);
2460 if (tcp_snd_wnd_test(tp, skb, mss)) {
2461 pcount = tp->packets_out;
2462 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2463 if (tp->packets_out > pcount)
2467 skb = tcp_write_queue_prev(sk, skb);
2469 skb = tcp_write_queue_tail(sk);
2472 /* At most one outstanding TLP retransmission. */
2473 if (tp->tlp_high_seq)
2476 /* Retransmit last segment. */
2480 if (skb_still_in_host_queue(sk, skb))
2483 pcount = tcp_skb_pcount(skb);
2484 if (WARN_ON(!pcount))
2487 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2488 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2491 skb = tcp_write_queue_next(sk, skb);
2494 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2497 if (__tcp_retransmit_skb(sk, skb, 1))
2500 /* Record snd_nxt for loss detection. */
2501 tp->tlp_high_seq = tp->snd_nxt;
2504 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2505 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2506 inet_csk(sk)->icsk_pending = 0;
2511 /* Push out any pending frames which were held back due to
2512 * TCP_CORK or attempt at coalescing tiny packets.
2513 * The socket must be locked by the caller.
2515 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2518 /* If we are closed, the bytes will have to remain here.
2519 * In time closedown will finish, we empty the write queue and
2520 * all will be happy.
2522 if (unlikely(sk->sk_state == TCP_CLOSE))
2525 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2526 sk_gfp_mask(sk, GFP_ATOMIC)))
2527 tcp_check_probe_timer(sk);
2530 /* Send _single_ skb sitting at the send head. This function requires
2531 * true push pending frames to setup probe timer etc.
2533 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2535 struct sk_buff *skb = tcp_send_head(sk);
2537 BUG_ON(!skb || skb->len < mss_now);
2539 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2542 /* This function returns the amount that we can raise the
2543 * usable window based on the following constraints
2545 * 1. The window can never be shrunk once it is offered (RFC 793)
2546 * 2. We limit memory per socket
2549 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2550 * RECV.NEXT + RCV.WIN fixed until:
2551 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2553 * i.e. don't raise the right edge of the window until you can raise
2554 * it at least MSS bytes.
2556 * Unfortunately, the recommended algorithm breaks header prediction,
2557 * since header prediction assumes th->window stays fixed.
2559 * Strictly speaking, keeping th->window fixed violates the receiver
2560 * side SWS prevention criteria. The problem is that under this rule
2561 * a stream of single byte packets will cause the right side of the
2562 * window to always advance by a single byte.
2564 * Of course, if the sender implements sender side SWS prevention
2565 * then this will not be a problem.
2567 * BSD seems to make the following compromise:
2569 * If the free space is less than the 1/4 of the maximum
2570 * space available and the free space is less than 1/2 mss,
2571 * then set the window to 0.
2572 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2573 * Otherwise, just prevent the window from shrinking
2574 * and from being larger than the largest representable value.
2576 * This prevents incremental opening of the window in the regime
2577 * where TCP is limited by the speed of the reader side taking
2578 * data out of the TCP receive queue. It does nothing about
2579 * those cases where the window is constrained on the sender side
2580 * because the pipeline is full.
2582 * BSD also seems to "accidentally" limit itself to windows that are a
2583 * multiple of MSS, at least until the free space gets quite small.
2584 * This would appear to be a side effect of the mbuf implementation.
2585 * Combining these two algorithms results in the observed behavior
2586 * of having a fixed window size at almost all times.
2588 * Below we obtain similar behavior by forcing the offered window to
2589 * a multiple of the mss when it is feasible to do so.
2591 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2592 * Regular options like TIMESTAMP are taken into account.
2594 u32 __tcp_select_window(struct sock *sk)
2596 struct inet_connection_sock *icsk = inet_csk(sk);
2597 struct tcp_sock *tp = tcp_sk(sk);
2598 /* MSS for the peer's data. Previous versions used mss_clamp
2599 * here. I don't know if the value based on our guesses
2600 * of peer's MSS is better for the performance. It's more correct
2601 * but may be worse for the performance because of rcv_mss
2602 * fluctuations. --SAW 1998/11/1
2604 int mss = icsk->icsk_ack.rcv_mss;
2605 int free_space = tcp_space(sk);
2606 int allowed_space = tcp_full_space(sk);
2607 int full_space = min_t(int, tp->window_clamp, allowed_space);
2610 if (unlikely(mss > full_space)) {
2615 if (free_space < (full_space >> 1)) {
2616 icsk->icsk_ack.quick = 0;
2618 if (tcp_under_memory_pressure(sk))
2619 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2622 /* free_space might become our new window, make sure we don't
2623 * increase it due to wscale.
2625 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2627 /* if free space is less than mss estimate, or is below 1/16th
2628 * of the maximum allowed, try to move to zero-window, else
2629 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2630 * new incoming data is dropped due to memory limits.
2631 * With large window, mss test triggers way too late in order
2632 * to announce zero window in time before rmem limit kicks in.
2634 if (free_space < (allowed_space >> 4) || free_space < mss)
2638 if (free_space > tp->rcv_ssthresh)
2639 free_space = tp->rcv_ssthresh;
2641 /* Don't do rounding if we are using window scaling, since the
2642 * scaled window will not line up with the MSS boundary anyway.
2644 if (tp->rx_opt.rcv_wscale) {
2645 window = free_space;
2647 /* Advertise enough space so that it won't get scaled away.
2648 * Import case: prevent zero window announcement if
2649 * 1<<rcv_wscale > mss.
2651 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2653 window = tp->rcv_wnd;
2654 /* Get the largest window that is a nice multiple of mss.
2655 * Window clamp already applied above.
2656 * If our current window offering is within 1 mss of the
2657 * free space we just keep it. This prevents the divide
2658 * and multiply from happening most of the time.
2659 * We also don't do any window rounding when the free space
2662 if (window <= free_space - mss || window > free_space)
2663 window = rounddown(free_space, mss);
2664 else if (mss == full_space &&
2665 free_space > window + (full_space >> 1))
2666 window = free_space;
2672 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2673 const struct sk_buff *next_skb)
2675 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2676 const struct skb_shared_info *next_shinfo =
2677 skb_shinfo(next_skb);
2678 struct skb_shared_info *shinfo = skb_shinfo(skb);
2680 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2681 shinfo->tskey = next_shinfo->tskey;
2682 TCP_SKB_CB(skb)->txstamp_ack |=
2683 TCP_SKB_CB(next_skb)->txstamp_ack;
2687 /* Collapses two adjacent SKB's during retransmission. */
2688 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2690 struct tcp_sock *tp = tcp_sk(sk);
2691 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2692 int skb_size, next_skb_size;
2694 skb_size = skb->len;
2695 next_skb_size = next_skb->len;
2697 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2699 if (next_skb_size) {
2700 if (next_skb_size <= skb_availroom(skb))
2701 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2703 else if (!skb_shift(skb, next_skb, next_skb_size))
2706 tcp_highest_sack_combine(sk, next_skb, skb);
2708 tcp_unlink_write_queue(next_skb, sk);
2710 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2711 skb->ip_summed = CHECKSUM_PARTIAL;
2713 if (skb->ip_summed != CHECKSUM_PARTIAL)
2714 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2716 /* Update sequence range on original skb. */
2717 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2719 /* Merge over control information. This moves PSH/FIN etc. over */
2720 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2722 /* All done, get rid of second SKB and account for it so
2723 * packet counting does not break.
2725 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2726 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2728 /* changed transmit queue under us so clear hints */
2729 tcp_clear_retrans_hints_partial(tp);
2730 if (next_skb == tp->retransmit_skb_hint)
2731 tp->retransmit_skb_hint = skb;
2733 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2735 tcp_skb_collapse_tstamp(skb, next_skb);
2737 sk_wmem_free_skb(sk, next_skb);
2741 /* Check if coalescing SKBs is legal. */
2742 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2744 if (tcp_skb_pcount(skb) > 1)
2746 if (skb_cloned(skb))
2748 if (skb == tcp_send_head(sk))
2750 /* Some heuristics for collapsing over SACK'd could be invented */
2751 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2757 /* Collapse packets in the retransmit queue to make to create
2758 * less packets on the wire. This is only done on retransmission.
2760 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2763 struct tcp_sock *tp = tcp_sk(sk);
2764 struct sk_buff *skb = to, *tmp;
2767 if (!sysctl_tcp_retrans_collapse)
2769 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2772 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2773 if (!tcp_can_collapse(sk, skb))
2776 if (!tcp_skb_can_collapse_to(to))
2789 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2792 if (!tcp_collapse_retrans(sk, to))
2797 /* This retransmits one SKB. Policy decisions and retransmit queue
2798 * state updates are done by the caller. Returns non-zero if an
2799 * error occurred which prevented the send.
2801 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2803 struct inet_connection_sock *icsk = inet_csk(sk);
2804 struct tcp_sock *tp = tcp_sk(sk);
2805 unsigned int cur_mss;
2809 /* Inconclusive MTU probe */
2810 if (icsk->icsk_mtup.probe_size)
2811 icsk->icsk_mtup.probe_size = 0;
2813 /* Do not sent more than we queued. 1/4 is reserved for possible
2814 * copying overhead: fragmentation, tunneling, mangling etc.
2816 if (refcount_read(&sk->sk_wmem_alloc) >
2817 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2821 if (skb_still_in_host_queue(sk, skb))
2824 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2825 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2827 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2831 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2832 return -EHOSTUNREACH; /* Routing failure or similar. */
2834 cur_mss = tcp_current_mss(sk);
2836 /* If receiver has shrunk his window, and skb is out of
2837 * new window, do not retransmit it. The exception is the
2838 * case, when window is shrunk to zero. In this case
2839 * our retransmit serves as a zero window probe.
2841 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2842 TCP_SKB_CB(skb)->seq != tp->snd_una)
2845 len = cur_mss * segs;
2846 if (skb->len > len) {
2847 if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC))
2848 return -ENOMEM; /* We'll try again later. */
2850 if (skb_unclone(skb, GFP_ATOMIC))
2853 diff = tcp_skb_pcount(skb);
2854 tcp_set_skb_tso_segs(skb, cur_mss);
2855 diff -= tcp_skb_pcount(skb);
2857 tcp_adjust_pcount(sk, skb, diff);
2858 if (skb->len < cur_mss)
2859 tcp_retrans_try_collapse(sk, skb, cur_mss);
2862 /* RFC3168, section 6.1.1.1. ECN fallback */
2863 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2864 tcp_ecn_clear_syn(sk, skb);
2866 /* Update global and local TCP statistics. */
2867 segs = tcp_skb_pcount(skb);
2868 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2869 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2870 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2871 tp->total_retrans += segs;
2873 /* make sure skb->data is aligned on arches that require it
2874 * and check if ack-trimming & collapsing extended the headroom
2875 * beyond what csum_start can cover.
2877 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2878 skb_headroom(skb) >= 0xFFFF)) {
2879 struct sk_buff *nskb;
2881 skb->skb_mstamp = tp->tcp_mstamp;
2882 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2883 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2886 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2890 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2891 } else if (err != -EBUSY) {
2892 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2897 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2899 struct tcp_sock *tp = tcp_sk(sk);
2900 int err = __tcp_retransmit_skb(sk, skb, segs);
2903 #if FASTRETRANS_DEBUG > 0
2904 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2905 net_dbg_ratelimited("retrans_out leaked\n");
2908 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2909 tp->retrans_out += tcp_skb_pcount(skb);
2911 /* Save stamp of the first retransmit. */
2912 if (!tp->retrans_stamp)
2913 tp->retrans_stamp = tcp_skb_timestamp(skb);
2917 if (tp->undo_retrans < 0)
2918 tp->undo_retrans = 0;
2919 tp->undo_retrans += tcp_skb_pcount(skb);
2923 /* This gets called after a retransmit timeout, and the initially
2924 * retransmitted data is acknowledged. It tries to continue
2925 * resending the rest of the retransmit queue, until either
2926 * we've sent it all or the congestion window limit is reached.
2927 * If doing SACK, the first ACK which comes back for a timeout
2928 * based retransmit packet might feed us FACK information again.
2929 * If so, we use it to avoid unnecessarily retransmissions.
2931 void tcp_xmit_retransmit_queue(struct sock *sk)
2933 const struct inet_connection_sock *icsk = inet_csk(sk);
2934 struct tcp_sock *tp = tcp_sk(sk);
2935 struct sk_buff *skb;
2936 struct sk_buff *hole = NULL;
2940 if (!tp->packets_out)
2943 if (tp->retransmit_skb_hint) {
2944 skb = tp->retransmit_skb_hint;
2946 skb = tcp_write_queue_head(sk);
2949 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2950 tcp_for_write_queue_from(skb, sk) {
2954 if (skb == tcp_send_head(sk))
2957 if (tcp_pacing_check(sk))
2960 /* we could do better than to assign each time */
2962 tp->retransmit_skb_hint = skb;
2964 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2967 sacked = TCP_SKB_CB(skb)->sacked;
2968 /* In case tcp_shift_skb_data() have aggregated large skbs,
2969 * we need to make sure not sending too bigs TSO packets
2971 segs = min_t(int, segs, max_segs);
2973 if (tp->retrans_out >= tp->lost_out) {
2975 } else if (!(sacked & TCPCB_LOST)) {
2976 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2981 if (icsk->icsk_ca_state != TCP_CA_Loss)
2982 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2984 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2987 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2990 if (tcp_small_queue_check(sk, skb, 1))
2993 if (tcp_retransmit_skb(sk, skb, segs))
2996 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
2998 if (tcp_in_cwnd_reduction(sk))
2999 tp->prr_out += tcp_skb_pcount(skb);
3001 if (skb == tcp_write_queue_head(sk) &&
3002 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3003 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3004 inet_csk(sk)->icsk_rto,
3009 /* We allow to exceed memory limits for FIN packets to expedite
3010 * connection tear down and (memory) recovery.
3011 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3012 * or even be forced to close flow without any FIN.
3013 * In general, we want to allow one skb per socket to avoid hangs
3014 * with edge trigger epoll()
3016 void sk_forced_mem_schedule(struct sock *sk, int size)
3020 if (size <= sk->sk_forward_alloc)
3022 amt = sk_mem_pages(size);
3023 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3024 sk_memory_allocated_add(sk, amt);
3026 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3027 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3030 /* Send a FIN. The caller locks the socket for us.
3031 * We should try to send a FIN packet really hard, but eventually give up.
3033 void tcp_send_fin(struct sock *sk)
3035 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3036 struct tcp_sock *tp = tcp_sk(sk);
3038 /* Optimization, tack on the FIN if we have one skb in write queue and
3039 * this skb was not yet sent, or we are under memory pressure.
3040 * Note: in the latter case, FIN packet will be sent after a timeout,
3041 * as TCP stack thinks it has already been transmitted.
3043 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
3045 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3046 TCP_SKB_CB(tskb)->end_seq++;
3048 if (!tcp_send_head(sk)) {
3049 /* This means tskb was already sent.
3050 * Pretend we included the FIN on previous transmit.
3051 * We need to set tp->snd_nxt to the value it would have
3052 * if FIN had been sent. This is because retransmit path
3053 * does not change tp->snd_nxt.
3059 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3060 if (unlikely(!skb)) {
3065 skb_reserve(skb, MAX_TCP_HEADER);
3066 sk_forced_mem_schedule(sk, skb->truesize);
3067 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3068 tcp_init_nondata_skb(skb, tp->write_seq,
3069 TCPHDR_ACK | TCPHDR_FIN);
3070 tcp_queue_skb(sk, skb);
3072 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3075 /* We get here when a process closes a file descriptor (either due to
3076 * an explicit close() or as a byproduct of exit()'ing) and there
3077 * was unread data in the receive queue. This behavior is recommended
3078 * by RFC 2525, section 2.17. -DaveM
3080 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3082 struct sk_buff *skb;
3084 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3086 /* NOTE: No TCP options attached and we never retransmit this. */
3087 skb = alloc_skb(MAX_TCP_HEADER, priority);
3089 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3093 /* Reserve space for headers and prepare control bits. */
3094 skb_reserve(skb, MAX_TCP_HEADER);
3095 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3096 TCPHDR_ACK | TCPHDR_RST);
3097 tcp_mstamp_refresh(tcp_sk(sk));
3099 if (tcp_transmit_skb(sk, skb, 0, priority))
3100 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3103 /* Send a crossed SYN-ACK during socket establishment.
3104 * WARNING: This routine must only be called when we have already sent
3105 * a SYN packet that crossed the incoming SYN that caused this routine
3106 * to get called. If this assumption fails then the initial rcv_wnd
3107 * and rcv_wscale values will not be correct.
3109 int tcp_send_synack(struct sock *sk)
3111 struct sk_buff *skb;
3113 skb = tcp_write_queue_head(sk);
3114 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3115 pr_debug("%s: wrong queue state\n", __func__);
3118 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3119 if (skb_cloned(skb)) {
3120 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
3123 tcp_unlink_write_queue(skb, sk);
3124 __skb_header_release(nskb);
3125 __tcp_add_write_queue_head(sk, nskb);
3126 sk_wmem_free_skb(sk, skb);
3127 sk->sk_wmem_queued += nskb->truesize;
3128 sk_mem_charge(sk, nskb->truesize);
3132 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3133 tcp_ecn_send_synack(sk, skb);
3135 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3139 * tcp_make_synack - Prepare a SYN-ACK.
3140 * sk: listener socket
3141 * dst: dst entry attached to the SYNACK
3142 * req: request_sock pointer
3144 * Allocate one skb and build a SYNACK packet.
3145 * @dst is consumed : Caller should not use it again.
3147 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3148 struct request_sock *req,
3149 struct tcp_fastopen_cookie *foc,
3150 enum tcp_synack_type synack_type)
3152 struct inet_request_sock *ireq = inet_rsk(req);
3153 const struct tcp_sock *tp = tcp_sk(sk);
3154 struct tcp_md5sig_key *md5 = NULL;
3155 struct tcp_out_options opts;
3156 struct sk_buff *skb;
3157 int tcp_header_size;
3161 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3162 if (unlikely(!skb)) {
3166 /* Reserve space for headers. */
3167 skb_reserve(skb, MAX_TCP_HEADER);
3169 switch (synack_type) {
3170 case TCP_SYNACK_NORMAL:
3171 skb_set_owner_w(skb, req_to_sk(req));
3173 case TCP_SYNACK_COOKIE:
3174 /* Under synflood, we do not attach skb to a socket,
3175 * to avoid false sharing.
3178 case TCP_SYNACK_FASTOPEN:
3179 /* sk is a const pointer, because we want to express multiple
3180 * cpu might call us concurrently.
3181 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3183 skb_set_owner_w(skb, (struct sock *)sk);
3186 skb_dst_set(skb, dst);
3188 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3190 memset(&opts, 0, sizeof(opts));
3191 #ifdef CONFIG_SYN_COOKIES
3192 if (unlikely(req->cookie_ts))
3193 skb->skb_mstamp = cookie_init_timestamp(req);
3196 skb->skb_mstamp = tcp_clock_us();
3198 #ifdef CONFIG_TCP_MD5SIG
3200 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3202 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3203 tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
3206 skb_push(skb, tcp_header_size);
3207 skb_reset_transport_header(skb);
3209 th = (struct tcphdr *)skb->data;
3210 memset(th, 0, sizeof(struct tcphdr));
3213 tcp_ecn_make_synack(req, th);
3214 th->source = htons(ireq->ir_num);
3215 th->dest = ireq->ir_rmt_port;
3216 skb->mark = ireq->ir_mark;
3217 /* Setting of flags are superfluous here for callers (and ECE is
3218 * not even correctly set)
3220 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3221 TCPHDR_SYN | TCPHDR_ACK);
3223 th->seq = htonl(TCP_SKB_CB(skb)->seq);
3224 /* XXX data is queued and acked as is. No buffer/window check */
3225 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3227 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3228 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3229 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3230 th->doff = (tcp_header_size >> 2);
3231 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3233 #ifdef CONFIG_TCP_MD5SIG
3234 /* Okay, we have all we need - do the md5 hash if needed */
3236 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3237 md5, req_to_sk(req), skb);
3241 /* Do not fool tcpdump (if any), clean our debris */
3245 EXPORT_SYMBOL(tcp_make_synack);
3247 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3249 struct inet_connection_sock *icsk = inet_csk(sk);
3250 const struct tcp_congestion_ops *ca;
3251 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3253 if (ca_key == TCP_CA_UNSPEC)
3257 ca = tcp_ca_find_key(ca_key);
3258 if (likely(ca && try_module_get(ca->owner))) {
3259 module_put(icsk->icsk_ca_ops->owner);
3260 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3261 icsk->icsk_ca_ops = ca;
3266 /* Do all connect socket setups that can be done AF independent. */
3267 static void tcp_connect_init(struct sock *sk)
3269 const struct dst_entry *dst = __sk_dst_get(sk);
3270 struct tcp_sock *tp = tcp_sk(sk);
3274 /* We'll fix this up when we get a response from the other end.
3275 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3277 tp->tcp_header_len = sizeof(struct tcphdr);
3278 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3279 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3281 #ifdef CONFIG_TCP_MD5SIG
3282 if (tp->af_specific->md5_lookup(sk, sk))
3283 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3286 /* If user gave his TCP_MAXSEG, record it to clamp */
3287 if (tp->rx_opt.user_mss)
3288 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3291 tcp_sync_mss(sk, dst_mtu(dst));
3293 tcp_ca_dst_init(sk, dst);
3295 if (!tp->window_clamp)
3296 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3297 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3299 tcp_initialize_rcv_mss(sk);
3301 /* limit the window selection if the user enforce a smaller rx buffer */
3302 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3303 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3304 tp->window_clamp = tcp_full_space(sk);
3306 rcv_wnd = tcp_rwnd_init_bpf(sk);
3308 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3310 tcp_select_initial_window(tcp_full_space(sk),
3311 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3314 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3318 tp->rx_opt.rcv_wscale = rcv_wscale;
3319 tp->rcv_ssthresh = tp->rcv_wnd;
3322 sock_reset_flag(sk, SOCK_DONE);
3325 tp->snd_una = tp->write_seq;
3326 tp->snd_sml = tp->write_seq;
3327 tp->snd_up = tp->write_seq;
3328 tp->snd_nxt = tp->write_seq;
3330 if (likely(!tp->repair))
3333 tp->rcv_tstamp = tcp_jiffies32;
3334 tp->rcv_wup = tp->rcv_nxt;
3335 tp->copied_seq = tp->rcv_nxt;
3337 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3338 inet_csk(sk)->icsk_retransmits = 0;
3339 tcp_clear_retrans(tp);
3342 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3344 struct tcp_sock *tp = tcp_sk(sk);
3345 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3347 tcb->end_seq += skb->len;
3348 __skb_header_release(skb);
3349 __tcp_add_write_queue_tail(sk, skb);
3350 sk->sk_wmem_queued += skb->truesize;
3351 sk_mem_charge(sk, skb->truesize);
3352 tp->write_seq = tcb->end_seq;
3353 tp->packets_out += tcp_skb_pcount(skb);
3356 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3357 * queue a data-only packet after the regular SYN, such that regular SYNs
3358 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3359 * only the SYN sequence, the data are retransmitted in the first ACK.
3360 * If cookie is not cached or other error occurs, falls back to send a
3361 * regular SYN with Fast Open cookie request option.
3363 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3365 struct tcp_sock *tp = tcp_sk(sk);
3366 struct tcp_fastopen_request *fo = tp->fastopen_req;
3368 struct sk_buff *syn_data;
3370 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3371 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3374 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3375 * user-MSS. Reserve maximum option space for middleboxes that add
3376 * private TCP options. The cost is reduced data space in SYN :(
3378 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3380 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3381 MAX_TCP_OPTION_SPACE;
3383 space = min_t(size_t, space, fo->size);
3385 /* limit to order-0 allocations */
3386 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3388 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3391 syn_data->ip_summed = CHECKSUM_PARTIAL;
3392 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3394 int copied = copy_from_iter(skb_put(syn_data, space), space,
3395 &fo->data->msg_iter);
3396 if (unlikely(!copied)) {
3397 kfree_skb(syn_data);
3400 if (copied != space) {
3401 skb_trim(syn_data, copied);
3405 /* No more data pending in inet_wait_for_connect() */
3406 if (space == fo->size)
3410 tcp_connect_queue_skb(sk, syn_data);
3412 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3414 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3416 syn->skb_mstamp = syn_data->skb_mstamp;
3418 /* Now full SYN+DATA was cloned and sent (or not),
3419 * remove the SYN from the original skb (syn_data)
3420 * we keep in write queue in case of a retransmit, as we
3421 * also have the SYN packet (with no data) in the same queue.
3423 TCP_SKB_CB(syn_data)->seq++;
3424 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3426 tp->syn_data = (fo->copied > 0);
3427 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3432 /* Send a regular SYN with Fast Open cookie request option */
3433 if (fo->cookie.len > 0)
3435 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3437 tp->syn_fastopen = 0;
3439 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3443 /* Build a SYN and send it off. */
3444 int tcp_connect(struct sock *sk)
3446 struct tcp_sock *tp = tcp_sk(sk);
3447 struct sk_buff *buff;
3450 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB);
3451 tcp_connect_init(sk);
3453 if (unlikely(tp->repair)) {
3454 tcp_finish_connect(sk, NULL);
3458 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3459 if (unlikely(!buff))
3462 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3463 tcp_mstamp_refresh(tp);
3464 tp->retrans_stamp = tcp_time_stamp(tp);
3465 tcp_connect_queue_skb(sk, buff);
3466 tcp_ecn_send_syn(sk, buff);
3468 /* Send off SYN; include data in Fast Open. */
3469 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3470 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3471 if (err == -ECONNREFUSED)
3474 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3475 * in order to make this packet get counted in tcpOutSegs.
3477 tp->snd_nxt = tp->write_seq;
3478 tp->pushed_seq = tp->write_seq;
3479 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3481 /* Timer for repeating the SYN until an answer. */
3482 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3483 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3486 EXPORT_SYMBOL(tcp_connect);
3488 /* Send out a delayed ack, the caller does the policy checking
3489 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3492 void tcp_send_delayed_ack(struct sock *sk)
3494 struct inet_connection_sock *icsk = inet_csk(sk);
3495 int ato = icsk->icsk_ack.ato;
3496 unsigned long timeout;
3498 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3500 if (ato > TCP_DELACK_MIN) {
3501 const struct tcp_sock *tp = tcp_sk(sk);
3502 int max_ato = HZ / 2;
3504 if (icsk->icsk_ack.pingpong ||
3505 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3506 max_ato = TCP_DELACK_MAX;
3508 /* Slow path, intersegment interval is "high". */
3510 /* If some rtt estimate is known, use it to bound delayed ack.
3511 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3515 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3522 ato = min(ato, max_ato);
3525 /* Stay within the limit we were given */
3526 timeout = jiffies + ato;
3528 /* Use new timeout only if there wasn't a older one earlier. */
3529 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3530 /* If delack timer was blocked or is about to expire,
3533 if (icsk->icsk_ack.blocked ||
3534 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3539 if (!time_before(timeout, icsk->icsk_ack.timeout))
3540 timeout = icsk->icsk_ack.timeout;
3542 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3543 icsk->icsk_ack.timeout = timeout;
3544 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3547 /* This routine sends an ack and also updates the window. */
3548 void tcp_send_ack(struct sock *sk)
3550 struct sk_buff *buff;
3552 /* If we have been reset, we may not send again. */
3553 if (sk->sk_state == TCP_CLOSE)
3556 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3558 /* We are not putting this on the write queue, so
3559 * tcp_transmit_skb() will set the ownership to this
3562 buff = alloc_skb(MAX_TCP_HEADER,
3563 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3564 if (unlikely(!buff)) {
3565 inet_csk_schedule_ack(sk);
3566 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3567 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3568 TCP_DELACK_MAX, TCP_RTO_MAX);
3572 /* Reserve space for headers and prepare control bits. */
3573 skb_reserve(buff, MAX_TCP_HEADER);
3574 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3576 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3578 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3580 skb_set_tcp_pure_ack(buff);
3582 /* Send it off, this clears delayed acks for us. */
3583 tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3585 EXPORT_SYMBOL_GPL(tcp_send_ack);
3587 /* This routine sends a packet with an out of date sequence
3588 * number. It assumes the other end will try to ack it.
3590 * Question: what should we make while urgent mode?
3591 * 4.4BSD forces sending single byte of data. We cannot send
3592 * out of window data, because we have SND.NXT==SND.MAX...
3594 * Current solution: to send TWO zero-length segments in urgent mode:
3595 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3596 * out-of-date with SND.UNA-1 to probe window.
3598 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3600 struct tcp_sock *tp = tcp_sk(sk);
3601 struct sk_buff *skb;
3603 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3604 skb = alloc_skb(MAX_TCP_HEADER,
3605 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3609 /* Reserve space for headers and set control bits. */
3610 skb_reserve(skb, MAX_TCP_HEADER);
3611 /* Use a previous sequence. This should cause the other
3612 * end to send an ack. Don't queue or clone SKB, just
3615 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3616 NET_INC_STATS(sock_net(sk), mib);
3617 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3620 /* Called from setsockopt( ... TCP_REPAIR ) */
3621 void tcp_send_window_probe(struct sock *sk)
3623 if (sk->sk_state == TCP_ESTABLISHED) {
3624 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3625 tcp_mstamp_refresh(tcp_sk(sk));
3626 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3630 /* Initiate keepalive or window probe from timer. */
3631 int tcp_write_wakeup(struct sock *sk, int mib)
3633 struct tcp_sock *tp = tcp_sk(sk);
3634 struct sk_buff *skb;
3636 if (sk->sk_state == TCP_CLOSE)
3639 skb = tcp_send_head(sk);
3640 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3642 unsigned int mss = tcp_current_mss(sk);
3643 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3645 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3646 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3648 /* We are probing the opening of a window
3649 * but the window size is != 0
3650 * must have been a result SWS avoidance ( sender )
3652 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3654 seg_size = min(seg_size, mss);
3655 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3656 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3658 } else if (!tcp_skb_pcount(skb))
3659 tcp_set_skb_tso_segs(skb, mss);
3661 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3662 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3664 tcp_event_new_data_sent(sk, skb);
3667 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3668 tcp_xmit_probe_skb(sk, 1, mib);
3669 return tcp_xmit_probe_skb(sk, 0, mib);
3673 /* A window probe timeout has occurred. If window is not closed send
3674 * a partial packet else a zero probe.
3676 void tcp_send_probe0(struct sock *sk)
3678 struct inet_connection_sock *icsk = inet_csk(sk);
3679 struct tcp_sock *tp = tcp_sk(sk);
3680 struct net *net = sock_net(sk);
3681 unsigned long probe_max;
3684 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3686 if (tp->packets_out || !tcp_send_head(sk)) {
3687 /* Cancel probe timer, if it is not required. */
3688 icsk->icsk_probes_out = 0;
3689 icsk->icsk_backoff = 0;
3694 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3695 icsk->icsk_backoff++;
3696 icsk->icsk_probes_out++;
3697 probe_max = TCP_RTO_MAX;
3699 /* If packet was not sent due to local congestion,
3700 * do not backoff and do not remember icsk_probes_out.
3701 * Let local senders to fight for local resources.
3703 * Use accumulated backoff yet.
3705 if (!icsk->icsk_probes_out)
3706 icsk->icsk_probes_out = 1;
3707 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3709 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3710 tcp_probe0_when(sk, probe_max),
3714 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3716 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3720 tcp_rsk(req)->txhash = net_tx_rndhash();
3721 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3723 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3724 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3725 if (unlikely(tcp_passive_fastopen(sk)))
3726 tcp_sk(sk)->total_retrans++;
3730 EXPORT_SYMBOL(tcp_rtx_synack);