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_time_stamp;
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_time_stamp;
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))
323 /* Packet ECN state for a SYN. */
324 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
326 struct tcp_sock *tp = tcp_sk(sk);
327 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
328 tcp_ca_needs_ecn(sk);
331 const struct dst_entry *dst = __sk_dst_get(sk);
333 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
340 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
341 tp->ecn_flags = TCP_ECN_OK;
342 if (tcp_ca_needs_ecn(sk))
347 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
349 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
350 /* tp->ecn_flags are cleared at a later point in time when
351 * SYN ACK is ultimatively being received.
353 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
357 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
359 if (inet_rsk(req)->ecn_ok)
363 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
366 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
367 struct tcphdr *th, int tcp_header_len)
369 struct tcp_sock *tp = tcp_sk(sk);
371 if (tp->ecn_flags & TCP_ECN_OK) {
372 /* Not-retransmitted data segment: set ECT and inject CWR. */
373 if (skb->len != tcp_header_len &&
374 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
376 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
377 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
379 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
381 } else if (!tcp_ca_needs_ecn(sk)) {
382 /* ACK or retransmitted segment: clear ECT|CE */
383 INET_ECN_dontxmit(sk);
385 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
390 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
391 * auto increment end seqno.
393 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
395 skb->ip_summed = CHECKSUM_PARTIAL;
398 TCP_SKB_CB(skb)->tcp_flags = flags;
399 TCP_SKB_CB(skb)->sacked = 0;
401 tcp_skb_pcount_set(skb, 1);
403 TCP_SKB_CB(skb)->seq = seq;
404 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
406 TCP_SKB_CB(skb)->end_seq = seq;
409 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
411 return tp->snd_una != tp->snd_up;
414 #define OPTION_SACK_ADVERTISE (1 << 0)
415 #define OPTION_TS (1 << 1)
416 #define OPTION_MD5 (1 << 2)
417 #define OPTION_WSCALE (1 << 3)
418 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
420 struct tcp_out_options {
421 u16 options; /* bit field of OPTION_* */
422 u16 mss; /* 0 to disable */
423 u8 ws; /* window scale, 0 to disable */
424 u8 num_sack_blocks; /* number of SACK blocks to include */
425 u8 hash_size; /* bytes in hash_location */
426 __u8 *hash_location; /* temporary pointer, overloaded */
427 __u32 tsval, tsecr; /* need to include OPTION_TS */
428 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
431 /* Write previously computed TCP options to the packet.
433 * Beware: Something in the Internet is very sensitive to the ordering of
434 * TCP options, we learned this through the hard way, so be careful here.
435 * Luckily we can at least blame others for their non-compliance but from
436 * inter-operability perspective it seems that we're somewhat stuck with
437 * the ordering which we have been using if we want to keep working with
438 * those broken things (not that it currently hurts anybody as there isn't
439 * particular reason why the ordering would need to be changed).
441 * At least SACK_PERM as the first option is known to lead to a disaster
442 * (but it may well be that other scenarios fail similarly).
444 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
445 struct tcp_out_options *opts)
447 u16 options = opts->options; /* mungable copy */
449 if (unlikely(OPTION_MD5 & options)) {
450 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
451 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
452 /* overload cookie hash location */
453 opts->hash_location = (__u8 *)ptr;
457 if (unlikely(opts->mss)) {
458 *ptr++ = htonl((TCPOPT_MSS << 24) |
459 (TCPOLEN_MSS << 16) |
463 if (likely(OPTION_TS & options)) {
464 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
465 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
466 (TCPOLEN_SACK_PERM << 16) |
467 (TCPOPT_TIMESTAMP << 8) |
469 options &= ~OPTION_SACK_ADVERTISE;
471 *ptr++ = htonl((TCPOPT_NOP << 24) |
473 (TCPOPT_TIMESTAMP << 8) |
476 *ptr++ = htonl(opts->tsval);
477 *ptr++ = htonl(opts->tsecr);
480 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
481 *ptr++ = htonl((TCPOPT_NOP << 24) |
483 (TCPOPT_SACK_PERM << 8) |
487 if (unlikely(OPTION_WSCALE & options)) {
488 *ptr++ = htonl((TCPOPT_NOP << 24) |
489 (TCPOPT_WINDOW << 16) |
490 (TCPOLEN_WINDOW << 8) |
494 if (unlikely(opts->num_sack_blocks)) {
495 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
496 tp->duplicate_sack : tp->selective_acks;
499 *ptr++ = htonl((TCPOPT_NOP << 24) |
502 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
503 TCPOLEN_SACK_PERBLOCK)));
505 for (this_sack = 0; this_sack < opts->num_sack_blocks;
507 *ptr++ = htonl(sp[this_sack].start_seq);
508 *ptr++ = htonl(sp[this_sack].end_seq);
511 tp->rx_opt.dsack = 0;
514 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
515 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
517 u32 len; /* Fast Open option length */
520 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
521 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
522 TCPOPT_FASTOPEN_MAGIC);
523 p += TCPOLEN_EXP_FASTOPEN_BASE;
525 len = TCPOLEN_FASTOPEN_BASE + foc->len;
526 *p++ = TCPOPT_FASTOPEN;
530 memcpy(p, foc->val, foc->len);
531 if ((len & 3) == 2) {
532 p[foc->len] = TCPOPT_NOP;
533 p[foc->len + 1] = TCPOPT_NOP;
535 ptr += (len + 3) >> 2;
539 /* Compute TCP options for SYN packets. This is not the final
540 * network wire format yet.
542 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
543 struct tcp_out_options *opts,
544 struct tcp_md5sig_key **md5)
546 struct tcp_sock *tp = tcp_sk(sk);
547 unsigned int remaining = MAX_TCP_OPTION_SPACE;
548 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
550 #ifdef CONFIG_TCP_MD5SIG
551 *md5 = tp->af_specific->md5_lookup(sk, sk);
553 opts->options |= OPTION_MD5;
554 remaining -= TCPOLEN_MD5SIG_ALIGNED;
560 /* We always get an MSS option. The option bytes which will be seen in
561 * normal data packets should timestamps be used, must be in the MSS
562 * advertised. But we subtract them from tp->mss_cache so that
563 * calculations in tcp_sendmsg are simpler etc. So account for this
564 * fact here if necessary. If we don't do this correctly, as a
565 * receiver we won't recognize data packets as being full sized when we
566 * should, and thus we won't abide by the delayed ACK rules correctly.
567 * SACKs don't matter, we never delay an ACK when we have any of those
569 opts->mss = tcp_advertise_mss(sk);
570 remaining -= TCPOLEN_MSS_ALIGNED;
572 if (likely(sysctl_tcp_timestamps && !*md5)) {
573 opts->options |= OPTION_TS;
574 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
575 opts->tsecr = tp->rx_opt.ts_recent;
576 remaining -= TCPOLEN_TSTAMP_ALIGNED;
578 if (likely(sysctl_tcp_window_scaling)) {
579 opts->ws = tp->rx_opt.rcv_wscale;
580 opts->options |= OPTION_WSCALE;
581 remaining -= TCPOLEN_WSCALE_ALIGNED;
583 if (likely(sysctl_tcp_sack)) {
584 opts->options |= OPTION_SACK_ADVERTISE;
585 if (unlikely(!(OPTION_TS & opts->options)))
586 remaining -= TCPOLEN_SACKPERM_ALIGNED;
589 if (fastopen && fastopen->cookie.len >= 0) {
590 u32 need = fastopen->cookie.len;
592 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
593 TCPOLEN_FASTOPEN_BASE;
594 need = (need + 3) & ~3U; /* Align to 32 bits */
595 if (remaining >= need) {
596 opts->options |= OPTION_FAST_OPEN_COOKIE;
597 opts->fastopen_cookie = &fastopen->cookie;
599 tp->syn_fastopen = 1;
600 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
604 return MAX_TCP_OPTION_SPACE - remaining;
607 /* Set up TCP options for SYN-ACKs. */
608 static unsigned int tcp_synack_options(struct request_sock *req,
609 unsigned int mss, struct sk_buff *skb,
610 struct tcp_out_options *opts,
611 const struct tcp_md5sig_key *md5,
612 struct tcp_fastopen_cookie *foc)
614 struct inet_request_sock *ireq = inet_rsk(req);
615 unsigned int remaining = MAX_TCP_OPTION_SPACE;
617 #ifdef CONFIG_TCP_MD5SIG
619 opts->options |= OPTION_MD5;
620 remaining -= TCPOLEN_MD5SIG_ALIGNED;
622 /* We can't fit any SACK blocks in a packet with MD5 + TS
623 * options. There was discussion about disabling SACK
624 * rather than TS in order to fit in better with old,
625 * buggy kernels, but that was deemed to be unnecessary.
627 ireq->tstamp_ok &= !ireq->sack_ok;
631 /* We always send an MSS option. */
633 remaining -= TCPOLEN_MSS_ALIGNED;
635 if (likely(ireq->wscale_ok)) {
636 opts->ws = ireq->rcv_wscale;
637 opts->options |= OPTION_WSCALE;
638 remaining -= TCPOLEN_WSCALE_ALIGNED;
640 if (likely(ireq->tstamp_ok)) {
641 opts->options |= OPTION_TS;
642 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
643 opts->tsecr = req->ts_recent;
644 remaining -= TCPOLEN_TSTAMP_ALIGNED;
646 if (likely(ireq->sack_ok)) {
647 opts->options |= OPTION_SACK_ADVERTISE;
648 if (unlikely(!ireq->tstamp_ok))
649 remaining -= TCPOLEN_SACKPERM_ALIGNED;
651 if (foc != NULL && foc->len >= 0) {
654 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
655 TCPOLEN_FASTOPEN_BASE;
656 need = (need + 3) & ~3U; /* Align to 32 bits */
657 if (remaining >= need) {
658 opts->options |= OPTION_FAST_OPEN_COOKIE;
659 opts->fastopen_cookie = foc;
664 return MAX_TCP_OPTION_SPACE - remaining;
667 /* Compute TCP options for ESTABLISHED sockets. This is not the
668 * final wire format yet.
670 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
671 struct tcp_out_options *opts,
672 struct tcp_md5sig_key **md5)
674 struct tcp_sock *tp = tcp_sk(sk);
675 unsigned int size = 0;
676 unsigned int eff_sacks;
680 #ifdef CONFIG_TCP_MD5SIG
681 *md5 = tp->af_specific->md5_lookup(sk, sk);
682 if (unlikely(*md5)) {
683 opts->options |= OPTION_MD5;
684 size += TCPOLEN_MD5SIG_ALIGNED;
690 if (likely(tp->rx_opt.tstamp_ok)) {
691 opts->options |= OPTION_TS;
692 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
693 opts->tsecr = tp->rx_opt.ts_recent;
694 size += TCPOLEN_TSTAMP_ALIGNED;
697 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
698 if (unlikely(eff_sacks)) {
699 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
700 opts->num_sack_blocks =
701 min_t(unsigned int, eff_sacks,
702 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
703 TCPOLEN_SACK_PERBLOCK);
704 size += TCPOLEN_SACK_BASE_ALIGNED +
705 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
712 /* TCP SMALL QUEUES (TSQ)
714 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
715 * to reduce RTT and bufferbloat.
716 * We do this using a special skb destructor (tcp_wfree).
718 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
719 * needs to be reallocated in a driver.
720 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
722 * Since transmit from skb destructor is forbidden, we use a tasklet
723 * to process all sockets that eventually need to send more skbs.
724 * We use one tasklet per cpu, with its own queue of sockets.
727 struct tasklet_struct tasklet;
728 struct list_head head; /* queue of tcp sockets */
730 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
732 static void tcp_tsq_handler(struct sock *sk)
734 if ((1 << sk->sk_state) &
735 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
736 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
737 struct tcp_sock *tp = tcp_sk(sk);
739 if (tp->lost_out > tp->retrans_out &&
740 tp->snd_cwnd > tcp_packets_in_flight(tp))
741 tcp_xmit_retransmit_queue(sk);
743 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
748 * One tasklet per cpu tries to send more skbs.
749 * We run in tasklet context but need to disable irqs when
750 * transferring tsq->head because tcp_wfree() might
751 * interrupt us (non NAPI drivers)
753 static void tcp_tasklet_func(unsigned long data)
755 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
758 struct list_head *q, *n;
762 local_irq_save(flags);
763 list_splice_init(&tsq->head, &list);
764 local_irq_restore(flags);
766 list_for_each_safe(q, n, &list) {
767 tp = list_entry(q, struct tcp_sock, tsq_node);
768 list_del(&tp->tsq_node);
770 sk = (struct sock *)tp;
771 smp_mb__before_atomic();
772 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
774 if (!sk->sk_lock.owned &&
775 test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) {
777 if (!sock_owned_by_user(sk)) {
778 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
788 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
789 TCPF_WRITE_TIMER_DEFERRED | \
790 TCPF_DELACK_TIMER_DEFERRED | \
791 TCPF_MTU_REDUCED_DEFERRED)
793 * tcp_release_cb - tcp release_sock() callback
796 * called from release_sock() to perform protocol dependent
797 * actions before socket release.
799 void tcp_release_cb(struct sock *sk)
801 unsigned long flags, nflags;
803 /* perform an atomic operation only if at least one flag is set */
805 flags = sk->sk_tsq_flags;
806 if (!(flags & TCP_DEFERRED_ALL))
808 nflags = flags & ~TCP_DEFERRED_ALL;
809 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
811 if (flags & TCPF_TSQ_DEFERRED)
814 /* Here begins the tricky part :
815 * We are called from release_sock() with :
817 * 2) sk_lock.slock spinlock held
818 * 3) socket owned by us (sk->sk_lock.owned == 1)
820 * But following code is meant to be called from BH handlers,
821 * so we should keep BH disabled, but early release socket ownership
823 sock_release_ownership(sk);
825 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
826 tcp_write_timer_handler(sk);
829 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
830 tcp_delack_timer_handler(sk);
833 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
834 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
838 EXPORT_SYMBOL(tcp_release_cb);
840 void __init tcp_tasklet_init(void)
844 for_each_possible_cpu(i) {
845 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
847 INIT_LIST_HEAD(&tsq->head);
848 tasklet_init(&tsq->tasklet,
855 * Write buffer destructor automatically called from kfree_skb.
856 * We can't xmit new skbs from this context, as we might already
859 void tcp_wfree(struct sk_buff *skb)
861 struct sock *sk = skb->sk;
862 struct tcp_sock *tp = tcp_sk(sk);
863 unsigned long flags, nval, oval;
866 /* Keep one reference on sk_wmem_alloc.
867 * Will be released by sk_free() from here or tcp_tasklet_func()
869 wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
871 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
872 * Wait until our queues (qdisc + devices) are drained.
874 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
875 * - chance for incoming ACK (processed by another cpu maybe)
876 * to migrate this flow (skb->ooo_okay will be eventually set)
878 if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
881 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
882 struct tsq_tasklet *tsq;
885 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
888 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
889 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
893 /* queue this socket to tasklet queue */
894 local_irq_save(flags);
895 tsq = this_cpu_ptr(&tsq_tasklet);
896 empty = list_empty(&tsq->head);
897 list_add(&tp->tsq_node, &tsq->head);
899 tasklet_schedule(&tsq->tasklet);
900 local_irq_restore(flags);
907 /* This routine actually transmits TCP packets queued in by
908 * tcp_do_sendmsg(). This is used by both the initial
909 * transmission and possible later retransmissions.
910 * All SKB's seen here are completely headerless. It is our
911 * job to build the TCP header, and pass the packet down to
912 * IP so it can do the same plus pass the packet off to the
915 * We are working here with either a clone of the original
916 * SKB, or a fresh unique copy made by the retransmit engine.
918 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
921 const struct inet_connection_sock *icsk = inet_csk(sk);
922 struct inet_sock *inet;
924 struct tcp_skb_cb *tcb;
925 struct tcp_out_options opts;
926 unsigned int tcp_options_size, tcp_header_size;
927 struct tcp_md5sig_key *md5;
931 BUG_ON(!skb || !tcp_skb_pcount(skb));
935 skb_mstamp_get(&skb->skb_mstamp);
936 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
938 tcp_rate_skb_sent(sk, skb);
940 if (unlikely(skb_cloned(skb)))
941 skb = pskb_copy(skb, gfp_mask);
943 skb = skb_clone(skb, gfp_mask);
949 tcb = TCP_SKB_CB(skb);
950 memset(&opts, 0, sizeof(opts));
952 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
953 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
955 tcp_options_size = tcp_established_options(sk, skb, &opts,
957 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
959 /* if no packet is in qdisc/device queue, then allow XPS to select
960 * another queue. We can be called from tcp_tsq_handler()
961 * which holds one reference to sk_wmem_alloc.
963 * TODO: Ideally, in-flight pure ACK packets should not matter here.
964 * One way to get this would be to set skb->truesize = 2 on them.
966 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
968 /* If we had to use memory reserve to allocate this skb,
969 * this might cause drops if packet is looped back :
970 * Other socket might not have SOCK_MEMALLOC.
971 * Packets not looped back do not care about pfmemalloc.
975 skb_push(skb, tcp_header_size);
976 skb_reset_transport_header(skb);
980 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
981 skb_set_hash_from_sk(skb, sk);
982 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
984 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
986 /* Build TCP header and checksum it. */
987 th = (struct tcphdr *)skb->data;
988 th->source = inet->inet_sport;
989 th->dest = inet->inet_dport;
990 th->seq = htonl(tcb->seq);
991 th->ack_seq = htonl(tp->rcv_nxt);
992 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
998 /* The urg_mode check is necessary during a below snd_una win probe */
999 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1000 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1001 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1003 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1004 th->urg_ptr = htons(0xFFFF);
1009 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1010 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1011 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1012 th->window = htons(tcp_select_window(sk));
1013 tcp_ecn_send(sk, skb, th, tcp_header_size);
1015 /* RFC1323: The window in SYN & SYN/ACK segments
1018 th->window = htons(min(tp->rcv_wnd, 65535U));
1020 #ifdef CONFIG_TCP_MD5SIG
1021 /* Calculate the MD5 hash, as we have all we need now */
1023 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1024 tp->af_specific->calc_md5_hash(opts.hash_location,
1029 icsk->icsk_af_ops->send_check(sk, skb);
1031 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1032 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1034 if (skb->len != tcp_header_size) {
1035 tcp_event_data_sent(tp, sk);
1036 tp->data_segs_out += tcp_skb_pcount(skb);
1039 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1040 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1041 tcp_skb_pcount(skb));
1043 tp->segs_out += tcp_skb_pcount(skb);
1044 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1045 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1046 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1048 /* Our usage of tstamp should remain private */
1051 /* Cleanup our debris for IP stacks */
1052 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1053 sizeof(struct inet6_skb_parm)));
1055 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1057 if (likely(err <= 0))
1062 return net_xmit_eval(err);
1065 /* This routine just queues the buffer for sending.
1067 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1068 * otherwise socket can stall.
1070 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1072 struct tcp_sock *tp = tcp_sk(sk);
1074 /* Advance write_seq and place onto the write_queue. */
1075 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1076 __skb_header_release(skb);
1077 tcp_add_write_queue_tail(sk, skb);
1078 sk->sk_wmem_queued += skb->truesize;
1079 sk_mem_charge(sk, skb->truesize);
1082 /* Initialize TSO segments for a packet. */
1083 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1085 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1086 /* Avoid the costly divide in the normal
1089 tcp_skb_pcount_set(skb, 1);
1090 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1092 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1093 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1097 /* When a modification to fackets out becomes necessary, we need to check
1098 * skb is counted to fackets_out or not.
1100 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1103 struct tcp_sock *tp = tcp_sk(sk);
1105 if (!tp->sacked_out || tcp_is_reno(tp))
1108 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1109 tp->fackets_out -= decr;
1112 /* Pcount in the middle of the write queue got changed, we need to do various
1113 * tweaks to fix counters
1115 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1117 struct tcp_sock *tp = tcp_sk(sk);
1119 tp->packets_out -= decr;
1121 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1122 tp->sacked_out -= decr;
1123 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1124 tp->retrans_out -= decr;
1125 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1126 tp->lost_out -= decr;
1128 /* Reno case is special. Sigh... */
1129 if (tcp_is_reno(tp) && decr > 0)
1130 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1132 tcp_adjust_fackets_out(sk, skb, decr);
1134 if (tp->lost_skb_hint &&
1135 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1136 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1137 tp->lost_cnt_hint -= decr;
1139 tcp_verify_left_out(tp);
1142 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1144 return TCP_SKB_CB(skb)->txstamp_ack ||
1145 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1148 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1150 struct skb_shared_info *shinfo = skb_shinfo(skb);
1152 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1153 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1154 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1155 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1157 shinfo->tx_flags &= ~tsflags;
1158 shinfo2->tx_flags |= tsflags;
1159 swap(shinfo->tskey, shinfo2->tskey);
1160 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1161 TCP_SKB_CB(skb)->txstamp_ack = 0;
1165 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1167 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1168 TCP_SKB_CB(skb)->eor = 0;
1171 /* Function to create two new TCP segments. Shrinks the given segment
1172 * to the specified size and appends a new segment with the rest of the
1173 * packet to the list. This won't be called frequently, I hope.
1174 * Remember, these are still headerless SKBs at this point.
1176 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1177 unsigned int mss_now, gfp_t gfp)
1179 struct tcp_sock *tp = tcp_sk(sk);
1180 struct sk_buff *buff;
1181 int nsize, old_factor;
1185 if (WARN_ON(len > skb->len))
1188 nsize = skb_headlen(skb) - len;
1192 if (skb_unclone(skb, gfp))
1195 /* Get a new skb... force flag on. */
1196 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1198 return -ENOMEM; /* We'll just try again later. */
1200 sk->sk_wmem_queued += buff->truesize;
1201 sk_mem_charge(sk, buff->truesize);
1202 nlen = skb->len - len - nsize;
1203 buff->truesize += nlen;
1204 skb->truesize -= nlen;
1206 /* Correct the sequence numbers. */
1207 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1208 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1209 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1211 /* PSH and FIN should only be set in the second packet. */
1212 flags = TCP_SKB_CB(skb)->tcp_flags;
1213 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1214 TCP_SKB_CB(buff)->tcp_flags = flags;
1215 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1216 tcp_skb_fragment_eor(skb, buff);
1218 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1219 /* Copy and checksum data tail into the new buffer. */
1220 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1221 skb_put(buff, nsize),
1226 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1228 skb->ip_summed = CHECKSUM_PARTIAL;
1229 skb_split(skb, buff, len);
1232 buff->ip_summed = skb->ip_summed;
1234 buff->tstamp = skb->tstamp;
1235 tcp_fragment_tstamp(skb, buff);
1237 old_factor = tcp_skb_pcount(skb);
1239 /* Fix up tso_factor for both original and new SKB. */
1240 tcp_set_skb_tso_segs(skb, mss_now);
1241 tcp_set_skb_tso_segs(buff, mss_now);
1243 /* Update delivered info for the new segment */
1244 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1246 /* If this packet has been sent out already, we must
1247 * adjust the various packet counters.
1249 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1250 int diff = old_factor - tcp_skb_pcount(skb) -
1251 tcp_skb_pcount(buff);
1254 tcp_adjust_pcount(sk, skb, diff);
1257 /* Link BUFF into the send queue. */
1258 __skb_header_release(buff);
1259 tcp_insert_write_queue_after(skb, buff, sk);
1264 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1265 * eventually). The difference is that pulled data not copied, but
1266 * immediately discarded.
1268 static void __pskb_trim_head(struct sk_buff *skb, int len)
1270 struct skb_shared_info *shinfo;
1273 eat = min_t(int, len, skb_headlen(skb));
1275 __skb_pull(skb, eat);
1282 shinfo = skb_shinfo(skb);
1283 for (i = 0; i < shinfo->nr_frags; i++) {
1284 int size = skb_frag_size(&shinfo->frags[i]);
1287 skb_frag_unref(skb, i);
1290 shinfo->frags[k] = shinfo->frags[i];
1292 shinfo->frags[k].page_offset += eat;
1293 skb_frag_size_sub(&shinfo->frags[k], eat);
1299 shinfo->nr_frags = k;
1301 skb_reset_tail_pointer(skb);
1302 skb->data_len -= len;
1303 skb->len = skb->data_len;
1306 /* Remove acked data from a packet in the transmit queue. */
1307 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1309 if (skb_unclone(skb, GFP_ATOMIC))
1312 __pskb_trim_head(skb, len);
1314 TCP_SKB_CB(skb)->seq += len;
1315 skb->ip_summed = CHECKSUM_PARTIAL;
1317 skb->truesize -= len;
1318 sk->sk_wmem_queued -= len;
1319 sk_mem_uncharge(sk, len);
1320 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1322 /* Any change of skb->len requires recalculation of tso factor. */
1323 if (tcp_skb_pcount(skb) > 1)
1324 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1329 /* Calculate MSS not accounting any TCP options. */
1330 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1332 const struct tcp_sock *tp = tcp_sk(sk);
1333 const struct inet_connection_sock *icsk = inet_csk(sk);
1336 /* Calculate base mss without TCP options:
1337 It is MMS_S - sizeof(tcphdr) of rfc1122
1339 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1341 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1342 if (icsk->icsk_af_ops->net_frag_header_len) {
1343 const struct dst_entry *dst = __sk_dst_get(sk);
1345 if (dst && dst_allfrag(dst))
1346 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1349 /* Clamp it (mss_clamp does not include tcp options) */
1350 if (mss_now > tp->rx_opt.mss_clamp)
1351 mss_now = tp->rx_opt.mss_clamp;
1353 /* Now subtract optional transport overhead */
1354 mss_now -= icsk->icsk_ext_hdr_len;
1356 /* Then reserve room for full set of TCP options and 8 bytes of data */
1362 /* Calculate MSS. Not accounting for SACKs here. */
1363 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1365 /* Subtract TCP options size, not including SACKs */
1366 return __tcp_mtu_to_mss(sk, pmtu) -
1367 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1370 /* Inverse of above */
1371 int tcp_mss_to_mtu(struct sock *sk, int mss)
1373 const struct tcp_sock *tp = tcp_sk(sk);
1374 const struct inet_connection_sock *icsk = inet_csk(sk);
1378 tp->tcp_header_len +
1379 icsk->icsk_ext_hdr_len +
1380 icsk->icsk_af_ops->net_header_len;
1382 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1383 if (icsk->icsk_af_ops->net_frag_header_len) {
1384 const struct dst_entry *dst = __sk_dst_get(sk);
1386 if (dst && dst_allfrag(dst))
1387 mtu += icsk->icsk_af_ops->net_frag_header_len;
1391 EXPORT_SYMBOL(tcp_mss_to_mtu);
1393 /* MTU probing init per socket */
1394 void tcp_mtup_init(struct sock *sk)
1396 struct tcp_sock *tp = tcp_sk(sk);
1397 struct inet_connection_sock *icsk = inet_csk(sk);
1398 struct net *net = sock_net(sk);
1400 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1401 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1402 icsk->icsk_af_ops->net_header_len;
1403 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1404 icsk->icsk_mtup.probe_size = 0;
1405 if (icsk->icsk_mtup.enabled)
1406 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1408 EXPORT_SYMBOL(tcp_mtup_init);
1410 /* This function synchronize snd mss to current pmtu/exthdr set.
1412 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1413 for TCP options, but includes only bare TCP header.
1415 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1416 It is minimum of user_mss and mss received with SYN.
1417 It also does not include TCP options.
1419 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1421 tp->mss_cache is current effective sending mss, including
1422 all tcp options except for SACKs. It is evaluated,
1423 taking into account current pmtu, but never exceeds
1424 tp->rx_opt.mss_clamp.
1426 NOTE1. rfc1122 clearly states that advertised MSS
1427 DOES NOT include either tcp or ip options.
1429 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1430 are READ ONLY outside this function. --ANK (980731)
1432 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1434 struct tcp_sock *tp = tcp_sk(sk);
1435 struct inet_connection_sock *icsk = inet_csk(sk);
1438 if (icsk->icsk_mtup.search_high > pmtu)
1439 icsk->icsk_mtup.search_high = pmtu;
1441 mss_now = tcp_mtu_to_mss(sk, pmtu);
1442 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1444 /* And store cached results */
1445 icsk->icsk_pmtu_cookie = pmtu;
1446 if (icsk->icsk_mtup.enabled)
1447 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1448 tp->mss_cache = mss_now;
1452 EXPORT_SYMBOL(tcp_sync_mss);
1454 /* Compute the current effective MSS, taking SACKs and IP options,
1455 * and even PMTU discovery events into account.
1457 unsigned int tcp_current_mss(struct sock *sk)
1459 const struct tcp_sock *tp = tcp_sk(sk);
1460 const struct dst_entry *dst = __sk_dst_get(sk);
1462 unsigned int header_len;
1463 struct tcp_out_options opts;
1464 struct tcp_md5sig_key *md5;
1466 mss_now = tp->mss_cache;
1469 u32 mtu = dst_mtu(dst);
1470 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1471 mss_now = tcp_sync_mss(sk, mtu);
1474 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1475 sizeof(struct tcphdr);
1476 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1477 * some common options. If this is an odd packet (because we have SACK
1478 * blocks etc) then our calculated header_len will be different, and
1479 * we have to adjust mss_now correspondingly */
1480 if (header_len != tp->tcp_header_len) {
1481 int delta = (int) header_len - tp->tcp_header_len;
1488 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1489 * As additional protections, we do not touch cwnd in retransmission phases,
1490 * and if application hit its sndbuf limit recently.
1492 static void tcp_cwnd_application_limited(struct sock *sk)
1494 struct tcp_sock *tp = tcp_sk(sk);
1496 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1497 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1498 /* Limited by application or receiver window. */
1499 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1500 u32 win_used = max(tp->snd_cwnd_used, init_win);
1501 if (win_used < tp->snd_cwnd) {
1502 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1503 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1505 tp->snd_cwnd_used = 0;
1507 tp->snd_cwnd_stamp = tcp_time_stamp;
1510 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1512 struct tcp_sock *tp = tcp_sk(sk);
1514 /* Track the maximum number of outstanding packets in each
1515 * window, and remember whether we were cwnd-limited then.
1517 if (!before(tp->snd_una, tp->max_packets_seq) ||
1518 tp->packets_out > tp->max_packets_out) {
1519 tp->max_packets_out = tp->packets_out;
1520 tp->max_packets_seq = tp->snd_nxt;
1521 tp->is_cwnd_limited = is_cwnd_limited;
1524 if (tcp_is_cwnd_limited(sk)) {
1525 /* Network is feed fully. */
1526 tp->snd_cwnd_used = 0;
1527 tp->snd_cwnd_stamp = tcp_time_stamp;
1529 /* Network starves. */
1530 if (tp->packets_out > tp->snd_cwnd_used)
1531 tp->snd_cwnd_used = tp->packets_out;
1533 if (sysctl_tcp_slow_start_after_idle &&
1534 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1535 tcp_cwnd_application_limited(sk);
1537 /* The following conditions together indicate the starvation
1538 * is caused by insufficient sender buffer:
1539 * 1) just sent some data (see tcp_write_xmit)
1540 * 2) not cwnd limited (this else condition)
1541 * 3) no more data to send (null tcp_send_head )
1542 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1544 if (!tcp_send_head(sk) && sk->sk_socket &&
1545 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1546 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1547 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1551 /* Minshall's variant of the Nagle send check. */
1552 static bool tcp_minshall_check(const struct tcp_sock *tp)
1554 return after(tp->snd_sml, tp->snd_una) &&
1555 !after(tp->snd_sml, tp->snd_nxt);
1558 /* Update snd_sml if this skb is under mss
1559 * Note that a TSO packet might end with a sub-mss segment
1560 * The test is really :
1561 * if ((skb->len % mss) != 0)
1562 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1563 * But we can avoid doing the divide again given we already have
1564 * skb_pcount = skb->len / mss_now
1566 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1567 const struct sk_buff *skb)
1569 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1570 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1573 /* Return false, if packet can be sent now without violation Nagle's rules:
1574 * 1. It is full sized. (provided by caller in %partial bool)
1575 * 2. Or it contains FIN. (already checked by caller)
1576 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1577 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1578 * With Minshall's modification: all sent small packets are ACKed.
1580 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1584 ((nonagle & TCP_NAGLE_CORK) ||
1585 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1588 /* Return how many segs we'd like on a TSO packet,
1589 * to send one TSO packet per ms
1591 u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1596 bytes = min(sk->sk_pacing_rate >> 10,
1597 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1599 /* Goal is to send at least one packet per ms,
1600 * not one big TSO packet every 100 ms.
1601 * This preserves ACK clocking and is consistent
1602 * with tcp_tso_should_defer() heuristic.
1604 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1606 return min_t(u32, segs, sk->sk_gso_max_segs);
1608 EXPORT_SYMBOL(tcp_tso_autosize);
1610 /* Return the number of segments we want in the skb we are transmitting.
1611 * See if congestion control module wants to decide; otherwise, autosize.
1613 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1615 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1616 u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0;
1619 tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs);
1622 /* Returns the portion of skb which can be sent right away */
1623 static unsigned int tcp_mss_split_point(const struct sock *sk,
1624 const struct sk_buff *skb,
1625 unsigned int mss_now,
1626 unsigned int max_segs,
1629 const struct tcp_sock *tp = tcp_sk(sk);
1630 u32 partial, needed, window, max_len;
1632 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1633 max_len = mss_now * max_segs;
1635 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1638 needed = min(skb->len, window);
1640 if (max_len <= needed)
1643 partial = needed % mss_now;
1644 /* If last segment is not a full MSS, check if Nagle rules allow us
1645 * to include this last segment in this skb.
1646 * Otherwise, we'll split the skb at last MSS boundary
1648 if (tcp_nagle_check(partial != 0, tp, nonagle))
1649 return needed - partial;
1654 /* Can at least one segment of SKB be sent right now, according to the
1655 * congestion window rules? If so, return how many segments are allowed.
1657 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1658 const struct sk_buff *skb)
1660 u32 in_flight, cwnd, halfcwnd;
1662 /* Don't be strict about the congestion window for the final FIN. */
1663 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1664 tcp_skb_pcount(skb) == 1)
1667 in_flight = tcp_packets_in_flight(tp);
1668 cwnd = tp->snd_cwnd;
1669 if (in_flight >= cwnd)
1672 /* For better scheduling, ensure we have at least
1673 * 2 GSO packets in flight.
1675 halfcwnd = max(cwnd >> 1, 1U);
1676 return min(halfcwnd, cwnd - in_flight);
1679 /* Initialize TSO state of a skb.
1680 * This must be invoked the first time we consider transmitting
1681 * SKB onto the wire.
1683 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1685 int tso_segs = tcp_skb_pcount(skb);
1687 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1688 tcp_set_skb_tso_segs(skb, mss_now);
1689 tso_segs = tcp_skb_pcount(skb);
1695 /* Return true if the Nagle test allows this packet to be
1698 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1699 unsigned int cur_mss, int nonagle)
1701 /* Nagle rule does not apply to frames, which sit in the middle of the
1702 * write_queue (they have no chances to get new data).
1704 * This is implemented in the callers, where they modify the 'nonagle'
1705 * argument based upon the location of SKB in the send queue.
1707 if (nonagle & TCP_NAGLE_PUSH)
1710 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1711 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1714 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1720 /* Does at least the first segment of SKB fit into the send window? */
1721 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1722 const struct sk_buff *skb,
1723 unsigned int cur_mss)
1725 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1727 if (skb->len > cur_mss)
1728 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1730 return !after(end_seq, tcp_wnd_end(tp));
1733 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1734 * should be put on the wire right now. If so, it returns the number of
1735 * packets allowed by the congestion window.
1737 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1738 unsigned int cur_mss, int nonagle)
1740 const struct tcp_sock *tp = tcp_sk(sk);
1741 unsigned int cwnd_quota;
1743 tcp_init_tso_segs(skb, cur_mss);
1745 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1748 cwnd_quota = tcp_cwnd_test(tp, skb);
1749 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1755 /* Test if sending is allowed right now. */
1756 bool tcp_may_send_now(struct sock *sk)
1758 const struct tcp_sock *tp = tcp_sk(sk);
1759 struct sk_buff *skb = tcp_send_head(sk);
1762 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1763 (tcp_skb_is_last(sk, skb) ?
1764 tp->nonagle : TCP_NAGLE_PUSH));
1767 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1768 * which is put after SKB on the list. It is very much like
1769 * tcp_fragment() except that it may make several kinds of assumptions
1770 * in order to speed up the splitting operation. In particular, we
1771 * know that all the data is in scatter-gather pages, and that the
1772 * packet has never been sent out before (and thus is not cloned).
1774 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1775 unsigned int mss_now, gfp_t gfp)
1777 struct sk_buff *buff;
1778 int nlen = skb->len - len;
1781 /* All of a TSO frame must be composed of paged data. */
1782 if (skb->len != skb->data_len)
1783 return tcp_fragment(sk, skb, len, mss_now, gfp);
1785 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1786 if (unlikely(!buff))
1789 sk->sk_wmem_queued += buff->truesize;
1790 sk_mem_charge(sk, buff->truesize);
1791 buff->truesize += nlen;
1792 skb->truesize -= nlen;
1794 /* Correct the sequence numbers. */
1795 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1796 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1797 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1799 /* PSH and FIN should only be set in the second packet. */
1800 flags = TCP_SKB_CB(skb)->tcp_flags;
1801 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1802 TCP_SKB_CB(buff)->tcp_flags = flags;
1804 /* This packet was never sent out yet, so no SACK bits. */
1805 TCP_SKB_CB(buff)->sacked = 0;
1807 tcp_skb_fragment_eor(skb, buff);
1809 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1810 skb_split(skb, buff, len);
1811 tcp_fragment_tstamp(skb, buff);
1813 /* Fix up tso_factor for both original and new SKB. */
1814 tcp_set_skb_tso_segs(skb, mss_now);
1815 tcp_set_skb_tso_segs(buff, mss_now);
1817 /* Link BUFF into the send queue. */
1818 __skb_header_release(buff);
1819 tcp_insert_write_queue_after(skb, buff, sk);
1824 /* Try to defer sending, if possible, in order to minimize the amount
1825 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1827 * This algorithm is from John Heffner.
1829 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1830 bool *is_cwnd_limited, u32 max_segs)
1832 const struct inet_connection_sock *icsk = inet_csk(sk);
1833 u32 age, send_win, cong_win, limit, in_flight;
1834 struct tcp_sock *tp = tcp_sk(sk);
1835 struct skb_mstamp now;
1836 struct sk_buff *head;
1839 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1842 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1845 /* Avoid bursty behavior by allowing defer
1846 * only if the last write was recent.
1848 if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1851 in_flight = tcp_packets_in_flight(tp);
1853 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1855 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1857 /* From in_flight test above, we know that cwnd > in_flight. */
1858 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1860 limit = min(send_win, cong_win);
1862 /* If a full-sized TSO skb can be sent, do it. */
1863 if (limit >= max_segs * tp->mss_cache)
1866 /* Middle in queue won't get any more data, full sendable already? */
1867 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1870 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1872 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1874 /* If at least some fraction of a window is available,
1877 chunk /= win_divisor;
1881 /* Different approach, try not to defer past a single
1882 * ACK. Receiver should ACK every other full sized
1883 * frame, so if we have space for more than 3 frames
1886 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1890 head = tcp_write_queue_head(sk);
1891 skb_mstamp_get(&now);
1892 age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1893 /* If next ACK is likely to come too late (half srtt), do not defer */
1894 if (age < (tp->srtt_us >> 4))
1897 /* Ok, it looks like it is advisable to defer. */
1899 if (cong_win < send_win && cong_win <= skb->len)
1900 *is_cwnd_limited = true;
1908 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1910 struct inet_connection_sock *icsk = inet_csk(sk);
1911 struct tcp_sock *tp = tcp_sk(sk);
1912 struct net *net = sock_net(sk);
1916 interval = net->ipv4.sysctl_tcp_probe_interval;
1917 delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1918 if (unlikely(delta >= interval * HZ)) {
1919 int mss = tcp_current_mss(sk);
1921 /* Update current search range */
1922 icsk->icsk_mtup.probe_size = 0;
1923 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1924 sizeof(struct tcphdr) +
1925 icsk->icsk_af_ops->net_header_len;
1926 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1928 /* Update probe time stamp */
1929 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1933 /* Create a new MTU probe if we are ready.
1934 * MTU probe is regularly attempting to increase the path MTU by
1935 * deliberately sending larger packets. This discovers routing
1936 * changes resulting in larger path MTUs.
1938 * Returns 0 if we should wait to probe (no cwnd available),
1939 * 1 if a probe was sent,
1942 static int tcp_mtu_probe(struct sock *sk)
1944 struct inet_connection_sock *icsk = inet_csk(sk);
1945 struct tcp_sock *tp = tcp_sk(sk);
1946 struct sk_buff *skb, *nskb, *next;
1947 struct net *net = sock_net(sk);
1954 /* Not currently probing/verifying,
1956 * have enough cwnd, and
1957 * not SACKing (the variable headers throw things off)
1959 if (likely(!icsk->icsk_mtup.enabled ||
1960 icsk->icsk_mtup.probe_size ||
1961 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1962 tp->snd_cwnd < 11 ||
1963 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
1966 /* Use binary search for probe_size between tcp_mss_base,
1967 * and current mss_clamp. if (search_high - search_low)
1968 * smaller than a threshold, backoff from probing.
1970 mss_now = tcp_current_mss(sk);
1971 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1972 icsk->icsk_mtup.search_low) >> 1);
1973 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1974 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
1975 /* When misfortune happens, we are reprobing actively,
1976 * and then reprobe timer has expired. We stick with current
1977 * probing process by not resetting search range to its orignal.
1979 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
1980 interval < net->ipv4.sysctl_tcp_probe_threshold) {
1981 /* Check whether enough time has elaplased for
1982 * another round of probing.
1984 tcp_mtu_check_reprobe(sk);
1988 /* Have enough data in the send queue to probe? */
1989 if (tp->write_seq - tp->snd_nxt < size_needed)
1992 if (tp->snd_wnd < size_needed)
1994 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1997 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1998 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1999 if (!tcp_packets_in_flight(tp))
2005 /* We're allowed to probe. Build it now. */
2006 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2009 sk->sk_wmem_queued += nskb->truesize;
2010 sk_mem_charge(sk, nskb->truesize);
2012 skb = tcp_send_head(sk);
2014 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2015 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2016 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2017 TCP_SKB_CB(nskb)->sacked = 0;
2019 nskb->ip_summed = skb->ip_summed;
2021 tcp_insert_write_queue_before(nskb, skb, sk);
2024 tcp_for_write_queue_from_safe(skb, next, sk) {
2025 copy = min_t(int, skb->len, probe_size - len);
2026 if (nskb->ip_summed) {
2027 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2029 __wsum csum = skb_copy_and_csum_bits(skb, 0,
2030 skb_put(nskb, copy),
2032 nskb->csum = csum_block_add(nskb->csum, csum, len);
2035 if (skb->len <= copy) {
2036 /* We've eaten all the data from this skb.
2038 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2039 tcp_unlink_write_queue(skb, sk);
2040 sk_wmem_free_skb(sk, skb);
2042 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2043 ~(TCPHDR_FIN|TCPHDR_PSH);
2044 if (!skb_shinfo(skb)->nr_frags) {
2045 skb_pull(skb, copy);
2046 if (skb->ip_summed != CHECKSUM_PARTIAL)
2047 skb->csum = csum_partial(skb->data,
2050 __pskb_trim_head(skb, copy);
2051 tcp_set_skb_tso_segs(skb, mss_now);
2053 TCP_SKB_CB(skb)->seq += copy;
2058 if (len >= probe_size)
2061 tcp_init_tso_segs(nskb, nskb->len);
2063 /* We're ready to send. If this fails, the probe will
2064 * be resegmented into mss-sized pieces by tcp_write_xmit().
2066 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2067 /* Decrement cwnd here because we are sending
2068 * effectively two packets. */
2070 tcp_event_new_data_sent(sk, nskb);
2072 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2073 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2074 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2082 /* TCP Small Queues :
2083 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2084 * (These limits are doubled for retransmits)
2086 * - better RTT estimation and ACK scheduling
2089 * Alas, some drivers / subsystems require a fair amount
2090 * of queued bytes to ensure line rate.
2091 * One example is wifi aggregation (802.11 AMPDU)
2093 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2094 unsigned int factor)
2098 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2099 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2102 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2103 /* Always send the 1st or 2nd skb in write queue.
2104 * No need to wait for TX completion to call us back,
2105 * after softirq/tasklet schedule.
2106 * This helps when TX completions are delayed too much.
2108 if (skb == sk->sk_write_queue.next ||
2109 skb->prev == sk->sk_write_queue.next)
2112 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2113 /* It is possible TX completion already happened
2114 * before we set TSQ_THROTTLED, so we must
2115 * test again the condition.
2117 smp_mb__after_atomic();
2118 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2124 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2126 const u32 now = tcp_time_stamp;
2128 if (tp->chrono_type > TCP_CHRONO_UNSPEC)
2129 tp->chrono_stat[tp->chrono_type - 1] += now - tp->chrono_start;
2130 tp->chrono_start = now;
2131 tp->chrono_type = new;
2134 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2136 struct tcp_sock *tp = tcp_sk(sk);
2138 /* If there are multiple conditions worthy of tracking in a
2139 * chronograph then the highest priority enum takes precedence
2140 * over the other conditions. So that if something "more interesting"
2141 * starts happening, stop the previous chrono and start a new one.
2143 if (type > tp->chrono_type)
2144 tcp_chrono_set(tp, type);
2147 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2149 struct tcp_sock *tp = tcp_sk(sk);
2152 /* There are multiple conditions worthy of tracking in a
2153 * chronograph, so that the highest priority enum takes
2154 * precedence over the other conditions (see tcp_chrono_start).
2155 * If a condition stops, we only stop chrono tracking if
2156 * it's the "most interesting" or current chrono we are
2157 * tracking and starts busy chrono if we have pending data.
2159 if (tcp_write_queue_empty(sk))
2160 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2161 else if (type == tp->chrono_type)
2162 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2165 /* This routine writes packets to the network. It advances the
2166 * send_head. This happens as incoming acks open up the remote
2169 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2170 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2171 * account rare use of URG, this is not a big flaw.
2173 * Send at most one packet when push_one > 0. Temporarily ignore
2174 * cwnd limit to force at most one packet out when push_one == 2.
2176 * Returns true, if no segments are in flight and we have queued segments,
2177 * but cannot send anything now because of SWS or another problem.
2179 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2180 int push_one, gfp_t gfp)
2182 struct tcp_sock *tp = tcp_sk(sk);
2183 struct sk_buff *skb;
2184 unsigned int tso_segs, sent_pkts;
2187 bool is_cwnd_limited = false, is_rwnd_limited = false;
2193 /* Do MTU probing. */
2194 result = tcp_mtu_probe(sk);
2197 } else if (result > 0) {
2202 max_segs = tcp_tso_segs(sk, mss_now);
2203 while ((skb = tcp_send_head(sk))) {
2206 tso_segs = tcp_init_tso_segs(skb, mss_now);
2209 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2210 /* "skb_mstamp" is used as a start point for the retransmit timer */
2211 skb_mstamp_get(&skb->skb_mstamp);
2212 goto repair; /* Skip network transmission */
2215 cwnd_quota = tcp_cwnd_test(tp, skb);
2218 /* Force out a loss probe pkt. */
2224 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2225 is_rwnd_limited = true;
2229 if (tso_segs == 1) {
2230 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2231 (tcp_skb_is_last(sk, skb) ?
2232 nonagle : TCP_NAGLE_PUSH))))
2236 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2242 if (tso_segs > 1 && !tcp_urg_mode(tp))
2243 limit = tcp_mss_split_point(sk, skb, mss_now,
2249 if (skb->len > limit &&
2250 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2253 if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
2254 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
2255 if (tcp_small_queue_check(sk, skb, 0))
2258 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2262 /* Advance the send_head. This one is sent out.
2263 * This call will increment packets_out.
2265 tcp_event_new_data_sent(sk, skb);
2267 tcp_minshall_update(tp, mss_now, skb);
2268 sent_pkts += tcp_skb_pcount(skb);
2274 if (is_rwnd_limited)
2275 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2277 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2279 if (likely(sent_pkts)) {
2280 if (tcp_in_cwnd_reduction(sk))
2281 tp->prr_out += sent_pkts;
2283 /* Send one loss probe per tail loss episode. */
2285 tcp_schedule_loss_probe(sk);
2286 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2287 tcp_cwnd_validate(sk, is_cwnd_limited);
2290 return !tp->packets_out && tcp_send_head(sk);
2293 bool tcp_schedule_loss_probe(struct sock *sk)
2295 struct inet_connection_sock *icsk = inet_csk(sk);
2296 struct tcp_sock *tp = tcp_sk(sk);
2297 u32 timeout, tlp_time_stamp, rto_time_stamp;
2298 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2300 /* No consecutive loss probes. */
2301 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2305 /* Don't do any loss probe on a Fast Open connection before 3WHS
2308 if (tp->fastopen_rsk)
2311 /* TLP is only scheduled when next timer event is RTO. */
2312 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2315 /* Schedule a loss probe in 2*RTT for SACK capable connections
2316 * in Open state, that are either limited by cwnd or application.
2318 if ((sysctl_tcp_early_retrans != 3 && sysctl_tcp_early_retrans != 4) ||
2319 !tp->packets_out || !tcp_is_sack(tp) ||
2320 icsk->icsk_ca_state != TCP_CA_Open)
2323 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2327 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2328 * for delayed ack when there's one outstanding packet. If no RTT
2329 * sample is available then probe after TCP_TIMEOUT_INIT.
2331 timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
2332 if (tp->packets_out == 1)
2333 timeout = max_t(u32, timeout,
2334 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2335 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2337 /* If RTO is shorter, just schedule TLP in its place. */
2338 tlp_time_stamp = tcp_time_stamp + timeout;
2339 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2340 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2341 s32 delta = rto_time_stamp - tcp_time_stamp;
2346 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2351 /* Thanks to skb fast clones, we can detect if a prior transmit of
2352 * a packet is still in a qdisc or driver queue.
2353 * In this case, there is very little point doing a retransmit !
2355 static bool skb_still_in_host_queue(const struct sock *sk,
2356 const struct sk_buff *skb)
2358 if (unlikely(skb_fclone_busy(sk, skb))) {
2359 NET_INC_STATS(sock_net(sk),
2360 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2366 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2367 * retransmit the last segment.
2369 void tcp_send_loss_probe(struct sock *sk)
2371 struct tcp_sock *tp = tcp_sk(sk);
2372 struct sk_buff *skb;
2374 int mss = tcp_current_mss(sk);
2376 skb = tcp_send_head(sk);
2378 if (tcp_snd_wnd_test(tp, skb, mss)) {
2379 pcount = tp->packets_out;
2380 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2381 if (tp->packets_out > pcount)
2385 skb = tcp_write_queue_prev(sk, skb);
2387 skb = tcp_write_queue_tail(sk);
2390 /* At most one outstanding TLP retransmission. */
2391 if (tp->tlp_high_seq)
2394 /* Retransmit last segment. */
2398 if (skb_still_in_host_queue(sk, skb))
2401 pcount = tcp_skb_pcount(skb);
2402 if (WARN_ON(!pcount))
2405 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2406 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2409 skb = tcp_write_queue_next(sk, skb);
2412 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2415 if (__tcp_retransmit_skb(sk, skb, 1))
2418 /* Record snd_nxt for loss detection. */
2419 tp->tlp_high_seq = tp->snd_nxt;
2422 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2423 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2424 inet_csk(sk)->icsk_pending = 0;
2429 /* Push out any pending frames which were held back due to
2430 * TCP_CORK or attempt at coalescing tiny packets.
2431 * The socket must be locked by the caller.
2433 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2436 /* If we are closed, the bytes will have to remain here.
2437 * In time closedown will finish, we empty the write queue and
2438 * all will be happy.
2440 if (unlikely(sk->sk_state == TCP_CLOSE))
2443 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2444 sk_gfp_mask(sk, GFP_ATOMIC)))
2445 tcp_check_probe_timer(sk);
2448 /* Send _single_ skb sitting at the send head. This function requires
2449 * true push pending frames to setup probe timer etc.
2451 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2453 struct sk_buff *skb = tcp_send_head(sk);
2455 BUG_ON(!skb || skb->len < mss_now);
2457 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2460 /* This function returns the amount that we can raise the
2461 * usable window based on the following constraints
2463 * 1. The window can never be shrunk once it is offered (RFC 793)
2464 * 2. We limit memory per socket
2467 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2468 * RECV.NEXT + RCV.WIN fixed until:
2469 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2471 * i.e. don't raise the right edge of the window until you can raise
2472 * it at least MSS bytes.
2474 * Unfortunately, the recommended algorithm breaks header prediction,
2475 * since header prediction assumes th->window stays fixed.
2477 * Strictly speaking, keeping th->window fixed violates the receiver
2478 * side SWS prevention criteria. The problem is that under this rule
2479 * a stream of single byte packets will cause the right side of the
2480 * window to always advance by a single byte.
2482 * Of course, if the sender implements sender side SWS prevention
2483 * then this will not be a problem.
2485 * BSD seems to make the following compromise:
2487 * If the free space is less than the 1/4 of the maximum
2488 * space available and the free space is less than 1/2 mss,
2489 * then set the window to 0.
2490 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2491 * Otherwise, just prevent the window from shrinking
2492 * and from being larger than the largest representable value.
2494 * This prevents incremental opening of the window in the regime
2495 * where TCP is limited by the speed of the reader side taking
2496 * data out of the TCP receive queue. It does nothing about
2497 * those cases where the window is constrained on the sender side
2498 * because the pipeline is full.
2500 * BSD also seems to "accidentally" limit itself to windows that are a
2501 * multiple of MSS, at least until the free space gets quite small.
2502 * This would appear to be a side effect of the mbuf implementation.
2503 * Combining these two algorithms results in the observed behavior
2504 * of having a fixed window size at almost all times.
2506 * Below we obtain similar behavior by forcing the offered window to
2507 * a multiple of the mss when it is feasible to do so.
2509 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2510 * Regular options like TIMESTAMP are taken into account.
2512 u32 __tcp_select_window(struct sock *sk)
2514 struct inet_connection_sock *icsk = inet_csk(sk);
2515 struct tcp_sock *tp = tcp_sk(sk);
2516 /* MSS for the peer's data. Previous versions used mss_clamp
2517 * here. I don't know if the value based on our guesses
2518 * of peer's MSS is better for the performance. It's more correct
2519 * but may be worse for the performance because of rcv_mss
2520 * fluctuations. --SAW 1998/11/1
2522 int mss = icsk->icsk_ack.rcv_mss;
2523 int free_space = tcp_space(sk);
2524 int allowed_space = tcp_full_space(sk);
2525 int full_space = min_t(int, tp->window_clamp, allowed_space);
2528 if (unlikely(mss > full_space)) {
2533 if (free_space < (full_space >> 1)) {
2534 icsk->icsk_ack.quick = 0;
2536 if (tcp_under_memory_pressure(sk))
2537 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2540 /* free_space might become our new window, make sure we don't
2541 * increase it due to wscale.
2543 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2545 /* if free space is less than mss estimate, or is below 1/16th
2546 * of the maximum allowed, try to move to zero-window, else
2547 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2548 * new incoming data is dropped due to memory limits.
2549 * With large window, mss test triggers way too late in order
2550 * to announce zero window in time before rmem limit kicks in.
2552 if (free_space < (allowed_space >> 4) || free_space < mss)
2556 if (free_space > tp->rcv_ssthresh)
2557 free_space = tp->rcv_ssthresh;
2559 /* Don't do rounding if we are using window scaling, since the
2560 * scaled window will not line up with the MSS boundary anyway.
2562 if (tp->rx_opt.rcv_wscale) {
2563 window = free_space;
2565 /* Advertise enough space so that it won't get scaled away.
2566 * Import case: prevent zero window announcement if
2567 * 1<<rcv_wscale > mss.
2569 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2571 window = tp->rcv_wnd;
2572 /* Get the largest window that is a nice multiple of mss.
2573 * Window clamp already applied above.
2574 * If our current window offering is within 1 mss of the
2575 * free space we just keep it. This prevents the divide
2576 * and multiply from happening most of the time.
2577 * We also don't do any window rounding when the free space
2580 if (window <= free_space - mss || window > free_space)
2581 window = rounddown(free_space, mss);
2582 else if (mss == full_space &&
2583 free_space > window + (full_space >> 1))
2584 window = free_space;
2590 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2591 const struct sk_buff *next_skb)
2593 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2594 const struct skb_shared_info *next_shinfo =
2595 skb_shinfo(next_skb);
2596 struct skb_shared_info *shinfo = skb_shinfo(skb);
2598 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2599 shinfo->tskey = next_shinfo->tskey;
2600 TCP_SKB_CB(skb)->txstamp_ack |=
2601 TCP_SKB_CB(next_skb)->txstamp_ack;
2605 /* Collapses two adjacent SKB's during retransmission. */
2606 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2608 struct tcp_sock *tp = tcp_sk(sk);
2609 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2610 int skb_size, next_skb_size;
2612 skb_size = skb->len;
2613 next_skb_size = next_skb->len;
2615 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2617 if (next_skb_size) {
2618 if (next_skb_size <= skb_availroom(skb))
2619 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2621 else if (!skb_shift(skb, next_skb, next_skb_size))
2624 tcp_highest_sack_combine(sk, next_skb, skb);
2626 tcp_unlink_write_queue(next_skb, sk);
2628 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2629 skb->ip_summed = CHECKSUM_PARTIAL;
2631 if (skb->ip_summed != CHECKSUM_PARTIAL)
2632 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2634 /* Update sequence range on original skb. */
2635 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2637 /* Merge over control information. This moves PSH/FIN etc. over */
2638 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2640 /* All done, get rid of second SKB and account for it so
2641 * packet counting does not break.
2643 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2644 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2646 /* changed transmit queue under us so clear hints */
2647 tcp_clear_retrans_hints_partial(tp);
2648 if (next_skb == tp->retransmit_skb_hint)
2649 tp->retransmit_skb_hint = skb;
2651 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2653 tcp_skb_collapse_tstamp(skb, next_skb);
2655 sk_wmem_free_skb(sk, next_skb);
2659 /* Check if coalescing SKBs is legal. */
2660 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2662 if (tcp_skb_pcount(skb) > 1)
2664 if (skb_cloned(skb))
2666 if (skb == tcp_send_head(sk))
2668 /* Some heuristics for collapsing over SACK'd could be invented */
2669 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2675 /* Collapse packets in the retransmit queue to make to create
2676 * less packets on the wire. This is only done on retransmission.
2678 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2681 struct tcp_sock *tp = tcp_sk(sk);
2682 struct sk_buff *skb = to, *tmp;
2685 if (!sysctl_tcp_retrans_collapse)
2687 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2690 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2691 if (!tcp_can_collapse(sk, skb))
2694 if (!tcp_skb_can_collapse_to(to))
2707 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2710 if (!tcp_collapse_retrans(sk, to))
2715 /* This retransmits one SKB. Policy decisions and retransmit queue
2716 * state updates are done by the caller. Returns non-zero if an
2717 * error occurred which prevented the send.
2719 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2721 struct inet_connection_sock *icsk = inet_csk(sk);
2722 struct tcp_sock *tp = tcp_sk(sk);
2723 unsigned int cur_mss;
2727 /* Inconclusive MTU probe */
2728 if (icsk->icsk_mtup.probe_size)
2729 icsk->icsk_mtup.probe_size = 0;
2731 /* Do not sent more than we queued. 1/4 is reserved for possible
2732 * copying overhead: fragmentation, tunneling, mangling etc.
2734 if (atomic_read(&sk->sk_wmem_alloc) >
2735 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2739 if (skb_still_in_host_queue(sk, skb))
2742 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2743 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2745 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2749 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2750 return -EHOSTUNREACH; /* Routing failure or similar. */
2752 cur_mss = tcp_current_mss(sk);
2754 /* If receiver has shrunk his window, and skb is out of
2755 * new window, do not retransmit it. The exception is the
2756 * case, when window is shrunk to zero. In this case
2757 * our retransmit serves as a zero window probe.
2759 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2760 TCP_SKB_CB(skb)->seq != tp->snd_una)
2763 len = cur_mss * segs;
2764 if (skb->len > len) {
2765 if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC))
2766 return -ENOMEM; /* We'll try again later. */
2768 if (skb_unclone(skb, GFP_ATOMIC))
2771 diff = tcp_skb_pcount(skb);
2772 tcp_set_skb_tso_segs(skb, cur_mss);
2773 diff -= tcp_skb_pcount(skb);
2775 tcp_adjust_pcount(sk, skb, diff);
2776 if (skb->len < cur_mss)
2777 tcp_retrans_try_collapse(sk, skb, cur_mss);
2780 /* RFC3168, section 6.1.1.1. ECN fallback */
2781 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2782 tcp_ecn_clear_syn(sk, skb);
2784 /* Update global and local TCP statistics. */
2785 segs = tcp_skb_pcount(skb);
2786 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2787 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2788 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2789 tp->total_retrans += segs;
2791 /* make sure skb->data is aligned on arches that require it
2792 * and check if ack-trimming & collapsing extended the headroom
2793 * beyond what csum_start can cover.
2795 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2796 skb_headroom(skb) >= 0xFFFF)) {
2797 struct sk_buff *nskb;
2799 skb_mstamp_get(&skb->skb_mstamp);
2800 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2801 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2804 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2808 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2809 } else if (err != -EBUSY) {
2810 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2815 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2817 struct tcp_sock *tp = tcp_sk(sk);
2818 int err = __tcp_retransmit_skb(sk, skb, segs);
2821 #if FASTRETRANS_DEBUG > 0
2822 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2823 net_dbg_ratelimited("retrans_out leaked\n");
2826 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2827 tp->retrans_out += tcp_skb_pcount(skb);
2829 /* Save stamp of the first retransmit. */
2830 if (!tp->retrans_stamp)
2831 tp->retrans_stamp = tcp_skb_timestamp(skb);
2835 if (tp->undo_retrans < 0)
2836 tp->undo_retrans = 0;
2837 tp->undo_retrans += tcp_skb_pcount(skb);
2841 /* This gets called after a retransmit timeout, and the initially
2842 * retransmitted data is acknowledged. It tries to continue
2843 * resending the rest of the retransmit queue, until either
2844 * we've sent it all or the congestion window limit is reached.
2845 * If doing SACK, the first ACK which comes back for a timeout
2846 * based retransmit packet might feed us FACK information again.
2847 * If so, we use it to avoid unnecessarily retransmissions.
2849 void tcp_xmit_retransmit_queue(struct sock *sk)
2851 const struct inet_connection_sock *icsk = inet_csk(sk);
2852 struct tcp_sock *tp = tcp_sk(sk);
2853 struct sk_buff *skb;
2854 struct sk_buff *hole = NULL;
2858 if (!tp->packets_out)
2861 if (tp->retransmit_skb_hint) {
2862 skb = tp->retransmit_skb_hint;
2864 skb = tcp_write_queue_head(sk);
2867 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2868 tcp_for_write_queue_from(skb, sk) {
2872 if (skb == tcp_send_head(sk))
2874 /* we could do better than to assign each time */
2876 tp->retransmit_skb_hint = skb;
2878 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2881 sacked = TCP_SKB_CB(skb)->sacked;
2882 /* In case tcp_shift_skb_data() have aggregated large skbs,
2883 * we need to make sure not sending too bigs TSO packets
2885 segs = min_t(int, segs, max_segs);
2887 if (tp->retrans_out >= tp->lost_out) {
2889 } else if (!(sacked & TCPCB_LOST)) {
2890 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2895 if (icsk->icsk_ca_state != TCP_CA_Loss)
2896 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2898 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2901 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2904 if (tcp_small_queue_check(sk, skb, 1))
2907 if (tcp_retransmit_skb(sk, skb, segs))
2910 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
2912 if (tcp_in_cwnd_reduction(sk))
2913 tp->prr_out += tcp_skb_pcount(skb);
2915 if (skb == tcp_write_queue_head(sk) &&
2916 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
2917 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2918 inet_csk(sk)->icsk_rto,
2923 /* We allow to exceed memory limits for FIN packets to expedite
2924 * connection tear down and (memory) recovery.
2925 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2926 * or even be forced to close flow without any FIN.
2927 * In general, we want to allow one skb per socket to avoid hangs
2928 * with edge trigger epoll()
2930 void sk_forced_mem_schedule(struct sock *sk, int size)
2934 if (size <= sk->sk_forward_alloc)
2936 amt = sk_mem_pages(size);
2937 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2938 sk_memory_allocated_add(sk, amt);
2940 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
2941 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
2944 /* Send a FIN. The caller locks the socket for us.
2945 * We should try to send a FIN packet really hard, but eventually give up.
2947 void tcp_send_fin(struct sock *sk)
2949 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
2950 struct tcp_sock *tp = tcp_sk(sk);
2952 /* Optimization, tack on the FIN if we have one skb in write queue and
2953 * this skb was not yet sent, or we are under memory pressure.
2954 * Note: in the latter case, FIN packet will be sent after a timeout,
2955 * as TCP stack thinks it has already been transmitted.
2957 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
2959 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2960 TCP_SKB_CB(tskb)->end_seq++;
2962 if (!tcp_send_head(sk)) {
2963 /* This means tskb was already sent.
2964 * Pretend we included the FIN on previous transmit.
2965 * We need to set tp->snd_nxt to the value it would have
2966 * if FIN had been sent. This is because retransmit path
2967 * does not change tp->snd_nxt.
2973 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2974 if (unlikely(!skb)) {
2979 skb_reserve(skb, MAX_TCP_HEADER);
2980 sk_forced_mem_schedule(sk, skb->truesize);
2981 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2982 tcp_init_nondata_skb(skb, tp->write_seq,
2983 TCPHDR_ACK | TCPHDR_FIN);
2984 tcp_queue_skb(sk, skb);
2986 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
2989 /* We get here when a process closes a file descriptor (either due to
2990 * an explicit close() or as a byproduct of exit()'ing) and there
2991 * was unread data in the receive queue. This behavior is recommended
2992 * by RFC 2525, section 2.17. -DaveM
2994 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2996 struct sk_buff *skb;
2998 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3000 /* NOTE: No TCP options attached and we never retransmit this. */
3001 skb = alloc_skb(MAX_TCP_HEADER, priority);
3003 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3007 /* Reserve space for headers and prepare control bits. */
3008 skb_reserve(skb, MAX_TCP_HEADER);
3009 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3010 TCPHDR_ACK | TCPHDR_RST);
3011 skb_mstamp_get(&skb->skb_mstamp);
3013 if (tcp_transmit_skb(sk, skb, 0, priority))
3014 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3017 /* Send a crossed SYN-ACK during socket establishment.
3018 * WARNING: This routine must only be called when we have already sent
3019 * a SYN packet that crossed the incoming SYN that caused this routine
3020 * to get called. If this assumption fails then the initial rcv_wnd
3021 * and rcv_wscale values will not be correct.
3023 int tcp_send_synack(struct sock *sk)
3025 struct sk_buff *skb;
3027 skb = tcp_write_queue_head(sk);
3028 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3029 pr_debug("%s: wrong queue state\n", __func__);
3032 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3033 if (skb_cloned(skb)) {
3034 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
3037 tcp_unlink_write_queue(skb, sk);
3038 __skb_header_release(nskb);
3039 __tcp_add_write_queue_head(sk, nskb);
3040 sk_wmem_free_skb(sk, skb);
3041 sk->sk_wmem_queued += nskb->truesize;
3042 sk_mem_charge(sk, nskb->truesize);
3046 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3047 tcp_ecn_send_synack(sk, skb);
3049 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3053 * tcp_make_synack - Prepare a SYN-ACK.
3054 * sk: listener socket
3055 * dst: dst entry attached to the SYNACK
3056 * req: request_sock pointer
3058 * Allocate one skb and build a SYNACK packet.
3059 * @dst is consumed : Caller should not use it again.
3061 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3062 struct request_sock *req,
3063 struct tcp_fastopen_cookie *foc,
3064 enum tcp_synack_type synack_type)
3066 struct inet_request_sock *ireq = inet_rsk(req);
3067 const struct tcp_sock *tp = tcp_sk(sk);
3068 struct tcp_md5sig_key *md5 = NULL;
3069 struct tcp_out_options opts;
3070 struct sk_buff *skb;
3071 int tcp_header_size;
3075 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3076 if (unlikely(!skb)) {
3080 /* Reserve space for headers. */
3081 skb_reserve(skb, MAX_TCP_HEADER);
3083 switch (synack_type) {
3084 case TCP_SYNACK_NORMAL:
3085 skb_set_owner_w(skb, req_to_sk(req));
3087 case TCP_SYNACK_COOKIE:
3088 /* Under synflood, we do not attach skb to a socket,
3089 * to avoid false sharing.
3092 case TCP_SYNACK_FASTOPEN:
3093 /* sk is a const pointer, because we want to express multiple
3094 * cpu might call us concurrently.
3095 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3097 skb_set_owner_w(skb, (struct sock *)sk);
3100 skb_dst_set(skb, dst);
3102 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3104 memset(&opts, 0, sizeof(opts));
3105 #ifdef CONFIG_SYN_COOKIES
3106 if (unlikely(req->cookie_ts))
3107 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
3110 skb_mstamp_get(&skb->skb_mstamp);
3112 #ifdef CONFIG_TCP_MD5SIG
3114 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3116 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3117 tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
3120 skb_push(skb, tcp_header_size);
3121 skb_reset_transport_header(skb);
3123 th = (struct tcphdr *)skb->data;
3124 memset(th, 0, sizeof(struct tcphdr));
3127 tcp_ecn_make_synack(req, th);
3128 th->source = htons(ireq->ir_num);
3129 th->dest = ireq->ir_rmt_port;
3130 /* Setting of flags are superfluous here for callers (and ECE is
3131 * not even correctly set)
3133 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3134 TCPHDR_SYN | TCPHDR_ACK);
3136 th->seq = htonl(TCP_SKB_CB(skb)->seq);
3137 /* XXX data is queued and acked as is. No buffer/window check */
3138 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3140 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3141 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3142 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3143 th->doff = (tcp_header_size >> 2);
3144 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3146 #ifdef CONFIG_TCP_MD5SIG
3147 /* Okay, we have all we need - do the md5 hash if needed */
3149 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3150 md5, req_to_sk(req), skb);
3154 /* Do not fool tcpdump (if any), clean our debris */
3158 EXPORT_SYMBOL(tcp_make_synack);
3160 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3162 struct inet_connection_sock *icsk = inet_csk(sk);
3163 const struct tcp_congestion_ops *ca;
3164 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3166 if (ca_key == TCP_CA_UNSPEC)
3170 ca = tcp_ca_find_key(ca_key);
3171 if (likely(ca && try_module_get(ca->owner))) {
3172 module_put(icsk->icsk_ca_ops->owner);
3173 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3174 icsk->icsk_ca_ops = ca;
3179 /* Do all connect socket setups that can be done AF independent. */
3180 static void tcp_connect_init(struct sock *sk)
3182 const struct dst_entry *dst = __sk_dst_get(sk);
3183 struct tcp_sock *tp = tcp_sk(sk);
3186 /* We'll fix this up when we get a response from the other end.
3187 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3189 tp->tcp_header_len = sizeof(struct tcphdr) +
3190 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3192 #ifdef CONFIG_TCP_MD5SIG
3193 if (tp->af_specific->md5_lookup(sk, sk))
3194 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3197 /* If user gave his TCP_MAXSEG, record it to clamp */
3198 if (tp->rx_opt.user_mss)
3199 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3202 tcp_sync_mss(sk, dst_mtu(dst));
3204 tcp_ca_dst_init(sk, dst);
3206 if (!tp->window_clamp)
3207 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3208 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3210 tcp_initialize_rcv_mss(sk);
3212 /* limit the window selection if the user enforce a smaller rx buffer */
3213 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3214 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3215 tp->window_clamp = tcp_full_space(sk);
3217 tcp_select_initial_window(tcp_full_space(sk),
3218 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3221 sysctl_tcp_window_scaling,
3223 dst_metric(dst, RTAX_INITRWND));
3225 tp->rx_opt.rcv_wscale = rcv_wscale;
3226 tp->rcv_ssthresh = tp->rcv_wnd;
3229 sock_reset_flag(sk, SOCK_DONE);
3232 tp->snd_una = tp->write_seq;
3233 tp->snd_sml = tp->write_seq;
3234 tp->snd_up = tp->write_seq;
3235 tp->snd_nxt = tp->write_seq;
3237 if (likely(!tp->repair))
3240 tp->rcv_tstamp = tcp_time_stamp;
3241 tp->rcv_wup = tp->rcv_nxt;
3242 tp->copied_seq = tp->rcv_nxt;
3244 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3245 inet_csk(sk)->icsk_retransmits = 0;
3246 tcp_clear_retrans(tp);
3249 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3251 struct tcp_sock *tp = tcp_sk(sk);
3252 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3254 tcb->end_seq += skb->len;
3255 __skb_header_release(skb);
3256 __tcp_add_write_queue_tail(sk, skb);
3257 sk->sk_wmem_queued += skb->truesize;
3258 sk_mem_charge(sk, skb->truesize);
3259 tp->write_seq = tcb->end_seq;
3260 tp->packets_out += tcp_skb_pcount(skb);
3263 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3264 * queue a data-only packet after the regular SYN, such that regular SYNs
3265 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3266 * only the SYN sequence, the data are retransmitted in the first ACK.
3267 * If cookie is not cached or other error occurs, falls back to send a
3268 * regular SYN with Fast Open cookie request option.
3270 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3272 struct tcp_sock *tp = tcp_sk(sk);
3273 struct tcp_fastopen_request *fo = tp->fastopen_req;
3275 struct sk_buff *syn_data;
3277 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3278 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3281 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3282 * user-MSS. Reserve maximum option space for middleboxes that add
3283 * private TCP options. The cost is reduced data space in SYN :(
3285 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3287 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3288 MAX_TCP_OPTION_SPACE;
3290 space = min_t(size_t, space, fo->size);
3292 /* limit to order-0 allocations */
3293 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3295 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3298 syn_data->ip_summed = CHECKSUM_PARTIAL;
3299 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3301 int copied = copy_from_iter(skb_put(syn_data, space), space,
3302 &fo->data->msg_iter);
3303 if (unlikely(!copied)) {
3304 kfree_skb(syn_data);
3307 if (copied != space) {
3308 skb_trim(syn_data, copied);
3312 /* No more data pending in inet_wait_for_connect() */
3313 if (space == fo->size)
3317 tcp_connect_queue_skb(sk, syn_data);
3319 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3321 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3323 syn->skb_mstamp = syn_data->skb_mstamp;
3325 /* Now full SYN+DATA was cloned and sent (or not),
3326 * remove the SYN from the original skb (syn_data)
3327 * we keep in write queue in case of a retransmit, as we
3328 * also have the SYN packet (with no data) in the same queue.
3330 TCP_SKB_CB(syn_data)->seq++;
3331 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3333 tp->syn_data = (fo->copied > 0);
3334 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3339 /* Send a regular SYN with Fast Open cookie request option */
3340 if (fo->cookie.len > 0)
3342 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3344 tp->syn_fastopen = 0;
3346 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3350 /* Build a SYN and send it off. */
3351 int tcp_connect(struct sock *sk)
3353 struct tcp_sock *tp = tcp_sk(sk);
3354 struct sk_buff *buff;
3357 tcp_connect_init(sk);
3359 if (unlikely(tp->repair)) {
3360 tcp_finish_connect(sk, NULL);
3364 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3365 if (unlikely(!buff))
3368 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3369 tp->retrans_stamp = tcp_time_stamp;
3370 tcp_connect_queue_skb(sk, buff);
3371 tcp_ecn_send_syn(sk, buff);
3373 /* Send off SYN; include data in Fast Open. */
3374 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3375 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3376 if (err == -ECONNREFUSED)
3379 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3380 * in order to make this packet get counted in tcpOutSegs.
3382 tp->snd_nxt = tp->write_seq;
3383 tp->pushed_seq = tp->write_seq;
3384 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3386 /* Timer for repeating the SYN until an answer. */
3387 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3388 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3391 EXPORT_SYMBOL(tcp_connect);
3393 /* Send out a delayed ack, the caller does the policy checking
3394 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3397 void tcp_send_delayed_ack(struct sock *sk)
3399 struct inet_connection_sock *icsk = inet_csk(sk);
3400 int ato = icsk->icsk_ack.ato;
3401 unsigned long timeout;
3403 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3405 if (ato > TCP_DELACK_MIN) {
3406 const struct tcp_sock *tp = tcp_sk(sk);
3407 int max_ato = HZ / 2;
3409 if (icsk->icsk_ack.pingpong ||
3410 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3411 max_ato = TCP_DELACK_MAX;
3413 /* Slow path, intersegment interval is "high". */
3415 /* If some rtt estimate is known, use it to bound delayed ack.
3416 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3420 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3427 ato = min(ato, max_ato);
3430 /* Stay within the limit we were given */
3431 timeout = jiffies + ato;
3433 /* Use new timeout only if there wasn't a older one earlier. */
3434 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3435 /* If delack timer was blocked or is about to expire,
3438 if (icsk->icsk_ack.blocked ||
3439 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3444 if (!time_before(timeout, icsk->icsk_ack.timeout))
3445 timeout = icsk->icsk_ack.timeout;
3447 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3448 icsk->icsk_ack.timeout = timeout;
3449 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3452 /* This routine sends an ack and also updates the window. */
3453 void tcp_send_ack(struct sock *sk)
3455 struct sk_buff *buff;
3457 /* If we have been reset, we may not send again. */
3458 if (sk->sk_state == TCP_CLOSE)
3461 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3463 /* We are not putting this on the write queue, so
3464 * tcp_transmit_skb() will set the ownership to this
3467 buff = alloc_skb(MAX_TCP_HEADER,
3468 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3469 if (unlikely(!buff)) {
3470 inet_csk_schedule_ack(sk);
3471 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3472 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3473 TCP_DELACK_MAX, TCP_RTO_MAX);
3477 /* Reserve space for headers and prepare control bits. */
3478 skb_reserve(buff, MAX_TCP_HEADER);
3479 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3481 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3483 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3485 skb_set_tcp_pure_ack(buff);
3487 /* Send it off, this clears delayed acks for us. */
3488 skb_mstamp_get(&buff->skb_mstamp);
3489 tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3491 EXPORT_SYMBOL_GPL(tcp_send_ack);
3493 /* This routine sends a packet with an out of date sequence
3494 * number. It assumes the other end will try to ack it.
3496 * Question: what should we make while urgent mode?
3497 * 4.4BSD forces sending single byte of data. We cannot send
3498 * out of window data, because we have SND.NXT==SND.MAX...
3500 * Current solution: to send TWO zero-length segments in urgent mode:
3501 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3502 * out-of-date with SND.UNA-1 to probe window.
3504 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3506 struct tcp_sock *tp = tcp_sk(sk);
3507 struct sk_buff *skb;
3509 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3510 skb = alloc_skb(MAX_TCP_HEADER,
3511 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3515 /* Reserve space for headers and set control bits. */
3516 skb_reserve(skb, MAX_TCP_HEADER);
3517 /* Use a previous sequence. This should cause the other
3518 * end to send an ack. Don't queue or clone SKB, just
3521 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3522 skb_mstamp_get(&skb->skb_mstamp);
3523 NET_INC_STATS(sock_net(sk), mib);
3524 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3527 void tcp_send_window_probe(struct sock *sk)
3529 if (sk->sk_state == TCP_ESTABLISHED) {
3530 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3531 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3535 /* Initiate keepalive or window probe from timer. */
3536 int tcp_write_wakeup(struct sock *sk, int mib)
3538 struct tcp_sock *tp = tcp_sk(sk);
3539 struct sk_buff *skb;
3541 if (sk->sk_state == TCP_CLOSE)
3544 skb = tcp_send_head(sk);
3545 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3547 unsigned int mss = tcp_current_mss(sk);
3548 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3550 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3551 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3553 /* We are probing the opening of a window
3554 * but the window size is != 0
3555 * must have been a result SWS avoidance ( sender )
3557 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3559 seg_size = min(seg_size, mss);
3560 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3561 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3563 } else if (!tcp_skb_pcount(skb))
3564 tcp_set_skb_tso_segs(skb, mss);
3566 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3567 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3569 tcp_event_new_data_sent(sk, skb);
3572 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3573 tcp_xmit_probe_skb(sk, 1, mib);
3574 return tcp_xmit_probe_skb(sk, 0, mib);
3578 /* A window probe timeout has occurred. If window is not closed send
3579 * a partial packet else a zero probe.
3581 void tcp_send_probe0(struct sock *sk)
3583 struct inet_connection_sock *icsk = inet_csk(sk);
3584 struct tcp_sock *tp = tcp_sk(sk);
3585 struct net *net = sock_net(sk);
3586 unsigned long probe_max;
3589 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3591 if (tp->packets_out || !tcp_send_head(sk)) {
3592 /* Cancel probe timer, if it is not required. */
3593 icsk->icsk_probes_out = 0;
3594 icsk->icsk_backoff = 0;
3599 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3600 icsk->icsk_backoff++;
3601 icsk->icsk_probes_out++;
3602 probe_max = TCP_RTO_MAX;
3604 /* If packet was not sent due to local congestion,
3605 * do not backoff and do not remember icsk_probes_out.
3606 * Let local senders to fight for local resources.
3608 * Use accumulated backoff yet.
3610 if (!icsk->icsk_probes_out)
3611 icsk->icsk_probes_out = 1;
3612 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3614 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3615 tcp_probe0_when(sk, probe_max),
3619 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3621 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3625 tcp_rsk(req)->txhash = net_tx_rndhash();
3626 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3628 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3629 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3630 if (unlikely(tcp_passive_fastopen(sk)))
3631 tcp_sk(sk)->total_retrans++;
3635 EXPORT_SYMBOL(tcp_rtx_synack);