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) = (65535 << 14);
216 space = min(*window_clamp, space);
218 /* Quantize space offering to a multiple of mss if possible. */
220 space = (space / mss) * 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 * See RFC1323 for an explanation of the limit to 14
240 space = max_t(u32, space, sysctl_tcp_rmem[2]);
241 space = max_t(u32, space, sysctl_rmem_max);
242 space = min_t(u32, space, *window_clamp);
243 while (space > 65535 && (*rcv_wscale) < 14) {
249 if (mss > (1 << *rcv_wscale)) {
250 if (!init_rcv_wnd) /* Use default unless specified otherwise */
251 init_rcv_wnd = tcp_default_init_rwnd(mss);
252 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
255 /* Set the clamp no higher than max representable value */
256 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
258 EXPORT_SYMBOL(tcp_select_initial_window);
260 /* Chose a new window to advertise, update state in tcp_sock for the
261 * socket, and return result with RFC1323 scaling applied. The return
262 * value can be stuffed directly into th->window for an outgoing
265 static u16 tcp_select_window(struct sock *sk)
267 struct tcp_sock *tp = tcp_sk(sk);
268 u32 old_win = tp->rcv_wnd;
269 u32 cur_win = tcp_receive_window(tp);
270 u32 new_win = __tcp_select_window(sk);
272 /* Never shrink the offered window */
273 if (new_win < cur_win) {
274 /* Danger Will Robinson!
275 * Don't update rcv_wup/rcv_wnd here or else
276 * we will not be able to advertise a zero
277 * window in time. --DaveM
279 * Relax Will Robinson.
282 NET_INC_STATS(sock_net(sk),
283 LINUX_MIB_TCPWANTZEROWINDOWADV);
284 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
286 tp->rcv_wnd = new_win;
287 tp->rcv_wup = tp->rcv_nxt;
289 /* Make sure we do not exceed the maximum possible
292 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
293 new_win = min(new_win, MAX_TCP_WINDOW);
295 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
297 /* RFC1323 scaling applied */
298 new_win >>= tp->rx_opt.rcv_wscale;
300 /* If we advertise zero window, disable fast path. */
304 NET_INC_STATS(sock_net(sk),
305 LINUX_MIB_TCPTOZEROWINDOWADV);
306 } else if (old_win == 0) {
307 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
313 /* Packet ECN state for a SYN-ACK */
314 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
316 const struct tcp_sock *tp = tcp_sk(sk);
318 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
319 if (!(tp->ecn_flags & TCP_ECN_OK))
320 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
321 else if (tcp_ca_needs_ecn(sk))
325 /* Packet ECN state for a SYN. */
326 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
328 struct tcp_sock *tp = tcp_sk(sk);
329 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
330 tcp_ca_needs_ecn(sk);
333 const struct dst_entry *dst = __sk_dst_get(sk);
335 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
342 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
343 tp->ecn_flags = TCP_ECN_OK;
344 if (tcp_ca_needs_ecn(sk))
349 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
351 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
352 /* tp->ecn_flags are cleared at a later point in time when
353 * SYN ACK is ultimatively being received.
355 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
359 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
361 if (inet_rsk(req)->ecn_ok)
365 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
368 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
369 struct tcphdr *th, int tcp_header_len)
371 struct tcp_sock *tp = tcp_sk(sk);
373 if (tp->ecn_flags & TCP_ECN_OK) {
374 /* Not-retransmitted data segment: set ECT and inject CWR. */
375 if (skb->len != tcp_header_len &&
376 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
378 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
379 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
381 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
383 } else if (!tcp_ca_needs_ecn(sk)) {
384 /* ACK or retransmitted segment: clear ECT|CE */
385 INET_ECN_dontxmit(sk);
387 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
392 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
393 * auto increment end seqno.
395 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
397 skb->ip_summed = CHECKSUM_PARTIAL;
400 TCP_SKB_CB(skb)->tcp_flags = flags;
401 TCP_SKB_CB(skb)->sacked = 0;
403 tcp_skb_pcount_set(skb, 1);
405 TCP_SKB_CB(skb)->seq = seq;
406 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
408 TCP_SKB_CB(skb)->end_seq = seq;
411 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
413 return tp->snd_una != tp->snd_up;
416 #define OPTION_SACK_ADVERTISE (1 << 0)
417 #define OPTION_TS (1 << 1)
418 #define OPTION_MD5 (1 << 2)
419 #define OPTION_WSCALE (1 << 3)
420 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
422 struct tcp_out_options {
423 u16 options; /* bit field of OPTION_* */
424 u16 mss; /* 0 to disable */
425 u8 ws; /* window scale, 0 to disable */
426 u8 num_sack_blocks; /* number of SACK blocks to include */
427 u8 hash_size; /* bytes in hash_location */
428 __u8 *hash_location; /* temporary pointer, overloaded */
429 __u32 tsval, tsecr; /* need to include OPTION_TS */
430 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
433 /* Write previously computed TCP options to the packet.
435 * Beware: Something in the Internet is very sensitive to the ordering of
436 * TCP options, we learned this through the hard way, so be careful here.
437 * Luckily we can at least blame others for their non-compliance but from
438 * inter-operability perspective it seems that we're somewhat stuck with
439 * the ordering which we have been using if we want to keep working with
440 * those broken things (not that it currently hurts anybody as there isn't
441 * particular reason why the ordering would need to be changed).
443 * At least SACK_PERM as the first option is known to lead to a disaster
444 * (but it may well be that other scenarios fail similarly).
446 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
447 struct tcp_out_options *opts)
449 u16 options = opts->options; /* mungable copy */
451 if (unlikely(OPTION_MD5 & options)) {
452 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
453 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
454 /* overload cookie hash location */
455 opts->hash_location = (__u8 *)ptr;
459 if (unlikely(opts->mss)) {
460 *ptr++ = htonl((TCPOPT_MSS << 24) |
461 (TCPOLEN_MSS << 16) |
465 if (likely(OPTION_TS & options)) {
466 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
467 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
468 (TCPOLEN_SACK_PERM << 16) |
469 (TCPOPT_TIMESTAMP << 8) |
471 options &= ~OPTION_SACK_ADVERTISE;
473 *ptr++ = htonl((TCPOPT_NOP << 24) |
475 (TCPOPT_TIMESTAMP << 8) |
478 *ptr++ = htonl(opts->tsval);
479 *ptr++ = htonl(opts->tsecr);
482 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
483 *ptr++ = htonl((TCPOPT_NOP << 24) |
485 (TCPOPT_SACK_PERM << 8) |
489 if (unlikely(OPTION_WSCALE & options)) {
490 *ptr++ = htonl((TCPOPT_NOP << 24) |
491 (TCPOPT_WINDOW << 16) |
492 (TCPOLEN_WINDOW << 8) |
496 if (unlikely(opts->num_sack_blocks)) {
497 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
498 tp->duplicate_sack : tp->selective_acks;
501 *ptr++ = htonl((TCPOPT_NOP << 24) |
504 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
505 TCPOLEN_SACK_PERBLOCK)));
507 for (this_sack = 0; this_sack < opts->num_sack_blocks;
509 *ptr++ = htonl(sp[this_sack].start_seq);
510 *ptr++ = htonl(sp[this_sack].end_seq);
513 tp->rx_opt.dsack = 0;
516 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
517 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
519 u32 len; /* Fast Open option length */
522 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
523 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
524 TCPOPT_FASTOPEN_MAGIC);
525 p += TCPOLEN_EXP_FASTOPEN_BASE;
527 len = TCPOLEN_FASTOPEN_BASE + foc->len;
528 *p++ = TCPOPT_FASTOPEN;
532 memcpy(p, foc->val, foc->len);
533 if ((len & 3) == 2) {
534 p[foc->len] = TCPOPT_NOP;
535 p[foc->len + 1] = TCPOPT_NOP;
537 ptr += (len + 3) >> 2;
541 /* Compute TCP options for SYN packets. This is not the final
542 * network wire format yet.
544 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
545 struct tcp_out_options *opts,
546 struct tcp_md5sig_key **md5)
548 struct tcp_sock *tp = tcp_sk(sk);
549 unsigned int remaining = MAX_TCP_OPTION_SPACE;
550 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
552 #ifdef CONFIG_TCP_MD5SIG
553 *md5 = tp->af_specific->md5_lookup(sk, sk);
555 opts->options |= OPTION_MD5;
556 remaining -= TCPOLEN_MD5SIG_ALIGNED;
562 /* We always get an MSS option. The option bytes which will be seen in
563 * normal data packets should timestamps be used, must be in the MSS
564 * advertised. But we subtract them from tp->mss_cache so that
565 * calculations in tcp_sendmsg are simpler etc. So account for this
566 * fact here if necessary. If we don't do this correctly, as a
567 * receiver we won't recognize data packets as being full sized when we
568 * should, and thus we won't abide by the delayed ACK rules correctly.
569 * SACKs don't matter, we never delay an ACK when we have any of those
571 opts->mss = tcp_advertise_mss(sk);
572 remaining -= TCPOLEN_MSS_ALIGNED;
574 if (likely(sysctl_tcp_timestamps && !*md5)) {
575 opts->options |= OPTION_TS;
576 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
577 opts->tsecr = tp->rx_opt.ts_recent;
578 remaining -= TCPOLEN_TSTAMP_ALIGNED;
580 if (likely(sysctl_tcp_window_scaling)) {
581 opts->ws = tp->rx_opt.rcv_wscale;
582 opts->options |= OPTION_WSCALE;
583 remaining -= TCPOLEN_WSCALE_ALIGNED;
585 if (likely(sysctl_tcp_sack)) {
586 opts->options |= OPTION_SACK_ADVERTISE;
587 if (unlikely(!(OPTION_TS & opts->options)))
588 remaining -= TCPOLEN_SACKPERM_ALIGNED;
591 if (fastopen && fastopen->cookie.len >= 0) {
592 u32 need = fastopen->cookie.len;
594 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
595 TCPOLEN_FASTOPEN_BASE;
596 need = (need + 3) & ~3U; /* Align to 32 bits */
597 if (remaining >= need) {
598 opts->options |= OPTION_FAST_OPEN_COOKIE;
599 opts->fastopen_cookie = &fastopen->cookie;
601 tp->syn_fastopen = 1;
602 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
606 return MAX_TCP_OPTION_SPACE - remaining;
609 /* Set up TCP options for SYN-ACKs. */
610 static unsigned int tcp_synack_options(struct request_sock *req,
611 unsigned int mss, struct sk_buff *skb,
612 struct tcp_out_options *opts,
613 const struct tcp_md5sig_key *md5,
614 struct tcp_fastopen_cookie *foc)
616 struct inet_request_sock *ireq = inet_rsk(req);
617 unsigned int remaining = MAX_TCP_OPTION_SPACE;
619 #ifdef CONFIG_TCP_MD5SIG
621 opts->options |= OPTION_MD5;
622 remaining -= TCPOLEN_MD5SIG_ALIGNED;
624 /* We can't fit any SACK blocks in a packet with MD5 + TS
625 * options. There was discussion about disabling SACK
626 * rather than TS in order to fit in better with old,
627 * buggy kernels, but that was deemed to be unnecessary.
629 ireq->tstamp_ok &= !ireq->sack_ok;
633 /* We always send an MSS option. */
635 remaining -= TCPOLEN_MSS_ALIGNED;
637 if (likely(ireq->wscale_ok)) {
638 opts->ws = ireq->rcv_wscale;
639 opts->options |= OPTION_WSCALE;
640 remaining -= TCPOLEN_WSCALE_ALIGNED;
642 if (likely(ireq->tstamp_ok)) {
643 opts->options |= OPTION_TS;
644 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
645 opts->tsecr = req->ts_recent;
646 remaining -= TCPOLEN_TSTAMP_ALIGNED;
648 if (likely(ireq->sack_ok)) {
649 opts->options |= OPTION_SACK_ADVERTISE;
650 if (unlikely(!ireq->tstamp_ok))
651 remaining -= TCPOLEN_SACKPERM_ALIGNED;
653 if (foc != NULL && foc->len >= 0) {
656 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
657 TCPOLEN_FASTOPEN_BASE;
658 need = (need + 3) & ~3U; /* Align to 32 bits */
659 if (remaining >= need) {
660 opts->options |= OPTION_FAST_OPEN_COOKIE;
661 opts->fastopen_cookie = foc;
666 return MAX_TCP_OPTION_SPACE - remaining;
669 /* Compute TCP options for ESTABLISHED sockets. This is not the
670 * final wire format yet.
672 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
673 struct tcp_out_options *opts,
674 struct tcp_md5sig_key **md5)
676 struct tcp_sock *tp = tcp_sk(sk);
677 unsigned int size = 0;
678 unsigned int eff_sacks;
682 #ifdef CONFIG_TCP_MD5SIG
683 *md5 = tp->af_specific->md5_lookup(sk, sk);
684 if (unlikely(*md5)) {
685 opts->options |= OPTION_MD5;
686 size += TCPOLEN_MD5SIG_ALIGNED;
692 if (likely(tp->rx_opt.tstamp_ok)) {
693 opts->options |= OPTION_TS;
694 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
695 opts->tsecr = tp->rx_opt.ts_recent;
696 size += TCPOLEN_TSTAMP_ALIGNED;
699 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
700 if (unlikely(eff_sacks)) {
701 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
702 opts->num_sack_blocks =
703 min_t(unsigned int, eff_sacks,
704 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
705 TCPOLEN_SACK_PERBLOCK);
706 size += TCPOLEN_SACK_BASE_ALIGNED +
707 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
714 /* TCP SMALL QUEUES (TSQ)
716 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
717 * to reduce RTT and bufferbloat.
718 * We do this using a special skb destructor (tcp_wfree).
720 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
721 * needs to be reallocated in a driver.
722 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
724 * Since transmit from skb destructor is forbidden, we use a tasklet
725 * to process all sockets that eventually need to send more skbs.
726 * We use one tasklet per cpu, with its own queue of sockets.
729 struct tasklet_struct tasklet;
730 struct list_head head; /* queue of tcp sockets */
732 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
734 static void tcp_tsq_handler(struct sock *sk)
736 if ((1 << sk->sk_state) &
737 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
738 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
739 struct tcp_sock *tp = tcp_sk(sk);
741 if (tp->lost_out > tp->retrans_out &&
742 tp->snd_cwnd > tcp_packets_in_flight(tp))
743 tcp_xmit_retransmit_queue(sk);
745 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
750 * One tasklet per cpu tries to send more skbs.
751 * We run in tasklet context but need to disable irqs when
752 * transferring tsq->head because tcp_wfree() might
753 * interrupt us (non NAPI drivers)
755 static void tcp_tasklet_func(unsigned long data)
757 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
760 struct list_head *q, *n;
764 local_irq_save(flags);
765 list_splice_init(&tsq->head, &list);
766 local_irq_restore(flags);
768 list_for_each_safe(q, n, &list) {
769 tp = list_entry(q, struct tcp_sock, tsq_node);
770 list_del(&tp->tsq_node);
772 sk = (struct sock *)tp;
773 smp_mb__before_atomic();
774 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
776 if (!sk->sk_lock.owned &&
777 test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) {
779 if (!sock_owned_by_user(sk)) {
780 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
790 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
791 TCPF_WRITE_TIMER_DEFERRED | \
792 TCPF_DELACK_TIMER_DEFERRED | \
793 TCPF_MTU_REDUCED_DEFERRED)
795 * tcp_release_cb - tcp release_sock() callback
798 * called from release_sock() to perform protocol dependent
799 * actions before socket release.
801 void tcp_release_cb(struct sock *sk)
803 unsigned long flags, nflags;
805 /* perform an atomic operation only if at least one flag is set */
807 flags = sk->sk_tsq_flags;
808 if (!(flags & TCP_DEFERRED_ALL))
810 nflags = flags & ~TCP_DEFERRED_ALL;
811 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
813 if (flags & TCPF_TSQ_DEFERRED)
816 /* Here begins the tricky part :
817 * We are called from release_sock() with :
819 * 2) sk_lock.slock spinlock held
820 * 3) socket owned by us (sk->sk_lock.owned == 1)
822 * But following code is meant to be called from BH handlers,
823 * so we should keep BH disabled, but early release socket ownership
825 sock_release_ownership(sk);
827 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
828 tcp_write_timer_handler(sk);
831 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
832 tcp_delack_timer_handler(sk);
835 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
836 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
840 EXPORT_SYMBOL(tcp_release_cb);
842 void __init tcp_tasklet_init(void)
846 for_each_possible_cpu(i) {
847 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
849 INIT_LIST_HEAD(&tsq->head);
850 tasklet_init(&tsq->tasklet,
857 * Write buffer destructor automatically called from kfree_skb.
858 * We can't xmit new skbs from this context, as we might already
861 void tcp_wfree(struct sk_buff *skb)
863 struct sock *sk = skb->sk;
864 struct tcp_sock *tp = tcp_sk(sk);
865 unsigned long flags, nval, oval;
868 /* Keep one reference on sk_wmem_alloc.
869 * Will be released by sk_free() from here or tcp_tasklet_func()
871 wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
873 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
874 * Wait until our queues (qdisc + devices) are drained.
876 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
877 * - chance for incoming ACK (processed by another cpu maybe)
878 * to migrate this flow (skb->ooo_okay will be eventually set)
880 if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
883 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
884 struct tsq_tasklet *tsq;
887 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
890 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
891 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
895 /* queue this socket to tasklet queue */
896 local_irq_save(flags);
897 tsq = this_cpu_ptr(&tsq_tasklet);
898 empty = list_empty(&tsq->head);
899 list_add(&tp->tsq_node, &tsq->head);
901 tasklet_schedule(&tsq->tasklet);
902 local_irq_restore(flags);
909 /* This routine actually transmits TCP packets queued in by
910 * tcp_do_sendmsg(). This is used by both the initial
911 * transmission and possible later retransmissions.
912 * All SKB's seen here are completely headerless. It is our
913 * job to build the TCP header, and pass the packet down to
914 * IP so it can do the same plus pass the packet off to the
917 * We are working here with either a clone of the original
918 * SKB, or a fresh unique copy made by the retransmit engine.
920 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
923 const struct inet_connection_sock *icsk = inet_csk(sk);
924 struct inet_sock *inet;
926 struct tcp_skb_cb *tcb;
927 struct tcp_out_options opts;
928 unsigned int tcp_options_size, tcp_header_size;
929 struct tcp_md5sig_key *md5;
933 BUG_ON(!skb || !tcp_skb_pcount(skb));
937 skb_mstamp_get(&skb->skb_mstamp);
938 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
940 tcp_rate_skb_sent(sk, skb);
942 if (unlikely(skb_cloned(skb)))
943 skb = pskb_copy(skb, gfp_mask);
945 skb = skb_clone(skb, gfp_mask);
951 tcb = TCP_SKB_CB(skb);
952 memset(&opts, 0, sizeof(opts));
954 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
955 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
957 tcp_options_size = tcp_established_options(sk, skb, &opts,
959 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
961 /* if no packet is in qdisc/device queue, then allow XPS to select
962 * another queue. We can be called from tcp_tsq_handler()
963 * which holds one reference to sk_wmem_alloc.
965 * TODO: Ideally, in-flight pure ACK packets should not matter here.
966 * One way to get this would be to set skb->truesize = 2 on them.
968 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
970 /* If we had to use memory reserve to allocate this skb,
971 * this might cause drops if packet is looped back :
972 * Other socket might not have SOCK_MEMALLOC.
973 * Packets not looped back do not care about pfmemalloc.
977 skb_push(skb, tcp_header_size);
978 skb_reset_transport_header(skb);
982 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
983 skb_set_hash_from_sk(skb, sk);
984 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
986 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
988 /* Build TCP header and checksum it. */
989 th = (struct tcphdr *)skb->data;
990 th->source = inet->inet_sport;
991 th->dest = inet->inet_dport;
992 th->seq = htonl(tcb->seq);
993 th->ack_seq = htonl(tp->rcv_nxt);
994 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1000 /* The urg_mode check is necessary during a below snd_una win probe */
1001 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1002 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1003 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1005 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1006 th->urg_ptr = htons(0xFFFF);
1011 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1012 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1013 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1014 th->window = htons(tcp_select_window(sk));
1015 tcp_ecn_send(sk, skb, th, tcp_header_size);
1017 /* RFC1323: The window in SYN & SYN/ACK segments
1020 th->window = htons(min(tp->rcv_wnd, 65535U));
1022 #ifdef CONFIG_TCP_MD5SIG
1023 /* Calculate the MD5 hash, as we have all we need now */
1025 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1026 tp->af_specific->calc_md5_hash(opts.hash_location,
1031 icsk->icsk_af_ops->send_check(sk, skb);
1033 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1034 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1036 if (skb->len != tcp_header_size) {
1037 tcp_event_data_sent(tp, sk);
1038 tp->data_segs_out += tcp_skb_pcount(skb);
1041 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1042 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1043 tcp_skb_pcount(skb));
1045 tp->segs_out += tcp_skb_pcount(skb);
1046 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1047 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1048 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1050 /* Our usage of tstamp should remain private */
1053 /* Cleanup our debris for IP stacks */
1054 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1055 sizeof(struct inet6_skb_parm)));
1057 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1059 if (likely(err <= 0))
1064 return net_xmit_eval(err);
1067 /* This routine just queues the buffer for sending.
1069 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1070 * otherwise socket can stall.
1072 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1074 struct tcp_sock *tp = tcp_sk(sk);
1076 /* Advance write_seq and place onto the write_queue. */
1077 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1078 __skb_header_release(skb);
1079 tcp_add_write_queue_tail(sk, skb);
1080 sk->sk_wmem_queued += skb->truesize;
1081 sk_mem_charge(sk, skb->truesize);
1084 /* Initialize TSO segments for a packet. */
1085 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1087 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1088 /* Avoid the costly divide in the normal
1091 tcp_skb_pcount_set(skb, 1);
1092 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1094 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1095 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1099 /* When a modification to fackets out becomes necessary, we need to check
1100 * skb is counted to fackets_out or not.
1102 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1105 struct tcp_sock *tp = tcp_sk(sk);
1107 if (!tp->sacked_out || tcp_is_reno(tp))
1110 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1111 tp->fackets_out -= decr;
1114 /* Pcount in the middle of the write queue got changed, we need to do various
1115 * tweaks to fix counters
1117 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1119 struct tcp_sock *tp = tcp_sk(sk);
1121 tp->packets_out -= decr;
1123 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1124 tp->sacked_out -= decr;
1125 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1126 tp->retrans_out -= decr;
1127 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1128 tp->lost_out -= decr;
1130 /* Reno case is special. Sigh... */
1131 if (tcp_is_reno(tp) && decr > 0)
1132 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1134 tcp_adjust_fackets_out(sk, skb, decr);
1136 if (tp->lost_skb_hint &&
1137 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1138 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1139 tp->lost_cnt_hint -= decr;
1141 tcp_verify_left_out(tp);
1144 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1146 return TCP_SKB_CB(skb)->txstamp_ack ||
1147 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1150 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1152 struct skb_shared_info *shinfo = skb_shinfo(skb);
1154 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1155 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1156 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1157 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1159 shinfo->tx_flags &= ~tsflags;
1160 shinfo2->tx_flags |= tsflags;
1161 swap(shinfo->tskey, shinfo2->tskey);
1162 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1163 TCP_SKB_CB(skb)->txstamp_ack = 0;
1167 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1169 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1170 TCP_SKB_CB(skb)->eor = 0;
1173 /* Function to create two new TCP segments. Shrinks the given segment
1174 * to the specified size and appends a new segment with the rest of the
1175 * packet to the list. This won't be called frequently, I hope.
1176 * Remember, these are still headerless SKBs at this point.
1178 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1179 unsigned int mss_now, gfp_t gfp)
1181 struct tcp_sock *tp = tcp_sk(sk);
1182 struct sk_buff *buff;
1183 int nsize, old_factor;
1187 if (WARN_ON(len > skb->len))
1190 nsize = skb_headlen(skb) - len;
1194 if (skb_unclone(skb, gfp))
1197 /* Get a new skb... force flag on. */
1198 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1200 return -ENOMEM; /* We'll just try again later. */
1202 sk->sk_wmem_queued += buff->truesize;
1203 sk_mem_charge(sk, buff->truesize);
1204 nlen = skb->len - len - nsize;
1205 buff->truesize += nlen;
1206 skb->truesize -= nlen;
1208 /* Correct the sequence numbers. */
1209 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1210 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1211 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1213 /* PSH and FIN should only be set in the second packet. */
1214 flags = TCP_SKB_CB(skb)->tcp_flags;
1215 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1216 TCP_SKB_CB(buff)->tcp_flags = flags;
1217 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1218 tcp_skb_fragment_eor(skb, buff);
1220 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1221 /* Copy and checksum data tail into the new buffer. */
1222 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1223 skb_put(buff, nsize),
1228 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1230 skb->ip_summed = CHECKSUM_PARTIAL;
1231 skb_split(skb, buff, len);
1234 buff->ip_summed = skb->ip_summed;
1236 buff->tstamp = skb->tstamp;
1237 tcp_fragment_tstamp(skb, buff);
1239 old_factor = tcp_skb_pcount(skb);
1241 /* Fix up tso_factor for both original and new SKB. */
1242 tcp_set_skb_tso_segs(skb, mss_now);
1243 tcp_set_skb_tso_segs(buff, mss_now);
1245 /* Update delivered info for the new segment */
1246 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1248 /* If this packet has been sent out already, we must
1249 * adjust the various packet counters.
1251 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1252 int diff = old_factor - tcp_skb_pcount(skb) -
1253 tcp_skb_pcount(buff);
1256 tcp_adjust_pcount(sk, skb, diff);
1259 /* Link BUFF into the send queue. */
1260 __skb_header_release(buff);
1261 tcp_insert_write_queue_after(skb, buff, sk);
1266 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1267 * eventually). The difference is that pulled data not copied, but
1268 * immediately discarded.
1270 static int __pskb_trim_head(struct sk_buff *skb, int len)
1272 struct skb_shared_info *shinfo;
1275 eat = min_t(int, len, skb_headlen(skb));
1277 __skb_pull(skb, eat);
1284 shinfo = skb_shinfo(skb);
1285 for (i = 0; i < shinfo->nr_frags; i++) {
1286 int size = skb_frag_size(&shinfo->frags[i]);
1289 skb_frag_unref(skb, i);
1292 shinfo->frags[k] = shinfo->frags[i];
1294 shinfo->frags[k].page_offset += eat;
1295 skb_frag_size_sub(&shinfo->frags[k], eat);
1301 shinfo->nr_frags = k;
1303 skb_reset_tail_pointer(skb);
1304 skb->data_len -= len;
1305 skb->len = skb->data_len;
1309 /* Remove acked data from a packet in the transmit queue. */
1310 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1314 if (skb_unclone(skb, GFP_ATOMIC))
1317 delta_truesize = __pskb_trim_head(skb, len);
1319 TCP_SKB_CB(skb)->seq += len;
1320 skb->ip_summed = CHECKSUM_PARTIAL;
1322 if (delta_truesize) {
1323 skb->truesize -= delta_truesize;
1324 sk->sk_wmem_queued -= delta_truesize;
1325 sk_mem_uncharge(sk, delta_truesize);
1326 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1329 /* Any change of skb->len requires recalculation of tso factor. */
1330 if (tcp_skb_pcount(skb) > 1)
1331 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1336 /* Calculate MSS not accounting any TCP options. */
1337 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1339 const struct tcp_sock *tp = tcp_sk(sk);
1340 const struct inet_connection_sock *icsk = inet_csk(sk);
1343 /* Calculate base mss without TCP options:
1344 It is MMS_S - sizeof(tcphdr) of rfc1122
1346 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1348 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1349 if (icsk->icsk_af_ops->net_frag_header_len) {
1350 const struct dst_entry *dst = __sk_dst_get(sk);
1352 if (dst && dst_allfrag(dst))
1353 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1356 /* Clamp it (mss_clamp does not include tcp options) */
1357 if (mss_now > tp->rx_opt.mss_clamp)
1358 mss_now = tp->rx_opt.mss_clamp;
1360 /* Now subtract optional transport overhead */
1361 mss_now -= icsk->icsk_ext_hdr_len;
1363 /* Then reserve room for full set of TCP options and 8 bytes of data */
1369 /* Calculate MSS. Not accounting for SACKs here. */
1370 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1372 /* Subtract TCP options size, not including SACKs */
1373 return __tcp_mtu_to_mss(sk, pmtu) -
1374 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1377 /* Inverse of above */
1378 int tcp_mss_to_mtu(struct sock *sk, int mss)
1380 const struct tcp_sock *tp = tcp_sk(sk);
1381 const struct inet_connection_sock *icsk = inet_csk(sk);
1385 tp->tcp_header_len +
1386 icsk->icsk_ext_hdr_len +
1387 icsk->icsk_af_ops->net_header_len;
1389 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1390 if (icsk->icsk_af_ops->net_frag_header_len) {
1391 const struct dst_entry *dst = __sk_dst_get(sk);
1393 if (dst && dst_allfrag(dst))
1394 mtu += icsk->icsk_af_ops->net_frag_header_len;
1398 EXPORT_SYMBOL(tcp_mss_to_mtu);
1400 /* MTU probing init per socket */
1401 void tcp_mtup_init(struct sock *sk)
1403 struct tcp_sock *tp = tcp_sk(sk);
1404 struct inet_connection_sock *icsk = inet_csk(sk);
1405 struct net *net = sock_net(sk);
1407 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1408 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1409 icsk->icsk_af_ops->net_header_len;
1410 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1411 icsk->icsk_mtup.probe_size = 0;
1412 if (icsk->icsk_mtup.enabled)
1413 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1415 EXPORT_SYMBOL(tcp_mtup_init);
1417 /* This function synchronize snd mss to current pmtu/exthdr set.
1419 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1420 for TCP options, but includes only bare TCP header.
1422 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1423 It is minimum of user_mss and mss received with SYN.
1424 It also does not include TCP options.
1426 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1428 tp->mss_cache is current effective sending mss, including
1429 all tcp options except for SACKs. It is evaluated,
1430 taking into account current pmtu, but never exceeds
1431 tp->rx_opt.mss_clamp.
1433 NOTE1. rfc1122 clearly states that advertised MSS
1434 DOES NOT include either tcp or ip options.
1436 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1437 are READ ONLY outside this function. --ANK (980731)
1439 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1441 struct tcp_sock *tp = tcp_sk(sk);
1442 struct inet_connection_sock *icsk = inet_csk(sk);
1445 if (icsk->icsk_mtup.search_high > pmtu)
1446 icsk->icsk_mtup.search_high = pmtu;
1448 mss_now = tcp_mtu_to_mss(sk, pmtu);
1449 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1451 /* And store cached results */
1452 icsk->icsk_pmtu_cookie = pmtu;
1453 if (icsk->icsk_mtup.enabled)
1454 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1455 tp->mss_cache = mss_now;
1459 EXPORT_SYMBOL(tcp_sync_mss);
1461 /* Compute the current effective MSS, taking SACKs and IP options,
1462 * and even PMTU discovery events into account.
1464 unsigned int tcp_current_mss(struct sock *sk)
1466 const struct tcp_sock *tp = tcp_sk(sk);
1467 const struct dst_entry *dst = __sk_dst_get(sk);
1469 unsigned int header_len;
1470 struct tcp_out_options opts;
1471 struct tcp_md5sig_key *md5;
1473 mss_now = tp->mss_cache;
1476 u32 mtu = dst_mtu(dst);
1477 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1478 mss_now = tcp_sync_mss(sk, mtu);
1481 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1482 sizeof(struct tcphdr);
1483 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1484 * some common options. If this is an odd packet (because we have SACK
1485 * blocks etc) then our calculated header_len will be different, and
1486 * we have to adjust mss_now correspondingly */
1487 if (header_len != tp->tcp_header_len) {
1488 int delta = (int) header_len - tp->tcp_header_len;
1495 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1496 * As additional protections, we do not touch cwnd in retransmission phases,
1497 * and if application hit its sndbuf limit recently.
1499 static void tcp_cwnd_application_limited(struct sock *sk)
1501 struct tcp_sock *tp = tcp_sk(sk);
1503 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1504 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1505 /* Limited by application or receiver window. */
1506 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1507 u32 win_used = max(tp->snd_cwnd_used, init_win);
1508 if (win_used < tp->snd_cwnd) {
1509 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1510 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1512 tp->snd_cwnd_used = 0;
1514 tp->snd_cwnd_stamp = tcp_time_stamp;
1517 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1519 struct tcp_sock *tp = tcp_sk(sk);
1521 /* Track the maximum number of outstanding packets in each
1522 * window, and remember whether we were cwnd-limited then.
1524 if (!before(tp->snd_una, tp->max_packets_seq) ||
1525 tp->packets_out > tp->max_packets_out) {
1526 tp->max_packets_out = tp->packets_out;
1527 tp->max_packets_seq = tp->snd_nxt;
1528 tp->is_cwnd_limited = is_cwnd_limited;
1531 if (tcp_is_cwnd_limited(sk)) {
1532 /* Network is feed fully. */
1533 tp->snd_cwnd_used = 0;
1534 tp->snd_cwnd_stamp = tcp_time_stamp;
1536 /* Network starves. */
1537 if (tp->packets_out > tp->snd_cwnd_used)
1538 tp->snd_cwnd_used = tp->packets_out;
1540 if (sysctl_tcp_slow_start_after_idle &&
1541 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1542 tcp_cwnd_application_limited(sk);
1544 /* The following conditions together indicate the starvation
1545 * is caused by insufficient sender buffer:
1546 * 1) just sent some data (see tcp_write_xmit)
1547 * 2) not cwnd limited (this else condition)
1548 * 3) no more data to send (null tcp_send_head )
1549 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1551 if (!tcp_send_head(sk) && sk->sk_socket &&
1552 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1553 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1554 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1558 /* Minshall's variant of the Nagle send check. */
1559 static bool tcp_minshall_check(const struct tcp_sock *tp)
1561 return after(tp->snd_sml, tp->snd_una) &&
1562 !after(tp->snd_sml, tp->snd_nxt);
1565 /* Update snd_sml if this skb is under mss
1566 * Note that a TSO packet might end with a sub-mss segment
1567 * The test is really :
1568 * if ((skb->len % mss) != 0)
1569 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1570 * But we can avoid doing the divide again given we already have
1571 * skb_pcount = skb->len / mss_now
1573 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1574 const struct sk_buff *skb)
1576 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1577 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1580 /* Return false, if packet can be sent now without violation Nagle's rules:
1581 * 1. It is full sized. (provided by caller in %partial bool)
1582 * 2. Or it contains FIN. (already checked by caller)
1583 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1584 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1585 * With Minshall's modification: all sent small packets are ACKed.
1587 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1591 ((nonagle & TCP_NAGLE_CORK) ||
1592 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1595 /* Return how many segs we'd like on a TSO packet,
1596 * to send one TSO packet per ms
1598 u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1603 bytes = min(sk->sk_pacing_rate >> 10,
1604 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1606 /* Goal is to send at least one packet per ms,
1607 * not one big TSO packet every 100 ms.
1608 * This preserves ACK clocking and is consistent
1609 * with tcp_tso_should_defer() heuristic.
1611 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1613 return min_t(u32, segs, sk->sk_gso_max_segs);
1615 EXPORT_SYMBOL(tcp_tso_autosize);
1617 /* Return the number of segments we want in the skb we are transmitting.
1618 * See if congestion control module wants to decide; otherwise, autosize.
1620 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1622 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1623 u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0;
1626 tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs);
1629 /* Returns the portion of skb which can be sent right away */
1630 static unsigned int tcp_mss_split_point(const struct sock *sk,
1631 const struct sk_buff *skb,
1632 unsigned int mss_now,
1633 unsigned int max_segs,
1636 const struct tcp_sock *tp = tcp_sk(sk);
1637 u32 partial, needed, window, max_len;
1639 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1640 max_len = mss_now * max_segs;
1642 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1645 needed = min(skb->len, window);
1647 if (max_len <= needed)
1650 partial = needed % mss_now;
1651 /* If last segment is not a full MSS, check if Nagle rules allow us
1652 * to include this last segment in this skb.
1653 * Otherwise, we'll split the skb at last MSS boundary
1655 if (tcp_nagle_check(partial != 0, tp, nonagle))
1656 return needed - partial;
1661 /* Can at least one segment of SKB be sent right now, according to the
1662 * congestion window rules? If so, return how many segments are allowed.
1664 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1665 const struct sk_buff *skb)
1667 u32 in_flight, cwnd, halfcwnd;
1669 /* Don't be strict about the congestion window for the final FIN. */
1670 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1671 tcp_skb_pcount(skb) == 1)
1674 in_flight = tcp_packets_in_flight(tp);
1675 cwnd = tp->snd_cwnd;
1676 if (in_flight >= cwnd)
1679 /* For better scheduling, ensure we have at least
1680 * 2 GSO packets in flight.
1682 halfcwnd = max(cwnd >> 1, 1U);
1683 return min(halfcwnd, cwnd - in_flight);
1686 /* Initialize TSO state of a skb.
1687 * This must be invoked the first time we consider transmitting
1688 * SKB onto the wire.
1690 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1692 int tso_segs = tcp_skb_pcount(skb);
1694 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1695 tcp_set_skb_tso_segs(skb, mss_now);
1696 tso_segs = tcp_skb_pcount(skb);
1702 /* Return true if the Nagle test allows this packet to be
1705 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1706 unsigned int cur_mss, int nonagle)
1708 /* Nagle rule does not apply to frames, which sit in the middle of the
1709 * write_queue (they have no chances to get new data).
1711 * This is implemented in the callers, where they modify the 'nonagle'
1712 * argument based upon the location of SKB in the send queue.
1714 if (nonagle & TCP_NAGLE_PUSH)
1717 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1718 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1721 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1727 /* Does at least the first segment of SKB fit into the send window? */
1728 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1729 const struct sk_buff *skb,
1730 unsigned int cur_mss)
1732 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1734 if (skb->len > cur_mss)
1735 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1737 return !after(end_seq, tcp_wnd_end(tp));
1740 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1741 * should be put on the wire right now. If so, it returns the number of
1742 * packets allowed by the congestion window.
1744 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1745 unsigned int cur_mss, int nonagle)
1747 const struct tcp_sock *tp = tcp_sk(sk);
1748 unsigned int cwnd_quota;
1750 tcp_init_tso_segs(skb, cur_mss);
1752 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1755 cwnd_quota = tcp_cwnd_test(tp, skb);
1756 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1762 /* Test if sending is allowed right now. */
1763 bool tcp_may_send_now(struct sock *sk)
1765 const struct tcp_sock *tp = tcp_sk(sk);
1766 struct sk_buff *skb = tcp_send_head(sk);
1769 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1770 (tcp_skb_is_last(sk, skb) ?
1771 tp->nonagle : TCP_NAGLE_PUSH));
1774 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1775 * which is put after SKB on the list. It is very much like
1776 * tcp_fragment() except that it may make several kinds of assumptions
1777 * in order to speed up the splitting operation. In particular, we
1778 * know that all the data is in scatter-gather pages, and that the
1779 * packet has never been sent out before (and thus is not cloned).
1781 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1782 unsigned int mss_now, gfp_t gfp)
1784 struct sk_buff *buff;
1785 int nlen = skb->len - len;
1788 /* All of a TSO frame must be composed of paged data. */
1789 if (skb->len != skb->data_len)
1790 return tcp_fragment(sk, skb, len, mss_now, gfp);
1792 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1793 if (unlikely(!buff))
1796 sk->sk_wmem_queued += buff->truesize;
1797 sk_mem_charge(sk, buff->truesize);
1798 buff->truesize += nlen;
1799 skb->truesize -= nlen;
1801 /* Correct the sequence numbers. */
1802 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1803 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1804 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1806 /* PSH and FIN should only be set in the second packet. */
1807 flags = TCP_SKB_CB(skb)->tcp_flags;
1808 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1809 TCP_SKB_CB(buff)->tcp_flags = flags;
1811 /* This packet was never sent out yet, so no SACK bits. */
1812 TCP_SKB_CB(buff)->sacked = 0;
1814 tcp_skb_fragment_eor(skb, buff);
1816 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1817 skb_split(skb, buff, len);
1818 tcp_fragment_tstamp(skb, buff);
1820 /* Fix up tso_factor for both original and new SKB. */
1821 tcp_set_skb_tso_segs(skb, mss_now);
1822 tcp_set_skb_tso_segs(buff, mss_now);
1824 /* Link BUFF into the send queue. */
1825 __skb_header_release(buff);
1826 tcp_insert_write_queue_after(skb, buff, sk);
1831 /* Try to defer sending, if possible, in order to minimize the amount
1832 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1834 * This algorithm is from John Heffner.
1836 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1837 bool *is_cwnd_limited, u32 max_segs)
1839 const struct inet_connection_sock *icsk = inet_csk(sk);
1840 u32 age, send_win, cong_win, limit, in_flight;
1841 struct tcp_sock *tp = tcp_sk(sk);
1842 struct skb_mstamp now;
1843 struct sk_buff *head;
1846 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1849 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1852 /* Avoid bursty behavior by allowing defer
1853 * only if the last write was recent.
1855 if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1858 in_flight = tcp_packets_in_flight(tp);
1860 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1862 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1864 /* From in_flight test above, we know that cwnd > in_flight. */
1865 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1867 limit = min(send_win, cong_win);
1869 /* If a full-sized TSO skb can be sent, do it. */
1870 if (limit >= max_segs * tp->mss_cache)
1873 /* Middle in queue won't get any more data, full sendable already? */
1874 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1877 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1879 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1881 /* If at least some fraction of a window is available,
1884 chunk /= win_divisor;
1888 /* Different approach, try not to defer past a single
1889 * ACK. Receiver should ACK every other full sized
1890 * frame, so if we have space for more than 3 frames
1893 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1897 head = tcp_write_queue_head(sk);
1898 skb_mstamp_get(&now);
1899 age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1900 /* If next ACK is likely to come too late (half srtt), do not defer */
1901 if (age < (tp->srtt_us >> 4))
1904 /* Ok, it looks like it is advisable to defer. */
1906 if (cong_win < send_win && cong_win <= skb->len)
1907 *is_cwnd_limited = true;
1915 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1917 struct inet_connection_sock *icsk = inet_csk(sk);
1918 struct tcp_sock *tp = tcp_sk(sk);
1919 struct net *net = sock_net(sk);
1923 interval = net->ipv4.sysctl_tcp_probe_interval;
1924 delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1925 if (unlikely(delta >= interval * HZ)) {
1926 int mss = tcp_current_mss(sk);
1928 /* Update current search range */
1929 icsk->icsk_mtup.probe_size = 0;
1930 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1931 sizeof(struct tcphdr) +
1932 icsk->icsk_af_ops->net_header_len;
1933 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1935 /* Update probe time stamp */
1936 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1940 /* Create a new MTU probe if we are ready.
1941 * MTU probe is regularly attempting to increase the path MTU by
1942 * deliberately sending larger packets. This discovers routing
1943 * changes resulting in larger path MTUs.
1945 * Returns 0 if we should wait to probe (no cwnd available),
1946 * 1 if a probe was sent,
1949 static int tcp_mtu_probe(struct sock *sk)
1951 struct inet_connection_sock *icsk = inet_csk(sk);
1952 struct tcp_sock *tp = tcp_sk(sk);
1953 struct sk_buff *skb, *nskb, *next;
1954 struct net *net = sock_net(sk);
1961 /* Not currently probing/verifying,
1963 * have enough cwnd, and
1964 * not SACKing (the variable headers throw things off)
1966 if (likely(!icsk->icsk_mtup.enabled ||
1967 icsk->icsk_mtup.probe_size ||
1968 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1969 tp->snd_cwnd < 11 ||
1970 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
1973 /* Use binary search for probe_size between tcp_mss_base,
1974 * and current mss_clamp. if (search_high - search_low)
1975 * smaller than a threshold, backoff from probing.
1977 mss_now = tcp_current_mss(sk);
1978 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1979 icsk->icsk_mtup.search_low) >> 1);
1980 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1981 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
1982 /* When misfortune happens, we are reprobing actively,
1983 * and then reprobe timer has expired. We stick with current
1984 * probing process by not resetting search range to its orignal.
1986 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
1987 interval < net->ipv4.sysctl_tcp_probe_threshold) {
1988 /* Check whether enough time has elaplased for
1989 * another round of probing.
1991 tcp_mtu_check_reprobe(sk);
1995 /* Have enough data in the send queue to probe? */
1996 if (tp->write_seq - tp->snd_nxt < size_needed)
1999 if (tp->snd_wnd < size_needed)
2001 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2004 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2005 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2006 if (!tcp_packets_in_flight(tp))
2012 /* We're allowed to probe. Build it now. */
2013 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2016 sk->sk_wmem_queued += nskb->truesize;
2017 sk_mem_charge(sk, nskb->truesize);
2019 skb = tcp_send_head(sk);
2021 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2022 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2023 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2024 TCP_SKB_CB(nskb)->sacked = 0;
2026 nskb->ip_summed = skb->ip_summed;
2028 tcp_insert_write_queue_before(nskb, skb, sk);
2031 tcp_for_write_queue_from_safe(skb, next, sk) {
2032 copy = min_t(int, skb->len, probe_size - len);
2033 if (nskb->ip_summed) {
2034 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2036 __wsum csum = skb_copy_and_csum_bits(skb, 0,
2037 skb_put(nskb, copy),
2039 nskb->csum = csum_block_add(nskb->csum, csum, len);
2042 if (skb->len <= copy) {
2043 /* We've eaten all the data from this skb.
2045 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2046 tcp_unlink_write_queue(skb, sk);
2047 sk_wmem_free_skb(sk, skb);
2049 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2050 ~(TCPHDR_FIN|TCPHDR_PSH);
2051 if (!skb_shinfo(skb)->nr_frags) {
2052 skb_pull(skb, copy);
2053 if (skb->ip_summed != CHECKSUM_PARTIAL)
2054 skb->csum = csum_partial(skb->data,
2057 __pskb_trim_head(skb, copy);
2058 tcp_set_skb_tso_segs(skb, mss_now);
2060 TCP_SKB_CB(skb)->seq += copy;
2065 if (len >= probe_size)
2068 tcp_init_tso_segs(nskb, nskb->len);
2070 /* We're ready to send. If this fails, the probe will
2071 * be resegmented into mss-sized pieces by tcp_write_xmit().
2073 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2074 /* Decrement cwnd here because we are sending
2075 * effectively two packets. */
2077 tcp_event_new_data_sent(sk, nskb);
2079 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2080 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2081 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2089 /* TCP Small Queues :
2090 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2091 * (These limits are doubled for retransmits)
2093 * - better RTT estimation and ACK scheduling
2096 * Alas, some drivers / subsystems require a fair amount
2097 * of queued bytes to ensure line rate.
2098 * One example is wifi aggregation (802.11 AMPDU)
2100 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2101 unsigned int factor)
2105 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2106 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2109 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2110 /* Always send the 1st or 2nd skb in write queue.
2111 * No need to wait for TX completion to call us back,
2112 * after softirq/tasklet schedule.
2113 * This helps when TX completions are delayed too much.
2115 if (skb == sk->sk_write_queue.next ||
2116 skb->prev == sk->sk_write_queue.next)
2119 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2120 /* It is possible TX completion already happened
2121 * before we set TSQ_THROTTLED, so we must
2122 * test again the condition.
2124 smp_mb__after_atomic();
2125 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2131 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2133 const u32 now = tcp_time_stamp;
2135 if (tp->chrono_type > TCP_CHRONO_UNSPEC)
2136 tp->chrono_stat[tp->chrono_type - 1] += now - tp->chrono_start;
2137 tp->chrono_start = now;
2138 tp->chrono_type = new;
2141 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2143 struct tcp_sock *tp = tcp_sk(sk);
2145 /* If there are multiple conditions worthy of tracking in a
2146 * chronograph then the highest priority enum takes precedence
2147 * over the other conditions. So that if something "more interesting"
2148 * starts happening, stop the previous chrono and start a new one.
2150 if (type > tp->chrono_type)
2151 tcp_chrono_set(tp, type);
2154 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2156 struct tcp_sock *tp = tcp_sk(sk);
2159 /* There are multiple conditions worthy of tracking in a
2160 * chronograph, so that the highest priority enum takes
2161 * precedence over the other conditions (see tcp_chrono_start).
2162 * If a condition stops, we only stop chrono tracking if
2163 * it's the "most interesting" or current chrono we are
2164 * tracking and starts busy chrono if we have pending data.
2166 if (tcp_write_queue_empty(sk))
2167 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2168 else if (type == tp->chrono_type)
2169 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2172 /* This routine writes packets to the network. It advances the
2173 * send_head. This happens as incoming acks open up the remote
2176 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2177 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2178 * account rare use of URG, this is not a big flaw.
2180 * Send at most one packet when push_one > 0. Temporarily ignore
2181 * cwnd limit to force at most one packet out when push_one == 2.
2183 * Returns true, if no segments are in flight and we have queued segments,
2184 * but cannot send anything now because of SWS or another problem.
2186 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2187 int push_one, gfp_t gfp)
2189 struct tcp_sock *tp = tcp_sk(sk);
2190 struct sk_buff *skb;
2191 unsigned int tso_segs, sent_pkts;
2194 bool is_cwnd_limited = false, is_rwnd_limited = false;
2200 /* Do MTU probing. */
2201 result = tcp_mtu_probe(sk);
2204 } else if (result > 0) {
2209 max_segs = tcp_tso_segs(sk, mss_now);
2210 while ((skb = tcp_send_head(sk))) {
2213 tso_segs = tcp_init_tso_segs(skb, mss_now);
2216 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2217 /* "skb_mstamp" is used as a start point for the retransmit timer */
2218 skb_mstamp_get(&skb->skb_mstamp);
2219 goto repair; /* Skip network transmission */
2222 cwnd_quota = tcp_cwnd_test(tp, skb);
2225 /* Force out a loss probe pkt. */
2231 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2232 is_rwnd_limited = true;
2236 if (tso_segs == 1) {
2237 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2238 (tcp_skb_is_last(sk, skb) ?
2239 nonagle : TCP_NAGLE_PUSH))))
2243 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2249 if (tso_segs > 1 && !tcp_urg_mode(tp))
2250 limit = tcp_mss_split_point(sk, skb, mss_now,
2256 if (skb->len > limit &&
2257 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2260 if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
2261 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
2262 if (tcp_small_queue_check(sk, skb, 0))
2265 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2269 /* Advance the send_head. This one is sent out.
2270 * This call will increment packets_out.
2272 tcp_event_new_data_sent(sk, skb);
2274 tcp_minshall_update(tp, mss_now, skb);
2275 sent_pkts += tcp_skb_pcount(skb);
2281 if (is_rwnd_limited)
2282 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2284 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2286 if (likely(sent_pkts)) {
2287 if (tcp_in_cwnd_reduction(sk))
2288 tp->prr_out += sent_pkts;
2290 /* Send one loss probe per tail loss episode. */
2292 tcp_schedule_loss_probe(sk);
2293 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2294 tcp_cwnd_validate(sk, is_cwnd_limited);
2297 return !tp->packets_out && tcp_send_head(sk);
2300 bool tcp_schedule_loss_probe(struct sock *sk)
2302 struct inet_connection_sock *icsk = inet_csk(sk);
2303 struct tcp_sock *tp = tcp_sk(sk);
2304 u32 timeout, tlp_time_stamp, rto_time_stamp;
2305 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2307 /* No consecutive loss probes. */
2308 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2312 /* Don't do any loss probe on a Fast Open connection before 3WHS
2315 if (tp->fastopen_rsk)
2318 /* TLP is only scheduled when next timer event is RTO. */
2319 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2322 /* Schedule a loss probe in 2*RTT for SACK capable connections
2323 * in Open state, that are either limited by cwnd or application.
2325 if ((sysctl_tcp_early_retrans != 3 && sysctl_tcp_early_retrans != 4) ||
2326 !tp->packets_out || !tcp_is_sack(tp) ||
2327 icsk->icsk_ca_state != TCP_CA_Open)
2330 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2334 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2335 * for delayed ack when there's one outstanding packet. If no RTT
2336 * sample is available then probe after TCP_TIMEOUT_INIT.
2338 timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
2339 if (tp->packets_out == 1)
2340 timeout = max_t(u32, timeout,
2341 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2342 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2344 /* If RTO is shorter, just schedule TLP in its place. */
2345 tlp_time_stamp = tcp_time_stamp + timeout;
2346 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2347 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2348 s32 delta = rto_time_stamp - tcp_time_stamp;
2353 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2358 /* Thanks to skb fast clones, we can detect if a prior transmit of
2359 * a packet is still in a qdisc or driver queue.
2360 * In this case, there is very little point doing a retransmit !
2362 static bool skb_still_in_host_queue(const struct sock *sk,
2363 const struct sk_buff *skb)
2365 if (unlikely(skb_fclone_busy(sk, skb))) {
2366 NET_INC_STATS(sock_net(sk),
2367 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2373 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2374 * retransmit the last segment.
2376 void tcp_send_loss_probe(struct sock *sk)
2378 struct tcp_sock *tp = tcp_sk(sk);
2379 struct sk_buff *skb;
2381 int mss = tcp_current_mss(sk);
2383 skb = tcp_send_head(sk);
2385 if (tcp_snd_wnd_test(tp, skb, mss)) {
2386 pcount = tp->packets_out;
2387 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2388 if (tp->packets_out > pcount)
2392 skb = tcp_write_queue_prev(sk, skb);
2394 skb = tcp_write_queue_tail(sk);
2397 /* At most one outstanding TLP retransmission. */
2398 if (tp->tlp_high_seq)
2401 /* Retransmit last segment. */
2405 if (skb_still_in_host_queue(sk, skb))
2408 pcount = tcp_skb_pcount(skb);
2409 if (WARN_ON(!pcount))
2412 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2413 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2416 skb = tcp_write_queue_next(sk, skb);
2419 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2422 if (__tcp_retransmit_skb(sk, skb, 1))
2425 /* Record snd_nxt for loss detection. */
2426 tp->tlp_high_seq = tp->snd_nxt;
2429 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2430 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2431 inet_csk(sk)->icsk_pending = 0;
2436 /* Push out any pending frames which were held back due to
2437 * TCP_CORK or attempt at coalescing tiny packets.
2438 * The socket must be locked by the caller.
2440 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2443 /* If we are closed, the bytes will have to remain here.
2444 * In time closedown will finish, we empty the write queue and
2445 * all will be happy.
2447 if (unlikely(sk->sk_state == TCP_CLOSE))
2450 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2451 sk_gfp_mask(sk, GFP_ATOMIC)))
2452 tcp_check_probe_timer(sk);
2455 /* Send _single_ skb sitting at the send head. This function requires
2456 * true push pending frames to setup probe timer etc.
2458 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2460 struct sk_buff *skb = tcp_send_head(sk);
2462 BUG_ON(!skb || skb->len < mss_now);
2464 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2467 /* This function returns the amount that we can raise the
2468 * usable window based on the following constraints
2470 * 1. The window can never be shrunk once it is offered (RFC 793)
2471 * 2. We limit memory per socket
2474 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2475 * RECV.NEXT + RCV.WIN fixed until:
2476 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2478 * i.e. don't raise the right edge of the window until you can raise
2479 * it at least MSS bytes.
2481 * Unfortunately, the recommended algorithm breaks header prediction,
2482 * since header prediction assumes th->window stays fixed.
2484 * Strictly speaking, keeping th->window fixed violates the receiver
2485 * side SWS prevention criteria. The problem is that under this rule
2486 * a stream of single byte packets will cause the right side of the
2487 * window to always advance by a single byte.
2489 * Of course, if the sender implements sender side SWS prevention
2490 * then this will not be a problem.
2492 * BSD seems to make the following compromise:
2494 * If the free space is less than the 1/4 of the maximum
2495 * space available and the free space is less than 1/2 mss,
2496 * then set the window to 0.
2497 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2498 * Otherwise, just prevent the window from shrinking
2499 * and from being larger than the largest representable value.
2501 * This prevents incremental opening of the window in the regime
2502 * where TCP is limited by the speed of the reader side taking
2503 * data out of the TCP receive queue. It does nothing about
2504 * those cases where the window is constrained on the sender side
2505 * because the pipeline is full.
2507 * BSD also seems to "accidentally" limit itself to windows that are a
2508 * multiple of MSS, at least until the free space gets quite small.
2509 * This would appear to be a side effect of the mbuf implementation.
2510 * Combining these two algorithms results in the observed behavior
2511 * of having a fixed window size at almost all times.
2513 * Below we obtain similar behavior by forcing the offered window to
2514 * a multiple of the mss when it is feasible to do so.
2516 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2517 * Regular options like TIMESTAMP are taken into account.
2519 u32 __tcp_select_window(struct sock *sk)
2521 struct inet_connection_sock *icsk = inet_csk(sk);
2522 struct tcp_sock *tp = tcp_sk(sk);
2523 /* MSS for the peer's data. Previous versions used mss_clamp
2524 * here. I don't know if the value based on our guesses
2525 * of peer's MSS is better for the performance. It's more correct
2526 * but may be worse for the performance because of rcv_mss
2527 * fluctuations. --SAW 1998/11/1
2529 int mss = icsk->icsk_ack.rcv_mss;
2530 int free_space = tcp_space(sk);
2531 int allowed_space = tcp_full_space(sk);
2532 int full_space = min_t(int, tp->window_clamp, allowed_space);
2535 if (unlikely(mss > full_space)) {
2540 if (free_space < (full_space >> 1)) {
2541 icsk->icsk_ack.quick = 0;
2543 if (tcp_under_memory_pressure(sk))
2544 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2547 /* free_space might become our new window, make sure we don't
2548 * increase it due to wscale.
2550 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2552 /* if free space is less than mss estimate, or is below 1/16th
2553 * of the maximum allowed, try to move to zero-window, else
2554 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2555 * new incoming data is dropped due to memory limits.
2556 * With large window, mss test triggers way too late in order
2557 * to announce zero window in time before rmem limit kicks in.
2559 if (free_space < (allowed_space >> 4) || free_space < mss)
2563 if (free_space > tp->rcv_ssthresh)
2564 free_space = tp->rcv_ssthresh;
2566 /* Don't do rounding if we are using window scaling, since the
2567 * scaled window will not line up with the MSS boundary anyway.
2569 window = tp->rcv_wnd;
2570 if (tp->rx_opt.rcv_wscale) {
2571 window = free_space;
2573 /* Advertise enough space so that it won't get scaled away.
2574 * Import case: prevent zero window announcement if
2575 * 1<<rcv_wscale > mss.
2577 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2578 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2579 << tp->rx_opt.rcv_wscale);
2581 /* Get the largest window that is a nice multiple of mss.
2582 * Window clamp already applied above.
2583 * If our current window offering is within 1 mss of the
2584 * free space we just keep it. This prevents the divide
2585 * and multiply from happening most of the time.
2586 * We also don't do any window rounding when the free space
2589 if (window <= free_space - mss || window > free_space)
2590 window = (free_space / mss) * mss;
2591 else if (mss == full_space &&
2592 free_space > window + (full_space >> 1))
2593 window = free_space;
2599 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2600 const struct sk_buff *next_skb)
2602 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2603 const struct skb_shared_info *next_shinfo =
2604 skb_shinfo(next_skb);
2605 struct skb_shared_info *shinfo = skb_shinfo(skb);
2607 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2608 shinfo->tskey = next_shinfo->tskey;
2609 TCP_SKB_CB(skb)->txstamp_ack |=
2610 TCP_SKB_CB(next_skb)->txstamp_ack;
2614 /* Collapses two adjacent SKB's during retransmission. */
2615 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2617 struct tcp_sock *tp = tcp_sk(sk);
2618 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2619 int skb_size, next_skb_size;
2621 skb_size = skb->len;
2622 next_skb_size = next_skb->len;
2624 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2626 if (next_skb_size) {
2627 if (next_skb_size <= skb_availroom(skb))
2628 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2630 else if (!skb_shift(skb, next_skb, next_skb_size))
2633 tcp_highest_sack_combine(sk, next_skb, skb);
2635 tcp_unlink_write_queue(next_skb, sk);
2637 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2638 skb->ip_summed = CHECKSUM_PARTIAL;
2640 if (skb->ip_summed != CHECKSUM_PARTIAL)
2641 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2643 /* Update sequence range on original skb. */
2644 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2646 /* Merge over control information. This moves PSH/FIN etc. over */
2647 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2649 /* All done, get rid of second SKB and account for it so
2650 * packet counting does not break.
2652 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2653 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2655 /* changed transmit queue under us so clear hints */
2656 tcp_clear_retrans_hints_partial(tp);
2657 if (next_skb == tp->retransmit_skb_hint)
2658 tp->retransmit_skb_hint = skb;
2660 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2662 tcp_skb_collapse_tstamp(skb, next_skb);
2664 sk_wmem_free_skb(sk, next_skb);
2668 /* Check if coalescing SKBs is legal. */
2669 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2671 if (tcp_skb_pcount(skb) > 1)
2673 if (skb_cloned(skb))
2675 if (skb == tcp_send_head(sk))
2677 /* Some heuristics for collapsing over SACK'd could be invented */
2678 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2684 /* Collapse packets in the retransmit queue to make to create
2685 * less packets on the wire. This is only done on retransmission.
2687 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2690 struct tcp_sock *tp = tcp_sk(sk);
2691 struct sk_buff *skb = to, *tmp;
2694 if (!sysctl_tcp_retrans_collapse)
2696 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2699 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2700 if (!tcp_can_collapse(sk, skb))
2703 if (!tcp_skb_can_collapse_to(to))
2716 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2719 if (!tcp_collapse_retrans(sk, to))
2724 /* This retransmits one SKB. Policy decisions and retransmit queue
2725 * state updates are done by the caller. Returns non-zero if an
2726 * error occurred which prevented the send.
2728 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2730 struct inet_connection_sock *icsk = inet_csk(sk);
2731 struct tcp_sock *tp = tcp_sk(sk);
2732 unsigned int cur_mss;
2736 /* Inconclusive MTU probe */
2737 if (icsk->icsk_mtup.probe_size)
2738 icsk->icsk_mtup.probe_size = 0;
2740 /* Do not sent more than we queued. 1/4 is reserved for possible
2741 * copying overhead: fragmentation, tunneling, mangling etc.
2743 if (atomic_read(&sk->sk_wmem_alloc) >
2744 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2748 if (skb_still_in_host_queue(sk, skb))
2751 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2752 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2754 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2758 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2759 return -EHOSTUNREACH; /* Routing failure or similar. */
2761 cur_mss = tcp_current_mss(sk);
2763 /* If receiver has shrunk his window, and skb is out of
2764 * new window, do not retransmit it. The exception is the
2765 * case, when window is shrunk to zero. In this case
2766 * our retransmit serves as a zero window probe.
2768 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2769 TCP_SKB_CB(skb)->seq != tp->snd_una)
2772 len = cur_mss * segs;
2773 if (skb->len > len) {
2774 if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC))
2775 return -ENOMEM; /* We'll try again later. */
2777 if (skb_unclone(skb, GFP_ATOMIC))
2780 diff = tcp_skb_pcount(skb);
2781 tcp_set_skb_tso_segs(skb, cur_mss);
2782 diff -= tcp_skb_pcount(skb);
2784 tcp_adjust_pcount(sk, skb, diff);
2785 if (skb->len < cur_mss)
2786 tcp_retrans_try_collapse(sk, skb, cur_mss);
2789 /* RFC3168, section 6.1.1.1. ECN fallback */
2790 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2791 tcp_ecn_clear_syn(sk, skb);
2793 /* Update global and local TCP statistics. */
2794 segs = tcp_skb_pcount(skb);
2795 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2796 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2797 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2798 tp->total_retrans += segs;
2800 /* make sure skb->data is aligned on arches that require it
2801 * and check if ack-trimming & collapsing extended the headroom
2802 * beyond what csum_start can cover.
2804 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2805 skb_headroom(skb) >= 0xFFFF)) {
2806 struct sk_buff *nskb;
2808 skb_mstamp_get(&skb->skb_mstamp);
2809 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2810 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2813 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2817 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2818 } else if (err != -EBUSY) {
2819 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2824 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2826 struct tcp_sock *tp = tcp_sk(sk);
2827 int err = __tcp_retransmit_skb(sk, skb, segs);
2830 #if FASTRETRANS_DEBUG > 0
2831 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2832 net_dbg_ratelimited("retrans_out leaked\n");
2835 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2836 tp->retrans_out += tcp_skb_pcount(skb);
2838 /* Save stamp of the first retransmit. */
2839 if (!tp->retrans_stamp)
2840 tp->retrans_stamp = tcp_skb_timestamp(skb);
2844 if (tp->undo_retrans < 0)
2845 tp->undo_retrans = 0;
2846 tp->undo_retrans += tcp_skb_pcount(skb);
2850 /* This gets called after a retransmit timeout, and the initially
2851 * retransmitted data is acknowledged. It tries to continue
2852 * resending the rest of the retransmit queue, until either
2853 * we've sent it all or the congestion window limit is reached.
2854 * If doing SACK, the first ACK which comes back for a timeout
2855 * based retransmit packet might feed us FACK information again.
2856 * If so, we use it to avoid unnecessarily retransmissions.
2858 void tcp_xmit_retransmit_queue(struct sock *sk)
2860 const struct inet_connection_sock *icsk = inet_csk(sk);
2861 struct tcp_sock *tp = tcp_sk(sk);
2862 struct sk_buff *skb;
2863 struct sk_buff *hole = NULL;
2867 if (!tp->packets_out)
2870 if (tp->retransmit_skb_hint) {
2871 skb = tp->retransmit_skb_hint;
2873 skb = tcp_write_queue_head(sk);
2876 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2877 tcp_for_write_queue_from(skb, sk) {
2881 if (skb == tcp_send_head(sk))
2883 /* we could do better than to assign each time */
2885 tp->retransmit_skb_hint = skb;
2887 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2890 sacked = TCP_SKB_CB(skb)->sacked;
2891 /* In case tcp_shift_skb_data() have aggregated large skbs,
2892 * we need to make sure not sending too bigs TSO packets
2894 segs = min_t(int, segs, max_segs);
2896 if (tp->retrans_out >= tp->lost_out) {
2898 } else if (!(sacked & TCPCB_LOST)) {
2899 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2904 if (icsk->icsk_ca_state != TCP_CA_Loss)
2905 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2907 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2910 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2913 if (tcp_small_queue_check(sk, skb, 1))
2916 if (tcp_retransmit_skb(sk, skb, segs))
2919 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
2921 if (tcp_in_cwnd_reduction(sk))
2922 tp->prr_out += tcp_skb_pcount(skb);
2924 if (skb == tcp_write_queue_head(sk) &&
2925 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
2926 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2927 inet_csk(sk)->icsk_rto,
2932 /* We allow to exceed memory limits for FIN packets to expedite
2933 * connection tear down and (memory) recovery.
2934 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2935 * or even be forced to close flow without any FIN.
2936 * In general, we want to allow one skb per socket to avoid hangs
2937 * with edge trigger epoll()
2939 void sk_forced_mem_schedule(struct sock *sk, int size)
2943 if (size <= sk->sk_forward_alloc)
2945 amt = sk_mem_pages(size);
2946 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2947 sk_memory_allocated_add(sk, amt);
2949 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
2950 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
2953 /* Send a FIN. The caller locks the socket for us.
2954 * We should try to send a FIN packet really hard, but eventually give up.
2956 void tcp_send_fin(struct sock *sk)
2958 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
2959 struct tcp_sock *tp = tcp_sk(sk);
2961 /* Optimization, tack on the FIN if we have one skb in write queue and
2962 * this skb was not yet sent, or we are under memory pressure.
2963 * Note: in the latter case, FIN packet will be sent after a timeout,
2964 * as TCP stack thinks it has already been transmitted.
2966 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
2968 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2969 TCP_SKB_CB(tskb)->end_seq++;
2971 if (!tcp_send_head(sk)) {
2972 /* This means tskb was already sent.
2973 * Pretend we included the FIN on previous transmit.
2974 * We need to set tp->snd_nxt to the value it would have
2975 * if FIN had been sent. This is because retransmit path
2976 * does not change tp->snd_nxt.
2982 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2983 if (unlikely(!skb)) {
2988 skb_reserve(skb, MAX_TCP_HEADER);
2989 sk_forced_mem_schedule(sk, skb->truesize);
2990 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2991 tcp_init_nondata_skb(skb, tp->write_seq,
2992 TCPHDR_ACK | TCPHDR_FIN);
2993 tcp_queue_skb(sk, skb);
2995 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
2998 /* We get here when a process closes a file descriptor (either due to
2999 * an explicit close() or as a byproduct of exit()'ing) and there
3000 * was unread data in the receive queue. This behavior is recommended
3001 * by RFC 2525, section 2.17. -DaveM
3003 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3005 struct sk_buff *skb;
3007 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3009 /* NOTE: No TCP options attached and we never retransmit this. */
3010 skb = alloc_skb(MAX_TCP_HEADER, priority);
3012 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3016 /* Reserve space for headers and prepare control bits. */
3017 skb_reserve(skb, MAX_TCP_HEADER);
3018 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3019 TCPHDR_ACK | TCPHDR_RST);
3020 skb_mstamp_get(&skb->skb_mstamp);
3022 if (tcp_transmit_skb(sk, skb, 0, priority))
3023 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3026 /* Send a crossed SYN-ACK during socket establishment.
3027 * WARNING: This routine must only be called when we have already sent
3028 * a SYN packet that crossed the incoming SYN that caused this routine
3029 * to get called. If this assumption fails then the initial rcv_wnd
3030 * and rcv_wscale values will not be correct.
3032 int tcp_send_synack(struct sock *sk)
3034 struct sk_buff *skb;
3036 skb = tcp_write_queue_head(sk);
3037 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3038 pr_debug("%s: wrong queue state\n", __func__);
3041 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3042 if (skb_cloned(skb)) {
3043 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
3046 tcp_unlink_write_queue(skb, sk);
3047 __skb_header_release(nskb);
3048 __tcp_add_write_queue_head(sk, nskb);
3049 sk_wmem_free_skb(sk, skb);
3050 sk->sk_wmem_queued += nskb->truesize;
3051 sk_mem_charge(sk, nskb->truesize);
3055 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3056 tcp_ecn_send_synack(sk, skb);
3058 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3062 * tcp_make_synack - Prepare a SYN-ACK.
3063 * sk: listener socket
3064 * dst: dst entry attached to the SYNACK
3065 * req: request_sock pointer
3067 * Allocate one skb and build a SYNACK packet.
3068 * @dst is consumed : Caller should not use it again.
3070 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3071 struct request_sock *req,
3072 struct tcp_fastopen_cookie *foc,
3073 enum tcp_synack_type synack_type)
3075 struct inet_request_sock *ireq = inet_rsk(req);
3076 const struct tcp_sock *tp = tcp_sk(sk);
3077 struct tcp_md5sig_key *md5 = NULL;
3078 struct tcp_out_options opts;
3079 struct sk_buff *skb;
3080 int tcp_header_size;
3084 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3085 if (unlikely(!skb)) {
3089 /* Reserve space for headers. */
3090 skb_reserve(skb, MAX_TCP_HEADER);
3092 switch (synack_type) {
3093 case TCP_SYNACK_NORMAL:
3094 skb_set_owner_w(skb, req_to_sk(req));
3096 case TCP_SYNACK_COOKIE:
3097 /* Under synflood, we do not attach skb to a socket,
3098 * to avoid false sharing.
3101 case TCP_SYNACK_FASTOPEN:
3102 /* sk is a const pointer, because we want to express multiple
3103 * cpu might call us concurrently.
3104 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3106 skb_set_owner_w(skb, (struct sock *)sk);
3109 skb_dst_set(skb, dst);
3111 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3113 memset(&opts, 0, sizeof(opts));
3114 #ifdef CONFIG_SYN_COOKIES
3115 if (unlikely(req->cookie_ts))
3116 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
3119 skb_mstamp_get(&skb->skb_mstamp);
3121 #ifdef CONFIG_TCP_MD5SIG
3123 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3125 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3126 tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
3129 skb_push(skb, tcp_header_size);
3130 skb_reset_transport_header(skb);
3132 th = (struct tcphdr *)skb->data;
3133 memset(th, 0, sizeof(struct tcphdr));
3136 tcp_ecn_make_synack(req, th);
3137 th->source = htons(ireq->ir_num);
3138 th->dest = ireq->ir_rmt_port;
3139 /* Setting of flags are superfluous here for callers (and ECE is
3140 * not even correctly set)
3142 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3143 TCPHDR_SYN | TCPHDR_ACK);
3145 th->seq = htonl(TCP_SKB_CB(skb)->seq);
3146 /* XXX data is queued and acked as is. No buffer/window check */
3147 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3149 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3150 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3151 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3152 th->doff = (tcp_header_size >> 2);
3153 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3155 #ifdef CONFIG_TCP_MD5SIG
3156 /* Okay, we have all we need - do the md5 hash if needed */
3158 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3159 md5, req_to_sk(req), skb);
3163 /* Do not fool tcpdump (if any), clean our debris */
3167 EXPORT_SYMBOL(tcp_make_synack);
3169 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3171 struct inet_connection_sock *icsk = inet_csk(sk);
3172 const struct tcp_congestion_ops *ca;
3173 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3175 if (ca_key == TCP_CA_UNSPEC)
3179 ca = tcp_ca_find_key(ca_key);
3180 if (likely(ca && try_module_get(ca->owner))) {
3181 module_put(icsk->icsk_ca_ops->owner);
3182 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3183 icsk->icsk_ca_ops = ca;
3188 /* Do all connect socket setups that can be done AF independent. */
3189 static void tcp_connect_init(struct sock *sk)
3191 const struct dst_entry *dst = __sk_dst_get(sk);
3192 struct tcp_sock *tp = tcp_sk(sk);
3195 /* We'll fix this up when we get a response from the other end.
3196 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3198 tp->tcp_header_len = sizeof(struct tcphdr) +
3199 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3201 #ifdef CONFIG_TCP_MD5SIG
3202 if (tp->af_specific->md5_lookup(sk, sk))
3203 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3206 /* If user gave his TCP_MAXSEG, record it to clamp */
3207 if (tp->rx_opt.user_mss)
3208 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3211 tcp_sync_mss(sk, dst_mtu(dst));
3213 tcp_ca_dst_init(sk, dst);
3215 if (!tp->window_clamp)
3216 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3217 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3219 tcp_initialize_rcv_mss(sk);
3221 /* limit the window selection if the user enforce a smaller rx buffer */
3222 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3223 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3224 tp->window_clamp = tcp_full_space(sk);
3226 tcp_select_initial_window(tcp_full_space(sk),
3227 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3230 sysctl_tcp_window_scaling,
3232 dst_metric(dst, RTAX_INITRWND));
3234 tp->rx_opt.rcv_wscale = rcv_wscale;
3235 tp->rcv_ssthresh = tp->rcv_wnd;
3238 sock_reset_flag(sk, SOCK_DONE);
3241 tp->snd_una = tp->write_seq;
3242 tp->snd_sml = tp->write_seq;
3243 tp->snd_up = tp->write_seq;
3244 tp->snd_nxt = tp->write_seq;
3246 if (likely(!tp->repair))
3249 tp->rcv_tstamp = tcp_time_stamp;
3250 tp->rcv_wup = tp->rcv_nxt;
3251 tp->copied_seq = tp->rcv_nxt;
3253 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3254 inet_csk(sk)->icsk_retransmits = 0;
3255 tcp_clear_retrans(tp);
3258 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3260 struct tcp_sock *tp = tcp_sk(sk);
3261 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3263 tcb->end_seq += skb->len;
3264 __skb_header_release(skb);
3265 __tcp_add_write_queue_tail(sk, skb);
3266 sk->sk_wmem_queued += skb->truesize;
3267 sk_mem_charge(sk, skb->truesize);
3268 tp->write_seq = tcb->end_seq;
3269 tp->packets_out += tcp_skb_pcount(skb);
3272 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3273 * queue a data-only packet after the regular SYN, such that regular SYNs
3274 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3275 * only the SYN sequence, the data are retransmitted in the first ACK.
3276 * If cookie is not cached or other error occurs, falls back to send a
3277 * regular SYN with Fast Open cookie request option.
3279 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3281 struct tcp_sock *tp = tcp_sk(sk);
3282 struct tcp_fastopen_request *fo = tp->fastopen_req;
3284 struct sk_buff *syn_data;
3286 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3287 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3290 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3291 * user-MSS. Reserve maximum option space for middleboxes that add
3292 * private TCP options. The cost is reduced data space in SYN :(
3294 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3296 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3297 MAX_TCP_OPTION_SPACE;
3299 space = min_t(size_t, space, fo->size);
3301 /* limit to order-0 allocations */
3302 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3304 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3307 syn_data->ip_summed = CHECKSUM_PARTIAL;
3308 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3310 int copied = copy_from_iter(skb_put(syn_data, space), space,
3311 &fo->data->msg_iter);
3312 if (unlikely(!copied)) {
3313 kfree_skb(syn_data);
3316 if (copied != space) {
3317 skb_trim(syn_data, copied);
3321 /* No more data pending in inet_wait_for_connect() */
3322 if (space == fo->size)
3326 tcp_connect_queue_skb(sk, syn_data);
3328 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3330 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3332 syn->skb_mstamp = syn_data->skb_mstamp;
3334 /* Now full SYN+DATA was cloned and sent (or not),
3335 * remove the SYN from the original skb (syn_data)
3336 * we keep in write queue in case of a retransmit, as we
3337 * also have the SYN packet (with no data) in the same queue.
3339 TCP_SKB_CB(syn_data)->seq++;
3340 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3342 tp->syn_data = (fo->copied > 0);
3343 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3348 /* Send a regular SYN with Fast Open cookie request option */
3349 if (fo->cookie.len > 0)
3351 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3353 tp->syn_fastopen = 0;
3355 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3359 /* Build a SYN and send it off. */
3360 int tcp_connect(struct sock *sk)
3362 struct tcp_sock *tp = tcp_sk(sk);
3363 struct sk_buff *buff;
3366 tcp_connect_init(sk);
3368 if (unlikely(tp->repair)) {
3369 tcp_finish_connect(sk, NULL);
3373 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3374 if (unlikely(!buff))
3377 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3378 tp->retrans_stamp = tcp_time_stamp;
3379 tcp_connect_queue_skb(sk, buff);
3380 tcp_ecn_send_syn(sk, buff);
3382 /* Send off SYN; include data in Fast Open. */
3383 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3384 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3385 if (err == -ECONNREFUSED)
3388 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3389 * in order to make this packet get counted in tcpOutSegs.
3391 tp->snd_nxt = tp->write_seq;
3392 tp->pushed_seq = tp->write_seq;
3393 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3395 /* Timer for repeating the SYN until an answer. */
3396 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3397 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3400 EXPORT_SYMBOL(tcp_connect);
3402 /* Send out a delayed ack, the caller does the policy checking
3403 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3406 void tcp_send_delayed_ack(struct sock *sk)
3408 struct inet_connection_sock *icsk = inet_csk(sk);
3409 int ato = icsk->icsk_ack.ato;
3410 unsigned long timeout;
3412 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3414 if (ato > TCP_DELACK_MIN) {
3415 const struct tcp_sock *tp = tcp_sk(sk);
3416 int max_ato = HZ / 2;
3418 if (icsk->icsk_ack.pingpong ||
3419 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3420 max_ato = TCP_DELACK_MAX;
3422 /* Slow path, intersegment interval is "high". */
3424 /* If some rtt estimate is known, use it to bound delayed ack.
3425 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3429 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3436 ato = min(ato, max_ato);
3439 /* Stay within the limit we were given */
3440 timeout = jiffies + ato;
3442 /* Use new timeout only if there wasn't a older one earlier. */
3443 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3444 /* If delack timer was blocked or is about to expire,
3447 if (icsk->icsk_ack.blocked ||
3448 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3453 if (!time_before(timeout, icsk->icsk_ack.timeout))
3454 timeout = icsk->icsk_ack.timeout;
3456 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3457 icsk->icsk_ack.timeout = timeout;
3458 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3461 /* This routine sends an ack and also updates the window. */
3462 void tcp_send_ack(struct sock *sk)
3464 struct sk_buff *buff;
3466 /* If we have been reset, we may not send again. */
3467 if (sk->sk_state == TCP_CLOSE)
3470 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3472 /* We are not putting this on the write queue, so
3473 * tcp_transmit_skb() will set the ownership to this
3476 buff = alloc_skb(MAX_TCP_HEADER,
3477 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3478 if (unlikely(!buff)) {
3479 inet_csk_schedule_ack(sk);
3480 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3481 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3482 TCP_DELACK_MAX, TCP_RTO_MAX);
3486 /* Reserve space for headers and prepare control bits. */
3487 skb_reserve(buff, MAX_TCP_HEADER);
3488 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3490 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3492 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3494 skb_set_tcp_pure_ack(buff);
3496 /* Send it off, this clears delayed acks for us. */
3497 skb_mstamp_get(&buff->skb_mstamp);
3498 tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3500 EXPORT_SYMBOL_GPL(tcp_send_ack);
3502 /* This routine sends a packet with an out of date sequence
3503 * number. It assumes the other end will try to ack it.
3505 * Question: what should we make while urgent mode?
3506 * 4.4BSD forces sending single byte of data. We cannot send
3507 * out of window data, because we have SND.NXT==SND.MAX...
3509 * Current solution: to send TWO zero-length segments in urgent mode:
3510 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3511 * out-of-date with SND.UNA-1 to probe window.
3513 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3515 struct tcp_sock *tp = tcp_sk(sk);
3516 struct sk_buff *skb;
3518 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3519 skb = alloc_skb(MAX_TCP_HEADER,
3520 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3524 /* Reserve space for headers and set control bits. */
3525 skb_reserve(skb, MAX_TCP_HEADER);
3526 /* Use a previous sequence. This should cause the other
3527 * end to send an ack. Don't queue or clone SKB, just
3530 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3531 skb_mstamp_get(&skb->skb_mstamp);
3532 NET_INC_STATS(sock_net(sk), mib);
3533 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3536 void tcp_send_window_probe(struct sock *sk)
3538 if (sk->sk_state == TCP_ESTABLISHED) {
3539 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3540 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3544 /* Initiate keepalive or window probe from timer. */
3545 int tcp_write_wakeup(struct sock *sk, int mib)
3547 struct tcp_sock *tp = tcp_sk(sk);
3548 struct sk_buff *skb;
3550 if (sk->sk_state == TCP_CLOSE)
3553 skb = tcp_send_head(sk);
3554 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3556 unsigned int mss = tcp_current_mss(sk);
3557 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3559 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3560 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3562 /* We are probing the opening of a window
3563 * but the window size is != 0
3564 * must have been a result SWS avoidance ( sender )
3566 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3568 seg_size = min(seg_size, mss);
3569 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3570 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3572 } else if (!tcp_skb_pcount(skb))
3573 tcp_set_skb_tso_segs(skb, mss);
3575 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3576 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3578 tcp_event_new_data_sent(sk, skb);
3581 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3582 tcp_xmit_probe_skb(sk, 1, mib);
3583 return tcp_xmit_probe_skb(sk, 0, mib);
3587 /* A window probe timeout has occurred. If window is not closed send
3588 * a partial packet else a zero probe.
3590 void tcp_send_probe0(struct sock *sk)
3592 struct inet_connection_sock *icsk = inet_csk(sk);
3593 struct tcp_sock *tp = tcp_sk(sk);
3594 struct net *net = sock_net(sk);
3595 unsigned long probe_max;
3598 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3600 if (tp->packets_out || !tcp_send_head(sk)) {
3601 /* Cancel probe timer, if it is not required. */
3602 icsk->icsk_probes_out = 0;
3603 icsk->icsk_backoff = 0;
3608 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3609 icsk->icsk_backoff++;
3610 icsk->icsk_probes_out++;
3611 probe_max = TCP_RTO_MAX;
3613 /* If packet was not sent due to local congestion,
3614 * do not backoff and do not remember icsk_probes_out.
3615 * Let local senders to fight for local resources.
3617 * Use accumulated backoff yet.
3619 if (!icsk->icsk_probes_out)
3620 icsk->icsk_probes_out = 1;
3621 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3623 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3624 tcp_probe0_when(sk, probe_max),
3628 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3630 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3634 tcp_rsk(req)->txhash = net_tx_rndhash();
3635 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3637 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3638 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3639 if (unlikely(tcp_passive_fastopen(sk)))
3640 tcp_sk(sk)->total_retrans++;
3644 EXPORT_SYMBOL(tcp_rtx_synack);