2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly = 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
53 /* Default TSQ limit of four TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly = 262144;
56 /* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
60 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
65 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
66 int push_one, gfp_t gfp);
68 /* Account for new data that has been sent to the network. */
69 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
71 struct inet_connection_sock *icsk = inet_csk(sk);
72 struct tcp_sock *tp = tcp_sk(sk);
73 unsigned int prior_packets = tp->packets_out;
75 tcp_advance_send_head(sk, skb);
76 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
78 tp->packets_out += tcp_skb_pcount(skb);
79 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
82 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
86 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
87 * window scaling factor due to loss of precision.
88 * If window has been shrunk, what should we make? It is not clear at all.
89 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
90 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
91 * invalid. OK, let's make this for now:
93 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
95 const struct tcp_sock *tp = tcp_sk(sk);
97 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
98 (tp->rx_opt.wscale_ok &&
99 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
102 return tcp_wnd_end(tp);
105 /* Calculate mss to advertise in SYN segment.
106 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
108 * 1. It is independent of path mtu.
109 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
110 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
111 * attached devices, because some buggy hosts are confused by
113 * 4. We do not make 3, we advertise MSS, calculated from first
114 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
115 * This may be overridden via information stored in routing table.
116 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
117 * probably even Jumbo".
119 static __u16 tcp_advertise_mss(struct sock *sk)
121 struct tcp_sock *tp = tcp_sk(sk);
122 const struct dst_entry *dst = __sk_dst_get(sk);
123 int mss = tp->advmss;
126 unsigned int metric = dst_metric_advmss(dst);
137 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
138 * This is the first part of cwnd validation mechanism.
140 void tcp_cwnd_restart(struct sock *sk, s32 delta)
142 struct tcp_sock *tp = tcp_sk(sk);
143 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
144 u32 cwnd = tp->snd_cwnd;
146 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
148 tp->snd_ssthresh = tcp_current_ssthresh(sk);
149 restart_cwnd = min(restart_cwnd, cwnd);
151 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
153 tp->snd_cwnd = max(cwnd, restart_cwnd);
154 tp->snd_cwnd_stamp = tcp_jiffies32;
155 tp->snd_cwnd_used = 0;
158 /* Congestion state accounting after a packet has been sent. */
159 static void tcp_event_data_sent(struct tcp_sock *tp,
162 struct inet_connection_sock *icsk = inet_csk(sk);
163 const u32 now = tcp_jiffies32;
165 if (tcp_packets_in_flight(tp) == 0)
166 tcp_ca_event(sk, CA_EVENT_TX_START);
170 /* If it is a reply for ato after last received
171 * packet, enter pingpong mode.
173 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
174 icsk->icsk_ack.pingpong = 1;
177 /* Account for an ACK we sent. */
178 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
180 tcp_dec_quickack_mode(sk, pkts);
181 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
185 u32 tcp_default_init_rwnd(u32 mss)
187 /* Initial receive window should be twice of TCP_INIT_CWND to
188 * enable proper sending of new unsent data during fast recovery
189 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
190 * limit when mss is larger than 1460.
192 u32 init_rwnd = TCP_INIT_CWND * 2;
195 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
199 /* Determine a window scaling and initial window to offer.
200 * Based on the assumption that the given amount of space
201 * will be offered. Store the results in the tp structure.
202 * NOTE: for smooth operation initial space offering should
203 * be a multiple of mss if possible. We assume here that mss >= 1.
204 * This MUST be enforced by all callers.
206 void tcp_select_initial_window(int __space, __u32 mss,
207 __u32 *rcv_wnd, __u32 *window_clamp,
208 int wscale_ok, __u8 *rcv_wscale,
211 unsigned int space = (__space < 0 ? 0 : __space);
213 /* If no clamp set the clamp to the max possible scaled window */
214 if (*window_clamp == 0)
215 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
216 space = min(*window_clamp, space);
218 /* Quantize space offering to a multiple of mss if possible. */
220 space = rounddown(space, mss);
222 /* NOTE: offering an initial window larger than 32767
223 * will break some buggy TCP stacks. If the admin tells us
224 * it is likely we could be speaking with such a buggy stack
225 * we will truncate our initial window offering to 32K-1
226 * unless the remote has sent us a window scaling option,
227 * which we interpret as a sign the remote TCP is not
228 * misinterpreting the window field as a signed quantity.
230 if (sysctl_tcp_workaround_signed_windows)
231 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
237 /* Set window scaling on max possible window */
238 space = max_t(u32, space, sysctl_tcp_rmem[2]);
239 space = max_t(u32, space, sysctl_rmem_max);
240 space = min_t(u32, space, *window_clamp);
241 while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
247 if (mss > (1 << *rcv_wscale)) {
248 if (!init_rcv_wnd) /* Use default unless specified otherwise */
249 init_rcv_wnd = tcp_default_init_rwnd(mss);
250 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
253 /* Set the clamp no higher than max representable value */
254 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
256 EXPORT_SYMBOL(tcp_select_initial_window);
258 /* Chose a new window to advertise, update state in tcp_sock for the
259 * socket, and return result with RFC1323 scaling applied. The return
260 * value can be stuffed directly into th->window for an outgoing
263 static u16 tcp_select_window(struct sock *sk)
265 struct tcp_sock *tp = tcp_sk(sk);
266 u32 old_win = tp->rcv_wnd;
267 u32 cur_win = tcp_receive_window(tp);
268 u32 new_win = __tcp_select_window(sk);
270 /* Never shrink the offered window */
271 if (new_win < cur_win) {
272 /* Danger Will Robinson!
273 * Don't update rcv_wup/rcv_wnd here or else
274 * we will not be able to advertise a zero
275 * window in time. --DaveM
277 * Relax Will Robinson.
280 NET_INC_STATS(sock_net(sk),
281 LINUX_MIB_TCPWANTZEROWINDOWADV);
282 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
284 tp->rcv_wnd = new_win;
285 tp->rcv_wup = tp->rcv_nxt;
287 /* Make sure we do not exceed the maximum possible
290 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
291 new_win = min(new_win, MAX_TCP_WINDOW);
293 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
295 /* RFC1323 scaling applied */
296 new_win >>= tp->rx_opt.rcv_wscale;
300 NET_INC_STATS(sock_net(sk),
301 LINUX_MIB_TCPTOZEROWINDOWADV);
302 } else if (old_win == 0) {
303 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
309 /* Packet ECN state for a SYN-ACK */
310 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
312 const struct tcp_sock *tp = tcp_sk(sk);
314 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
315 if (!(tp->ecn_flags & TCP_ECN_OK))
316 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
317 else if (tcp_ca_needs_ecn(sk) ||
318 tcp_bpf_ca_needs_ecn(sk))
322 /* Packet ECN state for a SYN. */
323 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
325 struct tcp_sock *tp = tcp_sk(sk);
326 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
327 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
328 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
331 const struct dst_entry *dst = __sk_dst_get(sk);
333 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
340 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
341 tp->ecn_flags = TCP_ECN_OK;
342 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
347 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
349 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
350 /* tp->ecn_flags are cleared at a later point in time when
351 * SYN ACK is ultimatively being received.
353 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
357 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
359 if (inet_rsk(req)->ecn_ok)
363 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
366 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
367 struct tcphdr *th, int tcp_header_len)
369 struct tcp_sock *tp = tcp_sk(sk);
371 if (tp->ecn_flags & TCP_ECN_OK) {
372 /* Not-retransmitted data segment: set ECT and inject CWR. */
373 if (skb->len != tcp_header_len &&
374 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
376 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
377 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
379 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
381 } else if (!tcp_ca_needs_ecn(sk)) {
382 /* ACK or retransmitted segment: clear ECT|CE */
383 INET_ECN_dontxmit(sk);
385 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
390 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
391 * auto increment end seqno.
393 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
395 skb->ip_summed = CHECKSUM_PARTIAL;
398 TCP_SKB_CB(skb)->tcp_flags = flags;
399 TCP_SKB_CB(skb)->sacked = 0;
401 tcp_skb_pcount_set(skb, 1);
403 TCP_SKB_CB(skb)->seq = seq;
404 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
406 TCP_SKB_CB(skb)->end_seq = seq;
409 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
411 return tp->snd_una != tp->snd_up;
414 #define OPTION_SACK_ADVERTISE (1 << 0)
415 #define OPTION_TS (1 << 1)
416 #define OPTION_MD5 (1 << 2)
417 #define OPTION_WSCALE (1 << 3)
418 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
420 struct tcp_out_options {
421 u16 options; /* bit field of OPTION_* */
422 u16 mss; /* 0 to disable */
423 u8 ws; /* window scale, 0 to disable */
424 u8 num_sack_blocks; /* number of SACK blocks to include */
425 u8 hash_size; /* bytes in hash_location */
426 __u8 *hash_location; /* temporary pointer, overloaded */
427 __u32 tsval, tsecr; /* need to include OPTION_TS */
428 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
431 /* Write previously computed TCP options to the packet.
433 * Beware: Something in the Internet is very sensitive to the ordering of
434 * TCP options, we learned this through the hard way, so be careful here.
435 * Luckily we can at least blame others for their non-compliance but from
436 * inter-operability perspective it seems that we're somewhat stuck with
437 * the ordering which we have been using if we want to keep working with
438 * those broken things (not that it currently hurts anybody as there isn't
439 * particular reason why the ordering would need to be changed).
441 * At least SACK_PERM as the first option is known to lead to a disaster
442 * (but it may well be that other scenarios fail similarly).
444 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
445 struct tcp_out_options *opts)
447 u16 options = opts->options; /* mungable copy */
449 if (unlikely(OPTION_MD5 & options)) {
450 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
451 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
452 /* overload cookie hash location */
453 opts->hash_location = (__u8 *)ptr;
457 if (unlikely(opts->mss)) {
458 *ptr++ = htonl((TCPOPT_MSS << 24) |
459 (TCPOLEN_MSS << 16) |
463 if (likely(OPTION_TS & options)) {
464 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
465 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
466 (TCPOLEN_SACK_PERM << 16) |
467 (TCPOPT_TIMESTAMP << 8) |
469 options &= ~OPTION_SACK_ADVERTISE;
471 *ptr++ = htonl((TCPOPT_NOP << 24) |
473 (TCPOPT_TIMESTAMP << 8) |
476 *ptr++ = htonl(opts->tsval);
477 *ptr++ = htonl(opts->tsecr);
480 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
481 *ptr++ = htonl((TCPOPT_NOP << 24) |
483 (TCPOPT_SACK_PERM << 8) |
487 if (unlikely(OPTION_WSCALE & options)) {
488 *ptr++ = htonl((TCPOPT_NOP << 24) |
489 (TCPOPT_WINDOW << 16) |
490 (TCPOLEN_WINDOW << 8) |
494 if (unlikely(opts->num_sack_blocks)) {
495 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
496 tp->duplicate_sack : tp->selective_acks;
499 *ptr++ = htonl((TCPOPT_NOP << 24) |
502 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
503 TCPOLEN_SACK_PERBLOCK)));
505 for (this_sack = 0; this_sack < opts->num_sack_blocks;
507 *ptr++ = htonl(sp[this_sack].start_seq);
508 *ptr++ = htonl(sp[this_sack].end_seq);
511 tp->rx_opt.dsack = 0;
514 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
515 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
517 u32 len; /* Fast Open option length */
520 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
521 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
522 TCPOPT_FASTOPEN_MAGIC);
523 p += TCPOLEN_EXP_FASTOPEN_BASE;
525 len = TCPOLEN_FASTOPEN_BASE + foc->len;
526 *p++ = TCPOPT_FASTOPEN;
530 memcpy(p, foc->val, foc->len);
531 if ((len & 3) == 2) {
532 p[foc->len] = TCPOPT_NOP;
533 p[foc->len + 1] = TCPOPT_NOP;
535 ptr += (len + 3) >> 2;
539 /* Compute TCP options for SYN packets. This is not the final
540 * network wire format yet.
542 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
543 struct tcp_out_options *opts,
544 struct tcp_md5sig_key **md5)
546 struct tcp_sock *tp = tcp_sk(sk);
547 unsigned int remaining = MAX_TCP_OPTION_SPACE;
548 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
550 #ifdef CONFIG_TCP_MD5SIG
551 *md5 = tp->af_specific->md5_lookup(sk, sk);
553 opts->options |= OPTION_MD5;
554 remaining -= TCPOLEN_MD5SIG_ALIGNED;
560 /* We always get an MSS option. The option bytes which will be seen in
561 * normal data packets should timestamps be used, must be in the MSS
562 * advertised. But we subtract them from tp->mss_cache so that
563 * calculations in tcp_sendmsg are simpler etc. So account for this
564 * fact here if necessary. If we don't do this correctly, as a
565 * receiver we won't recognize data packets as being full sized when we
566 * should, and thus we won't abide by the delayed ACK rules correctly.
567 * SACKs don't matter, we never delay an ACK when we have any of those
569 opts->mss = tcp_advertise_mss(sk);
570 remaining -= TCPOLEN_MSS_ALIGNED;
572 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
573 opts->options |= OPTION_TS;
574 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
575 opts->tsecr = tp->rx_opt.ts_recent;
576 remaining -= TCPOLEN_TSTAMP_ALIGNED;
578 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
579 opts->ws = tp->rx_opt.rcv_wscale;
580 opts->options |= OPTION_WSCALE;
581 remaining -= TCPOLEN_WSCALE_ALIGNED;
583 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
584 opts->options |= OPTION_SACK_ADVERTISE;
585 if (unlikely(!(OPTION_TS & opts->options)))
586 remaining -= TCPOLEN_SACKPERM_ALIGNED;
589 if (fastopen && fastopen->cookie.len >= 0) {
590 u32 need = fastopen->cookie.len;
592 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
593 TCPOLEN_FASTOPEN_BASE;
594 need = (need + 3) & ~3U; /* Align to 32 bits */
595 if (remaining >= need) {
596 opts->options |= OPTION_FAST_OPEN_COOKIE;
597 opts->fastopen_cookie = &fastopen->cookie;
599 tp->syn_fastopen = 1;
600 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
604 return MAX_TCP_OPTION_SPACE - remaining;
607 /* Set up TCP options for SYN-ACKs. */
608 static unsigned int tcp_synack_options(struct request_sock *req,
609 unsigned int mss, struct sk_buff *skb,
610 struct tcp_out_options *opts,
611 const struct tcp_md5sig_key *md5,
612 struct tcp_fastopen_cookie *foc)
614 struct inet_request_sock *ireq = inet_rsk(req);
615 unsigned int remaining = MAX_TCP_OPTION_SPACE;
617 #ifdef CONFIG_TCP_MD5SIG
619 opts->options |= OPTION_MD5;
620 remaining -= TCPOLEN_MD5SIG_ALIGNED;
622 /* We can't fit any SACK blocks in a packet with MD5 + TS
623 * options. There was discussion about disabling SACK
624 * rather than TS in order to fit in better with old,
625 * buggy kernels, but that was deemed to be unnecessary.
627 ireq->tstamp_ok &= !ireq->sack_ok;
631 /* We always send an MSS option. */
633 remaining -= TCPOLEN_MSS_ALIGNED;
635 if (likely(ireq->wscale_ok)) {
636 opts->ws = ireq->rcv_wscale;
637 opts->options |= OPTION_WSCALE;
638 remaining -= TCPOLEN_WSCALE_ALIGNED;
640 if (likely(ireq->tstamp_ok)) {
641 opts->options |= OPTION_TS;
642 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
643 opts->tsecr = req->ts_recent;
644 remaining -= TCPOLEN_TSTAMP_ALIGNED;
646 if (likely(ireq->sack_ok)) {
647 opts->options |= OPTION_SACK_ADVERTISE;
648 if (unlikely(!ireq->tstamp_ok))
649 remaining -= TCPOLEN_SACKPERM_ALIGNED;
651 if (foc != NULL && foc->len >= 0) {
654 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
655 TCPOLEN_FASTOPEN_BASE;
656 need = (need + 3) & ~3U; /* Align to 32 bits */
657 if (remaining >= need) {
658 opts->options |= OPTION_FAST_OPEN_COOKIE;
659 opts->fastopen_cookie = foc;
664 return MAX_TCP_OPTION_SPACE - remaining;
667 /* Compute TCP options for ESTABLISHED sockets. This is not the
668 * final wire format yet.
670 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
671 struct tcp_out_options *opts,
672 struct tcp_md5sig_key **md5)
674 struct tcp_sock *tp = tcp_sk(sk);
675 unsigned int size = 0;
676 unsigned int eff_sacks;
680 #ifdef CONFIG_TCP_MD5SIG
681 *md5 = tp->af_specific->md5_lookup(sk, sk);
682 if (unlikely(*md5)) {
683 opts->options |= OPTION_MD5;
684 size += TCPOLEN_MD5SIG_ALIGNED;
690 if (likely(tp->rx_opt.tstamp_ok)) {
691 opts->options |= OPTION_TS;
692 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
693 opts->tsecr = tp->rx_opt.ts_recent;
694 size += TCPOLEN_TSTAMP_ALIGNED;
697 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
698 if (unlikely(eff_sacks)) {
699 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
700 opts->num_sack_blocks =
701 min_t(unsigned int, eff_sacks,
702 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
703 TCPOLEN_SACK_PERBLOCK);
704 size += TCPOLEN_SACK_BASE_ALIGNED +
705 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
712 /* TCP SMALL QUEUES (TSQ)
714 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
715 * to reduce RTT and bufferbloat.
716 * We do this using a special skb destructor (tcp_wfree).
718 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
719 * needs to be reallocated in a driver.
720 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
722 * Since transmit from skb destructor is forbidden, we use a tasklet
723 * to process all sockets that eventually need to send more skbs.
724 * We use one tasklet per cpu, with its own queue of sockets.
727 struct tasklet_struct tasklet;
728 struct list_head head; /* queue of tcp sockets */
730 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
732 static void tcp_tsq_handler(struct sock *sk)
734 if ((1 << sk->sk_state) &
735 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
736 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
737 struct tcp_sock *tp = tcp_sk(sk);
739 if (tp->lost_out > tp->retrans_out &&
740 tp->snd_cwnd > tcp_packets_in_flight(tp))
741 tcp_xmit_retransmit_queue(sk);
743 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
748 * One tasklet per cpu tries to send more skbs.
749 * We run in tasklet context but need to disable irqs when
750 * transferring tsq->head because tcp_wfree() might
751 * interrupt us (non NAPI drivers)
753 static void tcp_tasklet_func(unsigned long data)
755 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
758 struct list_head *q, *n;
762 local_irq_save(flags);
763 list_splice_init(&tsq->head, &list);
764 local_irq_restore(flags);
766 list_for_each_safe(q, n, &list) {
767 tp = list_entry(q, struct tcp_sock, tsq_node);
768 list_del(&tp->tsq_node);
770 sk = (struct sock *)tp;
771 smp_mb__before_atomic();
772 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
774 if (!sk->sk_lock.owned &&
775 test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) {
777 if (!sock_owned_by_user(sk)) {
778 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
788 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
789 TCPF_WRITE_TIMER_DEFERRED | \
790 TCPF_DELACK_TIMER_DEFERRED | \
791 TCPF_MTU_REDUCED_DEFERRED)
793 * tcp_release_cb - tcp release_sock() callback
796 * called from release_sock() to perform protocol dependent
797 * actions before socket release.
799 void tcp_release_cb(struct sock *sk)
801 unsigned long flags, nflags;
803 /* perform an atomic operation only if at least one flag is set */
805 flags = sk->sk_tsq_flags;
806 if (!(flags & TCP_DEFERRED_ALL))
808 nflags = flags & ~TCP_DEFERRED_ALL;
809 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
811 if (flags & TCPF_TSQ_DEFERRED)
814 /* Here begins the tricky part :
815 * We are called from release_sock() with :
817 * 2) sk_lock.slock spinlock held
818 * 3) socket owned by us (sk->sk_lock.owned == 1)
820 * But following code is meant to be called from BH handlers,
821 * so we should keep BH disabled, but early release socket ownership
823 sock_release_ownership(sk);
825 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
826 tcp_write_timer_handler(sk);
829 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
830 tcp_delack_timer_handler(sk);
833 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
834 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
838 EXPORT_SYMBOL(tcp_release_cb);
840 void __init tcp_tasklet_init(void)
844 for_each_possible_cpu(i) {
845 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
847 INIT_LIST_HEAD(&tsq->head);
848 tasklet_init(&tsq->tasklet,
855 * Write buffer destructor automatically called from kfree_skb.
856 * We can't xmit new skbs from this context, as we might already
859 void tcp_wfree(struct sk_buff *skb)
861 struct sock *sk = skb->sk;
862 struct tcp_sock *tp = tcp_sk(sk);
863 unsigned long flags, nval, oval;
865 /* Keep one reference on sk_wmem_alloc.
866 * Will be released by sk_free() from here or tcp_tasklet_func()
868 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
870 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
871 * Wait until our queues (qdisc + devices) are drained.
873 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
874 * - chance for incoming ACK (processed by another cpu maybe)
875 * to migrate this flow (skb->ooo_okay will be eventually set)
877 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
880 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
881 struct tsq_tasklet *tsq;
884 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
887 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
888 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
892 /* queue this socket to tasklet queue */
893 local_irq_save(flags);
894 tsq = this_cpu_ptr(&tsq_tasklet);
895 empty = list_empty(&tsq->head);
896 list_add(&tp->tsq_node, &tsq->head);
898 tasklet_schedule(&tsq->tasklet);
899 local_irq_restore(flags);
906 /* Note: Called under hard irq.
907 * We can not call TCP stack right away.
909 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
911 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
912 struct sock *sk = (struct sock *)tp;
913 unsigned long nval, oval;
915 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
916 struct tsq_tasklet *tsq;
919 if (oval & TSQF_QUEUED)
922 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
923 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
927 if (!refcount_inc_not_zero(&sk->sk_wmem_alloc))
929 /* queue this socket to tasklet queue */
930 tsq = this_cpu_ptr(&tsq_tasklet);
931 empty = list_empty(&tsq->head);
932 list_add(&tp->tsq_node, &tsq->head);
934 tasklet_schedule(&tsq->tasklet);
937 return HRTIMER_NORESTART;
940 /* BBR congestion control needs pacing.
941 * Same remark for SO_MAX_PACING_RATE.
942 * sch_fq packet scheduler is efficiently handling pacing,
943 * but is not always installed/used.
944 * Return true if TCP stack should pace packets itself.
946 static bool tcp_needs_internal_pacing(const struct sock *sk)
948 return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
951 static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
956 if (!tcp_needs_internal_pacing(sk))
958 rate = sk->sk_pacing_rate;
959 if (!rate || rate == ~0U)
962 /* Should account for header sizes as sch_fq does,
963 * but lets make things simple.
965 len_ns = (u64)skb->len * NSEC_PER_SEC;
966 do_div(len_ns, rate);
967 hrtimer_start(&tcp_sk(sk)->pacing_timer,
968 ktime_add_ns(ktime_get(), len_ns),
969 HRTIMER_MODE_ABS_PINNED);
972 /* This routine actually transmits TCP packets queued in by
973 * tcp_do_sendmsg(). This is used by both the initial
974 * transmission and possible later retransmissions.
975 * All SKB's seen here are completely headerless. It is our
976 * job to build the TCP header, and pass the packet down to
977 * IP so it can do the same plus pass the packet off to the
980 * We are working here with either a clone of the original
981 * SKB, or a fresh unique copy made by the retransmit engine.
983 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
986 const struct inet_connection_sock *icsk = inet_csk(sk);
987 struct inet_sock *inet;
989 struct tcp_skb_cb *tcb;
990 struct tcp_out_options opts;
991 unsigned int tcp_options_size, tcp_header_size;
992 struct tcp_md5sig_key *md5;
996 BUG_ON(!skb || !tcp_skb_pcount(skb));
999 skb->skb_mstamp = tp->tcp_mstamp;
1001 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1003 tcp_rate_skb_sent(sk, skb);
1005 if (unlikely(skb_cloned(skb)))
1006 skb = pskb_copy(skb, gfp_mask);
1008 skb = skb_clone(skb, gfp_mask);
1014 tcb = TCP_SKB_CB(skb);
1015 memset(&opts, 0, sizeof(opts));
1017 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1018 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1020 tcp_options_size = tcp_established_options(sk, skb, &opts,
1022 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1024 /* if no packet is in qdisc/device queue, then allow XPS to select
1025 * another queue. We can be called from tcp_tsq_handler()
1026 * which holds one reference to sk_wmem_alloc.
1028 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1029 * One way to get this would be to set skb->truesize = 2 on them.
1031 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1033 /* If we had to use memory reserve to allocate this skb,
1034 * this might cause drops if packet is looped back :
1035 * Other socket might not have SOCK_MEMALLOC.
1036 * Packets not looped back do not care about pfmemalloc.
1038 skb->pfmemalloc = 0;
1040 skb_push(skb, tcp_header_size);
1041 skb_reset_transport_header(skb);
1045 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1046 skb_set_hash_from_sk(skb, sk);
1047 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1049 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1051 /* Build TCP header and checksum it. */
1052 th = (struct tcphdr *)skb->data;
1053 th->source = inet->inet_sport;
1054 th->dest = inet->inet_dport;
1055 th->seq = htonl(tcb->seq);
1056 th->ack_seq = htonl(tp->rcv_nxt);
1057 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1063 /* The urg_mode check is necessary during a below snd_una win probe */
1064 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1065 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1066 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1068 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1069 th->urg_ptr = htons(0xFFFF);
1074 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1075 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1076 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1077 th->window = htons(tcp_select_window(sk));
1078 tcp_ecn_send(sk, skb, th, tcp_header_size);
1080 /* RFC1323: The window in SYN & SYN/ACK segments
1083 th->window = htons(min(tp->rcv_wnd, 65535U));
1085 #ifdef CONFIG_TCP_MD5SIG
1086 /* Calculate the MD5 hash, as we have all we need now */
1088 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1089 tp->af_specific->calc_md5_hash(opts.hash_location,
1094 icsk->icsk_af_ops->send_check(sk, skb);
1096 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1097 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1099 if (skb->len != tcp_header_size) {
1100 tcp_event_data_sent(tp, sk);
1101 tp->data_segs_out += tcp_skb_pcount(skb);
1102 tcp_internal_pacing(sk, skb);
1105 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1106 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1107 tcp_skb_pcount(skb));
1109 tp->segs_out += tcp_skb_pcount(skb);
1110 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1111 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1112 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1114 /* Our usage of tstamp should remain private */
1117 /* Cleanup our debris for IP stacks */
1118 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1119 sizeof(struct inet6_skb_parm)));
1121 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1123 if (likely(err <= 0))
1128 return net_xmit_eval(err);
1131 /* This routine just queues the buffer for sending.
1133 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1134 * otherwise socket can stall.
1136 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1138 struct tcp_sock *tp = tcp_sk(sk);
1140 /* Advance write_seq and place onto the write_queue. */
1141 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1142 __skb_header_release(skb);
1143 tcp_add_write_queue_tail(sk, skb);
1144 sk->sk_wmem_queued += skb->truesize;
1145 sk_mem_charge(sk, skb->truesize);
1148 /* Initialize TSO segments for a packet. */
1149 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1151 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1152 /* Avoid the costly divide in the normal
1155 tcp_skb_pcount_set(skb, 1);
1156 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1158 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1159 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1163 /* When a modification to fackets out becomes necessary, we need to check
1164 * skb is counted to fackets_out or not.
1166 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1169 struct tcp_sock *tp = tcp_sk(sk);
1171 if (!tp->sacked_out || tcp_is_reno(tp))
1174 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1175 tp->fackets_out -= decr;
1178 /* Pcount in the middle of the write queue got changed, we need to do various
1179 * tweaks to fix counters
1181 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1183 struct tcp_sock *tp = tcp_sk(sk);
1185 tp->packets_out -= decr;
1187 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1188 tp->sacked_out -= decr;
1189 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1190 tp->retrans_out -= decr;
1191 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1192 tp->lost_out -= decr;
1194 /* Reno case is special. Sigh... */
1195 if (tcp_is_reno(tp) && decr > 0)
1196 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1198 tcp_adjust_fackets_out(sk, skb, decr);
1200 if (tp->lost_skb_hint &&
1201 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1202 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1203 tp->lost_cnt_hint -= decr;
1205 tcp_verify_left_out(tp);
1208 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1210 return TCP_SKB_CB(skb)->txstamp_ack ||
1211 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1214 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1216 struct skb_shared_info *shinfo = skb_shinfo(skb);
1218 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1219 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1220 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1221 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1223 shinfo->tx_flags &= ~tsflags;
1224 shinfo2->tx_flags |= tsflags;
1225 swap(shinfo->tskey, shinfo2->tskey);
1226 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1227 TCP_SKB_CB(skb)->txstamp_ack = 0;
1231 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1233 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1234 TCP_SKB_CB(skb)->eor = 0;
1237 /* Function to create two new TCP segments. Shrinks the given segment
1238 * to the specified size and appends a new segment with the rest of the
1239 * packet to the list. This won't be called frequently, I hope.
1240 * Remember, these are still headerless SKBs at this point.
1242 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1243 unsigned int mss_now, gfp_t gfp)
1245 struct tcp_sock *tp = tcp_sk(sk);
1246 struct sk_buff *buff;
1247 int nsize, old_factor;
1251 if (WARN_ON(len > skb->len))
1254 nsize = skb_headlen(skb) - len;
1258 if (skb_unclone(skb, gfp))
1261 /* Get a new skb... force flag on. */
1262 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1264 return -ENOMEM; /* We'll just try again later. */
1266 sk->sk_wmem_queued += buff->truesize;
1267 sk_mem_charge(sk, buff->truesize);
1268 nlen = skb->len - len - nsize;
1269 buff->truesize += nlen;
1270 skb->truesize -= nlen;
1272 /* Correct the sequence numbers. */
1273 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1274 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1275 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1277 /* PSH and FIN should only be set in the second packet. */
1278 flags = TCP_SKB_CB(skb)->tcp_flags;
1279 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1280 TCP_SKB_CB(buff)->tcp_flags = flags;
1281 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1282 tcp_skb_fragment_eor(skb, buff);
1284 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1285 /* Copy and checksum data tail into the new buffer. */
1286 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1287 skb_put(buff, nsize),
1292 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1294 skb->ip_summed = CHECKSUM_PARTIAL;
1295 skb_split(skb, buff, len);
1298 buff->ip_summed = skb->ip_summed;
1300 buff->tstamp = skb->tstamp;
1301 tcp_fragment_tstamp(skb, buff);
1303 old_factor = tcp_skb_pcount(skb);
1305 /* Fix up tso_factor for both original and new SKB. */
1306 tcp_set_skb_tso_segs(skb, mss_now);
1307 tcp_set_skb_tso_segs(buff, mss_now);
1309 /* Update delivered info for the new segment */
1310 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1312 /* If this packet has been sent out already, we must
1313 * adjust the various packet counters.
1315 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1316 int diff = old_factor - tcp_skb_pcount(skb) -
1317 tcp_skb_pcount(buff);
1320 tcp_adjust_pcount(sk, skb, diff);
1323 /* Link BUFF into the send queue. */
1324 __skb_header_release(buff);
1325 tcp_insert_write_queue_after(skb, buff, sk);
1330 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1331 * data is not copied, but immediately discarded.
1333 static int __pskb_trim_head(struct sk_buff *skb, int len)
1335 struct skb_shared_info *shinfo;
1338 eat = min_t(int, len, skb_headlen(skb));
1340 __skb_pull(skb, eat);
1347 shinfo = skb_shinfo(skb);
1348 for (i = 0; i < shinfo->nr_frags; i++) {
1349 int size = skb_frag_size(&shinfo->frags[i]);
1352 skb_frag_unref(skb, i);
1355 shinfo->frags[k] = shinfo->frags[i];
1357 shinfo->frags[k].page_offset += eat;
1358 skb_frag_size_sub(&shinfo->frags[k], eat);
1364 shinfo->nr_frags = k;
1366 skb->data_len -= len;
1367 skb->len = skb->data_len;
1371 /* Remove acked data from a packet in the transmit queue. */
1372 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1376 if (skb_unclone(skb, GFP_ATOMIC))
1379 delta_truesize = __pskb_trim_head(skb, len);
1381 TCP_SKB_CB(skb)->seq += len;
1382 skb->ip_summed = CHECKSUM_PARTIAL;
1384 if (delta_truesize) {
1385 skb->truesize -= delta_truesize;
1386 sk->sk_wmem_queued -= delta_truesize;
1387 sk_mem_uncharge(sk, delta_truesize);
1388 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1391 /* Any change of skb->len requires recalculation of tso factor. */
1392 if (tcp_skb_pcount(skb) > 1)
1393 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1398 /* Calculate MSS not accounting any TCP options. */
1399 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1401 const struct tcp_sock *tp = tcp_sk(sk);
1402 const struct inet_connection_sock *icsk = inet_csk(sk);
1405 /* Calculate base mss without TCP options:
1406 It is MMS_S - sizeof(tcphdr) of rfc1122
1408 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1410 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1411 if (icsk->icsk_af_ops->net_frag_header_len) {
1412 const struct dst_entry *dst = __sk_dst_get(sk);
1414 if (dst && dst_allfrag(dst))
1415 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1418 /* Clamp it (mss_clamp does not include tcp options) */
1419 if (mss_now > tp->rx_opt.mss_clamp)
1420 mss_now = tp->rx_opt.mss_clamp;
1422 /* Now subtract optional transport overhead */
1423 mss_now -= icsk->icsk_ext_hdr_len;
1425 /* Then reserve room for full set of TCP options and 8 bytes of data */
1431 /* Calculate MSS. Not accounting for SACKs here. */
1432 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1434 /* Subtract TCP options size, not including SACKs */
1435 return __tcp_mtu_to_mss(sk, pmtu) -
1436 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1439 /* Inverse of above */
1440 int tcp_mss_to_mtu(struct sock *sk, int mss)
1442 const struct tcp_sock *tp = tcp_sk(sk);
1443 const struct inet_connection_sock *icsk = inet_csk(sk);
1447 tp->tcp_header_len +
1448 icsk->icsk_ext_hdr_len +
1449 icsk->icsk_af_ops->net_header_len;
1451 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1452 if (icsk->icsk_af_ops->net_frag_header_len) {
1453 const struct dst_entry *dst = __sk_dst_get(sk);
1455 if (dst && dst_allfrag(dst))
1456 mtu += icsk->icsk_af_ops->net_frag_header_len;
1460 EXPORT_SYMBOL(tcp_mss_to_mtu);
1462 /* MTU probing init per socket */
1463 void tcp_mtup_init(struct sock *sk)
1465 struct tcp_sock *tp = tcp_sk(sk);
1466 struct inet_connection_sock *icsk = inet_csk(sk);
1467 struct net *net = sock_net(sk);
1469 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1470 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1471 icsk->icsk_af_ops->net_header_len;
1472 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1473 icsk->icsk_mtup.probe_size = 0;
1474 if (icsk->icsk_mtup.enabled)
1475 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1477 EXPORT_SYMBOL(tcp_mtup_init);
1479 /* This function synchronize snd mss to current pmtu/exthdr set.
1481 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1482 for TCP options, but includes only bare TCP header.
1484 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1485 It is minimum of user_mss and mss received with SYN.
1486 It also does not include TCP options.
1488 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1490 tp->mss_cache is current effective sending mss, including
1491 all tcp options except for SACKs. It is evaluated,
1492 taking into account current pmtu, but never exceeds
1493 tp->rx_opt.mss_clamp.
1495 NOTE1. rfc1122 clearly states that advertised MSS
1496 DOES NOT include either tcp or ip options.
1498 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1499 are READ ONLY outside this function. --ANK (980731)
1501 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1503 struct tcp_sock *tp = tcp_sk(sk);
1504 struct inet_connection_sock *icsk = inet_csk(sk);
1507 if (icsk->icsk_mtup.search_high > pmtu)
1508 icsk->icsk_mtup.search_high = pmtu;
1510 mss_now = tcp_mtu_to_mss(sk, pmtu);
1511 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1513 /* And store cached results */
1514 icsk->icsk_pmtu_cookie = pmtu;
1515 if (icsk->icsk_mtup.enabled)
1516 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1517 tp->mss_cache = mss_now;
1521 EXPORT_SYMBOL(tcp_sync_mss);
1523 /* Compute the current effective MSS, taking SACKs and IP options,
1524 * and even PMTU discovery events into account.
1526 unsigned int tcp_current_mss(struct sock *sk)
1528 const struct tcp_sock *tp = tcp_sk(sk);
1529 const struct dst_entry *dst = __sk_dst_get(sk);
1531 unsigned int header_len;
1532 struct tcp_out_options opts;
1533 struct tcp_md5sig_key *md5;
1535 mss_now = tp->mss_cache;
1538 u32 mtu = dst_mtu(dst);
1539 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1540 mss_now = tcp_sync_mss(sk, mtu);
1543 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1544 sizeof(struct tcphdr);
1545 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1546 * some common options. If this is an odd packet (because we have SACK
1547 * blocks etc) then our calculated header_len will be different, and
1548 * we have to adjust mss_now correspondingly */
1549 if (header_len != tp->tcp_header_len) {
1550 int delta = (int) header_len - tp->tcp_header_len;
1557 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1558 * As additional protections, we do not touch cwnd in retransmission phases,
1559 * and if application hit its sndbuf limit recently.
1561 static void tcp_cwnd_application_limited(struct sock *sk)
1563 struct tcp_sock *tp = tcp_sk(sk);
1565 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1566 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1567 /* Limited by application or receiver window. */
1568 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1569 u32 win_used = max(tp->snd_cwnd_used, init_win);
1570 if (win_used < tp->snd_cwnd) {
1571 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1572 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1574 tp->snd_cwnd_used = 0;
1576 tp->snd_cwnd_stamp = tcp_jiffies32;
1579 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1581 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1582 struct tcp_sock *tp = tcp_sk(sk);
1584 /* Track the maximum number of outstanding packets in each
1585 * window, and remember whether we were cwnd-limited then.
1587 if (!before(tp->snd_una, tp->max_packets_seq) ||
1588 tp->packets_out > tp->max_packets_out) {
1589 tp->max_packets_out = tp->packets_out;
1590 tp->max_packets_seq = tp->snd_nxt;
1591 tp->is_cwnd_limited = is_cwnd_limited;
1594 if (tcp_is_cwnd_limited(sk)) {
1595 /* Network is feed fully. */
1596 tp->snd_cwnd_used = 0;
1597 tp->snd_cwnd_stamp = tcp_jiffies32;
1599 /* Network starves. */
1600 if (tp->packets_out > tp->snd_cwnd_used)
1601 tp->snd_cwnd_used = tp->packets_out;
1603 if (sysctl_tcp_slow_start_after_idle &&
1604 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1605 !ca_ops->cong_control)
1606 tcp_cwnd_application_limited(sk);
1608 /* The following conditions together indicate the starvation
1609 * is caused by insufficient sender buffer:
1610 * 1) just sent some data (see tcp_write_xmit)
1611 * 2) not cwnd limited (this else condition)
1612 * 3) no more data to send (null tcp_send_head )
1613 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1615 if (!tcp_send_head(sk) && sk->sk_socket &&
1616 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1617 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1618 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1622 /* Minshall's variant of the Nagle send check. */
1623 static bool tcp_minshall_check(const struct tcp_sock *tp)
1625 return after(tp->snd_sml, tp->snd_una) &&
1626 !after(tp->snd_sml, tp->snd_nxt);
1629 /* Update snd_sml if this skb is under mss
1630 * Note that a TSO packet might end with a sub-mss segment
1631 * The test is really :
1632 * if ((skb->len % mss) != 0)
1633 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1634 * But we can avoid doing the divide again given we already have
1635 * skb_pcount = skb->len / mss_now
1637 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1638 const struct sk_buff *skb)
1640 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1641 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1644 /* Return false, if packet can be sent now without violation Nagle's rules:
1645 * 1. It is full sized. (provided by caller in %partial bool)
1646 * 2. Or it contains FIN. (already checked by caller)
1647 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1648 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1649 * With Minshall's modification: all sent small packets are ACKed.
1651 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1655 ((nonagle & TCP_NAGLE_CORK) ||
1656 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1659 /* Return how many segs we'd like on a TSO packet,
1660 * to send one TSO packet per ms
1662 u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1667 bytes = min(sk->sk_pacing_rate >> 10,
1668 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1670 /* Goal is to send at least one packet per ms,
1671 * not one big TSO packet every 100 ms.
1672 * This preserves ACK clocking and is consistent
1673 * with tcp_tso_should_defer() heuristic.
1675 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1677 return min_t(u32, segs, sk->sk_gso_max_segs);
1679 EXPORT_SYMBOL(tcp_tso_autosize);
1681 /* Return the number of segments we want in the skb we are transmitting.
1682 * See if congestion control module wants to decide; otherwise, autosize.
1684 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1686 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1687 u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0;
1690 tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs);
1693 /* Returns the portion of skb which can be sent right away */
1694 static unsigned int tcp_mss_split_point(const struct sock *sk,
1695 const struct sk_buff *skb,
1696 unsigned int mss_now,
1697 unsigned int max_segs,
1700 const struct tcp_sock *tp = tcp_sk(sk);
1701 u32 partial, needed, window, max_len;
1703 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1704 max_len = mss_now * max_segs;
1706 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1709 needed = min(skb->len, window);
1711 if (max_len <= needed)
1714 partial = needed % mss_now;
1715 /* If last segment is not a full MSS, check if Nagle rules allow us
1716 * to include this last segment in this skb.
1717 * Otherwise, we'll split the skb at last MSS boundary
1719 if (tcp_nagle_check(partial != 0, tp, nonagle))
1720 return needed - partial;
1725 /* Can at least one segment of SKB be sent right now, according to the
1726 * congestion window rules? If so, return how many segments are allowed.
1728 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1729 const struct sk_buff *skb)
1731 u32 in_flight, cwnd, halfcwnd;
1733 /* Don't be strict about the congestion window for the final FIN. */
1734 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1735 tcp_skb_pcount(skb) == 1)
1738 in_flight = tcp_packets_in_flight(tp);
1739 cwnd = tp->snd_cwnd;
1740 if (in_flight >= cwnd)
1743 /* For better scheduling, ensure we have at least
1744 * 2 GSO packets in flight.
1746 halfcwnd = max(cwnd >> 1, 1U);
1747 return min(halfcwnd, cwnd - in_flight);
1750 /* Initialize TSO state of a skb.
1751 * This must be invoked the first time we consider transmitting
1752 * SKB onto the wire.
1754 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1756 int tso_segs = tcp_skb_pcount(skb);
1758 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1759 tcp_set_skb_tso_segs(skb, mss_now);
1760 tso_segs = tcp_skb_pcount(skb);
1766 /* Return true if the Nagle test allows this packet to be
1769 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1770 unsigned int cur_mss, int nonagle)
1772 /* Nagle rule does not apply to frames, which sit in the middle of the
1773 * write_queue (they have no chances to get new data).
1775 * This is implemented in the callers, where they modify the 'nonagle'
1776 * argument based upon the location of SKB in the send queue.
1778 if (nonagle & TCP_NAGLE_PUSH)
1781 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1782 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1785 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1791 /* Does at least the first segment of SKB fit into the send window? */
1792 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1793 const struct sk_buff *skb,
1794 unsigned int cur_mss)
1796 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1798 if (skb->len > cur_mss)
1799 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1801 return !after(end_seq, tcp_wnd_end(tp));
1804 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1805 * should be put on the wire right now. If so, it returns the number of
1806 * packets allowed by the congestion window.
1808 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1809 unsigned int cur_mss, int nonagle)
1811 const struct tcp_sock *tp = tcp_sk(sk);
1812 unsigned int cwnd_quota;
1814 tcp_init_tso_segs(skb, cur_mss);
1816 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1819 cwnd_quota = tcp_cwnd_test(tp, skb);
1820 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1826 /* Test if sending is allowed right now. */
1827 bool tcp_may_send_now(struct sock *sk)
1829 const struct tcp_sock *tp = tcp_sk(sk);
1830 struct sk_buff *skb = tcp_send_head(sk);
1833 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1834 (tcp_skb_is_last(sk, skb) ?
1835 tp->nonagle : TCP_NAGLE_PUSH));
1838 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1839 * which is put after SKB on the list. It is very much like
1840 * tcp_fragment() except that it may make several kinds of assumptions
1841 * in order to speed up the splitting operation. In particular, we
1842 * know that all the data is in scatter-gather pages, and that the
1843 * packet has never been sent out before (and thus is not cloned).
1845 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1846 unsigned int mss_now, gfp_t gfp)
1848 struct sk_buff *buff;
1849 int nlen = skb->len - len;
1852 /* All of a TSO frame must be composed of paged data. */
1853 if (skb->len != skb->data_len)
1854 return tcp_fragment(sk, skb, len, mss_now, gfp);
1856 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1857 if (unlikely(!buff))
1860 sk->sk_wmem_queued += buff->truesize;
1861 sk_mem_charge(sk, buff->truesize);
1862 buff->truesize += nlen;
1863 skb->truesize -= nlen;
1865 /* Correct the sequence numbers. */
1866 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1867 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1868 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1870 /* PSH and FIN should only be set in the second packet. */
1871 flags = TCP_SKB_CB(skb)->tcp_flags;
1872 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1873 TCP_SKB_CB(buff)->tcp_flags = flags;
1875 /* This packet was never sent out yet, so no SACK bits. */
1876 TCP_SKB_CB(buff)->sacked = 0;
1878 tcp_skb_fragment_eor(skb, buff);
1880 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1881 skb_split(skb, buff, len);
1882 tcp_fragment_tstamp(skb, buff);
1884 /* Fix up tso_factor for both original and new SKB. */
1885 tcp_set_skb_tso_segs(skb, mss_now);
1886 tcp_set_skb_tso_segs(buff, mss_now);
1888 /* Link BUFF into the send queue. */
1889 __skb_header_release(buff);
1890 tcp_insert_write_queue_after(skb, buff, sk);
1895 /* Try to defer sending, if possible, in order to minimize the amount
1896 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1898 * This algorithm is from John Heffner.
1900 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1901 bool *is_cwnd_limited, u32 max_segs)
1903 const struct inet_connection_sock *icsk = inet_csk(sk);
1904 u32 age, send_win, cong_win, limit, in_flight;
1905 struct tcp_sock *tp = tcp_sk(sk);
1906 struct sk_buff *head;
1909 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1912 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1915 /* Avoid bursty behavior by allowing defer
1916 * only if the last write was recent.
1918 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
1921 in_flight = tcp_packets_in_flight(tp);
1923 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1925 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1927 /* From in_flight test above, we know that cwnd > in_flight. */
1928 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1930 limit = min(send_win, cong_win);
1932 /* If a full-sized TSO skb can be sent, do it. */
1933 if (limit >= max_segs * tp->mss_cache)
1936 /* Middle in queue won't get any more data, full sendable already? */
1937 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1940 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1942 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1944 /* If at least some fraction of a window is available,
1947 chunk /= win_divisor;
1951 /* Different approach, try not to defer past a single
1952 * ACK. Receiver should ACK every other full sized
1953 * frame, so if we have space for more than 3 frames
1956 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1960 head = tcp_write_queue_head(sk);
1962 age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
1963 /* If next ACK is likely to come too late (half srtt), do not defer */
1964 if (age < (tp->srtt_us >> 4))
1967 /* Ok, it looks like it is advisable to defer. */
1969 if (cong_win < send_win && cong_win <= skb->len)
1970 *is_cwnd_limited = true;
1978 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1980 struct inet_connection_sock *icsk = inet_csk(sk);
1981 struct tcp_sock *tp = tcp_sk(sk);
1982 struct net *net = sock_net(sk);
1986 interval = net->ipv4.sysctl_tcp_probe_interval;
1987 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
1988 if (unlikely(delta >= interval * HZ)) {
1989 int mss = tcp_current_mss(sk);
1991 /* Update current search range */
1992 icsk->icsk_mtup.probe_size = 0;
1993 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1994 sizeof(struct tcphdr) +
1995 icsk->icsk_af_ops->net_header_len;
1996 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1998 /* Update probe time stamp */
1999 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2003 /* Create a new MTU probe if we are ready.
2004 * MTU probe is regularly attempting to increase the path MTU by
2005 * deliberately sending larger packets. This discovers routing
2006 * changes resulting in larger path MTUs.
2008 * Returns 0 if we should wait to probe (no cwnd available),
2009 * 1 if a probe was sent,
2012 static int tcp_mtu_probe(struct sock *sk)
2014 struct inet_connection_sock *icsk = inet_csk(sk);
2015 struct tcp_sock *tp = tcp_sk(sk);
2016 struct sk_buff *skb, *nskb, *next;
2017 struct net *net = sock_net(sk);
2024 /* Not currently probing/verifying,
2026 * have enough cwnd, and
2027 * not SACKing (the variable headers throw things off)
2029 if (likely(!icsk->icsk_mtup.enabled ||
2030 icsk->icsk_mtup.probe_size ||
2031 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2032 tp->snd_cwnd < 11 ||
2033 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2036 /* Use binary search for probe_size between tcp_mss_base,
2037 * and current mss_clamp. if (search_high - search_low)
2038 * smaller than a threshold, backoff from probing.
2040 mss_now = tcp_current_mss(sk);
2041 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2042 icsk->icsk_mtup.search_low) >> 1);
2043 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2044 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2045 /* When misfortune happens, we are reprobing actively,
2046 * and then reprobe timer has expired. We stick with current
2047 * probing process by not resetting search range to its orignal.
2049 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2050 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2051 /* Check whether enough time has elaplased for
2052 * another round of probing.
2054 tcp_mtu_check_reprobe(sk);
2058 /* Have enough data in the send queue to probe? */
2059 if (tp->write_seq - tp->snd_nxt < size_needed)
2062 if (tp->snd_wnd < size_needed)
2064 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2067 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2068 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2069 if (!tcp_packets_in_flight(tp))
2075 /* We're allowed to probe. Build it now. */
2076 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2079 sk->sk_wmem_queued += nskb->truesize;
2080 sk_mem_charge(sk, nskb->truesize);
2082 skb = tcp_send_head(sk);
2084 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2085 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2086 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2087 TCP_SKB_CB(nskb)->sacked = 0;
2089 nskb->ip_summed = skb->ip_summed;
2091 tcp_insert_write_queue_before(nskb, skb, sk);
2094 tcp_for_write_queue_from_safe(skb, next, sk) {
2095 copy = min_t(int, skb->len, probe_size - len);
2096 if (nskb->ip_summed) {
2097 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2099 __wsum csum = skb_copy_and_csum_bits(skb, 0,
2100 skb_put(nskb, copy),
2102 nskb->csum = csum_block_add(nskb->csum, csum, len);
2105 if (skb->len <= copy) {
2106 /* We've eaten all the data from this skb.
2108 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2109 tcp_unlink_write_queue(skb, sk);
2110 sk_wmem_free_skb(sk, skb);
2112 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2113 ~(TCPHDR_FIN|TCPHDR_PSH);
2114 if (!skb_shinfo(skb)->nr_frags) {
2115 skb_pull(skb, copy);
2116 if (skb->ip_summed != CHECKSUM_PARTIAL)
2117 skb->csum = csum_partial(skb->data,
2120 __pskb_trim_head(skb, copy);
2121 tcp_set_skb_tso_segs(skb, mss_now);
2123 TCP_SKB_CB(skb)->seq += copy;
2128 if (len >= probe_size)
2131 tcp_init_tso_segs(nskb, nskb->len);
2133 /* We're ready to send. If this fails, the probe will
2134 * be resegmented into mss-sized pieces by tcp_write_xmit().
2136 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2137 /* Decrement cwnd here because we are sending
2138 * effectively two packets. */
2140 tcp_event_new_data_sent(sk, nskb);
2142 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2143 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2144 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2152 static bool tcp_pacing_check(const struct sock *sk)
2154 return tcp_needs_internal_pacing(sk) &&
2155 hrtimer_active(&tcp_sk(sk)->pacing_timer);
2158 /* TCP Small Queues :
2159 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2160 * (These limits are doubled for retransmits)
2162 * - better RTT estimation and ACK scheduling
2165 * Alas, some drivers / subsystems require a fair amount
2166 * of queued bytes to ensure line rate.
2167 * One example is wifi aggregation (802.11 AMPDU)
2169 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2170 unsigned int factor)
2174 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2175 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2178 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2179 /* Always send the 1st or 2nd skb in write queue.
2180 * No need to wait for TX completion to call us back,
2181 * after softirq/tasklet schedule.
2182 * This helps when TX completions are delayed too much.
2184 if (skb == sk->sk_write_queue.next ||
2185 skb->prev == sk->sk_write_queue.next)
2188 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2189 /* It is possible TX completion already happened
2190 * before we set TSQ_THROTTLED, so we must
2191 * test again the condition.
2193 smp_mb__after_atomic();
2194 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2200 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2202 const u32 now = tcp_jiffies32;
2203 enum tcp_chrono old = tp->chrono_type;
2205 if (old > TCP_CHRONO_UNSPEC)
2206 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2207 tp->chrono_start = now;
2208 tp->chrono_type = new;
2211 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2213 struct tcp_sock *tp = tcp_sk(sk);
2215 /* If there are multiple conditions worthy of tracking in a
2216 * chronograph then the highest priority enum takes precedence
2217 * over the other conditions. So that if something "more interesting"
2218 * starts happening, stop the previous chrono and start a new one.
2220 if (type > tp->chrono_type)
2221 tcp_chrono_set(tp, type);
2224 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2226 struct tcp_sock *tp = tcp_sk(sk);
2229 /* There are multiple conditions worthy of tracking in a
2230 * chronograph, so that the highest priority enum takes
2231 * precedence over the other conditions (see tcp_chrono_start).
2232 * If a condition stops, we only stop chrono tracking if
2233 * it's the "most interesting" or current chrono we are
2234 * tracking and starts busy chrono if we have pending data.
2236 if (tcp_write_queue_empty(sk))
2237 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2238 else if (type == tp->chrono_type)
2239 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2242 /* This routine writes packets to the network. It advances the
2243 * send_head. This happens as incoming acks open up the remote
2246 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2247 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2248 * account rare use of URG, this is not a big flaw.
2250 * Send at most one packet when push_one > 0. Temporarily ignore
2251 * cwnd limit to force at most one packet out when push_one == 2.
2253 * Returns true, if no segments are in flight and we have queued segments,
2254 * but cannot send anything now because of SWS or another problem.
2256 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2257 int push_one, gfp_t gfp)
2259 struct tcp_sock *tp = tcp_sk(sk);
2260 struct sk_buff *skb;
2261 unsigned int tso_segs, sent_pkts;
2264 bool is_cwnd_limited = false, is_rwnd_limited = false;
2270 /* Do MTU probing. */
2271 result = tcp_mtu_probe(sk);
2274 } else if (result > 0) {
2279 max_segs = tcp_tso_segs(sk, mss_now);
2280 tcp_mstamp_refresh(tp);
2281 while ((skb = tcp_send_head(sk))) {
2284 if (tcp_pacing_check(sk))
2287 tso_segs = tcp_init_tso_segs(skb, mss_now);
2290 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2291 /* "skb_mstamp" is used as a start point for the retransmit timer */
2292 skb->skb_mstamp = tp->tcp_mstamp;
2293 goto repair; /* Skip network transmission */
2296 cwnd_quota = tcp_cwnd_test(tp, skb);
2299 /* Force out a loss probe pkt. */
2305 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2306 is_rwnd_limited = true;
2310 if (tso_segs == 1) {
2311 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2312 (tcp_skb_is_last(sk, skb) ?
2313 nonagle : TCP_NAGLE_PUSH))))
2317 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2323 if (tso_segs > 1 && !tcp_urg_mode(tp))
2324 limit = tcp_mss_split_point(sk, skb, mss_now,
2330 if (skb->len > limit &&
2331 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2334 if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
2335 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
2336 if (tcp_small_queue_check(sk, skb, 0))
2339 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2343 /* Advance the send_head. This one is sent out.
2344 * This call will increment packets_out.
2346 tcp_event_new_data_sent(sk, skb);
2348 tcp_minshall_update(tp, mss_now, skb);
2349 sent_pkts += tcp_skb_pcount(skb);
2355 if (is_rwnd_limited)
2356 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2358 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2360 if (likely(sent_pkts)) {
2361 if (tcp_in_cwnd_reduction(sk))
2362 tp->prr_out += sent_pkts;
2364 /* Send one loss probe per tail loss episode. */
2366 tcp_schedule_loss_probe(sk);
2367 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2368 tcp_cwnd_validate(sk, is_cwnd_limited);
2371 return !tp->packets_out && tcp_send_head(sk);
2374 bool tcp_schedule_loss_probe(struct sock *sk)
2376 struct inet_connection_sock *icsk = inet_csk(sk);
2377 struct tcp_sock *tp = tcp_sk(sk);
2378 u32 timeout, tlp_time_stamp, rto_time_stamp;
2380 /* No consecutive loss probes. */
2381 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2385 /* Don't do any loss probe on a Fast Open connection before 3WHS
2388 if (tp->fastopen_rsk)
2391 /* TLP is only scheduled when next timer event is RTO. */
2392 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2395 /* Schedule a loss probe in 2*RTT for SACK capable connections
2396 * in Open state, that are either limited by cwnd or application.
2398 if ((sysctl_tcp_early_retrans != 3 && sysctl_tcp_early_retrans != 4) ||
2399 !tp->packets_out || !tcp_is_sack(tp) ||
2400 icsk->icsk_ca_state != TCP_CA_Open)
2403 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2407 /* Probe timeout is 2*rtt. Add minimum RTO to account
2408 * for delayed ack when there's one outstanding packet. If no RTT
2409 * sample is available then probe after TCP_TIMEOUT_INIT.
2412 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2413 if (tp->packets_out == 1)
2414 timeout += TCP_RTO_MIN;
2416 timeout += TCP_TIMEOUT_MIN;
2418 timeout = TCP_TIMEOUT_INIT;
2421 /* If RTO is shorter, just schedule TLP in its place. */
2422 tlp_time_stamp = tcp_jiffies32 + timeout;
2423 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2424 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2425 s32 delta = rto_time_stamp - tcp_jiffies32;
2430 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2435 /* Thanks to skb fast clones, we can detect if a prior transmit of
2436 * a packet is still in a qdisc or driver queue.
2437 * In this case, there is very little point doing a retransmit !
2439 static bool skb_still_in_host_queue(const struct sock *sk,
2440 const struct sk_buff *skb)
2442 if (unlikely(skb_fclone_busy(sk, skb))) {
2443 NET_INC_STATS(sock_net(sk),
2444 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2450 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2451 * retransmit the last segment.
2453 void tcp_send_loss_probe(struct sock *sk)
2455 struct tcp_sock *tp = tcp_sk(sk);
2456 struct sk_buff *skb;
2458 int mss = tcp_current_mss(sk);
2460 skb = tcp_send_head(sk);
2462 if (tcp_snd_wnd_test(tp, skb, mss)) {
2463 pcount = tp->packets_out;
2464 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2465 if (tp->packets_out > pcount)
2469 skb = tcp_write_queue_prev(sk, skb);
2471 skb = tcp_write_queue_tail(sk);
2474 /* At most one outstanding TLP retransmission. */
2475 if (tp->tlp_high_seq)
2478 /* Retransmit last segment. */
2482 if (skb_still_in_host_queue(sk, skb))
2485 pcount = tcp_skb_pcount(skb);
2486 if (WARN_ON(!pcount))
2489 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2490 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2493 skb = tcp_write_queue_next(sk, skb);
2496 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2499 if (__tcp_retransmit_skb(sk, skb, 1))
2502 /* Record snd_nxt for loss detection. */
2503 tp->tlp_high_seq = tp->snd_nxt;
2506 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2507 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2508 inet_csk(sk)->icsk_pending = 0;
2513 /* Push out any pending frames which were held back due to
2514 * TCP_CORK or attempt at coalescing tiny packets.
2515 * The socket must be locked by the caller.
2517 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2520 /* If we are closed, the bytes will have to remain here.
2521 * In time closedown will finish, we empty the write queue and
2522 * all will be happy.
2524 if (unlikely(sk->sk_state == TCP_CLOSE))
2527 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2528 sk_gfp_mask(sk, GFP_ATOMIC)))
2529 tcp_check_probe_timer(sk);
2532 /* Send _single_ skb sitting at the send head. This function requires
2533 * true push pending frames to setup probe timer etc.
2535 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2537 struct sk_buff *skb = tcp_send_head(sk);
2539 BUG_ON(!skb || skb->len < mss_now);
2541 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2544 /* This function returns the amount that we can raise the
2545 * usable window based on the following constraints
2547 * 1. The window can never be shrunk once it is offered (RFC 793)
2548 * 2. We limit memory per socket
2551 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2552 * RECV.NEXT + RCV.WIN fixed until:
2553 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2555 * i.e. don't raise the right edge of the window until you can raise
2556 * it at least MSS bytes.
2558 * Unfortunately, the recommended algorithm breaks header prediction,
2559 * since header prediction assumes th->window stays fixed.
2561 * Strictly speaking, keeping th->window fixed violates the receiver
2562 * side SWS prevention criteria. The problem is that under this rule
2563 * a stream of single byte packets will cause the right side of the
2564 * window to always advance by a single byte.
2566 * Of course, if the sender implements sender side SWS prevention
2567 * then this will not be a problem.
2569 * BSD seems to make the following compromise:
2571 * If the free space is less than the 1/4 of the maximum
2572 * space available and the free space is less than 1/2 mss,
2573 * then set the window to 0.
2574 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2575 * Otherwise, just prevent the window from shrinking
2576 * and from being larger than the largest representable value.
2578 * This prevents incremental opening of the window in the regime
2579 * where TCP is limited by the speed of the reader side taking
2580 * data out of the TCP receive queue. It does nothing about
2581 * those cases where the window is constrained on the sender side
2582 * because the pipeline is full.
2584 * BSD also seems to "accidentally" limit itself to windows that are a
2585 * multiple of MSS, at least until the free space gets quite small.
2586 * This would appear to be a side effect of the mbuf implementation.
2587 * Combining these two algorithms results in the observed behavior
2588 * of having a fixed window size at almost all times.
2590 * Below we obtain similar behavior by forcing the offered window to
2591 * a multiple of the mss when it is feasible to do so.
2593 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2594 * Regular options like TIMESTAMP are taken into account.
2596 u32 __tcp_select_window(struct sock *sk)
2598 struct inet_connection_sock *icsk = inet_csk(sk);
2599 struct tcp_sock *tp = tcp_sk(sk);
2600 /* MSS for the peer's data. Previous versions used mss_clamp
2601 * here. I don't know if the value based on our guesses
2602 * of peer's MSS is better for the performance. It's more correct
2603 * but may be worse for the performance because of rcv_mss
2604 * fluctuations. --SAW 1998/11/1
2606 int mss = icsk->icsk_ack.rcv_mss;
2607 int free_space = tcp_space(sk);
2608 int allowed_space = tcp_full_space(sk);
2609 int full_space = min_t(int, tp->window_clamp, allowed_space);
2612 if (unlikely(mss > full_space)) {
2617 if (free_space < (full_space >> 1)) {
2618 icsk->icsk_ack.quick = 0;
2620 if (tcp_under_memory_pressure(sk))
2621 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2624 /* free_space might become our new window, make sure we don't
2625 * increase it due to wscale.
2627 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2629 /* if free space is less than mss estimate, or is below 1/16th
2630 * of the maximum allowed, try to move to zero-window, else
2631 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2632 * new incoming data is dropped due to memory limits.
2633 * With large window, mss test triggers way too late in order
2634 * to announce zero window in time before rmem limit kicks in.
2636 if (free_space < (allowed_space >> 4) || free_space < mss)
2640 if (free_space > tp->rcv_ssthresh)
2641 free_space = tp->rcv_ssthresh;
2643 /* Don't do rounding if we are using window scaling, since the
2644 * scaled window will not line up with the MSS boundary anyway.
2646 if (tp->rx_opt.rcv_wscale) {
2647 window = free_space;
2649 /* Advertise enough space so that it won't get scaled away.
2650 * Import case: prevent zero window announcement if
2651 * 1<<rcv_wscale > mss.
2653 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2655 window = tp->rcv_wnd;
2656 /* Get the largest window that is a nice multiple of mss.
2657 * Window clamp already applied above.
2658 * If our current window offering is within 1 mss of the
2659 * free space we just keep it. This prevents the divide
2660 * and multiply from happening most of the time.
2661 * We also don't do any window rounding when the free space
2664 if (window <= free_space - mss || window > free_space)
2665 window = rounddown(free_space, mss);
2666 else if (mss == full_space &&
2667 free_space > window + (full_space >> 1))
2668 window = free_space;
2674 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2675 const struct sk_buff *next_skb)
2677 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2678 const struct skb_shared_info *next_shinfo =
2679 skb_shinfo(next_skb);
2680 struct skb_shared_info *shinfo = skb_shinfo(skb);
2682 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2683 shinfo->tskey = next_shinfo->tskey;
2684 TCP_SKB_CB(skb)->txstamp_ack |=
2685 TCP_SKB_CB(next_skb)->txstamp_ack;
2689 /* Collapses two adjacent SKB's during retransmission. */
2690 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2692 struct tcp_sock *tp = tcp_sk(sk);
2693 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2694 int skb_size, next_skb_size;
2696 skb_size = skb->len;
2697 next_skb_size = next_skb->len;
2699 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2701 if (next_skb_size) {
2702 if (next_skb_size <= skb_availroom(skb))
2703 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2705 else if (!skb_shift(skb, next_skb, next_skb_size))
2708 tcp_highest_sack_combine(sk, next_skb, skb);
2710 tcp_unlink_write_queue(next_skb, sk);
2712 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2713 skb->ip_summed = CHECKSUM_PARTIAL;
2715 if (skb->ip_summed != CHECKSUM_PARTIAL)
2716 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2718 /* Update sequence range on original skb. */
2719 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2721 /* Merge over control information. This moves PSH/FIN etc. over */
2722 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2724 /* All done, get rid of second SKB and account for it so
2725 * packet counting does not break.
2727 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2728 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2730 /* changed transmit queue under us so clear hints */
2731 tcp_clear_retrans_hints_partial(tp);
2732 if (next_skb == tp->retransmit_skb_hint)
2733 tp->retransmit_skb_hint = skb;
2735 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2737 tcp_skb_collapse_tstamp(skb, next_skb);
2739 sk_wmem_free_skb(sk, next_skb);
2743 /* Check if coalescing SKBs is legal. */
2744 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2746 if (tcp_skb_pcount(skb) > 1)
2748 if (skb_cloned(skb))
2750 if (skb == tcp_send_head(sk))
2752 /* Some heuristics for collapsing over SACK'd could be invented */
2753 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2759 /* Collapse packets in the retransmit queue to make to create
2760 * less packets on the wire. This is only done on retransmission.
2762 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2765 struct tcp_sock *tp = tcp_sk(sk);
2766 struct sk_buff *skb = to, *tmp;
2769 if (!sysctl_tcp_retrans_collapse)
2771 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2774 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2775 if (!tcp_can_collapse(sk, skb))
2778 if (!tcp_skb_can_collapse_to(to))
2791 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2794 if (!tcp_collapse_retrans(sk, to))
2799 /* This retransmits one SKB. Policy decisions and retransmit queue
2800 * state updates are done by the caller. Returns non-zero if an
2801 * error occurred which prevented the send.
2803 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2805 struct inet_connection_sock *icsk = inet_csk(sk);
2806 struct tcp_sock *tp = tcp_sk(sk);
2807 unsigned int cur_mss;
2811 /* Inconclusive MTU probe */
2812 if (icsk->icsk_mtup.probe_size)
2813 icsk->icsk_mtup.probe_size = 0;
2815 /* Do not sent more than we queued. 1/4 is reserved for possible
2816 * copying overhead: fragmentation, tunneling, mangling etc.
2818 if (refcount_read(&sk->sk_wmem_alloc) >
2819 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2823 if (skb_still_in_host_queue(sk, skb))
2826 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2827 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2829 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2833 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2834 return -EHOSTUNREACH; /* Routing failure or similar. */
2836 cur_mss = tcp_current_mss(sk);
2838 /* If receiver has shrunk his window, and skb is out of
2839 * new window, do not retransmit it. The exception is the
2840 * case, when window is shrunk to zero. In this case
2841 * our retransmit serves as a zero window probe.
2843 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2844 TCP_SKB_CB(skb)->seq != tp->snd_una)
2847 len = cur_mss * segs;
2848 if (skb->len > len) {
2849 if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC))
2850 return -ENOMEM; /* We'll try again later. */
2852 if (skb_unclone(skb, GFP_ATOMIC))
2855 diff = tcp_skb_pcount(skb);
2856 tcp_set_skb_tso_segs(skb, cur_mss);
2857 diff -= tcp_skb_pcount(skb);
2859 tcp_adjust_pcount(sk, skb, diff);
2860 if (skb->len < cur_mss)
2861 tcp_retrans_try_collapse(sk, skb, cur_mss);
2864 /* RFC3168, section 6.1.1.1. ECN fallback */
2865 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2866 tcp_ecn_clear_syn(sk, skb);
2868 /* Update global and local TCP statistics. */
2869 segs = tcp_skb_pcount(skb);
2870 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2871 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2872 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2873 tp->total_retrans += segs;
2875 /* make sure skb->data is aligned on arches that require it
2876 * and check if ack-trimming & collapsing extended the headroom
2877 * beyond what csum_start can cover.
2879 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2880 skb_headroom(skb) >= 0xFFFF)) {
2881 struct sk_buff *nskb;
2883 skb->skb_mstamp = tp->tcp_mstamp;
2884 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2885 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2888 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2892 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2893 } else if (err != -EBUSY) {
2894 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2899 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2901 struct tcp_sock *tp = tcp_sk(sk);
2902 int err = __tcp_retransmit_skb(sk, skb, segs);
2905 #if FASTRETRANS_DEBUG > 0
2906 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2907 net_dbg_ratelimited("retrans_out leaked\n");
2910 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2911 tp->retrans_out += tcp_skb_pcount(skb);
2913 /* Save stamp of the first retransmit. */
2914 if (!tp->retrans_stamp)
2915 tp->retrans_stamp = tcp_skb_timestamp(skb);
2919 if (tp->undo_retrans < 0)
2920 tp->undo_retrans = 0;
2921 tp->undo_retrans += tcp_skb_pcount(skb);
2925 /* This gets called after a retransmit timeout, and the initially
2926 * retransmitted data is acknowledged. It tries to continue
2927 * resending the rest of the retransmit queue, until either
2928 * we've sent it all or the congestion window limit is reached.
2929 * If doing SACK, the first ACK which comes back for a timeout
2930 * based retransmit packet might feed us FACK information again.
2931 * If so, we use it to avoid unnecessarily retransmissions.
2933 void tcp_xmit_retransmit_queue(struct sock *sk)
2935 const struct inet_connection_sock *icsk = inet_csk(sk);
2936 struct tcp_sock *tp = tcp_sk(sk);
2937 struct sk_buff *skb;
2938 struct sk_buff *hole = NULL;
2942 if (!tp->packets_out)
2945 if (tp->retransmit_skb_hint) {
2946 skb = tp->retransmit_skb_hint;
2948 skb = tcp_write_queue_head(sk);
2951 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2952 tcp_for_write_queue_from(skb, sk) {
2956 if (skb == tcp_send_head(sk))
2959 if (tcp_pacing_check(sk))
2962 /* we could do better than to assign each time */
2964 tp->retransmit_skb_hint = skb;
2966 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2969 sacked = TCP_SKB_CB(skb)->sacked;
2970 /* In case tcp_shift_skb_data() have aggregated large skbs,
2971 * we need to make sure not sending too bigs TSO packets
2973 segs = min_t(int, segs, max_segs);
2975 if (tp->retrans_out >= tp->lost_out) {
2977 } else if (!(sacked & TCPCB_LOST)) {
2978 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2983 if (icsk->icsk_ca_state != TCP_CA_Loss)
2984 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2986 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2989 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2992 if (tcp_small_queue_check(sk, skb, 1))
2995 if (tcp_retransmit_skb(sk, skb, segs))
2998 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3000 if (tcp_in_cwnd_reduction(sk))
3001 tp->prr_out += tcp_skb_pcount(skb);
3003 if (skb == tcp_write_queue_head(sk) &&
3004 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3005 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3006 inet_csk(sk)->icsk_rto,
3011 /* We allow to exceed memory limits for FIN packets to expedite
3012 * connection tear down and (memory) recovery.
3013 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3014 * or even be forced to close flow without any FIN.
3015 * In general, we want to allow one skb per socket to avoid hangs
3016 * with edge trigger epoll()
3018 void sk_forced_mem_schedule(struct sock *sk, int size)
3022 if (size <= sk->sk_forward_alloc)
3024 amt = sk_mem_pages(size);
3025 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3026 sk_memory_allocated_add(sk, amt);
3028 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3029 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3032 /* Send a FIN. The caller locks the socket for us.
3033 * We should try to send a FIN packet really hard, but eventually give up.
3035 void tcp_send_fin(struct sock *sk)
3037 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3038 struct tcp_sock *tp = tcp_sk(sk);
3040 /* Optimization, tack on the FIN if we have one skb in write queue and
3041 * this skb was not yet sent, or we are under memory pressure.
3042 * Note: in the latter case, FIN packet will be sent after a timeout,
3043 * as TCP stack thinks it has already been transmitted.
3045 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
3047 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3048 TCP_SKB_CB(tskb)->end_seq++;
3050 if (!tcp_send_head(sk)) {
3051 /* This means tskb was already sent.
3052 * Pretend we included the FIN on previous transmit.
3053 * We need to set tp->snd_nxt to the value it would have
3054 * if FIN had been sent. This is because retransmit path
3055 * does not change tp->snd_nxt.
3061 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3062 if (unlikely(!skb)) {
3067 skb_reserve(skb, MAX_TCP_HEADER);
3068 sk_forced_mem_schedule(sk, skb->truesize);
3069 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3070 tcp_init_nondata_skb(skb, tp->write_seq,
3071 TCPHDR_ACK | TCPHDR_FIN);
3072 tcp_queue_skb(sk, skb);
3074 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3077 /* We get here when a process closes a file descriptor (either due to
3078 * an explicit close() or as a byproduct of exit()'ing) and there
3079 * was unread data in the receive queue. This behavior is recommended
3080 * by RFC 2525, section 2.17. -DaveM
3082 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3084 struct sk_buff *skb;
3086 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3088 /* NOTE: No TCP options attached and we never retransmit this. */
3089 skb = alloc_skb(MAX_TCP_HEADER, priority);
3091 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3095 /* Reserve space for headers and prepare control bits. */
3096 skb_reserve(skb, MAX_TCP_HEADER);
3097 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3098 TCPHDR_ACK | TCPHDR_RST);
3099 tcp_mstamp_refresh(tcp_sk(sk));
3101 if (tcp_transmit_skb(sk, skb, 0, priority))
3102 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3105 /* Send a crossed SYN-ACK during socket establishment.
3106 * WARNING: This routine must only be called when we have already sent
3107 * a SYN packet that crossed the incoming SYN that caused this routine
3108 * to get called. If this assumption fails then the initial rcv_wnd
3109 * and rcv_wscale values will not be correct.
3111 int tcp_send_synack(struct sock *sk)
3113 struct sk_buff *skb;
3115 skb = tcp_write_queue_head(sk);
3116 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3117 pr_debug("%s: wrong queue state\n", __func__);
3120 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3121 if (skb_cloned(skb)) {
3122 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
3125 tcp_unlink_write_queue(skb, sk);
3126 __skb_header_release(nskb);
3127 __tcp_add_write_queue_head(sk, nskb);
3128 sk_wmem_free_skb(sk, skb);
3129 sk->sk_wmem_queued += nskb->truesize;
3130 sk_mem_charge(sk, nskb->truesize);
3134 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3135 tcp_ecn_send_synack(sk, skb);
3137 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3141 * tcp_make_synack - Prepare a SYN-ACK.
3142 * sk: listener socket
3143 * dst: dst entry attached to the SYNACK
3144 * req: request_sock pointer
3146 * Allocate one skb and build a SYNACK packet.
3147 * @dst is consumed : Caller should not use it again.
3149 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3150 struct request_sock *req,
3151 struct tcp_fastopen_cookie *foc,
3152 enum tcp_synack_type synack_type)
3154 struct inet_request_sock *ireq = inet_rsk(req);
3155 const struct tcp_sock *tp = tcp_sk(sk);
3156 struct tcp_md5sig_key *md5 = NULL;
3157 struct tcp_out_options opts;
3158 struct sk_buff *skb;
3159 int tcp_header_size;
3163 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3164 if (unlikely(!skb)) {
3168 /* Reserve space for headers. */
3169 skb_reserve(skb, MAX_TCP_HEADER);
3171 switch (synack_type) {
3172 case TCP_SYNACK_NORMAL:
3173 skb_set_owner_w(skb, req_to_sk(req));
3175 case TCP_SYNACK_COOKIE:
3176 /* Under synflood, we do not attach skb to a socket,
3177 * to avoid false sharing.
3180 case TCP_SYNACK_FASTOPEN:
3181 /* sk is a const pointer, because we want to express multiple
3182 * cpu might call us concurrently.
3183 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3185 skb_set_owner_w(skb, (struct sock *)sk);
3188 skb_dst_set(skb, dst);
3190 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3192 memset(&opts, 0, sizeof(opts));
3193 #ifdef CONFIG_SYN_COOKIES
3194 if (unlikely(req->cookie_ts))
3195 skb->skb_mstamp = cookie_init_timestamp(req);
3198 skb->skb_mstamp = tcp_clock_us();
3200 #ifdef CONFIG_TCP_MD5SIG
3202 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3204 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3205 tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
3208 skb_push(skb, tcp_header_size);
3209 skb_reset_transport_header(skb);
3211 th = (struct tcphdr *)skb->data;
3212 memset(th, 0, sizeof(struct tcphdr));
3215 tcp_ecn_make_synack(req, th);
3216 th->source = htons(ireq->ir_num);
3217 th->dest = ireq->ir_rmt_port;
3218 skb->mark = ireq->ir_mark;
3219 /* Setting of flags are superfluous here for callers (and ECE is
3220 * not even correctly set)
3222 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3223 TCPHDR_SYN | TCPHDR_ACK);
3225 th->seq = htonl(TCP_SKB_CB(skb)->seq);
3226 /* XXX data is queued and acked as is. No buffer/window check */
3227 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3229 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3230 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3231 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3232 th->doff = (tcp_header_size >> 2);
3233 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3235 #ifdef CONFIG_TCP_MD5SIG
3236 /* Okay, we have all we need - do the md5 hash if needed */
3238 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3239 md5, req_to_sk(req), skb);
3243 /* Do not fool tcpdump (if any), clean our debris */
3247 EXPORT_SYMBOL(tcp_make_synack);
3249 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3251 struct inet_connection_sock *icsk = inet_csk(sk);
3252 const struct tcp_congestion_ops *ca;
3253 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3255 if (ca_key == TCP_CA_UNSPEC)
3259 ca = tcp_ca_find_key(ca_key);
3260 if (likely(ca && try_module_get(ca->owner))) {
3261 module_put(icsk->icsk_ca_ops->owner);
3262 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3263 icsk->icsk_ca_ops = ca;
3268 /* Do all connect socket setups that can be done AF independent. */
3269 static void tcp_connect_init(struct sock *sk)
3271 const struct dst_entry *dst = __sk_dst_get(sk);
3272 struct tcp_sock *tp = tcp_sk(sk);
3276 /* We'll fix this up when we get a response from the other end.
3277 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3279 tp->tcp_header_len = sizeof(struct tcphdr);
3280 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3281 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3283 #ifdef CONFIG_TCP_MD5SIG
3284 if (tp->af_specific->md5_lookup(sk, sk))
3285 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3288 /* If user gave his TCP_MAXSEG, record it to clamp */
3289 if (tp->rx_opt.user_mss)
3290 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3293 tcp_sync_mss(sk, dst_mtu(dst));
3295 tcp_ca_dst_init(sk, dst);
3297 if (!tp->window_clamp)
3298 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3299 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3301 tcp_initialize_rcv_mss(sk);
3303 /* limit the window selection if the user enforce a smaller rx buffer */
3304 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3305 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3306 tp->window_clamp = tcp_full_space(sk);
3308 rcv_wnd = tcp_rwnd_init_bpf(sk);
3310 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3312 tcp_select_initial_window(tcp_full_space(sk),
3313 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3316 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3320 tp->rx_opt.rcv_wscale = rcv_wscale;
3321 tp->rcv_ssthresh = tp->rcv_wnd;
3324 sock_reset_flag(sk, SOCK_DONE);
3327 tp->snd_una = tp->write_seq;
3328 tp->snd_sml = tp->write_seq;
3329 tp->snd_up = tp->write_seq;
3330 tp->snd_nxt = tp->write_seq;
3332 if (likely(!tp->repair))
3335 tp->rcv_tstamp = tcp_jiffies32;
3336 tp->rcv_wup = tp->rcv_nxt;
3337 tp->copied_seq = tp->rcv_nxt;
3339 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3340 inet_csk(sk)->icsk_retransmits = 0;
3341 tcp_clear_retrans(tp);
3344 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3346 struct tcp_sock *tp = tcp_sk(sk);
3347 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3349 tcb->end_seq += skb->len;
3350 __skb_header_release(skb);
3351 __tcp_add_write_queue_tail(sk, skb);
3352 sk->sk_wmem_queued += skb->truesize;
3353 sk_mem_charge(sk, skb->truesize);
3354 tp->write_seq = tcb->end_seq;
3355 tp->packets_out += tcp_skb_pcount(skb);
3358 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3359 * queue a data-only packet after the regular SYN, such that regular SYNs
3360 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3361 * only the SYN sequence, the data are retransmitted in the first ACK.
3362 * If cookie is not cached or other error occurs, falls back to send a
3363 * regular SYN with Fast Open cookie request option.
3365 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3367 struct tcp_sock *tp = tcp_sk(sk);
3368 struct tcp_fastopen_request *fo = tp->fastopen_req;
3370 struct sk_buff *syn_data;
3372 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3373 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3376 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3377 * user-MSS. Reserve maximum option space for middleboxes that add
3378 * private TCP options. The cost is reduced data space in SYN :(
3380 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3382 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3383 MAX_TCP_OPTION_SPACE;
3385 space = min_t(size_t, space, fo->size);
3387 /* limit to order-0 allocations */
3388 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3390 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3393 syn_data->ip_summed = CHECKSUM_PARTIAL;
3394 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3396 int copied = copy_from_iter(skb_put(syn_data, space), space,
3397 &fo->data->msg_iter);
3398 if (unlikely(!copied)) {
3399 kfree_skb(syn_data);
3402 if (copied != space) {
3403 skb_trim(syn_data, copied);
3407 /* No more data pending in inet_wait_for_connect() */
3408 if (space == fo->size)
3412 tcp_connect_queue_skb(sk, syn_data);
3414 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3416 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3418 syn->skb_mstamp = syn_data->skb_mstamp;
3420 /* Now full SYN+DATA was cloned and sent (or not),
3421 * remove the SYN from the original skb (syn_data)
3422 * we keep in write queue in case of a retransmit, as we
3423 * also have the SYN packet (with no data) in the same queue.
3425 TCP_SKB_CB(syn_data)->seq++;
3426 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3428 tp->syn_data = (fo->copied > 0);
3429 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3434 /* Send a regular SYN with Fast Open cookie request option */
3435 if (fo->cookie.len > 0)
3437 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3439 tp->syn_fastopen = 0;
3441 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3445 /* Build a SYN and send it off. */
3446 int tcp_connect(struct sock *sk)
3448 struct tcp_sock *tp = tcp_sk(sk);
3449 struct sk_buff *buff;
3452 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB);
3453 tcp_connect_init(sk);
3455 if (unlikely(tp->repair)) {
3456 tcp_finish_connect(sk, NULL);
3460 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3461 if (unlikely(!buff))
3464 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3465 tcp_mstamp_refresh(tp);
3466 tp->retrans_stamp = tcp_time_stamp(tp);
3467 tcp_connect_queue_skb(sk, buff);
3468 tcp_ecn_send_syn(sk, buff);
3470 /* Send off SYN; include data in Fast Open. */
3471 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3472 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3473 if (err == -ECONNREFUSED)
3476 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3477 * in order to make this packet get counted in tcpOutSegs.
3479 tp->snd_nxt = tp->write_seq;
3480 tp->pushed_seq = tp->write_seq;
3481 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3483 /* Timer for repeating the SYN until an answer. */
3484 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3485 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3488 EXPORT_SYMBOL(tcp_connect);
3490 /* Send out a delayed ack, the caller does the policy checking
3491 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3494 void tcp_send_delayed_ack(struct sock *sk)
3496 struct inet_connection_sock *icsk = inet_csk(sk);
3497 int ato = icsk->icsk_ack.ato;
3498 unsigned long timeout;
3500 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3502 if (ato > TCP_DELACK_MIN) {
3503 const struct tcp_sock *tp = tcp_sk(sk);
3504 int max_ato = HZ / 2;
3506 if (icsk->icsk_ack.pingpong ||
3507 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3508 max_ato = TCP_DELACK_MAX;
3510 /* Slow path, intersegment interval is "high". */
3512 /* If some rtt estimate is known, use it to bound delayed ack.
3513 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3517 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3524 ato = min(ato, max_ato);
3527 /* Stay within the limit we were given */
3528 timeout = jiffies + ato;
3530 /* Use new timeout only if there wasn't a older one earlier. */
3531 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3532 /* If delack timer was blocked or is about to expire,
3535 if (icsk->icsk_ack.blocked ||
3536 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3541 if (!time_before(timeout, icsk->icsk_ack.timeout))
3542 timeout = icsk->icsk_ack.timeout;
3544 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3545 icsk->icsk_ack.timeout = timeout;
3546 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3549 /* This routine sends an ack and also updates the window. */
3550 void tcp_send_ack(struct sock *sk)
3552 struct sk_buff *buff;
3554 /* If we have been reset, we may not send again. */
3555 if (sk->sk_state == TCP_CLOSE)
3558 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3560 /* We are not putting this on the write queue, so
3561 * tcp_transmit_skb() will set the ownership to this
3564 buff = alloc_skb(MAX_TCP_HEADER,
3565 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3566 if (unlikely(!buff)) {
3567 inet_csk_schedule_ack(sk);
3568 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3569 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3570 TCP_DELACK_MAX, TCP_RTO_MAX);
3574 /* Reserve space for headers and prepare control bits. */
3575 skb_reserve(buff, MAX_TCP_HEADER);
3576 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3578 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3580 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3582 skb_set_tcp_pure_ack(buff);
3584 /* Send it off, this clears delayed acks for us. */
3585 tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3587 EXPORT_SYMBOL_GPL(tcp_send_ack);
3589 /* This routine sends a packet with an out of date sequence
3590 * number. It assumes the other end will try to ack it.
3592 * Question: what should we make while urgent mode?
3593 * 4.4BSD forces sending single byte of data. We cannot send
3594 * out of window data, because we have SND.NXT==SND.MAX...
3596 * Current solution: to send TWO zero-length segments in urgent mode:
3597 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3598 * out-of-date with SND.UNA-1 to probe window.
3600 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3602 struct tcp_sock *tp = tcp_sk(sk);
3603 struct sk_buff *skb;
3605 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3606 skb = alloc_skb(MAX_TCP_HEADER,
3607 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3611 /* Reserve space for headers and set control bits. */
3612 skb_reserve(skb, MAX_TCP_HEADER);
3613 /* Use a previous sequence. This should cause the other
3614 * end to send an ack. Don't queue or clone SKB, just
3617 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3618 NET_INC_STATS(sock_net(sk), mib);
3619 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3622 /* Called from setsockopt( ... TCP_REPAIR ) */
3623 void tcp_send_window_probe(struct sock *sk)
3625 if (sk->sk_state == TCP_ESTABLISHED) {
3626 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3627 tcp_mstamp_refresh(tcp_sk(sk));
3628 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3632 /* Initiate keepalive or window probe from timer. */
3633 int tcp_write_wakeup(struct sock *sk, int mib)
3635 struct tcp_sock *tp = tcp_sk(sk);
3636 struct sk_buff *skb;
3638 if (sk->sk_state == TCP_CLOSE)
3641 skb = tcp_send_head(sk);
3642 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3644 unsigned int mss = tcp_current_mss(sk);
3645 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3647 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3648 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3650 /* We are probing the opening of a window
3651 * but the window size is != 0
3652 * must have been a result SWS avoidance ( sender )
3654 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3656 seg_size = min(seg_size, mss);
3657 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3658 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3660 } else if (!tcp_skb_pcount(skb))
3661 tcp_set_skb_tso_segs(skb, mss);
3663 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3664 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3666 tcp_event_new_data_sent(sk, skb);
3669 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3670 tcp_xmit_probe_skb(sk, 1, mib);
3671 return tcp_xmit_probe_skb(sk, 0, mib);
3675 /* A window probe timeout has occurred. If window is not closed send
3676 * a partial packet else a zero probe.
3678 void tcp_send_probe0(struct sock *sk)
3680 struct inet_connection_sock *icsk = inet_csk(sk);
3681 struct tcp_sock *tp = tcp_sk(sk);
3682 struct net *net = sock_net(sk);
3683 unsigned long probe_max;
3686 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3688 if (tp->packets_out || !tcp_send_head(sk)) {
3689 /* Cancel probe timer, if it is not required. */
3690 icsk->icsk_probes_out = 0;
3691 icsk->icsk_backoff = 0;
3696 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3697 icsk->icsk_backoff++;
3698 icsk->icsk_probes_out++;
3699 probe_max = TCP_RTO_MAX;
3701 /* If packet was not sent due to local congestion,
3702 * do not backoff and do not remember icsk_probes_out.
3703 * Let local senders to fight for local resources.
3705 * Use accumulated backoff yet.
3707 if (!icsk->icsk_probes_out)
3708 icsk->icsk_probes_out = 1;
3709 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3711 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3712 tcp_probe0_when(sk, probe_max),
3716 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3718 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3722 tcp_rsk(req)->txhash = net_tx_rndhash();
3723 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3725 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3726 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3727 if (unlikely(tcp_passive_fastopen(sk)))
3728 tcp_sk(sk)->total_retrans++;
3732 EXPORT_SYMBOL(tcp_rtx_synack);