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,
87 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
88 * window scaling factor due to loss of precision.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
94 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
96 const struct tcp_sock *tp = tcp_sk(sk);
98 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
99 (tp->rx_opt.wscale_ok &&
100 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
103 return tcp_wnd_end(tp);
106 /* Calculate mss to advertise in SYN segment.
107 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
109 * 1. It is independent of path mtu.
110 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
111 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
112 * attached devices, because some buggy hosts are confused by
114 * 4. We do not make 3, we advertise MSS, calculated from first
115 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
116 * This may be overridden via information stored in routing table.
117 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
118 * probably even Jumbo".
120 static __u16 tcp_advertise_mss(struct sock *sk)
122 struct tcp_sock *tp = tcp_sk(sk);
123 const struct dst_entry *dst = __sk_dst_get(sk);
124 int mss = tp->advmss;
127 unsigned int metric = dst_metric_advmss(dst);
138 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
139 * This is the first part of cwnd validation mechanism.
141 void tcp_cwnd_restart(struct sock *sk, s32 delta)
143 struct tcp_sock *tp = tcp_sk(sk);
144 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
145 u32 cwnd = tp->snd_cwnd;
147 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
149 tp->snd_ssthresh = tcp_current_ssthresh(sk);
150 restart_cwnd = min(restart_cwnd, cwnd);
152 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
154 tp->snd_cwnd = max(cwnd, restart_cwnd);
155 tp->snd_cwnd_stamp = tcp_jiffies32;
156 tp->snd_cwnd_used = 0;
159 /* Congestion state accounting after a packet has been sent. */
160 static void tcp_event_data_sent(struct tcp_sock *tp,
163 struct inet_connection_sock *icsk = inet_csk(sk);
164 const u32 now = tcp_jiffies32;
166 if (tcp_packets_in_flight(tp) == 0)
167 tcp_ca_event(sk, CA_EVENT_TX_START);
171 /* If it is a reply for ato after last received
172 * packet, enter pingpong mode.
174 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
175 icsk->icsk_ack.pingpong = 1;
178 /* Account for an ACK we sent. */
179 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
182 struct tcp_sock *tp = tcp_sk(sk);
184 if (unlikely(rcv_nxt != tp->rcv_nxt))
185 return; /* Special ACK sent by DCTCP to reflect ECN */
186 tcp_dec_quickack_mode(sk, pkts);
187 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
191 u32 tcp_default_init_rwnd(u32 mss)
193 /* Initial receive window should be twice of TCP_INIT_CWND to
194 * enable proper sending of new unsent data during fast recovery
195 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
196 * limit when mss is larger than 1460.
198 u32 init_rwnd = TCP_INIT_CWND * 2;
201 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
205 /* Determine a window scaling and initial window to offer.
206 * Based on the assumption that the given amount of space
207 * will be offered. Store the results in the tp structure.
208 * NOTE: for smooth operation initial space offering should
209 * be a multiple of mss if possible. We assume here that mss >= 1.
210 * This MUST be enforced by all callers.
212 void tcp_select_initial_window(int __space, __u32 mss,
213 __u32 *rcv_wnd, __u32 *window_clamp,
214 int wscale_ok, __u8 *rcv_wscale,
217 unsigned int space = (__space < 0 ? 0 : __space);
219 /* If no clamp set the clamp to the max possible scaled window */
220 if (*window_clamp == 0)
221 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
222 space = min(*window_clamp, space);
224 /* Quantize space offering to a multiple of mss if possible. */
226 space = rounddown(space, mss);
228 /* NOTE: offering an initial window larger than 32767
229 * will break some buggy TCP stacks. If the admin tells us
230 * it is likely we could be speaking with such a buggy stack
231 * we will truncate our initial window offering to 32K-1
232 * unless the remote has sent us a window scaling option,
233 * which we interpret as a sign the remote TCP is not
234 * misinterpreting the window field as a signed quantity.
236 if (sysctl_tcp_workaround_signed_windows)
237 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
243 /* Set window scaling on max possible window */
244 space = max_t(u32, space, sysctl_tcp_rmem[2]);
245 space = max_t(u32, space, sysctl_rmem_max);
246 space = min_t(u32, space, *window_clamp);
247 while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
253 if (mss > (1 << *rcv_wscale)) {
254 if (!init_rcv_wnd) /* Use default unless specified otherwise */
255 init_rcv_wnd = tcp_default_init_rwnd(mss);
256 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
259 /* Set the clamp no higher than max representable value */
260 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
262 EXPORT_SYMBOL(tcp_select_initial_window);
264 /* Chose a new window to advertise, update state in tcp_sock for the
265 * socket, and return result with RFC1323 scaling applied. The return
266 * value can be stuffed directly into th->window for an outgoing
269 static u16 tcp_select_window(struct sock *sk)
271 struct tcp_sock *tp = tcp_sk(sk);
272 u32 old_win = tp->rcv_wnd;
273 u32 cur_win = tcp_receive_window(tp);
274 u32 new_win = __tcp_select_window(sk);
276 /* Never shrink the offered window */
277 if (new_win < cur_win) {
278 /* Danger Will Robinson!
279 * Don't update rcv_wup/rcv_wnd here or else
280 * we will not be able to advertise a zero
281 * window in time. --DaveM
283 * Relax Will Robinson.
286 NET_INC_STATS(sock_net(sk),
287 LINUX_MIB_TCPWANTZEROWINDOWADV);
288 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
290 tp->rcv_wnd = new_win;
291 tp->rcv_wup = tp->rcv_nxt;
293 /* Make sure we do not exceed the maximum possible
296 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
297 new_win = min(new_win, MAX_TCP_WINDOW);
299 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
301 /* RFC1323 scaling applied */
302 new_win >>= tp->rx_opt.rcv_wscale;
304 /* If we advertise zero window, disable fast path. */
308 NET_INC_STATS(sock_net(sk),
309 LINUX_MIB_TCPTOZEROWINDOWADV);
310 } else if (old_win == 0) {
311 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
317 /* Packet ECN state for a SYN-ACK */
318 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
320 const struct tcp_sock *tp = tcp_sk(sk);
322 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
323 if (!(tp->ecn_flags & TCP_ECN_OK))
324 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
325 else if (tcp_ca_needs_ecn(sk) ||
326 tcp_bpf_ca_needs_ecn(sk))
330 /* Packet ECN state for a SYN. */
331 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
333 struct tcp_sock *tp = tcp_sk(sk);
334 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
335 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
336 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
339 const struct dst_entry *dst = __sk_dst_get(sk);
341 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
348 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
349 tp->ecn_flags = TCP_ECN_OK;
350 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
355 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
357 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
358 /* tp->ecn_flags are cleared at a later point in time when
359 * SYN ACK is ultimatively being received.
361 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
365 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
367 if (inet_rsk(req)->ecn_ok)
371 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
374 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
375 struct tcphdr *th, int tcp_header_len)
377 struct tcp_sock *tp = tcp_sk(sk);
379 if (tp->ecn_flags & TCP_ECN_OK) {
380 /* Not-retransmitted data segment: set ECT and inject CWR. */
381 if (skb->len != tcp_header_len &&
382 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
384 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
385 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
387 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
389 } else if (!tcp_ca_needs_ecn(sk)) {
390 /* ACK or retransmitted segment: clear ECT|CE */
391 INET_ECN_dontxmit(sk);
393 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
398 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
399 * auto increment end seqno.
401 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
403 skb->ip_summed = CHECKSUM_PARTIAL;
406 TCP_SKB_CB(skb)->tcp_flags = flags;
407 TCP_SKB_CB(skb)->sacked = 0;
409 tcp_skb_pcount_set(skb, 1);
411 TCP_SKB_CB(skb)->seq = seq;
412 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
414 TCP_SKB_CB(skb)->end_seq = seq;
417 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
419 return tp->snd_una != tp->snd_up;
422 #define OPTION_SACK_ADVERTISE (1 << 0)
423 #define OPTION_TS (1 << 1)
424 #define OPTION_MD5 (1 << 2)
425 #define OPTION_WSCALE (1 << 3)
426 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
428 struct tcp_out_options {
429 u16 options; /* bit field of OPTION_* */
430 u16 mss; /* 0 to disable */
431 u8 ws; /* window scale, 0 to disable */
432 u8 num_sack_blocks; /* number of SACK blocks to include */
433 u8 hash_size; /* bytes in hash_location */
434 __u8 *hash_location; /* temporary pointer, overloaded */
435 __u32 tsval, tsecr; /* need to include OPTION_TS */
436 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
439 /* Write previously computed TCP options to the packet.
441 * Beware: Something in the Internet is very sensitive to the ordering of
442 * TCP options, we learned this through the hard way, so be careful here.
443 * Luckily we can at least blame others for their non-compliance but from
444 * inter-operability perspective it seems that we're somewhat stuck with
445 * the ordering which we have been using if we want to keep working with
446 * those broken things (not that it currently hurts anybody as there isn't
447 * particular reason why the ordering would need to be changed).
449 * At least SACK_PERM as the first option is known to lead to a disaster
450 * (but it may well be that other scenarios fail similarly).
452 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
453 struct tcp_out_options *opts)
455 u16 options = opts->options; /* mungable copy */
457 if (unlikely(OPTION_MD5 & options)) {
458 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
459 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
460 /* overload cookie hash location */
461 opts->hash_location = (__u8 *)ptr;
465 if (unlikely(opts->mss)) {
466 *ptr++ = htonl((TCPOPT_MSS << 24) |
467 (TCPOLEN_MSS << 16) |
471 if (likely(OPTION_TS & options)) {
472 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
473 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
474 (TCPOLEN_SACK_PERM << 16) |
475 (TCPOPT_TIMESTAMP << 8) |
477 options &= ~OPTION_SACK_ADVERTISE;
479 *ptr++ = htonl((TCPOPT_NOP << 24) |
481 (TCPOPT_TIMESTAMP << 8) |
484 *ptr++ = htonl(opts->tsval);
485 *ptr++ = htonl(opts->tsecr);
488 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
489 *ptr++ = htonl((TCPOPT_NOP << 24) |
491 (TCPOPT_SACK_PERM << 8) |
495 if (unlikely(OPTION_WSCALE & options)) {
496 *ptr++ = htonl((TCPOPT_NOP << 24) |
497 (TCPOPT_WINDOW << 16) |
498 (TCPOLEN_WINDOW << 8) |
502 if (unlikely(opts->num_sack_blocks)) {
503 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
504 tp->duplicate_sack : tp->selective_acks;
507 *ptr++ = htonl((TCPOPT_NOP << 24) |
510 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
511 TCPOLEN_SACK_PERBLOCK)));
513 for (this_sack = 0; this_sack < opts->num_sack_blocks;
515 *ptr++ = htonl(sp[this_sack].start_seq);
516 *ptr++ = htonl(sp[this_sack].end_seq);
519 tp->rx_opt.dsack = 0;
522 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
523 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
525 u32 len; /* Fast Open option length */
528 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
529 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
530 TCPOPT_FASTOPEN_MAGIC);
531 p += TCPOLEN_EXP_FASTOPEN_BASE;
533 len = TCPOLEN_FASTOPEN_BASE + foc->len;
534 *p++ = TCPOPT_FASTOPEN;
538 memcpy(p, foc->val, foc->len);
539 if ((len & 3) == 2) {
540 p[foc->len] = TCPOPT_NOP;
541 p[foc->len + 1] = TCPOPT_NOP;
543 ptr += (len + 3) >> 2;
547 /* Compute TCP options for SYN packets. This is not the final
548 * network wire format yet.
550 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
551 struct tcp_out_options *opts,
552 struct tcp_md5sig_key **md5)
554 struct tcp_sock *tp = tcp_sk(sk);
555 unsigned int remaining = MAX_TCP_OPTION_SPACE;
556 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
558 #ifdef CONFIG_TCP_MD5SIG
559 *md5 = tp->af_specific->md5_lookup(sk, sk);
561 opts->options |= OPTION_MD5;
562 remaining -= TCPOLEN_MD5SIG_ALIGNED;
568 /* We always get an MSS option. The option bytes which will be seen in
569 * normal data packets should timestamps be used, must be in the MSS
570 * advertised. But we subtract them from tp->mss_cache so that
571 * calculations in tcp_sendmsg are simpler etc. So account for this
572 * fact here if necessary. If we don't do this correctly, as a
573 * receiver we won't recognize data packets as being full sized when we
574 * should, and thus we won't abide by the delayed ACK rules correctly.
575 * SACKs don't matter, we never delay an ACK when we have any of those
577 opts->mss = tcp_advertise_mss(sk);
578 remaining -= TCPOLEN_MSS_ALIGNED;
580 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
581 opts->options |= OPTION_TS;
582 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
583 opts->tsecr = tp->rx_opt.ts_recent;
584 remaining -= TCPOLEN_TSTAMP_ALIGNED;
586 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
587 opts->ws = tp->rx_opt.rcv_wscale;
588 opts->options |= OPTION_WSCALE;
589 remaining -= TCPOLEN_WSCALE_ALIGNED;
591 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
592 opts->options |= OPTION_SACK_ADVERTISE;
593 if (unlikely(!(OPTION_TS & opts->options)))
594 remaining -= TCPOLEN_SACKPERM_ALIGNED;
597 if (fastopen && fastopen->cookie.len >= 0) {
598 u32 need = fastopen->cookie.len;
600 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
601 TCPOLEN_FASTOPEN_BASE;
602 need = (need + 3) & ~3U; /* Align to 32 bits */
603 if (remaining >= need) {
604 opts->options |= OPTION_FAST_OPEN_COOKIE;
605 opts->fastopen_cookie = &fastopen->cookie;
607 tp->syn_fastopen = 1;
608 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
612 return MAX_TCP_OPTION_SPACE - remaining;
615 /* Set up TCP options for SYN-ACKs. */
616 static unsigned int tcp_synack_options(struct request_sock *req,
617 unsigned int mss, struct sk_buff *skb,
618 struct tcp_out_options *opts,
619 const struct tcp_md5sig_key *md5,
620 struct tcp_fastopen_cookie *foc,
621 enum tcp_synack_type synack_type)
623 struct inet_request_sock *ireq = inet_rsk(req);
624 unsigned int remaining = MAX_TCP_OPTION_SPACE;
626 #ifdef CONFIG_TCP_MD5SIG
628 opts->options |= OPTION_MD5;
629 remaining -= TCPOLEN_MD5SIG_ALIGNED;
631 /* We can't fit any SACK blocks in a packet with MD5 + TS
632 * options. There was discussion about disabling SACK
633 * rather than TS in order to fit in better with old,
634 * buggy kernels, but that was deemed to be unnecessary.
636 if (synack_type != TCP_SYNACK_COOKIE)
637 ireq->tstamp_ok &= !ireq->sack_ok;
641 /* We always send an MSS option. */
643 remaining -= TCPOLEN_MSS_ALIGNED;
645 if (likely(ireq->wscale_ok)) {
646 opts->ws = ireq->rcv_wscale;
647 opts->options |= OPTION_WSCALE;
648 remaining -= TCPOLEN_WSCALE_ALIGNED;
650 if (likely(ireq->tstamp_ok)) {
651 opts->options |= OPTION_TS;
652 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
653 opts->tsecr = req->ts_recent;
654 remaining -= TCPOLEN_TSTAMP_ALIGNED;
656 if (likely(ireq->sack_ok)) {
657 opts->options |= OPTION_SACK_ADVERTISE;
658 if (unlikely(!ireq->tstamp_ok))
659 remaining -= TCPOLEN_SACKPERM_ALIGNED;
661 if (foc != NULL && foc->len >= 0) {
664 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
665 TCPOLEN_FASTOPEN_BASE;
666 need = (need + 3) & ~3U; /* Align to 32 bits */
667 if (remaining >= need) {
668 opts->options |= OPTION_FAST_OPEN_COOKIE;
669 opts->fastopen_cookie = foc;
674 return MAX_TCP_OPTION_SPACE - remaining;
677 /* Compute TCP options for ESTABLISHED sockets. This is not the
678 * final wire format yet.
680 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
681 struct tcp_out_options *opts,
682 struct tcp_md5sig_key **md5)
684 struct tcp_sock *tp = tcp_sk(sk);
685 unsigned int size = 0;
686 unsigned int eff_sacks;
690 #ifdef CONFIG_TCP_MD5SIG
691 *md5 = tp->af_specific->md5_lookup(sk, sk);
692 if (unlikely(*md5)) {
693 opts->options |= OPTION_MD5;
694 size += TCPOLEN_MD5SIG_ALIGNED;
700 if (likely(tp->rx_opt.tstamp_ok)) {
701 opts->options |= OPTION_TS;
702 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
703 opts->tsecr = tp->rx_opt.ts_recent;
704 size += TCPOLEN_TSTAMP_ALIGNED;
707 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
708 if (unlikely(eff_sacks)) {
709 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
710 opts->num_sack_blocks =
711 min_t(unsigned int, eff_sacks,
712 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
713 TCPOLEN_SACK_PERBLOCK);
714 if (likely(opts->num_sack_blocks))
715 size += TCPOLEN_SACK_BASE_ALIGNED +
716 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
723 /* TCP SMALL QUEUES (TSQ)
725 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
726 * to reduce RTT and bufferbloat.
727 * We do this using a special skb destructor (tcp_wfree).
729 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
730 * needs to be reallocated in a driver.
731 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
733 * Since transmit from skb destructor is forbidden, we use a tasklet
734 * to process all sockets that eventually need to send more skbs.
735 * We use one tasklet per cpu, with its own queue of sockets.
738 struct tasklet_struct tasklet;
739 struct list_head head; /* queue of tcp sockets */
741 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
743 static void tcp_tsq_handler(struct sock *sk)
745 if ((1 << sk->sk_state) &
746 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
747 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
748 struct tcp_sock *tp = tcp_sk(sk);
750 if (tp->lost_out > tp->retrans_out &&
751 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
752 tcp_mstamp_refresh(tp);
753 tcp_xmit_retransmit_queue(sk);
756 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
761 * One tasklet per cpu tries to send more skbs.
762 * We run in tasklet context but need to disable irqs when
763 * transferring tsq->head because tcp_wfree() might
764 * interrupt us (non NAPI drivers)
766 static void tcp_tasklet_func(unsigned long data)
768 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
771 struct list_head *q, *n;
775 local_irq_save(flags);
776 list_splice_init(&tsq->head, &list);
777 local_irq_restore(flags);
779 list_for_each_safe(q, n, &list) {
780 tp = list_entry(q, struct tcp_sock, tsq_node);
781 list_del(&tp->tsq_node);
783 sk = (struct sock *)tp;
784 smp_mb__before_atomic();
785 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
787 if (!sk->sk_lock.owned &&
788 test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) {
790 if (!sock_owned_by_user(sk)) {
791 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
801 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
802 TCPF_WRITE_TIMER_DEFERRED | \
803 TCPF_DELACK_TIMER_DEFERRED | \
804 TCPF_MTU_REDUCED_DEFERRED)
806 * tcp_release_cb - tcp release_sock() callback
809 * called from release_sock() to perform protocol dependent
810 * actions before socket release.
812 void tcp_release_cb(struct sock *sk)
814 unsigned long flags, nflags;
816 /* perform an atomic operation only if at least one flag is set */
818 flags = sk->sk_tsq_flags;
819 if (!(flags & TCP_DEFERRED_ALL))
821 nflags = flags & ~TCP_DEFERRED_ALL;
822 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
824 if (flags & TCPF_TSQ_DEFERRED)
827 /* Here begins the tricky part :
828 * We are called from release_sock() with :
830 * 2) sk_lock.slock spinlock held
831 * 3) socket owned by us (sk->sk_lock.owned == 1)
833 * But following code is meant to be called from BH handlers,
834 * so we should keep BH disabled, but early release socket ownership
836 sock_release_ownership(sk);
838 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
839 tcp_write_timer_handler(sk);
842 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
843 tcp_delack_timer_handler(sk);
846 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
847 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
851 EXPORT_SYMBOL(tcp_release_cb);
853 void __init tcp_tasklet_init(void)
857 for_each_possible_cpu(i) {
858 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
860 INIT_LIST_HEAD(&tsq->head);
861 tasklet_init(&tsq->tasklet,
868 * Write buffer destructor automatically called from kfree_skb.
869 * We can't xmit new skbs from this context, as we might already
872 void tcp_wfree(struct sk_buff *skb)
874 struct sock *sk = skb->sk;
875 struct tcp_sock *tp = tcp_sk(sk);
876 unsigned long flags, nval, oval;
878 /* Keep one reference on sk_wmem_alloc.
879 * Will be released by sk_free() from here or tcp_tasklet_func()
881 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
883 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
884 * Wait until our queues (qdisc + devices) are drained.
886 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
887 * - chance for incoming ACK (processed by another cpu maybe)
888 * to migrate this flow (skb->ooo_okay will be eventually set)
890 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
893 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
894 struct tsq_tasklet *tsq;
897 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
900 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
901 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
905 /* queue this socket to tasklet queue */
906 local_irq_save(flags);
907 tsq = this_cpu_ptr(&tsq_tasklet);
908 empty = list_empty(&tsq->head);
909 list_add(&tp->tsq_node, &tsq->head);
911 tasklet_schedule(&tsq->tasklet);
912 local_irq_restore(flags);
919 /* Note: Called under hard irq.
920 * We can not call TCP stack right away.
922 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
924 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
925 struct sock *sk = (struct sock *)tp;
926 unsigned long nval, oval;
928 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
929 struct tsq_tasklet *tsq;
932 if (oval & TSQF_QUEUED)
935 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
936 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
940 if (!refcount_inc_not_zero(&sk->sk_wmem_alloc))
942 /* queue this socket to tasklet queue */
943 tsq = this_cpu_ptr(&tsq_tasklet);
944 empty = list_empty(&tsq->head);
945 list_add(&tp->tsq_node, &tsq->head);
947 tasklet_schedule(&tsq->tasklet);
950 return HRTIMER_NORESTART;
953 /* BBR congestion control needs pacing.
954 * Same remark for SO_MAX_PACING_RATE.
955 * sch_fq packet scheduler is efficiently handling pacing,
956 * but is not always installed/used.
957 * Return true if TCP stack should pace packets itself.
959 static bool tcp_needs_internal_pacing(const struct sock *sk)
961 return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
964 static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
969 if (!tcp_needs_internal_pacing(sk))
971 rate = sk->sk_pacing_rate;
972 if (!rate || rate == ~0U)
975 /* Should account for header sizes as sch_fq does,
976 * but lets make things simple.
978 len_ns = (u64)skb->len * NSEC_PER_SEC;
979 do_div(len_ns, rate);
980 hrtimer_start(&tcp_sk(sk)->pacing_timer,
981 ktime_add_ns(ktime_get(), len_ns),
982 HRTIMER_MODE_ABS_PINNED);
985 /* This routine actually transmits TCP packets queued in by
986 * tcp_do_sendmsg(). This is used by both the initial
987 * transmission and possible later retransmissions.
988 * All SKB's seen here are completely headerless. It is our
989 * job to build the TCP header, and pass the packet down to
990 * IP so it can do the same plus pass the packet off to the
993 * We are working here with either a clone of the original
994 * SKB, or a fresh unique copy made by the retransmit engine.
996 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
997 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
999 const struct inet_connection_sock *icsk = inet_csk(sk);
1000 struct inet_sock *inet;
1001 struct tcp_sock *tp;
1002 struct tcp_skb_cb *tcb;
1003 struct tcp_out_options opts;
1004 unsigned int tcp_options_size, tcp_header_size;
1005 struct sk_buff *oskb = NULL;
1006 struct tcp_md5sig_key *md5;
1010 BUG_ON(!skb || !tcp_skb_pcount(skb));
1014 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1017 if (unlikely(skb_cloned(skb)))
1018 skb = pskb_copy(skb, gfp_mask);
1020 skb = skb_clone(skb, gfp_mask);
1024 skb->skb_mstamp = tp->tcp_mstamp;
1027 tcb = TCP_SKB_CB(skb);
1028 memset(&opts, 0, sizeof(opts));
1030 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1031 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1033 tcp_options_size = tcp_established_options(sk, skb, &opts,
1035 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1037 /* if no packet is in qdisc/device queue, then allow XPS to select
1038 * another queue. We can be called from tcp_tsq_handler()
1039 * which holds one reference to sk_wmem_alloc.
1041 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1042 * One way to get this would be to set skb->truesize = 2 on them.
1044 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1046 /* If we had to use memory reserve to allocate this skb,
1047 * this might cause drops if packet is looped back :
1048 * Other socket might not have SOCK_MEMALLOC.
1049 * Packets not looped back do not care about pfmemalloc.
1051 skb->pfmemalloc = 0;
1053 skb_push(skb, tcp_header_size);
1054 skb_reset_transport_header(skb);
1058 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1059 skb_set_hash_from_sk(skb, sk);
1060 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1062 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1064 /* Build TCP header and checksum it. */
1065 th = (struct tcphdr *)skb->data;
1066 th->source = inet->inet_sport;
1067 th->dest = inet->inet_dport;
1068 th->seq = htonl(tcb->seq);
1069 th->ack_seq = htonl(rcv_nxt);
1070 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1076 /* The urg_mode check is necessary during a below snd_una win probe */
1077 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1078 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1079 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1081 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1082 th->urg_ptr = htons(0xFFFF);
1087 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1088 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1089 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1090 th->window = htons(tcp_select_window(sk));
1091 tcp_ecn_send(sk, skb, th, tcp_header_size);
1093 /* RFC1323: The window in SYN & SYN/ACK segments
1096 th->window = htons(min(tp->rcv_wnd, 65535U));
1098 #ifdef CONFIG_TCP_MD5SIG
1099 /* Calculate the MD5 hash, as we have all we need now */
1101 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1102 tp->af_specific->calc_md5_hash(opts.hash_location,
1107 icsk->icsk_af_ops->send_check(sk, skb);
1109 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1110 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1112 if (skb->len != tcp_header_size) {
1113 tcp_event_data_sent(tp, sk);
1114 tp->data_segs_out += tcp_skb_pcount(skb);
1115 tcp_internal_pacing(sk, skb);
1118 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1119 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1120 tcp_skb_pcount(skb));
1122 tp->segs_out += tcp_skb_pcount(skb);
1123 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1124 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1125 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1127 /* Our usage of tstamp should remain private */
1130 /* Cleanup our debris for IP stacks */
1131 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1132 sizeof(struct inet6_skb_parm)));
1134 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1136 if (unlikely(err > 0)) {
1138 err = net_xmit_eval(err);
1141 oskb->skb_mstamp = tp->tcp_mstamp;
1142 tcp_rate_skb_sent(sk, oskb);
1147 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1150 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1151 tcp_sk(sk)->rcv_nxt);
1154 /* This routine just queues the buffer for sending.
1156 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1157 * otherwise socket can stall.
1159 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1161 struct tcp_sock *tp = tcp_sk(sk);
1163 /* Advance write_seq and place onto the write_queue. */
1164 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1165 __skb_header_release(skb);
1166 tcp_add_write_queue_tail(sk, skb);
1167 sk->sk_wmem_queued += skb->truesize;
1168 sk_mem_charge(sk, skb->truesize);
1171 /* Initialize TSO segments for a packet. */
1172 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1174 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1175 /* Avoid the costly divide in the normal
1178 tcp_skb_pcount_set(skb, 1);
1179 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1181 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1182 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1186 /* When a modification to fackets out becomes necessary, we need to check
1187 * skb is counted to fackets_out or not.
1189 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1192 struct tcp_sock *tp = tcp_sk(sk);
1194 if (!tp->sacked_out || tcp_is_reno(tp))
1197 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1198 tp->fackets_out -= decr;
1201 /* Pcount in the middle of the write queue got changed, we need to do various
1202 * tweaks to fix counters
1204 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1206 struct tcp_sock *tp = tcp_sk(sk);
1208 tp->packets_out -= decr;
1210 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1211 tp->sacked_out -= decr;
1212 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1213 tp->retrans_out -= decr;
1214 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1215 tp->lost_out -= decr;
1217 /* Reno case is special. Sigh... */
1218 if (tcp_is_reno(tp) && decr > 0)
1219 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1221 tcp_adjust_fackets_out(sk, skb, decr);
1223 if (tp->lost_skb_hint &&
1224 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1225 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1226 tp->lost_cnt_hint -= decr;
1228 tcp_verify_left_out(tp);
1231 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1233 return TCP_SKB_CB(skb)->txstamp_ack ||
1234 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1237 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1239 struct skb_shared_info *shinfo = skb_shinfo(skb);
1241 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1242 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1243 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1244 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1246 shinfo->tx_flags &= ~tsflags;
1247 shinfo2->tx_flags |= tsflags;
1248 swap(shinfo->tskey, shinfo2->tskey);
1249 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1250 TCP_SKB_CB(skb)->txstamp_ack = 0;
1254 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1256 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1257 TCP_SKB_CB(skb)->eor = 0;
1260 /* Function to create two new TCP segments. Shrinks the given segment
1261 * to the specified size and appends a new segment with the rest of the
1262 * packet to the list. This won't be called frequently, I hope.
1263 * Remember, these are still headerless SKBs at this point.
1265 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1266 unsigned int mss_now, gfp_t gfp)
1268 struct tcp_sock *tp = tcp_sk(sk);
1269 struct sk_buff *buff;
1270 int nsize, old_factor;
1275 if (WARN_ON(len > skb->len))
1278 nsize = skb_headlen(skb) - len;
1282 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1283 * We need some allowance to not penalize applications setting small
1285 * Also allow first and last skb in retransmit queue to be split.
1287 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1288 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1289 skb != tcp_rtx_queue_head(sk) &&
1290 skb != tcp_rtx_queue_tail(sk))) {
1291 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1295 if (skb_unclone(skb, gfp))
1298 /* Get a new skb... force flag on. */
1299 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1301 return -ENOMEM; /* We'll just try again later. */
1303 sk->sk_wmem_queued += buff->truesize;
1304 sk_mem_charge(sk, buff->truesize);
1305 nlen = skb->len - len - nsize;
1306 buff->truesize += nlen;
1307 skb->truesize -= nlen;
1309 /* Correct the sequence numbers. */
1310 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1311 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1312 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1314 /* PSH and FIN should only be set in the second packet. */
1315 flags = TCP_SKB_CB(skb)->tcp_flags;
1316 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1317 TCP_SKB_CB(buff)->tcp_flags = flags;
1318 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1319 tcp_skb_fragment_eor(skb, buff);
1321 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1322 /* Copy and checksum data tail into the new buffer. */
1323 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1324 skb_put(buff, nsize),
1329 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1331 skb->ip_summed = CHECKSUM_PARTIAL;
1332 skb_split(skb, buff, len);
1335 buff->ip_summed = skb->ip_summed;
1337 buff->tstamp = skb->tstamp;
1338 tcp_fragment_tstamp(skb, buff);
1340 old_factor = tcp_skb_pcount(skb);
1342 /* Fix up tso_factor for both original and new SKB. */
1343 tcp_set_skb_tso_segs(skb, mss_now);
1344 tcp_set_skb_tso_segs(buff, mss_now);
1346 /* Update delivered info for the new segment */
1347 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1349 /* If this packet has been sent out already, we must
1350 * adjust the various packet counters.
1352 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1353 int diff = old_factor - tcp_skb_pcount(skb) -
1354 tcp_skb_pcount(buff);
1357 tcp_adjust_pcount(sk, skb, diff);
1360 /* Link BUFF into the send queue. */
1361 __skb_header_release(buff);
1362 tcp_insert_write_queue_after(skb, buff, sk);
1367 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1368 * data is not copied, but immediately discarded.
1370 static int __pskb_trim_head(struct sk_buff *skb, int len)
1372 struct skb_shared_info *shinfo;
1375 eat = min_t(int, len, skb_headlen(skb));
1377 __skb_pull(skb, eat);
1384 shinfo = skb_shinfo(skb);
1385 for (i = 0; i < shinfo->nr_frags; i++) {
1386 int size = skb_frag_size(&shinfo->frags[i]);
1389 skb_frag_unref(skb, i);
1392 shinfo->frags[k] = shinfo->frags[i];
1394 shinfo->frags[k].page_offset += eat;
1395 skb_frag_size_sub(&shinfo->frags[k], eat);
1401 shinfo->nr_frags = k;
1403 skb->data_len -= len;
1404 skb->len = skb->data_len;
1408 /* Remove acked data from a packet in the transmit queue. */
1409 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1413 if (skb_unclone(skb, GFP_ATOMIC))
1416 delta_truesize = __pskb_trim_head(skb, len);
1418 TCP_SKB_CB(skb)->seq += len;
1419 skb->ip_summed = CHECKSUM_PARTIAL;
1421 if (delta_truesize) {
1422 skb->truesize -= delta_truesize;
1423 sk->sk_wmem_queued -= delta_truesize;
1424 sk_mem_uncharge(sk, delta_truesize);
1425 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1428 /* Any change of skb->len requires recalculation of tso factor. */
1429 if (tcp_skb_pcount(skb) > 1)
1430 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1435 /* Calculate MSS not accounting any TCP options. */
1436 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1438 const struct tcp_sock *tp = tcp_sk(sk);
1439 const struct inet_connection_sock *icsk = inet_csk(sk);
1442 /* Calculate base mss without TCP options:
1443 It is MMS_S - sizeof(tcphdr) of rfc1122
1445 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1447 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1448 if (icsk->icsk_af_ops->net_frag_header_len) {
1449 const struct dst_entry *dst = __sk_dst_get(sk);
1451 if (dst && dst_allfrag(dst))
1452 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1455 /* Clamp it (mss_clamp does not include tcp options) */
1456 if (mss_now > tp->rx_opt.mss_clamp)
1457 mss_now = tp->rx_opt.mss_clamp;
1459 /* Now subtract optional transport overhead */
1460 mss_now -= icsk->icsk_ext_hdr_len;
1462 /* Then reserve room for full set of TCP options and 8 bytes of data */
1463 mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1467 /* Calculate MSS. Not accounting for SACKs here. */
1468 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1470 /* Subtract TCP options size, not including SACKs */
1471 return __tcp_mtu_to_mss(sk, pmtu) -
1472 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1474 EXPORT_SYMBOL(tcp_mtu_to_mss);
1476 /* Inverse of above */
1477 int tcp_mss_to_mtu(struct sock *sk, int mss)
1479 const struct tcp_sock *tp = tcp_sk(sk);
1480 const struct inet_connection_sock *icsk = inet_csk(sk);
1484 tp->tcp_header_len +
1485 icsk->icsk_ext_hdr_len +
1486 icsk->icsk_af_ops->net_header_len;
1488 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1489 if (icsk->icsk_af_ops->net_frag_header_len) {
1490 const struct dst_entry *dst = __sk_dst_get(sk);
1492 if (dst && dst_allfrag(dst))
1493 mtu += icsk->icsk_af_ops->net_frag_header_len;
1497 EXPORT_SYMBOL(tcp_mss_to_mtu);
1499 /* MTU probing init per socket */
1500 void tcp_mtup_init(struct sock *sk)
1502 struct tcp_sock *tp = tcp_sk(sk);
1503 struct inet_connection_sock *icsk = inet_csk(sk);
1504 struct net *net = sock_net(sk);
1506 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1507 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1508 icsk->icsk_af_ops->net_header_len;
1509 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1510 icsk->icsk_mtup.probe_size = 0;
1511 if (icsk->icsk_mtup.enabled)
1512 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1514 EXPORT_SYMBOL(tcp_mtup_init);
1516 /* This function synchronize snd mss to current pmtu/exthdr set.
1518 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1519 for TCP options, but includes only bare TCP header.
1521 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1522 It is minimum of user_mss and mss received with SYN.
1523 It also does not include TCP options.
1525 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1527 tp->mss_cache is current effective sending mss, including
1528 all tcp options except for SACKs. It is evaluated,
1529 taking into account current pmtu, but never exceeds
1530 tp->rx_opt.mss_clamp.
1532 NOTE1. rfc1122 clearly states that advertised MSS
1533 DOES NOT include either tcp or ip options.
1535 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1536 are READ ONLY outside this function. --ANK (980731)
1538 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1540 struct tcp_sock *tp = tcp_sk(sk);
1541 struct inet_connection_sock *icsk = inet_csk(sk);
1544 if (icsk->icsk_mtup.search_high > pmtu)
1545 icsk->icsk_mtup.search_high = pmtu;
1547 mss_now = tcp_mtu_to_mss(sk, pmtu);
1548 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1550 /* And store cached results */
1551 icsk->icsk_pmtu_cookie = pmtu;
1552 if (icsk->icsk_mtup.enabled)
1553 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1554 tp->mss_cache = mss_now;
1558 EXPORT_SYMBOL(tcp_sync_mss);
1560 /* Compute the current effective MSS, taking SACKs and IP options,
1561 * and even PMTU discovery events into account.
1563 unsigned int tcp_current_mss(struct sock *sk)
1565 const struct tcp_sock *tp = tcp_sk(sk);
1566 const struct dst_entry *dst = __sk_dst_get(sk);
1568 unsigned int header_len;
1569 struct tcp_out_options opts;
1570 struct tcp_md5sig_key *md5;
1572 mss_now = tp->mss_cache;
1575 u32 mtu = dst_mtu(dst);
1576 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1577 mss_now = tcp_sync_mss(sk, mtu);
1580 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1581 sizeof(struct tcphdr);
1582 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1583 * some common options. If this is an odd packet (because we have SACK
1584 * blocks etc) then our calculated header_len will be different, and
1585 * we have to adjust mss_now correspondingly */
1586 if (header_len != tp->tcp_header_len) {
1587 int delta = (int) header_len - tp->tcp_header_len;
1594 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1595 * As additional protections, we do not touch cwnd in retransmission phases,
1596 * and if application hit its sndbuf limit recently.
1598 static void tcp_cwnd_application_limited(struct sock *sk)
1600 struct tcp_sock *tp = tcp_sk(sk);
1602 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1603 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1604 /* Limited by application or receiver window. */
1605 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1606 u32 win_used = max(tp->snd_cwnd_used, init_win);
1607 if (win_used < tp->snd_cwnd) {
1608 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1609 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1611 tp->snd_cwnd_used = 0;
1613 tp->snd_cwnd_stamp = tcp_jiffies32;
1616 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1618 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1619 struct tcp_sock *tp = tcp_sk(sk);
1621 /* Track the maximum number of outstanding packets in each
1622 * window, and remember whether we were cwnd-limited then.
1624 if (!before(tp->snd_una, tp->max_packets_seq) ||
1625 tp->packets_out > tp->max_packets_out ||
1627 tp->max_packets_out = tp->packets_out;
1628 tp->max_packets_seq = tp->snd_nxt;
1629 tp->is_cwnd_limited = is_cwnd_limited;
1632 if (tcp_is_cwnd_limited(sk)) {
1633 /* Network is feed fully. */
1634 tp->snd_cwnd_used = 0;
1635 tp->snd_cwnd_stamp = tcp_jiffies32;
1637 /* Network starves. */
1638 if (tp->packets_out > tp->snd_cwnd_used)
1639 tp->snd_cwnd_used = tp->packets_out;
1641 if (sysctl_tcp_slow_start_after_idle &&
1642 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1643 !ca_ops->cong_control)
1644 tcp_cwnd_application_limited(sk);
1646 /* The following conditions together indicate the starvation
1647 * is caused by insufficient sender buffer:
1648 * 1) just sent some data (see tcp_write_xmit)
1649 * 2) not cwnd limited (this else condition)
1650 * 3) no more data to send (null tcp_send_head )
1651 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1653 if (!tcp_send_head(sk) && sk->sk_socket &&
1654 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1655 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1656 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1660 /* Minshall's variant of the Nagle send check. */
1661 static bool tcp_minshall_check(const struct tcp_sock *tp)
1663 return after(tp->snd_sml, tp->snd_una) &&
1664 !after(tp->snd_sml, tp->snd_nxt);
1667 /* Update snd_sml if this skb is under mss
1668 * Note that a TSO packet might end with a sub-mss segment
1669 * The test is really :
1670 * if ((skb->len % mss) != 0)
1671 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1672 * But we can avoid doing the divide again given we already have
1673 * skb_pcount = skb->len / mss_now
1675 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1676 const struct sk_buff *skb)
1678 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1679 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1682 /* Return false, if packet can be sent now without violation Nagle's rules:
1683 * 1. It is full sized. (provided by caller in %partial bool)
1684 * 2. Or it contains FIN. (already checked by caller)
1685 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1686 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1687 * With Minshall's modification: all sent small packets are ACKed.
1689 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1693 ((nonagle & TCP_NAGLE_CORK) ||
1694 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1697 /* Return how many segs we'd like on a TSO packet,
1698 * to send one TSO packet per ms
1700 u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1705 bytes = min(sk->sk_pacing_rate >> 10,
1706 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1708 /* Goal is to send at least one packet per ms,
1709 * not one big TSO packet every 100 ms.
1710 * This preserves ACK clocking and is consistent
1711 * with tcp_tso_should_defer() heuristic.
1713 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1717 EXPORT_SYMBOL(tcp_tso_autosize);
1719 /* Return the number of segments we want in the skb we are transmitting.
1720 * See if congestion control module wants to decide; otherwise, autosize.
1722 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1724 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1725 u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0;
1728 tso_segs = tcp_tso_autosize(sk, mss_now,
1729 sysctl_tcp_min_tso_segs);
1730 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1733 /* Returns the portion of skb which can be sent right away */
1734 static unsigned int tcp_mss_split_point(const struct sock *sk,
1735 const struct sk_buff *skb,
1736 unsigned int mss_now,
1737 unsigned int max_segs,
1740 const struct tcp_sock *tp = tcp_sk(sk);
1741 u32 partial, needed, window, max_len;
1743 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1744 max_len = mss_now * max_segs;
1746 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1749 needed = min(skb->len, window);
1751 if (max_len <= needed)
1754 partial = needed % mss_now;
1755 /* If last segment is not a full MSS, check if Nagle rules allow us
1756 * to include this last segment in this skb.
1757 * Otherwise, we'll split the skb at last MSS boundary
1759 if (tcp_nagle_check(partial != 0, tp, nonagle))
1760 return needed - partial;
1765 /* Can at least one segment of SKB be sent right now, according to the
1766 * congestion window rules? If so, return how many segments are allowed.
1768 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1769 const struct sk_buff *skb)
1771 u32 in_flight, cwnd, halfcwnd;
1773 /* Don't be strict about the congestion window for the final FIN. */
1774 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1775 tcp_skb_pcount(skb) == 1)
1778 in_flight = tcp_packets_in_flight(tp);
1779 cwnd = tp->snd_cwnd;
1780 if (in_flight >= cwnd)
1783 /* For better scheduling, ensure we have at least
1784 * 2 GSO packets in flight.
1786 halfcwnd = max(cwnd >> 1, 1U);
1787 return min(halfcwnd, cwnd - in_flight);
1790 /* Initialize TSO state of a skb.
1791 * This must be invoked the first time we consider transmitting
1792 * SKB onto the wire.
1794 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1796 int tso_segs = tcp_skb_pcount(skb);
1798 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1799 tcp_set_skb_tso_segs(skb, mss_now);
1800 tso_segs = tcp_skb_pcount(skb);
1806 /* Return true if the Nagle test allows this packet to be
1809 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1810 unsigned int cur_mss, int nonagle)
1812 /* Nagle rule does not apply to frames, which sit in the middle of the
1813 * write_queue (they have no chances to get new data).
1815 * This is implemented in the callers, where they modify the 'nonagle'
1816 * argument based upon the location of SKB in the send queue.
1818 if (nonagle & TCP_NAGLE_PUSH)
1821 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1822 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1825 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1831 /* Does at least the first segment of SKB fit into the send window? */
1832 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1833 const struct sk_buff *skb,
1834 unsigned int cur_mss)
1836 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1838 if (skb->len > cur_mss)
1839 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1841 return !after(end_seq, tcp_wnd_end(tp));
1844 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1845 * which is put after SKB on the list. It is very much like
1846 * tcp_fragment() except that it may make several kinds of assumptions
1847 * in order to speed up the splitting operation. In particular, we
1848 * know that all the data is in scatter-gather pages, and that the
1849 * packet has never been sent out before (and thus is not cloned).
1851 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1852 unsigned int mss_now, gfp_t gfp)
1854 struct sk_buff *buff;
1855 int nlen = skb->len - len;
1858 /* All of a TSO frame must be composed of paged data. */
1859 if (skb->len != skb->data_len)
1860 return tcp_fragment(sk, skb, len, mss_now, gfp);
1862 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1863 if (unlikely(!buff))
1866 sk->sk_wmem_queued += buff->truesize;
1867 sk_mem_charge(sk, buff->truesize);
1868 buff->truesize += nlen;
1869 skb->truesize -= nlen;
1871 /* Correct the sequence numbers. */
1872 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1873 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1874 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1876 /* PSH and FIN should only be set in the second packet. */
1877 flags = TCP_SKB_CB(skb)->tcp_flags;
1878 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1879 TCP_SKB_CB(buff)->tcp_flags = flags;
1881 /* This packet was never sent out yet, so no SACK bits. */
1882 TCP_SKB_CB(buff)->sacked = 0;
1884 tcp_skb_fragment_eor(skb, buff);
1886 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1887 skb_split(skb, buff, len);
1888 tcp_fragment_tstamp(skb, buff);
1890 /* Fix up tso_factor for both original and new SKB. */
1891 tcp_set_skb_tso_segs(skb, mss_now);
1892 tcp_set_skb_tso_segs(buff, mss_now);
1894 /* Link BUFF into the send queue. */
1895 __skb_header_release(buff);
1896 tcp_insert_write_queue_after(skb, buff, sk);
1901 /* Try to defer sending, if possible, in order to minimize the amount
1902 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1904 * This algorithm is from John Heffner.
1906 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1907 bool *is_cwnd_limited,
1908 bool *is_rwnd_limited,
1911 const struct inet_connection_sock *icsk = inet_csk(sk);
1912 u32 age, send_win, cong_win, limit, in_flight;
1913 struct tcp_sock *tp = tcp_sk(sk);
1914 struct sk_buff *head;
1917 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1920 /* Avoid bursty behavior by allowing defer
1921 * only if the last write was recent.
1923 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
1926 in_flight = tcp_packets_in_flight(tp);
1928 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1930 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1932 /* From in_flight test above, we know that cwnd > in_flight. */
1933 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1935 limit = min(send_win, cong_win);
1937 /* If a full-sized TSO skb can be sent, do it. */
1938 if (limit >= max_segs * tp->mss_cache)
1941 /* Middle in queue won't get any more data, full sendable already? */
1942 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1945 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1947 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1949 /* If at least some fraction of a window is available,
1952 chunk /= win_divisor;
1956 /* Different approach, try not to defer past a single
1957 * ACK. Receiver should ACK every other full sized
1958 * frame, so if we have space for more than 3 frames
1961 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1965 head = tcp_write_queue_head(sk);
1967 age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
1968 /* If next ACK is likely to come too late (half srtt), do not defer */
1969 if (age < (tp->srtt_us >> 4))
1972 /* Ok, it looks like it is advisable to defer.
1973 * Three cases are tracked :
1974 * 1) We are cwnd-limited
1975 * 2) We are rwnd-limited
1976 * 3) We are application limited.
1978 if (cong_win < send_win) {
1979 if (cong_win <= skb->len) {
1980 *is_cwnd_limited = true;
1984 if (send_win <= skb->len) {
1985 *is_rwnd_limited = true;
1990 /* If this packet won't get more data, do not wait. */
1991 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2000 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2002 struct inet_connection_sock *icsk = inet_csk(sk);
2003 struct tcp_sock *tp = tcp_sk(sk);
2004 struct net *net = sock_net(sk);
2008 interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2009 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2010 if (unlikely(delta >= interval * HZ)) {
2011 int mss = tcp_current_mss(sk);
2013 /* Update current search range */
2014 icsk->icsk_mtup.probe_size = 0;
2015 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2016 sizeof(struct tcphdr) +
2017 icsk->icsk_af_ops->net_header_len;
2018 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2020 /* Update probe time stamp */
2021 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2025 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2027 struct sk_buff *skb, *next;
2029 skb = tcp_send_head(sk);
2030 tcp_for_write_queue_from_safe(skb, next, sk) {
2031 if (len <= skb->len)
2034 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2043 /* Create a new MTU probe if we are ready.
2044 * MTU probe is regularly attempting to increase the path MTU by
2045 * deliberately sending larger packets. This discovers routing
2046 * changes resulting in larger path MTUs.
2048 * Returns 0 if we should wait to probe (no cwnd available),
2049 * 1 if a probe was sent,
2052 static int tcp_mtu_probe(struct sock *sk)
2054 struct inet_connection_sock *icsk = inet_csk(sk);
2055 struct tcp_sock *tp = tcp_sk(sk);
2056 struct sk_buff *skb, *nskb, *next;
2057 struct net *net = sock_net(sk);
2064 /* Not currently probing/verifying,
2066 * have enough cwnd, and
2067 * not SACKing (the variable headers throw things off)
2069 if (likely(!icsk->icsk_mtup.enabled ||
2070 icsk->icsk_mtup.probe_size ||
2071 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2072 tp->snd_cwnd < 11 ||
2073 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2076 /* Use binary search for probe_size between tcp_mss_base,
2077 * and current mss_clamp. if (search_high - search_low)
2078 * smaller than a threshold, backoff from probing.
2080 mss_now = tcp_current_mss(sk);
2081 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2082 icsk->icsk_mtup.search_low) >> 1);
2083 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2084 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2085 /* When misfortune happens, we are reprobing actively,
2086 * and then reprobe timer has expired. We stick with current
2087 * probing process by not resetting search range to its orignal.
2089 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2090 interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2091 /* Check whether enough time has elaplased for
2092 * another round of probing.
2094 tcp_mtu_check_reprobe(sk);
2098 /* Have enough data in the send queue to probe? */
2099 if (tp->write_seq - tp->snd_nxt < size_needed)
2102 if (tp->snd_wnd < size_needed)
2104 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2107 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2108 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2109 if (!tcp_packets_in_flight(tp))
2115 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2118 /* We're allowed to probe. Build it now. */
2119 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2122 sk->sk_wmem_queued += nskb->truesize;
2123 sk_mem_charge(sk, nskb->truesize);
2125 skb = tcp_send_head(sk);
2127 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2128 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2129 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2130 TCP_SKB_CB(nskb)->sacked = 0;
2132 nskb->ip_summed = skb->ip_summed;
2134 tcp_insert_write_queue_before(nskb, skb, sk);
2135 tcp_highest_sack_replace(sk, skb, nskb);
2138 tcp_for_write_queue_from_safe(skb, next, sk) {
2139 copy = min_t(int, skb->len, probe_size - len);
2140 if (nskb->ip_summed) {
2141 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2143 __wsum csum = skb_copy_and_csum_bits(skb, 0,
2144 skb_put(nskb, copy),
2146 nskb->csum = csum_block_add(nskb->csum, csum, len);
2149 if (skb->len <= copy) {
2150 /* We've eaten all the data from this skb.
2152 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2153 /* If this is the last SKB we copy and eor is set
2154 * we need to propagate it to the new skb.
2156 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2157 tcp_skb_collapse_tstamp(nskb, skb);
2158 tcp_unlink_write_queue(skb, sk);
2159 sk_wmem_free_skb(sk, skb);
2161 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2162 ~(TCPHDR_FIN|TCPHDR_PSH);
2163 if (!skb_shinfo(skb)->nr_frags) {
2164 skb_pull(skb, copy);
2165 if (skb->ip_summed != CHECKSUM_PARTIAL)
2166 skb->csum = csum_partial(skb->data,
2169 __pskb_trim_head(skb, copy);
2170 tcp_set_skb_tso_segs(skb, mss_now);
2172 TCP_SKB_CB(skb)->seq += copy;
2177 if (len >= probe_size)
2180 tcp_init_tso_segs(nskb, nskb->len);
2182 /* We're ready to send. If this fails, the probe will
2183 * be resegmented into mss-sized pieces by tcp_write_xmit().
2185 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2186 /* Decrement cwnd here because we are sending
2187 * effectively two packets. */
2189 tcp_event_new_data_sent(sk, nskb);
2191 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2192 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2193 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2201 static bool tcp_pacing_check(const struct sock *sk)
2203 return tcp_needs_internal_pacing(sk) &&
2204 hrtimer_active(&tcp_sk(sk)->pacing_timer);
2207 /* TCP Small Queues :
2208 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2209 * (These limits are doubled for retransmits)
2211 * - better RTT estimation and ACK scheduling
2214 * Alas, some drivers / subsystems require a fair amount
2215 * of queued bytes to ensure line rate.
2216 * One example is wifi aggregation (802.11 AMPDU)
2218 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2219 unsigned int factor)
2223 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2224 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2227 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2228 /* Always send the 1st or 2nd skb in write queue.
2229 * No need to wait for TX completion to call us back,
2230 * after softirq/tasklet schedule.
2231 * This helps when TX completions are delayed too much.
2233 if (skb == sk->sk_write_queue.next ||
2234 skb->prev == sk->sk_write_queue.next)
2237 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2238 /* It is possible TX completion already happened
2239 * before we set TSQ_THROTTLED, so we must
2240 * test again the condition.
2242 smp_mb__after_atomic();
2243 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2249 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2251 const u32 now = tcp_jiffies32;
2252 enum tcp_chrono old = tp->chrono_type;
2254 if (old > TCP_CHRONO_UNSPEC)
2255 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2256 tp->chrono_start = now;
2257 tp->chrono_type = new;
2260 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2262 struct tcp_sock *tp = tcp_sk(sk);
2264 /* If there are multiple conditions worthy of tracking in a
2265 * chronograph then the highest priority enum takes precedence
2266 * over the other conditions. So that if something "more interesting"
2267 * starts happening, stop the previous chrono and start a new one.
2269 if (type > tp->chrono_type)
2270 tcp_chrono_set(tp, type);
2273 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2275 struct tcp_sock *tp = tcp_sk(sk);
2278 /* There are multiple conditions worthy of tracking in a
2279 * chronograph, so that the highest priority enum takes
2280 * precedence over the other conditions (see tcp_chrono_start).
2281 * If a condition stops, we only stop chrono tracking if
2282 * it's the "most interesting" or current chrono we are
2283 * tracking and starts busy chrono if we have pending data.
2285 if (tcp_write_queue_empty(sk))
2286 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2287 else if (type == tp->chrono_type)
2288 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2291 /* This routine writes packets to the network. It advances the
2292 * send_head. This happens as incoming acks open up the remote
2295 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2296 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2297 * account rare use of URG, this is not a big flaw.
2299 * Send at most one packet when push_one > 0. Temporarily ignore
2300 * cwnd limit to force at most one packet out when push_one == 2.
2302 * Returns true, if no segments are in flight and we have queued segments,
2303 * but cannot send anything now because of SWS or another problem.
2305 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2306 int push_one, gfp_t gfp)
2308 struct tcp_sock *tp = tcp_sk(sk);
2309 struct sk_buff *skb;
2310 unsigned int tso_segs, sent_pkts;
2313 bool is_cwnd_limited = false, is_rwnd_limited = false;
2318 tcp_mstamp_refresh(tp);
2320 /* Do MTU probing. */
2321 result = tcp_mtu_probe(sk);
2324 } else if (result > 0) {
2329 max_segs = tcp_tso_segs(sk, mss_now);
2330 while ((skb = tcp_send_head(sk))) {
2333 if (tcp_pacing_check(sk))
2336 tso_segs = tcp_init_tso_segs(skb, mss_now);
2339 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2340 /* "skb_mstamp" is used as a start point for the retransmit timer */
2341 skb->skb_mstamp = tp->tcp_mstamp;
2342 goto repair; /* Skip network transmission */
2345 cwnd_quota = tcp_cwnd_test(tp, skb);
2348 /* Force out a loss probe pkt. */
2354 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2355 is_rwnd_limited = true;
2359 if (tso_segs == 1) {
2360 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2361 (tcp_skb_is_last(sk, skb) ?
2362 nonagle : TCP_NAGLE_PUSH))))
2366 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2367 &is_rwnd_limited, max_segs))
2372 if (tso_segs > 1 && !tcp_urg_mode(tp))
2373 limit = tcp_mss_split_point(sk, skb, mss_now,
2379 if (skb->len > limit &&
2380 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2383 if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
2384 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
2385 if (tcp_small_queue_check(sk, skb, 0))
2388 /* Argh, we hit an empty skb(), presumably a thread
2389 * is sleeping in sendmsg()/sk_stream_wait_memory().
2390 * We do not want to send a pure-ack packet and have
2391 * a strange looking rtx queue with empty packet(s).
2393 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2396 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2400 /* Advance the send_head. This one is sent out.
2401 * This call will increment packets_out.
2403 tcp_event_new_data_sent(sk, skb);
2405 tcp_minshall_update(tp, mss_now, skb);
2406 sent_pkts += tcp_skb_pcount(skb);
2412 if (is_rwnd_limited)
2413 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2415 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2417 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2418 if (likely(sent_pkts || is_cwnd_limited))
2419 tcp_cwnd_validate(sk, is_cwnd_limited);
2421 if (likely(sent_pkts)) {
2422 if (tcp_in_cwnd_reduction(sk))
2423 tp->prr_out += sent_pkts;
2425 /* Send one loss probe per tail loss episode. */
2427 tcp_schedule_loss_probe(sk, false);
2430 return !tp->packets_out && tcp_send_head(sk);
2433 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2435 struct inet_connection_sock *icsk = inet_csk(sk);
2436 struct tcp_sock *tp = tcp_sk(sk);
2437 u32 timeout, rto_delta_us;
2439 /* Don't do any loss probe on a Fast Open connection before 3WHS
2442 if (tp->fastopen_rsk)
2445 /* Schedule a loss probe in 2*RTT for SACK capable connections
2446 * in Open state, that are either limited by cwnd or application.
2448 if ((sysctl_tcp_early_retrans != 3 && sysctl_tcp_early_retrans != 4) ||
2449 !tp->packets_out || !tcp_is_sack(tp) ||
2450 icsk->icsk_ca_state != TCP_CA_Open)
2453 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2457 /* Probe timeout is 2*rtt. Add minimum RTO to account
2458 * for delayed ack when there's one outstanding packet. If no RTT
2459 * sample is available then probe after TCP_TIMEOUT_INIT.
2462 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2463 if (tp->packets_out == 1)
2464 timeout += TCP_RTO_MIN;
2466 timeout += TCP_TIMEOUT_MIN;
2468 timeout = TCP_TIMEOUT_INIT;
2471 /* If the RTO formula yields an earlier time, then use that time. */
2472 rto_delta_us = advancing_rto ?
2473 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2474 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2475 if (rto_delta_us > 0)
2476 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2478 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2483 /* Thanks to skb fast clones, we can detect if a prior transmit of
2484 * a packet is still in a qdisc or driver queue.
2485 * In this case, there is very little point doing a retransmit !
2487 static bool skb_still_in_host_queue(const struct sock *sk,
2488 const struct sk_buff *skb)
2490 if (unlikely(skb_fclone_busy(sk, skb))) {
2491 NET_INC_STATS(sock_net(sk),
2492 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2498 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2499 * retransmit the last segment.
2501 void tcp_send_loss_probe(struct sock *sk)
2503 struct tcp_sock *tp = tcp_sk(sk);
2504 struct sk_buff *skb;
2506 int mss = tcp_current_mss(sk);
2508 /* At most one outstanding TLP */
2509 if (tp->tlp_high_seq)
2512 tp->tlp_retrans = 0;
2513 skb = tcp_send_head(sk);
2515 if (tcp_snd_wnd_test(tp, skb, mss)) {
2516 pcount = tp->packets_out;
2517 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2518 if (tp->packets_out > pcount)
2522 skb = tcp_write_queue_prev(sk, skb);
2524 skb = tcp_write_queue_tail(sk);
2527 if (unlikely(!skb)) {
2528 WARN_ONCE(tp->packets_out,
2529 "invalid inflight: %u state %u cwnd %u mss %d\n",
2530 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2531 inet_csk(sk)->icsk_pending = 0;
2535 if (skb_still_in_host_queue(sk, skb))
2538 pcount = tcp_skb_pcount(skb);
2539 if (WARN_ON(!pcount))
2542 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2543 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2546 skb = tcp_write_queue_next(sk, skb);
2549 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2552 if (__tcp_retransmit_skb(sk, skb, 1))
2555 tp->tlp_retrans = 1;
2558 /* Record snd_nxt for loss detection. */
2559 tp->tlp_high_seq = tp->snd_nxt;
2561 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2562 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2563 inet_csk(sk)->icsk_pending = 0;
2568 /* Push out any pending frames which were held back due to
2569 * TCP_CORK or attempt at coalescing tiny packets.
2570 * The socket must be locked by the caller.
2572 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2575 /* If we are closed, the bytes will have to remain here.
2576 * In time closedown will finish, we empty the write queue and
2577 * all will be happy.
2579 if (unlikely(sk->sk_state == TCP_CLOSE))
2582 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2583 sk_gfp_mask(sk, GFP_ATOMIC)))
2584 tcp_check_probe_timer(sk);
2587 /* Send _single_ skb sitting at the send head. This function requires
2588 * true push pending frames to setup probe timer etc.
2590 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2592 struct sk_buff *skb = tcp_send_head(sk);
2594 BUG_ON(!skb || skb->len < mss_now);
2596 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2599 /* This function returns the amount that we can raise the
2600 * usable window based on the following constraints
2602 * 1. The window can never be shrunk once it is offered (RFC 793)
2603 * 2. We limit memory per socket
2606 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2607 * RECV.NEXT + RCV.WIN fixed until:
2608 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2610 * i.e. don't raise the right edge of the window until you can raise
2611 * it at least MSS bytes.
2613 * Unfortunately, the recommended algorithm breaks header prediction,
2614 * since header prediction assumes th->window stays fixed.
2616 * Strictly speaking, keeping th->window fixed violates the receiver
2617 * side SWS prevention criteria. The problem is that under this rule
2618 * a stream of single byte packets will cause the right side of the
2619 * window to always advance by a single byte.
2621 * Of course, if the sender implements sender side SWS prevention
2622 * then this will not be a problem.
2624 * BSD seems to make the following compromise:
2626 * If the free space is less than the 1/4 of the maximum
2627 * space available and the free space is less than 1/2 mss,
2628 * then set the window to 0.
2629 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2630 * Otherwise, just prevent the window from shrinking
2631 * and from being larger than the largest representable value.
2633 * This prevents incremental opening of the window in the regime
2634 * where TCP is limited by the speed of the reader side taking
2635 * data out of the TCP receive queue. It does nothing about
2636 * those cases where the window is constrained on the sender side
2637 * because the pipeline is full.
2639 * BSD also seems to "accidentally" limit itself to windows that are a
2640 * multiple of MSS, at least until the free space gets quite small.
2641 * This would appear to be a side effect of the mbuf implementation.
2642 * Combining these two algorithms results in the observed behavior
2643 * of having a fixed window size at almost all times.
2645 * Below we obtain similar behavior by forcing the offered window to
2646 * a multiple of the mss when it is feasible to do so.
2648 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2649 * Regular options like TIMESTAMP are taken into account.
2651 u32 __tcp_select_window(struct sock *sk)
2653 struct inet_connection_sock *icsk = inet_csk(sk);
2654 struct tcp_sock *tp = tcp_sk(sk);
2655 /* MSS for the peer's data. Previous versions used mss_clamp
2656 * here. I don't know if the value based on our guesses
2657 * of peer's MSS is better for the performance. It's more correct
2658 * but may be worse for the performance because of rcv_mss
2659 * fluctuations. --SAW 1998/11/1
2661 int mss = icsk->icsk_ack.rcv_mss;
2662 int free_space = tcp_space(sk);
2663 int allowed_space = tcp_full_space(sk);
2664 int full_space = min_t(int, tp->window_clamp, allowed_space);
2667 if (unlikely(mss > full_space)) {
2672 if (free_space < (full_space >> 1)) {
2673 icsk->icsk_ack.quick = 0;
2675 if (tcp_under_memory_pressure(sk))
2676 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2679 /* free_space might become our new window, make sure we don't
2680 * increase it due to wscale.
2682 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2684 /* if free space is less than mss estimate, or is below 1/16th
2685 * of the maximum allowed, try to move to zero-window, else
2686 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2687 * new incoming data is dropped due to memory limits.
2688 * With large window, mss test triggers way too late in order
2689 * to announce zero window in time before rmem limit kicks in.
2691 if (free_space < (allowed_space >> 4) || free_space < mss)
2695 if (free_space > tp->rcv_ssthresh)
2696 free_space = tp->rcv_ssthresh;
2698 /* Don't do rounding if we are using window scaling, since the
2699 * scaled window will not line up with the MSS boundary anyway.
2701 if (tp->rx_opt.rcv_wscale) {
2702 window = free_space;
2704 /* Advertise enough space so that it won't get scaled away.
2705 * Import case: prevent zero window announcement if
2706 * 1<<rcv_wscale > mss.
2708 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2710 window = tp->rcv_wnd;
2711 /* Get the largest window that is a nice multiple of mss.
2712 * Window clamp already applied above.
2713 * If our current window offering is within 1 mss of the
2714 * free space we just keep it. This prevents the divide
2715 * and multiply from happening most of the time.
2716 * We also don't do any window rounding when the free space
2719 if (window <= free_space - mss || window > free_space)
2720 window = rounddown(free_space, mss);
2721 else if (mss == full_space &&
2722 free_space > window + (full_space >> 1))
2723 window = free_space;
2729 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2730 const struct sk_buff *next_skb)
2732 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2733 const struct skb_shared_info *next_shinfo =
2734 skb_shinfo(next_skb);
2735 struct skb_shared_info *shinfo = skb_shinfo(skb);
2737 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2738 shinfo->tskey = next_shinfo->tskey;
2739 TCP_SKB_CB(skb)->txstamp_ack |=
2740 TCP_SKB_CB(next_skb)->txstamp_ack;
2744 /* Collapses two adjacent SKB's during retransmission. */
2745 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2747 struct tcp_sock *tp = tcp_sk(sk);
2748 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2749 int skb_size, next_skb_size;
2751 skb_size = skb->len;
2752 next_skb_size = next_skb->len;
2754 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2756 if (next_skb_size) {
2757 if (next_skb_size <= skb_availroom(skb))
2758 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2760 else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
2763 tcp_highest_sack_replace(sk, next_skb, skb);
2765 tcp_unlink_write_queue(next_skb, sk);
2767 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2768 skb->ip_summed = CHECKSUM_PARTIAL;
2770 if (skb->ip_summed != CHECKSUM_PARTIAL)
2771 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2773 /* Update sequence range on original skb. */
2774 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2776 /* Merge over control information. This moves PSH/FIN etc. over */
2777 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2779 /* All done, get rid of second SKB and account for it so
2780 * packet counting does not break.
2782 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2783 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2785 /* changed transmit queue under us so clear hints */
2786 tcp_clear_retrans_hints_partial(tp);
2787 if (next_skb == tp->retransmit_skb_hint)
2788 tp->retransmit_skb_hint = skb;
2790 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2792 tcp_skb_collapse_tstamp(skb, next_skb);
2794 sk_wmem_free_skb(sk, next_skb);
2798 /* Check if coalescing SKBs is legal. */
2799 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2801 if (tcp_skb_pcount(skb) > 1)
2803 if (skb_cloned(skb))
2805 if (skb == tcp_send_head(sk))
2807 /* Some heuristics for collapsing over SACK'd could be invented */
2808 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2814 /* Collapse packets in the retransmit queue to make to create
2815 * less packets on the wire. This is only done on retransmission.
2817 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2820 struct tcp_sock *tp = tcp_sk(sk);
2821 struct sk_buff *skb = to, *tmp;
2824 if (!sysctl_tcp_retrans_collapse)
2826 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2829 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2830 if (!tcp_can_collapse(sk, skb))
2833 if (!tcp_skb_can_collapse_to(to))
2846 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2849 if (!tcp_collapse_retrans(sk, to))
2854 /* This retransmits one SKB. Policy decisions and retransmit queue
2855 * state updates are done by the caller. Returns non-zero if an
2856 * error occurred which prevented the send.
2858 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2860 struct inet_connection_sock *icsk = inet_csk(sk);
2861 struct tcp_sock *tp = tcp_sk(sk);
2862 unsigned int cur_mss;
2866 /* Inconclusive MTU probe */
2867 if (icsk->icsk_mtup.probe_size)
2868 icsk->icsk_mtup.probe_size = 0;
2870 /* Do not sent more than we queued. 1/4 is reserved for possible
2871 * copying overhead: fragmentation, tunneling, mangling etc.
2873 if (refcount_read(&sk->sk_wmem_alloc) >
2874 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2878 if (skb_still_in_host_queue(sk, skb))
2881 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2882 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2886 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2890 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2891 return -EHOSTUNREACH; /* Routing failure or similar. */
2893 cur_mss = tcp_current_mss(sk);
2895 /* If receiver has shrunk his window, and skb is out of
2896 * new window, do not retransmit it. The exception is the
2897 * case, when window is shrunk to zero. In this case
2898 * our retransmit serves as a zero window probe.
2900 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2901 TCP_SKB_CB(skb)->seq != tp->snd_una)
2904 len = cur_mss * segs;
2905 if (skb->len > len) {
2906 if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC))
2907 return -ENOMEM; /* We'll try again later. */
2909 if (skb_unclone(skb, GFP_ATOMIC))
2912 diff = tcp_skb_pcount(skb);
2913 tcp_set_skb_tso_segs(skb, cur_mss);
2914 diff -= tcp_skb_pcount(skb);
2916 tcp_adjust_pcount(sk, skb, diff);
2917 if (skb->len < cur_mss)
2918 tcp_retrans_try_collapse(sk, skb, cur_mss);
2921 /* RFC3168, section 6.1.1.1. ECN fallback */
2922 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2923 tcp_ecn_clear_syn(sk, skb);
2925 /* Update global and local TCP statistics. */
2926 segs = tcp_skb_pcount(skb);
2927 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2928 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2929 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2930 tp->total_retrans += segs;
2932 /* make sure skb->data is aligned on arches that require it
2933 * and check if ack-trimming & collapsing extended the headroom
2934 * beyond what csum_start can cover.
2936 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2937 skb_headroom(skb) >= 0xFFFF)) {
2938 struct sk_buff *nskb;
2940 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2941 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2944 skb->skb_mstamp = tp->tcp_mstamp;
2945 tcp_rate_skb_sent(sk, skb);
2948 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2952 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2953 } else if (err != -EBUSY) {
2954 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
2959 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2961 struct tcp_sock *tp = tcp_sk(sk);
2962 int err = __tcp_retransmit_skb(sk, skb, segs);
2965 #if FASTRETRANS_DEBUG > 0
2966 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2967 net_dbg_ratelimited("retrans_out leaked\n");
2970 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2971 tp->retrans_out += tcp_skb_pcount(skb);
2973 /* Save stamp of the first retransmit. */
2974 if (!tp->retrans_stamp)
2975 tp->retrans_stamp = tcp_skb_timestamp(skb);
2979 if (tp->undo_retrans < 0)
2980 tp->undo_retrans = 0;
2981 tp->undo_retrans += tcp_skb_pcount(skb);
2985 /* This gets called after a retransmit timeout, and the initially
2986 * retransmitted data is acknowledged. It tries to continue
2987 * resending the rest of the retransmit queue, until either
2988 * we've sent it all or the congestion window limit is reached.
2989 * If doing SACK, the first ACK which comes back for a timeout
2990 * based retransmit packet might feed us FACK information again.
2991 * If so, we use it to avoid unnecessarily retransmissions.
2993 void tcp_xmit_retransmit_queue(struct sock *sk)
2995 const struct inet_connection_sock *icsk = inet_csk(sk);
2996 struct tcp_sock *tp = tcp_sk(sk);
2997 struct sk_buff *skb;
2998 struct sk_buff *hole = NULL;
3002 if (!tp->packets_out)
3005 if (tp->retransmit_skb_hint) {
3006 skb = tp->retransmit_skb_hint;
3008 skb = tcp_write_queue_head(sk);
3011 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3012 tcp_for_write_queue_from(skb, sk) {
3016 if (skb == tcp_send_head(sk))
3019 if (tcp_pacing_check(sk))
3022 /* we could do better than to assign each time */
3024 tp->retransmit_skb_hint = skb;
3026 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
3029 sacked = TCP_SKB_CB(skb)->sacked;
3030 /* In case tcp_shift_skb_data() have aggregated large skbs,
3031 * we need to make sure not sending too bigs TSO packets
3033 segs = min_t(int, segs, max_segs);
3035 if (tp->retrans_out >= tp->lost_out) {
3037 } else if (!(sacked & TCPCB_LOST)) {
3038 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3043 if (icsk->icsk_ca_state != TCP_CA_Loss)
3044 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3046 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3049 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3052 if (tcp_small_queue_check(sk, skb, 1))
3055 if (tcp_retransmit_skb(sk, skb, segs))
3058 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3060 if (tcp_in_cwnd_reduction(sk))
3061 tp->prr_out += tcp_skb_pcount(skb);
3063 if (skb == tcp_write_queue_head(sk) &&
3064 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3065 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3066 inet_csk(sk)->icsk_rto,
3071 /* We allow to exceed memory limits for FIN packets to expedite
3072 * connection tear down and (memory) recovery.
3073 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3074 * or even be forced to close flow without any FIN.
3075 * In general, we want to allow one skb per socket to avoid hangs
3076 * with edge trigger epoll()
3078 void sk_forced_mem_schedule(struct sock *sk, int size)
3082 if (size <= sk->sk_forward_alloc)
3084 amt = sk_mem_pages(size);
3085 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3086 sk_memory_allocated_add(sk, amt);
3088 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3089 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3092 /* Send a FIN. The caller locks the socket for us.
3093 * We should try to send a FIN packet really hard, but eventually give up.
3095 void tcp_send_fin(struct sock *sk)
3097 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3098 struct tcp_sock *tp = tcp_sk(sk);
3100 /* Optimization, tack on the FIN if we have one skb in write queue and
3101 * this skb was not yet sent, or we are under memory pressure.
3102 * Note: in the latter case, FIN packet will be sent after a timeout,
3103 * as TCP stack thinks it has already been transmitted.
3105 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
3107 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3108 TCP_SKB_CB(tskb)->end_seq++;
3110 if (!tcp_send_head(sk)) {
3111 /* This means tskb was already sent.
3112 * Pretend we included the FIN on previous transmit.
3113 * We need to set tp->snd_nxt to the value it would have
3114 * if FIN had been sent. This is because retransmit path
3115 * does not change tp->snd_nxt.
3121 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3122 if (unlikely(!skb)) {
3127 skb_reserve(skb, MAX_TCP_HEADER);
3128 sk_forced_mem_schedule(sk, skb->truesize);
3129 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3130 tcp_init_nondata_skb(skb, tp->write_seq,
3131 TCPHDR_ACK | TCPHDR_FIN);
3132 tcp_queue_skb(sk, skb);
3134 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3137 /* We get here when a process closes a file descriptor (either due to
3138 * an explicit close() or as a byproduct of exit()'ing) and there
3139 * was unread data in the receive queue. This behavior is recommended
3140 * by RFC 2525, section 2.17. -DaveM
3142 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3144 struct sk_buff *skb;
3146 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3148 /* NOTE: No TCP options attached and we never retransmit this. */
3149 skb = alloc_skb(MAX_TCP_HEADER, priority);
3151 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3155 /* Reserve space for headers and prepare control bits. */
3156 skb_reserve(skb, MAX_TCP_HEADER);
3157 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3158 TCPHDR_ACK | TCPHDR_RST);
3159 tcp_mstamp_refresh(tcp_sk(sk));
3161 if (tcp_transmit_skb(sk, skb, 0, priority))
3162 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3165 /* Send a crossed SYN-ACK during socket establishment.
3166 * WARNING: This routine must only be called when we have already sent
3167 * a SYN packet that crossed the incoming SYN that caused this routine
3168 * to get called. If this assumption fails then the initial rcv_wnd
3169 * and rcv_wscale values will not be correct.
3171 int tcp_send_synack(struct sock *sk)
3173 struct sk_buff *skb;
3175 skb = tcp_write_queue_head(sk);
3176 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3177 pr_debug("%s: wrong queue state\n", __func__);
3180 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3181 if (skb_cloned(skb)) {
3182 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
3185 tcp_unlink_write_queue(skb, sk);
3186 __skb_header_release(nskb);
3187 __tcp_add_write_queue_head(sk, nskb);
3188 sk_wmem_free_skb(sk, skb);
3189 sk->sk_wmem_queued += nskb->truesize;
3190 sk_mem_charge(sk, nskb->truesize);
3194 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3195 tcp_ecn_send_synack(sk, skb);
3197 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3201 * tcp_make_synack - Prepare a SYN-ACK.
3202 * sk: listener socket
3203 * dst: dst entry attached to the SYNACK
3204 * req: request_sock pointer
3206 * Allocate one skb and build a SYNACK packet.
3207 * @dst is consumed : Caller should not use it again.
3209 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3210 struct request_sock *req,
3211 struct tcp_fastopen_cookie *foc,
3212 enum tcp_synack_type synack_type)
3214 struct inet_request_sock *ireq = inet_rsk(req);
3215 const struct tcp_sock *tp = tcp_sk(sk);
3216 struct tcp_md5sig_key *md5 = NULL;
3217 struct tcp_out_options opts;
3218 struct sk_buff *skb;
3219 int tcp_header_size;
3223 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3224 if (unlikely(!skb)) {
3228 /* Reserve space for headers. */
3229 skb_reserve(skb, MAX_TCP_HEADER);
3231 switch (synack_type) {
3232 case TCP_SYNACK_NORMAL:
3233 skb_set_owner_w(skb, req_to_sk(req));
3235 case TCP_SYNACK_COOKIE:
3236 /* Under synflood, we do not attach skb to a socket,
3237 * to avoid false sharing.
3240 case TCP_SYNACK_FASTOPEN:
3241 /* sk is a const pointer, because we want to express multiple
3242 * cpu might call us concurrently.
3243 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3245 skb_set_owner_w(skb, (struct sock *)sk);
3248 skb_dst_set(skb, dst);
3250 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3252 memset(&opts, 0, sizeof(opts));
3253 #ifdef CONFIG_SYN_COOKIES
3254 if (unlikely(req->cookie_ts))
3255 skb->skb_mstamp = cookie_init_timestamp(req);
3258 skb->skb_mstamp = tcp_clock_us();
3260 #ifdef CONFIG_TCP_MD5SIG
3262 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3264 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3265 tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5,
3266 foc, synack_type) + sizeof(*th);
3268 skb_push(skb, tcp_header_size);
3269 skb_reset_transport_header(skb);
3271 th = (struct tcphdr *)skb->data;
3272 memset(th, 0, sizeof(struct tcphdr));
3275 tcp_ecn_make_synack(req, th);
3276 th->source = htons(ireq->ir_num);
3277 th->dest = ireq->ir_rmt_port;
3278 skb->mark = ireq->ir_mark;
3279 skb->ip_summed = CHECKSUM_PARTIAL;
3280 th->seq = htonl(tcp_rsk(req)->snt_isn);
3281 /* XXX data is queued and acked as is. No buffer/window check */
3282 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3284 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3285 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3286 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3287 th->doff = (tcp_header_size >> 2);
3288 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3290 #ifdef CONFIG_TCP_MD5SIG
3291 /* Okay, we have all we need - do the md5 hash if needed */
3293 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3294 md5, req_to_sk(req), skb);
3298 /* Do not fool tcpdump (if any), clean our debris */
3302 EXPORT_SYMBOL(tcp_make_synack);
3304 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3306 struct inet_connection_sock *icsk = inet_csk(sk);
3307 const struct tcp_congestion_ops *ca;
3308 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3310 if (ca_key == TCP_CA_UNSPEC)
3314 ca = tcp_ca_find_key(ca_key);
3315 if (likely(ca && try_module_get(ca->owner))) {
3316 module_put(icsk->icsk_ca_ops->owner);
3317 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3318 icsk->icsk_ca_ops = ca;
3323 /* Do all connect socket setups that can be done AF independent. */
3324 static void tcp_connect_init(struct sock *sk)
3326 const struct dst_entry *dst = __sk_dst_get(sk);
3327 struct tcp_sock *tp = tcp_sk(sk);
3331 /* We'll fix this up when we get a response from the other end.
3332 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3334 tp->tcp_header_len = sizeof(struct tcphdr);
3335 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3336 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3338 #ifdef CONFIG_TCP_MD5SIG
3339 if (tp->af_specific->md5_lookup(sk, sk))
3340 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3343 /* If user gave his TCP_MAXSEG, record it to clamp */
3344 if (tp->rx_opt.user_mss)
3345 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3348 tcp_sync_mss(sk, dst_mtu(dst));
3350 tcp_ca_dst_init(sk, dst);
3352 if (!tp->window_clamp)
3353 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3354 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3356 tcp_initialize_rcv_mss(sk);
3358 /* limit the window selection if the user enforce a smaller rx buffer */
3359 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3360 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3361 tp->window_clamp = tcp_full_space(sk);
3363 rcv_wnd = tcp_rwnd_init_bpf(sk);
3365 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3367 tcp_select_initial_window(tcp_full_space(sk),
3368 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3371 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3375 tp->rx_opt.rcv_wscale = rcv_wscale;
3376 tp->rcv_ssthresh = tp->rcv_wnd;
3379 sock_reset_flag(sk, SOCK_DONE);
3382 tcp_write_queue_purge(sk);
3383 tp->snd_una = tp->write_seq;
3384 tp->snd_sml = tp->write_seq;
3385 tp->snd_up = tp->write_seq;
3386 tp->snd_nxt = tp->write_seq;
3388 if (likely(!tp->repair))
3391 tp->rcv_tstamp = tcp_jiffies32;
3392 tp->rcv_wup = tp->rcv_nxt;
3393 tp->copied_seq = tp->rcv_nxt;
3395 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3396 inet_csk(sk)->icsk_retransmits = 0;
3397 tcp_clear_retrans(tp);
3400 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3402 struct tcp_sock *tp = tcp_sk(sk);
3403 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3405 tcb->end_seq += skb->len;
3406 __skb_header_release(skb);
3407 __tcp_add_write_queue_tail(sk, skb);
3408 sk->sk_wmem_queued += skb->truesize;
3409 sk_mem_charge(sk, skb->truesize);
3410 tp->write_seq = tcb->end_seq;
3411 tp->packets_out += tcp_skb_pcount(skb);
3414 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3415 * queue a data-only packet after the regular SYN, such that regular SYNs
3416 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3417 * only the SYN sequence, the data are retransmitted in the first ACK.
3418 * If cookie is not cached or other error occurs, falls back to send a
3419 * regular SYN with Fast Open cookie request option.
3421 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3423 struct inet_connection_sock *icsk = inet_csk(sk);
3424 struct tcp_sock *tp = tcp_sk(sk);
3425 struct tcp_fastopen_request *fo = tp->fastopen_req;
3427 struct sk_buff *syn_data;
3429 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3430 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3433 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3434 * user-MSS. Reserve maximum option space for middleboxes that add
3435 * private TCP options. The cost is reduced data space in SYN :(
3437 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3438 /* Sync mss_cache after updating the mss_clamp */
3439 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3441 space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3442 MAX_TCP_OPTION_SPACE;
3444 space = min_t(size_t, space, fo->size);
3446 /* limit to order-0 allocations */
3447 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3449 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3452 syn_data->ip_summed = CHECKSUM_PARTIAL;
3453 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3455 int copied = copy_from_iter(skb_put(syn_data, space), space,
3456 &fo->data->msg_iter);
3457 if (unlikely(!copied)) {
3458 kfree_skb(syn_data);
3461 if (copied != space) {
3462 skb_trim(syn_data, copied);
3466 /* No more data pending in inet_wait_for_connect() */
3467 if (space == fo->size)
3471 tcp_connect_queue_skb(sk, syn_data);
3473 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3475 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3477 syn->skb_mstamp = syn_data->skb_mstamp;
3479 /* Now full SYN+DATA was cloned and sent (or not),
3480 * remove the SYN from the original skb (syn_data)
3481 * we keep in write queue in case of a retransmit, as we
3482 * also have the SYN packet (with no data) in the same queue.
3484 TCP_SKB_CB(syn_data)->seq++;
3485 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3487 tp->syn_data = (fo->copied > 0);
3488 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3492 /* data was not sent, this is our new send_head */
3493 sk->sk_send_head = syn_data;
3494 tp->packets_out -= tcp_skb_pcount(syn_data);
3497 /* Send a regular SYN with Fast Open cookie request option */
3498 if (fo->cookie.len > 0)
3500 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3502 tp->syn_fastopen = 0;
3504 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3508 /* Build a SYN and send it off. */
3509 int tcp_connect(struct sock *sk)
3511 struct tcp_sock *tp = tcp_sk(sk);
3512 struct sk_buff *buff;
3515 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB);
3517 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3518 return -EHOSTUNREACH; /* Routing failure or similar. */
3520 tcp_connect_init(sk);
3522 if (unlikely(tp->repair)) {
3523 tcp_finish_connect(sk, NULL);
3527 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3528 if (unlikely(!buff))
3531 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3532 tcp_mstamp_refresh(tp);
3533 tp->retrans_stamp = tcp_time_stamp(tp);
3534 tcp_connect_queue_skb(sk, buff);
3535 tcp_ecn_send_syn(sk, buff);
3537 /* Send off SYN; include data in Fast Open. */
3538 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3539 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3540 if (err == -ECONNREFUSED)
3543 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3544 * in order to make this packet get counted in tcpOutSegs.
3546 tp->snd_nxt = tp->write_seq;
3547 tp->pushed_seq = tp->write_seq;
3548 buff = tcp_send_head(sk);
3549 if (unlikely(buff)) {
3550 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3551 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3553 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3555 /* Timer for repeating the SYN until an answer. */
3556 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3557 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3560 EXPORT_SYMBOL(tcp_connect);
3562 /* Send out a delayed ack, the caller does the policy checking
3563 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3566 void tcp_send_delayed_ack(struct sock *sk)
3568 struct inet_connection_sock *icsk = inet_csk(sk);
3569 int ato = icsk->icsk_ack.ato;
3570 unsigned long timeout;
3572 if (ato > TCP_DELACK_MIN) {
3573 const struct tcp_sock *tp = tcp_sk(sk);
3574 int max_ato = HZ / 2;
3576 if (icsk->icsk_ack.pingpong ||
3577 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3578 max_ato = TCP_DELACK_MAX;
3580 /* Slow path, intersegment interval is "high". */
3582 /* If some rtt estimate is known, use it to bound delayed ack.
3583 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3587 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3594 ato = min(ato, max_ato);
3597 /* Stay within the limit we were given */
3598 timeout = jiffies + ato;
3600 /* Use new timeout only if there wasn't a older one earlier. */
3601 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3602 /* If delack timer was blocked or is about to expire,
3605 if (icsk->icsk_ack.blocked ||
3606 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3611 if (!time_before(timeout, icsk->icsk_ack.timeout))
3612 timeout = icsk->icsk_ack.timeout;
3614 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3615 icsk->icsk_ack.timeout = timeout;
3616 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3619 /* This routine sends an ack and also updates the window. */
3620 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3622 struct sk_buff *buff;
3624 /* If we have been reset, we may not send again. */
3625 if (sk->sk_state == TCP_CLOSE)
3628 /* We are not putting this on the write queue, so
3629 * tcp_transmit_skb() will set the ownership to this
3632 buff = alloc_skb(MAX_TCP_HEADER,
3633 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3634 if (unlikely(!buff)) {
3635 inet_csk_schedule_ack(sk);
3636 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3637 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3638 TCP_DELACK_MAX, TCP_RTO_MAX);
3642 /* Reserve space for headers and prepare control bits. */
3643 skb_reserve(buff, MAX_TCP_HEADER);
3644 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3646 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3648 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3650 skb_set_tcp_pure_ack(buff);
3652 /* Send it off, this clears delayed acks for us. */
3653 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3655 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3657 void tcp_send_ack(struct sock *sk)
3659 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3662 /* This routine sends a packet with an out of date sequence
3663 * number. It assumes the other end will try to ack it.
3665 * Question: what should we make while urgent mode?
3666 * 4.4BSD forces sending single byte of data. We cannot send
3667 * out of window data, because we have SND.NXT==SND.MAX...
3669 * Current solution: to send TWO zero-length segments in urgent mode:
3670 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3671 * out-of-date with SND.UNA-1 to probe window.
3673 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3675 struct tcp_sock *tp = tcp_sk(sk);
3676 struct sk_buff *skb;
3678 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3679 skb = alloc_skb(MAX_TCP_HEADER,
3680 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3684 /* Reserve space for headers and set control bits. */
3685 skb_reserve(skb, MAX_TCP_HEADER);
3686 /* Use a previous sequence. This should cause the other
3687 * end to send an ack. Don't queue or clone SKB, just
3690 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3691 NET_INC_STATS(sock_net(sk), mib);
3692 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3695 /* Called from setsockopt( ... TCP_REPAIR ) */
3696 void tcp_send_window_probe(struct sock *sk)
3698 if (sk->sk_state == TCP_ESTABLISHED) {
3699 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3700 tcp_mstamp_refresh(tcp_sk(sk));
3701 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3705 /* Initiate keepalive or window probe from timer. */
3706 int tcp_write_wakeup(struct sock *sk, int mib)
3708 struct tcp_sock *tp = tcp_sk(sk);
3709 struct sk_buff *skb;
3711 if (sk->sk_state == TCP_CLOSE)
3714 skb = tcp_send_head(sk);
3715 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3717 unsigned int mss = tcp_current_mss(sk);
3718 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3720 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3721 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3723 /* We are probing the opening of a window
3724 * but the window size is != 0
3725 * must have been a result SWS avoidance ( sender )
3727 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3729 seg_size = min(seg_size, mss);
3730 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3731 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3733 } else if (!tcp_skb_pcount(skb))
3734 tcp_set_skb_tso_segs(skb, mss);
3736 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3737 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3739 tcp_event_new_data_sent(sk, skb);
3742 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3743 tcp_xmit_probe_skb(sk, 1, mib);
3744 return tcp_xmit_probe_skb(sk, 0, mib);
3748 /* A window probe timeout has occurred. If window is not closed send
3749 * a partial packet else a zero probe.
3751 void tcp_send_probe0(struct sock *sk)
3753 struct inet_connection_sock *icsk = inet_csk(sk);
3754 struct tcp_sock *tp = tcp_sk(sk);
3755 struct net *net = sock_net(sk);
3756 unsigned long probe_max;
3759 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3761 if (tp->packets_out || !tcp_send_head(sk)) {
3762 /* Cancel probe timer, if it is not required. */
3763 icsk->icsk_probes_out = 0;
3764 icsk->icsk_backoff = 0;
3769 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3770 icsk->icsk_backoff++;
3771 icsk->icsk_probes_out++;
3772 probe_max = TCP_RTO_MAX;
3774 /* If packet was not sent due to local congestion,
3775 * do not backoff and do not remember icsk_probes_out.
3776 * Let local senders to fight for local resources.
3778 * Use accumulated backoff yet.
3780 if (!icsk->icsk_probes_out)
3781 icsk->icsk_probes_out = 1;
3782 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3784 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3785 tcp_probe0_when(sk, probe_max),
3789 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3791 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3795 tcp_rsk(req)->txhash = net_tx_rndhash();
3796 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3798 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3799 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3800 if (unlikely(tcp_passive_fastopen(sk)))
3801 tcp_sk(sk)->total_retrans++;
3805 EXPORT_SYMBOL(tcp_rtx_synack);