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>
44 #include <linux/static_key.h>
46 #include <trace/events/tcp.h>
48 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
49 int push_one, gfp_t gfp);
51 /* Account for new data that has been sent to the network. */
52 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
54 struct inet_connection_sock *icsk = inet_csk(sk);
55 struct tcp_sock *tp = tcp_sk(sk);
56 unsigned int prior_packets = tp->packets_out;
58 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
60 __skb_unlink(skb, &sk->sk_write_queue);
61 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
63 if (tp->highest_sack == NULL)
64 tp->highest_sack = skb;
66 tp->packets_out += tcp_skb_pcount(skb);
67 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
70 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
75 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
76 * window scaling factor due to loss of precision.
77 * If window has been shrunk, what should we make? It is not clear at all.
78 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
79 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
80 * invalid. OK, let's make this for now:
82 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
84 const struct tcp_sock *tp = tcp_sk(sk);
86 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
87 (tp->rx_opt.wscale_ok &&
88 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
91 return tcp_wnd_end(tp);
94 /* Calculate mss to advertise in SYN segment.
95 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
97 * 1. It is independent of path mtu.
98 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
99 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
100 * attached devices, because some buggy hosts are confused by
102 * 4. We do not make 3, we advertise MSS, calculated from first
103 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
104 * This may be overridden via information stored in routing table.
105 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
106 * probably even Jumbo".
108 static __u16 tcp_advertise_mss(struct sock *sk)
110 struct tcp_sock *tp = tcp_sk(sk);
111 const struct dst_entry *dst = __sk_dst_get(sk);
112 int mss = tp->advmss;
115 unsigned int metric = dst_metric_advmss(dst);
126 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
127 * This is the first part of cwnd validation mechanism.
129 void tcp_cwnd_restart(struct sock *sk, s32 delta)
131 struct tcp_sock *tp = tcp_sk(sk);
132 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
133 u32 cwnd = tp->snd_cwnd;
135 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
137 tp->snd_ssthresh = tcp_current_ssthresh(sk);
138 restart_cwnd = min(restart_cwnd, cwnd);
140 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
142 tp->snd_cwnd = max(cwnd, restart_cwnd);
143 tp->snd_cwnd_stamp = tcp_jiffies32;
144 tp->snd_cwnd_used = 0;
147 /* Congestion state accounting after a packet has been sent. */
148 static void tcp_event_data_sent(struct tcp_sock *tp,
151 struct inet_connection_sock *icsk = inet_csk(sk);
152 const u32 now = tcp_jiffies32;
154 if (tcp_packets_in_flight(tp) == 0)
155 tcp_ca_event(sk, CA_EVENT_TX_START);
159 /* If it is a reply for ato after last received
160 * packet, enter pingpong mode.
162 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
163 icsk->icsk_ack.pingpong = 1;
166 /* Account for an ACK we sent. */
167 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
170 struct tcp_sock *tp = tcp_sk(sk);
172 if (unlikely(tp->compressed_ack > TCP_FASTRETRANS_THRESH)) {
173 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
174 tp->compressed_ack - TCP_FASTRETRANS_THRESH);
175 tp->compressed_ack = TCP_FASTRETRANS_THRESH;
176 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
180 if (unlikely(rcv_nxt != tp->rcv_nxt))
181 return; /* Special ACK sent by DCTCP to reflect ECN */
182 tcp_dec_quickack_mode(sk, pkts);
183 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
186 /* Determine a window scaling and initial window to offer.
187 * Based on the assumption that the given amount of space
188 * will be offered. Store the results in the tp structure.
189 * NOTE: for smooth operation initial space offering should
190 * be a multiple of mss if possible. We assume here that mss >= 1.
191 * This MUST be enforced by all callers.
193 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
194 __u32 *rcv_wnd, __u32 *window_clamp,
195 int wscale_ok, __u8 *rcv_wscale,
198 unsigned int space = (__space < 0 ? 0 : __space);
200 /* If no clamp set the clamp to the max possible scaled window */
201 if (*window_clamp == 0)
202 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
203 space = min(*window_clamp, space);
205 /* Quantize space offering to a multiple of mss if possible. */
207 space = rounddown(space, mss);
209 /* NOTE: offering an initial window larger than 32767
210 * will break some buggy TCP stacks. If the admin tells us
211 * it is likely we could be speaking with such a buggy stack
212 * we will truncate our initial window offering to 32K-1
213 * unless the remote has sent us a window scaling option,
214 * which we interpret as a sign the remote TCP is not
215 * misinterpreting the window field as a signed quantity.
217 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
218 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
220 (*rcv_wnd) = min_t(u32, space, U16_MAX);
223 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
227 /* Set window scaling on max possible window */
228 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
229 space = max_t(u32, space, sysctl_rmem_max);
230 space = min_t(u32, space, *window_clamp);
231 while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
236 /* Set the clamp no higher than max representable value */
237 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
239 EXPORT_SYMBOL(tcp_select_initial_window);
241 /* Chose a new window to advertise, update state in tcp_sock for the
242 * socket, and return result with RFC1323 scaling applied. The return
243 * value can be stuffed directly into th->window for an outgoing
246 static u16 tcp_select_window(struct sock *sk)
248 struct tcp_sock *tp = tcp_sk(sk);
249 u32 old_win = tp->rcv_wnd;
250 u32 cur_win = tcp_receive_window(tp);
251 u32 new_win = __tcp_select_window(sk);
253 /* Never shrink the offered window */
254 if (new_win < cur_win) {
255 /* Danger Will Robinson!
256 * Don't update rcv_wup/rcv_wnd here or else
257 * we will not be able to advertise a zero
258 * window in time. --DaveM
260 * Relax Will Robinson.
263 NET_INC_STATS(sock_net(sk),
264 LINUX_MIB_TCPWANTZEROWINDOWADV);
265 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
267 tp->rcv_wnd = new_win;
268 tp->rcv_wup = tp->rcv_nxt;
270 /* Make sure we do not exceed the maximum possible
273 if (!tp->rx_opt.rcv_wscale &&
274 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
275 new_win = min(new_win, MAX_TCP_WINDOW);
277 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
279 /* RFC1323 scaling applied */
280 new_win >>= tp->rx_opt.rcv_wscale;
282 /* If we advertise zero window, disable fast path. */
286 NET_INC_STATS(sock_net(sk),
287 LINUX_MIB_TCPTOZEROWINDOWADV);
288 } else if (old_win == 0) {
289 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
295 /* Packet ECN state for a SYN-ACK */
296 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
298 const struct tcp_sock *tp = tcp_sk(sk);
300 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
301 if (!(tp->ecn_flags & TCP_ECN_OK))
302 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
303 else if (tcp_ca_needs_ecn(sk) ||
304 tcp_bpf_ca_needs_ecn(sk))
308 /* Packet ECN state for a SYN. */
309 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
311 struct tcp_sock *tp = tcp_sk(sk);
312 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
313 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
314 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
317 const struct dst_entry *dst = __sk_dst_get(sk);
319 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
326 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
327 tp->ecn_flags = TCP_ECN_OK;
328 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
333 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
335 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
336 /* tp->ecn_flags are cleared at a later point in time when
337 * SYN ACK is ultimatively being received.
339 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
343 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
345 if (inet_rsk(req)->ecn_ok)
349 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
352 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
353 struct tcphdr *th, int tcp_header_len)
355 struct tcp_sock *tp = tcp_sk(sk);
357 if (tp->ecn_flags & TCP_ECN_OK) {
358 /* Not-retransmitted data segment: set ECT and inject CWR. */
359 if (skb->len != tcp_header_len &&
360 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
362 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
363 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
365 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
367 } else if (!tcp_ca_needs_ecn(sk)) {
368 /* ACK or retransmitted segment: clear ECT|CE */
369 INET_ECN_dontxmit(sk);
371 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
376 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
377 * auto increment end seqno.
379 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
381 skb->ip_summed = CHECKSUM_PARTIAL;
383 TCP_SKB_CB(skb)->tcp_flags = flags;
384 TCP_SKB_CB(skb)->sacked = 0;
386 tcp_skb_pcount_set(skb, 1);
388 TCP_SKB_CB(skb)->seq = seq;
389 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
391 TCP_SKB_CB(skb)->end_seq = seq;
394 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
396 return tp->snd_una != tp->snd_up;
399 #define OPTION_SACK_ADVERTISE (1 << 0)
400 #define OPTION_TS (1 << 1)
401 #define OPTION_MD5 (1 << 2)
402 #define OPTION_WSCALE (1 << 3)
403 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
404 #define OPTION_SMC (1 << 9)
406 static void smc_options_write(__be32 *ptr, u16 *options)
408 #if IS_ENABLED(CONFIG_SMC)
409 if (static_branch_unlikely(&tcp_have_smc)) {
410 if (unlikely(OPTION_SMC & *options)) {
411 *ptr++ = htonl((TCPOPT_NOP << 24) |
414 (TCPOLEN_EXP_SMC_BASE));
415 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
421 struct tcp_out_options {
422 u16 options; /* bit field of OPTION_* */
423 u16 mss; /* 0 to disable */
424 u8 ws; /* window scale, 0 to disable */
425 u8 num_sack_blocks; /* number of SACK blocks to include */
426 u8 hash_size; /* bytes in hash_location */
427 __u8 *hash_location; /* temporary pointer, overloaded */
428 __u32 tsval, tsecr; /* need to include OPTION_TS */
429 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
432 /* Write previously computed TCP options to the packet.
434 * Beware: Something in the Internet is very sensitive to the ordering of
435 * TCP options, we learned this through the hard way, so be careful here.
436 * Luckily we can at least blame others for their non-compliance but from
437 * inter-operability perspective it seems that we're somewhat stuck with
438 * the ordering which we have been using if we want to keep working with
439 * those broken things (not that it currently hurts anybody as there isn't
440 * particular reason why the ordering would need to be changed).
442 * At least SACK_PERM as the first option is known to lead to a disaster
443 * (but it may well be that other scenarios fail similarly).
445 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
446 struct tcp_out_options *opts)
448 u16 options = opts->options; /* mungable copy */
450 if (unlikely(OPTION_MD5 & options)) {
451 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
452 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
453 /* overload cookie hash location */
454 opts->hash_location = (__u8 *)ptr;
458 if (unlikely(opts->mss)) {
459 *ptr++ = htonl((TCPOPT_MSS << 24) |
460 (TCPOLEN_MSS << 16) |
464 if (likely(OPTION_TS & options)) {
465 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
466 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
467 (TCPOLEN_SACK_PERM << 16) |
468 (TCPOPT_TIMESTAMP << 8) |
470 options &= ~OPTION_SACK_ADVERTISE;
472 *ptr++ = htonl((TCPOPT_NOP << 24) |
474 (TCPOPT_TIMESTAMP << 8) |
477 *ptr++ = htonl(opts->tsval);
478 *ptr++ = htonl(opts->tsecr);
481 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
482 *ptr++ = htonl((TCPOPT_NOP << 24) |
484 (TCPOPT_SACK_PERM << 8) |
488 if (unlikely(OPTION_WSCALE & options)) {
489 *ptr++ = htonl((TCPOPT_NOP << 24) |
490 (TCPOPT_WINDOW << 16) |
491 (TCPOLEN_WINDOW << 8) |
495 if (unlikely(opts->num_sack_blocks)) {
496 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
497 tp->duplicate_sack : tp->selective_acks;
500 *ptr++ = htonl((TCPOPT_NOP << 24) |
503 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
504 TCPOLEN_SACK_PERBLOCK)));
506 for (this_sack = 0; this_sack < opts->num_sack_blocks;
508 *ptr++ = htonl(sp[this_sack].start_seq);
509 *ptr++ = htonl(sp[this_sack].end_seq);
512 tp->rx_opt.dsack = 0;
515 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
516 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
518 u32 len; /* Fast Open option length */
521 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
522 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
523 TCPOPT_FASTOPEN_MAGIC);
524 p += TCPOLEN_EXP_FASTOPEN_BASE;
526 len = TCPOLEN_FASTOPEN_BASE + foc->len;
527 *p++ = TCPOPT_FASTOPEN;
531 memcpy(p, foc->val, foc->len);
532 if ((len & 3) == 2) {
533 p[foc->len] = TCPOPT_NOP;
534 p[foc->len + 1] = TCPOPT_NOP;
536 ptr += (len + 3) >> 2;
539 smc_options_write(ptr, &options);
542 static void smc_set_option(const struct tcp_sock *tp,
543 struct tcp_out_options *opts,
544 unsigned int *remaining)
546 #if IS_ENABLED(CONFIG_SMC)
547 if (static_branch_unlikely(&tcp_have_smc)) {
549 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
550 opts->options |= OPTION_SMC;
551 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
558 static void smc_set_option_cond(const struct tcp_sock *tp,
559 const struct inet_request_sock *ireq,
560 struct tcp_out_options *opts,
561 unsigned int *remaining)
563 #if IS_ENABLED(CONFIG_SMC)
564 if (static_branch_unlikely(&tcp_have_smc)) {
565 if (tp->syn_smc && ireq->smc_ok) {
566 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
567 opts->options |= OPTION_SMC;
568 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
575 /* Compute TCP options for SYN packets. This is not the final
576 * network wire format yet.
578 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
579 struct tcp_out_options *opts,
580 struct tcp_md5sig_key **md5)
582 struct tcp_sock *tp = tcp_sk(sk);
583 unsigned int remaining = MAX_TCP_OPTION_SPACE;
584 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
587 #ifdef CONFIG_TCP_MD5SIG
588 if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
589 *md5 = tp->af_specific->md5_lookup(sk, sk);
591 opts->options |= OPTION_MD5;
592 remaining -= TCPOLEN_MD5SIG_ALIGNED;
597 /* We always get an MSS option. The option bytes which will be seen in
598 * normal data packets should timestamps be used, must be in the MSS
599 * advertised. But we subtract them from tp->mss_cache so that
600 * calculations in tcp_sendmsg are simpler etc. So account for this
601 * fact here if necessary. If we don't do this correctly, as a
602 * receiver we won't recognize data packets as being full sized when we
603 * should, and thus we won't abide by the delayed ACK rules correctly.
604 * SACKs don't matter, we never delay an ACK when we have any of those
606 opts->mss = tcp_advertise_mss(sk);
607 remaining -= TCPOLEN_MSS_ALIGNED;
609 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
610 opts->options |= OPTION_TS;
611 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
612 opts->tsecr = tp->rx_opt.ts_recent;
613 remaining -= TCPOLEN_TSTAMP_ALIGNED;
615 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
616 opts->ws = tp->rx_opt.rcv_wscale;
617 opts->options |= OPTION_WSCALE;
618 remaining -= TCPOLEN_WSCALE_ALIGNED;
620 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
621 opts->options |= OPTION_SACK_ADVERTISE;
622 if (unlikely(!(OPTION_TS & opts->options)))
623 remaining -= TCPOLEN_SACKPERM_ALIGNED;
626 if (fastopen && fastopen->cookie.len >= 0) {
627 u32 need = fastopen->cookie.len;
629 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
630 TCPOLEN_FASTOPEN_BASE;
631 need = (need + 3) & ~3U; /* Align to 32 bits */
632 if (remaining >= need) {
633 opts->options |= OPTION_FAST_OPEN_COOKIE;
634 opts->fastopen_cookie = &fastopen->cookie;
636 tp->syn_fastopen = 1;
637 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
641 smc_set_option(tp, opts, &remaining);
643 return MAX_TCP_OPTION_SPACE - remaining;
646 /* Set up TCP options for SYN-ACKs. */
647 static unsigned int tcp_synack_options(const struct sock *sk,
648 struct request_sock *req,
649 unsigned int mss, struct sk_buff *skb,
650 struct tcp_out_options *opts,
651 const struct tcp_md5sig_key *md5,
652 struct tcp_fastopen_cookie *foc,
653 enum tcp_synack_type synack_type)
655 struct inet_request_sock *ireq = inet_rsk(req);
656 unsigned int remaining = MAX_TCP_OPTION_SPACE;
658 #ifdef CONFIG_TCP_MD5SIG
660 opts->options |= OPTION_MD5;
661 remaining -= TCPOLEN_MD5SIG_ALIGNED;
663 /* We can't fit any SACK blocks in a packet with MD5 + TS
664 * options. There was discussion about disabling SACK
665 * rather than TS in order to fit in better with old,
666 * buggy kernels, but that was deemed to be unnecessary.
668 if (synack_type != TCP_SYNACK_COOKIE)
669 ireq->tstamp_ok &= !ireq->sack_ok;
673 /* We always send an MSS option. */
675 remaining -= TCPOLEN_MSS_ALIGNED;
677 if (likely(ireq->wscale_ok)) {
678 opts->ws = ireq->rcv_wscale;
679 opts->options |= OPTION_WSCALE;
680 remaining -= TCPOLEN_WSCALE_ALIGNED;
682 if (likely(ireq->tstamp_ok)) {
683 opts->options |= OPTION_TS;
684 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
685 opts->tsecr = req->ts_recent;
686 remaining -= TCPOLEN_TSTAMP_ALIGNED;
688 if (likely(ireq->sack_ok)) {
689 opts->options |= OPTION_SACK_ADVERTISE;
690 if (unlikely(!ireq->tstamp_ok))
691 remaining -= TCPOLEN_SACKPERM_ALIGNED;
693 if (foc != NULL && foc->len >= 0) {
696 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
697 TCPOLEN_FASTOPEN_BASE;
698 need = (need + 3) & ~3U; /* Align to 32 bits */
699 if (remaining >= need) {
700 opts->options |= OPTION_FAST_OPEN_COOKIE;
701 opts->fastopen_cookie = foc;
706 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
708 return MAX_TCP_OPTION_SPACE - remaining;
711 /* Compute TCP options for ESTABLISHED sockets. This is not the
712 * final wire format yet.
714 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
715 struct tcp_out_options *opts,
716 struct tcp_md5sig_key **md5)
718 struct tcp_sock *tp = tcp_sk(sk);
719 unsigned int size = 0;
720 unsigned int eff_sacks;
725 #ifdef CONFIG_TCP_MD5SIG
726 if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
727 *md5 = tp->af_specific->md5_lookup(sk, sk);
729 opts->options |= OPTION_MD5;
730 size += TCPOLEN_MD5SIG_ALIGNED;
735 if (likely(tp->rx_opt.tstamp_ok)) {
736 opts->options |= OPTION_TS;
737 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
738 opts->tsecr = tp->rx_opt.ts_recent;
739 size += TCPOLEN_TSTAMP_ALIGNED;
742 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
743 if (unlikely(eff_sacks)) {
744 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
745 opts->num_sack_blocks =
746 min_t(unsigned int, eff_sacks,
747 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
748 TCPOLEN_SACK_PERBLOCK);
749 if (likely(opts->num_sack_blocks))
750 size += TCPOLEN_SACK_BASE_ALIGNED +
751 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
758 /* TCP SMALL QUEUES (TSQ)
760 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
761 * to reduce RTT and bufferbloat.
762 * We do this using a special skb destructor (tcp_wfree).
764 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
765 * needs to be reallocated in a driver.
766 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
768 * Since transmit from skb destructor is forbidden, we use a tasklet
769 * to process all sockets that eventually need to send more skbs.
770 * We use one tasklet per cpu, with its own queue of sockets.
773 struct tasklet_struct tasklet;
774 struct list_head head; /* queue of tcp sockets */
776 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
778 static void tcp_tsq_write(struct sock *sk)
780 if ((1 << sk->sk_state) &
781 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
782 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
783 struct tcp_sock *tp = tcp_sk(sk);
785 if (tp->lost_out > tp->retrans_out &&
786 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
787 tcp_mstamp_refresh(tp);
788 tcp_xmit_retransmit_queue(sk);
791 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
796 static void tcp_tsq_handler(struct sock *sk)
799 if (!sock_owned_by_user(sk))
801 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
806 * One tasklet per cpu tries to send more skbs.
807 * We run in tasklet context but need to disable irqs when
808 * transferring tsq->head because tcp_wfree() might
809 * interrupt us (non NAPI drivers)
811 static void tcp_tasklet_func(unsigned long data)
813 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
816 struct list_head *q, *n;
820 local_irq_save(flags);
821 list_splice_init(&tsq->head, &list);
822 local_irq_restore(flags);
824 list_for_each_safe(q, n, &list) {
825 tp = list_entry(q, struct tcp_sock, tsq_node);
826 list_del(&tp->tsq_node);
828 sk = (struct sock *)tp;
829 smp_mb__before_atomic();
830 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
837 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
838 TCPF_WRITE_TIMER_DEFERRED | \
839 TCPF_DELACK_TIMER_DEFERRED | \
840 TCPF_MTU_REDUCED_DEFERRED)
842 * tcp_release_cb - tcp release_sock() callback
845 * called from release_sock() to perform protocol dependent
846 * actions before socket release.
848 void tcp_release_cb(struct sock *sk)
850 unsigned long flags, nflags;
852 /* perform an atomic operation only if at least one flag is set */
854 flags = sk->sk_tsq_flags;
855 if (!(flags & TCP_DEFERRED_ALL))
857 nflags = flags & ~TCP_DEFERRED_ALL;
858 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
860 if (flags & TCPF_TSQ_DEFERRED) {
864 /* Here begins the tricky part :
865 * We are called from release_sock() with :
867 * 2) sk_lock.slock spinlock held
868 * 3) socket owned by us (sk->sk_lock.owned == 1)
870 * But following code is meant to be called from BH handlers,
871 * so we should keep BH disabled, but early release socket ownership
873 sock_release_ownership(sk);
875 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
876 tcp_write_timer_handler(sk);
879 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
880 tcp_delack_timer_handler(sk);
883 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
884 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
888 EXPORT_SYMBOL(tcp_release_cb);
890 void __init tcp_tasklet_init(void)
894 for_each_possible_cpu(i) {
895 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
897 INIT_LIST_HEAD(&tsq->head);
898 tasklet_init(&tsq->tasklet,
905 * Write buffer destructor automatically called from kfree_skb.
906 * We can't xmit new skbs from this context, as we might already
909 void tcp_wfree(struct sk_buff *skb)
911 struct sock *sk = skb->sk;
912 struct tcp_sock *tp = tcp_sk(sk);
913 unsigned long flags, nval, oval;
915 /* Keep one reference on sk_wmem_alloc.
916 * Will be released by sk_free() from here or tcp_tasklet_func()
918 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
920 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
921 * Wait until our queues (qdisc + devices) are drained.
923 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
924 * - chance for incoming ACK (processed by another cpu maybe)
925 * to migrate this flow (skb->ooo_okay will be eventually set)
927 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
930 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
931 struct tsq_tasklet *tsq;
934 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
937 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
938 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
942 /* queue this socket to tasklet queue */
943 local_irq_save(flags);
944 tsq = this_cpu_ptr(&tsq_tasklet);
945 empty = list_empty(&tsq->head);
946 list_add(&tp->tsq_node, &tsq->head);
948 tasklet_schedule(&tsq->tasklet);
949 local_irq_restore(flags);
956 /* Note: Called under soft irq.
957 * We can call TCP stack right away, unless socket is owned by user.
959 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
961 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
962 struct sock *sk = (struct sock *)tp;
967 return HRTIMER_NORESTART;
970 static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
972 struct tcp_sock *tp = tcp_sk(sk);
977 if (!tcp_needs_internal_pacing(sk))
979 rate = sk->sk_pacing_rate;
980 if (!rate || rate == ~0U)
983 len_ns = (u64)skb->len * NSEC_PER_SEC;
984 do_div(len_ns, rate);
986 /* If hrtimer is already armed, then our caller has not
987 * used tcp_pacing_check().
989 if (unlikely(hrtimer_is_queued(&tp->pacing_timer))) {
990 expire = hrtimer_get_softexpires(&tp->pacing_timer);
991 if (ktime_after(expire, now))
993 if (hrtimer_try_to_cancel(&tp->pacing_timer) == 1)
996 hrtimer_start(&tp->pacing_timer, ktime_add_ns(now, len_ns),
997 HRTIMER_MODE_ABS_PINNED_SOFT);
1001 static bool tcp_pacing_check(const struct sock *sk)
1003 return tcp_needs_internal_pacing(sk) &&
1004 hrtimer_is_queued(&tcp_sk(sk)->pacing_timer);
1007 static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb)
1009 skb->skb_mstamp = tp->tcp_mstamp;
1010 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1013 /* This routine actually transmits TCP packets queued in by
1014 * tcp_do_sendmsg(). This is used by both the initial
1015 * transmission and possible later retransmissions.
1016 * All SKB's seen here are completely headerless. It is our
1017 * job to build the TCP header, and pass the packet down to
1018 * IP so it can do the same plus pass the packet off to the
1021 * We are working here with either a clone of the original
1022 * SKB, or a fresh unique copy made by the retransmit engine.
1024 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1025 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1027 const struct inet_connection_sock *icsk = inet_csk(sk);
1028 struct inet_sock *inet;
1029 struct tcp_sock *tp;
1030 struct tcp_skb_cb *tcb;
1031 struct tcp_out_options opts;
1032 unsigned int tcp_options_size, tcp_header_size;
1033 struct sk_buff *oskb = NULL;
1034 struct tcp_md5sig_key *md5;
1038 BUG_ON(!skb || !tcp_skb_pcount(skb));
1042 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1046 tcp_skb_tsorted_save(oskb) {
1047 if (unlikely(skb_cloned(oskb)))
1048 skb = pskb_copy(oskb, gfp_mask);
1050 skb = skb_clone(oskb, gfp_mask);
1051 } tcp_skb_tsorted_restore(oskb);
1056 skb->skb_mstamp = tp->tcp_mstamp;
1059 tcb = TCP_SKB_CB(skb);
1060 memset(&opts, 0, sizeof(opts));
1062 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1063 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1065 tcp_options_size = tcp_established_options(sk, skb, &opts,
1067 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1069 /* if no packet is in qdisc/device queue, then allow XPS to select
1070 * another queue. We can be called from tcp_tsq_handler()
1071 * which holds one reference to sk.
1073 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1074 * One way to get this would be to set skb->truesize = 2 on them.
1076 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1078 /* If we had to use memory reserve to allocate this skb,
1079 * this might cause drops if packet is looped back :
1080 * Other socket might not have SOCK_MEMALLOC.
1081 * Packets not looped back do not care about pfmemalloc.
1083 skb->pfmemalloc = 0;
1085 skb_push(skb, tcp_header_size);
1086 skb_reset_transport_header(skb);
1090 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1091 skb_set_hash_from_sk(skb, sk);
1092 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1094 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1096 /* Build TCP header and checksum it. */
1097 th = (struct tcphdr *)skb->data;
1098 th->source = inet->inet_sport;
1099 th->dest = inet->inet_dport;
1100 th->seq = htonl(tcb->seq);
1101 th->ack_seq = htonl(rcv_nxt);
1102 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1108 /* The urg_mode check is necessary during a below snd_una win probe */
1109 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1110 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1111 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1113 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1114 th->urg_ptr = htons(0xFFFF);
1119 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1120 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1121 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1122 th->window = htons(tcp_select_window(sk));
1123 tcp_ecn_send(sk, skb, th, tcp_header_size);
1125 /* RFC1323: The window in SYN & SYN/ACK segments
1128 th->window = htons(min(tp->rcv_wnd, 65535U));
1130 #ifdef CONFIG_TCP_MD5SIG
1131 /* Calculate the MD5 hash, as we have all we need now */
1133 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1134 tp->af_specific->calc_md5_hash(opts.hash_location,
1139 icsk->icsk_af_ops->send_check(sk, skb);
1141 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1142 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1144 if (skb->len != tcp_header_size) {
1145 tcp_event_data_sent(tp, sk);
1146 tp->data_segs_out += tcp_skb_pcount(skb);
1147 tp->bytes_sent += skb->len - tcp_header_size;
1148 tcp_internal_pacing(sk, skb);
1151 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1152 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1153 tcp_skb_pcount(skb));
1155 tp->segs_out += tcp_skb_pcount(skb);
1156 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1157 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1158 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1160 /* Our usage of tstamp should remain private */
1163 /* Cleanup our debris for IP stacks */
1164 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1165 sizeof(struct inet6_skb_parm)));
1167 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1169 if (unlikely(err > 0)) {
1171 err = net_xmit_eval(err);
1174 tcp_update_skb_after_send(tp, oskb);
1175 tcp_rate_skb_sent(sk, oskb);
1180 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1183 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1184 tcp_sk(sk)->rcv_nxt);
1187 /* This routine just queues the buffer for sending.
1189 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1190 * otherwise socket can stall.
1192 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1194 struct tcp_sock *tp = tcp_sk(sk);
1196 /* Advance write_seq and place onto the write_queue. */
1197 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1198 __skb_header_release(skb);
1199 tcp_add_write_queue_tail(sk, skb);
1200 sk->sk_wmem_queued += skb->truesize;
1201 sk_mem_charge(sk, skb->truesize);
1204 /* Initialize TSO segments for a packet. */
1205 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1207 if (skb->len <= mss_now) {
1208 /* Avoid the costly divide in the normal
1211 tcp_skb_pcount_set(skb, 1);
1212 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1214 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1215 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1219 /* Pcount in the middle of the write queue got changed, we need to do various
1220 * tweaks to fix counters
1222 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1224 struct tcp_sock *tp = tcp_sk(sk);
1226 tp->packets_out -= decr;
1228 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1229 tp->sacked_out -= decr;
1230 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1231 tp->retrans_out -= decr;
1232 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1233 tp->lost_out -= decr;
1235 /* Reno case is special. Sigh... */
1236 if (tcp_is_reno(tp) && decr > 0)
1237 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1239 if (tp->lost_skb_hint &&
1240 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1241 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1242 tp->lost_cnt_hint -= decr;
1244 tcp_verify_left_out(tp);
1247 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1249 return TCP_SKB_CB(skb)->txstamp_ack ||
1250 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1253 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1255 struct skb_shared_info *shinfo = skb_shinfo(skb);
1257 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1258 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1259 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1260 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1262 shinfo->tx_flags &= ~tsflags;
1263 shinfo2->tx_flags |= tsflags;
1264 swap(shinfo->tskey, shinfo2->tskey);
1265 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1266 TCP_SKB_CB(skb)->txstamp_ack = 0;
1270 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1272 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1273 TCP_SKB_CB(skb)->eor = 0;
1276 /* Insert buff after skb on the write or rtx queue of sk. */
1277 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1278 struct sk_buff *buff,
1280 enum tcp_queue tcp_queue)
1282 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1283 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1285 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1288 /* Function to create two new TCP segments. Shrinks the given segment
1289 * to the specified size and appends a new segment with the rest of the
1290 * packet to the list. This won't be called frequently, I hope.
1291 * Remember, these are still headerless SKBs at this point.
1293 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1294 struct sk_buff *skb, u32 len,
1295 unsigned int mss_now, gfp_t gfp)
1297 struct tcp_sock *tp = tcp_sk(sk);
1298 struct sk_buff *buff;
1299 int nsize, old_factor;
1304 if (WARN_ON(len > skb->len))
1307 nsize = skb_headlen(skb) - len;
1311 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1312 * We need some allowance to not penalize applications setting small
1314 * Also allow first and last skb in retransmit queue to be split.
1316 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1317 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1318 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1319 skb != tcp_rtx_queue_head(sk) &&
1320 skb != tcp_rtx_queue_tail(sk))) {
1321 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1325 if (skb_unclone(skb, gfp))
1328 /* Get a new skb... force flag on. */
1329 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1331 return -ENOMEM; /* We'll just try again later. */
1333 sk->sk_wmem_queued += buff->truesize;
1334 sk_mem_charge(sk, buff->truesize);
1335 nlen = skb->len - len - nsize;
1336 buff->truesize += nlen;
1337 skb->truesize -= nlen;
1339 /* Correct the sequence numbers. */
1340 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1341 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1342 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1344 /* PSH and FIN should only be set in the second packet. */
1345 flags = TCP_SKB_CB(skb)->tcp_flags;
1346 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1347 TCP_SKB_CB(buff)->tcp_flags = flags;
1348 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1349 tcp_skb_fragment_eor(skb, buff);
1351 skb_split(skb, buff, len);
1353 buff->ip_summed = CHECKSUM_PARTIAL;
1355 buff->tstamp = skb->tstamp;
1356 tcp_fragment_tstamp(skb, buff);
1358 old_factor = tcp_skb_pcount(skb);
1360 /* Fix up tso_factor for both original and new SKB. */
1361 tcp_set_skb_tso_segs(skb, mss_now);
1362 tcp_set_skb_tso_segs(buff, mss_now);
1364 /* Update delivered info for the new segment */
1365 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1367 /* If this packet has been sent out already, we must
1368 * adjust the various packet counters.
1370 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1371 int diff = old_factor - tcp_skb_pcount(skb) -
1372 tcp_skb_pcount(buff);
1375 tcp_adjust_pcount(sk, skb, diff);
1378 /* Link BUFF into the send queue. */
1379 __skb_header_release(buff);
1380 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1381 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1382 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1387 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1388 * data is not copied, but immediately discarded.
1390 static int __pskb_trim_head(struct sk_buff *skb, int len)
1392 struct skb_shared_info *shinfo;
1395 eat = min_t(int, len, skb_headlen(skb));
1397 __skb_pull(skb, eat);
1404 shinfo = skb_shinfo(skb);
1405 for (i = 0; i < shinfo->nr_frags; i++) {
1406 int size = skb_frag_size(&shinfo->frags[i]);
1409 skb_frag_unref(skb, i);
1412 shinfo->frags[k] = shinfo->frags[i];
1414 shinfo->frags[k].page_offset += eat;
1415 skb_frag_size_sub(&shinfo->frags[k], eat);
1421 shinfo->nr_frags = k;
1423 skb->data_len -= len;
1424 skb->len = skb->data_len;
1428 /* Remove acked data from a packet in the transmit queue. */
1429 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1433 if (skb_unclone(skb, GFP_ATOMIC))
1436 delta_truesize = __pskb_trim_head(skb, len);
1438 TCP_SKB_CB(skb)->seq += len;
1439 skb->ip_summed = CHECKSUM_PARTIAL;
1441 if (delta_truesize) {
1442 skb->truesize -= delta_truesize;
1443 sk->sk_wmem_queued -= delta_truesize;
1444 sk_mem_uncharge(sk, delta_truesize);
1445 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1448 /* Any change of skb->len requires recalculation of tso factor. */
1449 if (tcp_skb_pcount(skb) > 1)
1450 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1455 /* Calculate MSS not accounting any TCP options. */
1456 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1458 const struct tcp_sock *tp = tcp_sk(sk);
1459 const struct inet_connection_sock *icsk = inet_csk(sk);
1462 /* Calculate base mss without TCP options:
1463 It is MMS_S - sizeof(tcphdr) of rfc1122
1465 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1467 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1468 if (icsk->icsk_af_ops->net_frag_header_len) {
1469 const struct dst_entry *dst = __sk_dst_get(sk);
1471 if (dst && dst_allfrag(dst))
1472 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1475 /* Clamp it (mss_clamp does not include tcp options) */
1476 if (mss_now > tp->rx_opt.mss_clamp)
1477 mss_now = tp->rx_opt.mss_clamp;
1479 /* Now subtract optional transport overhead */
1480 mss_now -= icsk->icsk_ext_hdr_len;
1482 /* Then reserve room for full set of TCP options and 8 bytes of data */
1483 mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1487 /* Calculate MSS. Not accounting for SACKs here. */
1488 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1490 /* Subtract TCP options size, not including SACKs */
1491 return __tcp_mtu_to_mss(sk, pmtu) -
1492 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1494 EXPORT_SYMBOL(tcp_mtu_to_mss);
1496 /* Inverse of above */
1497 int tcp_mss_to_mtu(struct sock *sk, int mss)
1499 const struct tcp_sock *tp = tcp_sk(sk);
1500 const struct inet_connection_sock *icsk = inet_csk(sk);
1504 tp->tcp_header_len +
1505 icsk->icsk_ext_hdr_len +
1506 icsk->icsk_af_ops->net_header_len;
1508 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1509 if (icsk->icsk_af_ops->net_frag_header_len) {
1510 const struct dst_entry *dst = __sk_dst_get(sk);
1512 if (dst && dst_allfrag(dst))
1513 mtu += icsk->icsk_af_ops->net_frag_header_len;
1517 EXPORT_SYMBOL(tcp_mss_to_mtu);
1519 /* MTU probing init per socket */
1520 void tcp_mtup_init(struct sock *sk)
1522 struct tcp_sock *tp = tcp_sk(sk);
1523 struct inet_connection_sock *icsk = inet_csk(sk);
1524 struct net *net = sock_net(sk);
1526 icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1527 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1528 icsk->icsk_af_ops->net_header_len;
1529 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1530 icsk->icsk_mtup.probe_size = 0;
1531 if (icsk->icsk_mtup.enabled)
1532 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1534 EXPORT_SYMBOL(tcp_mtup_init);
1536 /* This function synchronize snd mss to current pmtu/exthdr set.
1538 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1539 for TCP options, but includes only bare TCP header.
1541 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1542 It is minimum of user_mss and mss received with SYN.
1543 It also does not include TCP options.
1545 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1547 tp->mss_cache is current effective sending mss, including
1548 all tcp options except for SACKs. It is evaluated,
1549 taking into account current pmtu, but never exceeds
1550 tp->rx_opt.mss_clamp.
1552 NOTE1. rfc1122 clearly states that advertised MSS
1553 DOES NOT include either tcp or ip options.
1555 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1556 are READ ONLY outside this function. --ANK (980731)
1558 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1560 struct tcp_sock *tp = tcp_sk(sk);
1561 struct inet_connection_sock *icsk = inet_csk(sk);
1564 if (icsk->icsk_mtup.search_high > pmtu)
1565 icsk->icsk_mtup.search_high = pmtu;
1567 mss_now = tcp_mtu_to_mss(sk, pmtu);
1568 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1570 /* And store cached results */
1571 icsk->icsk_pmtu_cookie = pmtu;
1572 if (icsk->icsk_mtup.enabled)
1573 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1574 tp->mss_cache = mss_now;
1578 EXPORT_SYMBOL(tcp_sync_mss);
1580 /* Compute the current effective MSS, taking SACKs and IP options,
1581 * and even PMTU discovery events into account.
1583 unsigned int tcp_current_mss(struct sock *sk)
1585 const struct tcp_sock *tp = tcp_sk(sk);
1586 const struct dst_entry *dst = __sk_dst_get(sk);
1588 unsigned int header_len;
1589 struct tcp_out_options opts;
1590 struct tcp_md5sig_key *md5;
1592 mss_now = tp->mss_cache;
1595 u32 mtu = dst_mtu(dst);
1596 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1597 mss_now = tcp_sync_mss(sk, mtu);
1600 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1601 sizeof(struct tcphdr);
1602 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1603 * some common options. If this is an odd packet (because we have SACK
1604 * blocks etc) then our calculated header_len will be different, and
1605 * we have to adjust mss_now correspondingly */
1606 if (header_len != tp->tcp_header_len) {
1607 int delta = (int) header_len - tp->tcp_header_len;
1614 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1615 * As additional protections, we do not touch cwnd in retransmission phases,
1616 * and if application hit its sndbuf limit recently.
1618 static void tcp_cwnd_application_limited(struct sock *sk)
1620 struct tcp_sock *tp = tcp_sk(sk);
1622 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1623 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1624 /* Limited by application or receiver window. */
1625 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1626 u32 win_used = max(tp->snd_cwnd_used, init_win);
1627 if (win_used < tp->snd_cwnd) {
1628 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1629 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1631 tp->snd_cwnd_used = 0;
1633 tp->snd_cwnd_stamp = tcp_jiffies32;
1636 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1638 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1639 struct tcp_sock *tp = tcp_sk(sk);
1641 /* Track the strongest available signal of the degree to which the cwnd
1642 * is fully utilized. If cwnd-limited then remember that fact for the
1643 * current window. If not cwnd-limited then track the maximum number of
1644 * outstanding packets in the current window. (If cwnd-limited then we
1645 * chose to not update tp->max_packets_out to avoid an extra else
1646 * clause with no functional impact.)
1648 if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
1650 (!tp->is_cwnd_limited &&
1651 tp->packets_out > tp->max_packets_out)) {
1652 tp->is_cwnd_limited = is_cwnd_limited;
1653 tp->max_packets_out = tp->packets_out;
1654 tp->cwnd_usage_seq = tp->snd_nxt;
1657 if (tcp_is_cwnd_limited(sk)) {
1658 /* Network is feed fully. */
1659 tp->snd_cwnd_used = 0;
1660 tp->snd_cwnd_stamp = tcp_jiffies32;
1662 /* Network starves. */
1663 if (tp->packets_out > tp->snd_cwnd_used)
1664 tp->snd_cwnd_used = tp->packets_out;
1666 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
1667 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1668 !ca_ops->cong_control)
1669 tcp_cwnd_application_limited(sk);
1671 /* The following conditions together indicate the starvation
1672 * is caused by insufficient sender buffer:
1673 * 1) just sent some data (see tcp_write_xmit)
1674 * 2) not cwnd limited (this else condition)
1675 * 3) no more data to send (tcp_write_queue_empty())
1676 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1678 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1679 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1680 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1681 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1685 /* Minshall's variant of the Nagle send check. */
1686 static bool tcp_minshall_check(const struct tcp_sock *tp)
1688 return after(tp->snd_sml, tp->snd_una) &&
1689 !after(tp->snd_sml, tp->snd_nxt);
1692 /* Update snd_sml if this skb is under mss
1693 * Note that a TSO packet might end with a sub-mss segment
1694 * The test is really :
1695 * if ((skb->len % mss) != 0)
1696 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1697 * But we can avoid doing the divide again given we already have
1698 * skb_pcount = skb->len / mss_now
1700 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1701 const struct sk_buff *skb)
1703 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1704 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1707 /* Return false, if packet can be sent now without violation Nagle's rules:
1708 * 1. It is full sized. (provided by caller in %partial bool)
1709 * 2. Or it contains FIN. (already checked by caller)
1710 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1711 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1712 * With Minshall's modification: all sent small packets are ACKed.
1714 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1718 ((nonagle & TCP_NAGLE_CORK) ||
1719 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1722 /* Return how many segs we'd like on a TSO packet,
1723 * to send one TSO packet per ms
1725 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1730 bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift,
1731 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1733 /* Goal is to send at least one packet per ms,
1734 * not one big TSO packet every 100 ms.
1735 * This preserves ACK clocking and is consistent
1736 * with tcp_tso_should_defer() heuristic.
1738 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1743 /* Return the number of segments we want in the skb we are transmitting.
1744 * See if congestion control module wants to decide; otherwise, autosize.
1746 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1748 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1749 u32 min_tso, tso_segs;
1751 min_tso = ca_ops->min_tso_segs ?
1752 ca_ops->min_tso_segs(sk) :
1753 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
1755 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1756 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1759 /* Returns the portion of skb which can be sent right away */
1760 static unsigned int tcp_mss_split_point(const struct sock *sk,
1761 const struct sk_buff *skb,
1762 unsigned int mss_now,
1763 unsigned int max_segs,
1766 const struct tcp_sock *tp = tcp_sk(sk);
1767 u32 partial, needed, window, max_len;
1769 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1770 max_len = mss_now * max_segs;
1772 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1775 needed = min(skb->len, window);
1777 if (max_len <= needed)
1780 partial = needed % mss_now;
1781 /* If last segment is not a full MSS, check if Nagle rules allow us
1782 * to include this last segment in this skb.
1783 * Otherwise, we'll split the skb at last MSS boundary
1785 if (tcp_nagle_check(partial != 0, tp, nonagle))
1786 return needed - partial;
1791 /* Can at least one segment of SKB be sent right now, according to the
1792 * congestion window rules? If so, return how many segments are allowed.
1794 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1795 const struct sk_buff *skb)
1797 u32 in_flight, cwnd, halfcwnd;
1799 /* Don't be strict about the congestion window for the final FIN. */
1800 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1801 tcp_skb_pcount(skb) == 1)
1804 in_flight = tcp_packets_in_flight(tp);
1805 cwnd = tp->snd_cwnd;
1806 if (in_flight >= cwnd)
1809 /* For better scheduling, ensure we have at least
1810 * 2 GSO packets in flight.
1812 halfcwnd = max(cwnd >> 1, 1U);
1813 return min(halfcwnd, cwnd - in_flight);
1816 /* Initialize TSO state of a skb.
1817 * This must be invoked the first time we consider transmitting
1818 * SKB onto the wire.
1820 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1822 int tso_segs = tcp_skb_pcount(skb);
1824 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1825 tcp_set_skb_tso_segs(skb, mss_now);
1826 tso_segs = tcp_skb_pcount(skb);
1832 /* Return true if the Nagle test allows this packet to be
1835 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1836 unsigned int cur_mss, int nonagle)
1838 /* Nagle rule does not apply to frames, which sit in the middle of the
1839 * write_queue (they have no chances to get new data).
1841 * This is implemented in the callers, where they modify the 'nonagle'
1842 * argument based upon the location of SKB in the send queue.
1844 if (nonagle & TCP_NAGLE_PUSH)
1847 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1848 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1851 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1857 /* Does at least the first segment of SKB fit into the send window? */
1858 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1859 const struct sk_buff *skb,
1860 unsigned int cur_mss)
1862 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1864 if (skb->len > cur_mss)
1865 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1867 return !after(end_seq, tcp_wnd_end(tp));
1870 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1871 * which is put after SKB on the list. It is very much like
1872 * tcp_fragment() except that it may make several kinds of assumptions
1873 * in order to speed up the splitting operation. In particular, we
1874 * know that all the data is in scatter-gather pages, and that the
1875 * packet has never been sent out before (and thus is not cloned).
1877 static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1878 struct sk_buff *skb, unsigned int len,
1879 unsigned int mss_now, gfp_t gfp)
1881 struct sk_buff *buff;
1882 int nlen = skb->len - len;
1885 /* All of a TSO frame must be composed of paged data. */
1886 if (skb->len != skb->data_len)
1887 return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp);
1889 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1890 if (unlikely(!buff))
1893 sk->sk_wmem_queued += buff->truesize;
1894 sk_mem_charge(sk, buff->truesize);
1895 buff->truesize += nlen;
1896 skb->truesize -= nlen;
1898 /* Correct the sequence numbers. */
1899 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1900 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1901 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1903 /* PSH and FIN should only be set in the second packet. */
1904 flags = TCP_SKB_CB(skb)->tcp_flags;
1905 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1906 TCP_SKB_CB(buff)->tcp_flags = flags;
1908 /* This packet was never sent out yet, so no SACK bits. */
1909 TCP_SKB_CB(buff)->sacked = 0;
1911 tcp_skb_fragment_eor(skb, buff);
1913 buff->ip_summed = CHECKSUM_PARTIAL;
1914 skb_split(skb, buff, len);
1915 tcp_fragment_tstamp(skb, buff);
1917 /* Fix up tso_factor for both original and new SKB. */
1918 tcp_set_skb_tso_segs(skb, mss_now);
1919 tcp_set_skb_tso_segs(buff, mss_now);
1921 /* Link BUFF into the send queue. */
1922 __skb_header_release(buff);
1923 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1928 /* Try to defer sending, if possible, in order to minimize the amount
1929 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1931 * This algorithm is from John Heffner.
1933 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1934 bool *is_cwnd_limited,
1935 bool *is_rwnd_limited,
1938 const struct inet_connection_sock *icsk = inet_csk(sk);
1939 u32 age, send_win, cong_win, limit, in_flight;
1940 struct tcp_sock *tp = tcp_sk(sk);
1941 struct sk_buff *head;
1944 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1947 /* Avoid bursty behavior by allowing defer
1948 * only if the last write was recent.
1950 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
1953 in_flight = tcp_packets_in_flight(tp);
1955 BUG_ON(tcp_skb_pcount(skb) <= 1);
1956 BUG_ON(tp->snd_cwnd <= in_flight);
1958 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1960 /* From in_flight test above, we know that cwnd > in_flight. */
1961 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1963 limit = min(send_win, cong_win);
1965 /* If a full-sized TSO skb can be sent, do it. */
1966 if (limit >= max_segs * tp->mss_cache)
1969 /* Middle in queue won't get any more data, full sendable already? */
1970 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1973 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1975 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1977 /* If at least some fraction of a window is available,
1980 chunk /= win_divisor;
1984 /* Different approach, try not to defer past a single
1985 * ACK. Receiver should ACK every other full sized
1986 * frame, so if we have space for more than 3 frames
1989 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1993 /* TODO : use tsorted_sent_queue ? */
1994 head = tcp_rtx_queue_head(sk);
1997 age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
1998 /* If next ACK is likely to come too late (half srtt), do not defer */
1999 if (age < (tp->srtt_us >> 4))
2002 /* Ok, it looks like it is advisable to defer.
2003 * Three cases are tracked :
2004 * 1) We are cwnd-limited
2005 * 2) We are rwnd-limited
2006 * 3) We are application limited.
2008 if (cong_win < send_win) {
2009 if (cong_win <= skb->len) {
2010 *is_cwnd_limited = true;
2014 if (send_win <= skb->len) {
2015 *is_rwnd_limited = true;
2020 /* If this packet won't get more data, do not wait. */
2021 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2030 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2032 struct inet_connection_sock *icsk = inet_csk(sk);
2033 struct tcp_sock *tp = tcp_sk(sk);
2034 struct net *net = sock_net(sk);
2038 interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2039 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2040 if (unlikely(delta >= interval * HZ)) {
2041 int mss = tcp_current_mss(sk);
2043 /* Update current search range */
2044 icsk->icsk_mtup.probe_size = 0;
2045 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2046 sizeof(struct tcphdr) +
2047 icsk->icsk_af_ops->net_header_len;
2048 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2050 /* Update probe time stamp */
2051 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2055 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2057 struct sk_buff *skb, *next;
2059 skb = tcp_send_head(sk);
2060 tcp_for_write_queue_from_safe(skb, next, sk) {
2061 if (len <= skb->len)
2064 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2073 /* Create a new MTU probe if we are ready.
2074 * MTU probe is regularly attempting to increase the path MTU by
2075 * deliberately sending larger packets. This discovers routing
2076 * changes resulting in larger path MTUs.
2078 * Returns 0 if we should wait to probe (no cwnd available),
2079 * 1 if a probe was sent,
2082 static int tcp_mtu_probe(struct sock *sk)
2084 struct inet_connection_sock *icsk = inet_csk(sk);
2085 struct tcp_sock *tp = tcp_sk(sk);
2086 struct sk_buff *skb, *nskb, *next;
2087 struct net *net = sock_net(sk);
2094 /* Not currently probing/verifying,
2096 * have enough cwnd, and
2097 * not SACKing (the variable headers throw things off)
2099 if (likely(!icsk->icsk_mtup.enabled ||
2100 icsk->icsk_mtup.probe_size ||
2101 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2102 tp->snd_cwnd < 11 ||
2103 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2106 /* Use binary search for probe_size between tcp_mss_base,
2107 * and current mss_clamp. if (search_high - search_low)
2108 * smaller than a threshold, backoff from probing.
2110 mss_now = tcp_current_mss(sk);
2111 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2112 icsk->icsk_mtup.search_low) >> 1);
2113 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2114 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2115 /* When misfortune happens, we are reprobing actively,
2116 * and then reprobe timer has expired. We stick with current
2117 * probing process by not resetting search range to its orignal.
2119 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2120 interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2121 /* Check whether enough time has elaplased for
2122 * another round of probing.
2124 tcp_mtu_check_reprobe(sk);
2128 /* Have enough data in the send queue to probe? */
2129 if (tp->write_seq - tp->snd_nxt < size_needed)
2132 if (tp->snd_wnd < size_needed)
2134 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2137 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2138 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2139 if (!tcp_packets_in_flight(tp))
2145 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2148 if (tcp_pacing_check(sk))
2151 /* We're allowed to probe. Build it now. */
2152 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2155 sk->sk_wmem_queued += nskb->truesize;
2156 sk_mem_charge(sk, nskb->truesize);
2158 skb = tcp_send_head(sk);
2160 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2161 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2162 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2163 TCP_SKB_CB(nskb)->sacked = 0;
2165 nskb->ip_summed = CHECKSUM_PARTIAL;
2167 tcp_insert_write_queue_before(nskb, skb, sk);
2168 tcp_highest_sack_replace(sk, skb, nskb);
2171 tcp_for_write_queue_from_safe(skb, next, sk) {
2172 copy = min_t(int, skb->len, probe_size - len);
2173 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2175 if (skb->len <= copy) {
2176 /* We've eaten all the data from this skb.
2178 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2179 /* If this is the last SKB we copy and eor is set
2180 * we need to propagate it to the new skb.
2182 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2183 tcp_skb_collapse_tstamp(nskb, skb);
2184 tcp_unlink_write_queue(skb, sk);
2185 sk_wmem_free_skb(sk, skb);
2187 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2188 ~(TCPHDR_FIN|TCPHDR_PSH);
2189 if (!skb_shinfo(skb)->nr_frags) {
2190 skb_pull(skb, copy);
2192 __pskb_trim_head(skb, copy);
2193 tcp_set_skb_tso_segs(skb, mss_now);
2195 TCP_SKB_CB(skb)->seq += copy;
2200 if (len >= probe_size)
2203 tcp_init_tso_segs(nskb, nskb->len);
2205 /* We're ready to send. If this fails, the probe will
2206 * be resegmented into mss-sized pieces by tcp_write_xmit().
2208 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2209 /* Decrement cwnd here because we are sending
2210 * effectively two packets. */
2212 tcp_event_new_data_sent(sk, nskb);
2214 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2215 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2216 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2224 /* TCP Small Queues :
2225 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2226 * (These limits are doubled for retransmits)
2228 * - better RTT estimation and ACK scheduling
2231 * Alas, some drivers / subsystems require a fair amount
2232 * of queued bytes to ensure line rate.
2233 * One example is wifi aggregation (802.11 AMPDU)
2235 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2236 unsigned int factor)
2240 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift);
2241 limit = min_t(u32, limit,
2242 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2245 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2246 /* Always send skb if rtx queue is empty.
2247 * No need to wait for TX completion to call us back,
2248 * after softirq/tasklet schedule.
2249 * This helps when TX completions are delayed too much.
2251 if (tcp_rtx_queue_empty(sk))
2254 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2255 /* It is possible TX completion already happened
2256 * before we set TSQ_THROTTLED, so we must
2257 * test again the condition.
2259 smp_mb__after_atomic();
2260 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2266 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2268 const u32 now = tcp_jiffies32;
2269 enum tcp_chrono old = tp->chrono_type;
2271 if (old > TCP_CHRONO_UNSPEC)
2272 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2273 tp->chrono_start = now;
2274 tp->chrono_type = new;
2277 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2279 struct tcp_sock *tp = tcp_sk(sk);
2281 /* If there are multiple conditions worthy of tracking in a
2282 * chronograph then the highest priority enum takes precedence
2283 * over the other conditions. So that if something "more interesting"
2284 * starts happening, stop the previous chrono and start a new one.
2286 if (type > tp->chrono_type)
2287 tcp_chrono_set(tp, type);
2290 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2292 struct tcp_sock *tp = tcp_sk(sk);
2295 /* There are multiple conditions worthy of tracking in a
2296 * chronograph, so that the highest priority enum takes
2297 * precedence over the other conditions (see tcp_chrono_start).
2298 * If a condition stops, we only stop chrono tracking if
2299 * it's the "most interesting" or current chrono we are
2300 * tracking and starts busy chrono if we have pending data.
2302 if (tcp_rtx_and_write_queues_empty(sk))
2303 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2304 else if (type == tp->chrono_type)
2305 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2308 /* This routine writes packets to the network. It advances the
2309 * send_head. This happens as incoming acks open up the remote
2312 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2313 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2314 * account rare use of URG, this is not a big flaw.
2316 * Send at most one packet when push_one > 0. Temporarily ignore
2317 * cwnd limit to force at most one packet out when push_one == 2.
2319 * Returns true, if no segments are in flight and we have queued segments,
2320 * but cannot send anything now because of SWS or another problem.
2322 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2323 int push_one, gfp_t gfp)
2325 struct tcp_sock *tp = tcp_sk(sk);
2326 struct sk_buff *skb;
2327 unsigned int tso_segs, sent_pkts;
2330 bool is_cwnd_limited = false, is_rwnd_limited = false;
2335 tcp_mstamp_refresh(tp);
2337 /* Do MTU probing. */
2338 result = tcp_mtu_probe(sk);
2341 } else if (result > 0) {
2346 max_segs = tcp_tso_segs(sk, mss_now);
2347 while ((skb = tcp_send_head(sk))) {
2350 if (tcp_pacing_check(sk))
2353 tso_segs = tcp_init_tso_segs(skb, mss_now);
2356 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2357 /* "skb_mstamp" is used as a start point for the retransmit timer */
2358 tcp_update_skb_after_send(tp, skb);
2359 goto repair; /* Skip network transmission */
2362 cwnd_quota = tcp_cwnd_test(tp, skb);
2365 /* Force out a loss probe pkt. */
2371 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2372 is_rwnd_limited = true;
2376 if (tso_segs == 1) {
2377 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2378 (tcp_skb_is_last(sk, skb) ?
2379 nonagle : TCP_NAGLE_PUSH))))
2383 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2384 &is_rwnd_limited, max_segs))
2389 if (tso_segs > 1 && !tcp_urg_mode(tp))
2390 limit = tcp_mss_split_point(sk, skb, mss_now,
2396 if (skb->len > limit &&
2397 unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2398 skb, limit, mss_now, gfp)))
2401 if (tcp_small_queue_check(sk, skb, 0))
2404 /* Argh, we hit an empty skb(), presumably a thread
2405 * is sleeping in sendmsg()/sk_stream_wait_memory().
2406 * We do not want to send a pure-ack packet and have
2407 * a strange looking rtx queue with empty packet(s).
2409 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2412 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2416 /* Advance the send_head. This one is sent out.
2417 * This call will increment packets_out.
2419 tcp_event_new_data_sent(sk, skb);
2421 tcp_minshall_update(tp, mss_now, skb);
2422 sent_pkts += tcp_skb_pcount(skb);
2428 if (is_rwnd_limited)
2429 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2431 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2433 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2434 if (likely(sent_pkts || is_cwnd_limited))
2435 tcp_cwnd_validate(sk, is_cwnd_limited);
2437 if (likely(sent_pkts)) {
2438 if (tcp_in_cwnd_reduction(sk))
2439 tp->prr_out += sent_pkts;
2441 /* Send one loss probe per tail loss episode. */
2443 tcp_schedule_loss_probe(sk, false);
2446 return !tp->packets_out && !tcp_write_queue_empty(sk);
2449 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2451 struct inet_connection_sock *icsk = inet_csk(sk);
2452 struct tcp_sock *tp = tcp_sk(sk);
2453 u32 timeout, rto_delta_us;
2456 /* Don't do any loss probe on a Fast Open connection before 3WHS
2459 if (tp->fastopen_rsk)
2462 early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
2463 /* Schedule a loss probe in 2*RTT for SACK capable connections
2464 * not in loss recovery, that are either limited by cwnd or application.
2466 if ((early_retrans != 3 && early_retrans != 4) ||
2467 !tp->packets_out || !tcp_is_sack(tp) ||
2468 (icsk->icsk_ca_state != TCP_CA_Open &&
2469 icsk->icsk_ca_state != TCP_CA_CWR))
2472 /* Probe timeout is 2*rtt. Add minimum RTO to account
2473 * for delayed ack when there's one outstanding packet. If no RTT
2474 * sample is available then probe after TCP_TIMEOUT_INIT.
2477 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2478 if (tp->packets_out == 1)
2479 timeout += TCP_RTO_MIN;
2481 timeout += TCP_TIMEOUT_MIN;
2483 timeout = TCP_TIMEOUT_INIT;
2486 /* If the RTO formula yields an earlier time, then use that time. */
2487 rto_delta_us = advancing_rto ?
2488 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2489 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2490 if (rto_delta_us > 0)
2491 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2493 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2498 /* Thanks to skb fast clones, we can detect if a prior transmit of
2499 * a packet is still in a qdisc or driver queue.
2500 * In this case, there is very little point doing a retransmit !
2502 static bool skb_still_in_host_queue(const struct sock *sk,
2503 const struct sk_buff *skb)
2505 if (unlikely(skb_fclone_busy(sk, skb))) {
2506 NET_INC_STATS(sock_net(sk),
2507 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2513 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2514 * retransmit the last segment.
2516 void tcp_send_loss_probe(struct sock *sk)
2518 struct tcp_sock *tp = tcp_sk(sk);
2519 struct sk_buff *skb;
2521 int mss = tcp_current_mss(sk);
2523 /* At most one outstanding TLP */
2524 if (tp->tlp_high_seq)
2527 tp->tlp_retrans = 0;
2528 skb = tcp_send_head(sk);
2529 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2530 pcount = tp->packets_out;
2531 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2532 if (tp->packets_out > pcount)
2536 skb = skb_rb_last(&sk->tcp_rtx_queue);
2537 if (unlikely(!skb)) {
2538 WARN_ONCE(tp->packets_out,
2539 "invalid inflight: %u state %u cwnd %u mss %d\n",
2540 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2541 inet_csk(sk)->icsk_pending = 0;
2545 if (skb_still_in_host_queue(sk, skb))
2548 pcount = tcp_skb_pcount(skb);
2549 if (WARN_ON(!pcount))
2552 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2553 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2554 (pcount - 1) * mss, mss,
2557 skb = skb_rb_next(skb);
2560 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2563 if (__tcp_retransmit_skb(sk, skb, 1))
2566 tp->tlp_retrans = 1;
2569 /* Record snd_nxt for loss detection. */
2570 tp->tlp_high_seq = tp->snd_nxt;
2572 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2573 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2574 inet_csk(sk)->icsk_pending = 0;
2579 /* Push out any pending frames which were held back due to
2580 * TCP_CORK or attempt at coalescing tiny packets.
2581 * The socket must be locked by the caller.
2583 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2586 /* If we are closed, the bytes will have to remain here.
2587 * In time closedown will finish, we empty the write queue and
2588 * all will be happy.
2590 if (unlikely(sk->sk_state == TCP_CLOSE))
2593 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2594 sk_gfp_mask(sk, GFP_ATOMIC)))
2595 tcp_check_probe_timer(sk);
2598 /* Send _single_ skb sitting at the send head. This function requires
2599 * true push pending frames to setup probe timer etc.
2601 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2603 struct sk_buff *skb = tcp_send_head(sk);
2605 BUG_ON(!skb || skb->len < mss_now);
2607 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2610 /* This function returns the amount that we can raise the
2611 * usable window based on the following constraints
2613 * 1. The window can never be shrunk once it is offered (RFC 793)
2614 * 2. We limit memory per socket
2617 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2618 * RECV.NEXT + RCV.WIN fixed until:
2619 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2621 * i.e. don't raise the right edge of the window until you can raise
2622 * it at least MSS bytes.
2624 * Unfortunately, the recommended algorithm breaks header prediction,
2625 * since header prediction assumes th->window stays fixed.
2627 * Strictly speaking, keeping th->window fixed violates the receiver
2628 * side SWS prevention criteria. The problem is that under this rule
2629 * a stream of single byte packets will cause the right side of the
2630 * window to always advance by a single byte.
2632 * Of course, if the sender implements sender side SWS prevention
2633 * then this will not be a problem.
2635 * BSD seems to make the following compromise:
2637 * If the free space is less than the 1/4 of the maximum
2638 * space available and the free space is less than 1/2 mss,
2639 * then set the window to 0.
2640 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2641 * Otherwise, just prevent the window from shrinking
2642 * and from being larger than the largest representable value.
2644 * This prevents incremental opening of the window in the regime
2645 * where TCP is limited by the speed of the reader side taking
2646 * data out of the TCP receive queue. It does nothing about
2647 * those cases where the window is constrained on the sender side
2648 * because the pipeline is full.
2650 * BSD also seems to "accidentally" limit itself to windows that are a
2651 * multiple of MSS, at least until the free space gets quite small.
2652 * This would appear to be a side effect of the mbuf implementation.
2653 * Combining these two algorithms results in the observed behavior
2654 * of having a fixed window size at almost all times.
2656 * Below we obtain similar behavior by forcing the offered window to
2657 * a multiple of the mss when it is feasible to do so.
2659 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2660 * Regular options like TIMESTAMP are taken into account.
2662 u32 __tcp_select_window(struct sock *sk)
2664 struct inet_connection_sock *icsk = inet_csk(sk);
2665 struct tcp_sock *tp = tcp_sk(sk);
2666 /* MSS for the peer's data. Previous versions used mss_clamp
2667 * here. I don't know if the value based on our guesses
2668 * of peer's MSS is better for the performance. It's more correct
2669 * but may be worse for the performance because of rcv_mss
2670 * fluctuations. --SAW 1998/11/1
2672 int mss = icsk->icsk_ack.rcv_mss;
2673 int free_space = tcp_space(sk);
2674 int allowed_space = tcp_full_space(sk);
2675 int full_space = min_t(int, tp->window_clamp, allowed_space);
2678 if (unlikely(mss > full_space)) {
2683 if (free_space < (full_space >> 1)) {
2684 icsk->icsk_ack.quick = 0;
2686 if (tcp_under_memory_pressure(sk))
2687 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2690 /* free_space might become our new window, make sure we don't
2691 * increase it due to wscale.
2693 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2695 /* if free space is less than mss estimate, or is below 1/16th
2696 * of the maximum allowed, try to move to zero-window, else
2697 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2698 * new incoming data is dropped due to memory limits.
2699 * With large window, mss test triggers way too late in order
2700 * to announce zero window in time before rmem limit kicks in.
2702 if (free_space < (allowed_space >> 4) || free_space < mss)
2706 if (free_space > tp->rcv_ssthresh)
2707 free_space = tp->rcv_ssthresh;
2709 /* Don't do rounding if we are using window scaling, since the
2710 * scaled window will not line up with the MSS boundary anyway.
2712 if (tp->rx_opt.rcv_wscale) {
2713 window = free_space;
2715 /* Advertise enough space so that it won't get scaled away.
2716 * Import case: prevent zero window announcement if
2717 * 1<<rcv_wscale > mss.
2719 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2721 window = tp->rcv_wnd;
2722 /* Get the largest window that is a nice multiple of mss.
2723 * Window clamp already applied above.
2724 * If our current window offering is within 1 mss of the
2725 * free space we just keep it. This prevents the divide
2726 * and multiply from happening most of the time.
2727 * We also don't do any window rounding when the free space
2730 if (window <= free_space - mss || window > free_space)
2731 window = rounddown(free_space, mss);
2732 else if (mss == full_space &&
2733 free_space > window + (full_space >> 1))
2734 window = free_space;
2740 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2741 const struct sk_buff *next_skb)
2743 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2744 const struct skb_shared_info *next_shinfo =
2745 skb_shinfo(next_skb);
2746 struct skb_shared_info *shinfo = skb_shinfo(skb);
2748 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2749 shinfo->tskey = next_shinfo->tskey;
2750 TCP_SKB_CB(skb)->txstamp_ack |=
2751 TCP_SKB_CB(next_skb)->txstamp_ack;
2755 /* Collapses two adjacent SKB's during retransmission. */
2756 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2758 struct tcp_sock *tp = tcp_sk(sk);
2759 struct sk_buff *next_skb = skb_rb_next(skb);
2762 next_skb_size = next_skb->len;
2764 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2766 if (next_skb_size) {
2767 if (next_skb_size <= skb_availroom(skb))
2768 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2770 else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
2773 tcp_highest_sack_replace(sk, next_skb, skb);
2775 /* Update sequence range on original skb. */
2776 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2778 /* Merge over control information. This moves PSH/FIN etc. over */
2779 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2781 /* All done, get rid of second SKB and account for it so
2782 * packet counting does not break.
2784 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2785 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2787 /* changed transmit queue under us so clear hints */
2788 tcp_clear_retrans_hints_partial(tp);
2789 if (next_skb == tp->retransmit_skb_hint)
2790 tp->retransmit_skb_hint = skb;
2792 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2794 tcp_skb_collapse_tstamp(skb, next_skb);
2796 tcp_rtx_queue_unlink_and_free(next_skb, sk);
2800 /* Check if coalescing SKBs is legal. */
2801 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2803 if (tcp_skb_pcount(skb) > 1)
2805 if (skb_cloned(skb))
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 (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
2826 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2829 skb_rbtree_walk_from_safe(skb, tmp) {
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);
2894 avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2896 /* If receiver has shrunk his window, and skb is out of
2897 * new window, do not retransmit it. The exception is the
2898 * case, when window is shrunk to zero. In this case
2899 * our retransmit of one segment serves as a zero window probe.
2901 if (avail_wnd <= 0) {
2902 if (TCP_SKB_CB(skb)->seq != tp->snd_una)
2904 avail_wnd = cur_mss;
2907 len = cur_mss * segs;
2908 if (len > avail_wnd) {
2909 len = rounddown(avail_wnd, cur_mss);
2913 if (skb->len > len) {
2914 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2915 cur_mss, GFP_ATOMIC))
2916 return -ENOMEM; /* We'll try again later. */
2918 if (skb_unclone(skb, GFP_ATOMIC))
2921 diff = tcp_skb_pcount(skb);
2922 tcp_set_skb_tso_segs(skb, cur_mss);
2923 diff -= tcp_skb_pcount(skb);
2925 tcp_adjust_pcount(sk, skb, diff);
2926 avail_wnd = min_t(int, avail_wnd, cur_mss);
2927 if (skb->len < avail_wnd)
2928 tcp_retrans_try_collapse(sk, skb, avail_wnd);
2931 /* RFC3168, section 6.1.1.1. ECN fallback */
2932 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2933 tcp_ecn_clear_syn(sk, skb);
2935 /* Update global and local TCP statistics. */
2936 segs = tcp_skb_pcount(skb);
2937 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2938 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2939 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2940 tp->total_retrans += segs;
2941 tp->bytes_retrans += skb->len;
2943 /* make sure skb->data is aligned on arches that require it
2944 * and check if ack-trimming & collapsing extended the headroom
2945 * beyond what csum_start can cover.
2947 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2948 skb_headroom(skb) >= 0xFFFF)) {
2949 struct sk_buff *nskb;
2951 tcp_skb_tsorted_save(skb) {
2952 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2953 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2955 } tcp_skb_tsorted_restore(skb);
2958 tcp_update_skb_after_send(tp, skb);
2959 tcp_rate_skb_sent(sk, skb);
2962 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2965 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
2966 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
2967 TCP_SKB_CB(skb)->seq, segs, err);
2970 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2971 trace_tcp_retransmit_skb(sk, skb);
2972 } else if (err != -EBUSY) {
2973 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
2978 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2980 struct tcp_sock *tp = tcp_sk(sk);
2981 int err = __tcp_retransmit_skb(sk, skb, segs);
2984 #if FASTRETRANS_DEBUG > 0
2985 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2986 net_dbg_ratelimited("retrans_out leaked\n");
2989 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2990 tp->retrans_out += tcp_skb_pcount(skb);
2993 /* Save stamp of the first (attempted) retransmit. */
2994 if (!tp->retrans_stamp)
2995 tp->retrans_stamp = tcp_skb_timestamp(skb);
2997 if (tp->undo_retrans < 0)
2998 tp->undo_retrans = 0;
2999 tp->undo_retrans += tcp_skb_pcount(skb);
3003 /* This gets called after a retransmit timeout, and the initially
3004 * retransmitted data is acknowledged. It tries to continue
3005 * resending the rest of the retransmit queue, until either
3006 * we've sent it all or the congestion window limit is reached.
3008 void tcp_xmit_retransmit_queue(struct sock *sk)
3010 const struct inet_connection_sock *icsk = inet_csk(sk);
3011 struct sk_buff *skb, *rtx_head, *hole = NULL;
3012 struct tcp_sock *tp = tcp_sk(sk);
3016 if (!tp->packets_out)
3019 rtx_head = tcp_rtx_queue_head(sk);
3020 skb = tp->retransmit_skb_hint ?: rtx_head;
3021 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3022 skb_rbtree_walk_from(skb) {
3026 if (tcp_pacing_check(sk))
3029 /* we could do better than to assign each time */
3031 tp->retransmit_skb_hint = skb;
3033 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
3036 sacked = TCP_SKB_CB(skb)->sacked;
3037 /* In case tcp_shift_skb_data() have aggregated large skbs,
3038 * we need to make sure not sending too bigs TSO packets
3040 segs = min_t(int, segs, max_segs);
3042 if (tp->retrans_out >= tp->lost_out) {
3044 } else if (!(sacked & TCPCB_LOST)) {
3045 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3050 if (icsk->icsk_ca_state != TCP_CA_Loss)
3051 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3053 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3056 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3059 if (tcp_small_queue_check(sk, skb, 1))
3062 if (tcp_retransmit_skb(sk, skb, segs))
3065 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3067 if (tcp_in_cwnd_reduction(sk))
3068 tp->prr_out += tcp_skb_pcount(skb);
3070 if (skb == rtx_head &&
3071 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3072 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3073 inet_csk(sk)->icsk_rto,
3078 /* We allow to exceed memory limits for FIN packets to expedite
3079 * connection tear down and (memory) recovery.
3080 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3081 * or even be forced to close flow without any FIN.
3082 * In general, we want to allow one skb per socket to avoid hangs
3083 * with edge trigger epoll()
3085 void sk_forced_mem_schedule(struct sock *sk, int size)
3089 delta = size - sk->sk_forward_alloc;
3092 amt = sk_mem_pages(delta);
3093 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3094 sk_memory_allocated_add(sk, amt);
3096 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3097 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3100 /* Send a FIN. The caller locks the socket for us.
3101 * We should try to send a FIN packet really hard, but eventually give up.
3103 void tcp_send_fin(struct sock *sk)
3105 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3106 struct tcp_sock *tp = tcp_sk(sk);
3108 /* Optimization, tack on the FIN if we have one skb in write queue and
3109 * this skb was not yet sent, or we are under memory pressure.
3110 * Note: in the latter case, FIN packet will be sent after a timeout,
3111 * as TCP stack thinks it has already been transmitted.
3113 if (!tskb && tcp_under_memory_pressure(sk))
3114 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3118 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3119 TCP_SKB_CB(tskb)->end_seq++;
3121 if (tcp_write_queue_empty(sk)) {
3122 /* This means tskb was already sent.
3123 * Pretend we included the FIN on previous transmit.
3124 * We need to set tp->snd_nxt to the value it would have
3125 * if FIN had been sent. This is because retransmit path
3126 * does not change tp->snd_nxt.
3132 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3133 if (unlikely(!skb)) {
3138 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3139 skb_reserve(skb, MAX_TCP_HEADER);
3140 sk_forced_mem_schedule(sk, skb->truesize);
3141 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3142 tcp_init_nondata_skb(skb, tp->write_seq,
3143 TCPHDR_ACK | TCPHDR_FIN);
3144 tcp_queue_skb(sk, skb);
3146 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3149 /* We get here when a process closes a file descriptor (either due to
3150 * an explicit close() or as a byproduct of exit()'ing) and there
3151 * was unread data in the receive queue. This behavior is recommended
3152 * by RFC 2525, section 2.17. -DaveM
3154 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3156 struct sk_buff *skb;
3158 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3160 /* NOTE: No TCP options attached and we never retransmit this. */
3161 skb = alloc_skb(MAX_TCP_HEADER, priority);
3163 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3167 /* Reserve space for headers and prepare control bits. */
3168 skb_reserve(skb, MAX_TCP_HEADER);
3169 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3170 TCPHDR_ACK | TCPHDR_RST);
3171 tcp_mstamp_refresh(tcp_sk(sk));
3173 if (tcp_transmit_skb(sk, skb, 0, priority))
3174 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3176 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3177 * skb here is different to the troublesome skb, so use NULL
3179 trace_tcp_send_reset(sk, NULL);
3182 /* Send a crossed SYN-ACK during socket establishment.
3183 * WARNING: This routine must only be called when we have already sent
3184 * a SYN packet that crossed the incoming SYN that caused this routine
3185 * to get called. If this assumption fails then the initial rcv_wnd
3186 * and rcv_wscale values will not be correct.
3188 int tcp_send_synack(struct sock *sk)
3190 struct sk_buff *skb;
3192 skb = tcp_rtx_queue_head(sk);
3193 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3194 pr_err("%s: wrong queue state\n", __func__);
3197 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3198 if (skb_cloned(skb)) {
3199 struct sk_buff *nskb;
3201 tcp_skb_tsorted_save(skb) {
3202 nskb = skb_copy(skb, GFP_ATOMIC);
3203 } tcp_skb_tsorted_restore(skb);
3206 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3207 tcp_highest_sack_replace(sk, skb, nskb);
3208 tcp_rtx_queue_unlink_and_free(skb, sk);
3209 __skb_header_release(nskb);
3210 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3211 sk->sk_wmem_queued += nskb->truesize;
3212 sk_mem_charge(sk, nskb->truesize);
3216 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3217 tcp_ecn_send_synack(sk, skb);
3219 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3223 * tcp_make_synack - Prepare a SYN-ACK.
3224 * sk: listener socket
3225 * dst: dst entry attached to the SYNACK
3226 * req: request_sock pointer
3228 * Allocate one skb and build a SYNACK packet.
3229 * @dst is consumed : Caller should not use it again.
3231 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3232 struct request_sock *req,
3233 struct tcp_fastopen_cookie *foc,
3234 enum tcp_synack_type synack_type)
3236 struct inet_request_sock *ireq = inet_rsk(req);
3237 const struct tcp_sock *tp = tcp_sk(sk);
3238 struct tcp_md5sig_key *md5 = NULL;
3239 struct tcp_out_options opts;
3240 struct sk_buff *skb;
3241 int tcp_header_size;
3245 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3246 if (unlikely(!skb)) {
3250 /* Reserve space for headers. */
3251 skb_reserve(skb, MAX_TCP_HEADER);
3253 switch (synack_type) {
3254 case TCP_SYNACK_NORMAL:
3255 skb_set_owner_w(skb, req_to_sk(req));
3257 case TCP_SYNACK_COOKIE:
3258 /* Under synflood, we do not attach skb to a socket,
3259 * to avoid false sharing.
3262 case TCP_SYNACK_FASTOPEN:
3263 /* sk is a const pointer, because we want to express multiple
3264 * cpu might call us concurrently.
3265 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3267 skb_set_owner_w(skb, (struct sock *)sk);
3270 skb_dst_set(skb, dst);
3272 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3274 memset(&opts, 0, sizeof(opts));
3275 #ifdef CONFIG_SYN_COOKIES
3276 if (unlikely(req->cookie_ts))
3277 skb->skb_mstamp = cookie_init_timestamp(req);
3280 skb->skb_mstamp = tcp_clock_us();
3282 #ifdef CONFIG_TCP_MD5SIG
3284 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3286 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3287 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3288 foc, synack_type) + sizeof(*th);
3290 skb_push(skb, tcp_header_size);
3291 skb_reset_transport_header(skb);
3293 th = (struct tcphdr *)skb->data;
3294 memset(th, 0, sizeof(struct tcphdr));
3297 tcp_ecn_make_synack(req, th);
3298 th->source = htons(ireq->ir_num);
3299 th->dest = ireq->ir_rmt_port;
3300 skb->mark = ireq->ir_mark;
3301 skb->ip_summed = CHECKSUM_PARTIAL;
3302 th->seq = htonl(tcp_rsk(req)->snt_isn);
3303 /* XXX data is queued and acked as is. No buffer/window check */
3304 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3306 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3307 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3308 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3309 th->doff = (tcp_header_size >> 2);
3310 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3312 #ifdef CONFIG_TCP_MD5SIG
3313 /* Okay, we have all we need - do the md5 hash if needed */
3315 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3316 md5, req_to_sk(req), skb);
3320 /* Do not fool tcpdump (if any), clean our debris */
3324 EXPORT_SYMBOL(tcp_make_synack);
3326 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3328 struct inet_connection_sock *icsk = inet_csk(sk);
3329 const struct tcp_congestion_ops *ca;
3330 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3332 if (ca_key == TCP_CA_UNSPEC)
3336 ca = tcp_ca_find_key(ca_key);
3337 if (likely(ca && try_module_get(ca->owner))) {
3338 module_put(icsk->icsk_ca_ops->owner);
3339 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3340 icsk->icsk_ca_ops = ca;
3345 /* Do all connect socket setups that can be done AF independent. */
3346 static void tcp_connect_init(struct sock *sk)
3348 const struct dst_entry *dst = __sk_dst_get(sk);
3349 struct tcp_sock *tp = tcp_sk(sk);
3353 /* We'll fix this up when we get a response from the other end.
3354 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3356 tp->tcp_header_len = sizeof(struct tcphdr);
3357 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3358 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3360 #ifdef CONFIG_TCP_MD5SIG
3361 if (tp->af_specific->md5_lookup(sk, sk))
3362 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3365 /* If user gave his TCP_MAXSEG, record it to clamp */
3366 if (tp->rx_opt.user_mss)
3367 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3370 tcp_sync_mss(sk, dst_mtu(dst));
3372 tcp_ca_dst_init(sk, dst);
3374 if (!tp->window_clamp)
3375 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3376 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3378 tcp_initialize_rcv_mss(sk);
3380 /* limit the window selection if the user enforce a smaller rx buffer */
3381 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3382 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3383 tp->window_clamp = tcp_full_space(sk);
3385 rcv_wnd = tcp_rwnd_init_bpf(sk);
3387 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3389 tcp_select_initial_window(sk, tcp_full_space(sk),
3390 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3393 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3397 tp->rx_opt.rcv_wscale = rcv_wscale;
3398 tp->rcv_ssthresh = tp->rcv_wnd;
3401 sock_reset_flag(sk, SOCK_DONE);
3404 tcp_write_queue_purge(sk);
3405 tp->snd_una = tp->write_seq;
3406 tp->snd_sml = tp->write_seq;
3407 tp->snd_up = tp->write_seq;
3408 tp->snd_nxt = tp->write_seq;
3410 if (likely(!tp->repair))
3413 tp->rcv_tstamp = tcp_jiffies32;
3414 tp->rcv_wup = tp->rcv_nxt;
3415 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3417 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3418 inet_csk(sk)->icsk_retransmits = 0;
3419 tcp_clear_retrans(tp);
3422 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3424 struct tcp_sock *tp = tcp_sk(sk);
3425 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3427 tcb->end_seq += skb->len;
3428 __skb_header_release(skb);
3429 sk->sk_wmem_queued += skb->truesize;
3430 sk_mem_charge(sk, skb->truesize);
3431 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3432 tp->packets_out += tcp_skb_pcount(skb);
3435 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3436 * queue a data-only packet after the regular SYN, such that regular SYNs
3437 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3438 * only the SYN sequence, the data are retransmitted in the first ACK.
3439 * If cookie is not cached or other error occurs, falls back to send a
3440 * regular SYN with Fast Open cookie request option.
3442 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3444 struct inet_connection_sock *icsk = inet_csk(sk);
3445 struct tcp_sock *tp = tcp_sk(sk);
3446 struct tcp_fastopen_request *fo = tp->fastopen_req;
3448 struct sk_buff *syn_data;
3450 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3451 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3454 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3455 * user-MSS. Reserve maximum option space for middleboxes that add
3456 * private TCP options. The cost is reduced data space in SYN :(
3458 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3459 /* Sync mss_cache after updating the mss_clamp */
3460 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3462 space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3463 MAX_TCP_OPTION_SPACE;
3465 space = min_t(size_t, space, fo->size);
3467 /* limit to order-0 allocations */
3468 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3470 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3473 syn_data->ip_summed = CHECKSUM_PARTIAL;
3474 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3476 int copied = copy_from_iter(skb_put(syn_data, space), space,
3477 &fo->data->msg_iter);
3478 if (unlikely(!copied)) {
3479 tcp_skb_tsorted_anchor_cleanup(syn_data);
3480 kfree_skb(syn_data);
3483 if (copied != space) {
3484 skb_trim(syn_data, copied);
3488 /* No more data pending in inet_wait_for_connect() */
3489 if (space == fo->size)
3493 tcp_connect_queue_skb(sk, syn_data);
3495 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3497 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3499 syn->skb_mstamp = syn_data->skb_mstamp;
3501 /* Now full SYN+DATA was cloned and sent (or not),
3502 * remove the SYN from the original skb (syn_data)
3503 * we keep in write queue in case of a retransmit, as we
3504 * also have the SYN packet (with no data) in the same queue.
3506 TCP_SKB_CB(syn_data)->seq++;
3507 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3509 tp->syn_data = (fo->copied > 0);
3510 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3511 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3515 /* data was not sent, put it in write_queue */
3516 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3517 tp->packets_out -= tcp_skb_pcount(syn_data);
3520 /* Send a regular SYN with Fast Open cookie request option */
3521 if (fo->cookie.len > 0)
3523 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3525 tp->syn_fastopen = 0;
3527 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3531 /* Build a SYN and send it off. */
3532 int tcp_connect(struct sock *sk)
3534 struct tcp_sock *tp = tcp_sk(sk);
3535 struct sk_buff *buff;
3538 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3540 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3541 return -EHOSTUNREACH; /* Routing failure or similar. */
3543 tcp_connect_init(sk);
3545 if (unlikely(tp->repair)) {
3546 tcp_finish_connect(sk, NULL);
3550 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3551 if (unlikely(!buff))
3554 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3555 tcp_mstamp_refresh(tp);
3556 tp->retrans_stamp = tcp_time_stamp(tp);
3557 tcp_connect_queue_skb(sk, buff);
3558 tcp_ecn_send_syn(sk, buff);
3559 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3561 /* Send off SYN; include data in Fast Open. */
3562 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3563 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3564 if (err == -ECONNREFUSED)
3567 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3568 * in order to make this packet get counted in tcpOutSegs.
3570 tp->snd_nxt = tp->write_seq;
3571 tp->pushed_seq = tp->write_seq;
3572 buff = tcp_send_head(sk);
3573 if (unlikely(buff)) {
3574 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3575 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3577 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3579 /* Timer for repeating the SYN until an answer. */
3580 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3581 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3584 EXPORT_SYMBOL(tcp_connect);
3586 /* Send out a delayed ack, the caller does the policy checking
3587 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3590 void tcp_send_delayed_ack(struct sock *sk)
3592 struct inet_connection_sock *icsk = inet_csk(sk);
3593 int ato = icsk->icsk_ack.ato;
3594 unsigned long timeout;
3596 if (ato > TCP_DELACK_MIN) {
3597 const struct tcp_sock *tp = tcp_sk(sk);
3598 int max_ato = HZ / 2;
3600 if (icsk->icsk_ack.pingpong ||
3601 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3602 max_ato = TCP_DELACK_MAX;
3604 /* Slow path, intersegment interval is "high". */
3606 /* If some rtt estimate is known, use it to bound delayed ack.
3607 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3611 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3618 ato = min(ato, max_ato);
3621 /* Stay within the limit we were given */
3622 timeout = jiffies + ato;
3624 /* Use new timeout only if there wasn't a older one earlier. */
3625 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3626 /* If delack timer was blocked or is about to expire,
3629 if (icsk->icsk_ack.blocked ||
3630 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3635 if (!time_before(timeout, icsk->icsk_ack.timeout))
3636 timeout = icsk->icsk_ack.timeout;
3638 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3639 icsk->icsk_ack.timeout = timeout;
3640 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3643 /* This routine sends an ack and also updates the window. */
3644 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3646 struct sk_buff *buff;
3648 /* If we have been reset, we may not send again. */
3649 if (sk->sk_state == TCP_CLOSE)
3652 /* We are not putting this on the write queue, so
3653 * tcp_transmit_skb() will set the ownership to this
3656 buff = alloc_skb(MAX_TCP_HEADER,
3657 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3658 if (unlikely(!buff)) {
3659 inet_csk_schedule_ack(sk);
3660 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3661 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3662 TCP_DELACK_MAX, TCP_RTO_MAX);
3666 /* Reserve space for headers and prepare control bits. */
3667 skb_reserve(buff, MAX_TCP_HEADER);
3668 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3670 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3672 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3674 skb_set_tcp_pure_ack(buff);
3676 /* Send it off, this clears delayed acks for us. */
3677 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3679 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3681 void tcp_send_ack(struct sock *sk)
3683 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3686 /* This routine sends a packet with an out of date sequence
3687 * number. It assumes the other end will try to ack it.
3689 * Question: what should we make while urgent mode?
3690 * 4.4BSD forces sending single byte of data. We cannot send
3691 * out of window data, because we have SND.NXT==SND.MAX...
3693 * Current solution: to send TWO zero-length segments in urgent mode:
3694 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3695 * out-of-date with SND.UNA-1 to probe window.
3697 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3699 struct tcp_sock *tp = tcp_sk(sk);
3700 struct sk_buff *skb;
3702 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3703 skb = alloc_skb(MAX_TCP_HEADER,
3704 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3708 /* Reserve space for headers and set control bits. */
3709 skb_reserve(skb, MAX_TCP_HEADER);
3710 /* Use a previous sequence. This should cause the other
3711 * end to send an ack. Don't queue or clone SKB, just
3714 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3715 NET_INC_STATS(sock_net(sk), mib);
3716 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3719 /* Called from setsockopt( ... TCP_REPAIR ) */
3720 void tcp_send_window_probe(struct sock *sk)
3722 if (sk->sk_state == TCP_ESTABLISHED) {
3723 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3724 tcp_mstamp_refresh(tcp_sk(sk));
3725 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3729 /* Initiate keepalive or window probe from timer. */
3730 int tcp_write_wakeup(struct sock *sk, int mib)
3732 struct tcp_sock *tp = tcp_sk(sk);
3733 struct sk_buff *skb;
3735 if (sk->sk_state == TCP_CLOSE)
3738 skb = tcp_send_head(sk);
3739 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3741 unsigned int mss = tcp_current_mss(sk);
3742 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3744 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3745 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3747 /* We are probing the opening of a window
3748 * but the window size is != 0
3749 * must have been a result SWS avoidance ( sender )
3751 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3753 seg_size = min(seg_size, mss);
3754 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3755 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3756 skb, seg_size, mss, GFP_ATOMIC))
3758 } else if (!tcp_skb_pcount(skb))
3759 tcp_set_skb_tso_segs(skb, mss);
3761 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3762 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3764 tcp_event_new_data_sent(sk, skb);
3767 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3768 tcp_xmit_probe_skb(sk, 1, mib);
3769 return tcp_xmit_probe_skb(sk, 0, mib);
3773 /* A window probe timeout has occurred. If window is not closed send
3774 * a partial packet else a zero probe.
3776 void tcp_send_probe0(struct sock *sk)
3778 struct inet_connection_sock *icsk = inet_csk(sk);
3779 struct tcp_sock *tp = tcp_sk(sk);
3780 struct net *net = sock_net(sk);
3781 unsigned long probe_max;
3784 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3786 if (tp->packets_out || tcp_write_queue_empty(sk)) {
3787 /* Cancel probe timer, if it is not required. */
3788 icsk->icsk_probes_out = 0;
3789 icsk->icsk_backoff = 0;
3794 if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
3795 icsk->icsk_backoff++;
3796 icsk->icsk_probes_out++;
3797 probe_max = TCP_RTO_MAX;
3799 /* If packet was not sent due to local congestion,
3800 * do not backoff and do not remember icsk_probes_out.
3801 * Let local senders to fight for local resources.
3803 * Use accumulated backoff yet.
3805 if (!icsk->icsk_probes_out)
3806 icsk->icsk_probes_out = 1;
3807 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3809 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3810 tcp_probe0_when(sk, probe_max),
3814 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3816 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3820 tcp_rsk(req)->txhash = net_tx_rndhash();
3821 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3823 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3824 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3825 if (unlikely(tcp_passive_fastopen(sk)))
3826 tcp_sk(sk)->total_retrans++;
3827 trace_tcp_retransmit_synack(sk, req);
3831 EXPORT_SYMBOL(tcp_rtx_synack);