1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
9 * IPv4 specific functions
13 * linux/ipv4/tcp_input.c
14 * linux/ipv4/tcp_output.c
16 * See tcp.c for author information
21 * David S. Miller : New socket lookup architecture.
22 * This code is dedicated to John Dyson.
23 * David S. Miller : Change semantics of established hash,
24 * half is devoted to TIME_WAIT sockets
25 * and the rest go in the other half.
26 * Andi Kleen : Add support for syncookies and fixed
27 * some bugs: ip options weren't passed to
28 * the TCP layer, missed a check for an
30 * Andi Kleen : Implemented fast path mtu discovery.
31 * Fixed many serious bugs in the
32 * request_sock handling and moved
33 * most of it into the af independent code.
34 * Added tail drop and some other bugfixes.
35 * Added new listen semantics.
36 * Mike McLagan : Routing by source
37 * Juan Jose Ciarlante: ip_dynaddr bits
38 * Andi Kleen: various fixes.
39 * Vitaly E. Lavrov : Transparent proxy revived after year
41 * Andi Kleen : Fix new listen.
42 * Andi Kleen : Fix accept error reporting.
43 * YOSHIFUJI Hideaki @USAGI and: Support IPV6_V6ONLY socket option, which
44 * Alexey Kuznetsov allow both IPv4 and IPv6 sockets to bind
45 * a single port at the same time.
48 #define pr_fmt(fmt) "TCP: " fmt
50 #include <linux/bottom_half.h>
51 #include <linux/types.h>
52 #include <linux/fcntl.h>
53 #include <linux/module.h>
54 #include <linux/random.h>
55 #include <linux/cache.h>
56 #include <linux/jhash.h>
57 #include <linux/init.h>
58 #include <linux/times.h>
59 #include <linux/slab.h>
60 #include <linux/sched.h>
62 #include <net/net_namespace.h>
64 #include <net/inet_hashtables.h>
66 #include <net/transp_v6.h>
68 #include <net/inet_common.h>
69 #include <net/timewait_sock.h>
71 #include <net/secure_seq.h>
72 #include <net/busy_poll.h>
74 #include <linux/inet.h>
75 #include <linux/ipv6.h>
76 #include <linux/stddef.h>
77 #include <linux/proc_fs.h>
78 #include <linux/seq_file.h>
79 #include <linux/inetdevice.h>
80 #include <linux/btf_ids.h>
82 #include <crypto/hash.h>
83 #include <linux/scatterlist.h>
85 #include <trace/events/tcp.h>
87 #ifdef CONFIG_TCP_MD5SIG
88 static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
89 __be32 daddr, __be32 saddr, const struct tcphdr *th);
92 struct inet_hashinfo tcp_hashinfo;
93 EXPORT_SYMBOL(tcp_hashinfo);
95 static DEFINE_PER_CPU(struct sock *, ipv4_tcp_sk);
97 static u32 tcp_v4_init_seq(const struct sk_buff *skb)
99 return secure_tcp_seq(ip_hdr(skb)->daddr,
102 tcp_hdr(skb)->source);
105 static u32 tcp_v4_init_ts_off(const struct net *net, const struct sk_buff *skb)
107 return secure_tcp_ts_off(net, ip_hdr(skb)->daddr, ip_hdr(skb)->saddr);
110 int tcp_twsk_unique(struct sock *sk, struct sock *sktw, void *twp)
112 int reuse = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tw_reuse);
113 const struct inet_timewait_sock *tw = inet_twsk(sktw);
114 const struct tcp_timewait_sock *tcptw = tcp_twsk(sktw);
115 struct tcp_sock *tp = tcp_sk(sk);
118 /* Still does not detect *everything* that goes through
119 * lo, since we require a loopback src or dst address
120 * or direct binding to 'lo' interface.
122 bool loopback = false;
123 if (tw->tw_bound_dev_if == LOOPBACK_IFINDEX)
125 #if IS_ENABLED(CONFIG_IPV6)
126 if (tw->tw_family == AF_INET6) {
127 if (ipv6_addr_loopback(&tw->tw_v6_daddr) ||
128 ipv6_addr_v4mapped_loopback(&tw->tw_v6_daddr) ||
129 ipv6_addr_loopback(&tw->tw_v6_rcv_saddr) ||
130 ipv6_addr_v4mapped_loopback(&tw->tw_v6_rcv_saddr))
135 if (ipv4_is_loopback(tw->tw_daddr) ||
136 ipv4_is_loopback(tw->tw_rcv_saddr))
143 /* With PAWS, it is safe from the viewpoint
144 of data integrity. Even without PAWS it is safe provided sequence
145 spaces do not overlap i.e. at data rates <= 80Mbit/sec.
147 Actually, the idea is close to VJ's one, only timestamp cache is
148 held not per host, but per port pair and TW bucket is used as state
151 If TW bucket has been already destroyed we fall back to VJ's scheme
152 and use initial timestamp retrieved from peer table.
154 if (tcptw->tw_ts_recent_stamp &&
155 (!twp || (reuse && time_after32(ktime_get_seconds(),
156 tcptw->tw_ts_recent_stamp)))) {
157 /* In case of repair and re-using TIME-WAIT sockets we still
158 * want to be sure that it is safe as above but honor the
159 * sequence numbers and time stamps set as part of the repair
162 * Without this check re-using a TIME-WAIT socket with TCP
163 * repair would accumulate a -1 on the repair assigned
164 * sequence number. The first time it is reused the sequence
165 * is -1, the second time -2, etc. This fixes that issue
166 * without appearing to create any others.
168 if (likely(!tp->repair)) {
169 u32 seq = tcptw->tw_snd_nxt + 65535 + 2;
173 WRITE_ONCE(tp->write_seq, seq);
174 tp->rx_opt.ts_recent = tcptw->tw_ts_recent;
175 tp->rx_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
183 EXPORT_SYMBOL_GPL(tcp_twsk_unique);
185 static int tcp_v4_pre_connect(struct sock *sk, struct sockaddr *uaddr,
188 /* This check is replicated from tcp_v4_connect() and intended to
189 * prevent BPF program called below from accessing bytes that are out
190 * of the bound specified by user in addr_len.
192 if (addr_len < sizeof(struct sockaddr_in))
195 sock_owned_by_me(sk);
197 return BPF_CGROUP_RUN_PROG_INET4_CONNECT(sk, uaddr, &addr_len);
200 /* This will initiate an outgoing connection. */
201 int tcp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
203 struct sockaddr_in *usin = (struct sockaddr_in *)uaddr;
204 struct inet_timewait_death_row *tcp_death_row;
205 struct inet_sock *inet = inet_sk(sk);
206 struct tcp_sock *tp = tcp_sk(sk);
207 struct ip_options_rcu *inet_opt;
208 struct net *net = sock_net(sk);
209 __be16 orig_sport, orig_dport;
210 __be32 daddr, nexthop;
215 if (addr_len < sizeof(struct sockaddr_in))
218 if (usin->sin_family != AF_INET)
219 return -EAFNOSUPPORT;
221 nexthop = daddr = usin->sin_addr.s_addr;
222 inet_opt = rcu_dereference_protected(inet->inet_opt,
223 lockdep_sock_is_held(sk));
224 if (inet_opt && inet_opt->opt.srr) {
227 nexthop = inet_opt->opt.faddr;
230 orig_sport = inet->inet_sport;
231 orig_dport = usin->sin_port;
232 fl4 = &inet->cork.fl.u.ip4;
233 rt = ip_route_connect(fl4, nexthop, inet->inet_saddr,
234 sk->sk_bound_dev_if, IPPROTO_TCP, orig_sport,
238 if (err == -ENETUNREACH)
239 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
243 if (rt->rt_flags & (RTCF_MULTICAST | RTCF_BROADCAST)) {
248 if (!inet_opt || !inet_opt->opt.srr)
251 tcp_death_row = &sock_net(sk)->ipv4.tcp_death_row;
253 if (!inet->inet_saddr) {
254 err = inet_bhash2_update_saddr(sk, &fl4->saddr, AF_INET);
260 sk_rcv_saddr_set(sk, inet->inet_saddr);
263 if (tp->rx_opt.ts_recent_stamp && inet->inet_daddr != daddr) {
264 /* Reset inherited state */
265 tp->rx_opt.ts_recent = 0;
266 tp->rx_opt.ts_recent_stamp = 0;
267 if (likely(!tp->repair))
268 WRITE_ONCE(tp->write_seq, 0);
271 inet->inet_dport = usin->sin_port;
272 sk_daddr_set(sk, daddr);
274 inet_csk(sk)->icsk_ext_hdr_len = 0;
276 inet_csk(sk)->icsk_ext_hdr_len = inet_opt->opt.optlen;
278 tp->rx_opt.mss_clamp = TCP_MSS_DEFAULT;
280 /* Socket identity is still unknown (sport may be zero).
281 * However we set state to SYN-SENT and not releasing socket
282 * lock select source port, enter ourselves into the hash tables and
283 * complete initialization after this.
285 tcp_set_state(sk, TCP_SYN_SENT);
286 err = inet_hash_connect(tcp_death_row, sk);
292 rt = ip_route_newports(fl4, rt, orig_sport, orig_dport,
293 inet->inet_sport, inet->inet_dport, sk);
299 tp->tcp_usec_ts = dst_tcp_usec_ts(&rt->dst);
300 /* OK, now commit destination to socket. */
301 sk->sk_gso_type = SKB_GSO_TCPV4;
302 sk_setup_caps(sk, &rt->dst);
305 if (likely(!tp->repair)) {
307 WRITE_ONCE(tp->write_seq,
308 secure_tcp_seq(inet->inet_saddr,
312 WRITE_ONCE(tp->tsoffset,
313 secure_tcp_ts_off(net, inet->inet_saddr,
317 atomic_set(&inet->inet_id, get_random_u16());
319 if (tcp_fastopen_defer_connect(sk, &err))
324 err = tcp_connect(sk);
333 * This unhashes the socket and releases the local port,
336 tcp_set_state(sk, TCP_CLOSE);
337 inet_bhash2_reset_saddr(sk);
339 sk->sk_route_caps = 0;
340 inet->inet_dport = 0;
343 EXPORT_SYMBOL(tcp_v4_connect);
346 * This routine reacts to ICMP_FRAG_NEEDED mtu indications as defined in RFC1191.
347 * It can be called through tcp_release_cb() if socket was owned by user
348 * at the time tcp_v4_err() was called to handle ICMP message.
350 void tcp_v4_mtu_reduced(struct sock *sk)
352 struct inet_sock *inet = inet_sk(sk);
353 struct dst_entry *dst;
356 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE))
358 mtu = READ_ONCE(tcp_sk(sk)->mtu_info);
359 dst = inet_csk_update_pmtu(sk, mtu);
363 /* Something is about to be wrong... Remember soft error
364 * for the case, if this connection will not able to recover.
366 if (mtu < dst_mtu(dst) && ip_dont_fragment(sk, dst))
367 WRITE_ONCE(sk->sk_err_soft, EMSGSIZE);
371 if (inet->pmtudisc != IP_PMTUDISC_DONT &&
372 ip_sk_accept_pmtu(sk) &&
373 inet_csk(sk)->icsk_pmtu_cookie > mtu) {
374 tcp_sync_mss(sk, mtu);
376 /* Resend the TCP packet because it's
377 * clear that the old packet has been
378 * dropped. This is the new "fast" path mtu
381 tcp_simple_retransmit(sk);
382 } /* else let the usual retransmit timer handle it */
384 EXPORT_SYMBOL(tcp_v4_mtu_reduced);
386 static void do_redirect(struct sk_buff *skb, struct sock *sk)
388 struct dst_entry *dst = __sk_dst_check(sk, 0);
391 dst->ops->redirect(dst, sk, skb);
395 /* handle ICMP messages on TCP_NEW_SYN_RECV request sockets */
396 void tcp_req_err(struct sock *sk, u32 seq, bool abort)
398 struct request_sock *req = inet_reqsk(sk);
399 struct net *net = sock_net(sk);
401 /* ICMPs are not backlogged, hence we cannot get
402 * an established socket here.
404 if (seq != tcp_rsk(req)->snt_isn) {
405 __NET_INC_STATS(net, LINUX_MIB_OUTOFWINDOWICMPS);
408 * Still in SYN_RECV, just remove it silently.
409 * There is no good way to pass the error to the newly
410 * created socket, and POSIX does not want network
411 * errors returned from accept().
413 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
414 tcp_listendrop(req->rsk_listener);
418 EXPORT_SYMBOL(tcp_req_err);
420 /* TCP-LD (RFC 6069) logic */
421 void tcp_ld_RTO_revert(struct sock *sk, u32 seq)
423 struct inet_connection_sock *icsk = inet_csk(sk);
424 struct tcp_sock *tp = tcp_sk(sk);
429 if (sock_owned_by_user(sk))
432 if (seq != tp->snd_una || !icsk->icsk_retransmits ||
436 skb = tcp_rtx_queue_head(sk);
437 if (WARN_ON_ONCE(!skb))
440 icsk->icsk_backoff--;
441 icsk->icsk_rto = tp->srtt_us ? __tcp_set_rto(tp) : TCP_TIMEOUT_INIT;
442 icsk->icsk_rto = inet_csk_rto_backoff(icsk, TCP_RTO_MAX);
444 tcp_mstamp_refresh(tp);
445 delta_us = (u32)(tp->tcp_mstamp - tcp_skb_timestamp_us(skb));
446 remaining = icsk->icsk_rto - usecs_to_jiffies(delta_us);
449 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
450 remaining, TCP_RTO_MAX);
452 /* RTO revert clocked out retransmission.
453 * Will retransmit now.
455 tcp_retransmit_timer(sk);
458 EXPORT_SYMBOL(tcp_ld_RTO_revert);
461 * This routine is called by the ICMP module when it gets some
462 * sort of error condition. If err < 0 then the socket should
463 * be closed and the error returned to the user. If err > 0
464 * it's just the icmp type << 8 | icmp code. After adjustment
465 * header points to the first 8 bytes of the tcp header. We need
466 * to find the appropriate port.
468 * The locking strategy used here is very "optimistic". When
469 * someone else accesses the socket the ICMP is just dropped
470 * and for some paths there is no check at all.
471 * A more general error queue to queue errors for later handling
472 * is probably better.
476 int tcp_v4_err(struct sk_buff *skb, u32 info)
478 const struct iphdr *iph = (const struct iphdr *)skb->data;
479 struct tcphdr *th = (struct tcphdr *)(skb->data + (iph->ihl << 2));
481 const int type = icmp_hdr(skb)->type;
482 const int code = icmp_hdr(skb)->code;
484 struct request_sock *fastopen;
487 struct net *net = dev_net(skb->dev);
489 sk = __inet_lookup_established(net, net->ipv4.tcp_death_row.hashinfo,
490 iph->daddr, th->dest, iph->saddr,
491 ntohs(th->source), inet_iif(skb), 0);
493 __ICMP_INC_STATS(net, ICMP_MIB_INERRORS);
496 if (sk->sk_state == TCP_TIME_WAIT) {
497 /* To increase the counter of ignored icmps for TCP-AO */
498 tcp_ao_ignore_icmp(sk, AF_INET, type, code);
499 inet_twsk_put(inet_twsk(sk));
502 seq = ntohl(th->seq);
503 if (sk->sk_state == TCP_NEW_SYN_RECV) {
504 tcp_req_err(sk, seq, type == ICMP_PARAMETERPROB ||
505 type == ICMP_TIME_EXCEEDED ||
506 (type == ICMP_DEST_UNREACH &&
507 (code == ICMP_NET_UNREACH ||
508 code == ICMP_HOST_UNREACH)));
512 if (tcp_ao_ignore_icmp(sk, AF_INET, type, code)) {
518 /* If too many ICMPs get dropped on busy
519 * servers this needs to be solved differently.
520 * We do take care of PMTU discovery (RFC1191) special case :
521 * we can receive locally generated ICMP messages while socket is held.
523 if (sock_owned_by_user(sk)) {
524 if (!(type == ICMP_DEST_UNREACH && code == ICMP_FRAG_NEEDED))
525 __NET_INC_STATS(net, LINUX_MIB_LOCKDROPPEDICMPS);
527 if (sk->sk_state == TCP_CLOSE)
530 if (static_branch_unlikely(&ip4_min_ttl)) {
531 /* min_ttl can be changed concurrently from do_ip_setsockopt() */
532 if (unlikely(iph->ttl < READ_ONCE(inet_sk(sk)->min_ttl))) {
533 __NET_INC_STATS(net, LINUX_MIB_TCPMINTTLDROP);
539 /* XXX (TFO) - tp->snd_una should be ISN (tcp_create_openreq_child() */
540 fastopen = rcu_dereference(tp->fastopen_rsk);
541 snd_una = fastopen ? tcp_rsk(fastopen)->snt_isn : tp->snd_una;
542 if (sk->sk_state != TCP_LISTEN &&
543 !between(seq, snd_una, tp->snd_nxt)) {
544 __NET_INC_STATS(net, LINUX_MIB_OUTOFWINDOWICMPS);
550 if (!sock_owned_by_user(sk))
551 do_redirect(skb, sk);
553 case ICMP_SOURCE_QUENCH:
554 /* Just silently ignore these. */
556 case ICMP_PARAMETERPROB:
559 case ICMP_DEST_UNREACH:
560 if (code > NR_ICMP_UNREACH)
563 if (code == ICMP_FRAG_NEEDED) { /* PMTU discovery (RFC1191) */
564 /* We are not interested in TCP_LISTEN and open_requests
565 * (SYN-ACKs send out by Linux are always <576bytes so
566 * they should go through unfragmented).
568 if (sk->sk_state == TCP_LISTEN)
571 WRITE_ONCE(tp->mtu_info, info);
572 if (!sock_owned_by_user(sk)) {
573 tcp_v4_mtu_reduced(sk);
575 if (!test_and_set_bit(TCP_MTU_REDUCED_DEFERRED, &sk->sk_tsq_flags))
581 err = icmp_err_convert[code].errno;
582 /* check if this ICMP message allows revert of backoff.
586 (code == ICMP_NET_UNREACH || code == ICMP_HOST_UNREACH))
587 tcp_ld_RTO_revert(sk, seq);
589 case ICMP_TIME_EXCEEDED:
596 switch (sk->sk_state) {
599 /* Only in fast or simultaneous open. If a fast open socket is
600 * already accepted it is treated as a connected one below.
602 if (fastopen && !fastopen->sk)
605 ip_icmp_error(sk, skb, err, th->dest, info, (u8 *)th);
607 if (!sock_owned_by_user(sk)) {
608 WRITE_ONCE(sk->sk_err, err);
614 WRITE_ONCE(sk->sk_err_soft, err);
619 /* If we've already connected we will keep trying
620 * until we time out, or the user gives up.
622 * rfc1122 4.2.3.9 allows to consider as hard errors
623 * only PROTO_UNREACH and PORT_UNREACH (well, FRAG_FAILED too,
624 * but it is obsoleted by pmtu discovery).
626 * Note, that in modern internet, where routing is unreliable
627 * and in each dark corner broken firewalls sit, sending random
628 * errors ordered by their masters even this two messages finally lose
629 * their original sense (even Linux sends invalid PORT_UNREACHs)
631 * Now we are in compliance with RFCs.
635 if (!sock_owned_by_user(sk) &&
636 inet_test_bit(RECVERR, sk)) {
637 WRITE_ONCE(sk->sk_err, err);
639 } else { /* Only an error on timeout */
640 WRITE_ONCE(sk->sk_err_soft, err);
649 void __tcp_v4_send_check(struct sk_buff *skb, __be32 saddr, __be32 daddr)
651 struct tcphdr *th = tcp_hdr(skb);
653 th->check = ~tcp_v4_check(skb->len, saddr, daddr, 0);
654 skb->csum_start = skb_transport_header(skb) - skb->head;
655 skb->csum_offset = offsetof(struct tcphdr, check);
658 /* This routine computes an IPv4 TCP checksum. */
659 void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)
661 const struct inet_sock *inet = inet_sk(sk);
663 __tcp_v4_send_check(skb, inet->inet_saddr, inet->inet_daddr);
665 EXPORT_SYMBOL(tcp_v4_send_check);
667 #define REPLY_OPTIONS_LEN (MAX_TCP_OPTION_SPACE / sizeof(__be32))
669 static bool tcp_v4_ao_sign_reset(const struct sock *sk, struct sk_buff *skb,
670 const struct tcp_ao_hdr *aoh,
671 struct ip_reply_arg *arg, struct tcphdr *reply,
672 __be32 reply_options[REPLY_OPTIONS_LEN])
675 int sdif = tcp_v4_sdif(skb);
676 int dif = inet_iif(skb);
677 int l3index = sdif ? dif : 0;
678 bool allocated_traffic_key;
679 struct tcp_ao_key *key;
686 if (tcp_ao_prepare_reset(sk, skb, aoh, l3index, ntohl(reply->seq),
687 &key, &traffic_key, &allocated_traffic_key,
691 reply_options[0] = htonl((TCPOPT_AO << 24) | (tcp_ao_len(key) << 16) |
692 (aoh->rnext_keyid << 8) | keyid);
693 arg->iov[0].iov_len += tcp_ao_len_aligned(key);
694 reply->doff = arg->iov[0].iov_len / 4;
696 if (tcp_ao_hash_hdr(AF_INET, (char *)&reply_options[1],
698 (union tcp_ao_addr *)&ip_hdr(skb)->saddr,
699 (union tcp_ao_addr *)&ip_hdr(skb)->daddr,
705 if (allocated_traffic_key)
714 * This routine will send an RST to the other tcp.
716 * Someone asks: why I NEVER use socket parameters (TOS, TTL etc.)
718 * Answer: if a packet caused RST, it is not for a socket
719 * existing in our system, if it is matched to a socket,
720 * it is just duplicate segment or bug in other side's TCP.
721 * So that we build reply only basing on parameters
722 * arrived with segment.
723 * Exception: precedence violation. We do not implement it in any case.
726 static void tcp_v4_send_reset(const struct sock *sk, struct sk_buff *skb)
728 const struct tcphdr *th = tcp_hdr(skb);
731 __be32 opt[REPLY_OPTIONS_LEN];
733 const __u8 *md5_hash_location = NULL;
734 const struct tcp_ao_hdr *aoh;
735 struct ip_reply_arg arg;
736 #ifdef CONFIG_TCP_MD5SIG
737 struct tcp_md5sig_key *key = NULL;
738 unsigned char newhash[16];
739 struct sock *sk1 = NULL;
742 u64 transmit_time = 0;
747 /* Never send a reset in response to a reset. */
751 /* If sk not NULL, it means we did a successful lookup and incoming
752 * route had to be correct. prequeue might have dropped our dst.
754 if (!sk && skb_rtable(skb)->rt_type != RTN_LOCAL)
757 /* Swap the send and the receive. */
758 memset(&rep, 0, sizeof(rep));
759 rep.th.dest = th->source;
760 rep.th.source = th->dest;
761 rep.th.doff = sizeof(struct tcphdr) / 4;
765 rep.th.seq = th->ack_seq;
768 rep.th.ack_seq = htonl(ntohl(th->seq) + th->syn + th->fin +
769 skb->len - (th->doff << 2));
772 memset(&arg, 0, sizeof(arg));
773 arg.iov[0].iov_base = (unsigned char *)&rep;
774 arg.iov[0].iov_len = sizeof(rep.th);
776 net = sk ? sock_net(sk) : dev_net(skb_dst(skb)->dev);
778 /* Invalid TCP option size or twice included auth */
779 if (tcp_parse_auth_options(tcp_hdr(skb), &md5_hash_location, &aoh))
782 if (aoh && tcp_v4_ao_sign_reset(sk, skb, aoh, &arg, &rep.th, rep.opt))
785 #ifdef CONFIG_TCP_MD5SIG
787 if (sk && sk_fullsock(sk)) {
788 const union tcp_md5_addr *addr;
791 /* sdif set, means packet ingressed via a device
792 * in an L3 domain and inet_iif is set to it.
794 l3index = tcp_v4_sdif(skb) ? inet_iif(skb) : 0;
795 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
796 key = tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
797 } else if (md5_hash_location) {
798 const union tcp_md5_addr *addr;
799 int sdif = tcp_v4_sdif(skb);
800 int dif = inet_iif(skb);
804 * active side is lost. Try to find listening socket through
805 * source port, and then find md5 key through listening socket.
806 * we are not loose security here:
807 * Incoming packet is checked with md5 hash with finding key,
808 * no RST generated if md5 hash doesn't match.
810 sk1 = __inet_lookup_listener(net, net->ipv4.tcp_death_row.hashinfo,
811 NULL, 0, ip_hdr(skb)->saddr,
812 th->source, ip_hdr(skb)->daddr,
813 ntohs(th->source), dif, sdif);
814 /* don't send rst if it can't find key */
818 /* sdif set, means packet ingressed via a device
819 * in an L3 domain and dif is set to it.
821 l3index = sdif ? dif : 0;
822 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
823 key = tcp_md5_do_lookup(sk1, l3index, addr, AF_INET);
828 genhash = tcp_v4_md5_hash_skb(newhash, key, NULL, skb);
829 if (genhash || memcmp(md5_hash_location, newhash, 16) != 0)
835 rep.opt[0] = htonl((TCPOPT_NOP << 24) |
837 (TCPOPT_MD5SIG << 8) |
839 /* Update length and the length the header thinks exists */
840 arg.iov[0].iov_len += TCPOLEN_MD5SIG_ALIGNED;
841 rep.th.doff = arg.iov[0].iov_len / 4;
843 tcp_v4_md5_hash_hdr((__u8 *) &rep.opt[1],
844 key, ip_hdr(skb)->saddr,
845 ip_hdr(skb)->daddr, &rep.th);
848 /* Can't co-exist with TCPMD5, hence check rep.opt[0] */
849 if (rep.opt[0] == 0) {
850 __be32 mrst = mptcp_reset_option(skb);
854 arg.iov[0].iov_len += sizeof(mrst);
855 rep.th.doff = arg.iov[0].iov_len / 4;
859 arg.csum = csum_tcpudp_nofold(ip_hdr(skb)->daddr,
860 ip_hdr(skb)->saddr, /* XXX */
861 arg.iov[0].iov_len, IPPROTO_TCP, 0);
862 arg.csumoffset = offsetof(struct tcphdr, check) / 2;
863 arg.flags = (sk && inet_sk_transparent(sk)) ? IP_REPLY_ARG_NOSRCCHECK : 0;
865 /* When socket is gone, all binding information is lost.
866 * routing might fail in this case. No choice here, if we choose to force
867 * input interface, we will misroute in case of asymmetric route.
870 arg.bound_dev_if = sk->sk_bound_dev_if;
872 trace_tcp_send_reset(sk, skb);
875 BUILD_BUG_ON(offsetof(struct sock, sk_bound_dev_if) !=
876 offsetof(struct inet_timewait_sock, tw_bound_dev_if));
878 arg.tos = ip_hdr(skb)->tos;
879 arg.uid = sock_net_uid(net, sk && sk_fullsock(sk) ? sk : NULL);
881 ctl_sk = this_cpu_read(ipv4_tcp_sk);
882 sock_net_set(ctl_sk, net);
884 ctl_sk->sk_mark = (sk->sk_state == TCP_TIME_WAIT) ?
885 inet_twsk(sk)->tw_mark : sk->sk_mark;
886 ctl_sk->sk_priority = (sk->sk_state == TCP_TIME_WAIT) ?
887 inet_twsk(sk)->tw_priority : READ_ONCE(sk->sk_priority);
888 transmit_time = tcp_transmit_time(sk);
889 xfrm_sk_clone_policy(ctl_sk, sk);
890 txhash = (sk->sk_state == TCP_TIME_WAIT) ?
891 inet_twsk(sk)->tw_txhash : sk->sk_txhash;
894 ctl_sk->sk_priority = 0;
896 ip_send_unicast_reply(ctl_sk,
897 skb, &TCP_SKB_CB(skb)->header.h4.opt,
898 ip_hdr(skb)->saddr, ip_hdr(skb)->daddr,
899 &arg, arg.iov[0].iov_len,
900 transmit_time, txhash);
902 xfrm_sk_free_policy(ctl_sk);
903 sock_net_set(ctl_sk, &init_net);
904 __TCP_INC_STATS(net, TCP_MIB_OUTSEGS);
905 __TCP_INC_STATS(net, TCP_MIB_OUTRSTS);
908 #ifdef CONFIG_TCP_MD5SIG
914 /* The code following below sending ACKs in SYN-RECV and TIME-WAIT states
915 outside socket context is ugly, certainly. What can I do?
918 static void tcp_v4_send_ack(const struct sock *sk,
919 struct sk_buff *skb, u32 seq, u32 ack,
920 u32 win, u32 tsval, u32 tsecr, int oif,
922 int reply_flags, u8 tos, u32 txhash)
924 const struct tcphdr *th = tcp_hdr(skb);
927 __be32 opt[(MAX_TCP_OPTION_SPACE >> 2)];
929 struct net *net = sock_net(sk);
930 struct ip_reply_arg arg;
934 memset(&rep.th, 0, sizeof(struct tcphdr));
935 memset(&arg, 0, sizeof(arg));
937 arg.iov[0].iov_base = (unsigned char *)&rep;
938 arg.iov[0].iov_len = sizeof(rep.th);
940 rep.opt[0] = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
941 (TCPOPT_TIMESTAMP << 8) |
943 rep.opt[1] = htonl(tsval);
944 rep.opt[2] = htonl(tsecr);
945 arg.iov[0].iov_len += TCPOLEN_TSTAMP_ALIGNED;
948 /* Swap the send and the receive. */
949 rep.th.dest = th->source;
950 rep.th.source = th->dest;
951 rep.th.doff = arg.iov[0].iov_len / 4;
952 rep.th.seq = htonl(seq);
953 rep.th.ack_seq = htonl(ack);
955 rep.th.window = htons(win);
957 #ifdef CONFIG_TCP_MD5SIG
958 if (tcp_key_is_md5(key)) {
959 int offset = (tsecr) ? 3 : 0;
961 rep.opt[offset++] = htonl((TCPOPT_NOP << 24) |
963 (TCPOPT_MD5SIG << 8) |
965 arg.iov[0].iov_len += TCPOLEN_MD5SIG_ALIGNED;
966 rep.th.doff = arg.iov[0].iov_len/4;
968 tcp_v4_md5_hash_hdr((__u8 *) &rep.opt[offset],
969 key->md5_key, ip_hdr(skb)->saddr,
970 ip_hdr(skb)->daddr, &rep.th);
974 if (tcp_key_is_ao(key)) {
975 int offset = (tsecr) ? 3 : 0;
977 rep.opt[offset++] = htonl((TCPOPT_AO << 24) |
978 (tcp_ao_len(key->ao_key) << 16) |
979 (key->ao_key->sndid << 8) |
981 arg.iov[0].iov_len += tcp_ao_len_aligned(key->ao_key);
982 rep.th.doff = arg.iov[0].iov_len / 4;
984 tcp_ao_hash_hdr(AF_INET, (char *)&rep.opt[offset],
985 key->ao_key, key->traffic_key,
986 (union tcp_ao_addr *)&ip_hdr(skb)->saddr,
987 (union tcp_ao_addr *)&ip_hdr(skb)->daddr,
991 arg.flags = reply_flags;
992 arg.csum = csum_tcpudp_nofold(ip_hdr(skb)->daddr,
993 ip_hdr(skb)->saddr, /* XXX */
994 arg.iov[0].iov_len, IPPROTO_TCP, 0);
995 arg.csumoffset = offsetof(struct tcphdr, check) / 2;
997 arg.bound_dev_if = oif;
999 arg.uid = sock_net_uid(net, sk_fullsock(sk) ? sk : NULL);
1001 ctl_sk = this_cpu_read(ipv4_tcp_sk);
1002 sock_net_set(ctl_sk, net);
1003 ctl_sk->sk_mark = (sk->sk_state == TCP_TIME_WAIT) ?
1004 inet_twsk(sk)->tw_mark : READ_ONCE(sk->sk_mark);
1005 ctl_sk->sk_priority = (sk->sk_state == TCP_TIME_WAIT) ?
1006 inet_twsk(sk)->tw_priority : READ_ONCE(sk->sk_priority);
1007 transmit_time = tcp_transmit_time(sk);
1008 ip_send_unicast_reply(ctl_sk,
1009 skb, &TCP_SKB_CB(skb)->header.h4.opt,
1010 ip_hdr(skb)->saddr, ip_hdr(skb)->daddr,
1011 &arg, arg.iov[0].iov_len,
1012 transmit_time, txhash);
1014 sock_net_set(ctl_sk, &init_net);
1015 __TCP_INC_STATS(net, TCP_MIB_OUTSEGS);
1019 static void tcp_v4_timewait_ack(struct sock *sk, struct sk_buff *skb)
1021 struct inet_timewait_sock *tw = inet_twsk(sk);
1022 struct tcp_timewait_sock *tcptw = tcp_twsk(sk);
1023 struct tcp_key key = {};
1024 #ifdef CONFIG_TCP_AO
1025 struct tcp_ao_info *ao_info;
1027 if (static_branch_unlikely(&tcp_ao_needed.key)) {
1028 /* FIXME: the segment to-be-acked is not verified yet */
1029 ao_info = rcu_dereference(tcptw->ao_info);
1031 const struct tcp_ao_hdr *aoh;
1033 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh)) {
1039 key.ao_key = tcp_ao_established_key(ao_info, aoh->rnext_keyid, -1);
1043 struct tcp_ao_key *rnext_key;
1045 key.traffic_key = snd_other_key(key.ao_key);
1046 key.sne = READ_ONCE(ao_info->snd_sne);
1047 rnext_key = READ_ONCE(ao_info->rnext_key);
1048 key.rcv_next = rnext_key->rcvid;
1049 key.type = TCP_KEY_AO;
1053 #ifdef CONFIG_TCP_MD5SIG
1054 } else if (static_branch_unlikely(&tcp_md5_needed.key)) {
1055 key.md5_key = tcp_twsk_md5_key(tcptw);
1057 key.type = TCP_KEY_MD5;
1061 tcp_v4_send_ack(sk, skb,
1062 tcptw->tw_snd_nxt, tcptw->tw_rcv_nxt,
1063 tcptw->tw_rcv_wnd >> tw->tw_rcv_wscale,
1064 tcp_tw_tsval(tcptw),
1065 tcptw->tw_ts_recent,
1066 tw->tw_bound_dev_if, &key,
1067 tw->tw_transparent ? IP_REPLY_ARG_NOSRCCHECK : 0,
1074 static void tcp_v4_reqsk_send_ack(const struct sock *sk, struct sk_buff *skb,
1075 struct request_sock *req)
1077 struct tcp_key key = {};
1079 /* sk->sk_state == TCP_LISTEN -> for regular TCP_SYN_RECV
1080 * sk->sk_state == TCP_SYN_RECV -> for Fast Open.
1082 u32 seq = (sk->sk_state == TCP_LISTEN) ? tcp_rsk(req)->snt_isn + 1 :
1083 tcp_sk(sk)->snd_nxt;
1085 #ifdef CONFIG_TCP_AO
1086 if (static_branch_unlikely(&tcp_ao_needed.key) &&
1087 tcp_rsk_used_ao(req)) {
1088 const union tcp_md5_addr *addr;
1089 const struct tcp_ao_hdr *aoh;
1092 /* Invalid TCP option size or twice included auth */
1093 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh))
1098 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
1099 l3index = tcp_v4_sdif(skb) ? inet_iif(skb) : 0;
1100 key.ao_key = tcp_ao_do_lookup(sk, l3index, addr, AF_INET,
1101 aoh->rnext_keyid, -1);
1102 if (unlikely(!key.ao_key)) {
1103 /* Send ACK with any matching MKT for the peer */
1104 key.ao_key = tcp_ao_do_lookup(sk, l3index, addr, AF_INET, -1, -1);
1105 /* Matching key disappeared (user removed the key?)
1106 * let the handshake timeout.
1109 net_info_ratelimited("TCP-AO key for (%pI4, %d)->(%pI4, %d) suddenly disappeared, won't ACK new connection\n",
1111 ntohs(tcp_hdr(skb)->source),
1112 &ip_hdr(skb)->daddr,
1113 ntohs(tcp_hdr(skb)->dest));
1117 key.traffic_key = kmalloc(tcp_ao_digest_size(key.ao_key), GFP_ATOMIC);
1118 if (!key.traffic_key)
1121 key.type = TCP_KEY_AO;
1122 key.rcv_next = aoh->keyid;
1123 tcp_v4_ao_calc_key_rsk(key.ao_key, key.traffic_key, req);
1127 #ifdef CONFIG_TCP_MD5SIG
1128 } else if (static_branch_unlikely(&tcp_md5_needed.key)) {
1129 const union tcp_md5_addr *addr;
1132 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
1133 l3index = tcp_v4_sdif(skb) ? inet_iif(skb) : 0;
1134 key.md5_key = tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
1136 key.type = TCP_KEY_MD5;
1141 * The window field (SEG.WND) of every outgoing segment, with the
1142 * exception of <SYN> segments, MUST be right-shifted by
1143 * Rcv.Wind.Shift bits:
1145 tcp_v4_send_ack(sk, skb, seq,
1146 tcp_rsk(req)->rcv_nxt,
1147 req->rsk_rcv_wnd >> inet_rsk(req)->rcv_wscale,
1148 tcp_rsk_tsval(tcp_rsk(req)),
1149 READ_ONCE(req->ts_recent),
1151 inet_rsk(req)->no_srccheck ? IP_REPLY_ARG_NOSRCCHECK : 0,
1153 READ_ONCE(tcp_rsk(req)->txhash));
1154 if (tcp_key_is_ao(&key))
1155 kfree(key.traffic_key);
1159 * Send a SYN-ACK after having received a SYN.
1160 * This still operates on a request_sock only, not on a big
1163 static int tcp_v4_send_synack(const struct sock *sk, struct dst_entry *dst,
1165 struct request_sock *req,
1166 struct tcp_fastopen_cookie *foc,
1167 enum tcp_synack_type synack_type,
1168 struct sk_buff *syn_skb)
1170 const struct inet_request_sock *ireq = inet_rsk(req);
1173 struct sk_buff *skb;
1176 /* First, grab a route. */
1177 if (!dst && (dst = inet_csk_route_req(sk, &fl4, req)) == NULL)
1180 skb = tcp_make_synack(sk, dst, req, foc, synack_type, syn_skb);
1183 __tcp_v4_send_check(skb, ireq->ir_loc_addr, ireq->ir_rmt_addr);
1185 tos = READ_ONCE(inet_sk(sk)->tos);
1187 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reflect_tos))
1188 tos = (tcp_rsk(req)->syn_tos & ~INET_ECN_MASK) |
1189 (tos & INET_ECN_MASK);
1191 if (!INET_ECN_is_capable(tos) &&
1192 tcp_bpf_ca_needs_ecn((struct sock *)req))
1193 tos |= INET_ECN_ECT_0;
1196 err = ip_build_and_send_pkt(skb, sk, ireq->ir_loc_addr,
1198 rcu_dereference(ireq->ireq_opt),
1201 err = net_xmit_eval(err);
1208 * IPv4 request_sock destructor.
1210 static void tcp_v4_reqsk_destructor(struct request_sock *req)
1212 kfree(rcu_dereference_protected(inet_rsk(req)->ireq_opt, 1));
1215 #ifdef CONFIG_TCP_MD5SIG
1217 * RFC2385 MD5 checksumming requires a mapping of
1218 * IP address->MD5 Key.
1219 * We need to maintain these in the sk structure.
1222 DEFINE_STATIC_KEY_DEFERRED_FALSE(tcp_md5_needed, HZ);
1223 EXPORT_SYMBOL(tcp_md5_needed);
1225 static bool better_md5_match(struct tcp_md5sig_key *old, struct tcp_md5sig_key *new)
1230 /* l3index always overrides non-l3index */
1231 if (old->l3index && new->l3index == 0)
1233 if (old->l3index == 0 && new->l3index)
1236 return old->prefixlen < new->prefixlen;
1239 /* Find the Key structure for an address. */
1240 struct tcp_md5sig_key *__tcp_md5_do_lookup(const struct sock *sk, int l3index,
1241 const union tcp_md5_addr *addr,
1242 int family, bool any_l3index)
1244 const struct tcp_sock *tp = tcp_sk(sk);
1245 struct tcp_md5sig_key *key;
1246 const struct tcp_md5sig_info *md5sig;
1248 struct tcp_md5sig_key *best_match = NULL;
1251 /* caller either holds rcu_read_lock() or socket lock */
1252 md5sig = rcu_dereference_check(tp->md5sig_info,
1253 lockdep_sock_is_held(sk));
1257 hlist_for_each_entry_rcu(key, &md5sig->head, node,
1258 lockdep_sock_is_held(sk)) {
1259 if (key->family != family)
1261 if (!any_l3index && key->flags & TCP_MD5SIG_FLAG_IFINDEX &&
1262 key->l3index != l3index)
1264 if (family == AF_INET) {
1265 mask = inet_make_mask(key->prefixlen);
1266 match = (key->addr.a4.s_addr & mask) ==
1267 (addr->a4.s_addr & mask);
1268 #if IS_ENABLED(CONFIG_IPV6)
1269 } else if (family == AF_INET6) {
1270 match = ipv6_prefix_equal(&key->addr.a6, &addr->a6,
1277 if (match && better_md5_match(best_match, key))
1282 EXPORT_SYMBOL(__tcp_md5_do_lookup);
1284 static struct tcp_md5sig_key *tcp_md5_do_lookup_exact(const struct sock *sk,
1285 const union tcp_md5_addr *addr,
1286 int family, u8 prefixlen,
1287 int l3index, u8 flags)
1289 const struct tcp_sock *tp = tcp_sk(sk);
1290 struct tcp_md5sig_key *key;
1291 unsigned int size = sizeof(struct in_addr);
1292 const struct tcp_md5sig_info *md5sig;
1294 /* caller either holds rcu_read_lock() or socket lock */
1295 md5sig = rcu_dereference_check(tp->md5sig_info,
1296 lockdep_sock_is_held(sk));
1299 #if IS_ENABLED(CONFIG_IPV6)
1300 if (family == AF_INET6)
1301 size = sizeof(struct in6_addr);
1303 hlist_for_each_entry_rcu(key, &md5sig->head, node,
1304 lockdep_sock_is_held(sk)) {
1305 if (key->family != family)
1307 if ((key->flags & TCP_MD5SIG_FLAG_IFINDEX) != (flags & TCP_MD5SIG_FLAG_IFINDEX))
1309 if (key->l3index != l3index)
1311 if (!memcmp(&key->addr, addr, size) &&
1312 key->prefixlen == prefixlen)
1318 struct tcp_md5sig_key *tcp_v4_md5_lookup(const struct sock *sk,
1319 const struct sock *addr_sk)
1321 const union tcp_md5_addr *addr;
1324 l3index = l3mdev_master_ifindex_by_index(sock_net(sk),
1325 addr_sk->sk_bound_dev_if);
1326 addr = (const union tcp_md5_addr *)&addr_sk->sk_daddr;
1327 return tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
1329 EXPORT_SYMBOL(tcp_v4_md5_lookup);
1331 static int tcp_md5sig_info_add(struct sock *sk, gfp_t gfp)
1333 struct tcp_sock *tp = tcp_sk(sk);
1334 struct tcp_md5sig_info *md5sig;
1336 md5sig = kmalloc(sizeof(*md5sig), gfp);
1341 INIT_HLIST_HEAD(&md5sig->head);
1342 rcu_assign_pointer(tp->md5sig_info, md5sig);
1346 /* This can be called on a newly created socket, from other files */
1347 static int __tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
1348 int family, u8 prefixlen, int l3index, u8 flags,
1349 const u8 *newkey, u8 newkeylen, gfp_t gfp)
1351 /* Add Key to the list */
1352 struct tcp_md5sig_key *key;
1353 struct tcp_sock *tp = tcp_sk(sk);
1354 struct tcp_md5sig_info *md5sig;
1356 key = tcp_md5_do_lookup_exact(sk, addr, family, prefixlen, l3index, flags);
1358 /* Pre-existing entry - just update that one.
1359 * Note that the key might be used concurrently.
1360 * data_race() is telling kcsan that we do not care of
1361 * key mismatches, since changing MD5 key on live flows
1362 * can lead to packet drops.
1364 data_race(memcpy(key->key, newkey, newkeylen));
1366 /* Pairs with READ_ONCE() in tcp_md5_hash_key().
1367 * Also note that a reader could catch new key->keylen value
1368 * but old key->key[], this is the reason we use __GFP_ZERO
1369 * at sock_kmalloc() time below these lines.
1371 WRITE_ONCE(key->keylen, newkeylen);
1376 md5sig = rcu_dereference_protected(tp->md5sig_info,
1377 lockdep_sock_is_held(sk));
1379 key = sock_kmalloc(sk, sizeof(*key), gfp | __GFP_ZERO);
1383 memcpy(key->key, newkey, newkeylen);
1384 key->keylen = newkeylen;
1385 key->family = family;
1386 key->prefixlen = prefixlen;
1387 key->l3index = l3index;
1389 memcpy(&key->addr, addr,
1390 (IS_ENABLED(CONFIG_IPV6) && family == AF_INET6) ? sizeof(struct in6_addr) :
1391 sizeof(struct in_addr));
1392 hlist_add_head_rcu(&key->node, &md5sig->head);
1396 int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
1397 int family, u8 prefixlen, int l3index, u8 flags,
1398 const u8 *newkey, u8 newkeylen)
1400 struct tcp_sock *tp = tcp_sk(sk);
1402 if (!rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk))) {
1403 if (tcp_md5_alloc_sigpool())
1406 if (tcp_md5sig_info_add(sk, GFP_KERNEL)) {
1407 tcp_md5_release_sigpool();
1411 if (!static_branch_inc(&tcp_md5_needed.key)) {
1412 struct tcp_md5sig_info *md5sig;
1414 md5sig = rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk));
1415 rcu_assign_pointer(tp->md5sig_info, NULL);
1416 kfree_rcu(md5sig, rcu);
1417 tcp_md5_release_sigpool();
1422 return __tcp_md5_do_add(sk, addr, family, prefixlen, l3index, flags,
1423 newkey, newkeylen, GFP_KERNEL);
1425 EXPORT_SYMBOL(tcp_md5_do_add);
1427 int tcp_md5_key_copy(struct sock *sk, const union tcp_md5_addr *addr,
1428 int family, u8 prefixlen, int l3index,
1429 struct tcp_md5sig_key *key)
1431 struct tcp_sock *tp = tcp_sk(sk);
1433 if (!rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk))) {
1434 tcp_md5_add_sigpool();
1436 if (tcp_md5sig_info_add(sk, sk_gfp_mask(sk, GFP_ATOMIC))) {
1437 tcp_md5_release_sigpool();
1441 if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key)) {
1442 struct tcp_md5sig_info *md5sig;
1444 md5sig = rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk));
1445 net_warn_ratelimited("Too many TCP-MD5 keys in the system\n");
1446 rcu_assign_pointer(tp->md5sig_info, NULL);
1447 kfree_rcu(md5sig, rcu);
1448 tcp_md5_release_sigpool();
1453 return __tcp_md5_do_add(sk, addr, family, prefixlen, l3index,
1454 key->flags, key->key, key->keylen,
1455 sk_gfp_mask(sk, GFP_ATOMIC));
1457 EXPORT_SYMBOL(tcp_md5_key_copy);
1459 int tcp_md5_do_del(struct sock *sk, const union tcp_md5_addr *addr, int family,
1460 u8 prefixlen, int l3index, u8 flags)
1462 struct tcp_md5sig_key *key;
1464 key = tcp_md5_do_lookup_exact(sk, addr, family, prefixlen, l3index, flags);
1467 hlist_del_rcu(&key->node);
1468 atomic_sub(sizeof(*key), &sk->sk_omem_alloc);
1469 kfree_rcu(key, rcu);
1472 EXPORT_SYMBOL(tcp_md5_do_del);
1474 void tcp_clear_md5_list(struct sock *sk)
1476 struct tcp_sock *tp = tcp_sk(sk);
1477 struct tcp_md5sig_key *key;
1478 struct hlist_node *n;
1479 struct tcp_md5sig_info *md5sig;
1481 md5sig = rcu_dereference_protected(tp->md5sig_info, 1);
1483 hlist_for_each_entry_safe(key, n, &md5sig->head, node) {
1484 hlist_del_rcu(&key->node);
1485 atomic_sub(sizeof(*key), &sk->sk_omem_alloc);
1486 kfree_rcu(key, rcu);
1490 static int tcp_v4_parse_md5_keys(struct sock *sk, int optname,
1491 sockptr_t optval, int optlen)
1493 struct tcp_md5sig cmd;
1494 struct sockaddr_in *sin = (struct sockaddr_in *)&cmd.tcpm_addr;
1495 const union tcp_md5_addr *addr;
1501 if (optlen < sizeof(cmd))
1504 if (copy_from_sockptr(&cmd, optval, sizeof(cmd)))
1507 if (sin->sin_family != AF_INET)
1510 flags = cmd.tcpm_flags & TCP_MD5SIG_FLAG_IFINDEX;
1511 l3flag = cmd.tcpm_flags & TCP_MD5SIG_FLAG_IFINDEX;
1513 if (optname == TCP_MD5SIG_EXT &&
1514 cmd.tcpm_flags & TCP_MD5SIG_FLAG_PREFIX) {
1515 prefixlen = cmd.tcpm_prefixlen;
1520 if (optname == TCP_MD5SIG_EXT && cmd.tcpm_ifindex &&
1521 cmd.tcpm_flags & TCP_MD5SIG_FLAG_IFINDEX) {
1522 struct net_device *dev;
1525 dev = dev_get_by_index_rcu(sock_net(sk), cmd.tcpm_ifindex);
1526 if (dev && netif_is_l3_master(dev))
1527 l3index = dev->ifindex;
1531 /* ok to reference set/not set outside of rcu;
1532 * right now device MUST be an L3 master
1534 if (!dev || !l3index)
1538 addr = (union tcp_md5_addr *)&sin->sin_addr.s_addr;
1540 if (!cmd.tcpm_keylen)
1541 return tcp_md5_do_del(sk, addr, AF_INET, prefixlen, l3index, flags);
1543 if (cmd.tcpm_keylen > TCP_MD5SIG_MAXKEYLEN)
1546 /* Don't allow keys for peers that have a matching TCP-AO key.
1547 * See the comment in tcp_ao_add_cmd()
1549 if (tcp_ao_required(sk, addr, AF_INET, l3flag ? l3index : -1, false))
1550 return -EKEYREJECTED;
1552 return tcp_md5_do_add(sk, addr, AF_INET, prefixlen, l3index, flags,
1553 cmd.tcpm_key, cmd.tcpm_keylen);
1556 static int tcp_v4_md5_hash_headers(struct tcp_sigpool *hp,
1557 __be32 daddr, __be32 saddr,
1558 const struct tcphdr *th, int nbytes)
1560 struct tcp4_pseudohdr *bp;
1561 struct scatterlist sg;
1568 bp->protocol = IPPROTO_TCP;
1569 bp->len = cpu_to_be16(nbytes);
1571 _th = (struct tcphdr *)(bp + 1);
1572 memcpy(_th, th, sizeof(*th));
1575 sg_init_one(&sg, bp, sizeof(*bp) + sizeof(*th));
1576 ahash_request_set_crypt(hp->req, &sg, NULL,
1577 sizeof(*bp) + sizeof(*th));
1578 return crypto_ahash_update(hp->req);
1581 static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
1582 __be32 daddr, __be32 saddr, const struct tcphdr *th)
1584 struct tcp_sigpool hp;
1586 if (tcp_sigpool_start(tcp_md5_sigpool_id, &hp))
1587 goto clear_hash_nostart;
1589 if (crypto_ahash_init(hp.req))
1591 if (tcp_v4_md5_hash_headers(&hp, daddr, saddr, th, th->doff << 2))
1593 if (tcp_md5_hash_key(&hp, key))
1595 ahash_request_set_crypt(hp.req, NULL, md5_hash, 0);
1596 if (crypto_ahash_final(hp.req))
1599 tcp_sigpool_end(&hp);
1603 tcp_sigpool_end(&hp);
1605 memset(md5_hash, 0, 16);
1609 int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
1610 const struct sock *sk,
1611 const struct sk_buff *skb)
1613 const struct tcphdr *th = tcp_hdr(skb);
1614 struct tcp_sigpool hp;
1615 __be32 saddr, daddr;
1617 if (sk) { /* valid for establish/request sockets */
1618 saddr = sk->sk_rcv_saddr;
1619 daddr = sk->sk_daddr;
1621 const struct iphdr *iph = ip_hdr(skb);
1626 if (tcp_sigpool_start(tcp_md5_sigpool_id, &hp))
1627 goto clear_hash_nostart;
1629 if (crypto_ahash_init(hp.req))
1632 if (tcp_v4_md5_hash_headers(&hp, daddr, saddr, th, skb->len))
1634 if (tcp_sigpool_hash_skb_data(&hp, skb, th->doff << 2))
1636 if (tcp_md5_hash_key(&hp, key))
1638 ahash_request_set_crypt(hp.req, NULL, md5_hash, 0);
1639 if (crypto_ahash_final(hp.req))
1642 tcp_sigpool_end(&hp);
1646 tcp_sigpool_end(&hp);
1648 memset(md5_hash, 0, 16);
1651 EXPORT_SYMBOL(tcp_v4_md5_hash_skb);
1655 static void tcp_v4_init_req(struct request_sock *req,
1656 const struct sock *sk_listener,
1657 struct sk_buff *skb)
1659 struct inet_request_sock *ireq = inet_rsk(req);
1660 struct net *net = sock_net(sk_listener);
1662 sk_rcv_saddr_set(req_to_sk(req), ip_hdr(skb)->daddr);
1663 sk_daddr_set(req_to_sk(req), ip_hdr(skb)->saddr);
1664 RCU_INIT_POINTER(ireq->ireq_opt, tcp_v4_save_options(net, skb));
1667 static struct dst_entry *tcp_v4_route_req(const struct sock *sk,
1668 struct sk_buff *skb,
1670 struct request_sock *req)
1672 tcp_v4_init_req(req, sk, skb);
1674 if (security_inet_conn_request(sk, skb, req))
1677 return inet_csk_route_req(sk, &fl->u.ip4, req);
1680 struct request_sock_ops tcp_request_sock_ops __read_mostly = {
1682 .obj_size = sizeof(struct tcp_request_sock),
1683 .rtx_syn_ack = tcp_rtx_synack,
1684 .send_ack = tcp_v4_reqsk_send_ack,
1685 .destructor = tcp_v4_reqsk_destructor,
1686 .send_reset = tcp_v4_send_reset,
1687 .syn_ack_timeout = tcp_syn_ack_timeout,
1690 const struct tcp_request_sock_ops tcp_request_sock_ipv4_ops = {
1691 .mss_clamp = TCP_MSS_DEFAULT,
1692 #ifdef CONFIG_TCP_MD5SIG
1693 .req_md5_lookup = tcp_v4_md5_lookup,
1694 .calc_md5_hash = tcp_v4_md5_hash_skb,
1696 #ifdef CONFIG_TCP_AO
1697 .ao_lookup = tcp_v4_ao_lookup_rsk,
1698 .ao_calc_key = tcp_v4_ao_calc_key_rsk,
1699 .ao_synack_hash = tcp_v4_ao_synack_hash,
1701 #ifdef CONFIG_SYN_COOKIES
1702 .cookie_init_seq = cookie_v4_init_sequence,
1704 .route_req = tcp_v4_route_req,
1705 .init_seq = tcp_v4_init_seq,
1706 .init_ts_off = tcp_v4_init_ts_off,
1707 .send_synack = tcp_v4_send_synack,
1710 int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb)
1712 /* Never answer to SYNs send to broadcast or multicast */
1713 if (skb_rtable(skb)->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST))
1716 return tcp_conn_request(&tcp_request_sock_ops,
1717 &tcp_request_sock_ipv4_ops, sk, skb);
1723 EXPORT_SYMBOL(tcp_v4_conn_request);
1727 * The three way handshake has completed - we got a valid synack -
1728 * now create the new socket.
1730 struct sock *tcp_v4_syn_recv_sock(const struct sock *sk, struct sk_buff *skb,
1731 struct request_sock *req,
1732 struct dst_entry *dst,
1733 struct request_sock *req_unhash,
1736 struct inet_request_sock *ireq;
1737 bool found_dup_sk = false;
1738 struct inet_sock *newinet;
1739 struct tcp_sock *newtp;
1741 #ifdef CONFIG_TCP_MD5SIG
1742 const union tcp_md5_addr *addr;
1743 struct tcp_md5sig_key *key;
1746 struct ip_options_rcu *inet_opt;
1748 if (sk_acceptq_is_full(sk))
1751 newsk = tcp_create_openreq_child(sk, req, skb);
1755 newsk->sk_gso_type = SKB_GSO_TCPV4;
1756 inet_sk_rx_dst_set(newsk, skb);
1758 newtp = tcp_sk(newsk);
1759 newinet = inet_sk(newsk);
1760 ireq = inet_rsk(req);
1761 sk_daddr_set(newsk, ireq->ir_rmt_addr);
1762 sk_rcv_saddr_set(newsk, ireq->ir_loc_addr);
1763 newsk->sk_bound_dev_if = ireq->ir_iif;
1764 newinet->inet_saddr = ireq->ir_loc_addr;
1765 inet_opt = rcu_dereference(ireq->ireq_opt);
1766 RCU_INIT_POINTER(newinet->inet_opt, inet_opt);
1767 newinet->mc_index = inet_iif(skb);
1768 newinet->mc_ttl = ip_hdr(skb)->ttl;
1769 newinet->rcv_tos = ip_hdr(skb)->tos;
1770 inet_csk(newsk)->icsk_ext_hdr_len = 0;
1772 inet_csk(newsk)->icsk_ext_hdr_len = inet_opt->opt.optlen;
1773 atomic_set(&newinet->inet_id, get_random_u16());
1775 /* Set ToS of the new socket based upon the value of incoming SYN.
1776 * ECT bits are set later in tcp_init_transfer().
1778 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reflect_tos))
1779 newinet->tos = tcp_rsk(req)->syn_tos & ~INET_ECN_MASK;
1782 dst = inet_csk_route_child_sock(sk, newsk, req);
1786 /* syncookie case : see end of cookie_v4_check() */
1788 sk_setup_caps(newsk, dst);
1790 tcp_ca_openreq_child(newsk, dst);
1792 tcp_sync_mss(newsk, dst_mtu(dst));
1793 newtp->advmss = tcp_mss_clamp(tcp_sk(sk), dst_metric_advmss(dst));
1795 tcp_initialize_rcv_mss(newsk);
1797 #ifdef CONFIG_TCP_MD5SIG
1798 l3index = l3mdev_master_ifindex_by_index(sock_net(sk), ireq->ir_iif);
1799 /* Copy over the MD5 key from the original socket */
1800 addr = (union tcp_md5_addr *)&newinet->inet_daddr;
1801 key = tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
1802 if (key && !tcp_rsk_used_ao(req)) {
1803 if (tcp_md5_key_copy(newsk, addr, AF_INET, 32, l3index, key))
1805 sk_gso_disable(newsk);
1808 #ifdef CONFIG_TCP_AO
1809 if (tcp_ao_copy_all_matching(sk, newsk, req, skb, AF_INET))
1810 goto put_and_exit; /* OOM, release back memory */
1813 if (__inet_inherit_port(sk, newsk) < 0)
1815 *own_req = inet_ehash_nolisten(newsk, req_to_sk(req_unhash),
1817 if (likely(*own_req)) {
1818 tcp_move_syn(newtp, req);
1819 ireq->ireq_opt = NULL;
1821 newinet->inet_opt = NULL;
1823 if (!req_unhash && found_dup_sk) {
1824 /* This code path should only be executed in the
1825 * syncookie case only
1827 bh_unlock_sock(newsk);
1835 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
1842 newinet->inet_opt = NULL;
1843 inet_csk_prepare_forced_close(newsk);
1847 EXPORT_SYMBOL(tcp_v4_syn_recv_sock);
1849 static struct sock *tcp_v4_cookie_check(struct sock *sk, struct sk_buff *skb)
1851 #ifdef CONFIG_SYN_COOKIES
1852 const struct tcphdr *th = tcp_hdr(skb);
1855 sk = cookie_v4_check(sk, skb);
1860 u16 tcp_v4_get_syncookie(struct sock *sk, struct iphdr *iph,
1861 struct tcphdr *th, u32 *cookie)
1864 #ifdef CONFIG_SYN_COOKIES
1865 mss = tcp_get_syncookie_mss(&tcp_request_sock_ops,
1866 &tcp_request_sock_ipv4_ops, sk, th);
1868 *cookie = __cookie_v4_init_sequence(iph, th, &mss);
1869 tcp_synq_overflow(sk);
1875 INDIRECT_CALLABLE_DECLARE(struct dst_entry *ipv4_dst_check(struct dst_entry *,
1877 /* The socket must have it's spinlock held when we get
1878 * here, unless it is a TCP_LISTEN socket.
1880 * We have a potential double-lock case here, so even when
1881 * doing backlog processing we use the BH locking scheme.
1882 * This is because we cannot sleep with the original spinlock
1885 int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb)
1887 enum skb_drop_reason reason;
1890 if (sk->sk_state == TCP_ESTABLISHED) { /* Fast path */
1891 struct dst_entry *dst;
1893 dst = rcu_dereference_protected(sk->sk_rx_dst,
1894 lockdep_sock_is_held(sk));
1896 sock_rps_save_rxhash(sk, skb);
1897 sk_mark_napi_id(sk, skb);
1899 if (sk->sk_rx_dst_ifindex != skb->skb_iif ||
1900 !INDIRECT_CALL_1(dst->ops->check, ipv4_dst_check,
1902 RCU_INIT_POINTER(sk->sk_rx_dst, NULL);
1906 tcp_rcv_established(sk, skb);
1910 reason = SKB_DROP_REASON_NOT_SPECIFIED;
1911 if (tcp_checksum_complete(skb))
1914 if (sk->sk_state == TCP_LISTEN) {
1915 struct sock *nsk = tcp_v4_cookie_check(sk, skb);
1920 if (tcp_child_process(sk, nsk, skb)) {
1927 sock_rps_save_rxhash(sk, skb);
1929 if (tcp_rcv_state_process(sk, skb)) {
1936 tcp_v4_send_reset(rsk, skb);
1938 kfree_skb_reason(skb, reason);
1939 /* Be careful here. If this function gets more complicated and
1940 * gcc suffers from register pressure on the x86, sk (in %ebx)
1941 * might be destroyed here. This current version compiles correctly,
1942 * but you have been warned.
1947 reason = SKB_DROP_REASON_TCP_CSUM;
1948 trace_tcp_bad_csum(skb);
1949 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
1950 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
1953 EXPORT_SYMBOL(tcp_v4_do_rcv);
1955 int tcp_v4_early_demux(struct sk_buff *skb)
1957 struct net *net = dev_net(skb->dev);
1958 const struct iphdr *iph;
1959 const struct tcphdr *th;
1962 if (skb->pkt_type != PACKET_HOST)
1965 if (!pskb_may_pull(skb, skb_transport_offset(skb) + sizeof(struct tcphdr)))
1971 if (th->doff < sizeof(struct tcphdr) / 4)
1974 sk = __inet_lookup_established(net, net->ipv4.tcp_death_row.hashinfo,
1975 iph->saddr, th->source,
1976 iph->daddr, ntohs(th->dest),
1977 skb->skb_iif, inet_sdif(skb));
1980 skb->destructor = sock_edemux;
1981 if (sk_fullsock(sk)) {
1982 struct dst_entry *dst = rcu_dereference(sk->sk_rx_dst);
1985 dst = dst_check(dst, 0);
1987 sk->sk_rx_dst_ifindex == skb->skb_iif)
1988 skb_dst_set_noref(skb, dst);
1994 bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb,
1995 enum skb_drop_reason *reason)
1997 u32 limit, tail_gso_size, tail_gso_segs;
1998 struct skb_shared_info *shinfo;
1999 const struct tcphdr *th;
2000 struct tcphdr *thtail;
2001 struct sk_buff *tail;
2002 unsigned int hdrlen;
2008 /* In case all data was pulled from skb frags (in __pskb_pull_tail()),
2009 * we can fix skb->truesize to its real value to avoid future drops.
2010 * This is valid because skb is not yet charged to the socket.
2011 * It has been noticed pure SACK packets were sometimes dropped
2012 * (if cooked by drivers without copybreak feature).
2018 if (unlikely(tcp_checksum_complete(skb))) {
2020 trace_tcp_bad_csum(skb);
2021 *reason = SKB_DROP_REASON_TCP_CSUM;
2022 __TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
2023 __TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
2027 /* Attempt coalescing to last skb in backlog, even if we are
2029 * This is okay because skb capacity is limited to MAX_SKB_FRAGS.
2031 th = (const struct tcphdr *)skb->data;
2032 hdrlen = th->doff * 4;
2034 tail = sk->sk_backlog.tail;
2037 thtail = (struct tcphdr *)tail->data;
2039 if (TCP_SKB_CB(tail)->end_seq != TCP_SKB_CB(skb)->seq ||
2040 TCP_SKB_CB(tail)->ip_dsfield != TCP_SKB_CB(skb)->ip_dsfield ||
2041 ((TCP_SKB_CB(tail)->tcp_flags |
2042 TCP_SKB_CB(skb)->tcp_flags) & (TCPHDR_SYN | TCPHDR_RST | TCPHDR_URG)) ||
2043 !((TCP_SKB_CB(tail)->tcp_flags &
2044 TCP_SKB_CB(skb)->tcp_flags) & TCPHDR_ACK) ||
2045 ((TCP_SKB_CB(tail)->tcp_flags ^
2046 TCP_SKB_CB(skb)->tcp_flags) & (TCPHDR_ECE | TCPHDR_CWR)) ||
2047 #ifdef CONFIG_TLS_DEVICE
2048 tail->decrypted != skb->decrypted ||
2050 !mptcp_skb_can_collapse(tail, skb) ||
2051 thtail->doff != th->doff ||
2052 memcmp(thtail + 1, th + 1, hdrlen - sizeof(*th)))
2055 __skb_pull(skb, hdrlen);
2057 shinfo = skb_shinfo(skb);
2058 gso_size = shinfo->gso_size ?: skb->len;
2059 gso_segs = shinfo->gso_segs ?: 1;
2061 shinfo = skb_shinfo(tail);
2062 tail_gso_size = shinfo->gso_size ?: (tail->len - hdrlen);
2063 tail_gso_segs = shinfo->gso_segs ?: 1;
2065 if (skb_try_coalesce(tail, skb, &fragstolen, &delta)) {
2066 TCP_SKB_CB(tail)->end_seq = TCP_SKB_CB(skb)->end_seq;
2068 if (likely(!before(TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(tail)->ack_seq))) {
2069 TCP_SKB_CB(tail)->ack_seq = TCP_SKB_CB(skb)->ack_seq;
2070 thtail->window = th->window;
2073 /* We have to update both TCP_SKB_CB(tail)->tcp_flags and
2074 * thtail->fin, so that the fast path in tcp_rcv_established()
2075 * is not entered if we append a packet with a FIN.
2076 * SYN, RST, URG are not present.
2077 * ACK is set on both packets.
2078 * PSH : we do not really care in TCP stack,
2079 * at least for 'GRO' packets.
2081 thtail->fin |= th->fin;
2082 TCP_SKB_CB(tail)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2084 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2085 TCP_SKB_CB(tail)->has_rxtstamp = true;
2086 tail->tstamp = skb->tstamp;
2087 skb_hwtstamps(tail)->hwtstamp = skb_hwtstamps(skb)->hwtstamp;
2090 /* Not as strict as GRO. We only need to carry mss max value */
2091 shinfo->gso_size = max(gso_size, tail_gso_size);
2092 shinfo->gso_segs = min_t(u32, gso_segs + tail_gso_segs, 0xFFFF);
2094 sk->sk_backlog.len += delta;
2095 __NET_INC_STATS(sock_net(sk),
2096 LINUX_MIB_TCPBACKLOGCOALESCE);
2097 kfree_skb_partial(skb, fragstolen);
2100 __skb_push(skb, hdrlen);
2103 limit = (u32)READ_ONCE(sk->sk_rcvbuf) + (u32)(READ_ONCE(sk->sk_sndbuf) >> 1);
2105 /* Only socket owner can try to collapse/prune rx queues
2106 * to reduce memory overhead, so add a little headroom here.
2107 * Few sockets backlog are possibly concurrently non empty.
2111 if (unlikely(sk_add_backlog(sk, skb, limit))) {
2113 *reason = SKB_DROP_REASON_SOCKET_BACKLOG;
2114 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPBACKLOGDROP);
2119 EXPORT_SYMBOL(tcp_add_backlog);
2121 int tcp_filter(struct sock *sk, struct sk_buff *skb)
2123 struct tcphdr *th = (struct tcphdr *)skb->data;
2125 return sk_filter_trim_cap(sk, skb, th->doff * 4);
2127 EXPORT_SYMBOL(tcp_filter);
2129 static void tcp_v4_restore_cb(struct sk_buff *skb)
2131 memmove(IPCB(skb), &TCP_SKB_CB(skb)->header.h4,
2132 sizeof(struct inet_skb_parm));
2135 static void tcp_v4_fill_cb(struct sk_buff *skb, const struct iphdr *iph,
2136 const struct tcphdr *th)
2138 /* This is tricky : We move IPCB at its correct location into TCP_SKB_CB()
2139 * barrier() makes sure compiler wont play fool^Waliasing games.
2141 memmove(&TCP_SKB_CB(skb)->header.h4, IPCB(skb),
2142 sizeof(struct inet_skb_parm));
2145 TCP_SKB_CB(skb)->seq = ntohl(th->seq);
2146 TCP_SKB_CB(skb)->end_seq = (TCP_SKB_CB(skb)->seq + th->syn + th->fin +
2147 skb->len - th->doff * 4);
2148 TCP_SKB_CB(skb)->ack_seq = ntohl(th->ack_seq);
2149 TCP_SKB_CB(skb)->tcp_flags = tcp_flag_byte(th);
2150 TCP_SKB_CB(skb)->tcp_tw_isn = 0;
2151 TCP_SKB_CB(skb)->ip_dsfield = ipv4_get_dsfield(iph);
2152 TCP_SKB_CB(skb)->sacked = 0;
2153 TCP_SKB_CB(skb)->has_rxtstamp =
2154 skb->tstamp || skb_hwtstamps(skb)->hwtstamp;
2161 int tcp_v4_rcv(struct sk_buff *skb)
2163 struct net *net = dev_net(skb->dev);
2164 enum skb_drop_reason drop_reason;
2165 int sdif = inet_sdif(skb);
2166 int dif = inet_iif(skb);
2167 const struct iphdr *iph;
2168 const struct tcphdr *th;
2173 drop_reason = SKB_DROP_REASON_NOT_SPECIFIED;
2174 if (skb->pkt_type != PACKET_HOST)
2177 /* Count it even if it's bad */
2178 __TCP_INC_STATS(net, TCP_MIB_INSEGS);
2180 if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
2183 th = (const struct tcphdr *)skb->data;
2185 if (unlikely(th->doff < sizeof(struct tcphdr) / 4)) {
2186 drop_reason = SKB_DROP_REASON_PKT_TOO_SMALL;
2189 if (!pskb_may_pull(skb, th->doff * 4))
2192 /* An explanation is required here, I think.
2193 * Packet length and doff are validated by header prediction,
2194 * provided case of th->doff==0 is eliminated.
2195 * So, we defer the checks. */
2197 if (skb_checksum_init(skb, IPPROTO_TCP, inet_compute_pseudo))
2200 th = (const struct tcphdr *)skb->data;
2203 sk = __inet_lookup_skb(net->ipv4.tcp_death_row.hashinfo,
2204 skb, __tcp_hdrlen(th), th->source,
2205 th->dest, sdif, &refcounted);
2210 if (sk->sk_state == TCP_TIME_WAIT)
2213 if (sk->sk_state == TCP_NEW_SYN_RECV) {
2214 struct request_sock *req = inet_reqsk(sk);
2215 bool req_stolen = false;
2218 sk = req->rsk_listener;
2219 if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb))
2220 drop_reason = SKB_DROP_REASON_XFRM_POLICY;
2222 drop_reason = tcp_inbound_hash(sk, req, skb,
2223 &iph->saddr, &iph->daddr,
2224 AF_INET, dif, sdif);
2225 if (unlikely(drop_reason)) {
2226 sk_drops_add(sk, skb);
2230 if (tcp_checksum_complete(skb)) {
2234 if (unlikely(sk->sk_state != TCP_LISTEN)) {
2235 nsk = reuseport_migrate_sock(sk, req_to_sk(req), skb);
2237 inet_csk_reqsk_queue_drop_and_put(sk, req);
2241 /* reuseport_migrate_sock() has already held one sk_refcnt
2245 /* We own a reference on the listener, increase it again
2246 * as we might lose it too soon.
2252 if (!tcp_filter(sk, skb)) {
2253 th = (const struct tcphdr *)skb->data;
2255 tcp_v4_fill_cb(skb, iph, th);
2256 nsk = tcp_check_req(sk, skb, req, false, &req_stolen);
2258 drop_reason = SKB_DROP_REASON_SOCKET_FILTER;
2263 /* Another cpu got exclusive access to req
2264 * and created a full blown socket.
2265 * Try to feed this packet to this socket
2266 * instead of discarding it.
2268 tcp_v4_restore_cb(skb);
2272 goto discard_and_relse;
2277 tcp_v4_restore_cb(skb);
2278 } else if (tcp_child_process(sk, nsk, skb)) {
2279 tcp_v4_send_reset(nsk, skb);
2280 goto discard_and_relse;
2287 if (static_branch_unlikely(&ip4_min_ttl)) {
2288 /* min_ttl can be changed concurrently from do_ip_setsockopt() */
2289 if (unlikely(iph->ttl < READ_ONCE(inet_sk(sk)->min_ttl))) {
2290 __NET_INC_STATS(net, LINUX_MIB_TCPMINTTLDROP);
2291 drop_reason = SKB_DROP_REASON_TCP_MINTTL;
2292 goto discard_and_relse;
2296 if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb)) {
2297 drop_reason = SKB_DROP_REASON_XFRM_POLICY;
2298 goto discard_and_relse;
2301 drop_reason = tcp_inbound_hash(sk, NULL, skb, &iph->saddr, &iph->daddr,
2302 AF_INET, dif, sdif);
2304 goto discard_and_relse;
2308 if (tcp_filter(sk, skb)) {
2309 drop_reason = SKB_DROP_REASON_SOCKET_FILTER;
2310 goto discard_and_relse;
2312 th = (const struct tcphdr *)skb->data;
2314 tcp_v4_fill_cb(skb, iph, th);
2318 if (sk->sk_state == TCP_LISTEN) {
2319 ret = tcp_v4_do_rcv(sk, skb);
2320 goto put_and_return;
2323 sk_incoming_cpu_update(sk);
2325 bh_lock_sock_nested(sk);
2326 tcp_segs_in(tcp_sk(sk), skb);
2328 if (!sock_owned_by_user(sk)) {
2329 ret = tcp_v4_do_rcv(sk, skb);
2331 if (tcp_add_backlog(sk, skb, &drop_reason))
2332 goto discard_and_relse;
2343 drop_reason = SKB_DROP_REASON_NO_SOCKET;
2344 if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb))
2347 tcp_v4_fill_cb(skb, iph, th);
2349 if (tcp_checksum_complete(skb)) {
2351 drop_reason = SKB_DROP_REASON_TCP_CSUM;
2352 trace_tcp_bad_csum(skb);
2353 __TCP_INC_STATS(net, TCP_MIB_CSUMERRORS);
2355 __TCP_INC_STATS(net, TCP_MIB_INERRS);
2357 tcp_v4_send_reset(NULL, skb);
2361 SKB_DR_OR(drop_reason, NOT_SPECIFIED);
2362 /* Discard frame. */
2363 kfree_skb_reason(skb, drop_reason);
2367 sk_drops_add(sk, skb);
2373 if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb)) {
2374 drop_reason = SKB_DROP_REASON_XFRM_POLICY;
2375 inet_twsk_put(inet_twsk(sk));
2379 tcp_v4_fill_cb(skb, iph, th);
2381 if (tcp_checksum_complete(skb)) {
2382 inet_twsk_put(inet_twsk(sk));
2385 switch (tcp_timewait_state_process(inet_twsk(sk), skb, th)) {
2387 struct sock *sk2 = inet_lookup_listener(net,
2388 net->ipv4.tcp_death_row.hashinfo,
2389 skb, __tcp_hdrlen(th),
2390 iph->saddr, th->source,
2391 iph->daddr, th->dest,
2395 inet_twsk_deschedule_put(inet_twsk(sk));
2397 tcp_v4_restore_cb(skb);
2405 tcp_v4_timewait_ack(sk, skb);
2408 tcp_v4_send_reset(sk, skb);
2409 inet_twsk_deschedule_put(inet_twsk(sk));
2411 case TCP_TW_SUCCESS:;
2416 static struct timewait_sock_ops tcp_timewait_sock_ops = {
2417 .twsk_obj_size = sizeof(struct tcp_timewait_sock),
2418 .twsk_unique = tcp_twsk_unique,
2419 .twsk_destructor= tcp_twsk_destructor,
2422 void inet_sk_rx_dst_set(struct sock *sk, const struct sk_buff *skb)
2424 struct dst_entry *dst = skb_dst(skb);
2426 if (dst && dst_hold_safe(dst)) {
2427 rcu_assign_pointer(sk->sk_rx_dst, dst);
2428 sk->sk_rx_dst_ifindex = skb->skb_iif;
2431 EXPORT_SYMBOL(inet_sk_rx_dst_set);
2433 const struct inet_connection_sock_af_ops ipv4_specific = {
2434 .queue_xmit = ip_queue_xmit,
2435 .send_check = tcp_v4_send_check,
2436 .rebuild_header = inet_sk_rebuild_header,
2437 .sk_rx_dst_set = inet_sk_rx_dst_set,
2438 .conn_request = tcp_v4_conn_request,
2439 .syn_recv_sock = tcp_v4_syn_recv_sock,
2440 .net_header_len = sizeof(struct iphdr),
2441 .setsockopt = ip_setsockopt,
2442 .getsockopt = ip_getsockopt,
2443 .addr2sockaddr = inet_csk_addr2sockaddr,
2444 .sockaddr_len = sizeof(struct sockaddr_in),
2445 .mtu_reduced = tcp_v4_mtu_reduced,
2447 EXPORT_SYMBOL(ipv4_specific);
2449 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
2450 static const struct tcp_sock_af_ops tcp_sock_ipv4_specific = {
2451 #ifdef CONFIG_TCP_MD5SIG
2452 .md5_lookup = tcp_v4_md5_lookup,
2453 .calc_md5_hash = tcp_v4_md5_hash_skb,
2454 .md5_parse = tcp_v4_parse_md5_keys,
2456 #ifdef CONFIG_TCP_AO
2457 .ao_lookup = tcp_v4_ao_lookup,
2458 .calc_ao_hash = tcp_v4_ao_hash_skb,
2459 .ao_parse = tcp_v4_parse_ao,
2460 .ao_calc_key_sk = tcp_v4_ao_calc_key_sk,
2465 /* NOTE: A lot of things set to zero explicitly by call to
2466 * sk_alloc() so need not be done here.
2468 static int tcp_v4_init_sock(struct sock *sk)
2470 struct inet_connection_sock *icsk = inet_csk(sk);
2474 icsk->icsk_af_ops = &ipv4_specific;
2476 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
2477 tcp_sk(sk)->af_specific = &tcp_sock_ipv4_specific;
2483 #ifdef CONFIG_TCP_MD5SIG
2484 static void tcp_md5sig_info_free_rcu(struct rcu_head *head)
2486 struct tcp_md5sig_info *md5sig;
2488 md5sig = container_of(head, struct tcp_md5sig_info, rcu);
2490 static_branch_slow_dec_deferred(&tcp_md5_needed);
2491 tcp_md5_release_sigpool();
2495 void tcp_v4_destroy_sock(struct sock *sk)
2497 struct tcp_sock *tp = tcp_sk(sk);
2499 trace_tcp_destroy_sock(sk);
2501 tcp_clear_xmit_timers(sk);
2503 tcp_cleanup_congestion_control(sk);
2505 tcp_cleanup_ulp(sk);
2507 /* Cleanup up the write buffer. */
2508 tcp_write_queue_purge(sk);
2510 /* Check if we want to disable active TFO */
2511 tcp_fastopen_active_disable_ofo_check(sk);
2513 /* Cleans up our, hopefully empty, out_of_order_queue. */
2514 skb_rbtree_purge(&tp->out_of_order_queue);
2516 #ifdef CONFIG_TCP_MD5SIG
2517 /* Clean up the MD5 key list, if any */
2518 if (tp->md5sig_info) {
2519 struct tcp_md5sig_info *md5sig;
2521 md5sig = rcu_dereference_protected(tp->md5sig_info, 1);
2522 tcp_clear_md5_list(sk);
2523 call_rcu(&md5sig->rcu, tcp_md5sig_info_free_rcu);
2524 rcu_assign_pointer(tp->md5sig_info, NULL);
2527 tcp_ao_destroy_sock(sk, false);
2529 /* Clean up a referenced TCP bind bucket. */
2530 if (inet_csk(sk)->icsk_bind_hash)
2533 BUG_ON(rcu_access_pointer(tp->fastopen_rsk));
2535 /* If socket is aborted during connect operation */
2536 tcp_free_fastopen_req(tp);
2537 tcp_fastopen_destroy_cipher(sk);
2538 tcp_saved_syn_free(tp);
2540 sk_sockets_allocated_dec(sk);
2542 EXPORT_SYMBOL(tcp_v4_destroy_sock);
2544 #ifdef CONFIG_PROC_FS
2545 /* Proc filesystem TCP sock list dumping. */
2547 static unsigned short seq_file_family(const struct seq_file *seq);
2549 static bool seq_sk_match(struct seq_file *seq, const struct sock *sk)
2551 unsigned short family = seq_file_family(seq);
2553 /* AF_UNSPEC is used as a match all */
2554 return ((family == AF_UNSPEC || family == sk->sk_family) &&
2555 net_eq(sock_net(sk), seq_file_net(seq)));
2558 /* Find a non empty bucket (starting from st->bucket)
2559 * and return the first sk from it.
2561 static void *listening_get_first(struct seq_file *seq)
2563 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2564 struct tcp_iter_state *st = seq->private;
2567 for (; st->bucket <= hinfo->lhash2_mask; st->bucket++) {
2568 struct inet_listen_hashbucket *ilb2;
2569 struct hlist_nulls_node *node;
2572 ilb2 = &hinfo->lhash2[st->bucket];
2573 if (hlist_nulls_empty(&ilb2->nulls_head))
2576 spin_lock(&ilb2->lock);
2577 sk_nulls_for_each(sk, node, &ilb2->nulls_head) {
2578 if (seq_sk_match(seq, sk))
2581 spin_unlock(&ilb2->lock);
2587 /* Find the next sk of "cur" within the same bucket (i.e. st->bucket).
2588 * If "cur" is the last one in the st->bucket,
2589 * call listening_get_first() to return the first sk of the next
2592 static void *listening_get_next(struct seq_file *seq, void *cur)
2594 struct tcp_iter_state *st = seq->private;
2595 struct inet_listen_hashbucket *ilb2;
2596 struct hlist_nulls_node *node;
2597 struct inet_hashinfo *hinfo;
2598 struct sock *sk = cur;
2603 sk = sk_nulls_next(sk);
2604 sk_nulls_for_each_from(sk, node) {
2605 if (seq_sk_match(seq, sk))
2609 hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2610 ilb2 = &hinfo->lhash2[st->bucket];
2611 spin_unlock(&ilb2->lock);
2613 return listening_get_first(seq);
2616 static void *listening_get_idx(struct seq_file *seq, loff_t *pos)
2618 struct tcp_iter_state *st = seq->private;
2623 rc = listening_get_first(seq);
2625 while (rc && *pos) {
2626 rc = listening_get_next(seq, rc);
2632 static inline bool empty_bucket(struct inet_hashinfo *hinfo,
2633 const struct tcp_iter_state *st)
2635 return hlist_nulls_empty(&hinfo->ehash[st->bucket].chain);
2639 * Get first established socket starting from bucket given in st->bucket.
2640 * If st->bucket is zero, the very first socket in the hash is returned.
2642 static void *established_get_first(struct seq_file *seq)
2644 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2645 struct tcp_iter_state *st = seq->private;
2648 for (; st->bucket <= hinfo->ehash_mask; ++st->bucket) {
2650 struct hlist_nulls_node *node;
2651 spinlock_t *lock = inet_ehash_lockp(hinfo, st->bucket);
2655 /* Lockless fast path for the common case of empty buckets */
2656 if (empty_bucket(hinfo, st))
2660 sk_nulls_for_each(sk, node, &hinfo->ehash[st->bucket].chain) {
2661 if (seq_sk_match(seq, sk))
2664 spin_unlock_bh(lock);
2670 static void *established_get_next(struct seq_file *seq, void *cur)
2672 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2673 struct tcp_iter_state *st = seq->private;
2674 struct hlist_nulls_node *node;
2675 struct sock *sk = cur;
2680 sk = sk_nulls_next(sk);
2682 sk_nulls_for_each_from(sk, node) {
2683 if (seq_sk_match(seq, sk))
2687 spin_unlock_bh(inet_ehash_lockp(hinfo, st->bucket));
2689 return established_get_first(seq);
2692 static void *established_get_idx(struct seq_file *seq, loff_t pos)
2694 struct tcp_iter_state *st = seq->private;
2698 rc = established_get_first(seq);
2701 rc = established_get_next(seq, rc);
2707 static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
2710 struct tcp_iter_state *st = seq->private;
2712 st->state = TCP_SEQ_STATE_LISTENING;
2713 rc = listening_get_idx(seq, &pos);
2716 st->state = TCP_SEQ_STATE_ESTABLISHED;
2717 rc = established_get_idx(seq, pos);
2723 static void *tcp_seek_last_pos(struct seq_file *seq)
2725 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2726 struct tcp_iter_state *st = seq->private;
2727 int bucket = st->bucket;
2728 int offset = st->offset;
2729 int orig_num = st->num;
2732 switch (st->state) {
2733 case TCP_SEQ_STATE_LISTENING:
2734 if (st->bucket > hinfo->lhash2_mask)
2736 rc = listening_get_first(seq);
2737 while (offset-- && rc && bucket == st->bucket)
2738 rc = listening_get_next(seq, rc);
2742 st->state = TCP_SEQ_STATE_ESTABLISHED;
2744 case TCP_SEQ_STATE_ESTABLISHED:
2745 if (st->bucket > hinfo->ehash_mask)
2747 rc = established_get_first(seq);
2748 while (offset-- && rc && bucket == st->bucket)
2749 rc = established_get_next(seq, rc);
2757 void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
2759 struct tcp_iter_state *st = seq->private;
2762 if (*pos && *pos == st->last_pos) {
2763 rc = tcp_seek_last_pos(seq);
2768 st->state = TCP_SEQ_STATE_LISTENING;
2772 rc = *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
2775 st->last_pos = *pos;
2778 EXPORT_SYMBOL(tcp_seq_start);
2780 void *tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2782 struct tcp_iter_state *st = seq->private;
2785 if (v == SEQ_START_TOKEN) {
2786 rc = tcp_get_idx(seq, 0);
2790 switch (st->state) {
2791 case TCP_SEQ_STATE_LISTENING:
2792 rc = listening_get_next(seq, v);
2794 st->state = TCP_SEQ_STATE_ESTABLISHED;
2797 rc = established_get_first(seq);
2800 case TCP_SEQ_STATE_ESTABLISHED:
2801 rc = established_get_next(seq, v);
2806 st->last_pos = *pos;
2809 EXPORT_SYMBOL(tcp_seq_next);
2811 void tcp_seq_stop(struct seq_file *seq, void *v)
2813 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2814 struct tcp_iter_state *st = seq->private;
2816 switch (st->state) {
2817 case TCP_SEQ_STATE_LISTENING:
2818 if (v != SEQ_START_TOKEN)
2819 spin_unlock(&hinfo->lhash2[st->bucket].lock);
2821 case TCP_SEQ_STATE_ESTABLISHED:
2823 spin_unlock_bh(inet_ehash_lockp(hinfo, st->bucket));
2827 EXPORT_SYMBOL(tcp_seq_stop);
2829 static void get_openreq4(const struct request_sock *req,
2830 struct seq_file *f, int i)
2832 const struct inet_request_sock *ireq = inet_rsk(req);
2833 long delta = req->rsk_timer.expires - jiffies;
2835 seq_printf(f, "%4d: %08X:%04X %08X:%04X"
2836 " %02X %08X:%08X %02X:%08lX %08X %5u %8d %u %d %pK",
2841 ntohs(ireq->ir_rmt_port),
2843 0, 0, /* could print option size, but that is af dependent. */
2844 1, /* timers active (only the expire timer) */
2845 jiffies_delta_to_clock_t(delta),
2847 from_kuid_munged(seq_user_ns(f),
2848 sock_i_uid(req->rsk_listener)),
2849 0, /* non standard timer */
2850 0, /* open_requests have no inode */
2855 static void get_tcp4_sock(struct sock *sk, struct seq_file *f, int i)
2858 unsigned long timer_expires;
2859 const struct tcp_sock *tp = tcp_sk(sk);
2860 const struct inet_connection_sock *icsk = inet_csk(sk);
2861 const struct inet_sock *inet = inet_sk(sk);
2862 const struct fastopen_queue *fastopenq = &icsk->icsk_accept_queue.fastopenq;
2863 __be32 dest = inet->inet_daddr;
2864 __be32 src = inet->inet_rcv_saddr;
2865 __u16 destp = ntohs(inet->inet_dport);
2866 __u16 srcp = ntohs(inet->inet_sport);
2870 if (icsk->icsk_pending == ICSK_TIME_RETRANS ||
2871 icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2872 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2874 timer_expires = icsk->icsk_timeout;
2875 } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
2877 timer_expires = icsk->icsk_timeout;
2878 } else if (timer_pending(&sk->sk_timer)) {
2880 timer_expires = sk->sk_timer.expires;
2883 timer_expires = jiffies;
2886 state = inet_sk_state_load(sk);
2887 if (state == TCP_LISTEN)
2888 rx_queue = READ_ONCE(sk->sk_ack_backlog);
2890 /* Because we don't lock the socket,
2891 * we might find a transient negative value.
2893 rx_queue = max_t(int, READ_ONCE(tp->rcv_nxt) -
2894 READ_ONCE(tp->copied_seq), 0);
2896 seq_printf(f, "%4d: %08X:%04X %08X:%04X %02X %08X:%08X %02X:%08lX "
2897 "%08X %5u %8d %lu %d %pK %lu %lu %u %u %d",
2898 i, src, srcp, dest, destp, state,
2899 READ_ONCE(tp->write_seq) - tp->snd_una,
2902 jiffies_delta_to_clock_t(timer_expires - jiffies),
2903 icsk->icsk_retransmits,
2904 from_kuid_munged(seq_user_ns(f), sock_i_uid(sk)),
2905 icsk->icsk_probes_out,
2907 refcount_read(&sk->sk_refcnt), sk,
2908 jiffies_to_clock_t(icsk->icsk_rto),
2909 jiffies_to_clock_t(icsk->icsk_ack.ato),
2910 (icsk->icsk_ack.quick << 1) | inet_csk_in_pingpong_mode(sk),
2912 state == TCP_LISTEN ?
2913 fastopenq->max_qlen :
2914 (tcp_in_initial_slowstart(tp) ? -1 : tp->snd_ssthresh));
2917 static void get_timewait4_sock(const struct inet_timewait_sock *tw,
2918 struct seq_file *f, int i)
2920 long delta = tw->tw_timer.expires - jiffies;
2924 dest = tw->tw_daddr;
2925 src = tw->tw_rcv_saddr;
2926 destp = ntohs(tw->tw_dport);
2927 srcp = ntohs(tw->tw_sport);
2929 seq_printf(f, "%4d: %08X:%04X %08X:%04X"
2930 " %02X %08X:%08X %02X:%08lX %08X %5d %8d %d %d %pK",
2931 i, src, srcp, dest, destp, tw->tw_substate, 0, 0,
2932 3, jiffies_delta_to_clock_t(delta), 0, 0, 0, 0,
2933 refcount_read(&tw->tw_refcnt), tw);
2938 static int tcp4_seq_show(struct seq_file *seq, void *v)
2940 struct tcp_iter_state *st;
2941 struct sock *sk = v;
2943 seq_setwidth(seq, TMPSZ - 1);
2944 if (v == SEQ_START_TOKEN) {
2945 seq_puts(seq, " sl local_address rem_address st tx_queue "
2946 "rx_queue tr tm->when retrnsmt uid timeout "
2952 if (sk->sk_state == TCP_TIME_WAIT)
2953 get_timewait4_sock(v, seq, st->num);
2954 else if (sk->sk_state == TCP_NEW_SYN_RECV)
2955 get_openreq4(v, seq, st->num);
2957 get_tcp4_sock(v, seq, st->num);
2963 #ifdef CONFIG_BPF_SYSCALL
2964 struct bpf_tcp_iter_state {
2965 struct tcp_iter_state state;
2966 unsigned int cur_sk;
2967 unsigned int end_sk;
2968 unsigned int max_sk;
2969 struct sock **batch;
2970 bool st_bucket_done;
2973 struct bpf_iter__tcp {
2974 __bpf_md_ptr(struct bpf_iter_meta *, meta);
2975 __bpf_md_ptr(struct sock_common *, sk_common);
2976 uid_t uid __aligned(8);
2979 static int tcp_prog_seq_show(struct bpf_prog *prog, struct bpf_iter_meta *meta,
2980 struct sock_common *sk_common, uid_t uid)
2982 struct bpf_iter__tcp ctx;
2984 meta->seq_num--; /* skip SEQ_START_TOKEN */
2986 ctx.sk_common = sk_common;
2988 return bpf_iter_run_prog(prog, &ctx);
2991 static void bpf_iter_tcp_put_batch(struct bpf_tcp_iter_state *iter)
2993 while (iter->cur_sk < iter->end_sk)
2994 sock_gen_put(iter->batch[iter->cur_sk++]);
2997 static int bpf_iter_tcp_realloc_batch(struct bpf_tcp_iter_state *iter,
2998 unsigned int new_batch_sz)
3000 struct sock **new_batch;
3002 new_batch = kvmalloc(sizeof(*new_batch) * new_batch_sz,
3003 GFP_USER | __GFP_NOWARN);
3007 bpf_iter_tcp_put_batch(iter);
3008 kvfree(iter->batch);
3009 iter->batch = new_batch;
3010 iter->max_sk = new_batch_sz;
3015 static unsigned int bpf_iter_tcp_listening_batch(struct seq_file *seq,
3016 struct sock *start_sk)
3018 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
3019 struct bpf_tcp_iter_state *iter = seq->private;
3020 struct tcp_iter_state *st = &iter->state;
3021 struct hlist_nulls_node *node;
3022 unsigned int expected = 1;
3025 sock_hold(start_sk);
3026 iter->batch[iter->end_sk++] = start_sk;
3028 sk = sk_nulls_next(start_sk);
3029 sk_nulls_for_each_from(sk, node) {
3030 if (seq_sk_match(seq, sk)) {
3031 if (iter->end_sk < iter->max_sk) {
3033 iter->batch[iter->end_sk++] = sk;
3038 spin_unlock(&hinfo->lhash2[st->bucket].lock);
3043 static unsigned int bpf_iter_tcp_established_batch(struct seq_file *seq,
3044 struct sock *start_sk)
3046 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
3047 struct bpf_tcp_iter_state *iter = seq->private;
3048 struct tcp_iter_state *st = &iter->state;
3049 struct hlist_nulls_node *node;
3050 unsigned int expected = 1;
3053 sock_hold(start_sk);
3054 iter->batch[iter->end_sk++] = start_sk;
3056 sk = sk_nulls_next(start_sk);
3057 sk_nulls_for_each_from(sk, node) {
3058 if (seq_sk_match(seq, sk)) {
3059 if (iter->end_sk < iter->max_sk) {
3061 iter->batch[iter->end_sk++] = sk;
3066 spin_unlock_bh(inet_ehash_lockp(hinfo, st->bucket));
3071 static struct sock *bpf_iter_tcp_batch(struct seq_file *seq)
3073 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
3074 struct bpf_tcp_iter_state *iter = seq->private;
3075 struct tcp_iter_state *st = &iter->state;
3076 unsigned int expected;
3077 bool resized = false;
3080 /* The st->bucket is done. Directly advance to the next
3081 * bucket instead of having the tcp_seek_last_pos() to skip
3082 * one by one in the current bucket and eventually find out
3083 * it has to advance to the next bucket.
3085 if (iter->st_bucket_done) {
3088 if (st->state == TCP_SEQ_STATE_LISTENING &&
3089 st->bucket > hinfo->lhash2_mask) {
3090 st->state = TCP_SEQ_STATE_ESTABLISHED;
3096 /* Get a new batch */
3099 iter->st_bucket_done = false;
3101 sk = tcp_seek_last_pos(seq);
3103 return NULL; /* Done */
3105 if (st->state == TCP_SEQ_STATE_LISTENING)
3106 expected = bpf_iter_tcp_listening_batch(seq, sk);
3108 expected = bpf_iter_tcp_established_batch(seq, sk);
3110 if (iter->end_sk == expected) {
3111 iter->st_bucket_done = true;
3115 if (!resized && !bpf_iter_tcp_realloc_batch(iter, expected * 3 / 2)) {
3123 static void *bpf_iter_tcp_seq_start(struct seq_file *seq, loff_t *pos)
3125 /* bpf iter does not support lseek, so it always
3126 * continue from where it was stop()-ped.
3129 return bpf_iter_tcp_batch(seq);
3131 return SEQ_START_TOKEN;
3134 static void *bpf_iter_tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
3136 struct bpf_tcp_iter_state *iter = seq->private;
3137 struct tcp_iter_state *st = &iter->state;
3140 /* Whenever seq_next() is called, the iter->cur_sk is
3141 * done with seq_show(), so advance to the next sk in
3144 if (iter->cur_sk < iter->end_sk) {
3145 /* Keeping st->num consistent in tcp_iter_state.
3146 * bpf_iter_tcp does not use st->num.
3147 * meta.seq_num is used instead.
3150 /* Move st->offset to the next sk in the bucket such that
3151 * the future start() will resume at st->offset in
3152 * st->bucket. See tcp_seek_last_pos().
3155 sock_gen_put(iter->batch[iter->cur_sk++]);
3158 if (iter->cur_sk < iter->end_sk)
3159 sk = iter->batch[iter->cur_sk];
3161 sk = bpf_iter_tcp_batch(seq);
3164 /* Keeping st->last_pos consistent in tcp_iter_state.
3165 * bpf iter does not do lseek, so st->last_pos always equals to *pos.
3167 st->last_pos = *pos;
3171 static int bpf_iter_tcp_seq_show(struct seq_file *seq, void *v)
3173 struct bpf_iter_meta meta;
3174 struct bpf_prog *prog;
3175 struct sock *sk = v;
3179 if (v == SEQ_START_TOKEN)
3182 if (sk_fullsock(sk))
3185 if (unlikely(sk_unhashed(sk))) {
3190 if (sk->sk_state == TCP_TIME_WAIT) {
3192 } else if (sk->sk_state == TCP_NEW_SYN_RECV) {
3193 const struct request_sock *req = v;
3195 uid = from_kuid_munged(seq_user_ns(seq),
3196 sock_i_uid(req->rsk_listener));
3198 uid = from_kuid_munged(seq_user_ns(seq), sock_i_uid(sk));
3202 prog = bpf_iter_get_info(&meta, false);
3203 ret = tcp_prog_seq_show(prog, &meta, v, uid);
3206 if (sk_fullsock(sk))
3212 static void bpf_iter_tcp_seq_stop(struct seq_file *seq, void *v)
3214 struct bpf_tcp_iter_state *iter = seq->private;
3215 struct bpf_iter_meta meta;
3216 struct bpf_prog *prog;
3220 prog = bpf_iter_get_info(&meta, true);
3222 (void)tcp_prog_seq_show(prog, &meta, v, 0);
3225 if (iter->cur_sk < iter->end_sk) {
3226 bpf_iter_tcp_put_batch(iter);
3227 iter->st_bucket_done = false;
3231 static const struct seq_operations bpf_iter_tcp_seq_ops = {
3232 .show = bpf_iter_tcp_seq_show,
3233 .start = bpf_iter_tcp_seq_start,
3234 .next = bpf_iter_tcp_seq_next,
3235 .stop = bpf_iter_tcp_seq_stop,
3238 static unsigned short seq_file_family(const struct seq_file *seq)
3240 const struct tcp_seq_afinfo *afinfo;
3242 #ifdef CONFIG_BPF_SYSCALL
3243 /* Iterated from bpf_iter. Let the bpf prog to filter instead. */
3244 if (seq->op == &bpf_iter_tcp_seq_ops)
3248 /* Iterated from proc fs */
3249 afinfo = pde_data(file_inode(seq->file));
3250 return afinfo->family;
3253 static const struct seq_operations tcp4_seq_ops = {
3254 .show = tcp4_seq_show,
3255 .start = tcp_seq_start,
3256 .next = tcp_seq_next,
3257 .stop = tcp_seq_stop,
3260 static struct tcp_seq_afinfo tcp4_seq_afinfo = {
3264 static int __net_init tcp4_proc_init_net(struct net *net)
3266 if (!proc_create_net_data("tcp", 0444, net->proc_net, &tcp4_seq_ops,
3267 sizeof(struct tcp_iter_state), &tcp4_seq_afinfo))
3272 static void __net_exit tcp4_proc_exit_net(struct net *net)
3274 remove_proc_entry("tcp", net->proc_net);
3277 static struct pernet_operations tcp4_net_ops = {
3278 .init = tcp4_proc_init_net,
3279 .exit = tcp4_proc_exit_net,
3282 int __init tcp4_proc_init(void)
3284 return register_pernet_subsys(&tcp4_net_ops);
3287 void tcp4_proc_exit(void)
3289 unregister_pernet_subsys(&tcp4_net_ops);
3291 #endif /* CONFIG_PROC_FS */
3293 /* @wake is one when sk_stream_write_space() calls us.
3294 * This sends EPOLLOUT only if notsent_bytes is half the limit.
3295 * This mimics the strategy used in sock_def_write_space().
3297 bool tcp_stream_memory_free(const struct sock *sk, int wake)
3299 const struct tcp_sock *tp = tcp_sk(sk);
3300 u32 notsent_bytes = READ_ONCE(tp->write_seq) -
3301 READ_ONCE(tp->snd_nxt);
3303 return (notsent_bytes << wake) < tcp_notsent_lowat(tp);
3305 EXPORT_SYMBOL(tcp_stream_memory_free);
3307 struct proto tcp_prot = {
3309 .owner = THIS_MODULE,
3311 .pre_connect = tcp_v4_pre_connect,
3312 .connect = tcp_v4_connect,
3313 .disconnect = tcp_disconnect,
3314 .accept = inet_csk_accept,
3316 .init = tcp_v4_init_sock,
3317 .destroy = tcp_v4_destroy_sock,
3318 .shutdown = tcp_shutdown,
3319 .setsockopt = tcp_setsockopt,
3320 .getsockopt = tcp_getsockopt,
3321 .bpf_bypass_getsockopt = tcp_bpf_bypass_getsockopt,
3322 .keepalive = tcp_set_keepalive,
3323 .recvmsg = tcp_recvmsg,
3324 .sendmsg = tcp_sendmsg,
3325 .splice_eof = tcp_splice_eof,
3326 .backlog_rcv = tcp_v4_do_rcv,
3327 .release_cb = tcp_release_cb,
3329 .unhash = inet_unhash,
3330 .get_port = inet_csk_get_port,
3331 .put_port = inet_put_port,
3332 #ifdef CONFIG_BPF_SYSCALL
3333 .psock_update_sk_prot = tcp_bpf_update_proto,
3335 .enter_memory_pressure = tcp_enter_memory_pressure,
3336 .leave_memory_pressure = tcp_leave_memory_pressure,
3337 .stream_memory_free = tcp_stream_memory_free,
3338 .sockets_allocated = &tcp_sockets_allocated,
3339 .orphan_count = &tcp_orphan_count,
3341 .memory_allocated = &tcp_memory_allocated,
3342 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3344 .memory_pressure = &tcp_memory_pressure,
3345 .sysctl_mem = sysctl_tcp_mem,
3346 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3347 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3348 .max_header = MAX_TCP_HEADER,
3349 .obj_size = sizeof(struct tcp_sock),
3350 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3351 .twsk_prot = &tcp_timewait_sock_ops,
3352 .rsk_prot = &tcp_request_sock_ops,
3354 .no_autobind = true,
3355 .diag_destroy = tcp_abort,
3357 EXPORT_SYMBOL(tcp_prot);
3359 static void __net_exit tcp_sk_exit(struct net *net)
3361 if (net->ipv4.tcp_congestion_control)
3362 bpf_module_put(net->ipv4.tcp_congestion_control,
3363 net->ipv4.tcp_congestion_control->owner);
3366 static void __net_init tcp_set_hashinfo(struct net *net)
3368 struct inet_hashinfo *hinfo;
3369 unsigned int ehash_entries;
3370 struct net *old_net;
3372 if (net_eq(net, &init_net))
3375 old_net = current->nsproxy->net_ns;
3376 ehash_entries = READ_ONCE(old_net->ipv4.sysctl_tcp_child_ehash_entries);
3380 ehash_entries = roundup_pow_of_two(ehash_entries);
3381 hinfo = inet_pernet_hashinfo_alloc(&tcp_hashinfo, ehash_entries);
3383 pr_warn("Failed to allocate TCP ehash (entries: %u) "
3384 "for a netns, fallback to the global one\n",
3387 hinfo = &tcp_hashinfo;
3388 ehash_entries = tcp_hashinfo.ehash_mask + 1;
3391 net->ipv4.tcp_death_row.hashinfo = hinfo;
3392 net->ipv4.tcp_death_row.sysctl_max_tw_buckets = ehash_entries / 2;
3393 net->ipv4.sysctl_max_syn_backlog = max(128U, ehash_entries / 128);
3396 static int __net_init tcp_sk_init(struct net *net)
3398 net->ipv4.sysctl_tcp_ecn = 2;
3399 net->ipv4.sysctl_tcp_ecn_fallback = 1;
3401 net->ipv4.sysctl_tcp_base_mss = TCP_BASE_MSS;
3402 net->ipv4.sysctl_tcp_min_snd_mss = TCP_MIN_SND_MSS;
3403 net->ipv4.sysctl_tcp_probe_threshold = TCP_PROBE_THRESHOLD;
3404 net->ipv4.sysctl_tcp_probe_interval = TCP_PROBE_INTERVAL;
3405 net->ipv4.sysctl_tcp_mtu_probe_floor = TCP_MIN_SND_MSS;
3407 net->ipv4.sysctl_tcp_keepalive_time = TCP_KEEPALIVE_TIME;
3408 net->ipv4.sysctl_tcp_keepalive_probes = TCP_KEEPALIVE_PROBES;
3409 net->ipv4.sysctl_tcp_keepalive_intvl = TCP_KEEPALIVE_INTVL;
3411 net->ipv4.sysctl_tcp_syn_retries = TCP_SYN_RETRIES;
3412 net->ipv4.sysctl_tcp_synack_retries = TCP_SYNACK_RETRIES;
3413 net->ipv4.sysctl_tcp_syncookies = 1;
3414 net->ipv4.sysctl_tcp_reordering = TCP_FASTRETRANS_THRESH;
3415 net->ipv4.sysctl_tcp_retries1 = TCP_RETR1;
3416 net->ipv4.sysctl_tcp_retries2 = TCP_RETR2;
3417 net->ipv4.sysctl_tcp_orphan_retries = 0;
3418 net->ipv4.sysctl_tcp_fin_timeout = TCP_FIN_TIMEOUT;
3419 net->ipv4.sysctl_tcp_notsent_lowat = UINT_MAX;
3420 net->ipv4.sysctl_tcp_tw_reuse = 2;
3421 net->ipv4.sysctl_tcp_no_ssthresh_metrics_save = 1;
3423 refcount_set(&net->ipv4.tcp_death_row.tw_refcount, 1);
3424 tcp_set_hashinfo(net);
3426 net->ipv4.sysctl_tcp_sack = 1;
3427 net->ipv4.sysctl_tcp_window_scaling = 1;
3428 net->ipv4.sysctl_tcp_timestamps = 1;
3429 net->ipv4.sysctl_tcp_early_retrans = 3;
3430 net->ipv4.sysctl_tcp_recovery = TCP_RACK_LOSS_DETECTION;
3431 net->ipv4.sysctl_tcp_slow_start_after_idle = 1; /* By default, RFC2861 behavior. */
3432 net->ipv4.sysctl_tcp_retrans_collapse = 1;
3433 net->ipv4.sysctl_tcp_max_reordering = 300;
3434 net->ipv4.sysctl_tcp_dsack = 1;
3435 net->ipv4.sysctl_tcp_app_win = 31;
3436 net->ipv4.sysctl_tcp_adv_win_scale = 1;
3437 net->ipv4.sysctl_tcp_frto = 2;
3438 net->ipv4.sysctl_tcp_moderate_rcvbuf = 1;
3439 /* This limits the percentage of the congestion window which we
3440 * will allow a single TSO frame to consume. Building TSO frames
3441 * which are too large can cause TCP streams to be bursty.
3443 net->ipv4.sysctl_tcp_tso_win_divisor = 3;
3444 /* Default TSQ limit of 16 TSO segments */
3445 net->ipv4.sysctl_tcp_limit_output_bytes = 16 * 65536;
3447 /* rfc5961 challenge ack rate limiting, per net-ns, disabled by default. */
3448 net->ipv4.sysctl_tcp_challenge_ack_limit = INT_MAX;
3450 net->ipv4.sysctl_tcp_min_tso_segs = 2;
3451 net->ipv4.sysctl_tcp_tso_rtt_log = 9; /* 2^9 = 512 usec */
3452 net->ipv4.sysctl_tcp_min_rtt_wlen = 300;
3453 net->ipv4.sysctl_tcp_autocorking = 1;
3454 net->ipv4.sysctl_tcp_invalid_ratelimit = HZ/2;
3455 net->ipv4.sysctl_tcp_pacing_ss_ratio = 200;
3456 net->ipv4.sysctl_tcp_pacing_ca_ratio = 120;
3457 if (net != &init_net) {
3458 memcpy(net->ipv4.sysctl_tcp_rmem,
3459 init_net.ipv4.sysctl_tcp_rmem,
3460 sizeof(init_net.ipv4.sysctl_tcp_rmem));
3461 memcpy(net->ipv4.sysctl_tcp_wmem,
3462 init_net.ipv4.sysctl_tcp_wmem,
3463 sizeof(init_net.ipv4.sysctl_tcp_wmem));
3465 net->ipv4.sysctl_tcp_comp_sack_delay_ns = NSEC_PER_MSEC;
3466 net->ipv4.sysctl_tcp_comp_sack_slack_ns = 100 * NSEC_PER_USEC;
3467 net->ipv4.sysctl_tcp_comp_sack_nr = 44;
3468 net->ipv4.sysctl_tcp_backlog_ack_defer = 1;
3469 net->ipv4.sysctl_tcp_fastopen = TFO_CLIENT_ENABLE;
3470 net->ipv4.sysctl_tcp_fastopen_blackhole_timeout = 0;
3471 atomic_set(&net->ipv4.tfo_active_disable_times, 0);
3473 /* Set default values for PLB */
3474 net->ipv4.sysctl_tcp_plb_enabled = 0; /* Disabled by default */
3475 net->ipv4.sysctl_tcp_plb_idle_rehash_rounds = 3;
3476 net->ipv4.sysctl_tcp_plb_rehash_rounds = 12;
3477 net->ipv4.sysctl_tcp_plb_suspend_rto_sec = 60;
3478 /* Default congestion threshold for PLB to mark a round is 50% */
3479 net->ipv4.sysctl_tcp_plb_cong_thresh = (1 << TCP_PLB_SCALE) / 2;
3481 /* Reno is always built in */
3482 if (!net_eq(net, &init_net) &&
3483 bpf_try_module_get(init_net.ipv4.tcp_congestion_control,
3484 init_net.ipv4.tcp_congestion_control->owner))
3485 net->ipv4.tcp_congestion_control = init_net.ipv4.tcp_congestion_control;
3487 net->ipv4.tcp_congestion_control = &tcp_reno;
3489 net->ipv4.sysctl_tcp_syn_linear_timeouts = 4;
3490 net->ipv4.sysctl_tcp_shrink_window = 0;
3492 net->ipv4.sysctl_tcp_pingpong_thresh = 1;
3497 static void __net_exit tcp_sk_exit_batch(struct list_head *net_exit_list)
3501 tcp_twsk_purge(net_exit_list, AF_INET);
3503 list_for_each_entry(net, net_exit_list, exit_list) {
3504 inet_pernet_hashinfo_free(net->ipv4.tcp_death_row.hashinfo);
3505 WARN_ON_ONCE(!refcount_dec_and_test(&net->ipv4.tcp_death_row.tw_refcount));
3506 tcp_fastopen_ctx_destroy(net);
3510 static struct pernet_operations __net_initdata tcp_sk_ops = {
3511 .init = tcp_sk_init,
3512 .exit = tcp_sk_exit,
3513 .exit_batch = tcp_sk_exit_batch,
3516 #if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_PROC_FS)
3517 DEFINE_BPF_ITER_FUNC(tcp, struct bpf_iter_meta *meta,
3518 struct sock_common *sk_common, uid_t uid)
3520 #define INIT_BATCH_SZ 16
3522 static int bpf_iter_init_tcp(void *priv_data, struct bpf_iter_aux_info *aux)
3524 struct bpf_tcp_iter_state *iter = priv_data;
3527 err = bpf_iter_init_seq_net(priv_data, aux);
3531 err = bpf_iter_tcp_realloc_batch(iter, INIT_BATCH_SZ);
3533 bpf_iter_fini_seq_net(priv_data);
3540 static void bpf_iter_fini_tcp(void *priv_data)
3542 struct bpf_tcp_iter_state *iter = priv_data;
3544 bpf_iter_fini_seq_net(priv_data);
3545 kvfree(iter->batch);
3548 static const struct bpf_iter_seq_info tcp_seq_info = {
3549 .seq_ops = &bpf_iter_tcp_seq_ops,
3550 .init_seq_private = bpf_iter_init_tcp,
3551 .fini_seq_private = bpf_iter_fini_tcp,
3552 .seq_priv_size = sizeof(struct bpf_tcp_iter_state),
3555 static const struct bpf_func_proto *
3556 bpf_iter_tcp_get_func_proto(enum bpf_func_id func_id,
3557 const struct bpf_prog *prog)
3560 case BPF_FUNC_setsockopt:
3561 return &bpf_sk_setsockopt_proto;
3562 case BPF_FUNC_getsockopt:
3563 return &bpf_sk_getsockopt_proto;
3569 static struct bpf_iter_reg tcp_reg_info = {
3571 .ctx_arg_info_size = 1,
3573 { offsetof(struct bpf_iter__tcp, sk_common),
3574 PTR_TO_BTF_ID_OR_NULL | PTR_TRUSTED },
3576 .get_func_proto = bpf_iter_tcp_get_func_proto,
3577 .seq_info = &tcp_seq_info,
3580 static void __init bpf_iter_register(void)
3582 tcp_reg_info.ctx_arg_info[0].btf_id = btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON];
3583 if (bpf_iter_reg_target(&tcp_reg_info))
3584 pr_warn("Warning: could not register bpf iterator tcp\n");
3589 void __init tcp_v4_init(void)
3593 for_each_possible_cpu(cpu) {
3596 res = inet_ctl_sock_create(&sk, PF_INET, SOCK_RAW,
3597 IPPROTO_TCP, &init_net);
3599 panic("Failed to create the TCP control socket.\n");
3600 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
3602 /* Please enforce IP_DF and IPID==0 for RST and
3603 * ACK sent in SYN-RECV and TIME-WAIT state.
3605 inet_sk(sk)->pmtudisc = IP_PMTUDISC_DO;
3607 per_cpu(ipv4_tcp_sk, cpu) = sk;
3609 if (register_pernet_subsys(&tcp_sk_ops))
3610 panic("Failed to create the TCP control socket.\n");
3612 #if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_PROC_FS)
3613 bpf_iter_register();