Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
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.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * IPv4 specific functions
10 *
11 * code split from:
12 * linux/ipv4/tcp.c
13 * linux/ipv4/tcp_input.c
14 * linux/ipv4/tcp_output.c
15 *
16 * See tcp.c for author information
17 */
18
19/*
20 * Changes:
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
29 * ACK bit.
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
40 * coma.
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.
46 */
47
48#define pr_fmt(fmt) "TCP: " fmt
49
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>
61
62#include <net/net_namespace.h>
63#include <net/icmp.h>
64#include <net/inet_hashtables.h>
65#include <net/tcp.h>
66#include <net/transp_v6.h>
67#include <net/ipv6.h>
68#include <net/inet_common.h>
69#include <net/timewait_sock.h>
70#include <net/xfrm.h>
71#include <net/secure_seq.h>
72#include <net/busy_poll.h>
73#include <net/rstreason.h>
74
75#include <linux/inet.h>
76#include <linux/ipv6.h>
77#include <linux/stddef.h>
78#include <linux/proc_fs.h>
79#include <linux/seq_file.h>
80#include <linux/inetdevice.h>
81#include <linux/btf_ids.h>
82#include <linux/skbuff_ref.h>
83
84#include <crypto/hash.h>
85#include <linux/scatterlist.h>
86
87#include <trace/events/tcp.h>
88
89#ifdef CONFIG_TCP_MD5SIG
90static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
91 __be32 daddr, __be32 saddr, const struct tcphdr *th);
92#endif
93
94struct inet_hashinfo tcp_hashinfo;
95EXPORT_SYMBOL(tcp_hashinfo);
96
97static DEFINE_PER_CPU(struct sock_bh_locked, ipv4_tcp_sk) = {
98 .bh_lock = INIT_LOCAL_LOCK(bh_lock),
99};
100
101static DEFINE_MUTEX(tcp_exit_batch_mutex);
102
103static u32 tcp_v4_init_seq(const struct sk_buff *skb)
104{
105 return secure_tcp_seq(ip_hdr(skb)->daddr,
106 ip_hdr(skb)->saddr,
107 tcp_hdr(skb)->dest,
108 tcp_hdr(skb)->source);
109}
110
111static u32 tcp_v4_init_ts_off(const struct net *net, const struct sk_buff *skb)
112{
113 return secure_tcp_ts_off(net, ip_hdr(skb)->daddr, ip_hdr(skb)->saddr);
114}
115
116int tcp_twsk_unique(struct sock *sk, struct sock *sktw, void *twp)
117{
118 int reuse = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tw_reuse);
119 const struct inet_timewait_sock *tw = inet_twsk(sktw);
120 const struct tcp_timewait_sock *tcptw = tcp_twsk(sktw);
121 struct tcp_sock *tp = tcp_sk(sk);
122 int ts_recent_stamp;
123
124 if (READ_ONCE(tw->tw_substate) == TCP_FIN_WAIT2)
125 reuse = 0;
126
127 if (reuse == 2) {
128 /* Still does not detect *everything* that goes through
129 * lo, since we require a loopback src or dst address
130 * or direct binding to 'lo' interface.
131 */
132 bool loopback = false;
133 if (tw->tw_bound_dev_if == LOOPBACK_IFINDEX)
134 loopback = true;
135#if IS_ENABLED(CONFIG_IPV6)
136 if (tw->tw_family == AF_INET6) {
137 if (ipv6_addr_loopback(&tw->tw_v6_daddr) ||
138 ipv6_addr_v4mapped_loopback(&tw->tw_v6_daddr) ||
139 ipv6_addr_loopback(&tw->tw_v6_rcv_saddr) ||
140 ipv6_addr_v4mapped_loopback(&tw->tw_v6_rcv_saddr))
141 loopback = true;
142 } else
143#endif
144 {
145 if (ipv4_is_loopback(tw->tw_daddr) ||
146 ipv4_is_loopback(tw->tw_rcv_saddr))
147 loopback = true;
148 }
149 if (!loopback)
150 reuse = 0;
151 }
152
153 /* With PAWS, it is safe from the viewpoint
154 of data integrity. Even without PAWS it is safe provided sequence
155 spaces do not overlap i.e. at data rates <= 80Mbit/sec.
156
157 Actually, the idea is close to VJ's one, only timestamp cache is
158 held not per host, but per port pair and TW bucket is used as state
159 holder.
160
161 If TW bucket has been already destroyed we fall back to VJ's scheme
162 and use initial timestamp retrieved from peer table.
163 */
164 ts_recent_stamp = READ_ONCE(tcptw->tw_ts_recent_stamp);
165 if (ts_recent_stamp &&
166 (!twp || (reuse && time_after32(ktime_get_seconds(),
167 ts_recent_stamp)))) {
168 /* inet_twsk_hashdance_schedule() sets sk_refcnt after putting twsk
169 * and releasing the bucket lock.
170 */
171 if (unlikely(!refcount_inc_not_zero(&sktw->sk_refcnt)))
172 return 0;
173
174 /* In case of repair and re-using TIME-WAIT sockets we still
175 * want to be sure that it is safe as above but honor the
176 * sequence numbers and time stamps set as part of the repair
177 * process.
178 *
179 * Without this check re-using a TIME-WAIT socket with TCP
180 * repair would accumulate a -1 on the repair assigned
181 * sequence number. The first time it is reused the sequence
182 * is -1, the second time -2, etc. This fixes that issue
183 * without appearing to create any others.
184 */
185 if (likely(!tp->repair)) {
186 u32 seq = tcptw->tw_snd_nxt + 65535 + 2;
187
188 if (!seq)
189 seq = 1;
190 WRITE_ONCE(tp->write_seq, seq);
191 tp->rx_opt.ts_recent = READ_ONCE(tcptw->tw_ts_recent);
192 tp->rx_opt.ts_recent_stamp = ts_recent_stamp;
193 }
194
195 return 1;
196 }
197
198 return 0;
199}
200EXPORT_SYMBOL_GPL(tcp_twsk_unique);
201
202static int tcp_v4_pre_connect(struct sock *sk, struct sockaddr *uaddr,
203 int addr_len)
204{
205 /* This check is replicated from tcp_v4_connect() and intended to
206 * prevent BPF program called below from accessing bytes that are out
207 * of the bound specified by user in addr_len.
208 */
209 if (addr_len < sizeof(struct sockaddr_in))
210 return -EINVAL;
211
212 sock_owned_by_me(sk);
213
214 return BPF_CGROUP_RUN_PROG_INET4_CONNECT(sk, uaddr, &addr_len);
215}
216
217/* This will initiate an outgoing connection. */
218int tcp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
219{
220 struct sockaddr_in *usin = (struct sockaddr_in *)uaddr;
221 struct inet_timewait_death_row *tcp_death_row;
222 struct inet_sock *inet = inet_sk(sk);
223 struct tcp_sock *tp = tcp_sk(sk);
224 struct ip_options_rcu *inet_opt;
225 struct net *net = sock_net(sk);
226 __be16 orig_sport, orig_dport;
227 __be32 daddr, nexthop;
228 struct flowi4 *fl4;
229 struct rtable *rt;
230 int err;
231
232 if (addr_len < sizeof(struct sockaddr_in))
233 return -EINVAL;
234
235 if (usin->sin_family != AF_INET)
236 return -EAFNOSUPPORT;
237
238 nexthop = daddr = usin->sin_addr.s_addr;
239 inet_opt = rcu_dereference_protected(inet->inet_opt,
240 lockdep_sock_is_held(sk));
241 if (inet_opt && inet_opt->opt.srr) {
242 if (!daddr)
243 return -EINVAL;
244 nexthop = inet_opt->opt.faddr;
245 }
246
247 orig_sport = inet->inet_sport;
248 orig_dport = usin->sin_port;
249 fl4 = &inet->cork.fl.u.ip4;
250 rt = ip_route_connect(fl4, nexthop, inet->inet_saddr,
251 sk->sk_bound_dev_if, IPPROTO_TCP, orig_sport,
252 orig_dport, sk);
253 if (IS_ERR(rt)) {
254 err = PTR_ERR(rt);
255 if (err == -ENETUNREACH)
256 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
257 return err;
258 }
259
260 if (rt->rt_flags & (RTCF_MULTICAST | RTCF_BROADCAST)) {
261 ip_rt_put(rt);
262 return -ENETUNREACH;
263 }
264
265 if (!inet_opt || !inet_opt->opt.srr)
266 daddr = fl4->daddr;
267
268 tcp_death_row = &sock_net(sk)->ipv4.tcp_death_row;
269
270 if (!inet->inet_saddr) {
271 err = inet_bhash2_update_saddr(sk, &fl4->saddr, AF_INET);
272 if (err) {
273 ip_rt_put(rt);
274 return err;
275 }
276 } else {
277 sk_rcv_saddr_set(sk, inet->inet_saddr);
278 }
279
280 if (tp->rx_opt.ts_recent_stamp && inet->inet_daddr != daddr) {
281 /* Reset inherited state */
282 tp->rx_opt.ts_recent = 0;
283 tp->rx_opt.ts_recent_stamp = 0;
284 if (likely(!tp->repair))
285 WRITE_ONCE(tp->write_seq, 0);
286 }
287
288 inet->inet_dport = usin->sin_port;
289 sk_daddr_set(sk, daddr);
290
291 inet_csk(sk)->icsk_ext_hdr_len = 0;
292 if (inet_opt)
293 inet_csk(sk)->icsk_ext_hdr_len = inet_opt->opt.optlen;
294
295 tp->rx_opt.mss_clamp = TCP_MSS_DEFAULT;
296
297 /* Socket identity is still unknown (sport may be zero).
298 * However we set state to SYN-SENT and not releasing socket
299 * lock select source port, enter ourselves into the hash tables and
300 * complete initialization after this.
301 */
302 tcp_set_state(sk, TCP_SYN_SENT);
303 err = inet_hash_connect(tcp_death_row, sk);
304 if (err)
305 goto failure;
306
307 sk_set_txhash(sk);
308
309 rt = ip_route_newports(fl4, rt, orig_sport, orig_dport,
310 inet->inet_sport, inet->inet_dport, sk);
311 if (IS_ERR(rt)) {
312 err = PTR_ERR(rt);
313 rt = NULL;
314 goto failure;
315 }
316 tp->tcp_usec_ts = dst_tcp_usec_ts(&rt->dst);
317 /* OK, now commit destination to socket. */
318 sk->sk_gso_type = SKB_GSO_TCPV4;
319 sk_setup_caps(sk, &rt->dst);
320 rt = NULL;
321
322 if (likely(!tp->repair)) {
323 if (!tp->write_seq)
324 WRITE_ONCE(tp->write_seq,
325 secure_tcp_seq(inet->inet_saddr,
326 inet->inet_daddr,
327 inet->inet_sport,
328 usin->sin_port));
329 WRITE_ONCE(tp->tsoffset,
330 secure_tcp_ts_off(net, inet->inet_saddr,
331 inet->inet_daddr));
332 }
333
334 atomic_set(&inet->inet_id, get_random_u16());
335
336 if (tcp_fastopen_defer_connect(sk, &err))
337 return err;
338 if (err)
339 goto failure;
340
341 err = tcp_connect(sk);
342
343 if (err)
344 goto failure;
345
346 return 0;
347
348failure:
349 /*
350 * This unhashes the socket and releases the local port,
351 * if necessary.
352 */
353 tcp_set_state(sk, TCP_CLOSE);
354 inet_bhash2_reset_saddr(sk);
355 ip_rt_put(rt);
356 sk->sk_route_caps = 0;
357 inet->inet_dport = 0;
358 return err;
359}
360EXPORT_SYMBOL(tcp_v4_connect);
361
362/*
363 * This routine reacts to ICMP_FRAG_NEEDED mtu indications as defined in RFC1191.
364 * It can be called through tcp_release_cb() if socket was owned by user
365 * at the time tcp_v4_err() was called to handle ICMP message.
366 */
367void tcp_v4_mtu_reduced(struct sock *sk)
368{
369 struct inet_sock *inet = inet_sk(sk);
370 struct dst_entry *dst;
371 u32 mtu;
372
373 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE))
374 return;
375 mtu = READ_ONCE(tcp_sk(sk)->mtu_info);
376 dst = inet_csk_update_pmtu(sk, mtu);
377 if (!dst)
378 return;
379
380 /* Something is about to be wrong... Remember soft error
381 * for the case, if this connection will not able to recover.
382 */
383 if (mtu < dst_mtu(dst) && ip_dont_fragment(sk, dst))
384 WRITE_ONCE(sk->sk_err_soft, EMSGSIZE);
385
386 mtu = dst_mtu(dst);
387
388 if (inet->pmtudisc != IP_PMTUDISC_DONT &&
389 ip_sk_accept_pmtu(sk) &&
390 inet_csk(sk)->icsk_pmtu_cookie > mtu) {
391 tcp_sync_mss(sk, mtu);
392
393 /* Resend the TCP packet because it's
394 * clear that the old packet has been
395 * dropped. This is the new "fast" path mtu
396 * discovery.
397 */
398 tcp_simple_retransmit(sk);
399 } /* else let the usual retransmit timer handle it */
400}
401EXPORT_SYMBOL(tcp_v4_mtu_reduced);
402
403static void do_redirect(struct sk_buff *skb, struct sock *sk)
404{
405 struct dst_entry *dst = __sk_dst_check(sk, 0);
406
407 if (dst)
408 dst->ops->redirect(dst, sk, skb);
409}
410
411
412/* handle ICMP messages on TCP_NEW_SYN_RECV request sockets */
413void tcp_req_err(struct sock *sk, u32 seq, bool abort)
414{
415 struct request_sock *req = inet_reqsk(sk);
416 struct net *net = sock_net(sk);
417
418 /* ICMPs are not backlogged, hence we cannot get
419 * an established socket here.
420 */
421 if (seq != tcp_rsk(req)->snt_isn) {
422 __NET_INC_STATS(net, LINUX_MIB_OUTOFWINDOWICMPS);
423 } else if (abort) {
424 /*
425 * Still in SYN_RECV, just remove it silently.
426 * There is no good way to pass the error to the newly
427 * created socket, and POSIX does not want network
428 * errors returned from accept().
429 */
430 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
431 tcp_listendrop(req->rsk_listener);
432 }
433 reqsk_put(req);
434}
435EXPORT_SYMBOL(tcp_req_err);
436
437/* TCP-LD (RFC 6069) logic */
438void tcp_ld_RTO_revert(struct sock *sk, u32 seq)
439{
440 struct inet_connection_sock *icsk = inet_csk(sk);
441 struct tcp_sock *tp = tcp_sk(sk);
442 struct sk_buff *skb;
443 s32 remaining;
444 u32 delta_us;
445
446 if (sock_owned_by_user(sk))
447 return;
448
449 if (seq != tp->snd_una || !icsk->icsk_retransmits ||
450 !icsk->icsk_backoff)
451 return;
452
453 skb = tcp_rtx_queue_head(sk);
454 if (WARN_ON_ONCE(!skb))
455 return;
456
457 icsk->icsk_backoff--;
458 icsk->icsk_rto = tp->srtt_us ? __tcp_set_rto(tp) : TCP_TIMEOUT_INIT;
459 icsk->icsk_rto = inet_csk_rto_backoff(icsk, TCP_RTO_MAX);
460
461 tcp_mstamp_refresh(tp);
462 delta_us = (u32)(tp->tcp_mstamp - tcp_skb_timestamp_us(skb));
463 remaining = icsk->icsk_rto - usecs_to_jiffies(delta_us);
464
465 if (remaining > 0) {
466 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
467 remaining, TCP_RTO_MAX);
468 } else {
469 /* RTO revert clocked out retransmission.
470 * Will retransmit now.
471 */
472 tcp_retransmit_timer(sk);
473 }
474}
475EXPORT_SYMBOL(tcp_ld_RTO_revert);
476
477/*
478 * This routine is called by the ICMP module when it gets some
479 * sort of error condition. If err < 0 then the socket should
480 * be closed and the error returned to the user. If err > 0
481 * it's just the icmp type << 8 | icmp code. After adjustment
482 * header points to the first 8 bytes of the tcp header. We need
483 * to find the appropriate port.
484 *
485 * The locking strategy used here is very "optimistic". When
486 * someone else accesses the socket the ICMP is just dropped
487 * and for some paths there is no check at all.
488 * A more general error queue to queue errors for later handling
489 * is probably better.
490 *
491 */
492
493int tcp_v4_err(struct sk_buff *skb, u32 info)
494{
495 const struct iphdr *iph = (const struct iphdr *)skb->data;
496 struct tcphdr *th = (struct tcphdr *)(skb->data + (iph->ihl << 2));
497 struct tcp_sock *tp;
498 const int type = icmp_hdr(skb)->type;
499 const int code = icmp_hdr(skb)->code;
500 struct sock *sk;
501 struct request_sock *fastopen;
502 u32 seq, snd_una;
503 int err;
504 struct net *net = dev_net(skb->dev);
505
506 sk = __inet_lookup_established(net, net->ipv4.tcp_death_row.hashinfo,
507 iph->daddr, th->dest, iph->saddr,
508 ntohs(th->source), inet_iif(skb), 0);
509 if (!sk) {
510 __ICMP_INC_STATS(net, ICMP_MIB_INERRORS);
511 return -ENOENT;
512 }
513 if (sk->sk_state == TCP_TIME_WAIT) {
514 /* To increase the counter of ignored icmps for TCP-AO */
515 tcp_ao_ignore_icmp(sk, AF_INET, type, code);
516 inet_twsk_put(inet_twsk(sk));
517 return 0;
518 }
519 seq = ntohl(th->seq);
520 if (sk->sk_state == TCP_NEW_SYN_RECV) {
521 tcp_req_err(sk, seq, type == ICMP_PARAMETERPROB ||
522 type == ICMP_TIME_EXCEEDED ||
523 (type == ICMP_DEST_UNREACH &&
524 (code == ICMP_NET_UNREACH ||
525 code == ICMP_HOST_UNREACH)));
526 return 0;
527 }
528
529 if (tcp_ao_ignore_icmp(sk, AF_INET, type, code)) {
530 sock_put(sk);
531 return 0;
532 }
533
534 bh_lock_sock(sk);
535 /* If too many ICMPs get dropped on busy
536 * servers this needs to be solved differently.
537 * We do take care of PMTU discovery (RFC1191) special case :
538 * we can receive locally generated ICMP messages while socket is held.
539 */
540 if (sock_owned_by_user(sk)) {
541 if (!(type == ICMP_DEST_UNREACH && code == ICMP_FRAG_NEEDED))
542 __NET_INC_STATS(net, LINUX_MIB_LOCKDROPPEDICMPS);
543 }
544 if (sk->sk_state == TCP_CLOSE)
545 goto out;
546
547 if (static_branch_unlikely(&ip4_min_ttl)) {
548 /* min_ttl can be changed concurrently from do_ip_setsockopt() */
549 if (unlikely(iph->ttl < READ_ONCE(inet_sk(sk)->min_ttl))) {
550 __NET_INC_STATS(net, LINUX_MIB_TCPMINTTLDROP);
551 goto out;
552 }
553 }
554
555 tp = tcp_sk(sk);
556 /* XXX (TFO) - tp->snd_una should be ISN (tcp_create_openreq_child() */
557 fastopen = rcu_dereference(tp->fastopen_rsk);
558 snd_una = fastopen ? tcp_rsk(fastopen)->snt_isn : tp->snd_una;
559 if (sk->sk_state != TCP_LISTEN &&
560 !between(seq, snd_una, tp->snd_nxt)) {
561 __NET_INC_STATS(net, LINUX_MIB_OUTOFWINDOWICMPS);
562 goto out;
563 }
564
565 switch (type) {
566 case ICMP_REDIRECT:
567 if (!sock_owned_by_user(sk))
568 do_redirect(skb, sk);
569 goto out;
570 case ICMP_SOURCE_QUENCH:
571 /* Just silently ignore these. */
572 goto out;
573 case ICMP_PARAMETERPROB:
574 err = EPROTO;
575 break;
576 case ICMP_DEST_UNREACH:
577 if (code > NR_ICMP_UNREACH)
578 goto out;
579
580 if (code == ICMP_FRAG_NEEDED) { /* PMTU discovery (RFC1191) */
581 /* We are not interested in TCP_LISTEN and open_requests
582 * (SYN-ACKs send out by Linux are always <576bytes so
583 * they should go through unfragmented).
584 */
585 if (sk->sk_state == TCP_LISTEN)
586 goto out;
587
588 WRITE_ONCE(tp->mtu_info, info);
589 if (!sock_owned_by_user(sk)) {
590 tcp_v4_mtu_reduced(sk);
591 } else {
592 if (!test_and_set_bit(TCP_MTU_REDUCED_DEFERRED, &sk->sk_tsq_flags))
593 sock_hold(sk);
594 }
595 goto out;
596 }
597
598 err = icmp_err_convert[code].errno;
599 /* check if this ICMP message allows revert of backoff.
600 * (see RFC 6069)
601 */
602 if (!fastopen &&
603 (code == ICMP_NET_UNREACH || code == ICMP_HOST_UNREACH))
604 tcp_ld_RTO_revert(sk, seq);
605 break;
606 case ICMP_TIME_EXCEEDED:
607 err = EHOSTUNREACH;
608 break;
609 default:
610 goto out;
611 }
612
613 switch (sk->sk_state) {
614 case TCP_SYN_SENT:
615 case TCP_SYN_RECV:
616 /* Only in fast or simultaneous open. If a fast open socket is
617 * already accepted it is treated as a connected one below.
618 */
619 if (fastopen && !fastopen->sk)
620 break;
621
622 ip_icmp_error(sk, skb, err, th->dest, info, (u8 *)th);
623
624 if (!sock_owned_by_user(sk))
625 tcp_done_with_error(sk, err);
626 else
627 WRITE_ONCE(sk->sk_err_soft, err);
628 goto out;
629 }
630
631 /* If we've already connected we will keep trying
632 * until we time out, or the user gives up.
633 *
634 * rfc1122 4.2.3.9 allows to consider as hard errors
635 * only PROTO_UNREACH and PORT_UNREACH (well, FRAG_FAILED too,
636 * but it is obsoleted by pmtu discovery).
637 *
638 * Note, that in modern internet, where routing is unreliable
639 * and in each dark corner broken firewalls sit, sending random
640 * errors ordered by their masters even this two messages finally lose
641 * their original sense (even Linux sends invalid PORT_UNREACHs)
642 *
643 * Now we are in compliance with RFCs.
644 * --ANK (980905)
645 */
646
647 if (!sock_owned_by_user(sk) &&
648 inet_test_bit(RECVERR, sk)) {
649 WRITE_ONCE(sk->sk_err, err);
650 sk_error_report(sk);
651 } else { /* Only an error on timeout */
652 WRITE_ONCE(sk->sk_err_soft, err);
653 }
654
655out:
656 bh_unlock_sock(sk);
657 sock_put(sk);
658 return 0;
659}
660
661void __tcp_v4_send_check(struct sk_buff *skb, __be32 saddr, __be32 daddr)
662{
663 struct tcphdr *th = tcp_hdr(skb);
664
665 th->check = ~tcp_v4_check(skb->len, saddr, daddr, 0);
666 skb->csum_start = skb_transport_header(skb) - skb->head;
667 skb->csum_offset = offsetof(struct tcphdr, check);
668}
669
670/* This routine computes an IPv4 TCP checksum. */
671void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)
672{
673 const struct inet_sock *inet = inet_sk(sk);
674
675 __tcp_v4_send_check(skb, inet->inet_saddr, inet->inet_daddr);
676}
677EXPORT_SYMBOL(tcp_v4_send_check);
678
679#define REPLY_OPTIONS_LEN (MAX_TCP_OPTION_SPACE / sizeof(__be32))
680
681static bool tcp_v4_ao_sign_reset(const struct sock *sk, struct sk_buff *skb,
682 const struct tcp_ao_hdr *aoh,
683 struct ip_reply_arg *arg, struct tcphdr *reply,
684 __be32 reply_options[REPLY_OPTIONS_LEN])
685{
686#ifdef CONFIG_TCP_AO
687 int sdif = tcp_v4_sdif(skb);
688 int dif = inet_iif(skb);
689 int l3index = sdif ? dif : 0;
690 bool allocated_traffic_key;
691 struct tcp_ao_key *key;
692 char *traffic_key;
693 bool drop = true;
694 u32 ao_sne = 0;
695 u8 keyid;
696
697 rcu_read_lock();
698 if (tcp_ao_prepare_reset(sk, skb, aoh, l3index, ntohl(reply->seq),
699 &key, &traffic_key, &allocated_traffic_key,
700 &keyid, &ao_sne))
701 goto out;
702
703 reply_options[0] = htonl((TCPOPT_AO << 24) | (tcp_ao_len(key) << 16) |
704 (aoh->rnext_keyid << 8) | keyid);
705 arg->iov[0].iov_len += tcp_ao_len_aligned(key);
706 reply->doff = arg->iov[0].iov_len / 4;
707
708 if (tcp_ao_hash_hdr(AF_INET, (char *)&reply_options[1],
709 key, traffic_key,
710 (union tcp_ao_addr *)&ip_hdr(skb)->saddr,
711 (union tcp_ao_addr *)&ip_hdr(skb)->daddr,
712 reply, ao_sne))
713 goto out;
714 drop = false;
715out:
716 rcu_read_unlock();
717 if (allocated_traffic_key)
718 kfree(traffic_key);
719 return drop;
720#else
721 return true;
722#endif
723}
724
725/*
726 * This routine will send an RST to the other tcp.
727 *
728 * Someone asks: why I NEVER use socket parameters (TOS, TTL etc.)
729 * for reset.
730 * Answer: if a packet caused RST, it is not for a socket
731 * existing in our system, if it is matched to a socket,
732 * it is just duplicate segment or bug in other side's TCP.
733 * So that we build reply only basing on parameters
734 * arrived with segment.
735 * Exception: precedence violation. We do not implement it in any case.
736 */
737
738static void tcp_v4_send_reset(const struct sock *sk, struct sk_buff *skb,
739 enum sk_rst_reason reason)
740{
741 const struct tcphdr *th = tcp_hdr(skb);
742 struct {
743 struct tcphdr th;
744 __be32 opt[REPLY_OPTIONS_LEN];
745 } rep;
746 const __u8 *md5_hash_location = NULL;
747 const struct tcp_ao_hdr *aoh;
748 struct ip_reply_arg arg;
749#ifdef CONFIG_TCP_MD5SIG
750 struct tcp_md5sig_key *key = NULL;
751 unsigned char newhash[16];
752 struct sock *sk1 = NULL;
753 int genhash;
754#endif
755 u64 transmit_time = 0;
756 struct sock *ctl_sk;
757 struct net *net;
758 u32 txhash = 0;
759
760 /* Never send a reset in response to a reset. */
761 if (th->rst)
762 return;
763
764 /* If sk not NULL, it means we did a successful lookup and incoming
765 * route had to be correct. prequeue might have dropped our dst.
766 */
767 if (!sk && skb_rtable(skb)->rt_type != RTN_LOCAL)
768 return;
769
770 /* Swap the send and the receive. */
771 memset(&rep, 0, sizeof(rep));
772 rep.th.dest = th->source;
773 rep.th.source = th->dest;
774 rep.th.doff = sizeof(struct tcphdr) / 4;
775 rep.th.rst = 1;
776
777 if (th->ack) {
778 rep.th.seq = th->ack_seq;
779 } else {
780 rep.th.ack = 1;
781 rep.th.ack_seq = htonl(ntohl(th->seq) + th->syn + th->fin +
782 skb->len - (th->doff << 2));
783 }
784
785 memset(&arg, 0, sizeof(arg));
786 arg.iov[0].iov_base = (unsigned char *)&rep;
787 arg.iov[0].iov_len = sizeof(rep.th);
788
789 net = sk ? sock_net(sk) : dev_net(skb_dst(skb)->dev);
790
791 /* Invalid TCP option size or twice included auth */
792 if (tcp_parse_auth_options(tcp_hdr(skb), &md5_hash_location, &aoh))
793 return;
794
795 if (aoh && tcp_v4_ao_sign_reset(sk, skb, aoh, &arg, &rep.th, rep.opt))
796 return;
797
798#ifdef CONFIG_TCP_MD5SIG
799 rcu_read_lock();
800 if (sk && sk_fullsock(sk)) {
801 const union tcp_md5_addr *addr;
802 int l3index;
803
804 /* sdif set, means packet ingressed via a device
805 * in an L3 domain and inet_iif is set to it.
806 */
807 l3index = tcp_v4_sdif(skb) ? inet_iif(skb) : 0;
808 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
809 key = tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
810 } else if (md5_hash_location) {
811 const union tcp_md5_addr *addr;
812 int sdif = tcp_v4_sdif(skb);
813 int dif = inet_iif(skb);
814 int l3index;
815
816 /*
817 * active side is lost. Try to find listening socket through
818 * source port, and then find md5 key through listening socket.
819 * we are not loose security here:
820 * Incoming packet is checked with md5 hash with finding key,
821 * no RST generated if md5 hash doesn't match.
822 */
823 sk1 = __inet_lookup_listener(net, net->ipv4.tcp_death_row.hashinfo,
824 NULL, 0, ip_hdr(skb)->saddr,
825 th->source, ip_hdr(skb)->daddr,
826 ntohs(th->source), dif, sdif);
827 /* don't send rst if it can't find key */
828 if (!sk1)
829 goto out;
830
831 /* sdif set, means packet ingressed via a device
832 * in an L3 domain and dif is set to it.
833 */
834 l3index = sdif ? dif : 0;
835 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
836 key = tcp_md5_do_lookup(sk1, l3index, addr, AF_INET);
837 if (!key)
838 goto out;
839
840
841 genhash = tcp_v4_md5_hash_skb(newhash, key, NULL, skb);
842 if (genhash || memcmp(md5_hash_location, newhash, 16) != 0)
843 goto out;
844
845 }
846
847 if (key) {
848 rep.opt[0] = htonl((TCPOPT_NOP << 24) |
849 (TCPOPT_NOP << 16) |
850 (TCPOPT_MD5SIG << 8) |
851 TCPOLEN_MD5SIG);
852 /* Update length and the length the header thinks exists */
853 arg.iov[0].iov_len += TCPOLEN_MD5SIG_ALIGNED;
854 rep.th.doff = arg.iov[0].iov_len / 4;
855
856 tcp_v4_md5_hash_hdr((__u8 *) &rep.opt[1],
857 key, ip_hdr(skb)->saddr,
858 ip_hdr(skb)->daddr, &rep.th);
859 }
860#endif
861 /* Can't co-exist with TCPMD5, hence check rep.opt[0] */
862 if (rep.opt[0] == 0) {
863 __be32 mrst = mptcp_reset_option(skb);
864
865 if (mrst) {
866 rep.opt[0] = mrst;
867 arg.iov[0].iov_len += sizeof(mrst);
868 rep.th.doff = arg.iov[0].iov_len / 4;
869 }
870 }
871
872 arg.csum = csum_tcpudp_nofold(ip_hdr(skb)->daddr,
873 ip_hdr(skb)->saddr, /* XXX */
874 arg.iov[0].iov_len, IPPROTO_TCP, 0);
875 arg.csumoffset = offsetof(struct tcphdr, check) / 2;
876 arg.flags = (sk && inet_sk_transparent(sk)) ? IP_REPLY_ARG_NOSRCCHECK : 0;
877
878 /* When socket is gone, all binding information is lost.
879 * routing might fail in this case. No choice here, if we choose to force
880 * input interface, we will misroute in case of asymmetric route.
881 */
882 if (sk)
883 arg.bound_dev_if = sk->sk_bound_dev_if;
884
885 trace_tcp_send_reset(sk, skb, reason);
886
887 BUILD_BUG_ON(offsetof(struct sock, sk_bound_dev_if) !=
888 offsetof(struct inet_timewait_sock, tw_bound_dev_if));
889
890 arg.tos = ip_hdr(skb)->tos;
891 arg.uid = sock_net_uid(net, sk && sk_fullsock(sk) ? sk : NULL);
892 local_bh_disable();
893 local_lock_nested_bh(&ipv4_tcp_sk.bh_lock);
894 ctl_sk = this_cpu_read(ipv4_tcp_sk.sock);
895
896 sock_net_set(ctl_sk, net);
897 if (sk) {
898 ctl_sk->sk_mark = (sk->sk_state == TCP_TIME_WAIT) ?
899 inet_twsk(sk)->tw_mark : READ_ONCE(sk->sk_mark);
900 ctl_sk->sk_priority = (sk->sk_state == TCP_TIME_WAIT) ?
901 inet_twsk(sk)->tw_priority : READ_ONCE(sk->sk_priority);
902 transmit_time = tcp_transmit_time(sk);
903 xfrm_sk_clone_policy(ctl_sk, sk);
904 txhash = (sk->sk_state == TCP_TIME_WAIT) ?
905 inet_twsk(sk)->tw_txhash : sk->sk_txhash;
906 } else {
907 ctl_sk->sk_mark = 0;
908 ctl_sk->sk_priority = 0;
909 }
910 ip_send_unicast_reply(ctl_sk, sk,
911 skb, &TCP_SKB_CB(skb)->header.h4.opt,
912 ip_hdr(skb)->saddr, ip_hdr(skb)->daddr,
913 &arg, arg.iov[0].iov_len,
914 transmit_time, txhash);
915
916 xfrm_sk_free_policy(ctl_sk);
917 sock_net_set(ctl_sk, &init_net);
918 __TCP_INC_STATS(net, TCP_MIB_OUTSEGS);
919 __TCP_INC_STATS(net, TCP_MIB_OUTRSTS);
920 local_unlock_nested_bh(&ipv4_tcp_sk.bh_lock);
921 local_bh_enable();
922
923#ifdef CONFIG_TCP_MD5SIG
924out:
925 rcu_read_unlock();
926#endif
927}
928
929/* The code following below sending ACKs in SYN-RECV and TIME-WAIT states
930 outside socket context is ugly, certainly. What can I do?
931 */
932
933static void tcp_v4_send_ack(const struct sock *sk,
934 struct sk_buff *skb, u32 seq, u32 ack,
935 u32 win, u32 tsval, u32 tsecr, int oif,
936 struct tcp_key *key,
937 int reply_flags, u8 tos, u32 txhash)
938{
939 const struct tcphdr *th = tcp_hdr(skb);
940 struct {
941 struct tcphdr th;
942 __be32 opt[(MAX_TCP_OPTION_SPACE >> 2)];
943 } rep;
944 struct net *net = sock_net(sk);
945 struct ip_reply_arg arg;
946 struct sock *ctl_sk;
947 u64 transmit_time;
948
949 memset(&rep.th, 0, sizeof(struct tcphdr));
950 memset(&arg, 0, sizeof(arg));
951
952 arg.iov[0].iov_base = (unsigned char *)&rep;
953 arg.iov[0].iov_len = sizeof(rep.th);
954 if (tsecr) {
955 rep.opt[0] = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
956 (TCPOPT_TIMESTAMP << 8) |
957 TCPOLEN_TIMESTAMP);
958 rep.opt[1] = htonl(tsval);
959 rep.opt[2] = htonl(tsecr);
960 arg.iov[0].iov_len += TCPOLEN_TSTAMP_ALIGNED;
961 }
962
963 /* Swap the send and the receive. */
964 rep.th.dest = th->source;
965 rep.th.source = th->dest;
966 rep.th.doff = arg.iov[0].iov_len / 4;
967 rep.th.seq = htonl(seq);
968 rep.th.ack_seq = htonl(ack);
969 rep.th.ack = 1;
970 rep.th.window = htons(win);
971
972#ifdef CONFIG_TCP_MD5SIG
973 if (tcp_key_is_md5(key)) {
974 int offset = (tsecr) ? 3 : 0;
975
976 rep.opt[offset++] = htonl((TCPOPT_NOP << 24) |
977 (TCPOPT_NOP << 16) |
978 (TCPOPT_MD5SIG << 8) |
979 TCPOLEN_MD5SIG);
980 arg.iov[0].iov_len += TCPOLEN_MD5SIG_ALIGNED;
981 rep.th.doff = arg.iov[0].iov_len/4;
982
983 tcp_v4_md5_hash_hdr((__u8 *) &rep.opt[offset],
984 key->md5_key, ip_hdr(skb)->saddr,
985 ip_hdr(skb)->daddr, &rep.th);
986 }
987#endif
988#ifdef CONFIG_TCP_AO
989 if (tcp_key_is_ao(key)) {
990 int offset = (tsecr) ? 3 : 0;
991
992 rep.opt[offset++] = htonl((TCPOPT_AO << 24) |
993 (tcp_ao_len(key->ao_key) << 16) |
994 (key->ao_key->sndid << 8) |
995 key->rcv_next);
996 arg.iov[0].iov_len += tcp_ao_len_aligned(key->ao_key);
997 rep.th.doff = arg.iov[0].iov_len / 4;
998
999 tcp_ao_hash_hdr(AF_INET, (char *)&rep.opt[offset],
1000 key->ao_key, key->traffic_key,
1001 (union tcp_ao_addr *)&ip_hdr(skb)->saddr,
1002 (union tcp_ao_addr *)&ip_hdr(skb)->daddr,
1003 &rep.th, key->sne);
1004 }
1005#endif
1006 arg.flags = reply_flags;
1007 arg.csum = csum_tcpudp_nofold(ip_hdr(skb)->daddr,
1008 ip_hdr(skb)->saddr, /* XXX */
1009 arg.iov[0].iov_len, IPPROTO_TCP, 0);
1010 arg.csumoffset = offsetof(struct tcphdr, check) / 2;
1011 if (oif)
1012 arg.bound_dev_if = oif;
1013 arg.tos = tos;
1014 arg.uid = sock_net_uid(net, sk_fullsock(sk) ? sk : NULL);
1015 local_bh_disable();
1016 local_lock_nested_bh(&ipv4_tcp_sk.bh_lock);
1017 ctl_sk = this_cpu_read(ipv4_tcp_sk.sock);
1018 sock_net_set(ctl_sk, net);
1019 ctl_sk->sk_mark = (sk->sk_state == TCP_TIME_WAIT) ?
1020 inet_twsk(sk)->tw_mark : READ_ONCE(sk->sk_mark);
1021 ctl_sk->sk_priority = (sk->sk_state == TCP_TIME_WAIT) ?
1022 inet_twsk(sk)->tw_priority : READ_ONCE(sk->sk_priority);
1023 transmit_time = tcp_transmit_time(sk);
1024 ip_send_unicast_reply(ctl_sk, sk,
1025 skb, &TCP_SKB_CB(skb)->header.h4.opt,
1026 ip_hdr(skb)->saddr, ip_hdr(skb)->daddr,
1027 &arg, arg.iov[0].iov_len,
1028 transmit_time, txhash);
1029
1030 sock_net_set(ctl_sk, &init_net);
1031 __TCP_INC_STATS(net, TCP_MIB_OUTSEGS);
1032 local_unlock_nested_bh(&ipv4_tcp_sk.bh_lock);
1033 local_bh_enable();
1034}
1035
1036static void tcp_v4_timewait_ack(struct sock *sk, struct sk_buff *skb)
1037{
1038 struct inet_timewait_sock *tw = inet_twsk(sk);
1039 struct tcp_timewait_sock *tcptw = tcp_twsk(sk);
1040 struct tcp_key key = {};
1041#ifdef CONFIG_TCP_AO
1042 struct tcp_ao_info *ao_info;
1043
1044 if (static_branch_unlikely(&tcp_ao_needed.key)) {
1045 /* FIXME: the segment to-be-acked is not verified yet */
1046 ao_info = rcu_dereference(tcptw->ao_info);
1047 if (ao_info) {
1048 const struct tcp_ao_hdr *aoh;
1049
1050 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh)) {
1051 inet_twsk_put(tw);
1052 return;
1053 }
1054
1055 if (aoh)
1056 key.ao_key = tcp_ao_established_key(sk, ao_info,
1057 aoh->rnext_keyid, -1);
1058 }
1059 }
1060 if (key.ao_key) {
1061 struct tcp_ao_key *rnext_key;
1062
1063 key.traffic_key = snd_other_key(key.ao_key);
1064 key.sne = READ_ONCE(ao_info->snd_sne);
1065 rnext_key = READ_ONCE(ao_info->rnext_key);
1066 key.rcv_next = rnext_key->rcvid;
1067 key.type = TCP_KEY_AO;
1068#else
1069 if (0) {
1070#endif
1071 } else if (static_branch_tcp_md5()) {
1072 key.md5_key = tcp_twsk_md5_key(tcptw);
1073 if (key.md5_key)
1074 key.type = TCP_KEY_MD5;
1075 }
1076
1077 tcp_v4_send_ack(sk, skb,
1078 tcptw->tw_snd_nxt, READ_ONCE(tcptw->tw_rcv_nxt),
1079 tcptw->tw_rcv_wnd >> tw->tw_rcv_wscale,
1080 tcp_tw_tsval(tcptw),
1081 READ_ONCE(tcptw->tw_ts_recent),
1082 tw->tw_bound_dev_if, &key,
1083 tw->tw_transparent ? IP_REPLY_ARG_NOSRCCHECK : 0,
1084 tw->tw_tos,
1085 tw->tw_txhash);
1086
1087 inet_twsk_put(tw);
1088}
1089
1090static void tcp_v4_reqsk_send_ack(const struct sock *sk, struct sk_buff *skb,
1091 struct request_sock *req)
1092{
1093 struct tcp_key key = {};
1094
1095 /* sk->sk_state == TCP_LISTEN -> for regular TCP_SYN_RECV
1096 * sk->sk_state == TCP_SYN_RECV -> for Fast Open.
1097 */
1098 u32 seq = (sk->sk_state == TCP_LISTEN) ? tcp_rsk(req)->snt_isn + 1 :
1099 tcp_sk(sk)->snd_nxt;
1100
1101#ifdef CONFIG_TCP_AO
1102 if (static_branch_unlikely(&tcp_ao_needed.key) &&
1103 tcp_rsk_used_ao(req)) {
1104 const union tcp_md5_addr *addr;
1105 const struct tcp_ao_hdr *aoh;
1106 int l3index;
1107
1108 /* Invalid TCP option size or twice included auth */
1109 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh))
1110 return;
1111 if (!aoh)
1112 return;
1113
1114 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
1115 l3index = tcp_v4_sdif(skb) ? inet_iif(skb) : 0;
1116 key.ao_key = tcp_ao_do_lookup(sk, l3index, addr, AF_INET,
1117 aoh->rnext_keyid, -1);
1118 if (unlikely(!key.ao_key)) {
1119 /* Send ACK with any matching MKT for the peer */
1120 key.ao_key = tcp_ao_do_lookup(sk, l3index, addr, AF_INET, -1, -1);
1121 /* Matching key disappeared (user removed the key?)
1122 * let the handshake timeout.
1123 */
1124 if (!key.ao_key) {
1125 net_info_ratelimited("TCP-AO key for (%pI4, %d)->(%pI4, %d) suddenly disappeared, won't ACK new connection\n",
1126 addr,
1127 ntohs(tcp_hdr(skb)->source),
1128 &ip_hdr(skb)->daddr,
1129 ntohs(tcp_hdr(skb)->dest));
1130 return;
1131 }
1132 }
1133 key.traffic_key = kmalloc(tcp_ao_digest_size(key.ao_key), GFP_ATOMIC);
1134 if (!key.traffic_key)
1135 return;
1136
1137 key.type = TCP_KEY_AO;
1138 key.rcv_next = aoh->keyid;
1139 tcp_v4_ao_calc_key_rsk(key.ao_key, key.traffic_key, req);
1140#else
1141 if (0) {
1142#endif
1143 } else if (static_branch_tcp_md5()) {
1144 const union tcp_md5_addr *addr;
1145 int l3index;
1146
1147 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
1148 l3index = tcp_v4_sdif(skb) ? inet_iif(skb) : 0;
1149 key.md5_key = tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
1150 if (key.md5_key)
1151 key.type = TCP_KEY_MD5;
1152 }
1153
1154 tcp_v4_send_ack(sk, skb, seq,
1155 tcp_rsk(req)->rcv_nxt,
1156 tcp_synack_window(req) >> inet_rsk(req)->rcv_wscale,
1157 tcp_rsk_tsval(tcp_rsk(req)),
1158 READ_ONCE(req->ts_recent),
1159 0, &key,
1160 inet_rsk(req)->no_srccheck ? IP_REPLY_ARG_NOSRCCHECK : 0,
1161 ip_hdr(skb)->tos,
1162 READ_ONCE(tcp_rsk(req)->txhash));
1163 if (tcp_key_is_ao(&key))
1164 kfree(key.traffic_key);
1165}
1166
1167/*
1168 * Send a SYN-ACK after having received a SYN.
1169 * This still operates on a request_sock only, not on a big
1170 * socket.
1171 */
1172static int tcp_v4_send_synack(const struct sock *sk, struct dst_entry *dst,
1173 struct flowi *fl,
1174 struct request_sock *req,
1175 struct tcp_fastopen_cookie *foc,
1176 enum tcp_synack_type synack_type,
1177 struct sk_buff *syn_skb)
1178{
1179 const struct inet_request_sock *ireq = inet_rsk(req);
1180 struct flowi4 fl4;
1181 int err = -1;
1182 struct sk_buff *skb;
1183 u8 tos;
1184
1185 /* First, grab a route. */
1186 if (!dst && (dst = inet_csk_route_req(sk, &fl4, req)) == NULL)
1187 return -1;
1188
1189 skb = tcp_make_synack(sk, dst, req, foc, synack_type, syn_skb);
1190
1191 if (skb) {
1192 __tcp_v4_send_check(skb, ireq->ir_loc_addr, ireq->ir_rmt_addr);
1193
1194 tos = READ_ONCE(inet_sk(sk)->tos);
1195
1196 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reflect_tos))
1197 tos = (tcp_rsk(req)->syn_tos & ~INET_ECN_MASK) |
1198 (tos & INET_ECN_MASK);
1199
1200 if (!INET_ECN_is_capable(tos) &&
1201 tcp_bpf_ca_needs_ecn((struct sock *)req))
1202 tos |= INET_ECN_ECT_0;
1203
1204 rcu_read_lock();
1205 err = ip_build_and_send_pkt(skb, sk, ireq->ir_loc_addr,
1206 ireq->ir_rmt_addr,
1207 rcu_dereference(ireq->ireq_opt),
1208 tos);
1209 rcu_read_unlock();
1210 err = net_xmit_eval(err);
1211 }
1212
1213 return err;
1214}
1215
1216/*
1217 * IPv4 request_sock destructor.
1218 */
1219static void tcp_v4_reqsk_destructor(struct request_sock *req)
1220{
1221 kfree(rcu_dereference_protected(inet_rsk(req)->ireq_opt, 1));
1222}
1223
1224#ifdef CONFIG_TCP_MD5SIG
1225/*
1226 * RFC2385 MD5 checksumming requires a mapping of
1227 * IP address->MD5 Key.
1228 * We need to maintain these in the sk structure.
1229 */
1230
1231DEFINE_STATIC_KEY_DEFERRED_FALSE(tcp_md5_needed, HZ);
1232EXPORT_SYMBOL(tcp_md5_needed);
1233
1234static bool better_md5_match(struct tcp_md5sig_key *old, struct tcp_md5sig_key *new)
1235{
1236 if (!old)
1237 return true;
1238
1239 /* l3index always overrides non-l3index */
1240 if (old->l3index && new->l3index == 0)
1241 return false;
1242 if (old->l3index == 0 && new->l3index)
1243 return true;
1244
1245 return old->prefixlen < new->prefixlen;
1246}
1247
1248/* Find the Key structure for an address. */
1249struct tcp_md5sig_key *__tcp_md5_do_lookup(const struct sock *sk, int l3index,
1250 const union tcp_md5_addr *addr,
1251 int family, bool any_l3index)
1252{
1253 const struct tcp_sock *tp = tcp_sk(sk);
1254 struct tcp_md5sig_key *key;
1255 const struct tcp_md5sig_info *md5sig;
1256 __be32 mask;
1257 struct tcp_md5sig_key *best_match = NULL;
1258 bool match;
1259
1260 /* caller either holds rcu_read_lock() or socket lock */
1261 md5sig = rcu_dereference_check(tp->md5sig_info,
1262 lockdep_sock_is_held(sk));
1263 if (!md5sig)
1264 return NULL;
1265
1266 hlist_for_each_entry_rcu(key, &md5sig->head, node,
1267 lockdep_sock_is_held(sk)) {
1268 if (key->family != family)
1269 continue;
1270 if (!any_l3index && key->flags & TCP_MD5SIG_FLAG_IFINDEX &&
1271 key->l3index != l3index)
1272 continue;
1273 if (family == AF_INET) {
1274 mask = inet_make_mask(key->prefixlen);
1275 match = (key->addr.a4.s_addr & mask) ==
1276 (addr->a4.s_addr & mask);
1277#if IS_ENABLED(CONFIG_IPV6)
1278 } else if (family == AF_INET6) {
1279 match = ipv6_prefix_equal(&key->addr.a6, &addr->a6,
1280 key->prefixlen);
1281#endif
1282 } else {
1283 match = false;
1284 }
1285
1286 if (match && better_md5_match(best_match, key))
1287 best_match = key;
1288 }
1289 return best_match;
1290}
1291EXPORT_SYMBOL(__tcp_md5_do_lookup);
1292
1293static struct tcp_md5sig_key *tcp_md5_do_lookup_exact(const struct sock *sk,
1294 const union tcp_md5_addr *addr,
1295 int family, u8 prefixlen,
1296 int l3index, u8 flags)
1297{
1298 const struct tcp_sock *tp = tcp_sk(sk);
1299 struct tcp_md5sig_key *key;
1300 unsigned int size = sizeof(struct in_addr);
1301 const struct tcp_md5sig_info *md5sig;
1302
1303 /* caller either holds rcu_read_lock() or socket lock */
1304 md5sig = rcu_dereference_check(tp->md5sig_info,
1305 lockdep_sock_is_held(sk));
1306 if (!md5sig)
1307 return NULL;
1308#if IS_ENABLED(CONFIG_IPV6)
1309 if (family == AF_INET6)
1310 size = sizeof(struct in6_addr);
1311#endif
1312 hlist_for_each_entry_rcu(key, &md5sig->head, node,
1313 lockdep_sock_is_held(sk)) {
1314 if (key->family != family)
1315 continue;
1316 if ((key->flags & TCP_MD5SIG_FLAG_IFINDEX) != (flags & TCP_MD5SIG_FLAG_IFINDEX))
1317 continue;
1318 if (key->l3index != l3index)
1319 continue;
1320 if (!memcmp(&key->addr, addr, size) &&
1321 key->prefixlen == prefixlen)
1322 return key;
1323 }
1324 return NULL;
1325}
1326
1327struct tcp_md5sig_key *tcp_v4_md5_lookup(const struct sock *sk,
1328 const struct sock *addr_sk)
1329{
1330 const union tcp_md5_addr *addr;
1331 int l3index;
1332
1333 l3index = l3mdev_master_ifindex_by_index(sock_net(sk),
1334 addr_sk->sk_bound_dev_if);
1335 addr = (const union tcp_md5_addr *)&addr_sk->sk_daddr;
1336 return tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
1337}
1338EXPORT_SYMBOL(tcp_v4_md5_lookup);
1339
1340static int tcp_md5sig_info_add(struct sock *sk, gfp_t gfp)
1341{
1342 struct tcp_sock *tp = tcp_sk(sk);
1343 struct tcp_md5sig_info *md5sig;
1344
1345 md5sig = kmalloc(sizeof(*md5sig), gfp);
1346 if (!md5sig)
1347 return -ENOMEM;
1348
1349 sk_gso_disable(sk);
1350 INIT_HLIST_HEAD(&md5sig->head);
1351 rcu_assign_pointer(tp->md5sig_info, md5sig);
1352 return 0;
1353}
1354
1355/* This can be called on a newly created socket, from other files */
1356static int __tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
1357 int family, u8 prefixlen, int l3index, u8 flags,
1358 const u8 *newkey, u8 newkeylen, gfp_t gfp)
1359{
1360 /* Add Key to the list */
1361 struct tcp_md5sig_key *key;
1362 struct tcp_sock *tp = tcp_sk(sk);
1363 struct tcp_md5sig_info *md5sig;
1364
1365 key = tcp_md5_do_lookup_exact(sk, addr, family, prefixlen, l3index, flags);
1366 if (key) {
1367 /* Pre-existing entry - just update that one.
1368 * Note that the key might be used concurrently.
1369 * data_race() is telling kcsan that we do not care of
1370 * key mismatches, since changing MD5 key on live flows
1371 * can lead to packet drops.
1372 */
1373 data_race(memcpy(key->key, newkey, newkeylen));
1374
1375 /* Pairs with READ_ONCE() in tcp_md5_hash_key().
1376 * Also note that a reader could catch new key->keylen value
1377 * but old key->key[], this is the reason we use __GFP_ZERO
1378 * at sock_kmalloc() time below these lines.
1379 */
1380 WRITE_ONCE(key->keylen, newkeylen);
1381
1382 return 0;
1383 }
1384
1385 md5sig = rcu_dereference_protected(tp->md5sig_info,
1386 lockdep_sock_is_held(sk));
1387
1388 key = sock_kmalloc(sk, sizeof(*key), gfp | __GFP_ZERO);
1389 if (!key)
1390 return -ENOMEM;
1391
1392 memcpy(key->key, newkey, newkeylen);
1393 key->keylen = newkeylen;
1394 key->family = family;
1395 key->prefixlen = prefixlen;
1396 key->l3index = l3index;
1397 key->flags = flags;
1398 memcpy(&key->addr, addr,
1399 (IS_ENABLED(CONFIG_IPV6) && family == AF_INET6) ? sizeof(struct in6_addr) :
1400 sizeof(struct in_addr));
1401 hlist_add_head_rcu(&key->node, &md5sig->head);
1402 return 0;
1403}
1404
1405int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
1406 int family, u8 prefixlen, int l3index, u8 flags,
1407 const u8 *newkey, u8 newkeylen)
1408{
1409 struct tcp_sock *tp = tcp_sk(sk);
1410
1411 if (!rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk))) {
1412 if (tcp_md5_alloc_sigpool())
1413 return -ENOMEM;
1414
1415 if (tcp_md5sig_info_add(sk, GFP_KERNEL)) {
1416 tcp_md5_release_sigpool();
1417 return -ENOMEM;
1418 }
1419
1420 if (!static_branch_inc(&tcp_md5_needed.key)) {
1421 struct tcp_md5sig_info *md5sig;
1422
1423 md5sig = rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk));
1424 rcu_assign_pointer(tp->md5sig_info, NULL);
1425 kfree_rcu(md5sig, rcu);
1426 tcp_md5_release_sigpool();
1427 return -EUSERS;
1428 }
1429 }
1430
1431 return __tcp_md5_do_add(sk, addr, family, prefixlen, l3index, flags,
1432 newkey, newkeylen, GFP_KERNEL);
1433}
1434EXPORT_SYMBOL(tcp_md5_do_add);
1435
1436int tcp_md5_key_copy(struct sock *sk, const union tcp_md5_addr *addr,
1437 int family, u8 prefixlen, int l3index,
1438 struct tcp_md5sig_key *key)
1439{
1440 struct tcp_sock *tp = tcp_sk(sk);
1441
1442 if (!rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk))) {
1443 tcp_md5_add_sigpool();
1444
1445 if (tcp_md5sig_info_add(sk, sk_gfp_mask(sk, GFP_ATOMIC))) {
1446 tcp_md5_release_sigpool();
1447 return -ENOMEM;
1448 }
1449
1450 if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key)) {
1451 struct tcp_md5sig_info *md5sig;
1452
1453 md5sig = rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk));
1454 net_warn_ratelimited("Too many TCP-MD5 keys in the system\n");
1455 rcu_assign_pointer(tp->md5sig_info, NULL);
1456 kfree_rcu(md5sig, rcu);
1457 tcp_md5_release_sigpool();
1458 return -EUSERS;
1459 }
1460 }
1461
1462 return __tcp_md5_do_add(sk, addr, family, prefixlen, l3index,
1463 key->flags, key->key, key->keylen,
1464 sk_gfp_mask(sk, GFP_ATOMIC));
1465}
1466EXPORT_SYMBOL(tcp_md5_key_copy);
1467
1468int tcp_md5_do_del(struct sock *sk, const union tcp_md5_addr *addr, int family,
1469 u8 prefixlen, int l3index, u8 flags)
1470{
1471 struct tcp_md5sig_key *key;
1472
1473 key = tcp_md5_do_lookup_exact(sk, addr, family, prefixlen, l3index, flags);
1474 if (!key)
1475 return -ENOENT;
1476 hlist_del_rcu(&key->node);
1477 atomic_sub(sizeof(*key), &sk->sk_omem_alloc);
1478 kfree_rcu(key, rcu);
1479 return 0;
1480}
1481EXPORT_SYMBOL(tcp_md5_do_del);
1482
1483void tcp_clear_md5_list(struct sock *sk)
1484{
1485 struct tcp_sock *tp = tcp_sk(sk);
1486 struct tcp_md5sig_key *key;
1487 struct hlist_node *n;
1488 struct tcp_md5sig_info *md5sig;
1489
1490 md5sig = rcu_dereference_protected(tp->md5sig_info, 1);
1491
1492 hlist_for_each_entry_safe(key, n, &md5sig->head, node) {
1493 hlist_del_rcu(&key->node);
1494 atomic_sub(sizeof(*key), &sk->sk_omem_alloc);
1495 kfree_rcu(key, rcu);
1496 }
1497}
1498
1499static int tcp_v4_parse_md5_keys(struct sock *sk, int optname,
1500 sockptr_t optval, int optlen)
1501{
1502 struct tcp_md5sig cmd;
1503 struct sockaddr_in *sin = (struct sockaddr_in *)&cmd.tcpm_addr;
1504 const union tcp_md5_addr *addr;
1505 u8 prefixlen = 32;
1506 int l3index = 0;
1507 bool l3flag;
1508 u8 flags;
1509
1510 if (optlen < sizeof(cmd))
1511 return -EINVAL;
1512
1513 if (copy_from_sockptr(&cmd, optval, sizeof(cmd)))
1514 return -EFAULT;
1515
1516 if (sin->sin_family != AF_INET)
1517 return -EINVAL;
1518
1519 flags = cmd.tcpm_flags & TCP_MD5SIG_FLAG_IFINDEX;
1520 l3flag = cmd.tcpm_flags & TCP_MD5SIG_FLAG_IFINDEX;
1521
1522 if (optname == TCP_MD5SIG_EXT &&
1523 cmd.tcpm_flags & TCP_MD5SIG_FLAG_PREFIX) {
1524 prefixlen = cmd.tcpm_prefixlen;
1525 if (prefixlen > 32)
1526 return -EINVAL;
1527 }
1528
1529 if (optname == TCP_MD5SIG_EXT && cmd.tcpm_ifindex &&
1530 cmd.tcpm_flags & TCP_MD5SIG_FLAG_IFINDEX) {
1531 struct net_device *dev;
1532
1533 rcu_read_lock();
1534 dev = dev_get_by_index_rcu(sock_net(sk), cmd.tcpm_ifindex);
1535 if (dev && netif_is_l3_master(dev))
1536 l3index = dev->ifindex;
1537
1538 rcu_read_unlock();
1539
1540 /* ok to reference set/not set outside of rcu;
1541 * right now device MUST be an L3 master
1542 */
1543 if (!dev || !l3index)
1544 return -EINVAL;
1545 }
1546
1547 addr = (union tcp_md5_addr *)&sin->sin_addr.s_addr;
1548
1549 if (!cmd.tcpm_keylen)
1550 return tcp_md5_do_del(sk, addr, AF_INET, prefixlen, l3index, flags);
1551
1552 if (cmd.tcpm_keylen > TCP_MD5SIG_MAXKEYLEN)
1553 return -EINVAL;
1554
1555 /* Don't allow keys for peers that have a matching TCP-AO key.
1556 * See the comment in tcp_ao_add_cmd()
1557 */
1558 if (tcp_ao_required(sk, addr, AF_INET, l3flag ? l3index : -1, false))
1559 return -EKEYREJECTED;
1560
1561 return tcp_md5_do_add(sk, addr, AF_INET, prefixlen, l3index, flags,
1562 cmd.tcpm_key, cmd.tcpm_keylen);
1563}
1564
1565static int tcp_v4_md5_hash_headers(struct tcp_sigpool *hp,
1566 __be32 daddr, __be32 saddr,
1567 const struct tcphdr *th, int nbytes)
1568{
1569 struct tcp4_pseudohdr *bp;
1570 struct scatterlist sg;
1571 struct tcphdr *_th;
1572
1573 bp = hp->scratch;
1574 bp->saddr = saddr;
1575 bp->daddr = daddr;
1576 bp->pad = 0;
1577 bp->protocol = IPPROTO_TCP;
1578 bp->len = cpu_to_be16(nbytes);
1579
1580 _th = (struct tcphdr *)(bp + 1);
1581 memcpy(_th, th, sizeof(*th));
1582 _th->check = 0;
1583
1584 sg_init_one(&sg, bp, sizeof(*bp) + sizeof(*th));
1585 ahash_request_set_crypt(hp->req, &sg, NULL,
1586 sizeof(*bp) + sizeof(*th));
1587 return crypto_ahash_update(hp->req);
1588}
1589
1590static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
1591 __be32 daddr, __be32 saddr, const struct tcphdr *th)
1592{
1593 struct tcp_sigpool hp;
1594
1595 if (tcp_sigpool_start(tcp_md5_sigpool_id, &hp))
1596 goto clear_hash_nostart;
1597
1598 if (crypto_ahash_init(hp.req))
1599 goto clear_hash;
1600 if (tcp_v4_md5_hash_headers(&hp, daddr, saddr, th, th->doff << 2))
1601 goto clear_hash;
1602 if (tcp_md5_hash_key(&hp, key))
1603 goto clear_hash;
1604 ahash_request_set_crypt(hp.req, NULL, md5_hash, 0);
1605 if (crypto_ahash_final(hp.req))
1606 goto clear_hash;
1607
1608 tcp_sigpool_end(&hp);
1609 return 0;
1610
1611clear_hash:
1612 tcp_sigpool_end(&hp);
1613clear_hash_nostart:
1614 memset(md5_hash, 0, 16);
1615 return 1;
1616}
1617
1618int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
1619 const struct sock *sk,
1620 const struct sk_buff *skb)
1621{
1622 const struct tcphdr *th = tcp_hdr(skb);
1623 struct tcp_sigpool hp;
1624 __be32 saddr, daddr;
1625
1626 if (sk) { /* valid for establish/request sockets */
1627 saddr = sk->sk_rcv_saddr;
1628 daddr = sk->sk_daddr;
1629 } else {
1630 const struct iphdr *iph = ip_hdr(skb);
1631 saddr = iph->saddr;
1632 daddr = iph->daddr;
1633 }
1634
1635 if (tcp_sigpool_start(tcp_md5_sigpool_id, &hp))
1636 goto clear_hash_nostart;
1637
1638 if (crypto_ahash_init(hp.req))
1639 goto clear_hash;
1640
1641 if (tcp_v4_md5_hash_headers(&hp, daddr, saddr, th, skb->len))
1642 goto clear_hash;
1643 if (tcp_sigpool_hash_skb_data(&hp, skb, th->doff << 2))
1644 goto clear_hash;
1645 if (tcp_md5_hash_key(&hp, key))
1646 goto clear_hash;
1647 ahash_request_set_crypt(hp.req, NULL, md5_hash, 0);
1648 if (crypto_ahash_final(hp.req))
1649 goto clear_hash;
1650
1651 tcp_sigpool_end(&hp);
1652 return 0;
1653
1654clear_hash:
1655 tcp_sigpool_end(&hp);
1656clear_hash_nostart:
1657 memset(md5_hash, 0, 16);
1658 return 1;
1659}
1660EXPORT_SYMBOL(tcp_v4_md5_hash_skb);
1661
1662#endif
1663
1664static void tcp_v4_init_req(struct request_sock *req,
1665 const struct sock *sk_listener,
1666 struct sk_buff *skb)
1667{
1668 struct inet_request_sock *ireq = inet_rsk(req);
1669 struct net *net = sock_net(sk_listener);
1670
1671 sk_rcv_saddr_set(req_to_sk(req), ip_hdr(skb)->daddr);
1672 sk_daddr_set(req_to_sk(req), ip_hdr(skb)->saddr);
1673 RCU_INIT_POINTER(ireq->ireq_opt, tcp_v4_save_options(net, skb));
1674}
1675
1676static struct dst_entry *tcp_v4_route_req(const struct sock *sk,
1677 struct sk_buff *skb,
1678 struct flowi *fl,
1679 struct request_sock *req,
1680 u32 tw_isn)
1681{
1682 tcp_v4_init_req(req, sk, skb);
1683
1684 if (security_inet_conn_request(sk, skb, req))
1685 return NULL;
1686
1687 return inet_csk_route_req(sk, &fl->u.ip4, req);
1688}
1689
1690struct request_sock_ops tcp_request_sock_ops __read_mostly = {
1691 .family = PF_INET,
1692 .obj_size = sizeof(struct tcp_request_sock),
1693 .rtx_syn_ack = tcp_rtx_synack,
1694 .send_ack = tcp_v4_reqsk_send_ack,
1695 .destructor = tcp_v4_reqsk_destructor,
1696 .send_reset = tcp_v4_send_reset,
1697 .syn_ack_timeout = tcp_syn_ack_timeout,
1698};
1699
1700const struct tcp_request_sock_ops tcp_request_sock_ipv4_ops = {
1701 .mss_clamp = TCP_MSS_DEFAULT,
1702#ifdef CONFIG_TCP_MD5SIG
1703 .req_md5_lookup = tcp_v4_md5_lookup,
1704 .calc_md5_hash = tcp_v4_md5_hash_skb,
1705#endif
1706#ifdef CONFIG_TCP_AO
1707 .ao_lookup = tcp_v4_ao_lookup_rsk,
1708 .ao_calc_key = tcp_v4_ao_calc_key_rsk,
1709 .ao_synack_hash = tcp_v4_ao_synack_hash,
1710#endif
1711#ifdef CONFIG_SYN_COOKIES
1712 .cookie_init_seq = cookie_v4_init_sequence,
1713#endif
1714 .route_req = tcp_v4_route_req,
1715 .init_seq = tcp_v4_init_seq,
1716 .init_ts_off = tcp_v4_init_ts_off,
1717 .send_synack = tcp_v4_send_synack,
1718};
1719
1720int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb)
1721{
1722 /* Never answer to SYNs send to broadcast or multicast */
1723 if (skb_rtable(skb)->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST))
1724 goto drop;
1725
1726 return tcp_conn_request(&tcp_request_sock_ops,
1727 &tcp_request_sock_ipv4_ops, sk, skb);
1728
1729drop:
1730 tcp_listendrop(sk);
1731 return 0;
1732}
1733EXPORT_SYMBOL(tcp_v4_conn_request);
1734
1735
1736/*
1737 * The three way handshake has completed - we got a valid synack -
1738 * now create the new socket.
1739 */
1740struct sock *tcp_v4_syn_recv_sock(const struct sock *sk, struct sk_buff *skb,
1741 struct request_sock *req,
1742 struct dst_entry *dst,
1743 struct request_sock *req_unhash,
1744 bool *own_req)
1745{
1746 struct inet_request_sock *ireq;
1747 bool found_dup_sk = false;
1748 struct inet_sock *newinet;
1749 struct tcp_sock *newtp;
1750 struct sock *newsk;
1751#ifdef CONFIG_TCP_MD5SIG
1752 const union tcp_md5_addr *addr;
1753 struct tcp_md5sig_key *key;
1754 int l3index;
1755#endif
1756 struct ip_options_rcu *inet_opt;
1757
1758 if (sk_acceptq_is_full(sk))
1759 goto exit_overflow;
1760
1761 newsk = tcp_create_openreq_child(sk, req, skb);
1762 if (!newsk)
1763 goto exit_nonewsk;
1764
1765 newsk->sk_gso_type = SKB_GSO_TCPV4;
1766 inet_sk_rx_dst_set(newsk, skb);
1767
1768 newtp = tcp_sk(newsk);
1769 newinet = inet_sk(newsk);
1770 ireq = inet_rsk(req);
1771 sk_daddr_set(newsk, ireq->ir_rmt_addr);
1772 sk_rcv_saddr_set(newsk, ireq->ir_loc_addr);
1773 newsk->sk_bound_dev_if = ireq->ir_iif;
1774 newinet->inet_saddr = ireq->ir_loc_addr;
1775 inet_opt = rcu_dereference(ireq->ireq_opt);
1776 RCU_INIT_POINTER(newinet->inet_opt, inet_opt);
1777 newinet->mc_index = inet_iif(skb);
1778 newinet->mc_ttl = ip_hdr(skb)->ttl;
1779 newinet->rcv_tos = ip_hdr(skb)->tos;
1780 inet_csk(newsk)->icsk_ext_hdr_len = 0;
1781 if (inet_opt)
1782 inet_csk(newsk)->icsk_ext_hdr_len = inet_opt->opt.optlen;
1783 atomic_set(&newinet->inet_id, get_random_u16());
1784
1785 /* Set ToS of the new socket based upon the value of incoming SYN.
1786 * ECT bits are set later in tcp_init_transfer().
1787 */
1788 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reflect_tos))
1789 newinet->tos = tcp_rsk(req)->syn_tos & ~INET_ECN_MASK;
1790
1791 if (!dst) {
1792 dst = inet_csk_route_child_sock(sk, newsk, req);
1793 if (!dst)
1794 goto put_and_exit;
1795 } else {
1796 /* syncookie case : see end of cookie_v4_check() */
1797 }
1798 sk_setup_caps(newsk, dst);
1799
1800 tcp_ca_openreq_child(newsk, dst);
1801
1802 tcp_sync_mss(newsk, dst_mtu(dst));
1803 newtp->advmss = tcp_mss_clamp(tcp_sk(sk), dst_metric_advmss(dst));
1804
1805 tcp_initialize_rcv_mss(newsk);
1806
1807#ifdef CONFIG_TCP_MD5SIG
1808 l3index = l3mdev_master_ifindex_by_index(sock_net(sk), ireq->ir_iif);
1809 /* Copy over the MD5 key from the original socket */
1810 addr = (union tcp_md5_addr *)&newinet->inet_daddr;
1811 key = tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
1812 if (key && !tcp_rsk_used_ao(req)) {
1813 if (tcp_md5_key_copy(newsk, addr, AF_INET, 32, l3index, key))
1814 goto put_and_exit;
1815 sk_gso_disable(newsk);
1816 }
1817#endif
1818#ifdef CONFIG_TCP_AO
1819 if (tcp_ao_copy_all_matching(sk, newsk, req, skb, AF_INET))
1820 goto put_and_exit; /* OOM, release back memory */
1821#endif
1822
1823 if (__inet_inherit_port(sk, newsk) < 0)
1824 goto put_and_exit;
1825 *own_req = inet_ehash_nolisten(newsk, req_to_sk(req_unhash),
1826 &found_dup_sk);
1827 if (likely(*own_req)) {
1828 tcp_move_syn(newtp, req);
1829 ireq->ireq_opt = NULL;
1830 } else {
1831 newinet->inet_opt = NULL;
1832
1833 if (!req_unhash && found_dup_sk) {
1834 /* This code path should only be executed in the
1835 * syncookie case only
1836 */
1837 bh_unlock_sock(newsk);
1838 sock_put(newsk);
1839 newsk = NULL;
1840 }
1841 }
1842 return newsk;
1843
1844exit_overflow:
1845 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
1846exit_nonewsk:
1847 dst_release(dst);
1848exit:
1849 tcp_listendrop(sk);
1850 return NULL;
1851put_and_exit:
1852 newinet->inet_opt = NULL;
1853 inet_csk_prepare_forced_close(newsk);
1854 tcp_done(newsk);
1855 goto exit;
1856}
1857EXPORT_SYMBOL(tcp_v4_syn_recv_sock);
1858
1859static struct sock *tcp_v4_cookie_check(struct sock *sk, struct sk_buff *skb)
1860{
1861#ifdef CONFIG_SYN_COOKIES
1862 const struct tcphdr *th = tcp_hdr(skb);
1863
1864 if (!th->syn)
1865 sk = cookie_v4_check(sk, skb);
1866#endif
1867 return sk;
1868}
1869
1870u16 tcp_v4_get_syncookie(struct sock *sk, struct iphdr *iph,
1871 struct tcphdr *th, u32 *cookie)
1872{
1873 u16 mss = 0;
1874#ifdef CONFIG_SYN_COOKIES
1875 mss = tcp_get_syncookie_mss(&tcp_request_sock_ops,
1876 &tcp_request_sock_ipv4_ops, sk, th);
1877 if (mss) {
1878 *cookie = __cookie_v4_init_sequence(iph, th, &mss);
1879 tcp_synq_overflow(sk);
1880 }
1881#endif
1882 return mss;
1883}
1884
1885INDIRECT_CALLABLE_DECLARE(struct dst_entry *ipv4_dst_check(struct dst_entry *,
1886 u32));
1887/* The socket must have it's spinlock held when we get
1888 * here, unless it is a TCP_LISTEN socket.
1889 *
1890 * We have a potential double-lock case here, so even when
1891 * doing backlog processing we use the BH locking scheme.
1892 * This is because we cannot sleep with the original spinlock
1893 * held.
1894 */
1895int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb)
1896{
1897 enum skb_drop_reason reason;
1898 struct sock *rsk;
1899
1900 if (sk->sk_state == TCP_ESTABLISHED) { /* Fast path */
1901 struct dst_entry *dst;
1902
1903 dst = rcu_dereference_protected(sk->sk_rx_dst,
1904 lockdep_sock_is_held(sk));
1905
1906 sock_rps_save_rxhash(sk, skb);
1907 sk_mark_napi_id(sk, skb);
1908 if (dst) {
1909 if (sk->sk_rx_dst_ifindex != skb->skb_iif ||
1910 !INDIRECT_CALL_1(dst->ops->check, ipv4_dst_check,
1911 dst, 0)) {
1912 RCU_INIT_POINTER(sk->sk_rx_dst, NULL);
1913 dst_release(dst);
1914 }
1915 }
1916 tcp_rcv_established(sk, skb);
1917 return 0;
1918 }
1919
1920 if (tcp_checksum_complete(skb))
1921 goto csum_err;
1922
1923 if (sk->sk_state == TCP_LISTEN) {
1924 struct sock *nsk = tcp_v4_cookie_check(sk, skb);
1925
1926 if (!nsk)
1927 return 0;
1928 if (nsk != sk) {
1929 reason = tcp_child_process(sk, nsk, skb);
1930 if (reason) {
1931 rsk = nsk;
1932 goto reset;
1933 }
1934 return 0;
1935 }
1936 } else
1937 sock_rps_save_rxhash(sk, skb);
1938
1939 reason = tcp_rcv_state_process(sk, skb);
1940 if (reason) {
1941 rsk = sk;
1942 goto reset;
1943 }
1944 return 0;
1945
1946reset:
1947 tcp_v4_send_reset(rsk, skb, sk_rst_convert_drop_reason(reason));
1948discard:
1949 sk_skb_reason_drop(sk, skb, reason);
1950 /* Be careful here. If this function gets more complicated and
1951 * gcc suffers from register pressure on the x86, sk (in %ebx)
1952 * might be destroyed here. This current version compiles correctly,
1953 * but you have been warned.
1954 */
1955 return 0;
1956
1957csum_err:
1958 reason = SKB_DROP_REASON_TCP_CSUM;
1959 trace_tcp_bad_csum(skb);
1960 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
1961 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
1962 goto discard;
1963}
1964EXPORT_SYMBOL(tcp_v4_do_rcv);
1965
1966int tcp_v4_early_demux(struct sk_buff *skb)
1967{
1968 struct net *net = dev_net(skb->dev);
1969 const struct iphdr *iph;
1970 const struct tcphdr *th;
1971 struct sock *sk;
1972
1973 if (skb->pkt_type != PACKET_HOST)
1974 return 0;
1975
1976 if (!pskb_may_pull(skb, skb_transport_offset(skb) + sizeof(struct tcphdr)))
1977 return 0;
1978
1979 iph = ip_hdr(skb);
1980 th = tcp_hdr(skb);
1981
1982 if (th->doff < sizeof(struct tcphdr) / 4)
1983 return 0;
1984
1985 sk = __inet_lookup_established(net, net->ipv4.tcp_death_row.hashinfo,
1986 iph->saddr, th->source,
1987 iph->daddr, ntohs(th->dest),
1988 skb->skb_iif, inet_sdif(skb));
1989 if (sk) {
1990 skb->sk = sk;
1991 skb->destructor = sock_edemux;
1992 if (sk_fullsock(sk)) {
1993 struct dst_entry *dst = rcu_dereference(sk->sk_rx_dst);
1994
1995 if (dst)
1996 dst = dst_check(dst, 0);
1997 if (dst &&
1998 sk->sk_rx_dst_ifindex == skb->skb_iif)
1999 skb_dst_set_noref(skb, dst);
2000 }
2001 }
2002 return 0;
2003}
2004
2005bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb,
2006 enum skb_drop_reason *reason)
2007{
2008 u32 tail_gso_size, tail_gso_segs;
2009 struct skb_shared_info *shinfo;
2010 const struct tcphdr *th;
2011 struct tcphdr *thtail;
2012 struct sk_buff *tail;
2013 unsigned int hdrlen;
2014 bool fragstolen;
2015 u32 gso_segs;
2016 u32 gso_size;
2017 u64 limit;
2018 int delta;
2019
2020 /* In case all data was pulled from skb frags (in __pskb_pull_tail()),
2021 * we can fix skb->truesize to its real value to avoid future drops.
2022 * This is valid because skb is not yet charged to the socket.
2023 * It has been noticed pure SACK packets were sometimes dropped
2024 * (if cooked by drivers without copybreak feature).
2025 */
2026 skb_condense(skb);
2027
2028 tcp_cleanup_skb(skb);
2029
2030 if (unlikely(tcp_checksum_complete(skb))) {
2031 bh_unlock_sock(sk);
2032 trace_tcp_bad_csum(skb);
2033 *reason = SKB_DROP_REASON_TCP_CSUM;
2034 __TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
2035 __TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
2036 return true;
2037 }
2038
2039 /* Attempt coalescing to last skb in backlog, even if we are
2040 * above the limits.
2041 * This is okay because skb capacity is limited to MAX_SKB_FRAGS.
2042 */
2043 th = (const struct tcphdr *)skb->data;
2044 hdrlen = th->doff * 4;
2045
2046 tail = sk->sk_backlog.tail;
2047 if (!tail)
2048 goto no_coalesce;
2049 thtail = (struct tcphdr *)tail->data;
2050
2051 if (TCP_SKB_CB(tail)->end_seq != TCP_SKB_CB(skb)->seq ||
2052 TCP_SKB_CB(tail)->ip_dsfield != TCP_SKB_CB(skb)->ip_dsfield ||
2053 ((TCP_SKB_CB(tail)->tcp_flags |
2054 TCP_SKB_CB(skb)->tcp_flags) & (TCPHDR_SYN | TCPHDR_RST | TCPHDR_URG)) ||
2055 !((TCP_SKB_CB(tail)->tcp_flags &
2056 TCP_SKB_CB(skb)->tcp_flags) & TCPHDR_ACK) ||
2057 ((TCP_SKB_CB(tail)->tcp_flags ^
2058 TCP_SKB_CB(skb)->tcp_flags) & (TCPHDR_ECE | TCPHDR_CWR)) ||
2059 !tcp_skb_can_collapse_rx(tail, skb) ||
2060 thtail->doff != th->doff ||
2061 memcmp(thtail + 1, th + 1, hdrlen - sizeof(*th)))
2062 goto no_coalesce;
2063
2064 __skb_pull(skb, hdrlen);
2065
2066 shinfo = skb_shinfo(skb);
2067 gso_size = shinfo->gso_size ?: skb->len;
2068 gso_segs = shinfo->gso_segs ?: 1;
2069
2070 shinfo = skb_shinfo(tail);
2071 tail_gso_size = shinfo->gso_size ?: (tail->len - hdrlen);
2072 tail_gso_segs = shinfo->gso_segs ?: 1;
2073
2074 if (skb_try_coalesce(tail, skb, &fragstolen, &delta)) {
2075 TCP_SKB_CB(tail)->end_seq = TCP_SKB_CB(skb)->end_seq;
2076
2077 if (likely(!before(TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(tail)->ack_seq))) {
2078 TCP_SKB_CB(tail)->ack_seq = TCP_SKB_CB(skb)->ack_seq;
2079 thtail->window = th->window;
2080 }
2081
2082 /* We have to update both TCP_SKB_CB(tail)->tcp_flags and
2083 * thtail->fin, so that the fast path in tcp_rcv_established()
2084 * is not entered if we append a packet with a FIN.
2085 * SYN, RST, URG are not present.
2086 * ACK is set on both packets.
2087 * PSH : we do not really care in TCP stack,
2088 * at least for 'GRO' packets.
2089 */
2090 thtail->fin |= th->fin;
2091 TCP_SKB_CB(tail)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2092
2093 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2094 TCP_SKB_CB(tail)->has_rxtstamp = true;
2095 tail->tstamp = skb->tstamp;
2096 skb_hwtstamps(tail)->hwtstamp = skb_hwtstamps(skb)->hwtstamp;
2097 }
2098
2099 /* Not as strict as GRO. We only need to carry mss max value */
2100 shinfo->gso_size = max(gso_size, tail_gso_size);
2101 shinfo->gso_segs = min_t(u32, gso_segs + tail_gso_segs, 0xFFFF);
2102
2103 sk->sk_backlog.len += delta;
2104 __NET_INC_STATS(sock_net(sk),
2105 LINUX_MIB_TCPBACKLOGCOALESCE);
2106 kfree_skb_partial(skb, fragstolen);
2107 return false;
2108 }
2109 __skb_push(skb, hdrlen);
2110
2111no_coalesce:
2112 /* sk->sk_backlog.len is reset only at the end of __release_sock().
2113 * Both sk->sk_backlog.len and sk->sk_rmem_alloc could reach
2114 * sk_rcvbuf in normal conditions.
2115 */
2116 limit = ((u64)READ_ONCE(sk->sk_rcvbuf)) << 1;
2117
2118 limit += ((u32)READ_ONCE(sk->sk_sndbuf)) >> 1;
2119
2120 /* Only socket owner can try to collapse/prune rx queues
2121 * to reduce memory overhead, so add a little headroom here.
2122 * Few sockets backlog are possibly concurrently non empty.
2123 */
2124 limit += 64 * 1024;
2125
2126 limit = min_t(u64, limit, UINT_MAX);
2127
2128 if (unlikely(sk_add_backlog(sk, skb, limit))) {
2129 bh_unlock_sock(sk);
2130 *reason = SKB_DROP_REASON_SOCKET_BACKLOG;
2131 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPBACKLOGDROP);
2132 return true;
2133 }
2134 return false;
2135}
2136EXPORT_SYMBOL(tcp_add_backlog);
2137
2138int tcp_filter(struct sock *sk, struct sk_buff *skb)
2139{
2140 struct tcphdr *th = (struct tcphdr *)skb->data;
2141
2142 return sk_filter_trim_cap(sk, skb, th->doff * 4);
2143}
2144EXPORT_SYMBOL(tcp_filter);
2145
2146static void tcp_v4_restore_cb(struct sk_buff *skb)
2147{
2148 memmove(IPCB(skb), &TCP_SKB_CB(skb)->header.h4,
2149 sizeof(struct inet_skb_parm));
2150}
2151
2152static void tcp_v4_fill_cb(struct sk_buff *skb, const struct iphdr *iph,
2153 const struct tcphdr *th)
2154{
2155 /* This is tricky : We move IPCB at its correct location into TCP_SKB_CB()
2156 * barrier() makes sure compiler wont play fool^Waliasing games.
2157 */
2158 memmove(&TCP_SKB_CB(skb)->header.h4, IPCB(skb),
2159 sizeof(struct inet_skb_parm));
2160 barrier();
2161
2162 TCP_SKB_CB(skb)->seq = ntohl(th->seq);
2163 TCP_SKB_CB(skb)->end_seq = (TCP_SKB_CB(skb)->seq + th->syn + th->fin +
2164 skb->len - th->doff * 4);
2165 TCP_SKB_CB(skb)->ack_seq = ntohl(th->ack_seq);
2166 TCP_SKB_CB(skb)->tcp_flags = tcp_flag_byte(th);
2167 TCP_SKB_CB(skb)->ip_dsfield = ipv4_get_dsfield(iph);
2168 TCP_SKB_CB(skb)->sacked = 0;
2169 TCP_SKB_CB(skb)->has_rxtstamp =
2170 skb->tstamp || skb_hwtstamps(skb)->hwtstamp;
2171}
2172
2173/*
2174 * From tcp_input.c
2175 */
2176
2177int tcp_v4_rcv(struct sk_buff *skb)
2178{
2179 struct net *net = dev_net(skb->dev);
2180 enum skb_drop_reason drop_reason;
2181 int sdif = inet_sdif(skb);
2182 int dif = inet_iif(skb);
2183 const struct iphdr *iph;
2184 const struct tcphdr *th;
2185 struct sock *sk = NULL;
2186 bool refcounted;
2187 int ret;
2188 u32 isn;
2189
2190 drop_reason = SKB_DROP_REASON_NOT_SPECIFIED;
2191 if (skb->pkt_type != PACKET_HOST)
2192 goto discard_it;
2193
2194 /* Count it even if it's bad */
2195 __TCP_INC_STATS(net, TCP_MIB_INSEGS);
2196
2197 if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
2198 goto discard_it;
2199
2200 th = (const struct tcphdr *)skb->data;
2201
2202 if (unlikely(th->doff < sizeof(struct tcphdr) / 4)) {
2203 drop_reason = SKB_DROP_REASON_PKT_TOO_SMALL;
2204 goto bad_packet;
2205 }
2206 if (!pskb_may_pull(skb, th->doff * 4))
2207 goto discard_it;
2208
2209 /* An explanation is required here, I think.
2210 * Packet length and doff are validated by header prediction,
2211 * provided case of th->doff==0 is eliminated.
2212 * So, we defer the checks. */
2213
2214 if (skb_checksum_init(skb, IPPROTO_TCP, inet_compute_pseudo))
2215 goto csum_error;
2216
2217 th = (const struct tcphdr *)skb->data;
2218 iph = ip_hdr(skb);
2219lookup:
2220 sk = __inet_lookup_skb(net->ipv4.tcp_death_row.hashinfo,
2221 skb, __tcp_hdrlen(th), th->source,
2222 th->dest, sdif, &refcounted);
2223 if (!sk)
2224 goto no_tcp_socket;
2225
2226 if (sk->sk_state == TCP_TIME_WAIT)
2227 goto do_time_wait;
2228
2229 if (sk->sk_state == TCP_NEW_SYN_RECV) {
2230 struct request_sock *req = inet_reqsk(sk);
2231 bool req_stolen = false;
2232 struct sock *nsk;
2233
2234 sk = req->rsk_listener;
2235 if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb))
2236 drop_reason = SKB_DROP_REASON_XFRM_POLICY;
2237 else
2238 drop_reason = tcp_inbound_hash(sk, req, skb,
2239 &iph->saddr, &iph->daddr,
2240 AF_INET, dif, sdif);
2241 if (unlikely(drop_reason)) {
2242 sk_drops_add(sk, skb);
2243 reqsk_put(req);
2244 goto discard_it;
2245 }
2246 if (tcp_checksum_complete(skb)) {
2247 reqsk_put(req);
2248 goto csum_error;
2249 }
2250 if (unlikely(sk->sk_state != TCP_LISTEN)) {
2251 nsk = reuseport_migrate_sock(sk, req_to_sk(req), skb);
2252 if (!nsk) {
2253 inet_csk_reqsk_queue_drop_and_put(sk, req);
2254 goto lookup;
2255 }
2256 sk = nsk;
2257 /* reuseport_migrate_sock() has already held one sk_refcnt
2258 * before returning.
2259 */
2260 } else {
2261 /* We own a reference on the listener, increase it again
2262 * as we might lose it too soon.
2263 */
2264 sock_hold(sk);
2265 }
2266 refcounted = true;
2267 nsk = NULL;
2268 if (!tcp_filter(sk, skb)) {
2269 th = (const struct tcphdr *)skb->data;
2270 iph = ip_hdr(skb);
2271 tcp_v4_fill_cb(skb, iph, th);
2272 nsk = tcp_check_req(sk, skb, req, false, &req_stolen);
2273 } else {
2274 drop_reason = SKB_DROP_REASON_SOCKET_FILTER;
2275 }
2276 if (!nsk) {
2277 reqsk_put(req);
2278 if (req_stolen) {
2279 /* Another cpu got exclusive access to req
2280 * and created a full blown socket.
2281 * Try to feed this packet to this socket
2282 * instead of discarding it.
2283 */
2284 tcp_v4_restore_cb(skb);
2285 sock_put(sk);
2286 goto lookup;
2287 }
2288 goto discard_and_relse;
2289 }
2290 nf_reset_ct(skb);
2291 if (nsk == sk) {
2292 reqsk_put(req);
2293 tcp_v4_restore_cb(skb);
2294 } else {
2295 drop_reason = tcp_child_process(sk, nsk, skb);
2296 if (drop_reason) {
2297 enum sk_rst_reason rst_reason;
2298
2299 rst_reason = sk_rst_convert_drop_reason(drop_reason);
2300 tcp_v4_send_reset(nsk, skb, rst_reason);
2301 goto discard_and_relse;
2302 }
2303 sock_put(sk);
2304 return 0;
2305 }
2306 }
2307
2308process:
2309 if (static_branch_unlikely(&ip4_min_ttl)) {
2310 /* min_ttl can be changed concurrently from do_ip_setsockopt() */
2311 if (unlikely(iph->ttl < READ_ONCE(inet_sk(sk)->min_ttl))) {
2312 __NET_INC_STATS(net, LINUX_MIB_TCPMINTTLDROP);
2313 drop_reason = SKB_DROP_REASON_TCP_MINTTL;
2314 goto discard_and_relse;
2315 }
2316 }
2317
2318 if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb)) {
2319 drop_reason = SKB_DROP_REASON_XFRM_POLICY;
2320 goto discard_and_relse;
2321 }
2322
2323 drop_reason = tcp_inbound_hash(sk, NULL, skb, &iph->saddr, &iph->daddr,
2324 AF_INET, dif, sdif);
2325 if (drop_reason)
2326 goto discard_and_relse;
2327
2328 nf_reset_ct(skb);
2329
2330 if (tcp_filter(sk, skb)) {
2331 drop_reason = SKB_DROP_REASON_SOCKET_FILTER;
2332 goto discard_and_relse;
2333 }
2334 th = (const struct tcphdr *)skb->data;
2335 iph = ip_hdr(skb);
2336 tcp_v4_fill_cb(skb, iph, th);
2337
2338 skb->dev = NULL;
2339
2340 if (sk->sk_state == TCP_LISTEN) {
2341 ret = tcp_v4_do_rcv(sk, skb);
2342 goto put_and_return;
2343 }
2344
2345 sk_incoming_cpu_update(sk);
2346
2347 bh_lock_sock_nested(sk);
2348 tcp_segs_in(tcp_sk(sk), skb);
2349 ret = 0;
2350 if (!sock_owned_by_user(sk)) {
2351 ret = tcp_v4_do_rcv(sk, skb);
2352 } else {
2353 if (tcp_add_backlog(sk, skb, &drop_reason))
2354 goto discard_and_relse;
2355 }
2356 bh_unlock_sock(sk);
2357
2358put_and_return:
2359 if (refcounted)
2360 sock_put(sk);
2361
2362 return ret;
2363
2364no_tcp_socket:
2365 drop_reason = SKB_DROP_REASON_NO_SOCKET;
2366 if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb))
2367 goto discard_it;
2368
2369 tcp_v4_fill_cb(skb, iph, th);
2370
2371 if (tcp_checksum_complete(skb)) {
2372csum_error:
2373 drop_reason = SKB_DROP_REASON_TCP_CSUM;
2374 trace_tcp_bad_csum(skb);
2375 __TCP_INC_STATS(net, TCP_MIB_CSUMERRORS);
2376bad_packet:
2377 __TCP_INC_STATS(net, TCP_MIB_INERRS);
2378 } else {
2379 tcp_v4_send_reset(NULL, skb, sk_rst_convert_drop_reason(drop_reason));
2380 }
2381
2382discard_it:
2383 SKB_DR_OR(drop_reason, NOT_SPECIFIED);
2384 /* Discard frame. */
2385 sk_skb_reason_drop(sk, skb, drop_reason);
2386 return 0;
2387
2388discard_and_relse:
2389 sk_drops_add(sk, skb);
2390 if (refcounted)
2391 sock_put(sk);
2392 goto discard_it;
2393
2394do_time_wait:
2395 if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb)) {
2396 drop_reason = SKB_DROP_REASON_XFRM_POLICY;
2397 inet_twsk_put(inet_twsk(sk));
2398 goto discard_it;
2399 }
2400
2401 tcp_v4_fill_cb(skb, iph, th);
2402
2403 if (tcp_checksum_complete(skb)) {
2404 inet_twsk_put(inet_twsk(sk));
2405 goto csum_error;
2406 }
2407 switch (tcp_timewait_state_process(inet_twsk(sk), skb, th, &isn)) {
2408 case TCP_TW_SYN: {
2409 struct sock *sk2 = inet_lookup_listener(net,
2410 net->ipv4.tcp_death_row.hashinfo,
2411 skb, __tcp_hdrlen(th),
2412 iph->saddr, th->source,
2413 iph->daddr, th->dest,
2414 inet_iif(skb),
2415 sdif);
2416 if (sk2) {
2417 inet_twsk_deschedule_put(inet_twsk(sk));
2418 sk = sk2;
2419 tcp_v4_restore_cb(skb);
2420 refcounted = false;
2421 __this_cpu_write(tcp_tw_isn, isn);
2422 goto process;
2423 }
2424 }
2425 /* to ACK */
2426 fallthrough;
2427 case TCP_TW_ACK:
2428 tcp_v4_timewait_ack(sk, skb);
2429 break;
2430 case TCP_TW_RST:
2431 tcp_v4_send_reset(sk, skb, SK_RST_REASON_TCP_TIMEWAIT_SOCKET);
2432 inet_twsk_deschedule_put(inet_twsk(sk));
2433 goto discard_it;
2434 case TCP_TW_SUCCESS:;
2435 }
2436 goto discard_it;
2437}
2438
2439static struct timewait_sock_ops tcp_timewait_sock_ops = {
2440 .twsk_obj_size = sizeof(struct tcp_timewait_sock),
2441 .twsk_destructor= tcp_twsk_destructor,
2442};
2443
2444void inet_sk_rx_dst_set(struct sock *sk, const struct sk_buff *skb)
2445{
2446 struct dst_entry *dst = skb_dst(skb);
2447
2448 if (dst && dst_hold_safe(dst)) {
2449 rcu_assign_pointer(sk->sk_rx_dst, dst);
2450 sk->sk_rx_dst_ifindex = skb->skb_iif;
2451 }
2452}
2453EXPORT_SYMBOL(inet_sk_rx_dst_set);
2454
2455const struct inet_connection_sock_af_ops ipv4_specific = {
2456 .queue_xmit = ip_queue_xmit,
2457 .send_check = tcp_v4_send_check,
2458 .rebuild_header = inet_sk_rebuild_header,
2459 .sk_rx_dst_set = inet_sk_rx_dst_set,
2460 .conn_request = tcp_v4_conn_request,
2461 .syn_recv_sock = tcp_v4_syn_recv_sock,
2462 .net_header_len = sizeof(struct iphdr),
2463 .setsockopt = ip_setsockopt,
2464 .getsockopt = ip_getsockopt,
2465 .addr2sockaddr = inet_csk_addr2sockaddr,
2466 .sockaddr_len = sizeof(struct sockaddr_in),
2467 .mtu_reduced = tcp_v4_mtu_reduced,
2468};
2469EXPORT_SYMBOL(ipv4_specific);
2470
2471#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
2472static const struct tcp_sock_af_ops tcp_sock_ipv4_specific = {
2473#ifdef CONFIG_TCP_MD5SIG
2474 .md5_lookup = tcp_v4_md5_lookup,
2475 .calc_md5_hash = tcp_v4_md5_hash_skb,
2476 .md5_parse = tcp_v4_parse_md5_keys,
2477#endif
2478#ifdef CONFIG_TCP_AO
2479 .ao_lookup = tcp_v4_ao_lookup,
2480 .calc_ao_hash = tcp_v4_ao_hash_skb,
2481 .ao_parse = tcp_v4_parse_ao,
2482 .ao_calc_key_sk = tcp_v4_ao_calc_key_sk,
2483#endif
2484};
2485#endif
2486
2487/* NOTE: A lot of things set to zero explicitly by call to
2488 * sk_alloc() so need not be done here.
2489 */
2490static int tcp_v4_init_sock(struct sock *sk)
2491{
2492 struct inet_connection_sock *icsk = inet_csk(sk);
2493
2494 tcp_init_sock(sk);
2495
2496 icsk->icsk_af_ops = &ipv4_specific;
2497
2498#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
2499 tcp_sk(sk)->af_specific = &tcp_sock_ipv4_specific;
2500#endif
2501
2502 return 0;
2503}
2504
2505#ifdef CONFIG_TCP_MD5SIG
2506static void tcp_md5sig_info_free_rcu(struct rcu_head *head)
2507{
2508 struct tcp_md5sig_info *md5sig;
2509
2510 md5sig = container_of(head, struct tcp_md5sig_info, rcu);
2511 kfree(md5sig);
2512 static_branch_slow_dec_deferred(&tcp_md5_needed);
2513 tcp_md5_release_sigpool();
2514}
2515#endif
2516
2517static void tcp_release_user_frags(struct sock *sk)
2518{
2519#ifdef CONFIG_PAGE_POOL
2520 unsigned long index;
2521 void *netmem;
2522
2523 xa_for_each(&sk->sk_user_frags, index, netmem)
2524 WARN_ON_ONCE(!napi_pp_put_page((__force netmem_ref)netmem));
2525#endif
2526}
2527
2528void tcp_v4_destroy_sock(struct sock *sk)
2529{
2530 struct tcp_sock *tp = tcp_sk(sk);
2531
2532 tcp_release_user_frags(sk);
2533
2534 xa_destroy(&sk->sk_user_frags);
2535
2536 trace_tcp_destroy_sock(sk);
2537
2538 tcp_clear_xmit_timers(sk);
2539
2540 tcp_cleanup_congestion_control(sk);
2541
2542 tcp_cleanup_ulp(sk);
2543
2544 /* Cleanup up the write buffer. */
2545 tcp_write_queue_purge(sk);
2546
2547 /* Check if we want to disable active TFO */
2548 tcp_fastopen_active_disable_ofo_check(sk);
2549
2550 /* Cleans up our, hopefully empty, out_of_order_queue. */
2551 skb_rbtree_purge(&tp->out_of_order_queue);
2552
2553#ifdef CONFIG_TCP_MD5SIG
2554 /* Clean up the MD5 key list, if any */
2555 if (tp->md5sig_info) {
2556 struct tcp_md5sig_info *md5sig;
2557
2558 md5sig = rcu_dereference_protected(tp->md5sig_info, 1);
2559 tcp_clear_md5_list(sk);
2560 call_rcu(&md5sig->rcu, tcp_md5sig_info_free_rcu);
2561 rcu_assign_pointer(tp->md5sig_info, NULL);
2562 }
2563#endif
2564 tcp_ao_destroy_sock(sk, false);
2565
2566 /* Clean up a referenced TCP bind bucket. */
2567 if (inet_csk(sk)->icsk_bind_hash)
2568 inet_put_port(sk);
2569
2570 BUG_ON(rcu_access_pointer(tp->fastopen_rsk));
2571
2572 /* If socket is aborted during connect operation */
2573 tcp_free_fastopen_req(tp);
2574 tcp_fastopen_destroy_cipher(sk);
2575 tcp_saved_syn_free(tp);
2576
2577 sk_sockets_allocated_dec(sk);
2578}
2579EXPORT_SYMBOL(tcp_v4_destroy_sock);
2580
2581#ifdef CONFIG_PROC_FS
2582/* Proc filesystem TCP sock list dumping. */
2583
2584static unsigned short seq_file_family(const struct seq_file *seq);
2585
2586static bool seq_sk_match(struct seq_file *seq, const struct sock *sk)
2587{
2588 unsigned short family = seq_file_family(seq);
2589
2590 /* AF_UNSPEC is used as a match all */
2591 return ((family == AF_UNSPEC || family == sk->sk_family) &&
2592 net_eq(sock_net(sk), seq_file_net(seq)));
2593}
2594
2595/* Find a non empty bucket (starting from st->bucket)
2596 * and return the first sk from it.
2597 */
2598static void *listening_get_first(struct seq_file *seq)
2599{
2600 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2601 struct tcp_iter_state *st = seq->private;
2602
2603 st->offset = 0;
2604 for (; st->bucket <= hinfo->lhash2_mask; st->bucket++) {
2605 struct inet_listen_hashbucket *ilb2;
2606 struct hlist_nulls_node *node;
2607 struct sock *sk;
2608
2609 ilb2 = &hinfo->lhash2[st->bucket];
2610 if (hlist_nulls_empty(&ilb2->nulls_head))
2611 continue;
2612
2613 spin_lock(&ilb2->lock);
2614 sk_nulls_for_each(sk, node, &ilb2->nulls_head) {
2615 if (seq_sk_match(seq, sk))
2616 return sk;
2617 }
2618 spin_unlock(&ilb2->lock);
2619 }
2620
2621 return NULL;
2622}
2623
2624/* Find the next sk of "cur" within the same bucket (i.e. st->bucket).
2625 * If "cur" is the last one in the st->bucket,
2626 * call listening_get_first() to return the first sk of the next
2627 * non empty bucket.
2628 */
2629static void *listening_get_next(struct seq_file *seq, void *cur)
2630{
2631 struct tcp_iter_state *st = seq->private;
2632 struct inet_listen_hashbucket *ilb2;
2633 struct hlist_nulls_node *node;
2634 struct inet_hashinfo *hinfo;
2635 struct sock *sk = cur;
2636
2637 ++st->num;
2638 ++st->offset;
2639
2640 sk = sk_nulls_next(sk);
2641 sk_nulls_for_each_from(sk, node) {
2642 if (seq_sk_match(seq, sk))
2643 return sk;
2644 }
2645
2646 hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2647 ilb2 = &hinfo->lhash2[st->bucket];
2648 spin_unlock(&ilb2->lock);
2649 ++st->bucket;
2650 return listening_get_first(seq);
2651}
2652
2653static void *listening_get_idx(struct seq_file *seq, loff_t *pos)
2654{
2655 struct tcp_iter_state *st = seq->private;
2656 void *rc;
2657
2658 st->bucket = 0;
2659 st->offset = 0;
2660 rc = listening_get_first(seq);
2661
2662 while (rc && *pos) {
2663 rc = listening_get_next(seq, rc);
2664 --*pos;
2665 }
2666 return rc;
2667}
2668
2669static inline bool empty_bucket(struct inet_hashinfo *hinfo,
2670 const struct tcp_iter_state *st)
2671{
2672 return hlist_nulls_empty(&hinfo->ehash[st->bucket].chain);
2673}
2674
2675/*
2676 * Get first established socket starting from bucket given in st->bucket.
2677 * If st->bucket is zero, the very first socket in the hash is returned.
2678 */
2679static void *established_get_first(struct seq_file *seq)
2680{
2681 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2682 struct tcp_iter_state *st = seq->private;
2683
2684 st->offset = 0;
2685 for (; st->bucket <= hinfo->ehash_mask; ++st->bucket) {
2686 struct sock *sk;
2687 struct hlist_nulls_node *node;
2688 spinlock_t *lock = inet_ehash_lockp(hinfo, st->bucket);
2689
2690 cond_resched();
2691
2692 /* Lockless fast path for the common case of empty buckets */
2693 if (empty_bucket(hinfo, st))
2694 continue;
2695
2696 spin_lock_bh(lock);
2697 sk_nulls_for_each(sk, node, &hinfo->ehash[st->bucket].chain) {
2698 if (seq_sk_match(seq, sk))
2699 return sk;
2700 }
2701 spin_unlock_bh(lock);
2702 }
2703
2704 return NULL;
2705}
2706
2707static void *established_get_next(struct seq_file *seq, void *cur)
2708{
2709 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2710 struct tcp_iter_state *st = seq->private;
2711 struct hlist_nulls_node *node;
2712 struct sock *sk = cur;
2713
2714 ++st->num;
2715 ++st->offset;
2716
2717 sk = sk_nulls_next(sk);
2718
2719 sk_nulls_for_each_from(sk, node) {
2720 if (seq_sk_match(seq, sk))
2721 return sk;
2722 }
2723
2724 spin_unlock_bh(inet_ehash_lockp(hinfo, st->bucket));
2725 ++st->bucket;
2726 return established_get_first(seq);
2727}
2728
2729static void *established_get_idx(struct seq_file *seq, loff_t pos)
2730{
2731 struct tcp_iter_state *st = seq->private;
2732 void *rc;
2733
2734 st->bucket = 0;
2735 rc = established_get_first(seq);
2736
2737 while (rc && pos) {
2738 rc = established_get_next(seq, rc);
2739 --pos;
2740 }
2741 return rc;
2742}
2743
2744static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
2745{
2746 void *rc;
2747 struct tcp_iter_state *st = seq->private;
2748
2749 st->state = TCP_SEQ_STATE_LISTENING;
2750 rc = listening_get_idx(seq, &pos);
2751
2752 if (!rc) {
2753 st->state = TCP_SEQ_STATE_ESTABLISHED;
2754 rc = established_get_idx(seq, pos);
2755 }
2756
2757 return rc;
2758}
2759
2760static void *tcp_seek_last_pos(struct seq_file *seq)
2761{
2762 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2763 struct tcp_iter_state *st = seq->private;
2764 int bucket = st->bucket;
2765 int offset = st->offset;
2766 int orig_num = st->num;
2767 void *rc = NULL;
2768
2769 switch (st->state) {
2770 case TCP_SEQ_STATE_LISTENING:
2771 if (st->bucket > hinfo->lhash2_mask)
2772 break;
2773 rc = listening_get_first(seq);
2774 while (offset-- && rc && bucket == st->bucket)
2775 rc = listening_get_next(seq, rc);
2776 if (rc)
2777 break;
2778 st->bucket = 0;
2779 st->state = TCP_SEQ_STATE_ESTABLISHED;
2780 fallthrough;
2781 case TCP_SEQ_STATE_ESTABLISHED:
2782 if (st->bucket > hinfo->ehash_mask)
2783 break;
2784 rc = established_get_first(seq);
2785 while (offset-- && rc && bucket == st->bucket)
2786 rc = established_get_next(seq, rc);
2787 }
2788
2789 st->num = orig_num;
2790
2791 return rc;
2792}
2793
2794void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
2795{
2796 struct tcp_iter_state *st = seq->private;
2797 void *rc;
2798
2799 if (*pos && *pos == st->last_pos) {
2800 rc = tcp_seek_last_pos(seq);
2801 if (rc)
2802 goto out;
2803 }
2804
2805 st->state = TCP_SEQ_STATE_LISTENING;
2806 st->num = 0;
2807 st->bucket = 0;
2808 st->offset = 0;
2809 rc = *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
2810
2811out:
2812 st->last_pos = *pos;
2813 return rc;
2814}
2815EXPORT_SYMBOL(tcp_seq_start);
2816
2817void *tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2818{
2819 struct tcp_iter_state *st = seq->private;
2820 void *rc = NULL;
2821
2822 if (v == SEQ_START_TOKEN) {
2823 rc = tcp_get_idx(seq, 0);
2824 goto out;
2825 }
2826
2827 switch (st->state) {
2828 case TCP_SEQ_STATE_LISTENING:
2829 rc = listening_get_next(seq, v);
2830 if (!rc) {
2831 st->state = TCP_SEQ_STATE_ESTABLISHED;
2832 st->bucket = 0;
2833 st->offset = 0;
2834 rc = established_get_first(seq);
2835 }
2836 break;
2837 case TCP_SEQ_STATE_ESTABLISHED:
2838 rc = established_get_next(seq, v);
2839 break;
2840 }
2841out:
2842 ++*pos;
2843 st->last_pos = *pos;
2844 return rc;
2845}
2846EXPORT_SYMBOL(tcp_seq_next);
2847
2848void tcp_seq_stop(struct seq_file *seq, void *v)
2849{
2850 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2851 struct tcp_iter_state *st = seq->private;
2852
2853 switch (st->state) {
2854 case TCP_SEQ_STATE_LISTENING:
2855 if (v != SEQ_START_TOKEN)
2856 spin_unlock(&hinfo->lhash2[st->bucket].lock);
2857 break;
2858 case TCP_SEQ_STATE_ESTABLISHED:
2859 if (v)
2860 spin_unlock_bh(inet_ehash_lockp(hinfo, st->bucket));
2861 break;
2862 }
2863}
2864EXPORT_SYMBOL(tcp_seq_stop);
2865
2866static void get_openreq4(const struct request_sock *req,
2867 struct seq_file *f, int i)
2868{
2869 const struct inet_request_sock *ireq = inet_rsk(req);
2870 long delta = req->rsk_timer.expires - jiffies;
2871
2872 seq_printf(f, "%4d: %08X:%04X %08X:%04X"
2873 " %02X %08X:%08X %02X:%08lX %08X %5u %8d %u %d %pK",
2874 i,
2875 ireq->ir_loc_addr,
2876 ireq->ir_num,
2877 ireq->ir_rmt_addr,
2878 ntohs(ireq->ir_rmt_port),
2879 TCP_SYN_RECV,
2880 0, 0, /* could print option size, but that is af dependent. */
2881 1, /* timers active (only the expire timer) */
2882 jiffies_delta_to_clock_t(delta),
2883 req->num_timeout,
2884 from_kuid_munged(seq_user_ns(f),
2885 sock_i_uid(req->rsk_listener)),
2886 0, /* non standard timer */
2887 0, /* open_requests have no inode */
2888 0,
2889 req);
2890}
2891
2892static void get_tcp4_sock(struct sock *sk, struct seq_file *f, int i)
2893{
2894 int timer_active;
2895 unsigned long timer_expires;
2896 const struct tcp_sock *tp = tcp_sk(sk);
2897 const struct inet_connection_sock *icsk = inet_csk(sk);
2898 const struct inet_sock *inet = inet_sk(sk);
2899 const struct fastopen_queue *fastopenq = &icsk->icsk_accept_queue.fastopenq;
2900 __be32 dest = inet->inet_daddr;
2901 __be32 src = inet->inet_rcv_saddr;
2902 __u16 destp = ntohs(inet->inet_dport);
2903 __u16 srcp = ntohs(inet->inet_sport);
2904 u8 icsk_pending;
2905 int rx_queue;
2906 int state;
2907
2908 icsk_pending = smp_load_acquire(&icsk->icsk_pending);
2909 if (icsk_pending == ICSK_TIME_RETRANS ||
2910 icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2911 icsk_pending == ICSK_TIME_LOSS_PROBE) {
2912 timer_active = 1;
2913 timer_expires = icsk->icsk_timeout;
2914 } else if (icsk_pending == ICSK_TIME_PROBE0) {
2915 timer_active = 4;
2916 timer_expires = icsk->icsk_timeout;
2917 } else if (timer_pending(&sk->sk_timer)) {
2918 timer_active = 2;
2919 timer_expires = sk->sk_timer.expires;
2920 } else {
2921 timer_active = 0;
2922 timer_expires = jiffies;
2923 }
2924
2925 state = inet_sk_state_load(sk);
2926 if (state == TCP_LISTEN)
2927 rx_queue = READ_ONCE(sk->sk_ack_backlog);
2928 else
2929 /* Because we don't lock the socket,
2930 * we might find a transient negative value.
2931 */
2932 rx_queue = max_t(int, READ_ONCE(tp->rcv_nxt) -
2933 READ_ONCE(tp->copied_seq), 0);
2934
2935 seq_printf(f, "%4d: %08X:%04X %08X:%04X %02X %08X:%08X %02X:%08lX "
2936 "%08X %5u %8d %lu %d %pK %lu %lu %u %u %d",
2937 i, src, srcp, dest, destp, state,
2938 READ_ONCE(tp->write_seq) - tp->snd_una,
2939 rx_queue,
2940 timer_active,
2941 jiffies_delta_to_clock_t(timer_expires - jiffies),
2942 icsk->icsk_retransmits,
2943 from_kuid_munged(seq_user_ns(f), sock_i_uid(sk)),
2944 icsk->icsk_probes_out,
2945 sock_i_ino(sk),
2946 refcount_read(&sk->sk_refcnt), sk,
2947 jiffies_to_clock_t(icsk->icsk_rto),
2948 jiffies_to_clock_t(icsk->icsk_ack.ato),
2949 (icsk->icsk_ack.quick << 1) | inet_csk_in_pingpong_mode(sk),
2950 tcp_snd_cwnd(tp),
2951 state == TCP_LISTEN ?
2952 fastopenq->max_qlen :
2953 (tcp_in_initial_slowstart(tp) ? -1 : tp->snd_ssthresh));
2954}
2955
2956static void get_timewait4_sock(const struct inet_timewait_sock *tw,
2957 struct seq_file *f, int i)
2958{
2959 long delta = tw->tw_timer.expires - jiffies;
2960 __be32 dest, src;
2961 __u16 destp, srcp;
2962
2963 dest = tw->tw_daddr;
2964 src = tw->tw_rcv_saddr;
2965 destp = ntohs(tw->tw_dport);
2966 srcp = ntohs(tw->tw_sport);
2967
2968 seq_printf(f, "%4d: %08X:%04X %08X:%04X"
2969 " %02X %08X:%08X %02X:%08lX %08X %5d %8d %d %d %pK",
2970 i, src, srcp, dest, destp, READ_ONCE(tw->tw_substate), 0, 0,
2971 3, jiffies_delta_to_clock_t(delta), 0, 0, 0, 0,
2972 refcount_read(&tw->tw_refcnt), tw);
2973}
2974
2975#define TMPSZ 150
2976
2977static int tcp4_seq_show(struct seq_file *seq, void *v)
2978{
2979 struct tcp_iter_state *st;
2980 struct sock *sk = v;
2981
2982 seq_setwidth(seq, TMPSZ - 1);
2983 if (v == SEQ_START_TOKEN) {
2984 seq_puts(seq, " sl local_address rem_address st tx_queue "
2985 "rx_queue tr tm->when retrnsmt uid timeout "
2986 "inode");
2987 goto out;
2988 }
2989 st = seq->private;
2990
2991 if (sk->sk_state == TCP_TIME_WAIT)
2992 get_timewait4_sock(v, seq, st->num);
2993 else if (sk->sk_state == TCP_NEW_SYN_RECV)
2994 get_openreq4(v, seq, st->num);
2995 else
2996 get_tcp4_sock(v, seq, st->num);
2997out:
2998 seq_pad(seq, '\n');
2999 return 0;
3000}
3001
3002#ifdef CONFIG_BPF_SYSCALL
3003struct bpf_tcp_iter_state {
3004 struct tcp_iter_state state;
3005 unsigned int cur_sk;
3006 unsigned int end_sk;
3007 unsigned int max_sk;
3008 struct sock **batch;
3009 bool st_bucket_done;
3010};
3011
3012struct bpf_iter__tcp {
3013 __bpf_md_ptr(struct bpf_iter_meta *, meta);
3014 __bpf_md_ptr(struct sock_common *, sk_common);
3015 uid_t uid __aligned(8);
3016};
3017
3018static int tcp_prog_seq_show(struct bpf_prog *prog, struct bpf_iter_meta *meta,
3019 struct sock_common *sk_common, uid_t uid)
3020{
3021 struct bpf_iter__tcp ctx;
3022
3023 meta->seq_num--; /* skip SEQ_START_TOKEN */
3024 ctx.meta = meta;
3025 ctx.sk_common = sk_common;
3026 ctx.uid = uid;
3027 return bpf_iter_run_prog(prog, &ctx);
3028}
3029
3030static void bpf_iter_tcp_put_batch(struct bpf_tcp_iter_state *iter)
3031{
3032 while (iter->cur_sk < iter->end_sk)
3033 sock_gen_put(iter->batch[iter->cur_sk++]);
3034}
3035
3036static int bpf_iter_tcp_realloc_batch(struct bpf_tcp_iter_state *iter,
3037 unsigned int new_batch_sz)
3038{
3039 struct sock **new_batch;
3040
3041 new_batch = kvmalloc(sizeof(*new_batch) * new_batch_sz,
3042 GFP_USER | __GFP_NOWARN);
3043 if (!new_batch)
3044 return -ENOMEM;
3045
3046 bpf_iter_tcp_put_batch(iter);
3047 kvfree(iter->batch);
3048 iter->batch = new_batch;
3049 iter->max_sk = new_batch_sz;
3050
3051 return 0;
3052}
3053
3054static unsigned int bpf_iter_tcp_listening_batch(struct seq_file *seq,
3055 struct sock *start_sk)
3056{
3057 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
3058 struct bpf_tcp_iter_state *iter = seq->private;
3059 struct tcp_iter_state *st = &iter->state;
3060 struct hlist_nulls_node *node;
3061 unsigned int expected = 1;
3062 struct sock *sk;
3063
3064 sock_hold(start_sk);
3065 iter->batch[iter->end_sk++] = start_sk;
3066
3067 sk = sk_nulls_next(start_sk);
3068 sk_nulls_for_each_from(sk, node) {
3069 if (seq_sk_match(seq, sk)) {
3070 if (iter->end_sk < iter->max_sk) {
3071 sock_hold(sk);
3072 iter->batch[iter->end_sk++] = sk;
3073 }
3074 expected++;
3075 }
3076 }
3077 spin_unlock(&hinfo->lhash2[st->bucket].lock);
3078
3079 return expected;
3080}
3081
3082static unsigned int bpf_iter_tcp_established_batch(struct seq_file *seq,
3083 struct sock *start_sk)
3084{
3085 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
3086 struct bpf_tcp_iter_state *iter = seq->private;
3087 struct tcp_iter_state *st = &iter->state;
3088 struct hlist_nulls_node *node;
3089 unsigned int expected = 1;
3090 struct sock *sk;
3091
3092 sock_hold(start_sk);
3093 iter->batch[iter->end_sk++] = start_sk;
3094
3095 sk = sk_nulls_next(start_sk);
3096 sk_nulls_for_each_from(sk, node) {
3097 if (seq_sk_match(seq, sk)) {
3098 if (iter->end_sk < iter->max_sk) {
3099 sock_hold(sk);
3100 iter->batch[iter->end_sk++] = sk;
3101 }
3102 expected++;
3103 }
3104 }
3105 spin_unlock_bh(inet_ehash_lockp(hinfo, st->bucket));
3106
3107 return expected;
3108}
3109
3110static struct sock *bpf_iter_tcp_batch(struct seq_file *seq)
3111{
3112 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
3113 struct bpf_tcp_iter_state *iter = seq->private;
3114 struct tcp_iter_state *st = &iter->state;
3115 unsigned int expected;
3116 bool resized = false;
3117 struct sock *sk;
3118
3119 /* The st->bucket is done. Directly advance to the next
3120 * bucket instead of having the tcp_seek_last_pos() to skip
3121 * one by one in the current bucket and eventually find out
3122 * it has to advance to the next bucket.
3123 */
3124 if (iter->st_bucket_done) {
3125 st->offset = 0;
3126 st->bucket++;
3127 if (st->state == TCP_SEQ_STATE_LISTENING &&
3128 st->bucket > hinfo->lhash2_mask) {
3129 st->state = TCP_SEQ_STATE_ESTABLISHED;
3130 st->bucket = 0;
3131 }
3132 }
3133
3134again:
3135 /* Get a new batch */
3136 iter->cur_sk = 0;
3137 iter->end_sk = 0;
3138 iter->st_bucket_done = false;
3139
3140 sk = tcp_seek_last_pos(seq);
3141 if (!sk)
3142 return NULL; /* Done */
3143
3144 if (st->state == TCP_SEQ_STATE_LISTENING)
3145 expected = bpf_iter_tcp_listening_batch(seq, sk);
3146 else
3147 expected = bpf_iter_tcp_established_batch(seq, sk);
3148
3149 if (iter->end_sk == expected) {
3150 iter->st_bucket_done = true;
3151 return sk;
3152 }
3153
3154 if (!resized && !bpf_iter_tcp_realloc_batch(iter, expected * 3 / 2)) {
3155 resized = true;
3156 goto again;
3157 }
3158
3159 return sk;
3160}
3161
3162static void *bpf_iter_tcp_seq_start(struct seq_file *seq, loff_t *pos)
3163{
3164 /* bpf iter does not support lseek, so it always
3165 * continue from where it was stop()-ped.
3166 */
3167 if (*pos)
3168 return bpf_iter_tcp_batch(seq);
3169
3170 return SEQ_START_TOKEN;
3171}
3172
3173static void *bpf_iter_tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
3174{
3175 struct bpf_tcp_iter_state *iter = seq->private;
3176 struct tcp_iter_state *st = &iter->state;
3177 struct sock *sk;
3178
3179 /* Whenever seq_next() is called, the iter->cur_sk is
3180 * done with seq_show(), so advance to the next sk in
3181 * the batch.
3182 */
3183 if (iter->cur_sk < iter->end_sk) {
3184 /* Keeping st->num consistent in tcp_iter_state.
3185 * bpf_iter_tcp does not use st->num.
3186 * meta.seq_num is used instead.
3187 */
3188 st->num++;
3189 /* Move st->offset to the next sk in the bucket such that
3190 * the future start() will resume at st->offset in
3191 * st->bucket. See tcp_seek_last_pos().
3192 */
3193 st->offset++;
3194 sock_gen_put(iter->batch[iter->cur_sk++]);
3195 }
3196
3197 if (iter->cur_sk < iter->end_sk)
3198 sk = iter->batch[iter->cur_sk];
3199 else
3200 sk = bpf_iter_tcp_batch(seq);
3201
3202 ++*pos;
3203 /* Keeping st->last_pos consistent in tcp_iter_state.
3204 * bpf iter does not do lseek, so st->last_pos always equals to *pos.
3205 */
3206 st->last_pos = *pos;
3207 return sk;
3208}
3209
3210static int bpf_iter_tcp_seq_show(struct seq_file *seq, void *v)
3211{
3212 struct bpf_iter_meta meta;
3213 struct bpf_prog *prog;
3214 struct sock *sk = v;
3215 uid_t uid;
3216 int ret;
3217
3218 if (v == SEQ_START_TOKEN)
3219 return 0;
3220
3221 if (sk_fullsock(sk))
3222 lock_sock(sk);
3223
3224 if (unlikely(sk_unhashed(sk))) {
3225 ret = SEQ_SKIP;
3226 goto unlock;
3227 }
3228
3229 if (sk->sk_state == TCP_TIME_WAIT) {
3230 uid = 0;
3231 } else if (sk->sk_state == TCP_NEW_SYN_RECV) {
3232 const struct request_sock *req = v;
3233
3234 uid = from_kuid_munged(seq_user_ns(seq),
3235 sock_i_uid(req->rsk_listener));
3236 } else {
3237 uid = from_kuid_munged(seq_user_ns(seq), sock_i_uid(sk));
3238 }
3239
3240 meta.seq = seq;
3241 prog = bpf_iter_get_info(&meta, false);
3242 ret = tcp_prog_seq_show(prog, &meta, v, uid);
3243
3244unlock:
3245 if (sk_fullsock(sk))
3246 release_sock(sk);
3247 return ret;
3248
3249}
3250
3251static void bpf_iter_tcp_seq_stop(struct seq_file *seq, void *v)
3252{
3253 struct bpf_tcp_iter_state *iter = seq->private;
3254 struct bpf_iter_meta meta;
3255 struct bpf_prog *prog;
3256
3257 if (!v) {
3258 meta.seq = seq;
3259 prog = bpf_iter_get_info(&meta, true);
3260 if (prog)
3261 (void)tcp_prog_seq_show(prog, &meta, v, 0);
3262 }
3263
3264 if (iter->cur_sk < iter->end_sk) {
3265 bpf_iter_tcp_put_batch(iter);
3266 iter->st_bucket_done = false;
3267 }
3268}
3269
3270static const struct seq_operations bpf_iter_tcp_seq_ops = {
3271 .show = bpf_iter_tcp_seq_show,
3272 .start = bpf_iter_tcp_seq_start,
3273 .next = bpf_iter_tcp_seq_next,
3274 .stop = bpf_iter_tcp_seq_stop,
3275};
3276#endif
3277static unsigned short seq_file_family(const struct seq_file *seq)
3278{
3279 const struct tcp_seq_afinfo *afinfo;
3280
3281#ifdef CONFIG_BPF_SYSCALL
3282 /* Iterated from bpf_iter. Let the bpf prog to filter instead. */
3283 if (seq->op == &bpf_iter_tcp_seq_ops)
3284 return AF_UNSPEC;
3285#endif
3286
3287 /* Iterated from proc fs */
3288 afinfo = pde_data(file_inode(seq->file));
3289 return afinfo->family;
3290}
3291
3292static const struct seq_operations tcp4_seq_ops = {
3293 .show = tcp4_seq_show,
3294 .start = tcp_seq_start,
3295 .next = tcp_seq_next,
3296 .stop = tcp_seq_stop,
3297};
3298
3299static struct tcp_seq_afinfo tcp4_seq_afinfo = {
3300 .family = AF_INET,
3301};
3302
3303static int __net_init tcp4_proc_init_net(struct net *net)
3304{
3305 if (!proc_create_net_data("tcp", 0444, net->proc_net, &tcp4_seq_ops,
3306 sizeof(struct tcp_iter_state), &tcp4_seq_afinfo))
3307 return -ENOMEM;
3308 return 0;
3309}
3310
3311static void __net_exit tcp4_proc_exit_net(struct net *net)
3312{
3313 remove_proc_entry("tcp", net->proc_net);
3314}
3315
3316static struct pernet_operations tcp4_net_ops = {
3317 .init = tcp4_proc_init_net,
3318 .exit = tcp4_proc_exit_net,
3319};
3320
3321int __init tcp4_proc_init(void)
3322{
3323 return register_pernet_subsys(&tcp4_net_ops);
3324}
3325
3326void tcp4_proc_exit(void)
3327{
3328 unregister_pernet_subsys(&tcp4_net_ops);
3329}
3330#endif /* CONFIG_PROC_FS */
3331
3332/* @wake is one when sk_stream_write_space() calls us.
3333 * This sends EPOLLOUT only if notsent_bytes is half the limit.
3334 * This mimics the strategy used in sock_def_write_space().
3335 */
3336bool tcp_stream_memory_free(const struct sock *sk, int wake)
3337{
3338 const struct tcp_sock *tp = tcp_sk(sk);
3339 u32 notsent_bytes = READ_ONCE(tp->write_seq) -
3340 READ_ONCE(tp->snd_nxt);
3341
3342 return (notsent_bytes << wake) < tcp_notsent_lowat(tp);
3343}
3344EXPORT_SYMBOL(tcp_stream_memory_free);
3345
3346struct proto tcp_prot = {
3347 .name = "TCP",
3348 .owner = THIS_MODULE,
3349 .close = tcp_close,
3350 .pre_connect = tcp_v4_pre_connect,
3351 .connect = tcp_v4_connect,
3352 .disconnect = tcp_disconnect,
3353 .accept = inet_csk_accept,
3354 .ioctl = tcp_ioctl,
3355 .init = tcp_v4_init_sock,
3356 .destroy = tcp_v4_destroy_sock,
3357 .shutdown = tcp_shutdown,
3358 .setsockopt = tcp_setsockopt,
3359 .getsockopt = tcp_getsockopt,
3360 .bpf_bypass_getsockopt = tcp_bpf_bypass_getsockopt,
3361 .keepalive = tcp_set_keepalive,
3362 .recvmsg = tcp_recvmsg,
3363 .sendmsg = tcp_sendmsg,
3364 .splice_eof = tcp_splice_eof,
3365 .backlog_rcv = tcp_v4_do_rcv,
3366 .release_cb = tcp_release_cb,
3367 .hash = inet_hash,
3368 .unhash = inet_unhash,
3369 .get_port = inet_csk_get_port,
3370 .put_port = inet_put_port,
3371#ifdef CONFIG_BPF_SYSCALL
3372 .psock_update_sk_prot = tcp_bpf_update_proto,
3373#endif
3374 .enter_memory_pressure = tcp_enter_memory_pressure,
3375 .leave_memory_pressure = tcp_leave_memory_pressure,
3376 .stream_memory_free = tcp_stream_memory_free,
3377 .sockets_allocated = &tcp_sockets_allocated,
3378 .orphan_count = &tcp_orphan_count,
3379
3380 .memory_allocated = &tcp_memory_allocated,
3381 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3382
3383 .memory_pressure = &tcp_memory_pressure,
3384 .sysctl_mem = sysctl_tcp_mem,
3385 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3386 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3387 .max_header = MAX_TCP_HEADER,
3388 .obj_size = sizeof(struct tcp_sock),
3389 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3390 .twsk_prot = &tcp_timewait_sock_ops,
3391 .rsk_prot = &tcp_request_sock_ops,
3392 .h.hashinfo = NULL,
3393 .no_autobind = true,
3394 .diag_destroy = tcp_abort,
3395};
3396EXPORT_SYMBOL(tcp_prot);
3397
3398static void __net_exit tcp_sk_exit(struct net *net)
3399{
3400 if (net->ipv4.tcp_congestion_control)
3401 bpf_module_put(net->ipv4.tcp_congestion_control,
3402 net->ipv4.tcp_congestion_control->owner);
3403}
3404
3405static void __net_init tcp_set_hashinfo(struct net *net)
3406{
3407 struct inet_hashinfo *hinfo;
3408 unsigned int ehash_entries;
3409 struct net *old_net;
3410
3411 if (net_eq(net, &init_net))
3412 goto fallback;
3413
3414 old_net = current->nsproxy->net_ns;
3415 ehash_entries = READ_ONCE(old_net->ipv4.sysctl_tcp_child_ehash_entries);
3416 if (!ehash_entries)
3417 goto fallback;
3418
3419 ehash_entries = roundup_pow_of_two(ehash_entries);
3420 hinfo = inet_pernet_hashinfo_alloc(&tcp_hashinfo, ehash_entries);
3421 if (!hinfo) {
3422 pr_warn("Failed to allocate TCP ehash (entries: %u) "
3423 "for a netns, fallback to the global one\n",
3424 ehash_entries);
3425fallback:
3426 hinfo = &tcp_hashinfo;
3427 ehash_entries = tcp_hashinfo.ehash_mask + 1;
3428 }
3429
3430 net->ipv4.tcp_death_row.hashinfo = hinfo;
3431 net->ipv4.tcp_death_row.sysctl_max_tw_buckets = ehash_entries / 2;
3432 net->ipv4.sysctl_max_syn_backlog = max(128U, ehash_entries / 128);
3433}
3434
3435static int __net_init tcp_sk_init(struct net *net)
3436{
3437 net->ipv4.sysctl_tcp_ecn = 2;
3438 net->ipv4.sysctl_tcp_ecn_fallback = 1;
3439
3440 net->ipv4.sysctl_tcp_base_mss = TCP_BASE_MSS;
3441 net->ipv4.sysctl_tcp_min_snd_mss = TCP_MIN_SND_MSS;
3442 net->ipv4.sysctl_tcp_probe_threshold = TCP_PROBE_THRESHOLD;
3443 net->ipv4.sysctl_tcp_probe_interval = TCP_PROBE_INTERVAL;
3444 net->ipv4.sysctl_tcp_mtu_probe_floor = TCP_MIN_SND_MSS;
3445
3446 net->ipv4.sysctl_tcp_keepalive_time = TCP_KEEPALIVE_TIME;
3447 net->ipv4.sysctl_tcp_keepalive_probes = TCP_KEEPALIVE_PROBES;
3448 net->ipv4.sysctl_tcp_keepalive_intvl = TCP_KEEPALIVE_INTVL;
3449
3450 net->ipv4.sysctl_tcp_syn_retries = TCP_SYN_RETRIES;
3451 net->ipv4.sysctl_tcp_synack_retries = TCP_SYNACK_RETRIES;
3452 net->ipv4.sysctl_tcp_syncookies = 1;
3453 net->ipv4.sysctl_tcp_reordering = TCP_FASTRETRANS_THRESH;
3454 net->ipv4.sysctl_tcp_retries1 = TCP_RETR1;
3455 net->ipv4.sysctl_tcp_retries2 = TCP_RETR2;
3456 net->ipv4.sysctl_tcp_orphan_retries = 0;
3457 net->ipv4.sysctl_tcp_fin_timeout = TCP_FIN_TIMEOUT;
3458 net->ipv4.sysctl_tcp_notsent_lowat = UINT_MAX;
3459 net->ipv4.sysctl_tcp_tw_reuse = 2;
3460 net->ipv4.sysctl_tcp_no_ssthresh_metrics_save = 1;
3461
3462 refcount_set(&net->ipv4.tcp_death_row.tw_refcount, 1);
3463 tcp_set_hashinfo(net);
3464
3465 net->ipv4.sysctl_tcp_sack = 1;
3466 net->ipv4.sysctl_tcp_window_scaling = 1;
3467 net->ipv4.sysctl_tcp_timestamps = 1;
3468 net->ipv4.sysctl_tcp_early_retrans = 3;
3469 net->ipv4.sysctl_tcp_recovery = TCP_RACK_LOSS_DETECTION;
3470 net->ipv4.sysctl_tcp_slow_start_after_idle = 1; /* By default, RFC2861 behavior. */
3471 net->ipv4.sysctl_tcp_retrans_collapse = 1;
3472 net->ipv4.sysctl_tcp_max_reordering = 300;
3473 net->ipv4.sysctl_tcp_dsack = 1;
3474 net->ipv4.sysctl_tcp_app_win = 31;
3475 net->ipv4.sysctl_tcp_adv_win_scale = 1;
3476 net->ipv4.sysctl_tcp_frto = 2;
3477 net->ipv4.sysctl_tcp_moderate_rcvbuf = 1;
3478 /* This limits the percentage of the congestion window which we
3479 * will allow a single TSO frame to consume. Building TSO frames
3480 * which are too large can cause TCP streams to be bursty.
3481 */
3482 net->ipv4.sysctl_tcp_tso_win_divisor = 3;
3483 /* Default TSQ limit of 16 TSO segments */
3484 net->ipv4.sysctl_tcp_limit_output_bytes = 16 * 65536;
3485
3486 /* rfc5961 challenge ack rate limiting, per net-ns, disabled by default. */
3487 net->ipv4.sysctl_tcp_challenge_ack_limit = INT_MAX;
3488
3489 net->ipv4.sysctl_tcp_min_tso_segs = 2;
3490 net->ipv4.sysctl_tcp_tso_rtt_log = 9; /* 2^9 = 512 usec */
3491 net->ipv4.sysctl_tcp_min_rtt_wlen = 300;
3492 net->ipv4.sysctl_tcp_autocorking = 1;
3493 net->ipv4.sysctl_tcp_invalid_ratelimit = HZ/2;
3494 net->ipv4.sysctl_tcp_pacing_ss_ratio = 200;
3495 net->ipv4.sysctl_tcp_pacing_ca_ratio = 120;
3496 if (net != &init_net) {
3497 memcpy(net->ipv4.sysctl_tcp_rmem,
3498 init_net.ipv4.sysctl_tcp_rmem,
3499 sizeof(init_net.ipv4.sysctl_tcp_rmem));
3500 memcpy(net->ipv4.sysctl_tcp_wmem,
3501 init_net.ipv4.sysctl_tcp_wmem,
3502 sizeof(init_net.ipv4.sysctl_tcp_wmem));
3503 }
3504 net->ipv4.sysctl_tcp_comp_sack_delay_ns = NSEC_PER_MSEC;
3505 net->ipv4.sysctl_tcp_comp_sack_slack_ns = 100 * NSEC_PER_USEC;
3506 net->ipv4.sysctl_tcp_comp_sack_nr = 44;
3507 net->ipv4.sysctl_tcp_backlog_ack_defer = 1;
3508 net->ipv4.sysctl_tcp_fastopen = TFO_CLIENT_ENABLE;
3509 net->ipv4.sysctl_tcp_fastopen_blackhole_timeout = 0;
3510 atomic_set(&net->ipv4.tfo_active_disable_times, 0);
3511
3512 /* Set default values for PLB */
3513 net->ipv4.sysctl_tcp_plb_enabled = 0; /* Disabled by default */
3514 net->ipv4.sysctl_tcp_plb_idle_rehash_rounds = 3;
3515 net->ipv4.sysctl_tcp_plb_rehash_rounds = 12;
3516 net->ipv4.sysctl_tcp_plb_suspend_rto_sec = 60;
3517 /* Default congestion threshold for PLB to mark a round is 50% */
3518 net->ipv4.sysctl_tcp_plb_cong_thresh = (1 << TCP_PLB_SCALE) / 2;
3519
3520 /* Reno is always built in */
3521 if (!net_eq(net, &init_net) &&
3522 bpf_try_module_get(init_net.ipv4.tcp_congestion_control,
3523 init_net.ipv4.tcp_congestion_control->owner))
3524 net->ipv4.tcp_congestion_control = init_net.ipv4.tcp_congestion_control;
3525 else
3526 net->ipv4.tcp_congestion_control = &tcp_reno;
3527
3528 net->ipv4.sysctl_tcp_syn_linear_timeouts = 4;
3529 net->ipv4.sysctl_tcp_shrink_window = 0;
3530
3531 net->ipv4.sysctl_tcp_pingpong_thresh = 1;
3532 net->ipv4.sysctl_tcp_rto_min_us = jiffies_to_usecs(TCP_RTO_MIN);
3533
3534 return 0;
3535}
3536
3537static void __net_exit tcp_sk_exit_batch(struct list_head *net_exit_list)
3538{
3539 struct net *net;
3540
3541 /* make sure concurrent calls to tcp_sk_exit_batch from net_cleanup_work
3542 * and failed setup_net error unwinding path are serialized.
3543 *
3544 * tcp_twsk_purge() handles twsk in any dead netns, not just those in
3545 * net_exit_list, the thread that dismantles a particular twsk must
3546 * do so without other thread progressing to refcount_dec_and_test() of
3547 * tcp_death_row.tw_refcount.
3548 */
3549 mutex_lock(&tcp_exit_batch_mutex);
3550
3551 tcp_twsk_purge(net_exit_list);
3552
3553 list_for_each_entry(net, net_exit_list, exit_list) {
3554 inet_pernet_hashinfo_free(net->ipv4.tcp_death_row.hashinfo);
3555 WARN_ON_ONCE(!refcount_dec_and_test(&net->ipv4.tcp_death_row.tw_refcount));
3556 tcp_fastopen_ctx_destroy(net);
3557 }
3558
3559 mutex_unlock(&tcp_exit_batch_mutex);
3560}
3561
3562static struct pernet_operations __net_initdata tcp_sk_ops = {
3563 .init = tcp_sk_init,
3564 .exit = tcp_sk_exit,
3565 .exit_batch = tcp_sk_exit_batch,
3566};
3567
3568#if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_PROC_FS)
3569DEFINE_BPF_ITER_FUNC(tcp, struct bpf_iter_meta *meta,
3570 struct sock_common *sk_common, uid_t uid)
3571
3572#define INIT_BATCH_SZ 16
3573
3574static int bpf_iter_init_tcp(void *priv_data, struct bpf_iter_aux_info *aux)
3575{
3576 struct bpf_tcp_iter_state *iter = priv_data;
3577 int err;
3578
3579 err = bpf_iter_init_seq_net(priv_data, aux);
3580 if (err)
3581 return err;
3582
3583 err = bpf_iter_tcp_realloc_batch(iter, INIT_BATCH_SZ);
3584 if (err) {
3585 bpf_iter_fini_seq_net(priv_data);
3586 return err;
3587 }
3588
3589 return 0;
3590}
3591
3592static void bpf_iter_fini_tcp(void *priv_data)
3593{
3594 struct bpf_tcp_iter_state *iter = priv_data;
3595
3596 bpf_iter_fini_seq_net(priv_data);
3597 kvfree(iter->batch);
3598}
3599
3600static const struct bpf_iter_seq_info tcp_seq_info = {
3601 .seq_ops = &bpf_iter_tcp_seq_ops,
3602 .init_seq_private = bpf_iter_init_tcp,
3603 .fini_seq_private = bpf_iter_fini_tcp,
3604 .seq_priv_size = sizeof(struct bpf_tcp_iter_state),
3605};
3606
3607static const struct bpf_func_proto *
3608bpf_iter_tcp_get_func_proto(enum bpf_func_id func_id,
3609 const struct bpf_prog *prog)
3610{
3611 switch (func_id) {
3612 case BPF_FUNC_setsockopt:
3613 return &bpf_sk_setsockopt_proto;
3614 case BPF_FUNC_getsockopt:
3615 return &bpf_sk_getsockopt_proto;
3616 default:
3617 return NULL;
3618 }
3619}
3620
3621static struct bpf_iter_reg tcp_reg_info = {
3622 .target = "tcp",
3623 .ctx_arg_info_size = 1,
3624 .ctx_arg_info = {
3625 { offsetof(struct bpf_iter__tcp, sk_common),
3626 PTR_TO_BTF_ID_OR_NULL | PTR_TRUSTED },
3627 },
3628 .get_func_proto = bpf_iter_tcp_get_func_proto,
3629 .seq_info = &tcp_seq_info,
3630};
3631
3632static void __init bpf_iter_register(void)
3633{
3634 tcp_reg_info.ctx_arg_info[0].btf_id = btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON];
3635 if (bpf_iter_reg_target(&tcp_reg_info))
3636 pr_warn("Warning: could not register bpf iterator tcp\n");
3637}
3638
3639#endif
3640
3641void __init tcp_v4_init(void)
3642{
3643 int cpu, res;
3644
3645 for_each_possible_cpu(cpu) {
3646 struct sock *sk;
3647
3648 res = inet_ctl_sock_create(&sk, PF_INET, SOCK_RAW,
3649 IPPROTO_TCP, &init_net);
3650 if (res)
3651 panic("Failed to create the TCP control socket.\n");
3652 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
3653
3654 /* Please enforce IP_DF and IPID==0 for RST and
3655 * ACK sent in SYN-RECV and TIME-WAIT state.
3656 */
3657 inet_sk(sk)->pmtudisc = IP_PMTUDISC_DO;
3658
3659 sk->sk_clockid = CLOCK_MONOTONIC;
3660
3661 per_cpu(ipv4_tcp_sk.sock, cpu) = sk;
3662 }
3663 if (register_pernet_subsys(&tcp_sk_ops))
3664 panic("Failed to create the TCP control socket.\n");
3665
3666#if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_PROC_FS)
3667 bpf_iter_register();
3668#endif
3669}
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
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.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * IPv4 specific functions
10 *
11 * code split from:
12 * linux/ipv4/tcp.c
13 * linux/ipv4/tcp_input.c
14 * linux/ipv4/tcp_output.c
15 *
16 * See tcp.c for author information
17 */
18
19/*
20 * Changes:
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
29 * ACK bit.
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
40 * coma.
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.
46 */
47
48#define pr_fmt(fmt) "TCP: " fmt
49
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
61#include <net/net_namespace.h>
62#include <net/icmp.h>
63#include <net/inet_hashtables.h>
64#include <net/tcp.h>
65#include <net/transp_v6.h>
66#include <net/ipv6.h>
67#include <net/inet_common.h>
68#include <net/timewait_sock.h>
69#include <net/xfrm.h>
70#include <net/secure_seq.h>
71#include <net/busy_poll.h>
72
73#include <linux/inet.h>
74#include <linux/ipv6.h>
75#include <linux/stddef.h>
76#include <linux/proc_fs.h>
77#include <linux/seq_file.h>
78#include <linux/inetdevice.h>
79#include <linux/btf_ids.h>
80
81#include <crypto/hash.h>
82#include <linux/scatterlist.h>
83
84#include <trace/events/tcp.h>
85
86#ifdef CONFIG_TCP_MD5SIG
87static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
88 __be32 daddr, __be32 saddr, const struct tcphdr *th);
89#endif
90
91struct inet_hashinfo tcp_hashinfo;
92EXPORT_SYMBOL(tcp_hashinfo);
93
94static DEFINE_PER_CPU(struct sock *, ipv4_tcp_sk);
95
96static u32 tcp_v4_init_seq(const struct sk_buff *skb)
97{
98 return secure_tcp_seq(ip_hdr(skb)->daddr,
99 ip_hdr(skb)->saddr,
100 tcp_hdr(skb)->dest,
101 tcp_hdr(skb)->source);
102}
103
104static u32 tcp_v4_init_ts_off(const struct net *net, const struct sk_buff *skb)
105{
106 return secure_tcp_ts_off(net, ip_hdr(skb)->daddr, ip_hdr(skb)->saddr);
107}
108
109int tcp_twsk_unique(struct sock *sk, struct sock *sktw, void *twp)
110{
111 int reuse = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tw_reuse);
112 const struct inet_timewait_sock *tw = inet_twsk(sktw);
113 const struct tcp_timewait_sock *tcptw = tcp_twsk(sktw);
114 struct tcp_sock *tp = tcp_sk(sk);
115
116 if (reuse == 2) {
117 /* Still does not detect *everything* that goes through
118 * lo, since we require a loopback src or dst address
119 * or direct binding to 'lo' interface.
120 */
121 bool loopback = false;
122 if (tw->tw_bound_dev_if == LOOPBACK_IFINDEX)
123 loopback = true;
124#if IS_ENABLED(CONFIG_IPV6)
125 if (tw->tw_family == AF_INET6) {
126 if (ipv6_addr_loopback(&tw->tw_v6_daddr) ||
127 ipv6_addr_v4mapped_loopback(&tw->tw_v6_daddr) ||
128 ipv6_addr_loopback(&tw->tw_v6_rcv_saddr) ||
129 ipv6_addr_v4mapped_loopback(&tw->tw_v6_rcv_saddr))
130 loopback = true;
131 } else
132#endif
133 {
134 if (ipv4_is_loopback(tw->tw_daddr) ||
135 ipv4_is_loopback(tw->tw_rcv_saddr))
136 loopback = true;
137 }
138 if (!loopback)
139 reuse = 0;
140 }
141
142 /* With PAWS, it is safe from the viewpoint
143 of data integrity. Even without PAWS it is safe provided sequence
144 spaces do not overlap i.e. at data rates <= 80Mbit/sec.
145
146 Actually, the idea is close to VJ's one, only timestamp cache is
147 held not per host, but per port pair and TW bucket is used as state
148 holder.
149
150 If TW bucket has been already destroyed we fall back to VJ's scheme
151 and use initial timestamp retrieved from peer table.
152 */
153 if (tcptw->tw_ts_recent_stamp &&
154 (!twp || (reuse && time_after32(ktime_get_seconds(),
155 tcptw->tw_ts_recent_stamp)))) {
156 /* In case of repair and re-using TIME-WAIT sockets we still
157 * want to be sure that it is safe as above but honor the
158 * sequence numbers and time stamps set as part of the repair
159 * process.
160 *
161 * Without this check re-using a TIME-WAIT socket with TCP
162 * repair would accumulate a -1 on the repair assigned
163 * sequence number. The first time it is reused the sequence
164 * is -1, the second time -2, etc. This fixes that issue
165 * without appearing to create any others.
166 */
167 if (likely(!tp->repair)) {
168 u32 seq = tcptw->tw_snd_nxt + 65535 + 2;
169
170 if (!seq)
171 seq = 1;
172 WRITE_ONCE(tp->write_seq, seq);
173 tp->rx_opt.ts_recent = tcptw->tw_ts_recent;
174 tp->rx_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
175 }
176 sock_hold(sktw);
177 return 1;
178 }
179
180 return 0;
181}
182EXPORT_SYMBOL_GPL(tcp_twsk_unique);
183
184static int tcp_v4_pre_connect(struct sock *sk, struct sockaddr *uaddr,
185 int addr_len)
186{
187 /* This check is replicated from tcp_v4_connect() and intended to
188 * prevent BPF program called below from accessing bytes that are out
189 * of the bound specified by user in addr_len.
190 */
191 if (addr_len < sizeof(struct sockaddr_in))
192 return -EINVAL;
193
194 sock_owned_by_me(sk);
195
196 return BPF_CGROUP_RUN_PROG_INET4_CONNECT(sk, uaddr);
197}
198
199/* This will initiate an outgoing connection. */
200int tcp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
201{
202 struct sockaddr_in *usin = (struct sockaddr_in *)uaddr;
203 struct inet_timewait_death_row *tcp_death_row;
204 struct inet_sock *inet = inet_sk(sk);
205 struct tcp_sock *tp = tcp_sk(sk);
206 struct ip_options_rcu *inet_opt;
207 struct net *net = sock_net(sk);
208 __be16 orig_sport, orig_dport;
209 __be32 daddr, nexthop;
210 struct flowi4 *fl4;
211 struct rtable *rt;
212 int err;
213
214 if (addr_len < sizeof(struct sockaddr_in))
215 return -EINVAL;
216
217 if (usin->sin_family != AF_INET)
218 return -EAFNOSUPPORT;
219
220 nexthop = daddr = usin->sin_addr.s_addr;
221 inet_opt = rcu_dereference_protected(inet->inet_opt,
222 lockdep_sock_is_held(sk));
223 if (inet_opt && inet_opt->opt.srr) {
224 if (!daddr)
225 return -EINVAL;
226 nexthop = inet_opt->opt.faddr;
227 }
228
229 orig_sport = inet->inet_sport;
230 orig_dport = usin->sin_port;
231 fl4 = &inet->cork.fl.u.ip4;
232 rt = ip_route_connect(fl4, nexthop, inet->inet_saddr,
233 sk->sk_bound_dev_if, IPPROTO_TCP, orig_sport,
234 orig_dport, sk);
235 if (IS_ERR(rt)) {
236 err = PTR_ERR(rt);
237 if (err == -ENETUNREACH)
238 IP_INC_STATS(net, IPSTATS_MIB_OUTNOROUTES);
239 return err;
240 }
241
242 if (rt->rt_flags & (RTCF_MULTICAST | RTCF_BROADCAST)) {
243 ip_rt_put(rt);
244 return -ENETUNREACH;
245 }
246
247 if (!inet_opt || !inet_opt->opt.srr)
248 daddr = fl4->daddr;
249
250 tcp_death_row = &sock_net(sk)->ipv4.tcp_death_row;
251
252 if (!inet->inet_saddr) {
253 err = inet_bhash2_update_saddr(sk, &fl4->saddr, AF_INET);
254 if (err) {
255 ip_rt_put(rt);
256 return err;
257 }
258 } else {
259 sk_rcv_saddr_set(sk, inet->inet_saddr);
260 }
261
262 if (tp->rx_opt.ts_recent_stamp && inet->inet_daddr != daddr) {
263 /* Reset inherited state */
264 tp->rx_opt.ts_recent = 0;
265 tp->rx_opt.ts_recent_stamp = 0;
266 if (likely(!tp->repair))
267 WRITE_ONCE(tp->write_seq, 0);
268 }
269
270 inet->inet_dport = usin->sin_port;
271 sk_daddr_set(sk, daddr);
272
273 inet_csk(sk)->icsk_ext_hdr_len = 0;
274 if (inet_opt)
275 inet_csk(sk)->icsk_ext_hdr_len = inet_opt->opt.optlen;
276
277 tp->rx_opt.mss_clamp = TCP_MSS_DEFAULT;
278
279 /* Socket identity is still unknown (sport may be zero).
280 * However we set state to SYN-SENT and not releasing socket
281 * lock select source port, enter ourselves into the hash tables and
282 * complete initialization after this.
283 */
284 tcp_set_state(sk, TCP_SYN_SENT);
285 err = inet_hash_connect(tcp_death_row, sk);
286 if (err)
287 goto failure;
288
289 sk_set_txhash(sk);
290
291 rt = ip_route_newports(fl4, rt, orig_sport, orig_dport,
292 inet->inet_sport, inet->inet_dport, sk);
293 if (IS_ERR(rt)) {
294 err = PTR_ERR(rt);
295 rt = NULL;
296 goto failure;
297 }
298 /* OK, now commit destination to socket. */
299 sk->sk_gso_type = SKB_GSO_TCPV4;
300 sk_setup_caps(sk, &rt->dst);
301 rt = NULL;
302
303 if (likely(!tp->repair)) {
304 if (!tp->write_seq)
305 WRITE_ONCE(tp->write_seq,
306 secure_tcp_seq(inet->inet_saddr,
307 inet->inet_daddr,
308 inet->inet_sport,
309 usin->sin_port));
310 tp->tsoffset = secure_tcp_ts_off(net, inet->inet_saddr,
311 inet->inet_daddr);
312 }
313
314 inet->inet_id = get_random_u16();
315
316 if (tcp_fastopen_defer_connect(sk, &err))
317 return err;
318 if (err)
319 goto failure;
320
321 err = tcp_connect(sk);
322
323 if (err)
324 goto failure;
325
326 return 0;
327
328failure:
329 /*
330 * This unhashes the socket and releases the local port,
331 * if necessary.
332 */
333 tcp_set_state(sk, TCP_CLOSE);
334 inet_bhash2_reset_saddr(sk);
335 ip_rt_put(rt);
336 sk->sk_route_caps = 0;
337 inet->inet_dport = 0;
338 return err;
339}
340EXPORT_SYMBOL(tcp_v4_connect);
341
342/*
343 * This routine reacts to ICMP_FRAG_NEEDED mtu indications as defined in RFC1191.
344 * It can be called through tcp_release_cb() if socket was owned by user
345 * at the time tcp_v4_err() was called to handle ICMP message.
346 */
347void tcp_v4_mtu_reduced(struct sock *sk)
348{
349 struct inet_sock *inet = inet_sk(sk);
350 struct dst_entry *dst;
351 u32 mtu;
352
353 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE))
354 return;
355 mtu = READ_ONCE(tcp_sk(sk)->mtu_info);
356 dst = inet_csk_update_pmtu(sk, mtu);
357 if (!dst)
358 return;
359
360 /* Something is about to be wrong... Remember soft error
361 * for the case, if this connection will not able to recover.
362 */
363 if (mtu < dst_mtu(dst) && ip_dont_fragment(sk, dst))
364 sk->sk_err_soft = EMSGSIZE;
365
366 mtu = dst_mtu(dst);
367
368 if (inet->pmtudisc != IP_PMTUDISC_DONT &&
369 ip_sk_accept_pmtu(sk) &&
370 inet_csk(sk)->icsk_pmtu_cookie > mtu) {
371 tcp_sync_mss(sk, mtu);
372
373 /* Resend the TCP packet because it's
374 * clear that the old packet has been
375 * dropped. This is the new "fast" path mtu
376 * discovery.
377 */
378 tcp_simple_retransmit(sk);
379 } /* else let the usual retransmit timer handle it */
380}
381EXPORT_SYMBOL(tcp_v4_mtu_reduced);
382
383static void do_redirect(struct sk_buff *skb, struct sock *sk)
384{
385 struct dst_entry *dst = __sk_dst_check(sk, 0);
386
387 if (dst)
388 dst->ops->redirect(dst, sk, skb);
389}
390
391
392/* handle ICMP messages on TCP_NEW_SYN_RECV request sockets */
393void tcp_req_err(struct sock *sk, u32 seq, bool abort)
394{
395 struct request_sock *req = inet_reqsk(sk);
396 struct net *net = sock_net(sk);
397
398 /* ICMPs are not backlogged, hence we cannot get
399 * an established socket here.
400 */
401 if (seq != tcp_rsk(req)->snt_isn) {
402 __NET_INC_STATS(net, LINUX_MIB_OUTOFWINDOWICMPS);
403 } else if (abort) {
404 /*
405 * Still in SYN_RECV, just remove it silently.
406 * There is no good way to pass the error to the newly
407 * created socket, and POSIX does not want network
408 * errors returned from accept().
409 */
410 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
411 tcp_listendrop(req->rsk_listener);
412 }
413 reqsk_put(req);
414}
415EXPORT_SYMBOL(tcp_req_err);
416
417/* TCP-LD (RFC 6069) logic */
418void tcp_ld_RTO_revert(struct sock *sk, u32 seq)
419{
420 struct inet_connection_sock *icsk = inet_csk(sk);
421 struct tcp_sock *tp = tcp_sk(sk);
422 struct sk_buff *skb;
423 s32 remaining;
424 u32 delta_us;
425
426 if (sock_owned_by_user(sk))
427 return;
428
429 if (seq != tp->snd_una || !icsk->icsk_retransmits ||
430 !icsk->icsk_backoff)
431 return;
432
433 skb = tcp_rtx_queue_head(sk);
434 if (WARN_ON_ONCE(!skb))
435 return;
436
437 icsk->icsk_backoff--;
438 icsk->icsk_rto = tp->srtt_us ? __tcp_set_rto(tp) : TCP_TIMEOUT_INIT;
439 icsk->icsk_rto = inet_csk_rto_backoff(icsk, TCP_RTO_MAX);
440
441 tcp_mstamp_refresh(tp);
442 delta_us = (u32)(tp->tcp_mstamp - tcp_skb_timestamp_us(skb));
443 remaining = icsk->icsk_rto - usecs_to_jiffies(delta_us);
444
445 if (remaining > 0) {
446 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
447 remaining, TCP_RTO_MAX);
448 } else {
449 /* RTO revert clocked out retransmission.
450 * Will retransmit now.
451 */
452 tcp_retransmit_timer(sk);
453 }
454}
455EXPORT_SYMBOL(tcp_ld_RTO_revert);
456
457/*
458 * This routine is called by the ICMP module when it gets some
459 * sort of error condition. If err < 0 then the socket should
460 * be closed and the error returned to the user. If err > 0
461 * it's just the icmp type << 8 | icmp code. After adjustment
462 * header points to the first 8 bytes of the tcp header. We need
463 * to find the appropriate port.
464 *
465 * The locking strategy used here is very "optimistic". When
466 * someone else accesses the socket the ICMP is just dropped
467 * and for some paths there is no check at all.
468 * A more general error queue to queue errors for later handling
469 * is probably better.
470 *
471 */
472
473int tcp_v4_err(struct sk_buff *skb, u32 info)
474{
475 const struct iphdr *iph = (const struct iphdr *)skb->data;
476 struct tcphdr *th = (struct tcphdr *)(skb->data + (iph->ihl << 2));
477 struct tcp_sock *tp;
478 struct inet_sock *inet;
479 const int type = icmp_hdr(skb)->type;
480 const int code = icmp_hdr(skb)->code;
481 struct sock *sk;
482 struct request_sock *fastopen;
483 u32 seq, snd_una;
484 int err;
485 struct net *net = dev_net(skb->dev);
486
487 sk = __inet_lookup_established(net, net->ipv4.tcp_death_row.hashinfo,
488 iph->daddr, th->dest, iph->saddr,
489 ntohs(th->source), inet_iif(skb), 0);
490 if (!sk) {
491 __ICMP_INC_STATS(net, ICMP_MIB_INERRORS);
492 return -ENOENT;
493 }
494 if (sk->sk_state == TCP_TIME_WAIT) {
495 inet_twsk_put(inet_twsk(sk));
496 return 0;
497 }
498 seq = ntohl(th->seq);
499 if (sk->sk_state == TCP_NEW_SYN_RECV) {
500 tcp_req_err(sk, seq, type == ICMP_PARAMETERPROB ||
501 type == ICMP_TIME_EXCEEDED ||
502 (type == ICMP_DEST_UNREACH &&
503 (code == ICMP_NET_UNREACH ||
504 code == ICMP_HOST_UNREACH)));
505 return 0;
506 }
507
508 bh_lock_sock(sk);
509 /* If too many ICMPs get dropped on busy
510 * servers this needs to be solved differently.
511 * We do take care of PMTU discovery (RFC1191) special case :
512 * we can receive locally generated ICMP messages while socket is held.
513 */
514 if (sock_owned_by_user(sk)) {
515 if (!(type == ICMP_DEST_UNREACH && code == ICMP_FRAG_NEEDED))
516 __NET_INC_STATS(net, LINUX_MIB_LOCKDROPPEDICMPS);
517 }
518 if (sk->sk_state == TCP_CLOSE)
519 goto out;
520
521 if (static_branch_unlikely(&ip4_min_ttl)) {
522 /* min_ttl can be changed concurrently from do_ip_setsockopt() */
523 if (unlikely(iph->ttl < READ_ONCE(inet_sk(sk)->min_ttl))) {
524 __NET_INC_STATS(net, LINUX_MIB_TCPMINTTLDROP);
525 goto out;
526 }
527 }
528
529 tp = tcp_sk(sk);
530 /* XXX (TFO) - tp->snd_una should be ISN (tcp_create_openreq_child() */
531 fastopen = rcu_dereference(tp->fastopen_rsk);
532 snd_una = fastopen ? tcp_rsk(fastopen)->snt_isn : tp->snd_una;
533 if (sk->sk_state != TCP_LISTEN &&
534 !between(seq, snd_una, tp->snd_nxt)) {
535 __NET_INC_STATS(net, LINUX_MIB_OUTOFWINDOWICMPS);
536 goto out;
537 }
538
539 switch (type) {
540 case ICMP_REDIRECT:
541 if (!sock_owned_by_user(sk))
542 do_redirect(skb, sk);
543 goto out;
544 case ICMP_SOURCE_QUENCH:
545 /* Just silently ignore these. */
546 goto out;
547 case ICMP_PARAMETERPROB:
548 err = EPROTO;
549 break;
550 case ICMP_DEST_UNREACH:
551 if (code > NR_ICMP_UNREACH)
552 goto out;
553
554 if (code == ICMP_FRAG_NEEDED) { /* PMTU discovery (RFC1191) */
555 /* We are not interested in TCP_LISTEN and open_requests
556 * (SYN-ACKs send out by Linux are always <576bytes so
557 * they should go through unfragmented).
558 */
559 if (sk->sk_state == TCP_LISTEN)
560 goto out;
561
562 WRITE_ONCE(tp->mtu_info, info);
563 if (!sock_owned_by_user(sk)) {
564 tcp_v4_mtu_reduced(sk);
565 } else {
566 if (!test_and_set_bit(TCP_MTU_REDUCED_DEFERRED, &sk->sk_tsq_flags))
567 sock_hold(sk);
568 }
569 goto out;
570 }
571
572 err = icmp_err_convert[code].errno;
573 /* check if this ICMP message allows revert of backoff.
574 * (see RFC 6069)
575 */
576 if (!fastopen &&
577 (code == ICMP_NET_UNREACH || code == ICMP_HOST_UNREACH))
578 tcp_ld_RTO_revert(sk, seq);
579 break;
580 case ICMP_TIME_EXCEEDED:
581 err = EHOSTUNREACH;
582 break;
583 default:
584 goto out;
585 }
586
587 switch (sk->sk_state) {
588 case TCP_SYN_SENT:
589 case TCP_SYN_RECV:
590 /* Only in fast or simultaneous open. If a fast open socket is
591 * already accepted it is treated as a connected one below.
592 */
593 if (fastopen && !fastopen->sk)
594 break;
595
596 ip_icmp_error(sk, skb, err, th->dest, info, (u8 *)th);
597
598 if (!sock_owned_by_user(sk)) {
599 sk->sk_err = err;
600
601 sk_error_report(sk);
602
603 tcp_done(sk);
604 } else {
605 sk->sk_err_soft = err;
606 }
607 goto out;
608 }
609
610 /* If we've already connected we will keep trying
611 * until we time out, or the user gives up.
612 *
613 * rfc1122 4.2.3.9 allows to consider as hard errors
614 * only PROTO_UNREACH and PORT_UNREACH (well, FRAG_FAILED too,
615 * but it is obsoleted by pmtu discovery).
616 *
617 * Note, that in modern internet, where routing is unreliable
618 * and in each dark corner broken firewalls sit, sending random
619 * errors ordered by their masters even this two messages finally lose
620 * their original sense (even Linux sends invalid PORT_UNREACHs)
621 *
622 * Now we are in compliance with RFCs.
623 * --ANK (980905)
624 */
625
626 inet = inet_sk(sk);
627 if (!sock_owned_by_user(sk) && inet->recverr) {
628 sk->sk_err = err;
629 sk_error_report(sk);
630 } else { /* Only an error on timeout */
631 sk->sk_err_soft = err;
632 }
633
634out:
635 bh_unlock_sock(sk);
636 sock_put(sk);
637 return 0;
638}
639
640void __tcp_v4_send_check(struct sk_buff *skb, __be32 saddr, __be32 daddr)
641{
642 struct tcphdr *th = tcp_hdr(skb);
643
644 th->check = ~tcp_v4_check(skb->len, saddr, daddr, 0);
645 skb->csum_start = skb_transport_header(skb) - skb->head;
646 skb->csum_offset = offsetof(struct tcphdr, check);
647}
648
649/* This routine computes an IPv4 TCP checksum. */
650void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)
651{
652 const struct inet_sock *inet = inet_sk(sk);
653
654 __tcp_v4_send_check(skb, inet->inet_saddr, inet->inet_daddr);
655}
656EXPORT_SYMBOL(tcp_v4_send_check);
657
658/*
659 * This routine will send an RST to the other tcp.
660 *
661 * Someone asks: why I NEVER use socket parameters (TOS, TTL etc.)
662 * for reset.
663 * Answer: if a packet caused RST, it is not for a socket
664 * existing in our system, if it is matched to a socket,
665 * it is just duplicate segment or bug in other side's TCP.
666 * So that we build reply only basing on parameters
667 * arrived with segment.
668 * Exception: precedence violation. We do not implement it in any case.
669 */
670
671#ifdef CONFIG_TCP_MD5SIG
672#define OPTION_BYTES TCPOLEN_MD5SIG_ALIGNED
673#else
674#define OPTION_BYTES sizeof(__be32)
675#endif
676
677static void tcp_v4_send_reset(const struct sock *sk, struct sk_buff *skb)
678{
679 const struct tcphdr *th = tcp_hdr(skb);
680 struct {
681 struct tcphdr th;
682 __be32 opt[OPTION_BYTES / sizeof(__be32)];
683 } rep;
684 struct ip_reply_arg arg;
685#ifdef CONFIG_TCP_MD5SIG
686 struct tcp_md5sig_key *key = NULL;
687 const __u8 *hash_location = NULL;
688 unsigned char newhash[16];
689 int genhash;
690 struct sock *sk1 = NULL;
691#endif
692 u64 transmit_time = 0;
693 struct sock *ctl_sk;
694 struct net *net;
695
696 /* Never send a reset in response to a reset. */
697 if (th->rst)
698 return;
699
700 /* If sk not NULL, it means we did a successful lookup and incoming
701 * route had to be correct. prequeue might have dropped our dst.
702 */
703 if (!sk && skb_rtable(skb)->rt_type != RTN_LOCAL)
704 return;
705
706 /* Swap the send and the receive. */
707 memset(&rep, 0, sizeof(rep));
708 rep.th.dest = th->source;
709 rep.th.source = th->dest;
710 rep.th.doff = sizeof(struct tcphdr) / 4;
711 rep.th.rst = 1;
712
713 if (th->ack) {
714 rep.th.seq = th->ack_seq;
715 } else {
716 rep.th.ack = 1;
717 rep.th.ack_seq = htonl(ntohl(th->seq) + th->syn + th->fin +
718 skb->len - (th->doff << 2));
719 }
720
721 memset(&arg, 0, sizeof(arg));
722 arg.iov[0].iov_base = (unsigned char *)&rep;
723 arg.iov[0].iov_len = sizeof(rep.th);
724
725 net = sk ? sock_net(sk) : dev_net(skb_dst(skb)->dev);
726#ifdef CONFIG_TCP_MD5SIG
727 rcu_read_lock();
728 hash_location = tcp_parse_md5sig_option(th);
729 if (sk && sk_fullsock(sk)) {
730 const union tcp_md5_addr *addr;
731 int l3index;
732
733 /* sdif set, means packet ingressed via a device
734 * in an L3 domain and inet_iif is set to it.
735 */
736 l3index = tcp_v4_sdif(skb) ? inet_iif(skb) : 0;
737 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
738 key = tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
739 } else if (hash_location) {
740 const union tcp_md5_addr *addr;
741 int sdif = tcp_v4_sdif(skb);
742 int dif = inet_iif(skb);
743 int l3index;
744
745 /*
746 * active side is lost. Try to find listening socket through
747 * source port, and then find md5 key through listening socket.
748 * we are not loose security here:
749 * Incoming packet is checked with md5 hash with finding key,
750 * no RST generated if md5 hash doesn't match.
751 */
752 sk1 = __inet_lookup_listener(net, net->ipv4.tcp_death_row.hashinfo,
753 NULL, 0, ip_hdr(skb)->saddr,
754 th->source, ip_hdr(skb)->daddr,
755 ntohs(th->source), dif, sdif);
756 /* don't send rst if it can't find key */
757 if (!sk1)
758 goto out;
759
760 /* sdif set, means packet ingressed via a device
761 * in an L3 domain and dif is set to it.
762 */
763 l3index = sdif ? dif : 0;
764 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
765 key = tcp_md5_do_lookup(sk1, l3index, addr, AF_INET);
766 if (!key)
767 goto out;
768
769
770 genhash = tcp_v4_md5_hash_skb(newhash, key, NULL, skb);
771 if (genhash || memcmp(hash_location, newhash, 16) != 0)
772 goto out;
773
774 }
775
776 if (key) {
777 rep.opt[0] = htonl((TCPOPT_NOP << 24) |
778 (TCPOPT_NOP << 16) |
779 (TCPOPT_MD5SIG << 8) |
780 TCPOLEN_MD5SIG);
781 /* Update length and the length the header thinks exists */
782 arg.iov[0].iov_len += TCPOLEN_MD5SIG_ALIGNED;
783 rep.th.doff = arg.iov[0].iov_len / 4;
784
785 tcp_v4_md5_hash_hdr((__u8 *) &rep.opt[1],
786 key, ip_hdr(skb)->saddr,
787 ip_hdr(skb)->daddr, &rep.th);
788 }
789#endif
790 /* Can't co-exist with TCPMD5, hence check rep.opt[0] */
791 if (rep.opt[0] == 0) {
792 __be32 mrst = mptcp_reset_option(skb);
793
794 if (mrst) {
795 rep.opt[0] = mrst;
796 arg.iov[0].iov_len += sizeof(mrst);
797 rep.th.doff = arg.iov[0].iov_len / 4;
798 }
799 }
800
801 arg.csum = csum_tcpudp_nofold(ip_hdr(skb)->daddr,
802 ip_hdr(skb)->saddr, /* XXX */
803 arg.iov[0].iov_len, IPPROTO_TCP, 0);
804 arg.csumoffset = offsetof(struct tcphdr, check) / 2;
805 arg.flags = (sk && inet_sk_transparent(sk)) ? IP_REPLY_ARG_NOSRCCHECK : 0;
806
807 /* When socket is gone, all binding information is lost.
808 * routing might fail in this case. No choice here, if we choose to force
809 * input interface, we will misroute in case of asymmetric route.
810 */
811 if (sk) {
812 arg.bound_dev_if = sk->sk_bound_dev_if;
813 if (sk_fullsock(sk))
814 trace_tcp_send_reset(sk, skb);
815 }
816
817 BUILD_BUG_ON(offsetof(struct sock, sk_bound_dev_if) !=
818 offsetof(struct inet_timewait_sock, tw_bound_dev_if));
819
820 arg.tos = ip_hdr(skb)->tos;
821 arg.uid = sock_net_uid(net, sk && sk_fullsock(sk) ? sk : NULL);
822 local_bh_disable();
823 ctl_sk = this_cpu_read(ipv4_tcp_sk);
824 sock_net_set(ctl_sk, net);
825 if (sk) {
826 ctl_sk->sk_mark = (sk->sk_state == TCP_TIME_WAIT) ?
827 inet_twsk(sk)->tw_mark : sk->sk_mark;
828 ctl_sk->sk_priority = (sk->sk_state == TCP_TIME_WAIT) ?
829 inet_twsk(sk)->tw_priority : sk->sk_priority;
830 transmit_time = tcp_transmit_time(sk);
831 xfrm_sk_clone_policy(ctl_sk, sk);
832 }
833 ip_send_unicast_reply(ctl_sk,
834 skb, &TCP_SKB_CB(skb)->header.h4.opt,
835 ip_hdr(skb)->saddr, ip_hdr(skb)->daddr,
836 &arg, arg.iov[0].iov_len,
837 transmit_time);
838
839 ctl_sk->sk_mark = 0;
840 xfrm_sk_free_policy(ctl_sk);
841 sock_net_set(ctl_sk, &init_net);
842 __TCP_INC_STATS(net, TCP_MIB_OUTSEGS);
843 __TCP_INC_STATS(net, TCP_MIB_OUTRSTS);
844 local_bh_enable();
845
846#ifdef CONFIG_TCP_MD5SIG
847out:
848 rcu_read_unlock();
849#endif
850}
851
852/* The code following below sending ACKs in SYN-RECV and TIME-WAIT states
853 outside socket context is ugly, certainly. What can I do?
854 */
855
856static void tcp_v4_send_ack(const struct sock *sk,
857 struct sk_buff *skb, u32 seq, u32 ack,
858 u32 win, u32 tsval, u32 tsecr, int oif,
859 struct tcp_md5sig_key *key,
860 int reply_flags, u8 tos)
861{
862 const struct tcphdr *th = tcp_hdr(skb);
863 struct {
864 struct tcphdr th;
865 __be32 opt[(TCPOLEN_TSTAMP_ALIGNED >> 2)
866#ifdef CONFIG_TCP_MD5SIG
867 + (TCPOLEN_MD5SIG_ALIGNED >> 2)
868#endif
869 ];
870 } rep;
871 struct net *net = sock_net(sk);
872 struct ip_reply_arg arg;
873 struct sock *ctl_sk;
874 u64 transmit_time;
875
876 memset(&rep.th, 0, sizeof(struct tcphdr));
877 memset(&arg, 0, sizeof(arg));
878
879 arg.iov[0].iov_base = (unsigned char *)&rep;
880 arg.iov[0].iov_len = sizeof(rep.th);
881 if (tsecr) {
882 rep.opt[0] = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
883 (TCPOPT_TIMESTAMP << 8) |
884 TCPOLEN_TIMESTAMP);
885 rep.opt[1] = htonl(tsval);
886 rep.opt[2] = htonl(tsecr);
887 arg.iov[0].iov_len += TCPOLEN_TSTAMP_ALIGNED;
888 }
889
890 /* Swap the send and the receive. */
891 rep.th.dest = th->source;
892 rep.th.source = th->dest;
893 rep.th.doff = arg.iov[0].iov_len / 4;
894 rep.th.seq = htonl(seq);
895 rep.th.ack_seq = htonl(ack);
896 rep.th.ack = 1;
897 rep.th.window = htons(win);
898
899#ifdef CONFIG_TCP_MD5SIG
900 if (key) {
901 int offset = (tsecr) ? 3 : 0;
902
903 rep.opt[offset++] = htonl((TCPOPT_NOP << 24) |
904 (TCPOPT_NOP << 16) |
905 (TCPOPT_MD5SIG << 8) |
906 TCPOLEN_MD5SIG);
907 arg.iov[0].iov_len += TCPOLEN_MD5SIG_ALIGNED;
908 rep.th.doff = arg.iov[0].iov_len/4;
909
910 tcp_v4_md5_hash_hdr((__u8 *) &rep.opt[offset],
911 key, ip_hdr(skb)->saddr,
912 ip_hdr(skb)->daddr, &rep.th);
913 }
914#endif
915 arg.flags = reply_flags;
916 arg.csum = csum_tcpudp_nofold(ip_hdr(skb)->daddr,
917 ip_hdr(skb)->saddr, /* XXX */
918 arg.iov[0].iov_len, IPPROTO_TCP, 0);
919 arg.csumoffset = offsetof(struct tcphdr, check) / 2;
920 if (oif)
921 arg.bound_dev_if = oif;
922 arg.tos = tos;
923 arg.uid = sock_net_uid(net, sk_fullsock(sk) ? sk : NULL);
924 local_bh_disable();
925 ctl_sk = this_cpu_read(ipv4_tcp_sk);
926 sock_net_set(ctl_sk, net);
927 ctl_sk->sk_mark = (sk->sk_state == TCP_TIME_WAIT) ?
928 inet_twsk(sk)->tw_mark : sk->sk_mark;
929 ctl_sk->sk_priority = (sk->sk_state == TCP_TIME_WAIT) ?
930 inet_twsk(sk)->tw_priority : sk->sk_priority;
931 transmit_time = tcp_transmit_time(sk);
932 ip_send_unicast_reply(ctl_sk,
933 skb, &TCP_SKB_CB(skb)->header.h4.opt,
934 ip_hdr(skb)->saddr, ip_hdr(skb)->daddr,
935 &arg, arg.iov[0].iov_len,
936 transmit_time);
937
938 ctl_sk->sk_mark = 0;
939 sock_net_set(ctl_sk, &init_net);
940 __TCP_INC_STATS(net, TCP_MIB_OUTSEGS);
941 local_bh_enable();
942}
943
944static void tcp_v4_timewait_ack(struct sock *sk, struct sk_buff *skb)
945{
946 struct inet_timewait_sock *tw = inet_twsk(sk);
947 struct tcp_timewait_sock *tcptw = tcp_twsk(sk);
948
949 tcp_v4_send_ack(sk, skb,
950 tcptw->tw_snd_nxt, tcptw->tw_rcv_nxt,
951 tcptw->tw_rcv_wnd >> tw->tw_rcv_wscale,
952 tcp_time_stamp_raw() + tcptw->tw_ts_offset,
953 tcptw->tw_ts_recent,
954 tw->tw_bound_dev_if,
955 tcp_twsk_md5_key(tcptw),
956 tw->tw_transparent ? IP_REPLY_ARG_NOSRCCHECK : 0,
957 tw->tw_tos
958 );
959
960 inet_twsk_put(tw);
961}
962
963static void tcp_v4_reqsk_send_ack(const struct sock *sk, struct sk_buff *skb,
964 struct request_sock *req)
965{
966 const union tcp_md5_addr *addr;
967 int l3index;
968
969 /* sk->sk_state == TCP_LISTEN -> for regular TCP_SYN_RECV
970 * sk->sk_state == TCP_SYN_RECV -> for Fast Open.
971 */
972 u32 seq = (sk->sk_state == TCP_LISTEN) ? tcp_rsk(req)->snt_isn + 1 :
973 tcp_sk(sk)->snd_nxt;
974
975 /* RFC 7323 2.3
976 * The window field (SEG.WND) of every outgoing segment, with the
977 * exception of <SYN> segments, MUST be right-shifted by
978 * Rcv.Wind.Shift bits:
979 */
980 addr = (union tcp_md5_addr *)&ip_hdr(skb)->saddr;
981 l3index = tcp_v4_sdif(skb) ? inet_iif(skb) : 0;
982 tcp_v4_send_ack(sk, skb, seq,
983 tcp_rsk(req)->rcv_nxt,
984 req->rsk_rcv_wnd >> inet_rsk(req)->rcv_wscale,
985 tcp_time_stamp_raw() + tcp_rsk(req)->ts_off,
986 req->ts_recent,
987 0,
988 tcp_md5_do_lookup(sk, l3index, addr, AF_INET),
989 inet_rsk(req)->no_srccheck ? IP_REPLY_ARG_NOSRCCHECK : 0,
990 ip_hdr(skb)->tos);
991}
992
993/*
994 * Send a SYN-ACK after having received a SYN.
995 * This still operates on a request_sock only, not on a big
996 * socket.
997 */
998static int tcp_v4_send_synack(const struct sock *sk, struct dst_entry *dst,
999 struct flowi *fl,
1000 struct request_sock *req,
1001 struct tcp_fastopen_cookie *foc,
1002 enum tcp_synack_type synack_type,
1003 struct sk_buff *syn_skb)
1004{
1005 const struct inet_request_sock *ireq = inet_rsk(req);
1006 struct flowi4 fl4;
1007 int err = -1;
1008 struct sk_buff *skb;
1009 u8 tos;
1010
1011 /* First, grab a route. */
1012 if (!dst && (dst = inet_csk_route_req(sk, &fl4, req)) == NULL)
1013 return -1;
1014
1015 skb = tcp_make_synack(sk, dst, req, foc, synack_type, syn_skb);
1016
1017 if (skb) {
1018 __tcp_v4_send_check(skb, ireq->ir_loc_addr, ireq->ir_rmt_addr);
1019
1020 tos = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reflect_tos) ?
1021 (tcp_rsk(req)->syn_tos & ~INET_ECN_MASK) |
1022 (inet_sk(sk)->tos & INET_ECN_MASK) :
1023 inet_sk(sk)->tos;
1024
1025 if (!INET_ECN_is_capable(tos) &&
1026 tcp_bpf_ca_needs_ecn((struct sock *)req))
1027 tos |= INET_ECN_ECT_0;
1028
1029 rcu_read_lock();
1030 err = ip_build_and_send_pkt(skb, sk, ireq->ir_loc_addr,
1031 ireq->ir_rmt_addr,
1032 rcu_dereference(ireq->ireq_opt),
1033 tos);
1034 rcu_read_unlock();
1035 err = net_xmit_eval(err);
1036 }
1037
1038 return err;
1039}
1040
1041/*
1042 * IPv4 request_sock destructor.
1043 */
1044static void tcp_v4_reqsk_destructor(struct request_sock *req)
1045{
1046 kfree(rcu_dereference_protected(inet_rsk(req)->ireq_opt, 1));
1047}
1048
1049#ifdef CONFIG_TCP_MD5SIG
1050/*
1051 * RFC2385 MD5 checksumming requires a mapping of
1052 * IP address->MD5 Key.
1053 * We need to maintain these in the sk structure.
1054 */
1055
1056DEFINE_STATIC_KEY_DEFERRED_FALSE(tcp_md5_needed, HZ);
1057EXPORT_SYMBOL(tcp_md5_needed);
1058
1059static bool better_md5_match(struct tcp_md5sig_key *old, struct tcp_md5sig_key *new)
1060{
1061 if (!old)
1062 return true;
1063
1064 /* l3index always overrides non-l3index */
1065 if (old->l3index && new->l3index == 0)
1066 return false;
1067 if (old->l3index == 0 && new->l3index)
1068 return true;
1069
1070 return old->prefixlen < new->prefixlen;
1071}
1072
1073/* Find the Key structure for an address. */
1074struct tcp_md5sig_key *__tcp_md5_do_lookup(const struct sock *sk, int l3index,
1075 const union tcp_md5_addr *addr,
1076 int family)
1077{
1078 const struct tcp_sock *tp = tcp_sk(sk);
1079 struct tcp_md5sig_key *key;
1080 const struct tcp_md5sig_info *md5sig;
1081 __be32 mask;
1082 struct tcp_md5sig_key *best_match = NULL;
1083 bool match;
1084
1085 /* caller either holds rcu_read_lock() or socket lock */
1086 md5sig = rcu_dereference_check(tp->md5sig_info,
1087 lockdep_sock_is_held(sk));
1088 if (!md5sig)
1089 return NULL;
1090
1091 hlist_for_each_entry_rcu(key, &md5sig->head, node,
1092 lockdep_sock_is_held(sk)) {
1093 if (key->family != family)
1094 continue;
1095 if (key->flags & TCP_MD5SIG_FLAG_IFINDEX && key->l3index != l3index)
1096 continue;
1097 if (family == AF_INET) {
1098 mask = inet_make_mask(key->prefixlen);
1099 match = (key->addr.a4.s_addr & mask) ==
1100 (addr->a4.s_addr & mask);
1101#if IS_ENABLED(CONFIG_IPV6)
1102 } else if (family == AF_INET6) {
1103 match = ipv6_prefix_equal(&key->addr.a6, &addr->a6,
1104 key->prefixlen);
1105#endif
1106 } else {
1107 match = false;
1108 }
1109
1110 if (match && better_md5_match(best_match, key))
1111 best_match = key;
1112 }
1113 return best_match;
1114}
1115EXPORT_SYMBOL(__tcp_md5_do_lookup);
1116
1117static struct tcp_md5sig_key *tcp_md5_do_lookup_exact(const struct sock *sk,
1118 const union tcp_md5_addr *addr,
1119 int family, u8 prefixlen,
1120 int l3index, u8 flags)
1121{
1122 const struct tcp_sock *tp = tcp_sk(sk);
1123 struct tcp_md5sig_key *key;
1124 unsigned int size = sizeof(struct in_addr);
1125 const struct tcp_md5sig_info *md5sig;
1126
1127 /* caller either holds rcu_read_lock() or socket lock */
1128 md5sig = rcu_dereference_check(tp->md5sig_info,
1129 lockdep_sock_is_held(sk));
1130 if (!md5sig)
1131 return NULL;
1132#if IS_ENABLED(CONFIG_IPV6)
1133 if (family == AF_INET6)
1134 size = sizeof(struct in6_addr);
1135#endif
1136 hlist_for_each_entry_rcu(key, &md5sig->head, node,
1137 lockdep_sock_is_held(sk)) {
1138 if (key->family != family)
1139 continue;
1140 if ((key->flags & TCP_MD5SIG_FLAG_IFINDEX) != (flags & TCP_MD5SIG_FLAG_IFINDEX))
1141 continue;
1142 if (key->l3index != l3index)
1143 continue;
1144 if (!memcmp(&key->addr, addr, size) &&
1145 key->prefixlen == prefixlen)
1146 return key;
1147 }
1148 return NULL;
1149}
1150
1151struct tcp_md5sig_key *tcp_v4_md5_lookup(const struct sock *sk,
1152 const struct sock *addr_sk)
1153{
1154 const union tcp_md5_addr *addr;
1155 int l3index;
1156
1157 l3index = l3mdev_master_ifindex_by_index(sock_net(sk),
1158 addr_sk->sk_bound_dev_if);
1159 addr = (const union tcp_md5_addr *)&addr_sk->sk_daddr;
1160 return tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
1161}
1162EXPORT_SYMBOL(tcp_v4_md5_lookup);
1163
1164static int tcp_md5sig_info_add(struct sock *sk, gfp_t gfp)
1165{
1166 struct tcp_sock *tp = tcp_sk(sk);
1167 struct tcp_md5sig_info *md5sig;
1168
1169 md5sig = kmalloc(sizeof(*md5sig), gfp);
1170 if (!md5sig)
1171 return -ENOMEM;
1172
1173 sk_gso_disable(sk);
1174 INIT_HLIST_HEAD(&md5sig->head);
1175 rcu_assign_pointer(tp->md5sig_info, md5sig);
1176 return 0;
1177}
1178
1179/* This can be called on a newly created socket, from other files */
1180static int __tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
1181 int family, u8 prefixlen, int l3index, u8 flags,
1182 const u8 *newkey, u8 newkeylen, gfp_t gfp)
1183{
1184 /* Add Key to the list */
1185 struct tcp_md5sig_key *key;
1186 struct tcp_sock *tp = tcp_sk(sk);
1187 struct tcp_md5sig_info *md5sig;
1188
1189 key = tcp_md5_do_lookup_exact(sk, addr, family, prefixlen, l3index, flags);
1190 if (key) {
1191 /* Pre-existing entry - just update that one.
1192 * Note that the key might be used concurrently.
1193 * data_race() is telling kcsan that we do not care of
1194 * key mismatches, since changing MD5 key on live flows
1195 * can lead to packet drops.
1196 */
1197 data_race(memcpy(key->key, newkey, newkeylen));
1198
1199 /* Pairs with READ_ONCE() in tcp_md5_hash_key().
1200 * Also note that a reader could catch new key->keylen value
1201 * but old key->key[], this is the reason we use __GFP_ZERO
1202 * at sock_kmalloc() time below these lines.
1203 */
1204 WRITE_ONCE(key->keylen, newkeylen);
1205
1206 return 0;
1207 }
1208
1209 md5sig = rcu_dereference_protected(tp->md5sig_info,
1210 lockdep_sock_is_held(sk));
1211
1212 key = sock_kmalloc(sk, sizeof(*key), gfp | __GFP_ZERO);
1213 if (!key)
1214 return -ENOMEM;
1215 if (!tcp_alloc_md5sig_pool()) {
1216 sock_kfree_s(sk, key, sizeof(*key));
1217 return -ENOMEM;
1218 }
1219
1220 memcpy(key->key, newkey, newkeylen);
1221 key->keylen = newkeylen;
1222 key->family = family;
1223 key->prefixlen = prefixlen;
1224 key->l3index = l3index;
1225 key->flags = flags;
1226 memcpy(&key->addr, addr,
1227 (IS_ENABLED(CONFIG_IPV6) && family == AF_INET6) ? sizeof(struct in6_addr) :
1228 sizeof(struct in_addr));
1229 hlist_add_head_rcu(&key->node, &md5sig->head);
1230 return 0;
1231}
1232
1233int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
1234 int family, u8 prefixlen, int l3index, u8 flags,
1235 const u8 *newkey, u8 newkeylen)
1236{
1237 struct tcp_sock *tp = tcp_sk(sk);
1238
1239 if (!rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk))) {
1240 if (tcp_md5sig_info_add(sk, GFP_KERNEL))
1241 return -ENOMEM;
1242
1243 if (!static_branch_inc(&tcp_md5_needed.key)) {
1244 struct tcp_md5sig_info *md5sig;
1245
1246 md5sig = rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk));
1247 rcu_assign_pointer(tp->md5sig_info, NULL);
1248 kfree_rcu(md5sig, rcu);
1249 return -EUSERS;
1250 }
1251 }
1252
1253 return __tcp_md5_do_add(sk, addr, family, prefixlen, l3index, flags,
1254 newkey, newkeylen, GFP_KERNEL);
1255}
1256EXPORT_SYMBOL(tcp_md5_do_add);
1257
1258int tcp_md5_key_copy(struct sock *sk, const union tcp_md5_addr *addr,
1259 int family, u8 prefixlen, int l3index,
1260 struct tcp_md5sig_key *key)
1261{
1262 struct tcp_sock *tp = tcp_sk(sk);
1263
1264 if (!rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk))) {
1265 if (tcp_md5sig_info_add(sk, sk_gfp_mask(sk, GFP_ATOMIC)))
1266 return -ENOMEM;
1267
1268 if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key)) {
1269 struct tcp_md5sig_info *md5sig;
1270
1271 md5sig = rcu_dereference_protected(tp->md5sig_info, lockdep_sock_is_held(sk));
1272 net_warn_ratelimited("Too many TCP-MD5 keys in the system\n");
1273 rcu_assign_pointer(tp->md5sig_info, NULL);
1274 kfree_rcu(md5sig, rcu);
1275 return -EUSERS;
1276 }
1277 }
1278
1279 return __tcp_md5_do_add(sk, addr, family, prefixlen, l3index,
1280 key->flags, key->key, key->keylen,
1281 sk_gfp_mask(sk, GFP_ATOMIC));
1282}
1283EXPORT_SYMBOL(tcp_md5_key_copy);
1284
1285int tcp_md5_do_del(struct sock *sk, const union tcp_md5_addr *addr, int family,
1286 u8 prefixlen, int l3index, u8 flags)
1287{
1288 struct tcp_md5sig_key *key;
1289
1290 key = tcp_md5_do_lookup_exact(sk, addr, family, prefixlen, l3index, flags);
1291 if (!key)
1292 return -ENOENT;
1293 hlist_del_rcu(&key->node);
1294 atomic_sub(sizeof(*key), &sk->sk_omem_alloc);
1295 kfree_rcu(key, rcu);
1296 return 0;
1297}
1298EXPORT_SYMBOL(tcp_md5_do_del);
1299
1300static void tcp_clear_md5_list(struct sock *sk)
1301{
1302 struct tcp_sock *tp = tcp_sk(sk);
1303 struct tcp_md5sig_key *key;
1304 struct hlist_node *n;
1305 struct tcp_md5sig_info *md5sig;
1306
1307 md5sig = rcu_dereference_protected(tp->md5sig_info, 1);
1308
1309 hlist_for_each_entry_safe(key, n, &md5sig->head, node) {
1310 hlist_del_rcu(&key->node);
1311 atomic_sub(sizeof(*key), &sk->sk_omem_alloc);
1312 kfree_rcu(key, rcu);
1313 }
1314}
1315
1316static int tcp_v4_parse_md5_keys(struct sock *sk, int optname,
1317 sockptr_t optval, int optlen)
1318{
1319 struct tcp_md5sig cmd;
1320 struct sockaddr_in *sin = (struct sockaddr_in *)&cmd.tcpm_addr;
1321 const union tcp_md5_addr *addr;
1322 u8 prefixlen = 32;
1323 int l3index = 0;
1324 u8 flags;
1325
1326 if (optlen < sizeof(cmd))
1327 return -EINVAL;
1328
1329 if (copy_from_sockptr(&cmd, optval, sizeof(cmd)))
1330 return -EFAULT;
1331
1332 if (sin->sin_family != AF_INET)
1333 return -EINVAL;
1334
1335 flags = cmd.tcpm_flags & TCP_MD5SIG_FLAG_IFINDEX;
1336
1337 if (optname == TCP_MD5SIG_EXT &&
1338 cmd.tcpm_flags & TCP_MD5SIG_FLAG_PREFIX) {
1339 prefixlen = cmd.tcpm_prefixlen;
1340 if (prefixlen > 32)
1341 return -EINVAL;
1342 }
1343
1344 if (optname == TCP_MD5SIG_EXT && cmd.tcpm_ifindex &&
1345 cmd.tcpm_flags & TCP_MD5SIG_FLAG_IFINDEX) {
1346 struct net_device *dev;
1347
1348 rcu_read_lock();
1349 dev = dev_get_by_index_rcu(sock_net(sk), cmd.tcpm_ifindex);
1350 if (dev && netif_is_l3_master(dev))
1351 l3index = dev->ifindex;
1352
1353 rcu_read_unlock();
1354
1355 /* ok to reference set/not set outside of rcu;
1356 * right now device MUST be an L3 master
1357 */
1358 if (!dev || !l3index)
1359 return -EINVAL;
1360 }
1361
1362 addr = (union tcp_md5_addr *)&sin->sin_addr.s_addr;
1363
1364 if (!cmd.tcpm_keylen)
1365 return tcp_md5_do_del(sk, addr, AF_INET, prefixlen, l3index, flags);
1366
1367 if (cmd.tcpm_keylen > TCP_MD5SIG_MAXKEYLEN)
1368 return -EINVAL;
1369
1370 return tcp_md5_do_add(sk, addr, AF_INET, prefixlen, l3index, flags,
1371 cmd.tcpm_key, cmd.tcpm_keylen);
1372}
1373
1374static int tcp_v4_md5_hash_headers(struct tcp_md5sig_pool *hp,
1375 __be32 daddr, __be32 saddr,
1376 const struct tcphdr *th, int nbytes)
1377{
1378 struct tcp4_pseudohdr *bp;
1379 struct scatterlist sg;
1380 struct tcphdr *_th;
1381
1382 bp = hp->scratch;
1383 bp->saddr = saddr;
1384 bp->daddr = daddr;
1385 bp->pad = 0;
1386 bp->protocol = IPPROTO_TCP;
1387 bp->len = cpu_to_be16(nbytes);
1388
1389 _th = (struct tcphdr *)(bp + 1);
1390 memcpy(_th, th, sizeof(*th));
1391 _th->check = 0;
1392
1393 sg_init_one(&sg, bp, sizeof(*bp) + sizeof(*th));
1394 ahash_request_set_crypt(hp->md5_req, &sg, NULL,
1395 sizeof(*bp) + sizeof(*th));
1396 return crypto_ahash_update(hp->md5_req);
1397}
1398
1399static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
1400 __be32 daddr, __be32 saddr, const struct tcphdr *th)
1401{
1402 struct tcp_md5sig_pool *hp;
1403 struct ahash_request *req;
1404
1405 hp = tcp_get_md5sig_pool();
1406 if (!hp)
1407 goto clear_hash_noput;
1408 req = hp->md5_req;
1409
1410 if (crypto_ahash_init(req))
1411 goto clear_hash;
1412 if (tcp_v4_md5_hash_headers(hp, daddr, saddr, th, th->doff << 2))
1413 goto clear_hash;
1414 if (tcp_md5_hash_key(hp, key))
1415 goto clear_hash;
1416 ahash_request_set_crypt(req, NULL, md5_hash, 0);
1417 if (crypto_ahash_final(req))
1418 goto clear_hash;
1419
1420 tcp_put_md5sig_pool();
1421 return 0;
1422
1423clear_hash:
1424 tcp_put_md5sig_pool();
1425clear_hash_noput:
1426 memset(md5_hash, 0, 16);
1427 return 1;
1428}
1429
1430int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
1431 const struct sock *sk,
1432 const struct sk_buff *skb)
1433{
1434 struct tcp_md5sig_pool *hp;
1435 struct ahash_request *req;
1436 const struct tcphdr *th = tcp_hdr(skb);
1437 __be32 saddr, daddr;
1438
1439 if (sk) { /* valid for establish/request sockets */
1440 saddr = sk->sk_rcv_saddr;
1441 daddr = sk->sk_daddr;
1442 } else {
1443 const struct iphdr *iph = ip_hdr(skb);
1444 saddr = iph->saddr;
1445 daddr = iph->daddr;
1446 }
1447
1448 hp = tcp_get_md5sig_pool();
1449 if (!hp)
1450 goto clear_hash_noput;
1451 req = hp->md5_req;
1452
1453 if (crypto_ahash_init(req))
1454 goto clear_hash;
1455
1456 if (tcp_v4_md5_hash_headers(hp, daddr, saddr, th, skb->len))
1457 goto clear_hash;
1458 if (tcp_md5_hash_skb_data(hp, skb, th->doff << 2))
1459 goto clear_hash;
1460 if (tcp_md5_hash_key(hp, key))
1461 goto clear_hash;
1462 ahash_request_set_crypt(req, NULL, md5_hash, 0);
1463 if (crypto_ahash_final(req))
1464 goto clear_hash;
1465
1466 tcp_put_md5sig_pool();
1467 return 0;
1468
1469clear_hash:
1470 tcp_put_md5sig_pool();
1471clear_hash_noput:
1472 memset(md5_hash, 0, 16);
1473 return 1;
1474}
1475EXPORT_SYMBOL(tcp_v4_md5_hash_skb);
1476
1477#endif
1478
1479static void tcp_v4_init_req(struct request_sock *req,
1480 const struct sock *sk_listener,
1481 struct sk_buff *skb)
1482{
1483 struct inet_request_sock *ireq = inet_rsk(req);
1484 struct net *net = sock_net(sk_listener);
1485
1486 sk_rcv_saddr_set(req_to_sk(req), ip_hdr(skb)->daddr);
1487 sk_daddr_set(req_to_sk(req), ip_hdr(skb)->saddr);
1488 RCU_INIT_POINTER(ireq->ireq_opt, tcp_v4_save_options(net, skb));
1489}
1490
1491static struct dst_entry *tcp_v4_route_req(const struct sock *sk,
1492 struct sk_buff *skb,
1493 struct flowi *fl,
1494 struct request_sock *req)
1495{
1496 tcp_v4_init_req(req, sk, skb);
1497
1498 if (security_inet_conn_request(sk, skb, req))
1499 return NULL;
1500
1501 return inet_csk_route_req(sk, &fl->u.ip4, req);
1502}
1503
1504struct request_sock_ops tcp_request_sock_ops __read_mostly = {
1505 .family = PF_INET,
1506 .obj_size = sizeof(struct tcp_request_sock),
1507 .rtx_syn_ack = tcp_rtx_synack,
1508 .send_ack = tcp_v4_reqsk_send_ack,
1509 .destructor = tcp_v4_reqsk_destructor,
1510 .send_reset = tcp_v4_send_reset,
1511 .syn_ack_timeout = tcp_syn_ack_timeout,
1512};
1513
1514const struct tcp_request_sock_ops tcp_request_sock_ipv4_ops = {
1515 .mss_clamp = TCP_MSS_DEFAULT,
1516#ifdef CONFIG_TCP_MD5SIG
1517 .req_md5_lookup = tcp_v4_md5_lookup,
1518 .calc_md5_hash = tcp_v4_md5_hash_skb,
1519#endif
1520#ifdef CONFIG_SYN_COOKIES
1521 .cookie_init_seq = cookie_v4_init_sequence,
1522#endif
1523 .route_req = tcp_v4_route_req,
1524 .init_seq = tcp_v4_init_seq,
1525 .init_ts_off = tcp_v4_init_ts_off,
1526 .send_synack = tcp_v4_send_synack,
1527};
1528
1529int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb)
1530{
1531 /* Never answer to SYNs send to broadcast or multicast */
1532 if (skb_rtable(skb)->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST))
1533 goto drop;
1534
1535 return tcp_conn_request(&tcp_request_sock_ops,
1536 &tcp_request_sock_ipv4_ops, sk, skb);
1537
1538drop:
1539 tcp_listendrop(sk);
1540 return 0;
1541}
1542EXPORT_SYMBOL(tcp_v4_conn_request);
1543
1544
1545/*
1546 * The three way handshake has completed - we got a valid synack -
1547 * now create the new socket.
1548 */
1549struct sock *tcp_v4_syn_recv_sock(const struct sock *sk, struct sk_buff *skb,
1550 struct request_sock *req,
1551 struct dst_entry *dst,
1552 struct request_sock *req_unhash,
1553 bool *own_req)
1554{
1555 struct inet_request_sock *ireq;
1556 bool found_dup_sk = false;
1557 struct inet_sock *newinet;
1558 struct tcp_sock *newtp;
1559 struct sock *newsk;
1560#ifdef CONFIG_TCP_MD5SIG
1561 const union tcp_md5_addr *addr;
1562 struct tcp_md5sig_key *key;
1563 int l3index;
1564#endif
1565 struct ip_options_rcu *inet_opt;
1566
1567 if (sk_acceptq_is_full(sk))
1568 goto exit_overflow;
1569
1570 newsk = tcp_create_openreq_child(sk, req, skb);
1571 if (!newsk)
1572 goto exit_nonewsk;
1573
1574 newsk->sk_gso_type = SKB_GSO_TCPV4;
1575 inet_sk_rx_dst_set(newsk, skb);
1576
1577 newtp = tcp_sk(newsk);
1578 newinet = inet_sk(newsk);
1579 ireq = inet_rsk(req);
1580 sk_daddr_set(newsk, ireq->ir_rmt_addr);
1581 sk_rcv_saddr_set(newsk, ireq->ir_loc_addr);
1582 newsk->sk_bound_dev_if = ireq->ir_iif;
1583 newinet->inet_saddr = ireq->ir_loc_addr;
1584 inet_opt = rcu_dereference(ireq->ireq_opt);
1585 RCU_INIT_POINTER(newinet->inet_opt, inet_opt);
1586 newinet->mc_index = inet_iif(skb);
1587 newinet->mc_ttl = ip_hdr(skb)->ttl;
1588 newinet->rcv_tos = ip_hdr(skb)->tos;
1589 inet_csk(newsk)->icsk_ext_hdr_len = 0;
1590 if (inet_opt)
1591 inet_csk(newsk)->icsk_ext_hdr_len = inet_opt->opt.optlen;
1592 newinet->inet_id = get_random_u16();
1593
1594 /* Set ToS of the new socket based upon the value of incoming SYN.
1595 * ECT bits are set later in tcp_init_transfer().
1596 */
1597 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reflect_tos))
1598 newinet->tos = tcp_rsk(req)->syn_tos & ~INET_ECN_MASK;
1599
1600 if (!dst) {
1601 dst = inet_csk_route_child_sock(sk, newsk, req);
1602 if (!dst)
1603 goto put_and_exit;
1604 } else {
1605 /* syncookie case : see end of cookie_v4_check() */
1606 }
1607 sk_setup_caps(newsk, dst);
1608
1609 tcp_ca_openreq_child(newsk, dst);
1610
1611 tcp_sync_mss(newsk, dst_mtu(dst));
1612 newtp->advmss = tcp_mss_clamp(tcp_sk(sk), dst_metric_advmss(dst));
1613
1614 tcp_initialize_rcv_mss(newsk);
1615
1616#ifdef CONFIG_TCP_MD5SIG
1617 l3index = l3mdev_master_ifindex_by_index(sock_net(sk), ireq->ir_iif);
1618 /* Copy over the MD5 key from the original socket */
1619 addr = (union tcp_md5_addr *)&newinet->inet_daddr;
1620 key = tcp_md5_do_lookup(sk, l3index, addr, AF_INET);
1621 if (key) {
1622 if (tcp_md5_key_copy(newsk, addr, AF_INET, 32, l3index, key))
1623 goto put_and_exit;
1624 sk_gso_disable(newsk);
1625 }
1626#endif
1627
1628 if (__inet_inherit_port(sk, newsk) < 0)
1629 goto put_and_exit;
1630 *own_req = inet_ehash_nolisten(newsk, req_to_sk(req_unhash),
1631 &found_dup_sk);
1632 if (likely(*own_req)) {
1633 tcp_move_syn(newtp, req);
1634 ireq->ireq_opt = NULL;
1635 } else {
1636 newinet->inet_opt = NULL;
1637
1638 if (!req_unhash && found_dup_sk) {
1639 /* This code path should only be executed in the
1640 * syncookie case only
1641 */
1642 bh_unlock_sock(newsk);
1643 sock_put(newsk);
1644 newsk = NULL;
1645 }
1646 }
1647 return newsk;
1648
1649exit_overflow:
1650 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
1651exit_nonewsk:
1652 dst_release(dst);
1653exit:
1654 tcp_listendrop(sk);
1655 return NULL;
1656put_and_exit:
1657 newinet->inet_opt = NULL;
1658 inet_csk_prepare_forced_close(newsk);
1659 tcp_done(newsk);
1660 goto exit;
1661}
1662EXPORT_SYMBOL(tcp_v4_syn_recv_sock);
1663
1664static struct sock *tcp_v4_cookie_check(struct sock *sk, struct sk_buff *skb)
1665{
1666#ifdef CONFIG_SYN_COOKIES
1667 const struct tcphdr *th = tcp_hdr(skb);
1668
1669 if (!th->syn)
1670 sk = cookie_v4_check(sk, skb);
1671#endif
1672 return sk;
1673}
1674
1675u16 tcp_v4_get_syncookie(struct sock *sk, struct iphdr *iph,
1676 struct tcphdr *th, u32 *cookie)
1677{
1678 u16 mss = 0;
1679#ifdef CONFIG_SYN_COOKIES
1680 mss = tcp_get_syncookie_mss(&tcp_request_sock_ops,
1681 &tcp_request_sock_ipv4_ops, sk, th);
1682 if (mss) {
1683 *cookie = __cookie_v4_init_sequence(iph, th, &mss);
1684 tcp_synq_overflow(sk);
1685 }
1686#endif
1687 return mss;
1688}
1689
1690INDIRECT_CALLABLE_DECLARE(struct dst_entry *ipv4_dst_check(struct dst_entry *,
1691 u32));
1692/* The socket must have it's spinlock held when we get
1693 * here, unless it is a TCP_LISTEN socket.
1694 *
1695 * We have a potential double-lock case here, so even when
1696 * doing backlog processing we use the BH locking scheme.
1697 * This is because we cannot sleep with the original spinlock
1698 * held.
1699 */
1700int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb)
1701{
1702 enum skb_drop_reason reason;
1703 struct sock *rsk;
1704
1705 if (sk->sk_state == TCP_ESTABLISHED) { /* Fast path */
1706 struct dst_entry *dst;
1707
1708 dst = rcu_dereference_protected(sk->sk_rx_dst,
1709 lockdep_sock_is_held(sk));
1710
1711 sock_rps_save_rxhash(sk, skb);
1712 sk_mark_napi_id(sk, skb);
1713 if (dst) {
1714 if (sk->sk_rx_dst_ifindex != skb->skb_iif ||
1715 !INDIRECT_CALL_1(dst->ops->check, ipv4_dst_check,
1716 dst, 0)) {
1717 RCU_INIT_POINTER(sk->sk_rx_dst, NULL);
1718 dst_release(dst);
1719 }
1720 }
1721 tcp_rcv_established(sk, skb);
1722 return 0;
1723 }
1724
1725 reason = SKB_DROP_REASON_NOT_SPECIFIED;
1726 if (tcp_checksum_complete(skb))
1727 goto csum_err;
1728
1729 if (sk->sk_state == TCP_LISTEN) {
1730 struct sock *nsk = tcp_v4_cookie_check(sk, skb);
1731
1732 if (!nsk)
1733 goto discard;
1734 if (nsk != sk) {
1735 if (tcp_child_process(sk, nsk, skb)) {
1736 rsk = nsk;
1737 goto reset;
1738 }
1739 return 0;
1740 }
1741 } else
1742 sock_rps_save_rxhash(sk, skb);
1743
1744 if (tcp_rcv_state_process(sk, skb)) {
1745 rsk = sk;
1746 goto reset;
1747 }
1748 return 0;
1749
1750reset:
1751 tcp_v4_send_reset(rsk, skb);
1752discard:
1753 kfree_skb_reason(skb, reason);
1754 /* Be careful here. If this function gets more complicated and
1755 * gcc suffers from register pressure on the x86, sk (in %ebx)
1756 * might be destroyed here. This current version compiles correctly,
1757 * but you have been warned.
1758 */
1759 return 0;
1760
1761csum_err:
1762 reason = SKB_DROP_REASON_TCP_CSUM;
1763 trace_tcp_bad_csum(skb);
1764 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
1765 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
1766 goto discard;
1767}
1768EXPORT_SYMBOL(tcp_v4_do_rcv);
1769
1770int tcp_v4_early_demux(struct sk_buff *skb)
1771{
1772 struct net *net = dev_net(skb->dev);
1773 const struct iphdr *iph;
1774 const struct tcphdr *th;
1775 struct sock *sk;
1776
1777 if (skb->pkt_type != PACKET_HOST)
1778 return 0;
1779
1780 if (!pskb_may_pull(skb, skb_transport_offset(skb) + sizeof(struct tcphdr)))
1781 return 0;
1782
1783 iph = ip_hdr(skb);
1784 th = tcp_hdr(skb);
1785
1786 if (th->doff < sizeof(struct tcphdr) / 4)
1787 return 0;
1788
1789 sk = __inet_lookup_established(net, net->ipv4.tcp_death_row.hashinfo,
1790 iph->saddr, th->source,
1791 iph->daddr, ntohs(th->dest),
1792 skb->skb_iif, inet_sdif(skb));
1793 if (sk) {
1794 skb->sk = sk;
1795 skb->destructor = sock_edemux;
1796 if (sk_fullsock(sk)) {
1797 struct dst_entry *dst = rcu_dereference(sk->sk_rx_dst);
1798
1799 if (dst)
1800 dst = dst_check(dst, 0);
1801 if (dst &&
1802 sk->sk_rx_dst_ifindex == skb->skb_iif)
1803 skb_dst_set_noref(skb, dst);
1804 }
1805 }
1806 return 0;
1807}
1808
1809bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb,
1810 enum skb_drop_reason *reason)
1811{
1812 u32 limit, tail_gso_size, tail_gso_segs;
1813 struct skb_shared_info *shinfo;
1814 const struct tcphdr *th;
1815 struct tcphdr *thtail;
1816 struct sk_buff *tail;
1817 unsigned int hdrlen;
1818 bool fragstolen;
1819 u32 gso_segs;
1820 u32 gso_size;
1821 int delta;
1822
1823 /* In case all data was pulled from skb frags (in __pskb_pull_tail()),
1824 * we can fix skb->truesize to its real value to avoid future drops.
1825 * This is valid because skb is not yet charged to the socket.
1826 * It has been noticed pure SACK packets were sometimes dropped
1827 * (if cooked by drivers without copybreak feature).
1828 */
1829 skb_condense(skb);
1830
1831 skb_dst_drop(skb);
1832
1833 if (unlikely(tcp_checksum_complete(skb))) {
1834 bh_unlock_sock(sk);
1835 trace_tcp_bad_csum(skb);
1836 *reason = SKB_DROP_REASON_TCP_CSUM;
1837 __TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
1838 __TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
1839 return true;
1840 }
1841
1842 /* Attempt coalescing to last skb in backlog, even if we are
1843 * above the limits.
1844 * This is okay because skb capacity is limited to MAX_SKB_FRAGS.
1845 */
1846 th = (const struct tcphdr *)skb->data;
1847 hdrlen = th->doff * 4;
1848
1849 tail = sk->sk_backlog.tail;
1850 if (!tail)
1851 goto no_coalesce;
1852 thtail = (struct tcphdr *)tail->data;
1853
1854 if (TCP_SKB_CB(tail)->end_seq != TCP_SKB_CB(skb)->seq ||
1855 TCP_SKB_CB(tail)->ip_dsfield != TCP_SKB_CB(skb)->ip_dsfield ||
1856 ((TCP_SKB_CB(tail)->tcp_flags |
1857 TCP_SKB_CB(skb)->tcp_flags) & (TCPHDR_SYN | TCPHDR_RST | TCPHDR_URG)) ||
1858 !((TCP_SKB_CB(tail)->tcp_flags &
1859 TCP_SKB_CB(skb)->tcp_flags) & TCPHDR_ACK) ||
1860 ((TCP_SKB_CB(tail)->tcp_flags ^
1861 TCP_SKB_CB(skb)->tcp_flags) & (TCPHDR_ECE | TCPHDR_CWR)) ||
1862#ifdef CONFIG_TLS_DEVICE
1863 tail->decrypted != skb->decrypted ||
1864#endif
1865 thtail->doff != th->doff ||
1866 memcmp(thtail + 1, th + 1, hdrlen - sizeof(*th)))
1867 goto no_coalesce;
1868
1869 __skb_pull(skb, hdrlen);
1870
1871 shinfo = skb_shinfo(skb);
1872 gso_size = shinfo->gso_size ?: skb->len;
1873 gso_segs = shinfo->gso_segs ?: 1;
1874
1875 shinfo = skb_shinfo(tail);
1876 tail_gso_size = shinfo->gso_size ?: (tail->len - hdrlen);
1877 tail_gso_segs = shinfo->gso_segs ?: 1;
1878
1879 if (skb_try_coalesce(tail, skb, &fragstolen, &delta)) {
1880 TCP_SKB_CB(tail)->end_seq = TCP_SKB_CB(skb)->end_seq;
1881
1882 if (likely(!before(TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(tail)->ack_seq))) {
1883 TCP_SKB_CB(tail)->ack_seq = TCP_SKB_CB(skb)->ack_seq;
1884 thtail->window = th->window;
1885 }
1886
1887 /* We have to update both TCP_SKB_CB(tail)->tcp_flags and
1888 * thtail->fin, so that the fast path in tcp_rcv_established()
1889 * is not entered if we append a packet with a FIN.
1890 * SYN, RST, URG are not present.
1891 * ACK is set on both packets.
1892 * PSH : we do not really care in TCP stack,
1893 * at least for 'GRO' packets.
1894 */
1895 thtail->fin |= th->fin;
1896 TCP_SKB_CB(tail)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1897
1898 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1899 TCP_SKB_CB(tail)->has_rxtstamp = true;
1900 tail->tstamp = skb->tstamp;
1901 skb_hwtstamps(tail)->hwtstamp = skb_hwtstamps(skb)->hwtstamp;
1902 }
1903
1904 /* Not as strict as GRO. We only need to carry mss max value */
1905 shinfo->gso_size = max(gso_size, tail_gso_size);
1906 shinfo->gso_segs = min_t(u32, gso_segs + tail_gso_segs, 0xFFFF);
1907
1908 sk->sk_backlog.len += delta;
1909 __NET_INC_STATS(sock_net(sk),
1910 LINUX_MIB_TCPBACKLOGCOALESCE);
1911 kfree_skb_partial(skb, fragstolen);
1912 return false;
1913 }
1914 __skb_push(skb, hdrlen);
1915
1916no_coalesce:
1917 limit = (u32)READ_ONCE(sk->sk_rcvbuf) + (u32)(READ_ONCE(sk->sk_sndbuf) >> 1);
1918
1919 /* Only socket owner can try to collapse/prune rx queues
1920 * to reduce memory overhead, so add a little headroom here.
1921 * Few sockets backlog are possibly concurrently non empty.
1922 */
1923 limit += 64 * 1024;
1924
1925 if (unlikely(sk_add_backlog(sk, skb, limit))) {
1926 bh_unlock_sock(sk);
1927 *reason = SKB_DROP_REASON_SOCKET_BACKLOG;
1928 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPBACKLOGDROP);
1929 return true;
1930 }
1931 return false;
1932}
1933EXPORT_SYMBOL(tcp_add_backlog);
1934
1935int tcp_filter(struct sock *sk, struct sk_buff *skb)
1936{
1937 struct tcphdr *th = (struct tcphdr *)skb->data;
1938
1939 return sk_filter_trim_cap(sk, skb, th->doff * 4);
1940}
1941EXPORT_SYMBOL(tcp_filter);
1942
1943static void tcp_v4_restore_cb(struct sk_buff *skb)
1944{
1945 memmove(IPCB(skb), &TCP_SKB_CB(skb)->header.h4,
1946 sizeof(struct inet_skb_parm));
1947}
1948
1949static void tcp_v4_fill_cb(struct sk_buff *skb, const struct iphdr *iph,
1950 const struct tcphdr *th)
1951{
1952 /* This is tricky : We move IPCB at its correct location into TCP_SKB_CB()
1953 * barrier() makes sure compiler wont play fool^Waliasing games.
1954 */
1955 memmove(&TCP_SKB_CB(skb)->header.h4, IPCB(skb),
1956 sizeof(struct inet_skb_parm));
1957 barrier();
1958
1959 TCP_SKB_CB(skb)->seq = ntohl(th->seq);
1960 TCP_SKB_CB(skb)->end_seq = (TCP_SKB_CB(skb)->seq + th->syn + th->fin +
1961 skb->len - th->doff * 4);
1962 TCP_SKB_CB(skb)->ack_seq = ntohl(th->ack_seq);
1963 TCP_SKB_CB(skb)->tcp_flags = tcp_flag_byte(th);
1964 TCP_SKB_CB(skb)->tcp_tw_isn = 0;
1965 TCP_SKB_CB(skb)->ip_dsfield = ipv4_get_dsfield(iph);
1966 TCP_SKB_CB(skb)->sacked = 0;
1967 TCP_SKB_CB(skb)->has_rxtstamp =
1968 skb->tstamp || skb_hwtstamps(skb)->hwtstamp;
1969}
1970
1971/*
1972 * From tcp_input.c
1973 */
1974
1975int tcp_v4_rcv(struct sk_buff *skb)
1976{
1977 struct net *net = dev_net(skb->dev);
1978 enum skb_drop_reason drop_reason;
1979 int sdif = inet_sdif(skb);
1980 int dif = inet_iif(skb);
1981 const struct iphdr *iph;
1982 const struct tcphdr *th;
1983 bool refcounted;
1984 struct sock *sk;
1985 int ret;
1986
1987 drop_reason = SKB_DROP_REASON_NOT_SPECIFIED;
1988 if (skb->pkt_type != PACKET_HOST)
1989 goto discard_it;
1990
1991 /* Count it even if it's bad */
1992 __TCP_INC_STATS(net, TCP_MIB_INSEGS);
1993
1994 if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
1995 goto discard_it;
1996
1997 th = (const struct tcphdr *)skb->data;
1998
1999 if (unlikely(th->doff < sizeof(struct tcphdr) / 4)) {
2000 drop_reason = SKB_DROP_REASON_PKT_TOO_SMALL;
2001 goto bad_packet;
2002 }
2003 if (!pskb_may_pull(skb, th->doff * 4))
2004 goto discard_it;
2005
2006 /* An explanation is required here, I think.
2007 * Packet length and doff are validated by header prediction,
2008 * provided case of th->doff==0 is eliminated.
2009 * So, we defer the checks. */
2010
2011 if (skb_checksum_init(skb, IPPROTO_TCP, inet_compute_pseudo))
2012 goto csum_error;
2013
2014 th = (const struct tcphdr *)skb->data;
2015 iph = ip_hdr(skb);
2016lookup:
2017 sk = __inet_lookup_skb(net->ipv4.tcp_death_row.hashinfo,
2018 skb, __tcp_hdrlen(th), th->source,
2019 th->dest, sdif, &refcounted);
2020 if (!sk)
2021 goto no_tcp_socket;
2022
2023process:
2024 if (sk->sk_state == TCP_TIME_WAIT)
2025 goto do_time_wait;
2026
2027 if (sk->sk_state == TCP_NEW_SYN_RECV) {
2028 struct request_sock *req = inet_reqsk(sk);
2029 bool req_stolen = false;
2030 struct sock *nsk;
2031
2032 sk = req->rsk_listener;
2033 if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb))
2034 drop_reason = SKB_DROP_REASON_XFRM_POLICY;
2035 else
2036 drop_reason = tcp_inbound_md5_hash(sk, skb,
2037 &iph->saddr, &iph->daddr,
2038 AF_INET, dif, sdif);
2039 if (unlikely(drop_reason)) {
2040 sk_drops_add(sk, skb);
2041 reqsk_put(req);
2042 goto discard_it;
2043 }
2044 if (tcp_checksum_complete(skb)) {
2045 reqsk_put(req);
2046 goto csum_error;
2047 }
2048 if (unlikely(sk->sk_state != TCP_LISTEN)) {
2049 nsk = reuseport_migrate_sock(sk, req_to_sk(req), skb);
2050 if (!nsk) {
2051 inet_csk_reqsk_queue_drop_and_put(sk, req);
2052 goto lookup;
2053 }
2054 sk = nsk;
2055 /* reuseport_migrate_sock() has already held one sk_refcnt
2056 * before returning.
2057 */
2058 } else {
2059 /* We own a reference on the listener, increase it again
2060 * as we might lose it too soon.
2061 */
2062 sock_hold(sk);
2063 }
2064 refcounted = true;
2065 nsk = NULL;
2066 if (!tcp_filter(sk, skb)) {
2067 th = (const struct tcphdr *)skb->data;
2068 iph = ip_hdr(skb);
2069 tcp_v4_fill_cb(skb, iph, th);
2070 nsk = tcp_check_req(sk, skb, req, false, &req_stolen);
2071 } else {
2072 drop_reason = SKB_DROP_REASON_SOCKET_FILTER;
2073 }
2074 if (!nsk) {
2075 reqsk_put(req);
2076 if (req_stolen) {
2077 /* Another cpu got exclusive access to req
2078 * and created a full blown socket.
2079 * Try to feed this packet to this socket
2080 * instead of discarding it.
2081 */
2082 tcp_v4_restore_cb(skb);
2083 sock_put(sk);
2084 goto lookup;
2085 }
2086 goto discard_and_relse;
2087 }
2088 nf_reset_ct(skb);
2089 if (nsk == sk) {
2090 reqsk_put(req);
2091 tcp_v4_restore_cb(skb);
2092 } else if (tcp_child_process(sk, nsk, skb)) {
2093 tcp_v4_send_reset(nsk, skb);
2094 goto discard_and_relse;
2095 } else {
2096 sock_put(sk);
2097 return 0;
2098 }
2099 }
2100
2101 if (static_branch_unlikely(&ip4_min_ttl)) {
2102 /* min_ttl can be changed concurrently from do_ip_setsockopt() */
2103 if (unlikely(iph->ttl < READ_ONCE(inet_sk(sk)->min_ttl))) {
2104 __NET_INC_STATS(net, LINUX_MIB_TCPMINTTLDROP);
2105 goto discard_and_relse;
2106 }
2107 }
2108
2109 if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb)) {
2110 drop_reason = SKB_DROP_REASON_XFRM_POLICY;
2111 goto discard_and_relse;
2112 }
2113
2114 drop_reason = tcp_inbound_md5_hash(sk, skb, &iph->saddr,
2115 &iph->daddr, AF_INET, dif, sdif);
2116 if (drop_reason)
2117 goto discard_and_relse;
2118
2119 nf_reset_ct(skb);
2120
2121 if (tcp_filter(sk, skb)) {
2122 drop_reason = SKB_DROP_REASON_SOCKET_FILTER;
2123 goto discard_and_relse;
2124 }
2125 th = (const struct tcphdr *)skb->data;
2126 iph = ip_hdr(skb);
2127 tcp_v4_fill_cb(skb, iph, th);
2128
2129 skb->dev = NULL;
2130
2131 if (sk->sk_state == TCP_LISTEN) {
2132 ret = tcp_v4_do_rcv(sk, skb);
2133 goto put_and_return;
2134 }
2135
2136 sk_incoming_cpu_update(sk);
2137
2138 bh_lock_sock_nested(sk);
2139 tcp_segs_in(tcp_sk(sk), skb);
2140 ret = 0;
2141 if (!sock_owned_by_user(sk)) {
2142 ret = tcp_v4_do_rcv(sk, skb);
2143 } else {
2144 if (tcp_add_backlog(sk, skb, &drop_reason))
2145 goto discard_and_relse;
2146 }
2147 bh_unlock_sock(sk);
2148
2149put_and_return:
2150 if (refcounted)
2151 sock_put(sk);
2152
2153 return ret;
2154
2155no_tcp_socket:
2156 drop_reason = SKB_DROP_REASON_NO_SOCKET;
2157 if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb))
2158 goto discard_it;
2159
2160 tcp_v4_fill_cb(skb, iph, th);
2161
2162 if (tcp_checksum_complete(skb)) {
2163csum_error:
2164 drop_reason = SKB_DROP_REASON_TCP_CSUM;
2165 trace_tcp_bad_csum(skb);
2166 __TCP_INC_STATS(net, TCP_MIB_CSUMERRORS);
2167bad_packet:
2168 __TCP_INC_STATS(net, TCP_MIB_INERRS);
2169 } else {
2170 tcp_v4_send_reset(NULL, skb);
2171 }
2172
2173discard_it:
2174 SKB_DR_OR(drop_reason, NOT_SPECIFIED);
2175 /* Discard frame. */
2176 kfree_skb_reason(skb, drop_reason);
2177 return 0;
2178
2179discard_and_relse:
2180 sk_drops_add(sk, skb);
2181 if (refcounted)
2182 sock_put(sk);
2183 goto discard_it;
2184
2185do_time_wait:
2186 if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb)) {
2187 drop_reason = SKB_DROP_REASON_XFRM_POLICY;
2188 inet_twsk_put(inet_twsk(sk));
2189 goto discard_it;
2190 }
2191
2192 tcp_v4_fill_cb(skb, iph, th);
2193
2194 if (tcp_checksum_complete(skb)) {
2195 inet_twsk_put(inet_twsk(sk));
2196 goto csum_error;
2197 }
2198 switch (tcp_timewait_state_process(inet_twsk(sk), skb, th)) {
2199 case TCP_TW_SYN: {
2200 struct sock *sk2 = inet_lookup_listener(net,
2201 net->ipv4.tcp_death_row.hashinfo,
2202 skb, __tcp_hdrlen(th),
2203 iph->saddr, th->source,
2204 iph->daddr, th->dest,
2205 inet_iif(skb),
2206 sdif);
2207 if (sk2) {
2208 inet_twsk_deschedule_put(inet_twsk(sk));
2209 sk = sk2;
2210 tcp_v4_restore_cb(skb);
2211 refcounted = false;
2212 goto process;
2213 }
2214 }
2215 /* to ACK */
2216 fallthrough;
2217 case TCP_TW_ACK:
2218 tcp_v4_timewait_ack(sk, skb);
2219 break;
2220 case TCP_TW_RST:
2221 tcp_v4_send_reset(sk, skb);
2222 inet_twsk_deschedule_put(inet_twsk(sk));
2223 goto discard_it;
2224 case TCP_TW_SUCCESS:;
2225 }
2226 goto discard_it;
2227}
2228
2229static struct timewait_sock_ops tcp_timewait_sock_ops = {
2230 .twsk_obj_size = sizeof(struct tcp_timewait_sock),
2231 .twsk_unique = tcp_twsk_unique,
2232 .twsk_destructor= tcp_twsk_destructor,
2233};
2234
2235void inet_sk_rx_dst_set(struct sock *sk, const struct sk_buff *skb)
2236{
2237 struct dst_entry *dst = skb_dst(skb);
2238
2239 if (dst && dst_hold_safe(dst)) {
2240 rcu_assign_pointer(sk->sk_rx_dst, dst);
2241 sk->sk_rx_dst_ifindex = skb->skb_iif;
2242 }
2243}
2244EXPORT_SYMBOL(inet_sk_rx_dst_set);
2245
2246const struct inet_connection_sock_af_ops ipv4_specific = {
2247 .queue_xmit = ip_queue_xmit,
2248 .send_check = tcp_v4_send_check,
2249 .rebuild_header = inet_sk_rebuild_header,
2250 .sk_rx_dst_set = inet_sk_rx_dst_set,
2251 .conn_request = tcp_v4_conn_request,
2252 .syn_recv_sock = tcp_v4_syn_recv_sock,
2253 .net_header_len = sizeof(struct iphdr),
2254 .setsockopt = ip_setsockopt,
2255 .getsockopt = ip_getsockopt,
2256 .addr2sockaddr = inet_csk_addr2sockaddr,
2257 .sockaddr_len = sizeof(struct sockaddr_in),
2258 .mtu_reduced = tcp_v4_mtu_reduced,
2259};
2260EXPORT_SYMBOL(ipv4_specific);
2261
2262#ifdef CONFIG_TCP_MD5SIG
2263static const struct tcp_sock_af_ops tcp_sock_ipv4_specific = {
2264 .md5_lookup = tcp_v4_md5_lookup,
2265 .calc_md5_hash = tcp_v4_md5_hash_skb,
2266 .md5_parse = tcp_v4_parse_md5_keys,
2267};
2268#endif
2269
2270/* NOTE: A lot of things set to zero explicitly by call to
2271 * sk_alloc() so need not be done here.
2272 */
2273static int tcp_v4_init_sock(struct sock *sk)
2274{
2275 struct inet_connection_sock *icsk = inet_csk(sk);
2276
2277 tcp_init_sock(sk);
2278
2279 icsk->icsk_af_ops = &ipv4_specific;
2280
2281#ifdef CONFIG_TCP_MD5SIG
2282 tcp_sk(sk)->af_specific = &tcp_sock_ipv4_specific;
2283#endif
2284
2285 return 0;
2286}
2287
2288void tcp_v4_destroy_sock(struct sock *sk)
2289{
2290 struct tcp_sock *tp = tcp_sk(sk);
2291
2292 trace_tcp_destroy_sock(sk);
2293
2294 tcp_clear_xmit_timers(sk);
2295
2296 tcp_cleanup_congestion_control(sk);
2297
2298 tcp_cleanup_ulp(sk);
2299
2300 /* Cleanup up the write buffer. */
2301 tcp_write_queue_purge(sk);
2302
2303 /* Check if we want to disable active TFO */
2304 tcp_fastopen_active_disable_ofo_check(sk);
2305
2306 /* Cleans up our, hopefully empty, out_of_order_queue. */
2307 skb_rbtree_purge(&tp->out_of_order_queue);
2308
2309#ifdef CONFIG_TCP_MD5SIG
2310 /* Clean up the MD5 key list, if any */
2311 if (tp->md5sig_info) {
2312 tcp_clear_md5_list(sk);
2313 kfree_rcu(rcu_dereference_protected(tp->md5sig_info, 1), rcu);
2314 tp->md5sig_info = NULL;
2315 static_branch_slow_dec_deferred(&tcp_md5_needed);
2316 }
2317#endif
2318
2319 /* Clean up a referenced TCP bind bucket. */
2320 if (inet_csk(sk)->icsk_bind_hash)
2321 inet_put_port(sk);
2322
2323 BUG_ON(rcu_access_pointer(tp->fastopen_rsk));
2324
2325 /* If socket is aborted during connect operation */
2326 tcp_free_fastopen_req(tp);
2327 tcp_fastopen_destroy_cipher(sk);
2328 tcp_saved_syn_free(tp);
2329
2330 sk_sockets_allocated_dec(sk);
2331}
2332EXPORT_SYMBOL(tcp_v4_destroy_sock);
2333
2334#ifdef CONFIG_PROC_FS
2335/* Proc filesystem TCP sock list dumping. */
2336
2337static unsigned short seq_file_family(const struct seq_file *seq);
2338
2339static bool seq_sk_match(struct seq_file *seq, const struct sock *sk)
2340{
2341 unsigned short family = seq_file_family(seq);
2342
2343 /* AF_UNSPEC is used as a match all */
2344 return ((family == AF_UNSPEC || family == sk->sk_family) &&
2345 net_eq(sock_net(sk), seq_file_net(seq)));
2346}
2347
2348/* Find a non empty bucket (starting from st->bucket)
2349 * and return the first sk from it.
2350 */
2351static void *listening_get_first(struct seq_file *seq)
2352{
2353 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2354 struct tcp_iter_state *st = seq->private;
2355
2356 st->offset = 0;
2357 for (; st->bucket <= hinfo->lhash2_mask; st->bucket++) {
2358 struct inet_listen_hashbucket *ilb2;
2359 struct hlist_nulls_node *node;
2360 struct sock *sk;
2361
2362 ilb2 = &hinfo->lhash2[st->bucket];
2363 if (hlist_nulls_empty(&ilb2->nulls_head))
2364 continue;
2365
2366 spin_lock(&ilb2->lock);
2367 sk_nulls_for_each(sk, node, &ilb2->nulls_head) {
2368 if (seq_sk_match(seq, sk))
2369 return sk;
2370 }
2371 spin_unlock(&ilb2->lock);
2372 }
2373
2374 return NULL;
2375}
2376
2377/* Find the next sk of "cur" within the same bucket (i.e. st->bucket).
2378 * If "cur" is the last one in the st->bucket,
2379 * call listening_get_first() to return the first sk of the next
2380 * non empty bucket.
2381 */
2382static void *listening_get_next(struct seq_file *seq, void *cur)
2383{
2384 struct tcp_iter_state *st = seq->private;
2385 struct inet_listen_hashbucket *ilb2;
2386 struct hlist_nulls_node *node;
2387 struct inet_hashinfo *hinfo;
2388 struct sock *sk = cur;
2389
2390 ++st->num;
2391 ++st->offset;
2392
2393 sk = sk_nulls_next(sk);
2394 sk_nulls_for_each_from(sk, node) {
2395 if (seq_sk_match(seq, sk))
2396 return sk;
2397 }
2398
2399 hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2400 ilb2 = &hinfo->lhash2[st->bucket];
2401 spin_unlock(&ilb2->lock);
2402 ++st->bucket;
2403 return listening_get_first(seq);
2404}
2405
2406static void *listening_get_idx(struct seq_file *seq, loff_t *pos)
2407{
2408 struct tcp_iter_state *st = seq->private;
2409 void *rc;
2410
2411 st->bucket = 0;
2412 st->offset = 0;
2413 rc = listening_get_first(seq);
2414
2415 while (rc && *pos) {
2416 rc = listening_get_next(seq, rc);
2417 --*pos;
2418 }
2419 return rc;
2420}
2421
2422static inline bool empty_bucket(struct inet_hashinfo *hinfo,
2423 const struct tcp_iter_state *st)
2424{
2425 return hlist_nulls_empty(&hinfo->ehash[st->bucket].chain);
2426}
2427
2428/*
2429 * Get first established socket starting from bucket given in st->bucket.
2430 * If st->bucket is zero, the very first socket in the hash is returned.
2431 */
2432static void *established_get_first(struct seq_file *seq)
2433{
2434 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2435 struct tcp_iter_state *st = seq->private;
2436
2437 st->offset = 0;
2438 for (; st->bucket <= hinfo->ehash_mask; ++st->bucket) {
2439 struct sock *sk;
2440 struct hlist_nulls_node *node;
2441 spinlock_t *lock = inet_ehash_lockp(hinfo, st->bucket);
2442
2443 /* Lockless fast path for the common case of empty buckets */
2444 if (empty_bucket(hinfo, st))
2445 continue;
2446
2447 spin_lock_bh(lock);
2448 sk_nulls_for_each(sk, node, &hinfo->ehash[st->bucket].chain) {
2449 if (seq_sk_match(seq, sk))
2450 return sk;
2451 }
2452 spin_unlock_bh(lock);
2453 }
2454
2455 return NULL;
2456}
2457
2458static void *established_get_next(struct seq_file *seq, void *cur)
2459{
2460 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2461 struct tcp_iter_state *st = seq->private;
2462 struct hlist_nulls_node *node;
2463 struct sock *sk = cur;
2464
2465 ++st->num;
2466 ++st->offset;
2467
2468 sk = sk_nulls_next(sk);
2469
2470 sk_nulls_for_each_from(sk, node) {
2471 if (seq_sk_match(seq, sk))
2472 return sk;
2473 }
2474
2475 spin_unlock_bh(inet_ehash_lockp(hinfo, st->bucket));
2476 ++st->bucket;
2477 return established_get_first(seq);
2478}
2479
2480static void *established_get_idx(struct seq_file *seq, loff_t pos)
2481{
2482 struct tcp_iter_state *st = seq->private;
2483 void *rc;
2484
2485 st->bucket = 0;
2486 rc = established_get_first(seq);
2487
2488 while (rc && pos) {
2489 rc = established_get_next(seq, rc);
2490 --pos;
2491 }
2492 return rc;
2493}
2494
2495static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
2496{
2497 void *rc;
2498 struct tcp_iter_state *st = seq->private;
2499
2500 st->state = TCP_SEQ_STATE_LISTENING;
2501 rc = listening_get_idx(seq, &pos);
2502
2503 if (!rc) {
2504 st->state = TCP_SEQ_STATE_ESTABLISHED;
2505 rc = established_get_idx(seq, pos);
2506 }
2507
2508 return rc;
2509}
2510
2511static void *tcp_seek_last_pos(struct seq_file *seq)
2512{
2513 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2514 struct tcp_iter_state *st = seq->private;
2515 int bucket = st->bucket;
2516 int offset = st->offset;
2517 int orig_num = st->num;
2518 void *rc = NULL;
2519
2520 switch (st->state) {
2521 case TCP_SEQ_STATE_LISTENING:
2522 if (st->bucket > hinfo->lhash2_mask)
2523 break;
2524 rc = listening_get_first(seq);
2525 while (offset-- && rc && bucket == st->bucket)
2526 rc = listening_get_next(seq, rc);
2527 if (rc)
2528 break;
2529 st->bucket = 0;
2530 st->state = TCP_SEQ_STATE_ESTABLISHED;
2531 fallthrough;
2532 case TCP_SEQ_STATE_ESTABLISHED:
2533 if (st->bucket > hinfo->ehash_mask)
2534 break;
2535 rc = established_get_first(seq);
2536 while (offset-- && rc && bucket == st->bucket)
2537 rc = established_get_next(seq, rc);
2538 }
2539
2540 st->num = orig_num;
2541
2542 return rc;
2543}
2544
2545void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
2546{
2547 struct tcp_iter_state *st = seq->private;
2548 void *rc;
2549
2550 if (*pos && *pos == st->last_pos) {
2551 rc = tcp_seek_last_pos(seq);
2552 if (rc)
2553 goto out;
2554 }
2555
2556 st->state = TCP_SEQ_STATE_LISTENING;
2557 st->num = 0;
2558 st->bucket = 0;
2559 st->offset = 0;
2560 rc = *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
2561
2562out:
2563 st->last_pos = *pos;
2564 return rc;
2565}
2566EXPORT_SYMBOL(tcp_seq_start);
2567
2568void *tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2569{
2570 struct tcp_iter_state *st = seq->private;
2571 void *rc = NULL;
2572
2573 if (v == SEQ_START_TOKEN) {
2574 rc = tcp_get_idx(seq, 0);
2575 goto out;
2576 }
2577
2578 switch (st->state) {
2579 case TCP_SEQ_STATE_LISTENING:
2580 rc = listening_get_next(seq, v);
2581 if (!rc) {
2582 st->state = TCP_SEQ_STATE_ESTABLISHED;
2583 st->bucket = 0;
2584 st->offset = 0;
2585 rc = established_get_first(seq);
2586 }
2587 break;
2588 case TCP_SEQ_STATE_ESTABLISHED:
2589 rc = established_get_next(seq, v);
2590 break;
2591 }
2592out:
2593 ++*pos;
2594 st->last_pos = *pos;
2595 return rc;
2596}
2597EXPORT_SYMBOL(tcp_seq_next);
2598
2599void tcp_seq_stop(struct seq_file *seq, void *v)
2600{
2601 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2602 struct tcp_iter_state *st = seq->private;
2603
2604 switch (st->state) {
2605 case TCP_SEQ_STATE_LISTENING:
2606 if (v != SEQ_START_TOKEN)
2607 spin_unlock(&hinfo->lhash2[st->bucket].lock);
2608 break;
2609 case TCP_SEQ_STATE_ESTABLISHED:
2610 if (v)
2611 spin_unlock_bh(inet_ehash_lockp(hinfo, st->bucket));
2612 break;
2613 }
2614}
2615EXPORT_SYMBOL(tcp_seq_stop);
2616
2617static void get_openreq4(const struct request_sock *req,
2618 struct seq_file *f, int i)
2619{
2620 const struct inet_request_sock *ireq = inet_rsk(req);
2621 long delta = req->rsk_timer.expires - jiffies;
2622
2623 seq_printf(f, "%4d: %08X:%04X %08X:%04X"
2624 " %02X %08X:%08X %02X:%08lX %08X %5u %8d %u %d %pK",
2625 i,
2626 ireq->ir_loc_addr,
2627 ireq->ir_num,
2628 ireq->ir_rmt_addr,
2629 ntohs(ireq->ir_rmt_port),
2630 TCP_SYN_RECV,
2631 0, 0, /* could print option size, but that is af dependent. */
2632 1, /* timers active (only the expire timer) */
2633 jiffies_delta_to_clock_t(delta),
2634 req->num_timeout,
2635 from_kuid_munged(seq_user_ns(f),
2636 sock_i_uid(req->rsk_listener)),
2637 0, /* non standard timer */
2638 0, /* open_requests have no inode */
2639 0,
2640 req);
2641}
2642
2643static void get_tcp4_sock(struct sock *sk, struct seq_file *f, int i)
2644{
2645 int timer_active;
2646 unsigned long timer_expires;
2647 const struct tcp_sock *tp = tcp_sk(sk);
2648 const struct inet_connection_sock *icsk = inet_csk(sk);
2649 const struct inet_sock *inet = inet_sk(sk);
2650 const struct fastopen_queue *fastopenq = &icsk->icsk_accept_queue.fastopenq;
2651 __be32 dest = inet->inet_daddr;
2652 __be32 src = inet->inet_rcv_saddr;
2653 __u16 destp = ntohs(inet->inet_dport);
2654 __u16 srcp = ntohs(inet->inet_sport);
2655 int rx_queue;
2656 int state;
2657
2658 if (icsk->icsk_pending == ICSK_TIME_RETRANS ||
2659 icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2660 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2661 timer_active = 1;
2662 timer_expires = icsk->icsk_timeout;
2663 } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
2664 timer_active = 4;
2665 timer_expires = icsk->icsk_timeout;
2666 } else if (timer_pending(&sk->sk_timer)) {
2667 timer_active = 2;
2668 timer_expires = sk->sk_timer.expires;
2669 } else {
2670 timer_active = 0;
2671 timer_expires = jiffies;
2672 }
2673
2674 state = inet_sk_state_load(sk);
2675 if (state == TCP_LISTEN)
2676 rx_queue = READ_ONCE(sk->sk_ack_backlog);
2677 else
2678 /* Because we don't lock the socket,
2679 * we might find a transient negative value.
2680 */
2681 rx_queue = max_t(int, READ_ONCE(tp->rcv_nxt) -
2682 READ_ONCE(tp->copied_seq), 0);
2683
2684 seq_printf(f, "%4d: %08X:%04X %08X:%04X %02X %08X:%08X %02X:%08lX "
2685 "%08X %5u %8d %lu %d %pK %lu %lu %u %u %d",
2686 i, src, srcp, dest, destp, state,
2687 READ_ONCE(tp->write_seq) - tp->snd_una,
2688 rx_queue,
2689 timer_active,
2690 jiffies_delta_to_clock_t(timer_expires - jiffies),
2691 icsk->icsk_retransmits,
2692 from_kuid_munged(seq_user_ns(f), sock_i_uid(sk)),
2693 icsk->icsk_probes_out,
2694 sock_i_ino(sk),
2695 refcount_read(&sk->sk_refcnt), sk,
2696 jiffies_to_clock_t(icsk->icsk_rto),
2697 jiffies_to_clock_t(icsk->icsk_ack.ato),
2698 (icsk->icsk_ack.quick << 1) | inet_csk_in_pingpong_mode(sk),
2699 tcp_snd_cwnd(tp),
2700 state == TCP_LISTEN ?
2701 fastopenq->max_qlen :
2702 (tcp_in_initial_slowstart(tp) ? -1 : tp->snd_ssthresh));
2703}
2704
2705static void get_timewait4_sock(const struct inet_timewait_sock *tw,
2706 struct seq_file *f, int i)
2707{
2708 long delta = tw->tw_timer.expires - jiffies;
2709 __be32 dest, src;
2710 __u16 destp, srcp;
2711
2712 dest = tw->tw_daddr;
2713 src = tw->tw_rcv_saddr;
2714 destp = ntohs(tw->tw_dport);
2715 srcp = ntohs(tw->tw_sport);
2716
2717 seq_printf(f, "%4d: %08X:%04X %08X:%04X"
2718 " %02X %08X:%08X %02X:%08lX %08X %5d %8d %d %d %pK",
2719 i, src, srcp, dest, destp, tw->tw_substate, 0, 0,
2720 3, jiffies_delta_to_clock_t(delta), 0, 0, 0, 0,
2721 refcount_read(&tw->tw_refcnt), tw);
2722}
2723
2724#define TMPSZ 150
2725
2726static int tcp4_seq_show(struct seq_file *seq, void *v)
2727{
2728 struct tcp_iter_state *st;
2729 struct sock *sk = v;
2730
2731 seq_setwidth(seq, TMPSZ - 1);
2732 if (v == SEQ_START_TOKEN) {
2733 seq_puts(seq, " sl local_address rem_address st tx_queue "
2734 "rx_queue tr tm->when retrnsmt uid timeout "
2735 "inode");
2736 goto out;
2737 }
2738 st = seq->private;
2739
2740 if (sk->sk_state == TCP_TIME_WAIT)
2741 get_timewait4_sock(v, seq, st->num);
2742 else if (sk->sk_state == TCP_NEW_SYN_RECV)
2743 get_openreq4(v, seq, st->num);
2744 else
2745 get_tcp4_sock(v, seq, st->num);
2746out:
2747 seq_pad(seq, '\n');
2748 return 0;
2749}
2750
2751#ifdef CONFIG_BPF_SYSCALL
2752struct bpf_tcp_iter_state {
2753 struct tcp_iter_state state;
2754 unsigned int cur_sk;
2755 unsigned int end_sk;
2756 unsigned int max_sk;
2757 struct sock **batch;
2758 bool st_bucket_done;
2759};
2760
2761struct bpf_iter__tcp {
2762 __bpf_md_ptr(struct bpf_iter_meta *, meta);
2763 __bpf_md_ptr(struct sock_common *, sk_common);
2764 uid_t uid __aligned(8);
2765};
2766
2767static int tcp_prog_seq_show(struct bpf_prog *prog, struct bpf_iter_meta *meta,
2768 struct sock_common *sk_common, uid_t uid)
2769{
2770 struct bpf_iter__tcp ctx;
2771
2772 meta->seq_num--; /* skip SEQ_START_TOKEN */
2773 ctx.meta = meta;
2774 ctx.sk_common = sk_common;
2775 ctx.uid = uid;
2776 return bpf_iter_run_prog(prog, &ctx);
2777}
2778
2779static void bpf_iter_tcp_put_batch(struct bpf_tcp_iter_state *iter)
2780{
2781 while (iter->cur_sk < iter->end_sk)
2782 sock_put(iter->batch[iter->cur_sk++]);
2783}
2784
2785static int bpf_iter_tcp_realloc_batch(struct bpf_tcp_iter_state *iter,
2786 unsigned int new_batch_sz)
2787{
2788 struct sock **new_batch;
2789
2790 new_batch = kvmalloc(sizeof(*new_batch) * new_batch_sz,
2791 GFP_USER | __GFP_NOWARN);
2792 if (!new_batch)
2793 return -ENOMEM;
2794
2795 bpf_iter_tcp_put_batch(iter);
2796 kvfree(iter->batch);
2797 iter->batch = new_batch;
2798 iter->max_sk = new_batch_sz;
2799
2800 return 0;
2801}
2802
2803static unsigned int bpf_iter_tcp_listening_batch(struct seq_file *seq,
2804 struct sock *start_sk)
2805{
2806 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2807 struct bpf_tcp_iter_state *iter = seq->private;
2808 struct tcp_iter_state *st = &iter->state;
2809 struct hlist_nulls_node *node;
2810 unsigned int expected = 1;
2811 struct sock *sk;
2812
2813 sock_hold(start_sk);
2814 iter->batch[iter->end_sk++] = start_sk;
2815
2816 sk = sk_nulls_next(start_sk);
2817 sk_nulls_for_each_from(sk, node) {
2818 if (seq_sk_match(seq, sk)) {
2819 if (iter->end_sk < iter->max_sk) {
2820 sock_hold(sk);
2821 iter->batch[iter->end_sk++] = sk;
2822 }
2823 expected++;
2824 }
2825 }
2826 spin_unlock(&hinfo->lhash2[st->bucket].lock);
2827
2828 return expected;
2829}
2830
2831static unsigned int bpf_iter_tcp_established_batch(struct seq_file *seq,
2832 struct sock *start_sk)
2833{
2834 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2835 struct bpf_tcp_iter_state *iter = seq->private;
2836 struct tcp_iter_state *st = &iter->state;
2837 struct hlist_nulls_node *node;
2838 unsigned int expected = 1;
2839 struct sock *sk;
2840
2841 sock_hold(start_sk);
2842 iter->batch[iter->end_sk++] = start_sk;
2843
2844 sk = sk_nulls_next(start_sk);
2845 sk_nulls_for_each_from(sk, node) {
2846 if (seq_sk_match(seq, sk)) {
2847 if (iter->end_sk < iter->max_sk) {
2848 sock_hold(sk);
2849 iter->batch[iter->end_sk++] = sk;
2850 }
2851 expected++;
2852 }
2853 }
2854 spin_unlock_bh(inet_ehash_lockp(hinfo, st->bucket));
2855
2856 return expected;
2857}
2858
2859static struct sock *bpf_iter_tcp_batch(struct seq_file *seq)
2860{
2861 struct inet_hashinfo *hinfo = seq_file_net(seq)->ipv4.tcp_death_row.hashinfo;
2862 struct bpf_tcp_iter_state *iter = seq->private;
2863 struct tcp_iter_state *st = &iter->state;
2864 unsigned int expected;
2865 bool resized = false;
2866 struct sock *sk;
2867
2868 /* The st->bucket is done. Directly advance to the next
2869 * bucket instead of having the tcp_seek_last_pos() to skip
2870 * one by one in the current bucket and eventually find out
2871 * it has to advance to the next bucket.
2872 */
2873 if (iter->st_bucket_done) {
2874 st->offset = 0;
2875 st->bucket++;
2876 if (st->state == TCP_SEQ_STATE_LISTENING &&
2877 st->bucket > hinfo->lhash2_mask) {
2878 st->state = TCP_SEQ_STATE_ESTABLISHED;
2879 st->bucket = 0;
2880 }
2881 }
2882
2883again:
2884 /* Get a new batch */
2885 iter->cur_sk = 0;
2886 iter->end_sk = 0;
2887 iter->st_bucket_done = false;
2888
2889 sk = tcp_seek_last_pos(seq);
2890 if (!sk)
2891 return NULL; /* Done */
2892
2893 if (st->state == TCP_SEQ_STATE_LISTENING)
2894 expected = bpf_iter_tcp_listening_batch(seq, sk);
2895 else
2896 expected = bpf_iter_tcp_established_batch(seq, sk);
2897
2898 if (iter->end_sk == expected) {
2899 iter->st_bucket_done = true;
2900 return sk;
2901 }
2902
2903 if (!resized && !bpf_iter_tcp_realloc_batch(iter, expected * 3 / 2)) {
2904 resized = true;
2905 goto again;
2906 }
2907
2908 return sk;
2909}
2910
2911static void *bpf_iter_tcp_seq_start(struct seq_file *seq, loff_t *pos)
2912{
2913 /* bpf iter does not support lseek, so it always
2914 * continue from where it was stop()-ped.
2915 */
2916 if (*pos)
2917 return bpf_iter_tcp_batch(seq);
2918
2919 return SEQ_START_TOKEN;
2920}
2921
2922static void *bpf_iter_tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2923{
2924 struct bpf_tcp_iter_state *iter = seq->private;
2925 struct tcp_iter_state *st = &iter->state;
2926 struct sock *sk;
2927
2928 /* Whenever seq_next() is called, the iter->cur_sk is
2929 * done with seq_show(), so advance to the next sk in
2930 * the batch.
2931 */
2932 if (iter->cur_sk < iter->end_sk) {
2933 /* Keeping st->num consistent in tcp_iter_state.
2934 * bpf_iter_tcp does not use st->num.
2935 * meta.seq_num is used instead.
2936 */
2937 st->num++;
2938 /* Move st->offset to the next sk in the bucket such that
2939 * the future start() will resume at st->offset in
2940 * st->bucket. See tcp_seek_last_pos().
2941 */
2942 st->offset++;
2943 sock_put(iter->batch[iter->cur_sk++]);
2944 }
2945
2946 if (iter->cur_sk < iter->end_sk)
2947 sk = iter->batch[iter->cur_sk];
2948 else
2949 sk = bpf_iter_tcp_batch(seq);
2950
2951 ++*pos;
2952 /* Keeping st->last_pos consistent in tcp_iter_state.
2953 * bpf iter does not do lseek, so st->last_pos always equals to *pos.
2954 */
2955 st->last_pos = *pos;
2956 return sk;
2957}
2958
2959static int bpf_iter_tcp_seq_show(struct seq_file *seq, void *v)
2960{
2961 struct bpf_iter_meta meta;
2962 struct bpf_prog *prog;
2963 struct sock *sk = v;
2964 bool slow;
2965 uid_t uid;
2966 int ret;
2967
2968 if (v == SEQ_START_TOKEN)
2969 return 0;
2970
2971 if (sk_fullsock(sk))
2972 slow = lock_sock_fast(sk);
2973
2974 if (unlikely(sk_unhashed(sk))) {
2975 ret = SEQ_SKIP;
2976 goto unlock;
2977 }
2978
2979 if (sk->sk_state == TCP_TIME_WAIT) {
2980 uid = 0;
2981 } else if (sk->sk_state == TCP_NEW_SYN_RECV) {
2982 const struct request_sock *req = v;
2983
2984 uid = from_kuid_munged(seq_user_ns(seq),
2985 sock_i_uid(req->rsk_listener));
2986 } else {
2987 uid = from_kuid_munged(seq_user_ns(seq), sock_i_uid(sk));
2988 }
2989
2990 meta.seq = seq;
2991 prog = bpf_iter_get_info(&meta, false);
2992 ret = tcp_prog_seq_show(prog, &meta, v, uid);
2993
2994unlock:
2995 if (sk_fullsock(sk))
2996 unlock_sock_fast(sk, slow);
2997 return ret;
2998
2999}
3000
3001static void bpf_iter_tcp_seq_stop(struct seq_file *seq, void *v)
3002{
3003 struct bpf_tcp_iter_state *iter = seq->private;
3004 struct bpf_iter_meta meta;
3005 struct bpf_prog *prog;
3006
3007 if (!v) {
3008 meta.seq = seq;
3009 prog = bpf_iter_get_info(&meta, true);
3010 if (prog)
3011 (void)tcp_prog_seq_show(prog, &meta, v, 0);
3012 }
3013
3014 if (iter->cur_sk < iter->end_sk) {
3015 bpf_iter_tcp_put_batch(iter);
3016 iter->st_bucket_done = false;
3017 }
3018}
3019
3020static const struct seq_operations bpf_iter_tcp_seq_ops = {
3021 .show = bpf_iter_tcp_seq_show,
3022 .start = bpf_iter_tcp_seq_start,
3023 .next = bpf_iter_tcp_seq_next,
3024 .stop = bpf_iter_tcp_seq_stop,
3025};
3026#endif
3027static unsigned short seq_file_family(const struct seq_file *seq)
3028{
3029 const struct tcp_seq_afinfo *afinfo;
3030
3031#ifdef CONFIG_BPF_SYSCALL
3032 /* Iterated from bpf_iter. Let the bpf prog to filter instead. */
3033 if (seq->op == &bpf_iter_tcp_seq_ops)
3034 return AF_UNSPEC;
3035#endif
3036
3037 /* Iterated from proc fs */
3038 afinfo = pde_data(file_inode(seq->file));
3039 return afinfo->family;
3040}
3041
3042static const struct seq_operations tcp4_seq_ops = {
3043 .show = tcp4_seq_show,
3044 .start = tcp_seq_start,
3045 .next = tcp_seq_next,
3046 .stop = tcp_seq_stop,
3047};
3048
3049static struct tcp_seq_afinfo tcp4_seq_afinfo = {
3050 .family = AF_INET,
3051};
3052
3053static int __net_init tcp4_proc_init_net(struct net *net)
3054{
3055 if (!proc_create_net_data("tcp", 0444, net->proc_net, &tcp4_seq_ops,
3056 sizeof(struct tcp_iter_state), &tcp4_seq_afinfo))
3057 return -ENOMEM;
3058 return 0;
3059}
3060
3061static void __net_exit tcp4_proc_exit_net(struct net *net)
3062{
3063 remove_proc_entry("tcp", net->proc_net);
3064}
3065
3066static struct pernet_operations tcp4_net_ops = {
3067 .init = tcp4_proc_init_net,
3068 .exit = tcp4_proc_exit_net,
3069};
3070
3071int __init tcp4_proc_init(void)
3072{
3073 return register_pernet_subsys(&tcp4_net_ops);
3074}
3075
3076void tcp4_proc_exit(void)
3077{
3078 unregister_pernet_subsys(&tcp4_net_ops);
3079}
3080#endif /* CONFIG_PROC_FS */
3081
3082/* @wake is one when sk_stream_write_space() calls us.
3083 * This sends EPOLLOUT only if notsent_bytes is half the limit.
3084 * This mimics the strategy used in sock_def_write_space().
3085 */
3086bool tcp_stream_memory_free(const struct sock *sk, int wake)
3087{
3088 const struct tcp_sock *tp = tcp_sk(sk);
3089 u32 notsent_bytes = READ_ONCE(tp->write_seq) -
3090 READ_ONCE(tp->snd_nxt);
3091
3092 return (notsent_bytes << wake) < tcp_notsent_lowat(tp);
3093}
3094EXPORT_SYMBOL(tcp_stream_memory_free);
3095
3096struct proto tcp_prot = {
3097 .name = "TCP",
3098 .owner = THIS_MODULE,
3099 .close = tcp_close,
3100 .pre_connect = tcp_v4_pre_connect,
3101 .connect = tcp_v4_connect,
3102 .disconnect = tcp_disconnect,
3103 .accept = inet_csk_accept,
3104 .ioctl = tcp_ioctl,
3105 .init = tcp_v4_init_sock,
3106 .destroy = tcp_v4_destroy_sock,
3107 .shutdown = tcp_shutdown,
3108 .setsockopt = tcp_setsockopt,
3109 .getsockopt = tcp_getsockopt,
3110 .bpf_bypass_getsockopt = tcp_bpf_bypass_getsockopt,
3111 .keepalive = tcp_set_keepalive,
3112 .recvmsg = tcp_recvmsg,
3113 .sendmsg = tcp_sendmsg,
3114 .sendpage = tcp_sendpage,
3115 .backlog_rcv = tcp_v4_do_rcv,
3116 .release_cb = tcp_release_cb,
3117 .hash = inet_hash,
3118 .unhash = inet_unhash,
3119 .get_port = inet_csk_get_port,
3120 .put_port = inet_put_port,
3121#ifdef CONFIG_BPF_SYSCALL
3122 .psock_update_sk_prot = tcp_bpf_update_proto,
3123#endif
3124 .enter_memory_pressure = tcp_enter_memory_pressure,
3125 .leave_memory_pressure = tcp_leave_memory_pressure,
3126 .stream_memory_free = tcp_stream_memory_free,
3127 .sockets_allocated = &tcp_sockets_allocated,
3128 .orphan_count = &tcp_orphan_count,
3129
3130 .memory_allocated = &tcp_memory_allocated,
3131 .per_cpu_fw_alloc = &tcp_memory_per_cpu_fw_alloc,
3132
3133 .memory_pressure = &tcp_memory_pressure,
3134 .sysctl_mem = sysctl_tcp_mem,
3135 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
3136 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
3137 .max_header = MAX_TCP_HEADER,
3138 .obj_size = sizeof(struct tcp_sock),
3139 .slab_flags = SLAB_TYPESAFE_BY_RCU,
3140 .twsk_prot = &tcp_timewait_sock_ops,
3141 .rsk_prot = &tcp_request_sock_ops,
3142 .h.hashinfo = NULL,
3143 .no_autobind = true,
3144 .diag_destroy = tcp_abort,
3145};
3146EXPORT_SYMBOL(tcp_prot);
3147
3148static void __net_exit tcp_sk_exit(struct net *net)
3149{
3150 if (net->ipv4.tcp_congestion_control)
3151 bpf_module_put(net->ipv4.tcp_congestion_control,
3152 net->ipv4.tcp_congestion_control->owner);
3153}
3154
3155static void __net_init tcp_set_hashinfo(struct net *net)
3156{
3157 struct inet_hashinfo *hinfo;
3158 unsigned int ehash_entries;
3159 struct net *old_net;
3160
3161 if (net_eq(net, &init_net))
3162 goto fallback;
3163
3164 old_net = current->nsproxy->net_ns;
3165 ehash_entries = READ_ONCE(old_net->ipv4.sysctl_tcp_child_ehash_entries);
3166 if (!ehash_entries)
3167 goto fallback;
3168
3169 ehash_entries = roundup_pow_of_two(ehash_entries);
3170 hinfo = inet_pernet_hashinfo_alloc(&tcp_hashinfo, ehash_entries);
3171 if (!hinfo) {
3172 pr_warn("Failed to allocate TCP ehash (entries: %u) "
3173 "for a netns, fallback to the global one\n",
3174 ehash_entries);
3175fallback:
3176 hinfo = &tcp_hashinfo;
3177 ehash_entries = tcp_hashinfo.ehash_mask + 1;
3178 }
3179
3180 net->ipv4.tcp_death_row.hashinfo = hinfo;
3181 net->ipv4.tcp_death_row.sysctl_max_tw_buckets = ehash_entries / 2;
3182 net->ipv4.sysctl_max_syn_backlog = max(128U, ehash_entries / 128);
3183}
3184
3185static int __net_init tcp_sk_init(struct net *net)
3186{
3187 net->ipv4.sysctl_tcp_ecn = 2;
3188 net->ipv4.sysctl_tcp_ecn_fallback = 1;
3189
3190 net->ipv4.sysctl_tcp_base_mss = TCP_BASE_MSS;
3191 net->ipv4.sysctl_tcp_min_snd_mss = TCP_MIN_SND_MSS;
3192 net->ipv4.sysctl_tcp_probe_threshold = TCP_PROBE_THRESHOLD;
3193 net->ipv4.sysctl_tcp_probe_interval = TCP_PROBE_INTERVAL;
3194 net->ipv4.sysctl_tcp_mtu_probe_floor = TCP_MIN_SND_MSS;
3195
3196 net->ipv4.sysctl_tcp_keepalive_time = TCP_KEEPALIVE_TIME;
3197 net->ipv4.sysctl_tcp_keepalive_probes = TCP_KEEPALIVE_PROBES;
3198 net->ipv4.sysctl_tcp_keepalive_intvl = TCP_KEEPALIVE_INTVL;
3199
3200 net->ipv4.sysctl_tcp_syn_retries = TCP_SYN_RETRIES;
3201 net->ipv4.sysctl_tcp_synack_retries = TCP_SYNACK_RETRIES;
3202 net->ipv4.sysctl_tcp_syncookies = 1;
3203 net->ipv4.sysctl_tcp_reordering = TCP_FASTRETRANS_THRESH;
3204 net->ipv4.sysctl_tcp_retries1 = TCP_RETR1;
3205 net->ipv4.sysctl_tcp_retries2 = TCP_RETR2;
3206 net->ipv4.sysctl_tcp_orphan_retries = 0;
3207 net->ipv4.sysctl_tcp_fin_timeout = TCP_FIN_TIMEOUT;
3208 net->ipv4.sysctl_tcp_notsent_lowat = UINT_MAX;
3209 net->ipv4.sysctl_tcp_tw_reuse = 2;
3210 net->ipv4.sysctl_tcp_no_ssthresh_metrics_save = 1;
3211
3212 refcount_set(&net->ipv4.tcp_death_row.tw_refcount, 1);
3213 tcp_set_hashinfo(net);
3214
3215 net->ipv4.sysctl_tcp_sack = 1;
3216 net->ipv4.sysctl_tcp_window_scaling = 1;
3217 net->ipv4.sysctl_tcp_timestamps = 1;
3218 net->ipv4.sysctl_tcp_early_retrans = 3;
3219 net->ipv4.sysctl_tcp_recovery = TCP_RACK_LOSS_DETECTION;
3220 net->ipv4.sysctl_tcp_slow_start_after_idle = 1; /* By default, RFC2861 behavior. */
3221 net->ipv4.sysctl_tcp_retrans_collapse = 1;
3222 net->ipv4.sysctl_tcp_max_reordering = 300;
3223 net->ipv4.sysctl_tcp_dsack = 1;
3224 net->ipv4.sysctl_tcp_app_win = 31;
3225 net->ipv4.sysctl_tcp_adv_win_scale = 1;
3226 net->ipv4.sysctl_tcp_frto = 2;
3227 net->ipv4.sysctl_tcp_moderate_rcvbuf = 1;
3228 /* This limits the percentage of the congestion window which we
3229 * will allow a single TSO frame to consume. Building TSO frames
3230 * which are too large can cause TCP streams to be bursty.
3231 */
3232 net->ipv4.sysctl_tcp_tso_win_divisor = 3;
3233 /* Default TSQ limit of 16 TSO segments */
3234 net->ipv4.sysctl_tcp_limit_output_bytes = 16 * 65536;
3235
3236 /* rfc5961 challenge ack rate limiting, per net-ns, disabled by default. */
3237 net->ipv4.sysctl_tcp_challenge_ack_limit = INT_MAX;
3238
3239 net->ipv4.sysctl_tcp_min_tso_segs = 2;
3240 net->ipv4.sysctl_tcp_tso_rtt_log = 9; /* 2^9 = 512 usec */
3241 net->ipv4.sysctl_tcp_min_rtt_wlen = 300;
3242 net->ipv4.sysctl_tcp_autocorking = 1;
3243 net->ipv4.sysctl_tcp_invalid_ratelimit = HZ/2;
3244 net->ipv4.sysctl_tcp_pacing_ss_ratio = 200;
3245 net->ipv4.sysctl_tcp_pacing_ca_ratio = 120;
3246 if (net != &init_net) {
3247 memcpy(net->ipv4.sysctl_tcp_rmem,
3248 init_net.ipv4.sysctl_tcp_rmem,
3249 sizeof(init_net.ipv4.sysctl_tcp_rmem));
3250 memcpy(net->ipv4.sysctl_tcp_wmem,
3251 init_net.ipv4.sysctl_tcp_wmem,
3252 sizeof(init_net.ipv4.sysctl_tcp_wmem));
3253 }
3254 net->ipv4.sysctl_tcp_comp_sack_delay_ns = NSEC_PER_MSEC;
3255 net->ipv4.sysctl_tcp_comp_sack_slack_ns = 100 * NSEC_PER_USEC;
3256 net->ipv4.sysctl_tcp_comp_sack_nr = 44;
3257 net->ipv4.sysctl_tcp_fastopen = TFO_CLIENT_ENABLE;
3258 net->ipv4.sysctl_tcp_fastopen_blackhole_timeout = 0;
3259 atomic_set(&net->ipv4.tfo_active_disable_times, 0);
3260
3261 /* Set default values for PLB */
3262 net->ipv4.sysctl_tcp_plb_enabled = 0; /* Disabled by default */
3263 net->ipv4.sysctl_tcp_plb_idle_rehash_rounds = 3;
3264 net->ipv4.sysctl_tcp_plb_rehash_rounds = 12;
3265 net->ipv4.sysctl_tcp_plb_suspend_rto_sec = 60;
3266 /* Default congestion threshold for PLB to mark a round is 50% */
3267 net->ipv4.sysctl_tcp_plb_cong_thresh = (1 << TCP_PLB_SCALE) / 2;
3268
3269 /* Reno is always built in */
3270 if (!net_eq(net, &init_net) &&
3271 bpf_try_module_get(init_net.ipv4.tcp_congestion_control,
3272 init_net.ipv4.tcp_congestion_control->owner))
3273 net->ipv4.tcp_congestion_control = init_net.ipv4.tcp_congestion_control;
3274 else
3275 net->ipv4.tcp_congestion_control = &tcp_reno;
3276
3277 return 0;
3278}
3279
3280static void __net_exit tcp_sk_exit_batch(struct list_head *net_exit_list)
3281{
3282 struct net *net;
3283
3284 tcp_twsk_purge(net_exit_list, AF_INET);
3285
3286 list_for_each_entry(net, net_exit_list, exit_list) {
3287 inet_pernet_hashinfo_free(net->ipv4.tcp_death_row.hashinfo);
3288 WARN_ON_ONCE(!refcount_dec_and_test(&net->ipv4.tcp_death_row.tw_refcount));
3289 tcp_fastopen_ctx_destroy(net);
3290 }
3291}
3292
3293static struct pernet_operations __net_initdata tcp_sk_ops = {
3294 .init = tcp_sk_init,
3295 .exit = tcp_sk_exit,
3296 .exit_batch = tcp_sk_exit_batch,
3297};
3298
3299#if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_PROC_FS)
3300DEFINE_BPF_ITER_FUNC(tcp, struct bpf_iter_meta *meta,
3301 struct sock_common *sk_common, uid_t uid)
3302
3303#define INIT_BATCH_SZ 16
3304
3305static int bpf_iter_init_tcp(void *priv_data, struct bpf_iter_aux_info *aux)
3306{
3307 struct bpf_tcp_iter_state *iter = priv_data;
3308 int err;
3309
3310 err = bpf_iter_init_seq_net(priv_data, aux);
3311 if (err)
3312 return err;
3313
3314 err = bpf_iter_tcp_realloc_batch(iter, INIT_BATCH_SZ);
3315 if (err) {
3316 bpf_iter_fini_seq_net(priv_data);
3317 return err;
3318 }
3319
3320 return 0;
3321}
3322
3323static void bpf_iter_fini_tcp(void *priv_data)
3324{
3325 struct bpf_tcp_iter_state *iter = priv_data;
3326
3327 bpf_iter_fini_seq_net(priv_data);
3328 kvfree(iter->batch);
3329}
3330
3331static const struct bpf_iter_seq_info tcp_seq_info = {
3332 .seq_ops = &bpf_iter_tcp_seq_ops,
3333 .init_seq_private = bpf_iter_init_tcp,
3334 .fini_seq_private = bpf_iter_fini_tcp,
3335 .seq_priv_size = sizeof(struct bpf_tcp_iter_state),
3336};
3337
3338static const struct bpf_func_proto *
3339bpf_iter_tcp_get_func_proto(enum bpf_func_id func_id,
3340 const struct bpf_prog *prog)
3341{
3342 switch (func_id) {
3343 case BPF_FUNC_setsockopt:
3344 return &bpf_sk_setsockopt_proto;
3345 case BPF_FUNC_getsockopt:
3346 return &bpf_sk_getsockopt_proto;
3347 default:
3348 return NULL;
3349 }
3350}
3351
3352static struct bpf_iter_reg tcp_reg_info = {
3353 .target = "tcp",
3354 .ctx_arg_info_size = 1,
3355 .ctx_arg_info = {
3356 { offsetof(struct bpf_iter__tcp, sk_common),
3357 PTR_TO_BTF_ID_OR_NULL },
3358 },
3359 .get_func_proto = bpf_iter_tcp_get_func_proto,
3360 .seq_info = &tcp_seq_info,
3361};
3362
3363static void __init bpf_iter_register(void)
3364{
3365 tcp_reg_info.ctx_arg_info[0].btf_id = btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON];
3366 if (bpf_iter_reg_target(&tcp_reg_info))
3367 pr_warn("Warning: could not register bpf iterator tcp\n");
3368}
3369
3370#endif
3371
3372void __init tcp_v4_init(void)
3373{
3374 int cpu, res;
3375
3376 for_each_possible_cpu(cpu) {
3377 struct sock *sk;
3378
3379 res = inet_ctl_sock_create(&sk, PF_INET, SOCK_RAW,
3380 IPPROTO_TCP, &init_net);
3381 if (res)
3382 panic("Failed to create the TCP control socket.\n");
3383 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
3384
3385 /* Please enforce IP_DF and IPID==0 for RST and
3386 * ACK sent in SYN-RECV and TIME-WAIT state.
3387 */
3388 inet_sk(sk)->pmtudisc = IP_PMTUDISC_DO;
3389
3390 per_cpu(ipv4_tcp_sk, cpu) = sk;
3391 }
3392 if (register_pernet_subsys(&tcp_sk_ops))
3393 panic("Failed to create the TCP control socket.\n");
3394
3395#if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_PROC_FS)
3396 bpf_iter_register();
3397#endif
3398}