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/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * IPv4 specific functions
9 *
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 * This program is free software; you can redistribute it and/or
19 * modify it under the terms of the GNU General Public License
20 * as published by the Free Software Foundation; either version
21 * 2 of the License, or (at your option) any later version.
22 */
23
24/*
25 * Changes:
26 * David S. Miller : New socket lookup architecture.
27 * This code is dedicated to John Dyson.
28 * David S. Miller : Change semantics of established hash,
29 * half is devoted to TIME_WAIT sockets
30 * and the rest go in the other half.
31 * Andi Kleen : Add support for syncookies and fixed
32 * some bugs: ip options weren't passed to
33 * the TCP layer, missed a check for an
34 * ACK bit.
35 * Andi Kleen : Implemented fast path mtu discovery.
36 * Fixed many serious bugs in the
37 * request_sock handling and moved
38 * most of it into the af independent code.
39 * Added tail drop and some other bugfixes.
40 * Added new listen semantics.
41 * Mike McLagan : Routing by source
42 * Juan Jose Ciarlante: ip_dynaddr bits
43 * Andi Kleen: various fixes.
44 * Vitaly E. Lavrov : Transparent proxy revived after year
45 * coma.
46 * Andi Kleen : Fix new listen.
47 * Andi Kleen : Fix accept error reporting.
48 * YOSHIFUJI Hideaki @USAGI and: Support IPV6_V6ONLY socket option, which
49 * Alexey Kuznetsov allow both IPv4 and IPv6 sockets to bind
50 * a single port at the same time.
51 */
52
53#define pr_fmt(fmt) "TCP: " fmt
54
55#include <linux/bottom_half.h>
56#include <linux/types.h>
57#include <linux/fcntl.h>
58#include <linux/module.h>
59#include <linux/random.h>
60#include <linux/cache.h>
61#include <linux/jhash.h>
62#include <linux/init.h>
63#include <linux/times.h>
64#include <linux/slab.h>
65
66#include <net/net_namespace.h>
67#include <net/icmp.h>
68#include <net/inet_hashtables.h>
69#include <net/tcp.h>
70#include <net/transp_v6.h>
71#include <net/ipv6.h>
72#include <net/inet_common.h>
73#include <net/timewait_sock.h>
74#include <net/xfrm.h>
75#include <net/secure_seq.h>
76#include <net/busy_poll.h>
77
78#include <linux/inet.h>
79#include <linux/ipv6.h>
80#include <linux/stddef.h>
81#include <linux/proc_fs.h>
82#include <linux/seq_file.h>
83#include <linux/inetdevice.h>
84
85#include <crypto/hash.h>
86#include <linux/scatterlist.h>
87
88#include <trace/events/tcp.h>
89
90#ifdef CONFIG_TCP_MD5SIG
91static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
92 __be32 daddr, __be32 saddr, const struct tcphdr *th);
93#endif
94
95struct inet_hashinfo tcp_hashinfo;
96EXPORT_SYMBOL(tcp_hashinfo);
97
98static u32 tcp_v4_init_seq(const struct sk_buff *skb)
99{
100 return secure_tcp_seq(ip_hdr(skb)->daddr,
101 ip_hdr(skb)->saddr,
102 tcp_hdr(skb)->dest,
103 tcp_hdr(skb)->source);
104}
105
106static u32 tcp_v4_init_ts_off(const struct net *net, const struct sk_buff *skb)
107{
108 return secure_tcp_ts_off(net, ip_hdr(skb)->daddr, ip_hdr(skb)->saddr);
109}
110
111int tcp_twsk_unique(struct sock *sk, struct sock *sktw, void *twp)
112{
113 const struct tcp_timewait_sock *tcptw = tcp_twsk(sktw);
114 struct tcp_sock *tp = tcp_sk(sk);
115
116 /* With PAWS, it is safe from the viewpoint
117 of data integrity. Even without PAWS it is safe provided sequence
118 spaces do not overlap i.e. at data rates <= 80Mbit/sec.
119
120 Actually, the idea is close to VJ's one, only timestamp cache is
121 held not per host, but per port pair and TW bucket is used as state
122 holder.
123
124 If TW bucket has been already destroyed we fall back to VJ's scheme
125 and use initial timestamp retrieved from peer table.
126 */
127 if (tcptw->tw_ts_recent_stamp &&
128 (!twp || (sock_net(sk)->ipv4.sysctl_tcp_tw_reuse &&
129 get_seconds() - tcptw->tw_ts_recent_stamp > 1))) {
130 tp->write_seq = tcptw->tw_snd_nxt + 65535 + 2;
131 if (tp->write_seq == 0)
132 tp->write_seq = 1;
133 tp->rx_opt.ts_recent = tcptw->tw_ts_recent;
134 tp->rx_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
135 sock_hold(sktw);
136 return 1;
137 }
138
139 return 0;
140}
141EXPORT_SYMBOL_GPL(tcp_twsk_unique);
142
143static int tcp_v4_pre_connect(struct sock *sk, struct sockaddr *uaddr,
144 int addr_len)
145{
146 /* This check is replicated from tcp_v4_connect() and intended to
147 * prevent BPF program called below from accessing bytes that are out
148 * of the bound specified by user in addr_len.
149 */
150 if (addr_len < sizeof(struct sockaddr_in))
151 return -EINVAL;
152
153 sock_owned_by_me(sk);
154
155 return BPF_CGROUP_RUN_PROG_INET4_CONNECT(sk, uaddr);
156}
157
158/* This will initiate an outgoing connection. */
159int tcp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
160{
161 struct sockaddr_in *usin = (struct sockaddr_in *)uaddr;
162 struct inet_sock *inet = inet_sk(sk);
163 struct tcp_sock *tp = tcp_sk(sk);
164 __be16 orig_sport, orig_dport;
165 __be32 daddr, nexthop;
166 struct flowi4 *fl4;
167 struct rtable *rt;
168 int err;
169 struct ip_options_rcu *inet_opt;
170 struct inet_timewait_death_row *tcp_death_row = &sock_net(sk)->ipv4.tcp_death_row;
171
172 if (addr_len < sizeof(struct sockaddr_in))
173 return -EINVAL;
174
175 if (usin->sin_family != AF_INET)
176 return -EAFNOSUPPORT;
177
178 nexthop = daddr = usin->sin_addr.s_addr;
179 inet_opt = rcu_dereference_protected(inet->inet_opt,
180 lockdep_sock_is_held(sk));
181 if (inet_opt && inet_opt->opt.srr) {
182 if (!daddr)
183 return -EINVAL;
184 nexthop = inet_opt->opt.faddr;
185 }
186
187 orig_sport = inet->inet_sport;
188 orig_dport = usin->sin_port;
189 fl4 = &inet->cork.fl.u.ip4;
190 rt = ip_route_connect(fl4, nexthop, inet->inet_saddr,
191 RT_CONN_FLAGS(sk), sk->sk_bound_dev_if,
192 IPPROTO_TCP,
193 orig_sport, orig_dport, sk);
194 if (IS_ERR(rt)) {
195 err = PTR_ERR(rt);
196 if (err == -ENETUNREACH)
197 IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
198 return err;
199 }
200
201 if (rt->rt_flags & (RTCF_MULTICAST | RTCF_BROADCAST)) {
202 ip_rt_put(rt);
203 return -ENETUNREACH;
204 }
205
206 if (!inet_opt || !inet_opt->opt.srr)
207 daddr = fl4->daddr;
208
209 if (!inet->inet_saddr)
210 inet->inet_saddr = fl4->saddr;
211 sk_rcv_saddr_set(sk, inet->inet_saddr);
212
213 if (tp->rx_opt.ts_recent_stamp && inet->inet_daddr != daddr) {
214 /* Reset inherited state */
215 tp->rx_opt.ts_recent = 0;
216 tp->rx_opt.ts_recent_stamp = 0;
217 if (likely(!tp->repair))
218 tp->write_seq = 0;
219 }
220
221 inet->inet_dport = usin->sin_port;
222 sk_daddr_set(sk, daddr);
223
224 inet_csk(sk)->icsk_ext_hdr_len = 0;
225 if (inet_opt)
226 inet_csk(sk)->icsk_ext_hdr_len = inet_opt->opt.optlen;
227
228 tp->rx_opt.mss_clamp = TCP_MSS_DEFAULT;
229
230 /* Socket identity is still unknown (sport may be zero).
231 * However we set state to SYN-SENT and not releasing socket
232 * lock select source port, enter ourselves into the hash tables and
233 * complete initialization after this.
234 */
235 tcp_set_state(sk, TCP_SYN_SENT);
236 err = inet_hash_connect(tcp_death_row, sk);
237 if (err)
238 goto failure;
239
240 sk_set_txhash(sk);
241
242 rt = ip_route_newports(fl4, rt, orig_sport, orig_dport,
243 inet->inet_sport, inet->inet_dport, sk);
244 if (IS_ERR(rt)) {
245 err = PTR_ERR(rt);
246 rt = NULL;
247 goto failure;
248 }
249 /* OK, now commit destination to socket. */
250 sk->sk_gso_type = SKB_GSO_TCPV4;
251 sk_setup_caps(sk, &rt->dst);
252 rt = NULL;
253
254 if (likely(!tp->repair)) {
255 if (!tp->write_seq)
256 tp->write_seq = secure_tcp_seq(inet->inet_saddr,
257 inet->inet_daddr,
258 inet->inet_sport,
259 usin->sin_port);
260 tp->tsoffset = secure_tcp_ts_off(sock_net(sk),
261 inet->inet_saddr,
262 inet->inet_daddr);
263 }
264
265 inet->inet_id = tp->write_seq ^ jiffies;
266
267 if (tcp_fastopen_defer_connect(sk, &err))
268 return err;
269 if (err)
270 goto failure;
271
272 err = tcp_connect(sk);
273
274 if (err)
275 goto failure;
276
277 return 0;
278
279failure:
280 /*
281 * This unhashes the socket and releases the local port,
282 * if necessary.
283 */
284 tcp_set_state(sk, TCP_CLOSE);
285 ip_rt_put(rt);
286 sk->sk_route_caps = 0;
287 inet->inet_dport = 0;
288 return err;
289}
290EXPORT_SYMBOL(tcp_v4_connect);
291
292/*
293 * This routine reacts to ICMP_FRAG_NEEDED mtu indications as defined in RFC1191.
294 * It can be called through tcp_release_cb() if socket was owned by user
295 * at the time tcp_v4_err() was called to handle ICMP message.
296 */
297void tcp_v4_mtu_reduced(struct sock *sk)
298{
299 struct inet_sock *inet = inet_sk(sk);
300 struct dst_entry *dst;
301 u32 mtu;
302
303 if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE))
304 return;
305 mtu = tcp_sk(sk)->mtu_info;
306 dst = inet_csk_update_pmtu(sk, mtu);
307 if (!dst)
308 return;
309
310 /* Something is about to be wrong... Remember soft error
311 * for the case, if this connection will not able to recover.
312 */
313 if (mtu < dst_mtu(dst) && ip_dont_fragment(sk, dst))
314 sk->sk_err_soft = EMSGSIZE;
315
316 mtu = dst_mtu(dst);
317
318 if (inet->pmtudisc != IP_PMTUDISC_DONT &&
319 ip_sk_accept_pmtu(sk) &&
320 inet_csk(sk)->icsk_pmtu_cookie > mtu) {
321 tcp_sync_mss(sk, mtu);
322
323 /* Resend the TCP packet because it's
324 * clear that the old packet has been
325 * dropped. This is the new "fast" path mtu
326 * discovery.
327 */
328 tcp_simple_retransmit(sk);
329 } /* else let the usual retransmit timer handle it */
330}
331EXPORT_SYMBOL(tcp_v4_mtu_reduced);
332
333static void do_redirect(struct sk_buff *skb, struct sock *sk)
334{
335 struct dst_entry *dst = __sk_dst_check(sk, 0);
336
337 if (dst)
338 dst->ops->redirect(dst, sk, skb);
339}
340
341
342/* handle ICMP messages on TCP_NEW_SYN_RECV request sockets */
343void tcp_req_err(struct sock *sk, u32 seq, bool abort)
344{
345 struct request_sock *req = inet_reqsk(sk);
346 struct net *net = sock_net(sk);
347
348 /* ICMPs are not backlogged, hence we cannot get
349 * an established socket here.
350 */
351 if (seq != tcp_rsk(req)->snt_isn) {
352 __NET_INC_STATS(net, LINUX_MIB_OUTOFWINDOWICMPS);
353 } else if (abort) {
354 /*
355 * Still in SYN_RECV, just remove it silently.
356 * There is no good way to pass the error to the newly
357 * created socket, and POSIX does not want network
358 * errors returned from accept().
359 */
360 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
361 tcp_listendrop(req->rsk_listener);
362 }
363 reqsk_put(req);
364}
365EXPORT_SYMBOL(tcp_req_err);
366
367/*
368 * This routine is called by the ICMP module when it gets some
369 * sort of error condition. If err < 0 then the socket should
370 * be closed and the error returned to the user. If err > 0
371 * it's just the icmp type << 8 | icmp code. After adjustment
372 * header points to the first 8 bytes of the tcp header. We need
373 * to find the appropriate port.
374 *
375 * The locking strategy used here is very "optimistic". When
376 * someone else accesses the socket the ICMP is just dropped
377 * and for some paths there is no check at all.
378 * A more general error queue to queue errors for later handling
379 * is probably better.
380 *
381 */
382
383void tcp_v4_err(struct sk_buff *icmp_skb, u32 info)
384{
385 const struct iphdr *iph = (const struct iphdr *)icmp_skb->data;
386 struct tcphdr *th = (struct tcphdr *)(icmp_skb->data + (iph->ihl << 2));
387 struct inet_connection_sock *icsk;
388 struct tcp_sock *tp;
389 struct inet_sock *inet;
390 const int type = icmp_hdr(icmp_skb)->type;
391 const int code = icmp_hdr(icmp_skb)->code;
392 struct sock *sk;
393 struct sk_buff *skb;
394 struct request_sock *fastopen;
395 u32 seq, snd_una;
396 s32 remaining;
397 u32 delta_us;
398 int err;
399 struct net *net = dev_net(icmp_skb->dev);
400
401 sk = __inet_lookup_established(net, &tcp_hashinfo, iph->daddr,
402 th->dest, iph->saddr, ntohs(th->source),
403 inet_iif(icmp_skb), 0);
404 if (!sk) {
405 __ICMP_INC_STATS(net, ICMP_MIB_INERRORS);
406 return;
407 }
408 if (sk->sk_state == TCP_TIME_WAIT) {
409 inet_twsk_put(inet_twsk(sk));
410 return;
411 }
412 seq = ntohl(th->seq);
413 if (sk->sk_state == TCP_NEW_SYN_RECV)
414 return tcp_req_err(sk, seq,
415 type == ICMP_PARAMETERPROB ||
416 type == ICMP_TIME_EXCEEDED ||
417 (type == ICMP_DEST_UNREACH &&
418 (code == ICMP_NET_UNREACH ||
419 code == ICMP_HOST_UNREACH)));
420
421 bh_lock_sock(sk);
422 /* If too many ICMPs get dropped on busy
423 * servers this needs to be solved differently.
424 * We do take care of PMTU discovery (RFC1191) special case :
425 * we can receive locally generated ICMP messages while socket is held.
426 */
427 if (sock_owned_by_user(sk)) {
428 if (!(type == ICMP_DEST_UNREACH && code == ICMP_FRAG_NEEDED))
429 __NET_INC_STATS(net, LINUX_MIB_LOCKDROPPEDICMPS);
430 }
431 if (sk->sk_state == TCP_CLOSE)
432 goto out;
433
434 if (unlikely(iph->ttl < inet_sk(sk)->min_ttl)) {
435 __NET_INC_STATS(net, LINUX_MIB_TCPMINTTLDROP);
436 goto out;
437 }
438
439 icsk = inet_csk(sk);
440 tp = tcp_sk(sk);
441 /* XXX (TFO) - tp->snd_una should be ISN (tcp_create_openreq_child() */
442 fastopen = tp->fastopen_rsk;
443 snd_una = fastopen ? tcp_rsk(fastopen)->snt_isn : tp->snd_una;
444 if (sk->sk_state != TCP_LISTEN &&
445 !between(seq, snd_una, tp->snd_nxt)) {
446 __NET_INC_STATS(net, LINUX_MIB_OUTOFWINDOWICMPS);
447 goto out;
448 }
449
450 switch (type) {
451 case ICMP_REDIRECT:
452 if (!sock_owned_by_user(sk))
453 do_redirect(icmp_skb, sk);
454 goto out;
455 case ICMP_SOURCE_QUENCH:
456 /* Just silently ignore these. */
457 goto out;
458 case ICMP_PARAMETERPROB:
459 err = EPROTO;
460 break;
461 case ICMP_DEST_UNREACH:
462 if (code > NR_ICMP_UNREACH)
463 goto out;
464
465 if (code == ICMP_FRAG_NEEDED) { /* PMTU discovery (RFC1191) */
466 /* We are not interested in TCP_LISTEN and open_requests
467 * (SYN-ACKs send out by Linux are always <576bytes so
468 * they should go through unfragmented).
469 */
470 if (sk->sk_state == TCP_LISTEN)
471 goto out;
472
473 tp->mtu_info = info;
474 if (!sock_owned_by_user(sk)) {
475 tcp_v4_mtu_reduced(sk);
476 } else {
477 if (!test_and_set_bit(TCP_MTU_REDUCED_DEFERRED, &sk->sk_tsq_flags))
478 sock_hold(sk);
479 }
480 goto out;
481 }
482
483 err = icmp_err_convert[code].errno;
484 /* check if icmp_skb allows revert of backoff
485 * (see draft-zimmermann-tcp-lcd) */
486 if (code != ICMP_NET_UNREACH && code != ICMP_HOST_UNREACH)
487 break;
488 if (seq != tp->snd_una || !icsk->icsk_retransmits ||
489 !icsk->icsk_backoff || fastopen)
490 break;
491
492 if (sock_owned_by_user(sk))
493 break;
494
495 icsk->icsk_backoff--;
496 icsk->icsk_rto = tp->srtt_us ? __tcp_set_rto(tp) :
497 TCP_TIMEOUT_INIT;
498 icsk->icsk_rto = inet_csk_rto_backoff(icsk, TCP_RTO_MAX);
499
500 skb = tcp_rtx_queue_head(sk);
501 BUG_ON(!skb);
502
503 tcp_mstamp_refresh(tp);
504 delta_us = (u32)(tp->tcp_mstamp - skb->skb_mstamp);
505 remaining = icsk->icsk_rto -
506 usecs_to_jiffies(delta_us);
507
508 if (remaining > 0) {
509 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
510 remaining, TCP_RTO_MAX);
511 } else {
512 /* RTO revert clocked out retransmission.
513 * Will retransmit now */
514 tcp_retransmit_timer(sk);
515 }
516
517 break;
518 case ICMP_TIME_EXCEEDED:
519 err = EHOSTUNREACH;
520 break;
521 default:
522 goto out;
523 }
524
525 switch (sk->sk_state) {
526 case TCP_SYN_SENT:
527 case TCP_SYN_RECV:
528 /* Only in fast or simultaneous open. If a fast open socket is
529 * is already accepted it is treated as a connected one below.
530 */
531 if (fastopen && !fastopen->sk)
532 break;
533
534 if (!sock_owned_by_user(sk)) {
535 sk->sk_err = err;
536
537 sk->sk_error_report(sk);
538
539 tcp_done(sk);
540 } else {
541 sk->sk_err_soft = err;
542 }
543 goto out;
544 }
545
546 /* If we've already connected we will keep trying
547 * until we time out, or the user gives up.
548 *
549 * rfc1122 4.2.3.9 allows to consider as hard errors
550 * only PROTO_UNREACH and PORT_UNREACH (well, FRAG_FAILED too,
551 * but it is obsoleted by pmtu discovery).
552 *
553 * Note, that in modern internet, where routing is unreliable
554 * and in each dark corner broken firewalls sit, sending random
555 * errors ordered by their masters even this two messages finally lose
556 * their original sense (even Linux sends invalid PORT_UNREACHs)
557 *
558 * Now we are in compliance with RFCs.
559 * --ANK (980905)
560 */
561
562 inet = inet_sk(sk);
563 if (!sock_owned_by_user(sk) && inet->recverr) {
564 sk->sk_err = err;
565 sk->sk_error_report(sk);
566 } else { /* Only an error on timeout */
567 sk->sk_err_soft = err;
568 }
569
570out:
571 bh_unlock_sock(sk);
572 sock_put(sk);
573}
574
575void __tcp_v4_send_check(struct sk_buff *skb, __be32 saddr, __be32 daddr)
576{
577 struct tcphdr *th = tcp_hdr(skb);
578
579 th->check = ~tcp_v4_check(skb->len, saddr, daddr, 0);
580 skb->csum_start = skb_transport_header(skb) - skb->head;
581 skb->csum_offset = offsetof(struct tcphdr, check);
582}
583
584/* This routine computes an IPv4 TCP checksum. */
585void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)
586{
587 const struct inet_sock *inet = inet_sk(sk);
588
589 __tcp_v4_send_check(skb, inet->inet_saddr, inet->inet_daddr);
590}
591EXPORT_SYMBOL(tcp_v4_send_check);
592
593/*
594 * This routine will send an RST to the other tcp.
595 *
596 * Someone asks: why I NEVER use socket parameters (TOS, TTL etc.)
597 * for reset.
598 * Answer: if a packet caused RST, it is not for a socket
599 * existing in our system, if it is matched to a socket,
600 * it is just duplicate segment or bug in other side's TCP.
601 * So that we build reply only basing on parameters
602 * arrived with segment.
603 * Exception: precedence violation. We do not implement it in any case.
604 */
605
606static void tcp_v4_send_reset(const struct sock *sk, struct sk_buff *skb)
607{
608 const struct tcphdr *th = tcp_hdr(skb);
609 struct {
610 struct tcphdr th;
611#ifdef CONFIG_TCP_MD5SIG
612 __be32 opt[(TCPOLEN_MD5SIG_ALIGNED >> 2)];
613#endif
614 } rep;
615 struct ip_reply_arg arg;
616#ifdef CONFIG_TCP_MD5SIG
617 struct tcp_md5sig_key *key = NULL;
618 const __u8 *hash_location = NULL;
619 unsigned char newhash[16];
620 int genhash;
621 struct sock *sk1 = NULL;
622#endif
623 struct net *net;
624
625 /* Never send a reset in response to a reset. */
626 if (th->rst)
627 return;
628
629 /* If sk not NULL, it means we did a successful lookup and incoming
630 * route had to be correct. prequeue might have dropped our dst.
631 */
632 if (!sk && skb_rtable(skb)->rt_type != RTN_LOCAL)
633 return;
634
635 /* Swap the send and the receive. */
636 memset(&rep, 0, sizeof(rep));
637 rep.th.dest = th->source;
638 rep.th.source = th->dest;
639 rep.th.doff = sizeof(struct tcphdr) / 4;
640 rep.th.rst = 1;
641
642 if (th->ack) {
643 rep.th.seq = th->ack_seq;
644 } else {
645 rep.th.ack = 1;
646 rep.th.ack_seq = htonl(ntohl(th->seq) + th->syn + th->fin +
647 skb->len - (th->doff << 2));
648 }
649
650 memset(&arg, 0, sizeof(arg));
651 arg.iov[0].iov_base = (unsigned char *)&rep;
652 arg.iov[0].iov_len = sizeof(rep.th);
653
654 net = sk ? sock_net(sk) : dev_net(skb_dst(skb)->dev);
655#ifdef CONFIG_TCP_MD5SIG
656 rcu_read_lock();
657 hash_location = tcp_parse_md5sig_option(th);
658 if (sk && sk_fullsock(sk)) {
659 key = tcp_md5_do_lookup(sk, (union tcp_md5_addr *)
660 &ip_hdr(skb)->saddr, AF_INET);
661 } else if (hash_location) {
662 /*
663 * active side is lost. Try to find listening socket through
664 * source port, and then find md5 key through listening socket.
665 * we are not loose security here:
666 * Incoming packet is checked with md5 hash with finding key,
667 * no RST generated if md5 hash doesn't match.
668 */
669 sk1 = __inet_lookup_listener(net, &tcp_hashinfo, NULL, 0,
670 ip_hdr(skb)->saddr,
671 th->source, ip_hdr(skb)->daddr,
672 ntohs(th->source), inet_iif(skb),
673 tcp_v4_sdif(skb));
674 /* don't send rst if it can't find key */
675 if (!sk1)
676 goto out;
677
678 key = tcp_md5_do_lookup(sk1, (union tcp_md5_addr *)
679 &ip_hdr(skb)->saddr, AF_INET);
680 if (!key)
681 goto out;
682
683
684 genhash = tcp_v4_md5_hash_skb(newhash, key, NULL, skb);
685 if (genhash || memcmp(hash_location, newhash, 16) != 0)
686 goto out;
687
688 }
689
690 if (key) {
691 rep.opt[0] = htonl((TCPOPT_NOP << 24) |
692 (TCPOPT_NOP << 16) |
693 (TCPOPT_MD5SIG << 8) |
694 TCPOLEN_MD5SIG);
695 /* Update length and the length the header thinks exists */
696 arg.iov[0].iov_len += TCPOLEN_MD5SIG_ALIGNED;
697 rep.th.doff = arg.iov[0].iov_len / 4;
698
699 tcp_v4_md5_hash_hdr((__u8 *) &rep.opt[1],
700 key, ip_hdr(skb)->saddr,
701 ip_hdr(skb)->daddr, &rep.th);
702 }
703#endif
704 arg.csum = csum_tcpudp_nofold(ip_hdr(skb)->daddr,
705 ip_hdr(skb)->saddr, /* XXX */
706 arg.iov[0].iov_len, IPPROTO_TCP, 0);
707 arg.csumoffset = offsetof(struct tcphdr, check) / 2;
708 arg.flags = (sk && inet_sk_transparent(sk)) ? IP_REPLY_ARG_NOSRCCHECK : 0;
709
710 /* When socket is gone, all binding information is lost.
711 * routing might fail in this case. No choice here, if we choose to force
712 * input interface, we will misroute in case of asymmetric route.
713 */
714 if (sk) {
715 arg.bound_dev_if = sk->sk_bound_dev_if;
716 if (sk_fullsock(sk))
717 trace_tcp_send_reset(sk, skb);
718 }
719
720 BUILD_BUG_ON(offsetof(struct sock, sk_bound_dev_if) !=
721 offsetof(struct inet_timewait_sock, tw_bound_dev_if));
722
723 arg.tos = ip_hdr(skb)->tos;
724 arg.uid = sock_net_uid(net, sk && sk_fullsock(sk) ? sk : NULL);
725 local_bh_disable();
726 ip_send_unicast_reply(*this_cpu_ptr(net->ipv4.tcp_sk),
727 skb, &TCP_SKB_CB(skb)->header.h4.opt,
728 ip_hdr(skb)->saddr, ip_hdr(skb)->daddr,
729 &arg, arg.iov[0].iov_len);
730
731 __TCP_INC_STATS(net, TCP_MIB_OUTSEGS);
732 __TCP_INC_STATS(net, TCP_MIB_OUTRSTS);
733 local_bh_enable();
734
735#ifdef CONFIG_TCP_MD5SIG
736out:
737 rcu_read_unlock();
738#endif
739}
740
741/* The code following below sending ACKs in SYN-RECV and TIME-WAIT states
742 outside socket context is ugly, certainly. What can I do?
743 */
744
745static void tcp_v4_send_ack(const struct sock *sk,
746 struct sk_buff *skb, u32 seq, u32 ack,
747 u32 win, u32 tsval, u32 tsecr, int oif,
748 struct tcp_md5sig_key *key,
749 int reply_flags, u8 tos)
750{
751 const struct tcphdr *th = tcp_hdr(skb);
752 struct {
753 struct tcphdr th;
754 __be32 opt[(TCPOLEN_TSTAMP_ALIGNED >> 2)
755#ifdef CONFIG_TCP_MD5SIG
756 + (TCPOLEN_MD5SIG_ALIGNED >> 2)
757#endif
758 ];
759 } rep;
760 struct net *net = sock_net(sk);
761 struct ip_reply_arg arg;
762
763 memset(&rep.th, 0, sizeof(struct tcphdr));
764 memset(&arg, 0, sizeof(arg));
765
766 arg.iov[0].iov_base = (unsigned char *)&rep;
767 arg.iov[0].iov_len = sizeof(rep.th);
768 if (tsecr) {
769 rep.opt[0] = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
770 (TCPOPT_TIMESTAMP << 8) |
771 TCPOLEN_TIMESTAMP);
772 rep.opt[1] = htonl(tsval);
773 rep.opt[2] = htonl(tsecr);
774 arg.iov[0].iov_len += TCPOLEN_TSTAMP_ALIGNED;
775 }
776
777 /* Swap the send and the receive. */
778 rep.th.dest = th->source;
779 rep.th.source = th->dest;
780 rep.th.doff = arg.iov[0].iov_len / 4;
781 rep.th.seq = htonl(seq);
782 rep.th.ack_seq = htonl(ack);
783 rep.th.ack = 1;
784 rep.th.window = htons(win);
785
786#ifdef CONFIG_TCP_MD5SIG
787 if (key) {
788 int offset = (tsecr) ? 3 : 0;
789
790 rep.opt[offset++] = htonl((TCPOPT_NOP << 24) |
791 (TCPOPT_NOP << 16) |
792 (TCPOPT_MD5SIG << 8) |
793 TCPOLEN_MD5SIG);
794 arg.iov[0].iov_len += TCPOLEN_MD5SIG_ALIGNED;
795 rep.th.doff = arg.iov[0].iov_len/4;
796
797 tcp_v4_md5_hash_hdr((__u8 *) &rep.opt[offset],
798 key, ip_hdr(skb)->saddr,
799 ip_hdr(skb)->daddr, &rep.th);
800 }
801#endif
802 arg.flags = reply_flags;
803 arg.csum = csum_tcpudp_nofold(ip_hdr(skb)->daddr,
804 ip_hdr(skb)->saddr, /* XXX */
805 arg.iov[0].iov_len, IPPROTO_TCP, 0);
806 arg.csumoffset = offsetof(struct tcphdr, check) / 2;
807 if (oif)
808 arg.bound_dev_if = oif;
809 arg.tos = tos;
810 arg.uid = sock_net_uid(net, sk_fullsock(sk) ? sk : NULL);
811 local_bh_disable();
812 ip_send_unicast_reply(*this_cpu_ptr(net->ipv4.tcp_sk),
813 skb, &TCP_SKB_CB(skb)->header.h4.opt,
814 ip_hdr(skb)->saddr, ip_hdr(skb)->daddr,
815 &arg, arg.iov[0].iov_len);
816
817 __TCP_INC_STATS(net, TCP_MIB_OUTSEGS);
818 local_bh_enable();
819}
820
821static void tcp_v4_timewait_ack(struct sock *sk, struct sk_buff *skb)
822{
823 struct inet_timewait_sock *tw = inet_twsk(sk);
824 struct tcp_timewait_sock *tcptw = tcp_twsk(sk);
825
826 tcp_v4_send_ack(sk, skb,
827 tcptw->tw_snd_nxt, tcptw->tw_rcv_nxt,
828 tcptw->tw_rcv_wnd >> tw->tw_rcv_wscale,
829 tcp_time_stamp_raw() + tcptw->tw_ts_offset,
830 tcptw->tw_ts_recent,
831 tw->tw_bound_dev_if,
832 tcp_twsk_md5_key(tcptw),
833 tw->tw_transparent ? IP_REPLY_ARG_NOSRCCHECK : 0,
834 tw->tw_tos
835 );
836
837 inet_twsk_put(tw);
838}
839
840static void tcp_v4_reqsk_send_ack(const struct sock *sk, struct sk_buff *skb,
841 struct request_sock *req)
842{
843 /* sk->sk_state == TCP_LISTEN -> for regular TCP_SYN_RECV
844 * sk->sk_state == TCP_SYN_RECV -> for Fast Open.
845 */
846 u32 seq = (sk->sk_state == TCP_LISTEN) ? tcp_rsk(req)->snt_isn + 1 :
847 tcp_sk(sk)->snd_nxt;
848
849 /* RFC 7323 2.3
850 * The window field (SEG.WND) of every outgoing segment, with the
851 * exception of <SYN> segments, MUST be right-shifted by
852 * Rcv.Wind.Shift bits:
853 */
854 tcp_v4_send_ack(sk, skb, seq,
855 tcp_rsk(req)->rcv_nxt,
856 req->rsk_rcv_wnd >> inet_rsk(req)->rcv_wscale,
857 tcp_time_stamp_raw() + tcp_rsk(req)->ts_off,
858 req->ts_recent,
859 0,
860 tcp_md5_do_lookup(sk, (union tcp_md5_addr *)&ip_hdr(skb)->saddr,
861 AF_INET),
862 inet_rsk(req)->no_srccheck ? IP_REPLY_ARG_NOSRCCHECK : 0,
863 ip_hdr(skb)->tos);
864}
865
866/*
867 * Send a SYN-ACK after having received a SYN.
868 * This still operates on a request_sock only, not on a big
869 * socket.
870 */
871static int tcp_v4_send_synack(const struct sock *sk, struct dst_entry *dst,
872 struct flowi *fl,
873 struct request_sock *req,
874 struct tcp_fastopen_cookie *foc,
875 enum tcp_synack_type synack_type)
876{
877 const struct inet_request_sock *ireq = inet_rsk(req);
878 struct flowi4 fl4;
879 int err = -1;
880 struct sk_buff *skb;
881
882 /* First, grab a route. */
883 if (!dst && (dst = inet_csk_route_req(sk, &fl4, req)) == NULL)
884 return -1;
885
886 skb = tcp_make_synack(sk, dst, req, foc, synack_type);
887
888 if (skb) {
889 __tcp_v4_send_check(skb, ireq->ir_loc_addr, ireq->ir_rmt_addr);
890
891 err = ip_build_and_send_pkt(skb, sk, ireq->ir_loc_addr,
892 ireq->ir_rmt_addr,
893 ireq_opt_deref(ireq));
894 err = net_xmit_eval(err);
895 }
896
897 return err;
898}
899
900/*
901 * IPv4 request_sock destructor.
902 */
903static void tcp_v4_reqsk_destructor(struct request_sock *req)
904{
905 kfree(rcu_dereference_protected(inet_rsk(req)->ireq_opt, 1));
906}
907
908#ifdef CONFIG_TCP_MD5SIG
909/*
910 * RFC2385 MD5 checksumming requires a mapping of
911 * IP address->MD5 Key.
912 * We need to maintain these in the sk structure.
913 */
914
915/* Find the Key structure for an address. */
916struct tcp_md5sig_key *tcp_md5_do_lookup(const struct sock *sk,
917 const union tcp_md5_addr *addr,
918 int family)
919{
920 const struct tcp_sock *tp = tcp_sk(sk);
921 struct tcp_md5sig_key *key;
922 const struct tcp_md5sig_info *md5sig;
923 __be32 mask;
924 struct tcp_md5sig_key *best_match = NULL;
925 bool match;
926
927 /* caller either holds rcu_read_lock() or socket lock */
928 md5sig = rcu_dereference_check(tp->md5sig_info,
929 lockdep_sock_is_held(sk));
930 if (!md5sig)
931 return NULL;
932
933 hlist_for_each_entry_rcu(key, &md5sig->head, node) {
934 if (key->family != family)
935 continue;
936
937 if (family == AF_INET) {
938 mask = inet_make_mask(key->prefixlen);
939 match = (key->addr.a4.s_addr & mask) ==
940 (addr->a4.s_addr & mask);
941#if IS_ENABLED(CONFIG_IPV6)
942 } else if (family == AF_INET6) {
943 match = ipv6_prefix_equal(&key->addr.a6, &addr->a6,
944 key->prefixlen);
945#endif
946 } else {
947 match = false;
948 }
949
950 if (match && (!best_match ||
951 key->prefixlen > best_match->prefixlen))
952 best_match = key;
953 }
954 return best_match;
955}
956EXPORT_SYMBOL(tcp_md5_do_lookup);
957
958static struct tcp_md5sig_key *tcp_md5_do_lookup_exact(const struct sock *sk,
959 const union tcp_md5_addr *addr,
960 int family, u8 prefixlen)
961{
962 const struct tcp_sock *tp = tcp_sk(sk);
963 struct tcp_md5sig_key *key;
964 unsigned int size = sizeof(struct in_addr);
965 const struct tcp_md5sig_info *md5sig;
966
967 /* caller either holds rcu_read_lock() or socket lock */
968 md5sig = rcu_dereference_check(tp->md5sig_info,
969 lockdep_sock_is_held(sk));
970 if (!md5sig)
971 return NULL;
972#if IS_ENABLED(CONFIG_IPV6)
973 if (family == AF_INET6)
974 size = sizeof(struct in6_addr);
975#endif
976 hlist_for_each_entry_rcu(key, &md5sig->head, node) {
977 if (key->family != family)
978 continue;
979 if (!memcmp(&key->addr, addr, size) &&
980 key->prefixlen == prefixlen)
981 return key;
982 }
983 return NULL;
984}
985
986struct tcp_md5sig_key *tcp_v4_md5_lookup(const struct sock *sk,
987 const struct sock *addr_sk)
988{
989 const union tcp_md5_addr *addr;
990
991 addr = (const union tcp_md5_addr *)&addr_sk->sk_daddr;
992 return tcp_md5_do_lookup(sk, addr, AF_INET);
993}
994EXPORT_SYMBOL(tcp_v4_md5_lookup);
995
996/* This can be called on a newly created socket, from other files */
997int tcp_md5_do_add(struct sock *sk, const union tcp_md5_addr *addr,
998 int family, u8 prefixlen, const u8 *newkey, u8 newkeylen,
999 gfp_t gfp)
1000{
1001 /* Add Key to the list */
1002 struct tcp_md5sig_key *key;
1003 struct tcp_sock *tp = tcp_sk(sk);
1004 struct tcp_md5sig_info *md5sig;
1005
1006 key = tcp_md5_do_lookup_exact(sk, addr, family, prefixlen);
1007 if (key) {
1008 /* Pre-existing entry - just update that one. */
1009 memcpy(key->key, newkey, newkeylen);
1010 key->keylen = newkeylen;
1011 return 0;
1012 }
1013
1014 md5sig = rcu_dereference_protected(tp->md5sig_info,
1015 lockdep_sock_is_held(sk));
1016 if (!md5sig) {
1017 md5sig = kmalloc(sizeof(*md5sig), gfp);
1018 if (!md5sig)
1019 return -ENOMEM;
1020
1021 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1022 INIT_HLIST_HEAD(&md5sig->head);
1023 rcu_assign_pointer(tp->md5sig_info, md5sig);
1024 }
1025
1026 key = sock_kmalloc(sk, sizeof(*key), gfp);
1027 if (!key)
1028 return -ENOMEM;
1029 if (!tcp_alloc_md5sig_pool()) {
1030 sock_kfree_s(sk, key, sizeof(*key));
1031 return -ENOMEM;
1032 }
1033
1034 memcpy(key->key, newkey, newkeylen);
1035 key->keylen = newkeylen;
1036 key->family = family;
1037 key->prefixlen = prefixlen;
1038 memcpy(&key->addr, addr,
1039 (family == AF_INET6) ? sizeof(struct in6_addr) :
1040 sizeof(struct in_addr));
1041 hlist_add_head_rcu(&key->node, &md5sig->head);
1042 return 0;
1043}
1044EXPORT_SYMBOL(tcp_md5_do_add);
1045
1046int tcp_md5_do_del(struct sock *sk, const union tcp_md5_addr *addr, int family,
1047 u8 prefixlen)
1048{
1049 struct tcp_md5sig_key *key;
1050
1051 key = tcp_md5_do_lookup_exact(sk, addr, family, prefixlen);
1052 if (!key)
1053 return -ENOENT;
1054 hlist_del_rcu(&key->node);
1055 atomic_sub(sizeof(*key), &sk->sk_omem_alloc);
1056 kfree_rcu(key, rcu);
1057 return 0;
1058}
1059EXPORT_SYMBOL(tcp_md5_do_del);
1060
1061static void tcp_clear_md5_list(struct sock *sk)
1062{
1063 struct tcp_sock *tp = tcp_sk(sk);
1064 struct tcp_md5sig_key *key;
1065 struct hlist_node *n;
1066 struct tcp_md5sig_info *md5sig;
1067
1068 md5sig = rcu_dereference_protected(tp->md5sig_info, 1);
1069
1070 hlist_for_each_entry_safe(key, n, &md5sig->head, node) {
1071 hlist_del_rcu(&key->node);
1072 atomic_sub(sizeof(*key), &sk->sk_omem_alloc);
1073 kfree_rcu(key, rcu);
1074 }
1075}
1076
1077static int tcp_v4_parse_md5_keys(struct sock *sk, int optname,
1078 char __user *optval, int optlen)
1079{
1080 struct tcp_md5sig cmd;
1081 struct sockaddr_in *sin = (struct sockaddr_in *)&cmd.tcpm_addr;
1082 u8 prefixlen = 32;
1083
1084 if (optlen < sizeof(cmd))
1085 return -EINVAL;
1086
1087 if (copy_from_user(&cmd, optval, sizeof(cmd)))
1088 return -EFAULT;
1089
1090 if (sin->sin_family != AF_INET)
1091 return -EINVAL;
1092
1093 if (optname == TCP_MD5SIG_EXT &&
1094 cmd.tcpm_flags & TCP_MD5SIG_FLAG_PREFIX) {
1095 prefixlen = cmd.tcpm_prefixlen;
1096 if (prefixlen > 32)
1097 return -EINVAL;
1098 }
1099
1100 if (!cmd.tcpm_keylen)
1101 return tcp_md5_do_del(sk, (union tcp_md5_addr *)&sin->sin_addr.s_addr,
1102 AF_INET, prefixlen);
1103
1104 if (cmd.tcpm_keylen > TCP_MD5SIG_MAXKEYLEN)
1105 return -EINVAL;
1106
1107 return tcp_md5_do_add(sk, (union tcp_md5_addr *)&sin->sin_addr.s_addr,
1108 AF_INET, prefixlen, cmd.tcpm_key, cmd.tcpm_keylen,
1109 GFP_KERNEL);
1110}
1111
1112static int tcp_v4_md5_hash_headers(struct tcp_md5sig_pool *hp,
1113 __be32 daddr, __be32 saddr,
1114 const struct tcphdr *th, int nbytes)
1115{
1116 struct tcp4_pseudohdr *bp;
1117 struct scatterlist sg;
1118 struct tcphdr *_th;
1119
1120 bp = hp->scratch;
1121 bp->saddr = saddr;
1122 bp->daddr = daddr;
1123 bp->pad = 0;
1124 bp->protocol = IPPROTO_TCP;
1125 bp->len = cpu_to_be16(nbytes);
1126
1127 _th = (struct tcphdr *)(bp + 1);
1128 memcpy(_th, th, sizeof(*th));
1129 _th->check = 0;
1130
1131 sg_init_one(&sg, bp, sizeof(*bp) + sizeof(*th));
1132 ahash_request_set_crypt(hp->md5_req, &sg, NULL,
1133 sizeof(*bp) + sizeof(*th));
1134 return crypto_ahash_update(hp->md5_req);
1135}
1136
1137static int tcp_v4_md5_hash_hdr(char *md5_hash, const struct tcp_md5sig_key *key,
1138 __be32 daddr, __be32 saddr, const struct tcphdr *th)
1139{
1140 struct tcp_md5sig_pool *hp;
1141 struct ahash_request *req;
1142
1143 hp = tcp_get_md5sig_pool();
1144 if (!hp)
1145 goto clear_hash_noput;
1146 req = hp->md5_req;
1147
1148 if (crypto_ahash_init(req))
1149 goto clear_hash;
1150 if (tcp_v4_md5_hash_headers(hp, daddr, saddr, th, th->doff << 2))
1151 goto clear_hash;
1152 if (tcp_md5_hash_key(hp, key))
1153 goto clear_hash;
1154 ahash_request_set_crypt(req, NULL, md5_hash, 0);
1155 if (crypto_ahash_final(req))
1156 goto clear_hash;
1157
1158 tcp_put_md5sig_pool();
1159 return 0;
1160
1161clear_hash:
1162 tcp_put_md5sig_pool();
1163clear_hash_noput:
1164 memset(md5_hash, 0, 16);
1165 return 1;
1166}
1167
1168int tcp_v4_md5_hash_skb(char *md5_hash, const struct tcp_md5sig_key *key,
1169 const struct sock *sk,
1170 const struct sk_buff *skb)
1171{
1172 struct tcp_md5sig_pool *hp;
1173 struct ahash_request *req;
1174 const struct tcphdr *th = tcp_hdr(skb);
1175 __be32 saddr, daddr;
1176
1177 if (sk) { /* valid for establish/request sockets */
1178 saddr = sk->sk_rcv_saddr;
1179 daddr = sk->sk_daddr;
1180 } else {
1181 const struct iphdr *iph = ip_hdr(skb);
1182 saddr = iph->saddr;
1183 daddr = iph->daddr;
1184 }
1185
1186 hp = tcp_get_md5sig_pool();
1187 if (!hp)
1188 goto clear_hash_noput;
1189 req = hp->md5_req;
1190
1191 if (crypto_ahash_init(req))
1192 goto clear_hash;
1193
1194 if (tcp_v4_md5_hash_headers(hp, daddr, saddr, th, skb->len))
1195 goto clear_hash;
1196 if (tcp_md5_hash_skb_data(hp, skb, th->doff << 2))
1197 goto clear_hash;
1198 if (tcp_md5_hash_key(hp, key))
1199 goto clear_hash;
1200 ahash_request_set_crypt(req, NULL, md5_hash, 0);
1201 if (crypto_ahash_final(req))
1202 goto clear_hash;
1203
1204 tcp_put_md5sig_pool();
1205 return 0;
1206
1207clear_hash:
1208 tcp_put_md5sig_pool();
1209clear_hash_noput:
1210 memset(md5_hash, 0, 16);
1211 return 1;
1212}
1213EXPORT_SYMBOL(tcp_v4_md5_hash_skb);
1214
1215#endif
1216
1217/* Called with rcu_read_lock() */
1218static bool tcp_v4_inbound_md5_hash(const struct sock *sk,
1219 const struct sk_buff *skb)
1220{
1221#ifdef CONFIG_TCP_MD5SIG
1222 /*
1223 * This gets called for each TCP segment that arrives
1224 * so we want to be efficient.
1225 * We have 3 drop cases:
1226 * o No MD5 hash and one expected.
1227 * o MD5 hash and we're not expecting one.
1228 * o MD5 hash and its wrong.
1229 */
1230 const __u8 *hash_location = NULL;
1231 struct tcp_md5sig_key *hash_expected;
1232 const struct iphdr *iph = ip_hdr(skb);
1233 const struct tcphdr *th = tcp_hdr(skb);
1234 int genhash;
1235 unsigned char newhash[16];
1236
1237 hash_expected = tcp_md5_do_lookup(sk, (union tcp_md5_addr *)&iph->saddr,
1238 AF_INET);
1239 hash_location = tcp_parse_md5sig_option(th);
1240
1241 /* We've parsed the options - do we have a hash? */
1242 if (!hash_expected && !hash_location)
1243 return false;
1244
1245 if (hash_expected && !hash_location) {
1246 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
1247 return true;
1248 }
1249
1250 if (!hash_expected && hash_location) {
1251 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
1252 return true;
1253 }
1254
1255 /* Okay, so this is hash_expected and hash_location -
1256 * so we need to calculate the checksum.
1257 */
1258 genhash = tcp_v4_md5_hash_skb(newhash,
1259 hash_expected,
1260 NULL, skb);
1261
1262 if (genhash || memcmp(hash_location, newhash, 16) != 0) {
1263 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
1264 net_info_ratelimited("MD5 Hash failed for (%pI4, %d)->(%pI4, %d)%s\n",
1265 &iph->saddr, ntohs(th->source),
1266 &iph->daddr, ntohs(th->dest),
1267 genhash ? " tcp_v4_calc_md5_hash failed"
1268 : "");
1269 return true;
1270 }
1271 return false;
1272#endif
1273 return false;
1274}
1275
1276static void tcp_v4_init_req(struct request_sock *req,
1277 const struct sock *sk_listener,
1278 struct sk_buff *skb)
1279{
1280 struct inet_request_sock *ireq = inet_rsk(req);
1281 struct net *net = sock_net(sk_listener);
1282
1283 sk_rcv_saddr_set(req_to_sk(req), ip_hdr(skb)->daddr);
1284 sk_daddr_set(req_to_sk(req), ip_hdr(skb)->saddr);
1285 RCU_INIT_POINTER(ireq->ireq_opt, tcp_v4_save_options(net, skb));
1286}
1287
1288static struct dst_entry *tcp_v4_route_req(const struct sock *sk,
1289 struct flowi *fl,
1290 const struct request_sock *req)
1291{
1292 return inet_csk_route_req(sk, &fl->u.ip4, req);
1293}
1294
1295struct request_sock_ops tcp_request_sock_ops __read_mostly = {
1296 .family = PF_INET,
1297 .obj_size = sizeof(struct tcp_request_sock),
1298 .rtx_syn_ack = tcp_rtx_synack,
1299 .send_ack = tcp_v4_reqsk_send_ack,
1300 .destructor = tcp_v4_reqsk_destructor,
1301 .send_reset = tcp_v4_send_reset,
1302 .syn_ack_timeout = tcp_syn_ack_timeout,
1303};
1304
1305static const struct tcp_request_sock_ops tcp_request_sock_ipv4_ops = {
1306 .mss_clamp = TCP_MSS_DEFAULT,
1307#ifdef CONFIG_TCP_MD5SIG
1308 .req_md5_lookup = tcp_v4_md5_lookup,
1309 .calc_md5_hash = tcp_v4_md5_hash_skb,
1310#endif
1311 .init_req = tcp_v4_init_req,
1312#ifdef CONFIG_SYN_COOKIES
1313 .cookie_init_seq = cookie_v4_init_sequence,
1314#endif
1315 .route_req = tcp_v4_route_req,
1316 .init_seq = tcp_v4_init_seq,
1317 .init_ts_off = tcp_v4_init_ts_off,
1318 .send_synack = tcp_v4_send_synack,
1319};
1320
1321int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb)
1322{
1323 /* Never answer to SYNs send to broadcast or multicast */
1324 if (skb_rtable(skb)->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST))
1325 goto drop;
1326
1327 return tcp_conn_request(&tcp_request_sock_ops,
1328 &tcp_request_sock_ipv4_ops, sk, skb);
1329
1330drop:
1331 tcp_listendrop(sk);
1332 return 0;
1333}
1334EXPORT_SYMBOL(tcp_v4_conn_request);
1335
1336
1337/*
1338 * The three way handshake has completed - we got a valid synack -
1339 * now create the new socket.
1340 */
1341struct sock *tcp_v4_syn_recv_sock(const struct sock *sk, struct sk_buff *skb,
1342 struct request_sock *req,
1343 struct dst_entry *dst,
1344 struct request_sock *req_unhash,
1345 bool *own_req)
1346{
1347 struct inet_request_sock *ireq;
1348 struct inet_sock *newinet;
1349 struct tcp_sock *newtp;
1350 struct sock *newsk;
1351#ifdef CONFIG_TCP_MD5SIG
1352 struct tcp_md5sig_key *key;
1353#endif
1354 struct ip_options_rcu *inet_opt;
1355
1356 if (sk_acceptq_is_full(sk))
1357 goto exit_overflow;
1358
1359 newsk = tcp_create_openreq_child(sk, req, skb);
1360 if (!newsk)
1361 goto exit_nonewsk;
1362
1363 newsk->sk_gso_type = SKB_GSO_TCPV4;
1364 inet_sk_rx_dst_set(newsk, skb);
1365
1366 newtp = tcp_sk(newsk);
1367 newinet = inet_sk(newsk);
1368 ireq = inet_rsk(req);
1369 sk_daddr_set(newsk, ireq->ir_rmt_addr);
1370 sk_rcv_saddr_set(newsk, ireq->ir_loc_addr);
1371 newsk->sk_bound_dev_if = ireq->ir_iif;
1372 newinet->inet_saddr = ireq->ir_loc_addr;
1373 inet_opt = rcu_dereference(ireq->ireq_opt);
1374 RCU_INIT_POINTER(newinet->inet_opt, inet_opt);
1375 newinet->mc_index = inet_iif(skb);
1376 newinet->mc_ttl = ip_hdr(skb)->ttl;
1377 newinet->rcv_tos = ip_hdr(skb)->tos;
1378 inet_csk(newsk)->icsk_ext_hdr_len = 0;
1379 if (inet_opt)
1380 inet_csk(newsk)->icsk_ext_hdr_len = inet_opt->opt.optlen;
1381 newinet->inet_id = newtp->write_seq ^ jiffies;
1382
1383 if (!dst) {
1384 dst = inet_csk_route_child_sock(sk, newsk, req);
1385 if (!dst)
1386 goto put_and_exit;
1387 } else {
1388 /* syncookie case : see end of cookie_v4_check() */
1389 }
1390 sk_setup_caps(newsk, dst);
1391
1392 tcp_ca_openreq_child(newsk, dst);
1393
1394 tcp_sync_mss(newsk, dst_mtu(dst));
1395 newtp->advmss = tcp_mss_clamp(tcp_sk(sk), dst_metric_advmss(dst));
1396
1397 tcp_initialize_rcv_mss(newsk);
1398
1399#ifdef CONFIG_TCP_MD5SIG
1400 /* Copy over the MD5 key from the original socket */
1401 key = tcp_md5_do_lookup(sk, (union tcp_md5_addr *)&newinet->inet_daddr,
1402 AF_INET);
1403 if (key) {
1404 /*
1405 * We're using one, so create a matching key
1406 * on the newsk structure. If we fail to get
1407 * memory, then we end up not copying the key
1408 * across. Shucks.
1409 */
1410 tcp_md5_do_add(newsk, (union tcp_md5_addr *)&newinet->inet_daddr,
1411 AF_INET, 32, key->key, key->keylen, GFP_ATOMIC);
1412 sk_nocaps_add(newsk, NETIF_F_GSO_MASK);
1413 }
1414#endif
1415
1416 if (__inet_inherit_port(sk, newsk) < 0)
1417 goto put_and_exit;
1418 *own_req = inet_ehash_nolisten(newsk, req_to_sk(req_unhash));
1419 if (likely(*own_req)) {
1420 tcp_move_syn(newtp, req);
1421 ireq->ireq_opt = NULL;
1422 } else {
1423 newinet->inet_opt = NULL;
1424 }
1425 return newsk;
1426
1427exit_overflow:
1428 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
1429exit_nonewsk:
1430 dst_release(dst);
1431exit:
1432 tcp_listendrop(sk);
1433 return NULL;
1434put_and_exit:
1435 newinet->inet_opt = NULL;
1436 inet_csk_prepare_forced_close(newsk);
1437 tcp_done(newsk);
1438 goto exit;
1439}
1440EXPORT_SYMBOL(tcp_v4_syn_recv_sock);
1441
1442static struct sock *tcp_v4_cookie_check(struct sock *sk, struct sk_buff *skb)
1443{
1444#ifdef CONFIG_SYN_COOKIES
1445 const struct tcphdr *th = tcp_hdr(skb);
1446
1447 if (!th->syn)
1448 sk = cookie_v4_check(sk, skb);
1449#endif
1450 return sk;
1451}
1452
1453/* The socket must have it's spinlock held when we get
1454 * here, unless it is a TCP_LISTEN socket.
1455 *
1456 * We have a potential double-lock case here, so even when
1457 * doing backlog processing we use the BH locking scheme.
1458 * This is because we cannot sleep with the original spinlock
1459 * held.
1460 */
1461int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb)
1462{
1463 struct sock *rsk;
1464
1465 if (sk->sk_state == TCP_ESTABLISHED) { /* Fast path */
1466 struct dst_entry *dst = sk->sk_rx_dst;
1467
1468 sock_rps_save_rxhash(sk, skb);
1469 sk_mark_napi_id(sk, skb);
1470 if (dst) {
1471 if (inet_sk(sk)->rx_dst_ifindex != skb->skb_iif ||
1472 !dst->ops->check(dst, 0)) {
1473 dst_release(dst);
1474 sk->sk_rx_dst = NULL;
1475 }
1476 }
1477 tcp_rcv_established(sk, skb, tcp_hdr(skb));
1478 return 0;
1479 }
1480
1481 if (tcp_checksum_complete(skb))
1482 goto csum_err;
1483
1484 if (sk->sk_state == TCP_LISTEN) {
1485 struct sock *nsk = tcp_v4_cookie_check(sk, skb);
1486
1487 if (!nsk)
1488 goto discard;
1489 if (nsk != sk) {
1490 if (tcp_child_process(sk, nsk, skb)) {
1491 rsk = nsk;
1492 goto reset;
1493 }
1494 return 0;
1495 }
1496 } else
1497 sock_rps_save_rxhash(sk, skb);
1498
1499 if (tcp_rcv_state_process(sk, skb)) {
1500 rsk = sk;
1501 goto reset;
1502 }
1503 return 0;
1504
1505reset:
1506 tcp_v4_send_reset(rsk, skb);
1507discard:
1508 kfree_skb(skb);
1509 /* Be careful here. If this function gets more complicated and
1510 * gcc suffers from register pressure on the x86, sk (in %ebx)
1511 * might be destroyed here. This current version compiles correctly,
1512 * but you have been warned.
1513 */
1514 return 0;
1515
1516csum_err:
1517 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS);
1518 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS);
1519 goto discard;
1520}
1521EXPORT_SYMBOL(tcp_v4_do_rcv);
1522
1523int tcp_v4_early_demux(struct sk_buff *skb)
1524{
1525 const struct iphdr *iph;
1526 const struct tcphdr *th;
1527 struct sock *sk;
1528
1529 if (skb->pkt_type != PACKET_HOST)
1530 return 0;
1531
1532 if (!pskb_may_pull(skb, skb_transport_offset(skb) + sizeof(struct tcphdr)))
1533 return 0;
1534
1535 iph = ip_hdr(skb);
1536 th = tcp_hdr(skb);
1537
1538 if (th->doff < sizeof(struct tcphdr) / 4)
1539 return 0;
1540
1541 sk = __inet_lookup_established(dev_net(skb->dev), &tcp_hashinfo,
1542 iph->saddr, th->source,
1543 iph->daddr, ntohs(th->dest),
1544 skb->skb_iif, inet_sdif(skb));
1545 if (sk) {
1546 skb->sk = sk;
1547 skb->destructor = sock_edemux;
1548 if (sk_fullsock(sk)) {
1549 struct dst_entry *dst = READ_ONCE(sk->sk_rx_dst);
1550
1551 if (dst)
1552 dst = dst_check(dst, 0);
1553 if (dst &&
1554 inet_sk(sk)->rx_dst_ifindex == skb->skb_iif)
1555 skb_dst_set_noref(skb, dst);
1556 }
1557 }
1558 return 0;
1559}
1560
1561bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb)
1562{
1563 u32 limit = sk->sk_rcvbuf + sk->sk_sndbuf;
1564
1565 /* Only socket owner can try to collapse/prune rx queues
1566 * to reduce memory overhead, so add a little headroom here.
1567 * Few sockets backlog are possibly concurrently non empty.
1568 */
1569 limit += 64*1024;
1570
1571 /* In case all data was pulled from skb frags (in __pskb_pull_tail()),
1572 * we can fix skb->truesize to its real value to avoid future drops.
1573 * This is valid because skb is not yet charged to the socket.
1574 * It has been noticed pure SACK packets were sometimes dropped
1575 * (if cooked by drivers without copybreak feature).
1576 */
1577 skb_condense(skb);
1578
1579 if (unlikely(sk_add_backlog(sk, skb, limit))) {
1580 bh_unlock_sock(sk);
1581 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPBACKLOGDROP);
1582 return true;
1583 }
1584 return false;
1585}
1586EXPORT_SYMBOL(tcp_add_backlog);
1587
1588int tcp_filter(struct sock *sk, struct sk_buff *skb)
1589{
1590 struct tcphdr *th = (struct tcphdr *)skb->data;
1591 unsigned int eaten = skb->len;
1592 int err;
1593
1594 err = sk_filter_trim_cap(sk, skb, th->doff * 4);
1595 if (!err) {
1596 eaten -= skb->len;
1597 TCP_SKB_CB(skb)->end_seq -= eaten;
1598 }
1599 return err;
1600}
1601EXPORT_SYMBOL(tcp_filter);
1602
1603static void tcp_v4_restore_cb(struct sk_buff *skb)
1604{
1605 memmove(IPCB(skb), &TCP_SKB_CB(skb)->header.h4,
1606 sizeof(struct inet_skb_parm));
1607}
1608
1609static void tcp_v4_fill_cb(struct sk_buff *skb, const struct iphdr *iph,
1610 const struct tcphdr *th)
1611{
1612 /* This is tricky : We move IPCB at its correct location into TCP_SKB_CB()
1613 * barrier() makes sure compiler wont play fool^Waliasing games.
1614 */
1615 memmove(&TCP_SKB_CB(skb)->header.h4, IPCB(skb),
1616 sizeof(struct inet_skb_parm));
1617 barrier();
1618
1619 TCP_SKB_CB(skb)->seq = ntohl(th->seq);
1620 TCP_SKB_CB(skb)->end_seq = (TCP_SKB_CB(skb)->seq + th->syn + th->fin +
1621 skb->len - th->doff * 4);
1622 TCP_SKB_CB(skb)->ack_seq = ntohl(th->ack_seq);
1623 TCP_SKB_CB(skb)->tcp_flags = tcp_flag_byte(th);
1624 TCP_SKB_CB(skb)->tcp_tw_isn = 0;
1625 TCP_SKB_CB(skb)->ip_dsfield = ipv4_get_dsfield(iph);
1626 TCP_SKB_CB(skb)->sacked = 0;
1627 TCP_SKB_CB(skb)->has_rxtstamp =
1628 skb->tstamp || skb_hwtstamps(skb)->hwtstamp;
1629}
1630
1631/*
1632 * From tcp_input.c
1633 */
1634
1635int tcp_v4_rcv(struct sk_buff *skb)
1636{
1637 struct net *net = dev_net(skb->dev);
1638 int sdif = inet_sdif(skb);
1639 const struct iphdr *iph;
1640 const struct tcphdr *th;
1641 bool refcounted;
1642 struct sock *sk;
1643 int ret;
1644
1645 if (skb->pkt_type != PACKET_HOST)
1646 goto discard_it;
1647
1648 /* Count it even if it's bad */
1649 __TCP_INC_STATS(net, TCP_MIB_INSEGS);
1650
1651 if (!pskb_may_pull(skb, sizeof(struct tcphdr)))
1652 goto discard_it;
1653
1654 th = (const struct tcphdr *)skb->data;
1655
1656 if (unlikely(th->doff < sizeof(struct tcphdr) / 4))
1657 goto bad_packet;
1658 if (!pskb_may_pull(skb, th->doff * 4))
1659 goto discard_it;
1660
1661 /* An explanation is required here, I think.
1662 * Packet length and doff are validated by header prediction,
1663 * provided case of th->doff==0 is eliminated.
1664 * So, we defer the checks. */
1665
1666 if (skb_checksum_init(skb, IPPROTO_TCP, inet_compute_pseudo))
1667 goto csum_error;
1668
1669 th = (const struct tcphdr *)skb->data;
1670 iph = ip_hdr(skb);
1671lookup:
1672 sk = __inet_lookup_skb(&tcp_hashinfo, skb, __tcp_hdrlen(th), th->source,
1673 th->dest, sdif, &refcounted);
1674 if (!sk)
1675 goto no_tcp_socket;
1676
1677process:
1678 if (sk->sk_state == TCP_TIME_WAIT)
1679 goto do_time_wait;
1680
1681 if (sk->sk_state == TCP_NEW_SYN_RECV) {
1682 struct request_sock *req = inet_reqsk(sk);
1683 bool req_stolen = false;
1684 struct sock *nsk;
1685
1686 sk = req->rsk_listener;
1687 if (unlikely(tcp_v4_inbound_md5_hash(sk, skb))) {
1688 sk_drops_add(sk, skb);
1689 reqsk_put(req);
1690 goto discard_it;
1691 }
1692 if (unlikely(sk->sk_state != TCP_LISTEN)) {
1693 inet_csk_reqsk_queue_drop_and_put(sk, req);
1694 goto lookup;
1695 }
1696 /* We own a reference on the listener, increase it again
1697 * as we might lose it too soon.
1698 */
1699 sock_hold(sk);
1700 refcounted = true;
1701 nsk = NULL;
1702 if (!tcp_filter(sk, skb)) {
1703 th = (const struct tcphdr *)skb->data;
1704 iph = ip_hdr(skb);
1705 tcp_v4_fill_cb(skb, iph, th);
1706 nsk = tcp_check_req(sk, skb, req, false, &req_stolen);
1707 }
1708 if (!nsk) {
1709 reqsk_put(req);
1710 if (req_stolen) {
1711 /* Another cpu got exclusive access to req
1712 * and created a full blown socket.
1713 * Try to feed this packet to this socket
1714 * instead of discarding it.
1715 */
1716 tcp_v4_restore_cb(skb);
1717 sock_put(sk);
1718 goto lookup;
1719 }
1720 goto discard_and_relse;
1721 }
1722 if (nsk == sk) {
1723 reqsk_put(req);
1724 tcp_v4_restore_cb(skb);
1725 } else if (tcp_child_process(sk, nsk, skb)) {
1726 tcp_v4_send_reset(nsk, skb);
1727 goto discard_and_relse;
1728 } else {
1729 sock_put(sk);
1730 return 0;
1731 }
1732 }
1733 if (unlikely(iph->ttl < inet_sk(sk)->min_ttl)) {
1734 __NET_INC_STATS(net, LINUX_MIB_TCPMINTTLDROP);
1735 goto discard_and_relse;
1736 }
1737
1738 if (!xfrm4_policy_check(sk, XFRM_POLICY_IN, skb))
1739 goto discard_and_relse;
1740
1741 if (tcp_v4_inbound_md5_hash(sk, skb))
1742 goto discard_and_relse;
1743
1744 nf_reset(skb);
1745
1746 if (tcp_filter(sk, skb))
1747 goto discard_and_relse;
1748 th = (const struct tcphdr *)skb->data;
1749 iph = ip_hdr(skb);
1750 tcp_v4_fill_cb(skb, iph, th);
1751
1752 skb->dev = NULL;
1753
1754 if (sk->sk_state == TCP_LISTEN) {
1755 ret = tcp_v4_do_rcv(sk, skb);
1756 goto put_and_return;
1757 }
1758
1759 sk_incoming_cpu_update(sk);
1760
1761 bh_lock_sock_nested(sk);
1762 tcp_segs_in(tcp_sk(sk), skb);
1763 ret = 0;
1764 if (!sock_owned_by_user(sk)) {
1765 ret = tcp_v4_do_rcv(sk, skb);
1766 } else if (tcp_add_backlog(sk, skb)) {
1767 goto discard_and_relse;
1768 }
1769 bh_unlock_sock(sk);
1770
1771put_and_return:
1772 if (refcounted)
1773 sock_put(sk);
1774
1775 return ret;
1776
1777no_tcp_socket:
1778 if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb))
1779 goto discard_it;
1780
1781 tcp_v4_fill_cb(skb, iph, th);
1782
1783 if (tcp_checksum_complete(skb)) {
1784csum_error:
1785 __TCP_INC_STATS(net, TCP_MIB_CSUMERRORS);
1786bad_packet:
1787 __TCP_INC_STATS(net, TCP_MIB_INERRS);
1788 } else {
1789 tcp_v4_send_reset(NULL, skb);
1790 }
1791
1792discard_it:
1793 /* Discard frame. */
1794 kfree_skb(skb);
1795 return 0;
1796
1797discard_and_relse:
1798 sk_drops_add(sk, skb);
1799 if (refcounted)
1800 sock_put(sk);
1801 goto discard_it;
1802
1803do_time_wait:
1804 if (!xfrm4_policy_check(NULL, XFRM_POLICY_IN, skb)) {
1805 inet_twsk_put(inet_twsk(sk));
1806 goto discard_it;
1807 }
1808
1809 tcp_v4_fill_cb(skb, iph, th);
1810
1811 if (tcp_checksum_complete(skb)) {
1812 inet_twsk_put(inet_twsk(sk));
1813 goto csum_error;
1814 }
1815 switch (tcp_timewait_state_process(inet_twsk(sk), skb, th)) {
1816 case TCP_TW_SYN: {
1817 struct sock *sk2 = inet_lookup_listener(dev_net(skb->dev),
1818 &tcp_hashinfo, skb,
1819 __tcp_hdrlen(th),
1820 iph->saddr, th->source,
1821 iph->daddr, th->dest,
1822 inet_iif(skb),
1823 sdif);
1824 if (sk2) {
1825 inet_twsk_deschedule_put(inet_twsk(sk));
1826 sk = sk2;
1827 tcp_v4_restore_cb(skb);
1828 refcounted = false;
1829 goto process;
1830 }
1831 }
1832 /* to ACK */
1833 /* fall through */
1834 case TCP_TW_ACK:
1835 tcp_v4_timewait_ack(sk, skb);
1836 break;
1837 case TCP_TW_RST:
1838 tcp_v4_send_reset(sk, skb);
1839 inet_twsk_deschedule_put(inet_twsk(sk));
1840 goto discard_it;
1841 case TCP_TW_SUCCESS:;
1842 }
1843 goto discard_it;
1844}
1845
1846static struct timewait_sock_ops tcp_timewait_sock_ops = {
1847 .twsk_obj_size = sizeof(struct tcp_timewait_sock),
1848 .twsk_unique = tcp_twsk_unique,
1849 .twsk_destructor= tcp_twsk_destructor,
1850};
1851
1852void inet_sk_rx_dst_set(struct sock *sk, const struct sk_buff *skb)
1853{
1854 struct dst_entry *dst = skb_dst(skb);
1855
1856 if (dst && dst_hold_safe(dst)) {
1857 sk->sk_rx_dst = dst;
1858 inet_sk(sk)->rx_dst_ifindex = skb->skb_iif;
1859 }
1860}
1861EXPORT_SYMBOL(inet_sk_rx_dst_set);
1862
1863const struct inet_connection_sock_af_ops ipv4_specific = {
1864 .queue_xmit = ip_queue_xmit,
1865 .send_check = tcp_v4_send_check,
1866 .rebuild_header = inet_sk_rebuild_header,
1867 .sk_rx_dst_set = inet_sk_rx_dst_set,
1868 .conn_request = tcp_v4_conn_request,
1869 .syn_recv_sock = tcp_v4_syn_recv_sock,
1870 .net_header_len = sizeof(struct iphdr),
1871 .setsockopt = ip_setsockopt,
1872 .getsockopt = ip_getsockopt,
1873 .addr2sockaddr = inet_csk_addr2sockaddr,
1874 .sockaddr_len = sizeof(struct sockaddr_in),
1875#ifdef CONFIG_COMPAT
1876 .compat_setsockopt = compat_ip_setsockopt,
1877 .compat_getsockopt = compat_ip_getsockopt,
1878#endif
1879 .mtu_reduced = tcp_v4_mtu_reduced,
1880};
1881EXPORT_SYMBOL(ipv4_specific);
1882
1883#ifdef CONFIG_TCP_MD5SIG
1884static const struct tcp_sock_af_ops tcp_sock_ipv4_specific = {
1885 .md5_lookup = tcp_v4_md5_lookup,
1886 .calc_md5_hash = tcp_v4_md5_hash_skb,
1887 .md5_parse = tcp_v4_parse_md5_keys,
1888};
1889#endif
1890
1891/* NOTE: A lot of things set to zero explicitly by call to
1892 * sk_alloc() so need not be done here.
1893 */
1894static int tcp_v4_init_sock(struct sock *sk)
1895{
1896 struct inet_connection_sock *icsk = inet_csk(sk);
1897
1898 tcp_init_sock(sk);
1899
1900 icsk->icsk_af_ops = &ipv4_specific;
1901
1902#ifdef CONFIG_TCP_MD5SIG
1903 tcp_sk(sk)->af_specific = &tcp_sock_ipv4_specific;
1904#endif
1905
1906 return 0;
1907}
1908
1909void tcp_v4_destroy_sock(struct sock *sk)
1910{
1911 struct tcp_sock *tp = tcp_sk(sk);
1912
1913 trace_tcp_destroy_sock(sk);
1914
1915 tcp_clear_xmit_timers(sk);
1916
1917 tcp_cleanup_congestion_control(sk);
1918
1919 tcp_cleanup_ulp(sk);
1920
1921 /* Cleanup up the write buffer. */
1922 tcp_write_queue_purge(sk);
1923
1924 /* Check if we want to disable active TFO */
1925 tcp_fastopen_active_disable_ofo_check(sk);
1926
1927 /* Cleans up our, hopefully empty, out_of_order_queue. */
1928 skb_rbtree_purge(&tp->out_of_order_queue);
1929
1930#ifdef CONFIG_TCP_MD5SIG
1931 /* Clean up the MD5 key list, if any */
1932 if (tp->md5sig_info) {
1933 tcp_clear_md5_list(sk);
1934 kfree_rcu(rcu_dereference_protected(tp->md5sig_info, 1), rcu);
1935 tp->md5sig_info = NULL;
1936 }
1937#endif
1938
1939 /* Clean up a referenced TCP bind bucket. */
1940 if (inet_csk(sk)->icsk_bind_hash)
1941 inet_put_port(sk);
1942
1943 BUG_ON(tp->fastopen_rsk);
1944
1945 /* If socket is aborted during connect operation */
1946 tcp_free_fastopen_req(tp);
1947 tcp_fastopen_destroy_cipher(sk);
1948 tcp_saved_syn_free(tp);
1949
1950 sk_sockets_allocated_dec(sk);
1951}
1952EXPORT_SYMBOL(tcp_v4_destroy_sock);
1953
1954#ifdef CONFIG_PROC_FS
1955/* Proc filesystem TCP sock list dumping. */
1956
1957/*
1958 * Get next listener socket follow cur. If cur is NULL, get first socket
1959 * starting from bucket given in st->bucket; when st->bucket is zero the
1960 * very first socket in the hash table is returned.
1961 */
1962static void *listening_get_next(struct seq_file *seq, void *cur)
1963{
1964 struct tcp_iter_state *st = seq->private;
1965 struct net *net = seq_file_net(seq);
1966 struct inet_listen_hashbucket *ilb;
1967 struct sock *sk = cur;
1968
1969 if (!sk) {
1970get_head:
1971 ilb = &tcp_hashinfo.listening_hash[st->bucket];
1972 spin_lock(&ilb->lock);
1973 sk = sk_head(&ilb->head);
1974 st->offset = 0;
1975 goto get_sk;
1976 }
1977 ilb = &tcp_hashinfo.listening_hash[st->bucket];
1978 ++st->num;
1979 ++st->offset;
1980
1981 sk = sk_next(sk);
1982get_sk:
1983 sk_for_each_from(sk) {
1984 if (!net_eq(sock_net(sk), net))
1985 continue;
1986 if (sk->sk_family == st->family)
1987 return sk;
1988 }
1989 spin_unlock(&ilb->lock);
1990 st->offset = 0;
1991 if (++st->bucket < INET_LHTABLE_SIZE)
1992 goto get_head;
1993 return NULL;
1994}
1995
1996static void *listening_get_idx(struct seq_file *seq, loff_t *pos)
1997{
1998 struct tcp_iter_state *st = seq->private;
1999 void *rc;
2000
2001 st->bucket = 0;
2002 st->offset = 0;
2003 rc = listening_get_next(seq, NULL);
2004
2005 while (rc && *pos) {
2006 rc = listening_get_next(seq, rc);
2007 --*pos;
2008 }
2009 return rc;
2010}
2011
2012static inline bool empty_bucket(const struct tcp_iter_state *st)
2013{
2014 return hlist_nulls_empty(&tcp_hashinfo.ehash[st->bucket].chain);
2015}
2016
2017/*
2018 * Get first established socket starting from bucket given in st->bucket.
2019 * If st->bucket is zero, the very first socket in the hash is returned.
2020 */
2021static void *established_get_first(struct seq_file *seq)
2022{
2023 struct tcp_iter_state *st = seq->private;
2024 struct net *net = seq_file_net(seq);
2025 void *rc = NULL;
2026
2027 st->offset = 0;
2028 for (; st->bucket <= tcp_hashinfo.ehash_mask; ++st->bucket) {
2029 struct sock *sk;
2030 struct hlist_nulls_node *node;
2031 spinlock_t *lock = inet_ehash_lockp(&tcp_hashinfo, st->bucket);
2032
2033 /* Lockless fast path for the common case of empty buckets */
2034 if (empty_bucket(st))
2035 continue;
2036
2037 spin_lock_bh(lock);
2038 sk_nulls_for_each(sk, node, &tcp_hashinfo.ehash[st->bucket].chain) {
2039 if (sk->sk_family != st->family ||
2040 !net_eq(sock_net(sk), net)) {
2041 continue;
2042 }
2043 rc = sk;
2044 goto out;
2045 }
2046 spin_unlock_bh(lock);
2047 }
2048out:
2049 return rc;
2050}
2051
2052static void *established_get_next(struct seq_file *seq, void *cur)
2053{
2054 struct sock *sk = cur;
2055 struct hlist_nulls_node *node;
2056 struct tcp_iter_state *st = seq->private;
2057 struct net *net = seq_file_net(seq);
2058
2059 ++st->num;
2060 ++st->offset;
2061
2062 sk = sk_nulls_next(sk);
2063
2064 sk_nulls_for_each_from(sk, node) {
2065 if (sk->sk_family == st->family && net_eq(sock_net(sk), net))
2066 return sk;
2067 }
2068
2069 spin_unlock_bh(inet_ehash_lockp(&tcp_hashinfo, st->bucket));
2070 ++st->bucket;
2071 return established_get_first(seq);
2072}
2073
2074static void *established_get_idx(struct seq_file *seq, loff_t pos)
2075{
2076 struct tcp_iter_state *st = seq->private;
2077 void *rc;
2078
2079 st->bucket = 0;
2080 rc = established_get_first(seq);
2081
2082 while (rc && pos) {
2083 rc = established_get_next(seq, rc);
2084 --pos;
2085 }
2086 return rc;
2087}
2088
2089static void *tcp_get_idx(struct seq_file *seq, loff_t pos)
2090{
2091 void *rc;
2092 struct tcp_iter_state *st = seq->private;
2093
2094 st->state = TCP_SEQ_STATE_LISTENING;
2095 rc = listening_get_idx(seq, &pos);
2096
2097 if (!rc) {
2098 st->state = TCP_SEQ_STATE_ESTABLISHED;
2099 rc = established_get_idx(seq, pos);
2100 }
2101
2102 return rc;
2103}
2104
2105static void *tcp_seek_last_pos(struct seq_file *seq)
2106{
2107 struct tcp_iter_state *st = seq->private;
2108 int offset = st->offset;
2109 int orig_num = st->num;
2110 void *rc = NULL;
2111
2112 switch (st->state) {
2113 case TCP_SEQ_STATE_LISTENING:
2114 if (st->bucket >= INET_LHTABLE_SIZE)
2115 break;
2116 st->state = TCP_SEQ_STATE_LISTENING;
2117 rc = listening_get_next(seq, NULL);
2118 while (offset-- && rc)
2119 rc = listening_get_next(seq, rc);
2120 if (rc)
2121 break;
2122 st->bucket = 0;
2123 st->state = TCP_SEQ_STATE_ESTABLISHED;
2124 /* Fallthrough */
2125 case TCP_SEQ_STATE_ESTABLISHED:
2126 if (st->bucket > tcp_hashinfo.ehash_mask)
2127 break;
2128 rc = established_get_first(seq);
2129 while (offset-- && rc)
2130 rc = established_get_next(seq, rc);
2131 }
2132
2133 st->num = orig_num;
2134
2135 return rc;
2136}
2137
2138static void *tcp_seq_start(struct seq_file *seq, loff_t *pos)
2139{
2140 struct tcp_iter_state *st = seq->private;
2141 void *rc;
2142
2143 if (*pos && *pos == st->last_pos) {
2144 rc = tcp_seek_last_pos(seq);
2145 if (rc)
2146 goto out;
2147 }
2148
2149 st->state = TCP_SEQ_STATE_LISTENING;
2150 st->num = 0;
2151 st->bucket = 0;
2152 st->offset = 0;
2153 rc = *pos ? tcp_get_idx(seq, *pos - 1) : SEQ_START_TOKEN;
2154
2155out:
2156 st->last_pos = *pos;
2157 return rc;
2158}
2159
2160static void *tcp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2161{
2162 struct tcp_iter_state *st = seq->private;
2163 void *rc = NULL;
2164
2165 if (v == SEQ_START_TOKEN) {
2166 rc = tcp_get_idx(seq, 0);
2167 goto out;
2168 }
2169
2170 switch (st->state) {
2171 case TCP_SEQ_STATE_LISTENING:
2172 rc = listening_get_next(seq, v);
2173 if (!rc) {
2174 st->state = TCP_SEQ_STATE_ESTABLISHED;
2175 st->bucket = 0;
2176 st->offset = 0;
2177 rc = established_get_first(seq);
2178 }
2179 break;
2180 case TCP_SEQ_STATE_ESTABLISHED:
2181 rc = established_get_next(seq, v);
2182 break;
2183 }
2184out:
2185 ++*pos;
2186 st->last_pos = *pos;
2187 return rc;
2188}
2189
2190static void tcp_seq_stop(struct seq_file *seq, void *v)
2191{
2192 struct tcp_iter_state *st = seq->private;
2193
2194 switch (st->state) {
2195 case TCP_SEQ_STATE_LISTENING:
2196 if (v != SEQ_START_TOKEN)
2197 spin_unlock(&tcp_hashinfo.listening_hash[st->bucket].lock);
2198 break;
2199 case TCP_SEQ_STATE_ESTABLISHED:
2200 if (v)
2201 spin_unlock_bh(inet_ehash_lockp(&tcp_hashinfo, st->bucket));
2202 break;
2203 }
2204}
2205
2206int tcp_seq_open(struct inode *inode, struct file *file)
2207{
2208 struct tcp_seq_afinfo *afinfo = PDE_DATA(inode);
2209 struct tcp_iter_state *s;
2210 int err;
2211
2212 err = seq_open_net(inode, file, &afinfo->seq_ops,
2213 sizeof(struct tcp_iter_state));
2214 if (err < 0)
2215 return err;
2216
2217 s = ((struct seq_file *)file->private_data)->private;
2218 s->family = afinfo->family;
2219 s->last_pos = 0;
2220 return 0;
2221}
2222EXPORT_SYMBOL(tcp_seq_open);
2223
2224int tcp_proc_register(struct net *net, struct tcp_seq_afinfo *afinfo)
2225{
2226 int rc = 0;
2227 struct proc_dir_entry *p;
2228
2229 afinfo->seq_ops.start = tcp_seq_start;
2230 afinfo->seq_ops.next = tcp_seq_next;
2231 afinfo->seq_ops.stop = tcp_seq_stop;
2232
2233 p = proc_create_data(afinfo->name, 0444, net->proc_net,
2234 afinfo->seq_fops, afinfo);
2235 if (!p)
2236 rc = -ENOMEM;
2237 return rc;
2238}
2239EXPORT_SYMBOL(tcp_proc_register);
2240
2241void tcp_proc_unregister(struct net *net, struct tcp_seq_afinfo *afinfo)
2242{
2243 remove_proc_entry(afinfo->name, net->proc_net);
2244}
2245EXPORT_SYMBOL(tcp_proc_unregister);
2246
2247static void get_openreq4(const struct request_sock *req,
2248 struct seq_file *f, int i)
2249{
2250 const struct inet_request_sock *ireq = inet_rsk(req);
2251 long delta = req->rsk_timer.expires - jiffies;
2252
2253 seq_printf(f, "%4d: %08X:%04X %08X:%04X"
2254 " %02X %08X:%08X %02X:%08lX %08X %5u %8d %u %d %pK",
2255 i,
2256 ireq->ir_loc_addr,
2257 ireq->ir_num,
2258 ireq->ir_rmt_addr,
2259 ntohs(ireq->ir_rmt_port),
2260 TCP_SYN_RECV,
2261 0, 0, /* could print option size, but that is af dependent. */
2262 1, /* timers active (only the expire timer) */
2263 jiffies_delta_to_clock_t(delta),
2264 req->num_timeout,
2265 from_kuid_munged(seq_user_ns(f),
2266 sock_i_uid(req->rsk_listener)),
2267 0, /* non standard timer */
2268 0, /* open_requests have no inode */
2269 0,
2270 req);
2271}
2272
2273static void get_tcp4_sock(struct sock *sk, struct seq_file *f, int i)
2274{
2275 int timer_active;
2276 unsigned long timer_expires;
2277 const struct tcp_sock *tp = tcp_sk(sk);
2278 const struct inet_connection_sock *icsk = inet_csk(sk);
2279 const struct inet_sock *inet = inet_sk(sk);
2280 const struct fastopen_queue *fastopenq = &icsk->icsk_accept_queue.fastopenq;
2281 __be32 dest = inet->inet_daddr;
2282 __be32 src = inet->inet_rcv_saddr;
2283 __u16 destp = ntohs(inet->inet_dport);
2284 __u16 srcp = ntohs(inet->inet_sport);
2285 int rx_queue;
2286 int state;
2287
2288 if (icsk->icsk_pending == ICSK_TIME_RETRANS ||
2289 icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
2290 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2291 timer_active = 1;
2292 timer_expires = icsk->icsk_timeout;
2293 } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
2294 timer_active = 4;
2295 timer_expires = icsk->icsk_timeout;
2296 } else if (timer_pending(&sk->sk_timer)) {
2297 timer_active = 2;
2298 timer_expires = sk->sk_timer.expires;
2299 } else {
2300 timer_active = 0;
2301 timer_expires = jiffies;
2302 }
2303
2304 state = inet_sk_state_load(sk);
2305 if (state == TCP_LISTEN)
2306 rx_queue = sk->sk_ack_backlog;
2307 else
2308 /* Because we don't lock the socket,
2309 * we might find a transient negative value.
2310 */
2311 rx_queue = max_t(int, tp->rcv_nxt - tp->copied_seq, 0);
2312
2313 seq_printf(f, "%4d: %08X:%04X %08X:%04X %02X %08X:%08X %02X:%08lX "
2314 "%08X %5u %8d %lu %d %pK %lu %lu %u %u %d",
2315 i, src, srcp, dest, destp, state,
2316 tp->write_seq - tp->snd_una,
2317 rx_queue,
2318 timer_active,
2319 jiffies_delta_to_clock_t(timer_expires - jiffies),
2320 icsk->icsk_retransmits,
2321 from_kuid_munged(seq_user_ns(f), sock_i_uid(sk)),
2322 icsk->icsk_probes_out,
2323 sock_i_ino(sk),
2324 refcount_read(&sk->sk_refcnt), sk,
2325 jiffies_to_clock_t(icsk->icsk_rto),
2326 jiffies_to_clock_t(icsk->icsk_ack.ato),
2327 (icsk->icsk_ack.quick << 1) | icsk->icsk_ack.pingpong,
2328 tp->snd_cwnd,
2329 state == TCP_LISTEN ?
2330 fastopenq->max_qlen :
2331 (tcp_in_initial_slowstart(tp) ? -1 : tp->snd_ssthresh));
2332}
2333
2334static void get_timewait4_sock(const struct inet_timewait_sock *tw,
2335 struct seq_file *f, int i)
2336{
2337 long delta = tw->tw_timer.expires - jiffies;
2338 __be32 dest, src;
2339 __u16 destp, srcp;
2340
2341 dest = tw->tw_daddr;
2342 src = tw->tw_rcv_saddr;
2343 destp = ntohs(tw->tw_dport);
2344 srcp = ntohs(tw->tw_sport);
2345
2346 seq_printf(f, "%4d: %08X:%04X %08X:%04X"
2347 " %02X %08X:%08X %02X:%08lX %08X %5d %8d %d %d %pK",
2348 i, src, srcp, dest, destp, tw->tw_substate, 0, 0,
2349 3, jiffies_delta_to_clock_t(delta), 0, 0, 0, 0,
2350 refcount_read(&tw->tw_refcnt), tw);
2351}
2352
2353#define TMPSZ 150
2354
2355static int tcp4_seq_show(struct seq_file *seq, void *v)
2356{
2357 struct tcp_iter_state *st;
2358 struct sock *sk = v;
2359
2360 seq_setwidth(seq, TMPSZ - 1);
2361 if (v == SEQ_START_TOKEN) {
2362 seq_puts(seq, " sl local_address rem_address st tx_queue "
2363 "rx_queue tr tm->when retrnsmt uid timeout "
2364 "inode");
2365 goto out;
2366 }
2367 st = seq->private;
2368
2369 if (sk->sk_state == TCP_TIME_WAIT)
2370 get_timewait4_sock(v, seq, st->num);
2371 else if (sk->sk_state == TCP_NEW_SYN_RECV)
2372 get_openreq4(v, seq, st->num);
2373 else
2374 get_tcp4_sock(v, seq, st->num);
2375out:
2376 seq_pad(seq, '\n');
2377 return 0;
2378}
2379
2380static const struct file_operations tcp_afinfo_seq_fops = {
2381 .open = tcp_seq_open,
2382 .read = seq_read,
2383 .llseek = seq_lseek,
2384 .release = seq_release_net
2385};
2386
2387static struct tcp_seq_afinfo tcp4_seq_afinfo = {
2388 .name = "tcp",
2389 .family = AF_INET,
2390 .seq_fops = &tcp_afinfo_seq_fops,
2391 .seq_ops = {
2392 .show = tcp4_seq_show,
2393 },
2394};
2395
2396static int __net_init tcp4_proc_init_net(struct net *net)
2397{
2398 return tcp_proc_register(net, &tcp4_seq_afinfo);
2399}
2400
2401static void __net_exit tcp4_proc_exit_net(struct net *net)
2402{
2403 tcp_proc_unregister(net, &tcp4_seq_afinfo);
2404}
2405
2406static struct pernet_operations tcp4_net_ops = {
2407 .init = tcp4_proc_init_net,
2408 .exit = tcp4_proc_exit_net,
2409};
2410
2411int __init tcp4_proc_init(void)
2412{
2413 return register_pernet_subsys(&tcp4_net_ops);
2414}
2415
2416void tcp4_proc_exit(void)
2417{
2418 unregister_pernet_subsys(&tcp4_net_ops);
2419}
2420#endif /* CONFIG_PROC_FS */
2421
2422struct proto tcp_prot = {
2423 .name = "TCP",
2424 .owner = THIS_MODULE,
2425 .close = tcp_close,
2426 .pre_connect = tcp_v4_pre_connect,
2427 .connect = tcp_v4_connect,
2428 .disconnect = tcp_disconnect,
2429 .accept = inet_csk_accept,
2430 .ioctl = tcp_ioctl,
2431 .init = tcp_v4_init_sock,
2432 .destroy = tcp_v4_destroy_sock,
2433 .shutdown = tcp_shutdown,
2434 .setsockopt = tcp_setsockopt,
2435 .getsockopt = tcp_getsockopt,
2436 .keepalive = tcp_set_keepalive,
2437 .recvmsg = tcp_recvmsg,
2438 .sendmsg = tcp_sendmsg,
2439 .sendpage = tcp_sendpage,
2440 .backlog_rcv = tcp_v4_do_rcv,
2441 .release_cb = tcp_release_cb,
2442 .hash = inet_hash,
2443 .unhash = inet_unhash,
2444 .get_port = inet_csk_get_port,
2445 .enter_memory_pressure = tcp_enter_memory_pressure,
2446 .leave_memory_pressure = tcp_leave_memory_pressure,
2447 .stream_memory_free = tcp_stream_memory_free,
2448 .sockets_allocated = &tcp_sockets_allocated,
2449 .orphan_count = &tcp_orphan_count,
2450 .memory_allocated = &tcp_memory_allocated,
2451 .memory_pressure = &tcp_memory_pressure,
2452 .sysctl_mem = sysctl_tcp_mem,
2453 .sysctl_wmem_offset = offsetof(struct net, ipv4.sysctl_tcp_wmem),
2454 .sysctl_rmem_offset = offsetof(struct net, ipv4.sysctl_tcp_rmem),
2455 .max_header = MAX_TCP_HEADER,
2456 .obj_size = sizeof(struct tcp_sock),
2457 .slab_flags = SLAB_TYPESAFE_BY_RCU,
2458 .twsk_prot = &tcp_timewait_sock_ops,
2459 .rsk_prot = &tcp_request_sock_ops,
2460 .h.hashinfo = &tcp_hashinfo,
2461 .no_autobind = true,
2462#ifdef CONFIG_COMPAT
2463 .compat_setsockopt = compat_tcp_setsockopt,
2464 .compat_getsockopt = compat_tcp_getsockopt,
2465#endif
2466 .diag_destroy = tcp_abort,
2467};
2468EXPORT_SYMBOL(tcp_prot);
2469
2470static void __net_exit tcp_sk_exit(struct net *net)
2471{
2472 int cpu;
2473
2474 module_put(net->ipv4.tcp_congestion_control->owner);
2475
2476 for_each_possible_cpu(cpu)
2477 inet_ctl_sock_destroy(*per_cpu_ptr(net->ipv4.tcp_sk, cpu));
2478 free_percpu(net->ipv4.tcp_sk);
2479}
2480
2481static int __net_init tcp_sk_init(struct net *net)
2482{
2483 int res, cpu, cnt;
2484
2485 net->ipv4.tcp_sk = alloc_percpu(struct sock *);
2486 if (!net->ipv4.tcp_sk)
2487 return -ENOMEM;
2488
2489 for_each_possible_cpu(cpu) {
2490 struct sock *sk;
2491
2492 res = inet_ctl_sock_create(&sk, PF_INET, SOCK_RAW,
2493 IPPROTO_TCP, net);
2494 if (res)
2495 goto fail;
2496 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
2497 *per_cpu_ptr(net->ipv4.tcp_sk, cpu) = sk;
2498 }
2499
2500 net->ipv4.sysctl_tcp_ecn = 2;
2501 net->ipv4.sysctl_tcp_ecn_fallback = 1;
2502
2503 net->ipv4.sysctl_tcp_base_mss = TCP_BASE_MSS;
2504 net->ipv4.sysctl_tcp_probe_threshold = TCP_PROBE_THRESHOLD;
2505 net->ipv4.sysctl_tcp_probe_interval = TCP_PROBE_INTERVAL;
2506
2507 net->ipv4.sysctl_tcp_keepalive_time = TCP_KEEPALIVE_TIME;
2508 net->ipv4.sysctl_tcp_keepalive_probes = TCP_KEEPALIVE_PROBES;
2509 net->ipv4.sysctl_tcp_keepalive_intvl = TCP_KEEPALIVE_INTVL;
2510
2511 net->ipv4.sysctl_tcp_syn_retries = TCP_SYN_RETRIES;
2512 net->ipv4.sysctl_tcp_synack_retries = TCP_SYNACK_RETRIES;
2513 net->ipv4.sysctl_tcp_syncookies = 1;
2514 net->ipv4.sysctl_tcp_reordering = TCP_FASTRETRANS_THRESH;
2515 net->ipv4.sysctl_tcp_retries1 = TCP_RETR1;
2516 net->ipv4.sysctl_tcp_retries2 = TCP_RETR2;
2517 net->ipv4.sysctl_tcp_orphan_retries = 0;
2518 net->ipv4.sysctl_tcp_fin_timeout = TCP_FIN_TIMEOUT;
2519 net->ipv4.sysctl_tcp_notsent_lowat = UINT_MAX;
2520 net->ipv4.sysctl_tcp_tw_reuse = 0;
2521
2522 cnt = tcp_hashinfo.ehash_mask + 1;
2523 net->ipv4.tcp_death_row.sysctl_max_tw_buckets = (cnt + 1) / 2;
2524 net->ipv4.tcp_death_row.hashinfo = &tcp_hashinfo;
2525
2526 net->ipv4.sysctl_max_syn_backlog = max(128, cnt / 256);
2527 net->ipv4.sysctl_tcp_sack = 1;
2528 net->ipv4.sysctl_tcp_window_scaling = 1;
2529 net->ipv4.sysctl_tcp_timestamps = 1;
2530 net->ipv4.sysctl_tcp_early_retrans = 3;
2531 net->ipv4.sysctl_tcp_recovery = TCP_RACK_LOSS_DETECTION;
2532 net->ipv4.sysctl_tcp_slow_start_after_idle = 1; /* By default, RFC2861 behavior. */
2533 net->ipv4.sysctl_tcp_retrans_collapse = 1;
2534 net->ipv4.sysctl_tcp_max_reordering = 300;
2535 net->ipv4.sysctl_tcp_dsack = 1;
2536 net->ipv4.sysctl_tcp_app_win = 31;
2537 net->ipv4.sysctl_tcp_adv_win_scale = 1;
2538 net->ipv4.sysctl_tcp_frto = 2;
2539 net->ipv4.sysctl_tcp_moderate_rcvbuf = 1;
2540 /* This limits the percentage of the congestion window which we
2541 * will allow a single TSO frame to consume. Building TSO frames
2542 * which are too large can cause TCP streams to be bursty.
2543 */
2544 net->ipv4.sysctl_tcp_tso_win_divisor = 3;
2545 /* Default TSQ limit of four TSO segments */
2546 net->ipv4.sysctl_tcp_limit_output_bytes = 262144;
2547 /* rfc5961 challenge ack rate limiting */
2548 net->ipv4.sysctl_tcp_challenge_ack_limit = 1000;
2549 net->ipv4.sysctl_tcp_min_tso_segs = 2;
2550 net->ipv4.sysctl_tcp_min_rtt_wlen = 300;
2551 net->ipv4.sysctl_tcp_autocorking = 1;
2552 net->ipv4.sysctl_tcp_invalid_ratelimit = HZ/2;
2553 net->ipv4.sysctl_tcp_pacing_ss_ratio = 200;
2554 net->ipv4.sysctl_tcp_pacing_ca_ratio = 120;
2555 if (net != &init_net) {
2556 memcpy(net->ipv4.sysctl_tcp_rmem,
2557 init_net.ipv4.sysctl_tcp_rmem,
2558 sizeof(init_net.ipv4.sysctl_tcp_rmem));
2559 memcpy(net->ipv4.sysctl_tcp_wmem,
2560 init_net.ipv4.sysctl_tcp_wmem,
2561 sizeof(init_net.ipv4.sysctl_tcp_wmem));
2562 }
2563 net->ipv4.sysctl_tcp_fastopen = TFO_CLIENT_ENABLE;
2564 spin_lock_init(&net->ipv4.tcp_fastopen_ctx_lock);
2565 net->ipv4.sysctl_tcp_fastopen_blackhole_timeout = 60 * 60;
2566 atomic_set(&net->ipv4.tfo_active_disable_times, 0);
2567
2568 /* Reno is always built in */
2569 if (!net_eq(net, &init_net) &&
2570 try_module_get(init_net.ipv4.tcp_congestion_control->owner))
2571 net->ipv4.tcp_congestion_control = init_net.ipv4.tcp_congestion_control;
2572 else
2573 net->ipv4.tcp_congestion_control = &tcp_reno;
2574
2575 return 0;
2576fail:
2577 tcp_sk_exit(net);
2578
2579 return res;
2580}
2581
2582static void __net_exit tcp_sk_exit_batch(struct list_head *net_exit_list)
2583{
2584 struct net *net;
2585
2586 inet_twsk_purge(&tcp_hashinfo, AF_INET);
2587
2588 list_for_each_entry(net, net_exit_list, exit_list)
2589 tcp_fastopen_ctx_destroy(net);
2590}
2591
2592static struct pernet_operations __net_initdata tcp_sk_ops = {
2593 .init = tcp_sk_init,
2594 .exit = tcp_sk_exit,
2595 .exit_batch = tcp_sk_exit_batch,
2596};
2597
2598void __init tcp_v4_init(void)
2599{
2600 if (register_pernet_subsys(&tcp_sk_ops))
2601 panic("Failed to create the TCP control socket.\n");
2602}