xref: /openbmc/linux/net/ipv4/inet_diag.c (revision a8da474e) 
1 /*
2  * inet_diag.c	Module for monitoring INET transport protocols sockets.
3  *
4  * Authors:	Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
5  *
6  *	This program is free software; you can redistribute it and/or
7  *      modify it under the terms of the GNU General Public License
8  *      as published by the Free Software Foundation; either version
9  *      2 of the License, or (at your option) any later version.
10  */
11 
12 #include <linux/kernel.h>
13 #include <linux/module.h>
14 #include <linux/types.h>
15 #include <linux/fcntl.h>
16 #include <linux/random.h>
17 #include <linux/slab.h>
18 #include <linux/cache.h>
19 #include <linux/init.h>
20 #include <linux/time.h>
21 
22 #include <net/icmp.h>
23 #include <net/tcp.h>
24 #include <net/ipv6.h>
25 #include <net/inet_common.h>
26 #include <net/inet_connection_sock.h>
27 #include <net/inet_hashtables.h>
28 #include <net/inet_timewait_sock.h>
29 #include <net/inet6_hashtables.h>
30 #include <net/netlink.h>
31 
32 #include <linux/inet.h>
33 #include <linux/stddef.h>
34 
35 #include <linux/inet_diag.h>
36 #include <linux/sock_diag.h>
37 
38 static const struct inet_diag_handler **inet_diag_table;
39 
40 struct inet_diag_entry {
41 	const __be32 *saddr;
42 	const __be32 *daddr;
43 	u16 sport;
44 	u16 dport;
45 	u16 family;
46 	u16 userlocks;
47 };
48 
49 static DEFINE_MUTEX(inet_diag_table_mutex);
50 
51 static const struct inet_diag_handler *inet_diag_lock_handler(int proto)
52 {
53 	if (!inet_diag_table[proto])
54 		request_module("net-pf-%d-proto-%d-type-%d-%d", PF_NETLINK,
55 			       NETLINK_SOCK_DIAG, AF_INET, proto);
56 
57 	mutex_lock(&inet_diag_table_mutex);
58 	if (!inet_diag_table[proto])
59 		return ERR_PTR(-ENOENT);
60 
61 	return inet_diag_table[proto];
62 }
63 
64 static void inet_diag_unlock_handler(const struct inet_diag_handler *handler)
65 {
66 	mutex_unlock(&inet_diag_table_mutex);
67 }
68 
69 static void inet_diag_msg_common_fill(struct inet_diag_msg *r, struct sock *sk)
70 {
71 	r->idiag_family = sk->sk_family;
72 
73 	r->id.idiag_sport = htons(sk->sk_num);
74 	r->id.idiag_dport = sk->sk_dport;
75 	r->id.idiag_if = sk->sk_bound_dev_if;
76 	sock_diag_save_cookie(sk, r->id.idiag_cookie);
77 
78 #if IS_ENABLED(CONFIG_IPV6)
79 	if (sk->sk_family == AF_INET6) {
80 		*(struct in6_addr *)r->id.idiag_src = sk->sk_v6_rcv_saddr;
81 		*(struct in6_addr *)r->id.idiag_dst = sk->sk_v6_daddr;
82 	} else
83 #endif
84 	{
85 	memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
86 	memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
87 
88 	r->id.idiag_src[0] = sk->sk_rcv_saddr;
89 	r->id.idiag_dst[0] = sk->sk_daddr;
90 	}
91 }
92 
93 static size_t inet_sk_attr_size(void)
94 {
95 	return	  nla_total_size(sizeof(struct tcp_info))
96 		+ nla_total_size(1) /* INET_DIAG_SHUTDOWN */
97 		+ nla_total_size(1) /* INET_DIAG_TOS */
98 		+ nla_total_size(1) /* INET_DIAG_TCLASS */
99 		+ nla_total_size(sizeof(struct inet_diag_meminfo))
100 		+ nla_total_size(sizeof(struct inet_diag_msg))
101 		+ nla_total_size(SK_MEMINFO_VARS * sizeof(u32))
102 		+ nla_total_size(TCP_CA_NAME_MAX)
103 		+ nla_total_size(sizeof(struct tcpvegas_info))
104 		+ 64;
105 }
106 
107 int inet_sk_diag_fill(struct sock *sk, struct inet_connection_sock *icsk,
108 		      struct sk_buff *skb, const struct inet_diag_req_v2 *req,
109 		      struct user_namespace *user_ns,
110 		      u32 portid, u32 seq, u16 nlmsg_flags,
111 		      const struct nlmsghdr *unlh)
112 {
113 	const struct inet_sock *inet = inet_sk(sk);
114 	const struct tcp_congestion_ops *ca_ops;
115 	const struct inet_diag_handler *handler;
116 	int ext = req->idiag_ext;
117 	struct inet_diag_msg *r;
118 	struct nlmsghdr  *nlh;
119 	struct nlattr *attr;
120 	void *info = NULL;
121 
122 	handler = inet_diag_table[req->sdiag_protocol];
123 	BUG_ON(!handler);
124 
125 	nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
126 			nlmsg_flags);
127 	if (!nlh)
128 		return -EMSGSIZE;
129 
130 	r = nlmsg_data(nlh);
131 	BUG_ON(!sk_fullsock(sk));
132 
133 	inet_diag_msg_common_fill(r, sk);
134 	r->idiag_state = sk->sk_state;
135 	r->idiag_timer = 0;
136 	r->idiag_retrans = 0;
137 
138 	if (nla_put_u8(skb, INET_DIAG_SHUTDOWN, sk->sk_shutdown))
139 		goto errout;
140 
141 	/* IPv6 dual-stack sockets use inet->tos for IPv4 connections,
142 	 * hence this needs to be included regardless of socket family.
143 	 */
144 	if (ext & (1 << (INET_DIAG_TOS - 1)))
145 		if (nla_put_u8(skb, INET_DIAG_TOS, inet->tos) < 0)
146 			goto errout;
147 
148 #if IS_ENABLED(CONFIG_IPV6)
149 	if (r->idiag_family == AF_INET6) {
150 		if (ext & (1 << (INET_DIAG_TCLASS - 1)))
151 			if (nla_put_u8(skb, INET_DIAG_TCLASS,
152 				       inet6_sk(sk)->tclass) < 0)
153 				goto errout;
154 
155 		if (((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) &&
156 		    nla_put_u8(skb, INET_DIAG_SKV6ONLY, ipv6_only_sock(sk)))
157 			goto errout;
158 	}
159 #endif
160 
161 	r->idiag_uid = from_kuid_munged(user_ns, sock_i_uid(sk));
162 	r->idiag_inode = sock_i_ino(sk);
163 
164 	if (ext & (1 << (INET_DIAG_MEMINFO - 1))) {
165 		struct inet_diag_meminfo minfo = {
166 			.idiag_rmem = sk_rmem_alloc_get(sk),
167 			.idiag_wmem = sk->sk_wmem_queued,
168 			.idiag_fmem = sk->sk_forward_alloc,
169 			.idiag_tmem = sk_wmem_alloc_get(sk),
170 		};
171 
172 		if (nla_put(skb, INET_DIAG_MEMINFO, sizeof(minfo), &minfo) < 0)
173 			goto errout;
174 	}
175 
176 	if (ext & (1 << (INET_DIAG_SKMEMINFO - 1)))
177 		if (sock_diag_put_meminfo(sk, skb, INET_DIAG_SKMEMINFO))
178 			goto errout;
179 
180 	if (!icsk) {
181 		handler->idiag_get_info(sk, r, NULL);
182 		goto out;
183 	}
184 
185 #define EXPIRES_IN_MS(tmo)  DIV_ROUND_UP((tmo - jiffies) * 1000, HZ)
186 
187 	if (icsk->icsk_pending == ICSK_TIME_RETRANS ||
188 	    icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
189 	    icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
190 		r->idiag_timer = 1;
191 		r->idiag_retrans = icsk->icsk_retransmits;
192 		r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
193 	} else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
194 		r->idiag_timer = 4;
195 		r->idiag_retrans = icsk->icsk_probes_out;
196 		r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
197 	} else if (timer_pending(&sk->sk_timer)) {
198 		r->idiag_timer = 2;
199 		r->idiag_retrans = icsk->icsk_probes_out;
200 		r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
201 	} else {
202 		r->idiag_timer = 0;
203 		r->idiag_expires = 0;
204 	}
205 #undef EXPIRES_IN_MS
206 
207 	if ((ext & (1 << (INET_DIAG_INFO - 1))) && handler->idiag_info_size) {
208 		attr = nla_reserve(skb, INET_DIAG_INFO,
209 				   handler->idiag_info_size);
210 		if (!attr)
211 			goto errout;
212 
213 		info = nla_data(attr);
214 	}
215 
216 	if (ext & (1 << (INET_DIAG_CONG - 1))) {
217 		int err = 0;
218 
219 		rcu_read_lock();
220 		ca_ops = READ_ONCE(icsk->icsk_ca_ops);
221 		if (ca_ops)
222 			err = nla_put_string(skb, INET_DIAG_CONG, ca_ops->name);
223 		rcu_read_unlock();
224 		if (err < 0)
225 			goto errout;
226 	}
227 
228 	handler->idiag_get_info(sk, r, info);
229 
230 	if (sk->sk_state < TCP_TIME_WAIT) {
231 		union tcp_cc_info info;
232 		size_t sz = 0;
233 		int attr;
234 
235 		rcu_read_lock();
236 		ca_ops = READ_ONCE(icsk->icsk_ca_ops);
237 		if (ca_ops && ca_ops->get_info)
238 			sz = ca_ops->get_info(sk, ext, &attr, &info);
239 		rcu_read_unlock();
240 		if (sz && nla_put(skb, attr, sz, &info) < 0)
241 			goto errout;
242 	}
243 
244 out:
245 	nlmsg_end(skb, nlh);
246 	return 0;
247 
248 errout:
249 	nlmsg_cancel(skb, nlh);
250 	return -EMSGSIZE;
251 }
252 EXPORT_SYMBOL_GPL(inet_sk_diag_fill);
253 
254 static int inet_csk_diag_fill(struct sock *sk,
255 			      struct sk_buff *skb,
256 			      const struct inet_diag_req_v2 *req,
257 			      struct user_namespace *user_ns,
258 			      u32 portid, u32 seq, u16 nlmsg_flags,
259 			      const struct nlmsghdr *unlh)
260 {
261 	return inet_sk_diag_fill(sk, inet_csk(sk), skb, req,
262 				 user_ns, portid, seq, nlmsg_flags, unlh);
263 }
264 
265 static int inet_twsk_diag_fill(struct sock *sk,
266 			       struct sk_buff *skb,
267 			       u32 portid, u32 seq, u16 nlmsg_flags,
268 			       const struct nlmsghdr *unlh)
269 {
270 	struct inet_timewait_sock *tw = inet_twsk(sk);
271 	struct inet_diag_msg *r;
272 	struct nlmsghdr *nlh;
273 	long tmo;
274 
275 	nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
276 			nlmsg_flags);
277 	if (!nlh)
278 		return -EMSGSIZE;
279 
280 	r = nlmsg_data(nlh);
281 	BUG_ON(tw->tw_state != TCP_TIME_WAIT);
282 
283 	tmo = tw->tw_timer.expires - jiffies;
284 	if (tmo < 0)
285 		tmo = 0;
286 
287 	inet_diag_msg_common_fill(r, sk);
288 	r->idiag_retrans      = 0;
289 
290 	r->idiag_state	      = tw->tw_substate;
291 	r->idiag_timer	      = 3;
292 	r->idiag_expires      = jiffies_to_msecs(tmo);
293 	r->idiag_rqueue	      = 0;
294 	r->idiag_wqueue	      = 0;
295 	r->idiag_uid	      = 0;
296 	r->idiag_inode	      = 0;
297 
298 	nlmsg_end(skb, nlh);
299 	return 0;
300 }
301 
302 static int inet_req_diag_fill(struct sock *sk, struct sk_buff *skb,
303 			      u32 portid, u32 seq, u16 nlmsg_flags,
304 			      const struct nlmsghdr *unlh)
305 {
306 	struct inet_diag_msg *r;
307 	struct nlmsghdr *nlh;
308 	long tmo;
309 
310 	nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
311 			nlmsg_flags);
312 	if (!nlh)
313 		return -EMSGSIZE;
314 
315 	r = nlmsg_data(nlh);
316 	inet_diag_msg_common_fill(r, sk);
317 	r->idiag_state = TCP_SYN_RECV;
318 	r->idiag_timer = 1;
319 	r->idiag_retrans = inet_reqsk(sk)->num_retrans;
320 
321 	BUILD_BUG_ON(offsetof(struct inet_request_sock, ir_cookie) !=
322 		     offsetof(struct sock, sk_cookie));
323 
324 	tmo = inet_reqsk(sk)->rsk_timer.expires - jiffies;
325 	r->idiag_expires = (tmo >= 0) ? jiffies_to_msecs(tmo) : 0;
326 	r->idiag_rqueue	= 0;
327 	r->idiag_wqueue	= 0;
328 	r->idiag_uid	= 0;
329 	r->idiag_inode	= 0;
330 
331 	nlmsg_end(skb, nlh);
332 	return 0;
333 }
334 
335 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
336 			const struct inet_diag_req_v2 *r,
337 			struct user_namespace *user_ns,
338 			u32 portid, u32 seq, u16 nlmsg_flags,
339 			const struct nlmsghdr *unlh)
340 {
341 	if (sk->sk_state == TCP_TIME_WAIT)
342 		return inet_twsk_diag_fill(sk, skb, portid, seq,
343 					   nlmsg_flags, unlh);
344 
345 	if (sk->sk_state == TCP_NEW_SYN_RECV)
346 		return inet_req_diag_fill(sk, skb, portid, seq,
347 					  nlmsg_flags, unlh);
348 
349 	return inet_csk_diag_fill(sk, skb, r, user_ns, portid, seq,
350 				  nlmsg_flags, unlh);
351 }
352 
353 int inet_diag_dump_one_icsk(struct inet_hashinfo *hashinfo,
354 			    struct sk_buff *in_skb,
355 			    const struct nlmsghdr *nlh,
356 			    const struct inet_diag_req_v2 *req)
357 {
358 	struct net *net = sock_net(in_skb->sk);
359 	struct sk_buff *rep;
360 	struct sock *sk;
361 	int err;
362 
363 	err = -EINVAL;
364 	if (req->sdiag_family == AF_INET)
365 		sk = inet_lookup(net, hashinfo, req->id.idiag_dst[0],
366 				 req->id.idiag_dport, req->id.idiag_src[0],
367 				 req->id.idiag_sport, req->id.idiag_if);
368 #if IS_ENABLED(CONFIG_IPV6)
369 	else if (req->sdiag_family == AF_INET6)
370 		sk = inet6_lookup(net, hashinfo,
371 				  (struct in6_addr *)req->id.idiag_dst,
372 				  req->id.idiag_dport,
373 				  (struct in6_addr *)req->id.idiag_src,
374 				  req->id.idiag_sport,
375 				  req->id.idiag_if);
376 #endif
377 	else
378 		goto out_nosk;
379 
380 	err = -ENOENT;
381 	if (!sk)
382 		goto out_nosk;
383 
384 	err = sock_diag_check_cookie(sk, req->id.idiag_cookie);
385 	if (err)
386 		goto out;
387 
388 	rep = nlmsg_new(inet_sk_attr_size(), GFP_KERNEL);
389 	if (!rep) {
390 		err = -ENOMEM;
391 		goto out;
392 	}
393 
394 	err = sk_diag_fill(sk, rep, req,
395 			   sk_user_ns(NETLINK_CB(in_skb).sk),
396 			   NETLINK_CB(in_skb).portid,
397 			   nlh->nlmsg_seq, 0, nlh);
398 	if (err < 0) {
399 		WARN_ON(err == -EMSGSIZE);
400 		nlmsg_free(rep);
401 		goto out;
402 	}
403 	err = netlink_unicast(net->diag_nlsk, rep, NETLINK_CB(in_skb).portid,
404 			      MSG_DONTWAIT);
405 	if (err > 0)
406 		err = 0;
407 
408 out:
409 	if (sk)
410 		sock_gen_put(sk);
411 
412 out_nosk:
413 	return err;
414 }
415 EXPORT_SYMBOL_GPL(inet_diag_dump_one_icsk);
416 
417 static int inet_diag_get_exact(struct sk_buff *in_skb,
418 			       const struct nlmsghdr *nlh,
419 			       const struct inet_diag_req_v2 *req)
420 {
421 	const struct inet_diag_handler *handler;
422 	int err;
423 
424 	handler = inet_diag_lock_handler(req->sdiag_protocol);
425 	if (IS_ERR(handler))
426 		err = PTR_ERR(handler);
427 	else
428 		err = handler->dump_one(in_skb, nlh, req);
429 	inet_diag_unlock_handler(handler);
430 
431 	return err;
432 }
433 
434 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
435 {
436 	int words = bits >> 5;
437 
438 	bits &= 0x1f;
439 
440 	if (words) {
441 		if (memcmp(a1, a2, words << 2))
442 			return 0;
443 	}
444 	if (bits) {
445 		__be32 w1, w2;
446 		__be32 mask;
447 
448 		w1 = a1[words];
449 		w2 = a2[words];
450 
451 		mask = htonl((0xffffffff) << (32 - bits));
452 
453 		if ((w1 ^ w2) & mask)
454 			return 0;
455 	}
456 
457 	return 1;
458 }
459 
460 static int inet_diag_bc_run(const struct nlattr *_bc,
461 			    const struct inet_diag_entry *entry)
462 {
463 	const void *bc = nla_data(_bc);
464 	int len = nla_len(_bc);
465 
466 	while (len > 0) {
467 		int yes = 1;
468 		const struct inet_diag_bc_op *op = bc;
469 
470 		switch (op->code) {
471 		case INET_DIAG_BC_NOP:
472 			break;
473 		case INET_DIAG_BC_JMP:
474 			yes = 0;
475 			break;
476 		case INET_DIAG_BC_S_GE:
477 			yes = entry->sport >= op[1].no;
478 			break;
479 		case INET_DIAG_BC_S_LE:
480 			yes = entry->sport <= op[1].no;
481 			break;
482 		case INET_DIAG_BC_D_GE:
483 			yes = entry->dport >= op[1].no;
484 			break;
485 		case INET_DIAG_BC_D_LE:
486 			yes = entry->dport <= op[1].no;
487 			break;
488 		case INET_DIAG_BC_AUTO:
489 			yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
490 			break;
491 		case INET_DIAG_BC_S_COND:
492 		case INET_DIAG_BC_D_COND: {
493 			const struct inet_diag_hostcond *cond;
494 			const __be32 *addr;
495 
496 			cond = (const struct inet_diag_hostcond *)(op + 1);
497 			if (cond->port != -1 &&
498 			    cond->port != (op->code == INET_DIAG_BC_S_COND ?
499 					     entry->sport : entry->dport)) {
500 				yes = 0;
501 				break;
502 			}
503 
504 			if (op->code == INET_DIAG_BC_S_COND)
505 				addr = entry->saddr;
506 			else
507 				addr = entry->daddr;
508 
509 			if (cond->family != AF_UNSPEC &&
510 			    cond->family != entry->family) {
511 				if (entry->family == AF_INET6 &&
512 				    cond->family == AF_INET) {
513 					if (addr[0] == 0 && addr[1] == 0 &&
514 					    addr[2] == htonl(0xffff) &&
515 					    bitstring_match(addr + 3,
516 							    cond->addr,
517 							    cond->prefix_len))
518 						break;
519 				}
520 				yes = 0;
521 				break;
522 			}
523 
524 			if (cond->prefix_len == 0)
525 				break;
526 			if (bitstring_match(addr, cond->addr,
527 					    cond->prefix_len))
528 				break;
529 			yes = 0;
530 			break;
531 		}
532 		}
533 
534 		if (yes) {
535 			len -= op->yes;
536 			bc += op->yes;
537 		} else {
538 			len -= op->no;
539 			bc += op->no;
540 		}
541 	}
542 	return len == 0;
543 }
544 
545 /* This helper is available for all sockets (ESTABLISH, TIMEWAIT, SYN_RECV)
546  */
547 static void entry_fill_addrs(struct inet_diag_entry *entry,
548 			     const struct sock *sk)
549 {
550 #if IS_ENABLED(CONFIG_IPV6)
551 	if (sk->sk_family == AF_INET6) {
552 		entry->saddr = sk->sk_v6_rcv_saddr.s6_addr32;
553 		entry->daddr = sk->sk_v6_daddr.s6_addr32;
554 	} else
555 #endif
556 	{
557 		entry->saddr = &sk->sk_rcv_saddr;
558 		entry->daddr = &sk->sk_daddr;
559 	}
560 }
561 
562 int inet_diag_bc_sk(const struct nlattr *bc, struct sock *sk)
563 {
564 	struct inet_sock *inet = inet_sk(sk);
565 	struct inet_diag_entry entry;
566 
567 	if (!bc)
568 		return 1;
569 
570 	entry.family = sk->sk_family;
571 	entry_fill_addrs(&entry, sk);
572 	entry.sport = inet->inet_num;
573 	entry.dport = ntohs(inet->inet_dport);
574 	entry.userlocks = sk_fullsock(sk) ? sk->sk_userlocks : 0;
575 
576 	return inet_diag_bc_run(bc, &entry);
577 }
578 EXPORT_SYMBOL_GPL(inet_diag_bc_sk);
579 
580 static int valid_cc(const void *bc, int len, int cc)
581 {
582 	while (len >= 0) {
583 		const struct inet_diag_bc_op *op = bc;
584 
585 		if (cc > len)
586 			return 0;
587 		if (cc == len)
588 			return 1;
589 		if (op->yes < 4 || op->yes & 3)
590 			return 0;
591 		len -= op->yes;
592 		bc  += op->yes;
593 	}
594 	return 0;
595 }
596 
597 /* Validate an inet_diag_hostcond. */
598 static bool valid_hostcond(const struct inet_diag_bc_op *op, int len,
599 			   int *min_len)
600 {
601 	struct inet_diag_hostcond *cond;
602 	int addr_len;
603 
604 	/* Check hostcond space. */
605 	*min_len += sizeof(struct inet_diag_hostcond);
606 	if (len < *min_len)
607 		return false;
608 	cond = (struct inet_diag_hostcond *)(op + 1);
609 
610 	/* Check address family and address length. */
611 	switch (cond->family) {
612 	case AF_UNSPEC:
613 		addr_len = 0;
614 		break;
615 	case AF_INET:
616 		addr_len = sizeof(struct in_addr);
617 		break;
618 	case AF_INET6:
619 		addr_len = sizeof(struct in6_addr);
620 		break;
621 	default:
622 		return false;
623 	}
624 	*min_len += addr_len;
625 	if (len < *min_len)
626 		return false;
627 
628 	/* Check prefix length (in bits) vs address length (in bytes). */
629 	if (cond->prefix_len > 8 * addr_len)
630 		return false;
631 
632 	return true;
633 }
634 
635 /* Validate a port comparison operator. */
636 static bool valid_port_comparison(const struct inet_diag_bc_op *op,
637 				  int len, int *min_len)
638 {
639 	/* Port comparisons put the port in a follow-on inet_diag_bc_op. */
640 	*min_len += sizeof(struct inet_diag_bc_op);
641 	if (len < *min_len)
642 		return false;
643 	return true;
644 }
645 
646 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
647 {
648 	const void *bc = bytecode;
649 	int  len = bytecode_len;
650 
651 	while (len > 0) {
652 		int min_len = sizeof(struct inet_diag_bc_op);
653 		const struct inet_diag_bc_op *op = bc;
654 
655 		switch (op->code) {
656 		case INET_DIAG_BC_S_COND:
657 		case INET_DIAG_BC_D_COND:
658 			if (!valid_hostcond(bc, len, &min_len))
659 				return -EINVAL;
660 			break;
661 		case INET_DIAG_BC_S_GE:
662 		case INET_DIAG_BC_S_LE:
663 		case INET_DIAG_BC_D_GE:
664 		case INET_DIAG_BC_D_LE:
665 			if (!valid_port_comparison(bc, len, &min_len))
666 				return -EINVAL;
667 			break;
668 		case INET_DIAG_BC_AUTO:
669 		case INET_DIAG_BC_JMP:
670 		case INET_DIAG_BC_NOP:
671 			break;
672 		default:
673 			return -EINVAL;
674 		}
675 
676 		if (op->code != INET_DIAG_BC_NOP) {
677 			if (op->no < min_len || op->no > len + 4 || op->no & 3)
678 				return -EINVAL;
679 			if (op->no < len &&
680 			    !valid_cc(bytecode, bytecode_len, len - op->no))
681 				return -EINVAL;
682 		}
683 
684 		if (op->yes < min_len || op->yes > len + 4 || op->yes & 3)
685 			return -EINVAL;
686 		bc  += op->yes;
687 		len -= op->yes;
688 	}
689 	return len == 0 ? 0 : -EINVAL;
690 }
691 
692 static int inet_csk_diag_dump(struct sock *sk,
693 			      struct sk_buff *skb,
694 			      struct netlink_callback *cb,
695 			      const struct inet_diag_req_v2 *r,
696 			      const struct nlattr *bc)
697 {
698 	if (!inet_diag_bc_sk(bc, sk))
699 		return 0;
700 
701 	return inet_csk_diag_fill(sk, skb, r,
702 				  sk_user_ns(NETLINK_CB(cb->skb).sk),
703 				  NETLINK_CB(cb->skb).portid,
704 				  cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
705 }
706 
707 static void twsk_build_assert(void)
708 {
709 	BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_family) !=
710 		     offsetof(struct sock, sk_family));
711 
712 	BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_num) !=
713 		     offsetof(struct inet_sock, inet_num));
714 
715 	BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_dport) !=
716 		     offsetof(struct inet_sock, inet_dport));
717 
718 	BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_rcv_saddr) !=
719 		     offsetof(struct inet_sock, inet_rcv_saddr));
720 
721 	BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_daddr) !=
722 		     offsetof(struct inet_sock, inet_daddr));
723 
724 #if IS_ENABLED(CONFIG_IPV6)
725 	BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_v6_rcv_saddr) !=
726 		     offsetof(struct sock, sk_v6_rcv_saddr));
727 
728 	BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_v6_daddr) !=
729 		     offsetof(struct sock, sk_v6_daddr));
730 #endif
731 }
732 
733 void inet_diag_dump_icsk(struct inet_hashinfo *hashinfo, struct sk_buff *skb,
734 			 struct netlink_callback *cb,
735 			 const struct inet_diag_req_v2 *r, struct nlattr *bc)
736 {
737 	struct net *net = sock_net(skb->sk);
738 	int i, num, s_i, s_num;
739 	u32 idiag_states = r->idiag_states;
740 
741 	if (idiag_states & TCPF_SYN_RECV)
742 		idiag_states |= TCPF_NEW_SYN_RECV;
743 	s_i = cb->args[1];
744 	s_num = num = cb->args[2];
745 
746 	if (cb->args[0] == 0) {
747 		if (!(idiag_states & TCPF_LISTEN))
748 			goto skip_listen_ht;
749 
750 		for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
751 			struct inet_listen_hashbucket *ilb;
752 			struct hlist_nulls_node *node;
753 			struct sock *sk;
754 
755 			num = 0;
756 			ilb = &hashinfo->listening_hash[i];
757 			spin_lock_bh(&ilb->lock);
758 			sk_nulls_for_each(sk, node, &ilb->head) {
759 				struct inet_sock *inet = inet_sk(sk);
760 
761 				if (!net_eq(sock_net(sk), net))
762 					continue;
763 
764 				if (num < s_num) {
765 					num++;
766 					continue;
767 				}
768 
769 				if (r->sdiag_family != AF_UNSPEC &&
770 				    sk->sk_family != r->sdiag_family)
771 					goto next_listen;
772 
773 				if (r->id.idiag_sport != inet->inet_sport &&
774 				    r->id.idiag_sport)
775 					goto next_listen;
776 
777 				if (r->id.idiag_dport ||
778 				    cb->args[3] > 0)
779 					goto next_listen;
780 
781 				if (inet_csk_diag_dump(sk, skb, cb, r, bc) < 0) {
782 					spin_unlock_bh(&ilb->lock);
783 					goto done;
784 				}
785 
786 next_listen:
787 				cb->args[3] = 0;
788 				cb->args[4] = 0;
789 				++num;
790 			}
791 			spin_unlock_bh(&ilb->lock);
792 
793 			s_num = 0;
794 			cb->args[3] = 0;
795 			cb->args[4] = 0;
796 		}
797 skip_listen_ht:
798 		cb->args[0] = 1;
799 		s_i = num = s_num = 0;
800 	}
801 
802 	if (!(idiag_states & ~TCPF_LISTEN))
803 		goto out;
804 
805 	for (i = s_i; i <= hashinfo->ehash_mask; i++) {
806 		struct inet_ehash_bucket *head = &hashinfo->ehash[i];
807 		spinlock_t *lock = inet_ehash_lockp(hashinfo, i);
808 		struct hlist_nulls_node *node;
809 		struct sock *sk;
810 
811 		num = 0;
812 
813 		if (hlist_nulls_empty(&head->chain))
814 			continue;
815 
816 		if (i > s_i)
817 			s_num = 0;
818 
819 		spin_lock_bh(lock);
820 		sk_nulls_for_each(sk, node, &head->chain) {
821 			int state, res;
822 
823 			if (!net_eq(sock_net(sk), net))
824 				continue;
825 			if (num < s_num)
826 				goto next_normal;
827 			state = (sk->sk_state == TCP_TIME_WAIT) ?
828 				inet_twsk(sk)->tw_substate : sk->sk_state;
829 			if (!(idiag_states & (1 << state)))
830 				goto next_normal;
831 			if (r->sdiag_family != AF_UNSPEC &&
832 			    sk->sk_family != r->sdiag_family)
833 				goto next_normal;
834 			if (r->id.idiag_sport != htons(sk->sk_num) &&
835 			    r->id.idiag_sport)
836 				goto next_normal;
837 			if (r->id.idiag_dport != sk->sk_dport &&
838 			    r->id.idiag_dport)
839 				goto next_normal;
840 			twsk_build_assert();
841 
842 			if (!inet_diag_bc_sk(bc, sk))
843 				goto next_normal;
844 
845 			res = sk_diag_fill(sk, skb, r,
846 					   sk_user_ns(NETLINK_CB(cb->skb).sk),
847 					   NETLINK_CB(cb->skb).portid,
848 					   cb->nlh->nlmsg_seq, NLM_F_MULTI,
849 					   cb->nlh);
850 			if (res < 0) {
851 				spin_unlock_bh(lock);
852 				goto done;
853 			}
854 next_normal:
855 			++num;
856 		}
857 
858 		spin_unlock_bh(lock);
859 	}
860 
861 done:
862 	cb->args[1] = i;
863 	cb->args[2] = num;
864 out:
865 	;
866 }
867 EXPORT_SYMBOL_GPL(inet_diag_dump_icsk);
868 
869 static int __inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb,
870 			    const struct inet_diag_req_v2 *r,
871 			    struct nlattr *bc)
872 {
873 	const struct inet_diag_handler *handler;
874 	int err = 0;
875 
876 	handler = inet_diag_lock_handler(r->sdiag_protocol);
877 	if (!IS_ERR(handler))
878 		handler->dump(skb, cb, r, bc);
879 	else
880 		err = PTR_ERR(handler);
881 	inet_diag_unlock_handler(handler);
882 
883 	return err ? : skb->len;
884 }
885 
886 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
887 {
888 	int hdrlen = sizeof(struct inet_diag_req_v2);
889 	struct nlattr *bc = NULL;
890 
891 	if (nlmsg_attrlen(cb->nlh, hdrlen))
892 		bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
893 
894 	return __inet_diag_dump(skb, cb, nlmsg_data(cb->nlh), bc);
895 }
896 
897 static int inet_diag_type2proto(int type)
898 {
899 	switch (type) {
900 	case TCPDIAG_GETSOCK:
901 		return IPPROTO_TCP;
902 	case DCCPDIAG_GETSOCK:
903 		return IPPROTO_DCCP;
904 	default:
905 		return 0;
906 	}
907 }
908 
909 static int inet_diag_dump_compat(struct sk_buff *skb,
910 				 struct netlink_callback *cb)
911 {
912 	struct inet_diag_req *rc = nlmsg_data(cb->nlh);
913 	int hdrlen = sizeof(struct inet_diag_req);
914 	struct inet_diag_req_v2 req;
915 	struct nlattr *bc = NULL;
916 
917 	req.sdiag_family = AF_UNSPEC; /* compatibility */
918 	req.sdiag_protocol = inet_diag_type2proto(cb->nlh->nlmsg_type);
919 	req.idiag_ext = rc->idiag_ext;
920 	req.idiag_states = rc->idiag_states;
921 	req.id = rc->id;
922 
923 	if (nlmsg_attrlen(cb->nlh, hdrlen))
924 		bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
925 
926 	return __inet_diag_dump(skb, cb, &req, bc);
927 }
928 
929 static int inet_diag_get_exact_compat(struct sk_buff *in_skb,
930 				      const struct nlmsghdr *nlh)
931 {
932 	struct inet_diag_req *rc = nlmsg_data(nlh);
933 	struct inet_diag_req_v2 req;
934 
935 	req.sdiag_family = rc->idiag_family;
936 	req.sdiag_protocol = inet_diag_type2proto(nlh->nlmsg_type);
937 	req.idiag_ext = rc->idiag_ext;
938 	req.idiag_states = rc->idiag_states;
939 	req.id = rc->id;
940 
941 	return inet_diag_get_exact(in_skb, nlh, &req);
942 }
943 
944 static int inet_diag_rcv_msg_compat(struct sk_buff *skb, struct nlmsghdr *nlh)
945 {
946 	int hdrlen = sizeof(struct inet_diag_req);
947 	struct net *net = sock_net(skb->sk);
948 
949 	if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
950 	    nlmsg_len(nlh) < hdrlen)
951 		return -EINVAL;
952 
953 	if (nlh->nlmsg_flags & NLM_F_DUMP) {
954 		if (nlmsg_attrlen(nlh, hdrlen)) {
955 			struct nlattr *attr;
956 
957 			attr = nlmsg_find_attr(nlh, hdrlen,
958 					       INET_DIAG_REQ_BYTECODE);
959 			if (!attr ||
960 			    nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
961 			    inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
962 				return -EINVAL;
963 		}
964 		{
965 			struct netlink_dump_control c = {
966 				.dump = inet_diag_dump_compat,
967 			};
968 			return netlink_dump_start(net->diag_nlsk, skb, nlh, &c);
969 		}
970 	}
971 
972 	return inet_diag_get_exact_compat(skb, nlh);
973 }
974 
975 static int inet_diag_handler_dump(struct sk_buff *skb, struct nlmsghdr *h)
976 {
977 	int hdrlen = sizeof(struct inet_diag_req_v2);
978 	struct net *net = sock_net(skb->sk);
979 
980 	if (nlmsg_len(h) < hdrlen)
981 		return -EINVAL;
982 
983 	if (h->nlmsg_flags & NLM_F_DUMP) {
984 		if (nlmsg_attrlen(h, hdrlen)) {
985 			struct nlattr *attr;
986 
987 			attr = nlmsg_find_attr(h, hdrlen,
988 					       INET_DIAG_REQ_BYTECODE);
989 			if (!attr ||
990 			    nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
991 			    inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
992 				return -EINVAL;
993 		}
994 		{
995 			struct netlink_dump_control c = {
996 				.dump = inet_diag_dump,
997 			};
998 			return netlink_dump_start(net->diag_nlsk, skb, h, &c);
999 		}
1000 	}
1001 
1002 	return inet_diag_get_exact(skb, h, nlmsg_data(h));
1003 }
1004 
1005 static
1006 int inet_diag_handler_get_info(struct sk_buff *skb, struct sock *sk)
1007 {
1008 	const struct inet_diag_handler *handler;
1009 	struct nlmsghdr *nlh;
1010 	struct nlattr *attr;
1011 	struct inet_diag_msg *r;
1012 	void *info = NULL;
1013 	int err = 0;
1014 
1015 	nlh = nlmsg_put(skb, 0, 0, SOCK_DIAG_BY_FAMILY, sizeof(*r), 0);
1016 	if (!nlh)
1017 		return -ENOMEM;
1018 
1019 	r = nlmsg_data(nlh);
1020 	memset(r, 0, sizeof(*r));
1021 	inet_diag_msg_common_fill(r, sk);
1022 	if (sk->sk_type == SOCK_DGRAM || sk->sk_type == SOCK_STREAM)
1023 		r->id.idiag_sport = inet_sk(sk)->inet_sport;
1024 	r->idiag_state = sk->sk_state;
1025 
1026 	if ((err = nla_put_u8(skb, INET_DIAG_PROTOCOL, sk->sk_protocol))) {
1027 		nlmsg_cancel(skb, nlh);
1028 		return err;
1029 	}
1030 
1031 	handler = inet_diag_lock_handler(sk->sk_protocol);
1032 	if (IS_ERR(handler)) {
1033 		inet_diag_unlock_handler(handler);
1034 		nlmsg_cancel(skb, nlh);
1035 		return PTR_ERR(handler);
1036 	}
1037 
1038 	attr = handler->idiag_info_size
1039 		? nla_reserve(skb, INET_DIAG_INFO, handler->idiag_info_size)
1040 		: NULL;
1041 	if (attr)
1042 		info = nla_data(attr);
1043 
1044 	handler->idiag_get_info(sk, r, info);
1045 	inet_diag_unlock_handler(handler);
1046 
1047 	nlmsg_end(skb, nlh);
1048 	return 0;
1049 }
1050 
1051 static const struct sock_diag_handler inet_diag_handler = {
1052 	.family = AF_INET,
1053 	.dump = inet_diag_handler_dump,
1054 	.get_info = inet_diag_handler_get_info,
1055 };
1056 
1057 static const struct sock_diag_handler inet6_diag_handler = {
1058 	.family = AF_INET6,
1059 	.dump = inet_diag_handler_dump,
1060 	.get_info = inet_diag_handler_get_info,
1061 };
1062 
1063 int inet_diag_register(const struct inet_diag_handler *h)
1064 {
1065 	const __u16 type = h->idiag_type;
1066 	int err = -EINVAL;
1067 
1068 	if (type >= IPPROTO_MAX)
1069 		goto out;
1070 
1071 	mutex_lock(&inet_diag_table_mutex);
1072 	err = -EEXIST;
1073 	if (!inet_diag_table[type]) {
1074 		inet_diag_table[type] = h;
1075 		err = 0;
1076 	}
1077 	mutex_unlock(&inet_diag_table_mutex);
1078 out:
1079 	return err;
1080 }
1081 EXPORT_SYMBOL_GPL(inet_diag_register);
1082 
1083 void inet_diag_unregister(const struct inet_diag_handler *h)
1084 {
1085 	const __u16 type = h->idiag_type;
1086 
1087 	if (type >= IPPROTO_MAX)
1088 		return;
1089 
1090 	mutex_lock(&inet_diag_table_mutex);
1091 	inet_diag_table[type] = NULL;
1092 	mutex_unlock(&inet_diag_table_mutex);
1093 }
1094 EXPORT_SYMBOL_GPL(inet_diag_unregister);
1095 
1096 static int __init inet_diag_init(void)
1097 {
1098 	const int inet_diag_table_size = (IPPROTO_MAX *
1099 					  sizeof(struct inet_diag_handler *));
1100 	int err = -ENOMEM;
1101 
1102 	inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
1103 	if (!inet_diag_table)
1104 		goto out;
1105 
1106 	err = sock_diag_register(&inet_diag_handler);
1107 	if (err)
1108 		goto out_free_nl;
1109 
1110 	err = sock_diag_register(&inet6_diag_handler);
1111 	if (err)
1112 		goto out_free_inet;
1113 
1114 	sock_diag_register_inet_compat(inet_diag_rcv_msg_compat);
1115 out:
1116 	return err;
1117 
1118 out_free_inet:
1119 	sock_diag_unregister(&inet_diag_handler);
1120 out_free_nl:
1121 	kfree(inet_diag_table);
1122 	goto out;
1123 }
1124 
1125 static void __exit inet_diag_exit(void)
1126 {
1127 	sock_diag_unregister(&inet6_diag_handler);
1128 	sock_diag_unregister(&inet_diag_handler);
1129 	sock_diag_unregister_inet_compat(inet_diag_rcv_msg_compat);
1130 	kfree(inet_diag_table);
1131 }
1132 
1133 module_init(inet_diag_init);
1134 module_exit(inet_diag_exit);
1135 MODULE_LICENSE("GPL");
1136 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 2 /* AF_INET */);
1137 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 10 /* AF_INET6 */);
1138