?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
ext4: Optimize ext4 DIO overwrites
[linux/fpc-iii.git] / net / ipv4 / inet_diag.c
blobf11e997e517b6652479acd892ffea6cdeb940e33
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * inet_diag.c Module for monitoring INET transport protocols sockets.
4 *
5 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
6 */
7
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/types.h>
11 #include <linux/fcntl.h>
12 #include <linux/random.h>
13 #include <linux/slab.h>
14 #include <linux/cache.h>
15 #include <linux/init.h>
16 #include <linux/time.h>
17
18 #include <net/icmp.h>
19 #include <net/tcp.h>
20 #include <net/ipv6.h>
21 #include <net/inet_common.h>
22 #include <net/inet_connection_sock.h>
23 #include <net/inet_hashtables.h>
24 #include <net/inet_timewait_sock.h>
25 #include <net/inet6_hashtables.h>
26 #include <net/netlink.h>
27
28 #include <linux/inet.h>
29 #include <linux/stddef.h>
30
31 #include <linux/inet_diag.h>
32 #include <linux/sock_diag.h>
33
34 static const struct inet_diag_handler **inet_diag_table;
35
36 struct inet_diag_entry {
37 const __be32 *saddr;
38 const __be32 *daddr;
39 u16 sport;
40 u16 dport;
41 u16 family;
42 u16 userlocks;
43 u32 ifindex;
44 u32 mark;
45 };
46
47 static DEFINE_MUTEX(inet_diag_table_mutex);
48
49 static const struct inet_diag_handler *inet_diag_lock_handler(int proto)
50 {
51 if (!inet_diag_table[proto])
52 sock_load_diag_module(AF_INET, proto);
53
54 mutex_lock(&inet_diag_table_mutex);
55 if (!inet_diag_table[proto])
56 return ERR_PTR(-ENOENT);
57
58 return inet_diag_table[proto];
59 }
60
61 static void inet_diag_unlock_handler(const struct inet_diag_handler *handler)
62 {
63 mutex_unlock(&inet_diag_table_mutex);
64 }
65
66 void inet_diag_msg_common_fill(struct inet_diag_msg *r, struct sock *sk)
67 {
68 r->idiag_family = sk->sk_family;
69
70 r->id.idiag_sport = htons(sk->sk_num);
71 r->id.idiag_dport = sk->sk_dport;
72 r->id.idiag_if = sk->sk_bound_dev_if;
73 sock_diag_save_cookie(sk, r->id.idiag_cookie);
74
75 #if IS_ENABLED(CONFIG_IPV6)
76 if (sk->sk_family == AF_INET6) {
77 *(struct in6_addr *)r->id.idiag_src = sk->sk_v6_rcv_saddr;
78 *(struct in6_addr *)r->id.idiag_dst = sk->sk_v6_daddr;
79 } else
80 #endif
81 {
82 memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
83 memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
84
85 r->id.idiag_src[0] = sk->sk_rcv_saddr;
86 r->id.idiag_dst[0] = sk->sk_daddr;
87 }
88 }
89 EXPORT_SYMBOL_GPL(inet_diag_msg_common_fill);
90
91 static size_t inet_sk_attr_size(struct sock *sk,
92 const struct inet_diag_req_v2 *req,
93 bool net_admin)
94 {
95 const struct inet_diag_handler *handler;
96 size_t aux = 0;
97
98 handler = inet_diag_table[req->sdiag_protocol];
99 if (handler && handler->idiag_get_aux_size)
100 aux = handler->idiag_get_aux_size(sk, net_admin);
101
102 return nla_total_size(sizeof(struct tcp_info))
103 + nla_total_size(1) /* INET_DIAG_SHUTDOWN */
104 + nla_total_size(1) /* INET_DIAG_TOS */
105 + nla_total_size(1) /* INET_DIAG_TCLASS */
106 + nla_total_size(4) /* INET_DIAG_MARK */
107 + nla_total_size(4) /* INET_DIAG_CLASS_ID */
108 + nla_total_size(sizeof(struct inet_diag_meminfo))
109 + nla_total_size(sizeof(struct inet_diag_msg))
110 + nla_total_size(SK_MEMINFO_VARS * sizeof(u32))
111 + nla_total_size(TCP_CA_NAME_MAX)
112 + nla_total_size(sizeof(struct tcpvegas_info))
113 + aux
114 + 64;
115 }
116
117 int inet_diag_msg_attrs_fill(struct sock *sk, struct sk_buff *skb,
118 struct inet_diag_msg *r, int ext,
119 struct user_namespace *user_ns,
120 bool net_admin)
121 {
122 const struct inet_sock *inet = inet_sk(sk);
123
124 if (nla_put_u8(skb, INET_DIAG_SHUTDOWN, sk->sk_shutdown))
125 goto errout;
126
127 /* IPv6 dual-stack sockets use inet->tos for IPv4 connections,
128 * hence this needs to be included regardless of socket family.
129 */
130 if (ext & (1 << (INET_DIAG_TOS - 1)))
131 if (nla_put_u8(skb, INET_DIAG_TOS, inet->tos) < 0)
132 goto errout;
133
134 #if IS_ENABLED(CONFIG_IPV6)
135 if (r->idiag_family == AF_INET6) {
136 if (ext & (1 << (INET_DIAG_TCLASS - 1)))
137 if (nla_put_u8(skb, INET_DIAG_TCLASS,
138 inet6_sk(sk)->tclass) < 0)
139 goto errout;
140
141 if (((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) &&
142 nla_put_u8(skb, INET_DIAG_SKV6ONLY, ipv6_only_sock(sk)))
143 goto errout;
144 }
145 #endif
146
147 if (net_admin && nla_put_u32(skb, INET_DIAG_MARK, sk->sk_mark))
148 goto errout;
149
150 r->idiag_uid = from_kuid_munged(user_ns, sock_i_uid(sk));
151 r->idiag_inode = sock_i_ino(sk);
152
153 return 0;
154 errout:
155 return 1;
156 }
157 EXPORT_SYMBOL_GPL(inet_diag_msg_attrs_fill);
158
159 int inet_sk_diag_fill(struct sock *sk, struct inet_connection_sock *icsk,
160 struct sk_buff *skb, const struct inet_diag_req_v2 *req,
161 struct user_namespace *user_ns,
162 u32 portid, u32 seq, u16 nlmsg_flags,
163 const struct nlmsghdr *unlh,
164 bool net_admin)
165 {
166 const struct tcp_congestion_ops *ca_ops;
167 const struct inet_diag_handler *handler;
168 int ext = req->idiag_ext;
169 struct inet_diag_msg *r;
170 struct nlmsghdr *nlh;
171 struct nlattr *attr;
172 void *info = NULL;
173
174 handler = inet_diag_table[req->sdiag_protocol];
175 BUG_ON(!handler);
176
177 nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
178 nlmsg_flags);
179 if (!nlh)
180 return -EMSGSIZE;
181
182 r = nlmsg_data(nlh);
183 BUG_ON(!sk_fullsock(sk));
184
185 inet_diag_msg_common_fill(r, sk);
186 r->idiag_state = sk->sk_state;
187 r->idiag_timer = 0;
188 r->idiag_retrans = 0;
189
190 if (inet_diag_msg_attrs_fill(sk, skb, r, ext, user_ns, net_admin))
191 goto errout;
192
193 if (ext & (1 << (INET_DIAG_MEMINFO - 1))) {
194 struct inet_diag_meminfo minfo = {
195 .idiag_rmem = sk_rmem_alloc_get(sk),
196 .idiag_wmem = READ_ONCE(sk->sk_wmem_queued),
197 .idiag_fmem = sk->sk_forward_alloc,
198 .idiag_tmem = sk_wmem_alloc_get(sk),
199 };
200
201 if (nla_put(skb, INET_DIAG_MEMINFO, sizeof(minfo), &minfo) < 0)
202 goto errout;
203 }
204
205 if (ext & (1 << (INET_DIAG_SKMEMINFO - 1)))
206 if (sock_diag_put_meminfo(sk, skb, INET_DIAG_SKMEMINFO))
207 goto errout;
208
209 /*
210 * RAW sockets might have user-defined protocols assigned,
211 * so report the one supplied on socket creation.
212 */
213 if (sk->sk_type == SOCK_RAW) {
214 if (nla_put_u8(skb, INET_DIAG_PROTOCOL, sk->sk_protocol))
215 goto errout;
216 }
217
218 if (!icsk) {
219 handler->idiag_get_info(sk, r, NULL);
220 goto out;
221 }
222
223 if (icsk->icsk_pending == ICSK_TIME_RETRANS ||
224 icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT ||
225 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
226 r->idiag_timer = 1;
227 r->idiag_retrans = icsk->icsk_retransmits;
228 r->idiag_expires =
229 jiffies_delta_to_msecs(icsk->icsk_timeout - jiffies);
230 } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
231 r->idiag_timer = 4;
232 r->idiag_retrans = icsk->icsk_probes_out;
233 r->idiag_expires =
234 jiffies_delta_to_msecs(icsk->icsk_timeout - jiffies);
235 } else if (timer_pending(&sk->sk_timer)) {
236 r->idiag_timer = 2;
237 r->idiag_retrans = icsk->icsk_probes_out;
238 r->idiag_expires =
239 jiffies_delta_to_msecs(sk->sk_timer.expires - jiffies);
240 } else {
241 r->idiag_timer = 0;
242 r->idiag_expires = 0;
243 }
244
245 if ((ext & (1 << (INET_DIAG_INFO - 1))) && handler->idiag_info_size) {
246 attr = nla_reserve_64bit(skb, INET_DIAG_INFO,
247 handler->idiag_info_size,
248 INET_DIAG_PAD);
249 if (!attr)
250 goto errout;
251
252 info = nla_data(attr);
253 }
254
255 if (ext & (1 << (INET_DIAG_CONG - 1))) {
256 int err = 0;
257
258 rcu_read_lock();
259 ca_ops = READ_ONCE(icsk->icsk_ca_ops);
260 if (ca_ops)
261 err = nla_put_string(skb, INET_DIAG_CONG, ca_ops->name);
262 rcu_read_unlock();
263 if (err < 0)
264 goto errout;
265 }
266
267 handler->idiag_get_info(sk, r, info);
268
269 if (ext & (1 << (INET_DIAG_INFO - 1)) && handler->idiag_get_aux)
270 if (handler->idiag_get_aux(sk, net_admin, skb) < 0)
271 goto errout;
272
273 if (sk->sk_state < TCP_TIME_WAIT) {
274 union tcp_cc_info info;
275 size_t sz = 0;
276 int attr;
277
278 rcu_read_lock();
279 ca_ops = READ_ONCE(icsk->icsk_ca_ops);
280 if (ca_ops && ca_ops->get_info)
281 sz = ca_ops->get_info(sk, ext, &attr, &info);
282 rcu_read_unlock();
283 if (sz && nla_put(skb, attr, sz, &info) < 0)
284 goto errout;
285 }
286
287 if (ext & (1 << (INET_DIAG_CLASS_ID - 1)) ||
288 ext & (1 << (INET_DIAG_TCLASS - 1))) {
289 u32 classid = 0;
290
291 #ifdef CONFIG_SOCK_CGROUP_DATA
292 classid = sock_cgroup_classid(&sk->sk_cgrp_data);
293 #endif
294 /* Fallback to socket priority if class id isn't set.
295 * Classful qdiscs use it as direct reference to class.
296 * For cgroup2 classid is always zero.
297 */
298 if (!classid)
299 classid = sk->sk_priority;
300
301 if (nla_put_u32(skb, INET_DIAG_CLASS_ID, classid))
302 goto errout;
303 }
304
305 out:
306 nlmsg_end(skb, nlh);
307 return 0;
308
309 errout:
310 nlmsg_cancel(skb, nlh);
311 return -EMSGSIZE;
312 }
313 EXPORT_SYMBOL_GPL(inet_sk_diag_fill);
314
315 static int inet_csk_diag_fill(struct sock *sk,
316 struct sk_buff *skb,
317 const struct inet_diag_req_v2 *req,
318 struct user_namespace *user_ns,
319 u32 portid, u32 seq, u16 nlmsg_flags,
320 const struct nlmsghdr *unlh,
321 bool net_admin)
322 {
323 return inet_sk_diag_fill(sk, inet_csk(sk), skb, req, user_ns,
324 portid, seq, nlmsg_flags, unlh, net_admin);
325 }
326
327 static int inet_twsk_diag_fill(struct sock *sk,
328 struct sk_buff *skb,
329 u32 portid, u32 seq, u16 nlmsg_flags,
330 const struct nlmsghdr *unlh)
331 {
332 struct inet_timewait_sock *tw = inet_twsk(sk);
333 struct inet_diag_msg *r;
334 struct nlmsghdr *nlh;
335 long tmo;
336
337 nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
338 nlmsg_flags);
339 if (!nlh)
340 return -EMSGSIZE;
341
342 r = nlmsg_data(nlh);
343 BUG_ON(tw->tw_state != TCP_TIME_WAIT);
344
345 inet_diag_msg_common_fill(r, sk);
346 r->idiag_retrans = 0;
347
348 r->idiag_state = tw->tw_substate;
349 r->idiag_timer = 3;
350 tmo = tw->tw_timer.expires - jiffies;
351 r->idiag_expires = jiffies_delta_to_msecs(tmo);
352 r->idiag_rqueue = 0;
353 r->idiag_wqueue = 0;
354 r->idiag_uid = 0;
355 r->idiag_inode = 0;
356
357 nlmsg_end(skb, nlh);
358 return 0;
359 }
360
361 static int inet_req_diag_fill(struct sock *sk, struct sk_buff *skb,
362 u32 portid, u32 seq, u16 nlmsg_flags,
363 const struct nlmsghdr *unlh, bool net_admin)
364 {
365 struct request_sock *reqsk = inet_reqsk(sk);
366 struct inet_diag_msg *r;
367 struct nlmsghdr *nlh;
368 long tmo;
369
370 nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
371 nlmsg_flags);
372 if (!nlh)
373 return -EMSGSIZE;
374
375 r = nlmsg_data(nlh);
376 inet_diag_msg_common_fill(r, sk);
377 r->idiag_state = TCP_SYN_RECV;
378 r->idiag_timer = 1;
379 r->idiag_retrans = reqsk->num_retrans;
380
381 BUILD_BUG_ON(offsetof(struct inet_request_sock, ir_cookie) !=
382 offsetof(struct sock, sk_cookie));
383
384 tmo = inet_reqsk(sk)->rsk_timer.expires - jiffies;
385 r->idiag_expires = jiffies_delta_to_msecs(tmo);
386 r->idiag_rqueue = 0;
387 r->idiag_wqueue = 0;
388 r->idiag_uid = 0;
389 r->idiag_inode = 0;
390
391 if (net_admin && nla_put_u32(skb, INET_DIAG_MARK,
392 inet_rsk(reqsk)->ir_mark))
393 return -EMSGSIZE;
394
395 nlmsg_end(skb, nlh);
396 return 0;
397 }
398
399 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
400 const struct inet_diag_req_v2 *r,
401 struct user_namespace *user_ns,
402 u32 portid, u32 seq, u16 nlmsg_flags,
403 const struct nlmsghdr *unlh, bool net_admin)
404 {
405 if (sk->sk_state == TCP_TIME_WAIT)
406 return inet_twsk_diag_fill(sk, skb, portid, seq,
407 nlmsg_flags, unlh);
408
409 if (sk->sk_state == TCP_NEW_SYN_RECV)
410 return inet_req_diag_fill(sk, skb, portid, seq,
411 nlmsg_flags, unlh, net_admin);
412
413 return inet_csk_diag_fill(sk, skb, r, user_ns, portid, seq,
414 nlmsg_flags, unlh, net_admin);
415 }
416
417 struct sock *inet_diag_find_one_icsk(struct net *net,
418 struct inet_hashinfo *hashinfo,
419 const struct inet_diag_req_v2 *req)
420 {
421 struct sock *sk;
422
423 rcu_read_lock();
424 if (req->sdiag_family == AF_INET)
425 sk = inet_lookup(net, hashinfo, NULL, 0, req->id.idiag_dst[0],
426 req->id.idiag_dport, req->id.idiag_src[0],
427 req->id.idiag_sport, req->id.idiag_if);
428 #if IS_ENABLED(CONFIG_IPV6)
429 else if (req->sdiag_family == AF_INET6) {
430 if (ipv6_addr_v4mapped((struct in6_addr *)req->id.idiag_dst) &&
431 ipv6_addr_v4mapped((struct in6_addr *)req->id.idiag_src))
432 sk = inet_lookup(net, hashinfo, NULL, 0, req->id.idiag_dst[3],
433 req->id.idiag_dport, req->id.idiag_src[3],
434 req->id.idiag_sport, req->id.idiag_if);
435 else
436 sk = inet6_lookup(net, hashinfo, NULL, 0,
437 (struct in6_addr *)req->id.idiag_dst,
438 req->id.idiag_dport,
439 (struct in6_addr *)req->id.idiag_src,
440 req->id.idiag_sport,
441 req->id.idiag_if);
442 }
443 #endif
444 else {
445 rcu_read_unlock();
446 return ERR_PTR(-EINVAL);
447 }
448 rcu_read_unlock();
449 if (!sk)
450 return ERR_PTR(-ENOENT);
451
452 if (sock_diag_check_cookie(sk, req->id.idiag_cookie)) {
453 sock_gen_put(sk);
454 return ERR_PTR(-ENOENT);
455 }
456
457 return sk;
458 }
459 EXPORT_SYMBOL_GPL(inet_diag_find_one_icsk);
460
461 int inet_diag_dump_one_icsk(struct inet_hashinfo *hashinfo,
462 struct sk_buff *in_skb,
463 const struct nlmsghdr *nlh,
464 const struct inet_diag_req_v2 *req)
465 {
466 bool net_admin = netlink_net_capable(in_skb, CAP_NET_ADMIN);
467 struct net *net = sock_net(in_skb->sk);
468 struct sk_buff *rep;
469 struct sock *sk;
470 int err;
471
472 sk = inet_diag_find_one_icsk(net, hashinfo, req);
473 if (IS_ERR(sk))
474 return PTR_ERR(sk);
475
476 rep = nlmsg_new(inet_sk_attr_size(sk, req, net_admin), GFP_KERNEL);
477 if (!rep) {
478 err = -ENOMEM;
479 goto out;
480 }
481
482 err = sk_diag_fill(sk, rep, req,
483 sk_user_ns(NETLINK_CB(in_skb).sk),
484 NETLINK_CB(in_skb).portid,
485 nlh->nlmsg_seq, 0, nlh, net_admin);
486 if (err < 0) {
487 WARN_ON(err == -EMSGSIZE);
488 nlmsg_free(rep);
489 goto out;
490 }
491 err = netlink_unicast(net->diag_nlsk, rep, NETLINK_CB(in_skb).portid,
492 MSG_DONTWAIT);
493 if (err > 0)
494 err = 0;
495
496 out:
497 if (sk)
498 sock_gen_put(sk);
499
500 return err;
501 }
502 EXPORT_SYMBOL_GPL(inet_diag_dump_one_icsk);
503
504 static int inet_diag_cmd_exact(int cmd, struct sk_buff *in_skb,
505 const struct nlmsghdr *nlh,
506 const struct inet_diag_req_v2 *req)
507 {
508 const struct inet_diag_handler *handler;
509 int err;
510
511 handler = inet_diag_lock_handler(req->sdiag_protocol);
512 if (IS_ERR(handler))
513 err = PTR_ERR(handler);
514 else if (cmd == SOCK_DIAG_BY_FAMILY)
515 err = handler->dump_one(in_skb, nlh, req);
516 else if (cmd == SOCK_DESTROY && handler->destroy)
517 err = handler->destroy(in_skb, req);
518 else
519 err = -EOPNOTSUPP;
520 inet_diag_unlock_handler(handler);
521
522 return err;
523 }
524
525 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
526 {
527 int words = bits >> 5;
528
529 bits &= 0x1f;
530
531 if (words) {
532 if (memcmp(a1, a2, words << 2))
533 return 0;
534 }
535 if (bits) {
536 __be32 w1, w2;
537 __be32 mask;
538
539 w1 = a1[words];
540 w2 = a2[words];
541
542 mask = htonl((0xffffffff) << (32 - bits));
543
544 if ((w1 ^ w2) & mask)
545 return 0;
546 }
547
548 return 1;
549 }
550
551 static int inet_diag_bc_run(const struct nlattr *_bc,
552 const struct inet_diag_entry *entry)
553 {
554 const void *bc = nla_data(_bc);
555 int len = nla_len(_bc);
556
557 while (len > 0) {
558 int yes = 1;
559 const struct inet_diag_bc_op *op = bc;
560
561 switch (op->code) {
562 case INET_DIAG_BC_NOP:
563 break;
564 case INET_DIAG_BC_JMP:
565 yes = 0;
566 break;
567 case INET_DIAG_BC_S_EQ:
568 yes = entry->sport == op[1].no;
569 break;
570 case INET_DIAG_BC_S_GE:
571 yes = entry->sport >= op[1].no;
572 break;
573 case INET_DIAG_BC_S_LE:
574 yes = entry->sport <= op[1].no;
575 break;
576 case INET_DIAG_BC_D_EQ:
577 yes = entry->dport == op[1].no;
578 break;
579 case INET_DIAG_BC_D_GE:
580 yes = entry->dport >= op[1].no;
581 break;
582 case INET_DIAG_BC_D_LE:
583 yes = entry->dport <= op[1].no;
584 break;
585 case INET_DIAG_BC_AUTO:
586 yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
587 break;
588 case INET_DIAG_BC_S_COND:
589 case INET_DIAG_BC_D_COND: {
590 const struct inet_diag_hostcond *cond;
591 const __be32 *addr;
592
593 cond = (const struct inet_diag_hostcond *)(op + 1);
594 if (cond->port != -1 &&
595 cond->port != (op->code == INET_DIAG_BC_S_COND ?
596 entry->sport : entry->dport)) {
597 yes = 0;
598 break;
599 }
600
601 if (op->code == INET_DIAG_BC_S_COND)
602 addr = entry->saddr;
603 else
604 addr = entry->daddr;
605
606 if (cond->family != AF_UNSPEC &&
607 cond->family != entry->family) {
608 if (entry->family == AF_INET6 &&
609 cond->family == AF_INET) {
610 if (addr[0] == 0 && addr[1] == 0 &&
611 addr[2] == htonl(0xffff) &&
612 bitstring_match(addr + 3,
613 cond->addr,
614 cond->prefix_len))
615 break;
616 }
617 yes = 0;
618 break;
619 }
620
621 if (cond->prefix_len == 0)
622 break;
623 if (bitstring_match(addr, cond->addr,
624 cond->prefix_len))
625 break;
626 yes = 0;
627 break;
628 }
629 case INET_DIAG_BC_DEV_COND: {
630 u32 ifindex;
631
632 ifindex = *((const u32 *)(op + 1));
633 if (ifindex != entry->ifindex)
634 yes = 0;
635 break;
636 }
637 case INET_DIAG_BC_MARK_COND: {
638 struct inet_diag_markcond *cond;
639
640 cond = (struct inet_diag_markcond *)(op + 1);
641 if ((entry->mark & cond->mask) != cond->mark)
642 yes = 0;
643 break;
644 }
645 }
646
647 if (yes) {
648 len -= op->yes;
649 bc += op->yes;
650 } else {
651 len -= op->no;
652 bc += op->no;
653 }
654 }
655 return len == 0;
656 }
657
658 /* This helper is available for all sockets (ESTABLISH, TIMEWAIT, SYN_RECV)
659 */
660 static void entry_fill_addrs(struct inet_diag_entry *entry,
661 const struct sock *sk)
662 {
663 #if IS_ENABLED(CONFIG_IPV6)
664 if (sk->sk_family == AF_INET6) {
665 entry->saddr = sk->sk_v6_rcv_saddr.s6_addr32;
666 entry->daddr = sk->sk_v6_daddr.s6_addr32;
667 } else
668 #endif
669 {
670 entry->saddr = &sk->sk_rcv_saddr;
671 entry->daddr = &sk->sk_daddr;
672 }
673 }
674
675 int inet_diag_bc_sk(const struct nlattr *bc, struct sock *sk)
676 {
677 struct inet_sock *inet = inet_sk(sk);
678 struct inet_diag_entry entry;
679
680 if (!bc)
681 return 1;
682
683 entry.family = sk->sk_family;
684 entry_fill_addrs(&entry, sk);
685 entry.sport = inet->inet_num;
686 entry.dport = ntohs(inet->inet_dport);
687 entry.ifindex = sk->sk_bound_dev_if;
688 entry.userlocks = sk_fullsock(sk) ? sk->sk_userlocks : 0;
689 if (sk_fullsock(sk))
690 entry.mark = sk->sk_mark;
691 else if (sk->sk_state == TCP_NEW_SYN_RECV)
692 entry.mark = inet_rsk(inet_reqsk(sk))->ir_mark;
693 else
694 entry.mark = 0;
695
696 return inet_diag_bc_run(bc, &entry);
697 }
698 EXPORT_SYMBOL_GPL(inet_diag_bc_sk);
699
700 static int valid_cc(const void *bc, int len, int cc)
701 {
702 while (len >= 0) {
703 const struct inet_diag_bc_op *op = bc;
704
705 if (cc > len)
706 return 0;
707 if (cc == len)
708 return 1;
709 if (op->yes < 4 || op->yes & 3)
710 return 0;
711 len -= op->yes;
712 bc += op->yes;
713 }
714 return 0;
715 }
716
717 /* data is u32 ifindex */
718 static bool valid_devcond(const struct inet_diag_bc_op *op, int len,
719 int *min_len)
720 {
721 /* Check ifindex space. */
722 *min_len += sizeof(u32);
723 if (len < *min_len)
724 return false;
725
726 return true;
727 }
728 /* Validate an inet_diag_hostcond. */
729 static bool valid_hostcond(const struct inet_diag_bc_op *op, int len,
730 int *min_len)
731 {
732 struct inet_diag_hostcond *cond;
733 int addr_len;
734
735 /* Check hostcond space. */
736 *min_len += sizeof(struct inet_diag_hostcond);
737 if (len < *min_len)
738 return false;
739 cond = (struct inet_diag_hostcond *)(op + 1);
740
741 /* Check address family and address length. */
742 switch (cond->family) {
743 case AF_UNSPEC:
744 addr_len = 0;
745 break;
746 case AF_INET:
747 addr_len = sizeof(struct in_addr);
748 break;
749 case AF_INET6:
750 addr_len = sizeof(struct in6_addr);
751 break;
752 default:
753 return false;
754 }
755 *min_len += addr_len;
756 if (len < *min_len)
757 return false;
758
759 /* Check prefix length (in bits) vs address length (in bytes). */
760 if (cond->prefix_len > 8 * addr_len)
761 return false;
762
763 return true;
764 }
765
766 /* Validate a port comparison operator. */
767 static bool valid_port_comparison(const struct inet_diag_bc_op *op,
768 int len, int *min_len)
769 {
770 /* Port comparisons put the port in a follow-on inet_diag_bc_op. */
771 *min_len += sizeof(struct inet_diag_bc_op);
772 if (len < *min_len)
773 return false;
774 return true;
775 }
776
777 static bool valid_markcond(const struct inet_diag_bc_op *op, int len,
778 int *min_len)
779 {
780 *min_len += sizeof(struct inet_diag_markcond);
781 return len >= *min_len;
782 }
783
784 static int inet_diag_bc_audit(const struct nlattr *attr,
785 const struct sk_buff *skb)
786 {
787 bool net_admin = netlink_net_capable(skb, CAP_NET_ADMIN);
788 const void *bytecode, *bc;
789 int bytecode_len, len;
790
791 if (!attr || nla_len(attr) < sizeof(struct inet_diag_bc_op))
792 return -EINVAL;
793
794 bytecode = bc = nla_data(attr);
795 len = bytecode_len = nla_len(attr);
796
797 while (len > 0) {
798 int min_len = sizeof(struct inet_diag_bc_op);
799 const struct inet_diag_bc_op *op = bc;
800
801 switch (op->code) {
802 case INET_DIAG_BC_S_COND:
803 case INET_DIAG_BC_D_COND:
804 if (!valid_hostcond(bc, len, &min_len))
805 return -EINVAL;
806 break;
807 case INET_DIAG_BC_DEV_COND:
808 if (!valid_devcond(bc, len, &min_len))
809 return -EINVAL;
810 break;
811 case INET_DIAG_BC_S_EQ:
812 case INET_DIAG_BC_S_GE:
813 case INET_DIAG_BC_S_LE:
814 case INET_DIAG_BC_D_EQ:
815 case INET_DIAG_BC_D_GE:
816 case INET_DIAG_BC_D_LE:
817 if (!valid_port_comparison(bc, len, &min_len))
818 return -EINVAL;
819 break;
820 case INET_DIAG_BC_MARK_COND:
821 if (!net_admin)
822 return -EPERM;
823 if (!valid_markcond(bc, len, &min_len))
824 return -EINVAL;
825 break;
826 case INET_DIAG_BC_AUTO:
827 case INET_DIAG_BC_JMP:
828 case INET_DIAG_BC_NOP:
829 break;
830 default:
831 return -EINVAL;
832 }
833
834 if (op->code != INET_DIAG_BC_NOP) {
835 if (op->no < min_len || op->no > len + 4 || op->no & 3)
836 return -EINVAL;
837 if (op->no < len &&
838 !valid_cc(bytecode, bytecode_len, len - op->no))
839 return -EINVAL;
840 }
841
842 if (op->yes < min_len || op->yes > len + 4 || op->yes & 3)
843 return -EINVAL;
844 bc += op->yes;
845 len -= op->yes;
846 }
847 return len == 0 ? 0 : -EINVAL;
848 }
849
850 static int inet_csk_diag_dump(struct sock *sk,
851 struct sk_buff *skb,
852 struct netlink_callback *cb,
853 const struct inet_diag_req_v2 *r,
854 const struct nlattr *bc,
855 bool net_admin)
856 {
857 if (!inet_diag_bc_sk(bc, sk))
858 return 0;
859
860 return inet_csk_diag_fill(sk, skb, r,
861 sk_user_ns(NETLINK_CB(cb->skb).sk),
862 NETLINK_CB(cb->skb).portid,
863 cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh,
864 net_admin);
865 }
866
867 static void twsk_build_assert(void)
868 {
869 BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_family) !=
870 offsetof(struct sock, sk_family));
871
872 BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_num) !=
873 offsetof(struct inet_sock, inet_num));
874
875 BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_dport) !=
876 offsetof(struct inet_sock, inet_dport));
877
878 BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_rcv_saddr) !=
879 offsetof(struct inet_sock, inet_rcv_saddr));
880
881 BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_daddr) !=
882 offsetof(struct inet_sock, inet_daddr));
883
884 #if IS_ENABLED(CONFIG_IPV6)
885 BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_v6_rcv_saddr) !=
886 offsetof(struct sock, sk_v6_rcv_saddr));
887
888 BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_v6_daddr) !=
889 offsetof(struct sock, sk_v6_daddr));
890 #endif
891 }
892
893 void inet_diag_dump_icsk(struct inet_hashinfo *hashinfo, struct sk_buff *skb,
894 struct netlink_callback *cb,
895 const struct inet_diag_req_v2 *r, struct nlattr *bc)
896 {
897 bool net_admin = netlink_net_capable(cb->skb, CAP_NET_ADMIN);
898 struct net *net = sock_net(skb->sk);
899 u32 idiag_states = r->idiag_states;
900 int i, num, s_i, s_num;
901 struct sock *sk;
902
903 if (idiag_states & TCPF_SYN_RECV)
904 idiag_states |= TCPF_NEW_SYN_RECV;
905 s_i = cb->args[1];
906 s_num = num = cb->args[2];
907
908 if (cb->args[0] == 0) {
909 if (!(idiag_states & TCPF_LISTEN) || r->id.idiag_dport)
910 goto skip_listen_ht;
911
912 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
913 struct inet_listen_hashbucket *ilb;
914 struct hlist_nulls_node *node;
915
916 num = 0;
917 ilb = &hashinfo->listening_hash[i];
918 spin_lock(&ilb->lock);
919 sk_nulls_for_each(sk, node, &ilb->nulls_head) {
920 struct inet_sock *inet = inet_sk(sk);
921
922 if (!net_eq(sock_net(sk), net))
923 continue;
924
925 if (num < s_num) {
926 num++;
927 continue;
928 }
929
930 if (r->sdiag_family != AF_UNSPEC &&
931 sk->sk_family != r->sdiag_family)
932 goto next_listen;
933
934 if (r->id.idiag_sport != inet->inet_sport &&
935 r->id.idiag_sport)
936 goto next_listen;
937
938 if (inet_csk_diag_dump(sk, skb, cb, r,
939 bc, net_admin) < 0) {
940 spin_unlock(&ilb->lock);
941 goto done;
942 }
943
944 next_listen:
945 ++num;
946 }
947 spin_unlock(&ilb->lock);
948
949 s_num = 0;
950 }
951 skip_listen_ht:
952 cb->args[0] = 1;
953 s_i = num = s_num = 0;
954 }
955
956 if (!(idiag_states & ~TCPF_LISTEN))
957 goto out;
958
959 #define SKARR_SZ 16
960 for (i = s_i; i <= hashinfo->ehash_mask; i++) {
961 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
962 spinlock_t *lock = inet_ehash_lockp(hashinfo, i);
963 struct hlist_nulls_node *node;
964 struct sock *sk_arr[SKARR_SZ];
965 int num_arr[SKARR_SZ];
966 int idx, accum, res;
967
968 if (hlist_nulls_empty(&head->chain))
969 continue;
970
971 if (i > s_i)
972 s_num = 0;
973
974 next_chunk:
975 num = 0;
976 accum = 0;
977 spin_lock_bh(lock);
978 sk_nulls_for_each(sk, node, &head->chain) {
979 int state;
980
981 if (!net_eq(sock_net(sk), net))
982 continue;
983 if (num < s_num)
984 goto next_normal;
985 state = (sk->sk_state == TCP_TIME_WAIT) ?
986 inet_twsk(sk)->tw_substate : sk->sk_state;
987 if (!(idiag_states & (1 << state)))
988 goto next_normal;
989 if (r->sdiag_family != AF_UNSPEC &&
990 sk->sk_family != r->sdiag_family)
991 goto next_normal;
992 if (r->id.idiag_sport != htons(sk->sk_num) &&
993 r->id.idiag_sport)
994 goto next_normal;
995 if (r->id.idiag_dport != sk->sk_dport &&
996 r->id.idiag_dport)
997 goto next_normal;
998 twsk_build_assert();
999
1000 if (!inet_diag_bc_sk(bc, sk))
1001 goto next_normal;
1002
1003 if (!refcount_inc_not_zero(&sk->sk_refcnt))
1004 goto next_normal;
1005
1006 num_arr[accum] = num;
1007 sk_arr[accum] = sk;
1008 if (++accum == SKARR_SZ)
1009 break;
1010 next_normal:
1011 ++num;
1012 }
1013 spin_unlock_bh(lock);
1014 res = 0;
1015 for (idx = 0; idx < accum; idx++) {
1016 if (res >= 0) {
1017 res = sk_diag_fill(sk_arr[idx], skb, r,
1018 sk_user_ns(NETLINK_CB(cb->skb).sk),
1019 NETLINK_CB(cb->skb).portid,
1020 cb->nlh->nlmsg_seq, NLM_F_MULTI,
1021 cb->nlh, net_admin);
1022 if (res < 0)
1023 num = num_arr[idx];
1024 }
1025 sock_gen_put(sk_arr[idx]);
1026 }
1027 if (res < 0)
1028 break;
1029 cond_resched();
1030 if (accum == SKARR_SZ) {
1031 s_num = num + 1;
1032 goto next_chunk;
1033 }
1034 }
1035
1036 done:
1037 cb->args[1] = i;
1038 cb->args[2] = num;
1039 out:
1040 ;
1041 }
1042 EXPORT_SYMBOL_GPL(inet_diag_dump_icsk);
1043
1044 static int __inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb,
1045 const struct inet_diag_req_v2 *r,
1046 struct nlattr *bc)
1047 {
1048 const struct inet_diag_handler *handler;
1049 int err = 0;
1050
1051 handler = inet_diag_lock_handler(r->sdiag_protocol);
1052 if (!IS_ERR(handler))
1053 handler->dump(skb, cb, r, bc);
1054 else
1055 err = PTR_ERR(handler);
1056 inet_diag_unlock_handler(handler);
1057
1058 return err ? : skb->len;
1059 }
1060
1061 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
1062 {
1063 int hdrlen = sizeof(struct inet_diag_req_v2);
1064 struct nlattr *bc = NULL;
1065
1066 if (nlmsg_attrlen(cb->nlh, hdrlen))
1067 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
1068
1069 return __inet_diag_dump(skb, cb, nlmsg_data(cb->nlh), bc);
1070 }
1071
1072 static int inet_diag_type2proto(int type)
1073 {
1074 switch (type) {
1075 case TCPDIAG_GETSOCK:
1076 return IPPROTO_TCP;
1077 case DCCPDIAG_GETSOCK:
1078 return IPPROTO_DCCP;
1079 default:
1080 return 0;
1081 }
1082 }
1083
1084 static int inet_diag_dump_compat(struct sk_buff *skb,
1085 struct netlink_callback *cb)
1086 {
1087 struct inet_diag_req *rc = nlmsg_data(cb->nlh);
1088 int hdrlen = sizeof(struct inet_diag_req);
1089 struct inet_diag_req_v2 req;
1090 struct nlattr *bc = NULL;
1091
1092 req.sdiag_family = AF_UNSPEC; /* compatibility */
1093 req.sdiag_protocol = inet_diag_type2proto(cb->nlh->nlmsg_type);
1094 req.idiag_ext = rc->idiag_ext;
1095 req.idiag_states = rc->idiag_states;
1096 req.id = rc->id;
1097
1098 if (nlmsg_attrlen(cb->nlh, hdrlen))
1099 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
1100
1101 return __inet_diag_dump(skb, cb, &req, bc);
1102 }
1103
1104 static int inet_diag_get_exact_compat(struct sk_buff *in_skb,
1105 const struct nlmsghdr *nlh)
1106 {
1107 struct inet_diag_req *rc = nlmsg_data(nlh);
1108 struct inet_diag_req_v2 req;
1109
1110 req.sdiag_family = rc->idiag_family;
1111 req.sdiag_protocol = inet_diag_type2proto(nlh->nlmsg_type);
1112 req.idiag_ext = rc->idiag_ext;
1113 req.idiag_states = rc->idiag_states;
1114 req.id = rc->id;
1115
1116 return inet_diag_cmd_exact(SOCK_DIAG_BY_FAMILY, in_skb, nlh, &req);
1117 }
1118
1119 static int inet_diag_rcv_msg_compat(struct sk_buff *skb, struct nlmsghdr *nlh)
1120 {
1121 int hdrlen = sizeof(struct inet_diag_req);
1122 struct net *net = sock_net(skb->sk);
1123
1124 if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
1125 nlmsg_len(nlh) < hdrlen)
1126 return -EINVAL;
1127
1128 if (nlh->nlmsg_flags & NLM_F_DUMP) {
1129 if (nlmsg_attrlen(nlh, hdrlen)) {
1130 struct nlattr *attr;
1131 int err;
1132
1133 attr = nlmsg_find_attr(nlh, hdrlen,
1134 INET_DIAG_REQ_BYTECODE);
1135 err = inet_diag_bc_audit(attr, skb);
1136 if (err)
1137 return err;
1138 }
1139 {
1140 struct netlink_dump_control c = {
1141 .dump = inet_diag_dump_compat,
1142 };
1143 return netlink_dump_start(net->diag_nlsk, skb, nlh, &c);
1144 }
1145 }
1146
1147 return inet_diag_get_exact_compat(skb, nlh);
1148 }
1149
1150 static int inet_diag_handler_cmd(struct sk_buff *skb, struct nlmsghdr *h)
1151 {
1152 int hdrlen = sizeof(struct inet_diag_req_v2);
1153 struct net *net = sock_net(skb->sk);
1154
1155 if (nlmsg_len(h) < hdrlen)
1156 return -EINVAL;
1157
1158 if (h->nlmsg_type == SOCK_DIAG_BY_FAMILY &&
1159 h->nlmsg_flags & NLM_F_DUMP) {
1160 if (nlmsg_attrlen(h, hdrlen)) {
1161 struct nlattr *attr;
1162 int err;
1163
1164 attr = nlmsg_find_attr(h, hdrlen,
1165 INET_DIAG_REQ_BYTECODE);
1166 err = inet_diag_bc_audit(attr, skb);
1167 if (err)
1168 return err;
1169 }
1170 {
1171 struct netlink_dump_control c = {
1172 .dump = inet_diag_dump,
1173 };
1174 return netlink_dump_start(net->diag_nlsk, skb, h, &c);
1175 }
1176 }
1177
1178 return inet_diag_cmd_exact(h->nlmsg_type, skb, h, nlmsg_data(h));
1179 }
1180
1181 static
1182 int inet_diag_handler_get_info(struct sk_buff *skb, struct sock *sk)
1183 {
1184 const struct inet_diag_handler *handler;
1185 struct nlmsghdr *nlh;
1186 struct nlattr *attr;
1187 struct inet_diag_msg *r;
1188 void *info = NULL;
1189 int err = 0;
1190
1191 nlh = nlmsg_put(skb, 0, 0, SOCK_DIAG_BY_FAMILY, sizeof(*r), 0);
1192 if (!nlh)
1193 return -ENOMEM;
1194
1195 r = nlmsg_data(nlh);
1196 memset(r, 0, sizeof(*r));
1197 inet_diag_msg_common_fill(r, sk);
1198 if (sk->sk_type == SOCK_DGRAM || sk->sk_type == SOCK_STREAM)
1199 r->id.idiag_sport = inet_sk(sk)->inet_sport;
1200 r->idiag_state = sk->sk_state;
1201
1202 if ((err = nla_put_u8(skb, INET_DIAG_PROTOCOL, sk->sk_protocol))) {
1203 nlmsg_cancel(skb, nlh);
1204 return err;
1205 }
1206
1207 handler = inet_diag_lock_handler(sk->sk_protocol);
1208 if (IS_ERR(handler)) {
1209 inet_diag_unlock_handler(handler);
1210 nlmsg_cancel(skb, nlh);
1211 return PTR_ERR(handler);
1212 }
1213
1214 attr = handler->idiag_info_size
1215 ? nla_reserve_64bit(skb, INET_DIAG_INFO,
1216 handler->idiag_info_size,
1217 INET_DIAG_PAD)
1218 : NULL;
1219 if (attr)
1220 info = nla_data(attr);
1221
1222 handler->idiag_get_info(sk, r, info);
1223 inet_diag_unlock_handler(handler);
1224
1225 nlmsg_end(skb, nlh);
1226 return 0;
1227 }
1228
1229 static const struct sock_diag_handler inet_diag_handler = {
1230 .family = AF_INET,
1231 .dump = inet_diag_handler_cmd,
1232 .get_info = inet_diag_handler_get_info,
1233 .destroy = inet_diag_handler_cmd,
1234 };
1235
1236 static const struct sock_diag_handler inet6_diag_handler = {
1237 .family = AF_INET6,
1238 .dump = inet_diag_handler_cmd,
1239 .get_info = inet_diag_handler_get_info,
1240 .destroy = inet_diag_handler_cmd,
1241 };
1242
1243 int inet_diag_register(const struct inet_diag_handler *h)
1244 {
1245 const __u16 type = h->idiag_type;
1246 int err = -EINVAL;
1247
1248 if (type >= IPPROTO_MAX)
1249 goto out;
1250
1251 mutex_lock(&inet_diag_table_mutex);
1252 err = -EEXIST;
1253 if (!inet_diag_table[type]) {
1254 inet_diag_table[type] = h;
1255 err = 0;
1256 }
1257 mutex_unlock(&inet_diag_table_mutex);
1258 out:
1259 return err;
1260 }
1261 EXPORT_SYMBOL_GPL(inet_diag_register);
1262
1263 void inet_diag_unregister(const struct inet_diag_handler *h)
1264 {
1265 const __u16 type = h->idiag_type;
1266
1267 if (type >= IPPROTO_MAX)
1268 return;
1269
1270 mutex_lock(&inet_diag_table_mutex);
1271 inet_diag_table[type] = NULL;
1272 mutex_unlock(&inet_diag_table_mutex);
1273 }
1274 EXPORT_SYMBOL_GPL(inet_diag_unregister);
1275
1276 static int __init inet_diag_init(void)
1277 {
1278 const int inet_diag_table_size = (IPPROTO_MAX *
1279 sizeof(struct inet_diag_handler *));
1280 int err = -ENOMEM;
1281
1282 inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
1283 if (!inet_diag_table)
1284 goto out;
1285
1286 err = sock_diag_register(&inet_diag_handler);
1287 if (err)
1288 goto out_free_nl;
1289
1290 err = sock_diag_register(&inet6_diag_handler);
1291 if (err)
1292 goto out_free_inet;
1293
1294 sock_diag_register_inet_compat(inet_diag_rcv_msg_compat);
1295 out:
1296 return err;
1297
1298 out_free_inet:
1299 sock_diag_unregister(&inet_diag_handler);
1300 out_free_nl:
1301 kfree(inet_diag_table);
1302 goto out;
1303 }
1304
1305 static void __exit inet_diag_exit(void)
1306 {
1307 sock_diag_unregister(&inet6_diag_handler);
1308 sock_diag_unregister(&inet_diag_handler);
1309 sock_diag_unregister_inet_compat(inet_diag_rcv_msg_compat);
1310 kfree(inet_diag_table);
1311 }
1312
1313 module_init(inet_diag_init);
1314 module_exit(inet_diag_exit);
1315 MODULE_LICENSE("GPL");
1316 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 2 /* AF_INET */);
1317 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 10 /* AF_INET6 */);
Linux with added FPC-III support