net/ipv4/inet_diag.c

   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 struct sock *inet_diag_find_one_icsk(struct net *net,
 354                                      struct inet_hashinfo *hashinfo,
 355                                      const struct inet_diag_req_v2 *req)
 356 {
 357         struct sock *sk;
 358
 359         rcu_read_lock();
 360         if (req->sdiag_family == AF_INET)
 361                 sk = inet_lookup(net, hashinfo, NULL, 0, req->id.idiag_dst[0],
 362                                  req->id.idiag_dport, req->id.idiag_src[0],
 363                                  req->id.idiag_sport, req->id.idiag_if);
 364 #if IS_ENABLED(CONFIG_IPV6)
 365         else if (req->sdiag_family == AF_INET6) {
 366                 if (ipv6_addr_v4mapped((struct in6_addr *)req->id.idiag_dst) &&
 367                     ipv6_addr_v4mapped((struct in6_addr *)req->id.idiag_src))
 368                         sk = inet_lookup(net, hashinfo, NULL, 0, req->id.idiag_dst[3],
 369                                          req->id.idiag_dport, req->id.idiag_src[3],
 370                                          req->id.idiag_sport, req->id.idiag_if);
 371                 else
 372                         sk = inet6_lookup(net, hashinfo, NULL, 0,
 373                                           (struct in6_addr *)req->id.idiag_dst,
 374                                           req->id.idiag_dport,
 375                                           (struct in6_addr *)req->id.idiag_src,
 376                                           req->id.idiag_sport,
 377                                           req->id.idiag_if);
 378         }
 379 #endif
 380         else {
 381                 rcu_read_unlock();
 382                 return ERR_PTR(-EINVAL);
 383         }
 384         rcu_read_unlock();
 385         if (!sk)
 386                 return ERR_PTR(-ENOENT);
 387
 388         if (sock_diag_check_cookie(sk, req->id.idiag_cookie)) {
 389                 sock_gen_put(sk);
 390                 return ERR_PTR(-ENOENT);
 391         }
 392
 393         return sk;
 394 }
 395 EXPORT_SYMBOL_GPL(inet_diag_find_one_icsk);
 396
 397 int inet_diag_dump_one_icsk(struct inet_hashinfo *hashinfo,
 398                             struct sk_buff *in_skb,
 399                             const struct nlmsghdr *nlh,
 400                             const struct inet_diag_req_v2 *req)
 401 {
 402         struct net *net = sock_net(in_skb->sk);
 403         struct sk_buff *rep;
 404         struct sock *sk;
 405         int err;
 406
 407         sk = inet_diag_find_one_icsk(net, hashinfo, req);
 408         if (IS_ERR(sk))
 409                 return PTR_ERR(sk);
 410
 411         rep = nlmsg_new(inet_sk_attr_size(), GFP_KERNEL);
 412         if (!rep) {
 413                 err = -ENOMEM;
 414                 goto out;
 415         }
 416
 417         err = sk_diag_fill(sk, rep, req,
 418                            sk_user_ns(NETLINK_CB(in_skb).sk),
 419                            NETLINK_CB(in_skb).portid,
 420                            nlh->nlmsg_seq, 0, nlh);
 421         if (err < 0) {
 422                 WARN_ON(err == -EMSGSIZE);
 423                 nlmsg_free(rep);
 424                 goto out;
 425         }
 426         err = netlink_unicast(net->diag_nlsk, rep, NETLINK_CB(in_skb).portid,
 427                               MSG_DONTWAIT);
 428         if (err > 0)
 429                 err = 0;
 430
 431 out:
 432         if (sk)
 433                 sock_gen_put(sk);
 434
 435         return err;
 436 }
 437 EXPORT_SYMBOL_GPL(inet_diag_dump_one_icsk);
 438
 439 static int inet_diag_cmd_exact(int cmd, struct sk_buff *in_skb,
 440                                const struct nlmsghdr *nlh,
 441                                const struct inet_diag_req_v2 *req)
 442 {
 443         const struct inet_diag_handler *handler;
 444         int err;
 445
 446         handler = inet_diag_lock_handler(req->sdiag_protocol);
 447         if (IS_ERR(handler))
 448                 err = PTR_ERR(handler);
 449         else if (cmd == SOCK_DIAG_BY_FAMILY)
 450                 err = handler->dump_one(in_skb, nlh, req);
 451         else if (cmd == SOCK_DESTROY && handler->destroy)
 452                 err = handler->destroy(in_skb, req);
 453         else
 454                 err = -EOPNOTSUPP;
 455         inet_diag_unlock_handler(handler);
 456
 457         return err;
 458 }
 459
 460 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
 461 {
 462         int words = bits >> 5;
 463
 464         bits &= 0x1f;
 465
 466         if (words) {
 467                 if (memcmp(a1, a2, words << 2))
 468                         return 0;
 469         }
 470         if (bits) {
 471                 __be32 w1, w2;
 472                 __be32 mask;
 473
 474                 w1 = a1[words];
 475                 w2 = a2[words];
 476
 477                 mask = htonl((0xffffffff) << (32 - bits));
 478
 479                 if ((w1 ^ w2) & mask)
 480                         return 0;
 481         }
 482
 483         return 1;
 484 }
 485
 486 static int inet_diag_bc_run(const struct nlattr *_bc,
 487                             const struct inet_diag_entry *entry)
 488 {
 489         const void *bc = nla_data(_bc);
 490         int len = nla_len(_bc);
 491
 492         while (len > 0) {
 493                 int yes = 1;
 494                 const struct inet_diag_bc_op *op = bc;
 495
 496                 switch (op->code) {
 497                 case INET_DIAG_BC_NOP:
 498                         break;
 499                 case INET_DIAG_BC_JMP:
 500                         yes = 0;
 501                         break;
 502                 case INET_DIAG_BC_S_GE:
 503                         yes = entry->sport >= op[1].no;
 504                         break;
 505                 case INET_DIAG_BC_S_LE:
 506                         yes = entry->sport <= op[1].no;
 507                         break;
 508                 case INET_DIAG_BC_D_GE:
 509                         yes = entry->dport >= op[1].no;
 510                         break;
 511                 case INET_DIAG_BC_D_LE:
 512                         yes = entry->dport <= op[1].no;
 513                         break;
 514                 case INET_DIAG_BC_AUTO:
 515                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 516                         break;
 517                 case INET_DIAG_BC_S_COND:
 518                 case INET_DIAG_BC_D_COND: {
 519                         const struct inet_diag_hostcond *cond;
 520                         const __be32 *addr;
 521
 522                         cond = (const struct inet_diag_hostcond *)(op + 1);
 523                         if (cond->port != -1 &&
 524                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 525                                              entry->sport : entry->dport)) {
 526                                 yes = 0;
 527                                 break;
 528                         }
 529
 530                         if (op->code == INET_DIAG_BC_S_COND)
 531                                 addr = entry->saddr;
 532                         else
 533                                 addr = entry->daddr;
 534
 535                         if (cond->family != AF_UNSPEC &&
 536                             cond->family != entry->family) {
 537                                 if (entry->family == AF_INET6 &&
 538                                     cond->family == AF_INET) {
 539                                         if (addr[0] == 0 && addr[1] == 0 &&
 540                                             addr[2] == htonl(0xffff) &&
 541                                             bitstring_match(addr + 3,
 542                                                             cond->addr,
 543                                                             cond->prefix_len))
 544                                                 break;
 545                                 }
 546                                 yes = 0;
 547                                 break;
 548                         }
 549
 550                         if (cond->prefix_len == 0)
 551                                 break;
 552                         if (bitstring_match(addr, cond->addr,
 553                                             cond->prefix_len))
 554                                 break;
 555                         yes = 0;
 556                         break;
 557                 }
 558                 }
 559
 560                 if (yes) {
 561                         len -= op->yes;
 562                         bc += op->yes;
 563                 } else {
 564                         len -= op->no;
 565                         bc += op->no;
 566                 }
 567         }
 568         return len == 0;
 569 }
 570
 571 /* This helper is available for all sockets (ESTABLISH, TIMEWAIT, SYN_RECV)
 572  */
 573 static void entry_fill_addrs(struct inet_diag_entry *entry,
 574                              const struct sock *sk)
 575 {
 576 #if IS_ENABLED(CONFIG_IPV6)
 577         if (sk->sk_family == AF_INET6) {
 578                 entry->saddr = sk->sk_v6_rcv_saddr.s6_addr32;
 579                 entry->daddr = sk->sk_v6_daddr.s6_addr32;
 580         } else
 581 #endif
 582         {
 583                 entry->saddr = &sk->sk_rcv_saddr;
 584                 entry->daddr = &sk->sk_daddr;
 585         }
 586 }
 587
 588 int inet_diag_bc_sk(const struct nlattr *bc, struct sock *sk)
 589 {
 590         struct inet_sock *inet = inet_sk(sk);
 591         struct inet_diag_entry entry;
 592
 593         if (!bc)
 594                 return 1;
 595
 596         entry.family = sk->sk_family;
 597         entry_fill_addrs(&entry, sk);
 598         entry.sport = inet->inet_num;
 599         entry.dport = ntohs(inet->inet_dport);
 600         entry.userlocks = sk_fullsock(sk) ? sk->sk_userlocks : 0;
 601
 602         return inet_diag_bc_run(bc, &entry);
 603 }
 604 EXPORT_SYMBOL_GPL(inet_diag_bc_sk);
 605
 606 static int valid_cc(const void *bc, int len, int cc)
 607 {
 608         while (len >= 0) {
 609                 const struct inet_diag_bc_op *op = bc;
 610
 611                 if (cc > len)
 612                         return 0;
 613                 if (cc == len)
 614                         return 1;
 615                 if (op->yes < 4 || op->yes & 3)
 616                         return 0;
 617                 len -= op->yes;
 618                 bc  += op->yes;
 619         }
 620         return 0;
 621 }
 622
 623 /* Validate an inet_diag_hostcond. */
 624 static bool valid_hostcond(const struct inet_diag_bc_op *op, int len,
 625                            int *min_len)
 626 {
 627         struct inet_diag_hostcond *cond;
 628         int addr_len;
 629
 630         /* Check hostcond space. */
 631         *min_len += sizeof(struct inet_diag_hostcond);
 632         if (len < *min_len)
 633                 return false;
 634         cond = (struct inet_diag_hostcond *)(op + 1);
 635
 636         /* Check address family and address length. */
 637         switch (cond->family) {
 638         case AF_UNSPEC:
 639                 addr_len = 0;
 640                 break;
 641         case AF_INET:
 642                 addr_len = sizeof(struct in_addr);
 643                 break;
 644         case AF_INET6:
 645                 addr_len = sizeof(struct in6_addr);
 646                 break;
 647         default:
 648                 return false;
 649         }
 650         *min_len += addr_len;
 651         if (len < *min_len)
 652                 return false;
 653
 654         /* Check prefix length (in bits) vs address length (in bytes). */
 655         if (cond->prefix_len > 8 * addr_len)
 656                 return false;
 657
 658         return true;
 659 }
 660
 661 /* Validate a port comparison operator. */
 662 static bool valid_port_comparison(const struct inet_diag_bc_op *op,
 663                                   int len, int *min_len)
 664 {
 665         /* Port comparisons put the port in a follow-on inet_diag_bc_op. */
 666         *min_len += sizeof(struct inet_diag_bc_op);
 667         if (len < *min_len)
 668                 return false;
 669         return true;
 670 }
 671
 672 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
 673 {
 674         const void *bc = bytecode;
 675         int  len = bytecode_len;
 676
 677         while (len > 0) {
 678                 int min_len = sizeof(struct inet_diag_bc_op);
 679                 const struct inet_diag_bc_op *op = bc;
 680
 681                 switch (op->code) {
 682                 case INET_DIAG_BC_S_COND:
 683                 case INET_DIAG_BC_D_COND:
 684                         if (!valid_hostcond(bc, len, &min_len))
 685                                 return -EINVAL;
 686                         break;
 687                 case INET_DIAG_BC_S_GE:
 688                 case INET_DIAG_BC_S_LE:
 689                 case INET_DIAG_BC_D_GE:
 690                 case INET_DIAG_BC_D_LE:
 691                         if (!valid_port_comparison(bc, len, &min_len))
 692                                 return -EINVAL;
 693                         break;
 694                 case INET_DIAG_BC_AUTO:
 695                 case INET_DIAG_BC_JMP:
 696                 case INET_DIAG_BC_NOP:
 697                         break;
 698                 default:
 699                         return -EINVAL;
 700                 }
 701
 702                 if (op->code != INET_DIAG_BC_NOP) {
 703                         if (op->no < min_len || op->no > len + 4 || op->no & 3)
 704                                 return -EINVAL;
 705                         if (op->no < len &&
 706                             !valid_cc(bytecode, bytecode_len, len - op->no))
 707                                 return -EINVAL;
 708                 }
 709
 710                 if (op->yes < min_len || op->yes > len + 4 || op->yes & 3)
 711                         return -EINVAL;
 712                 bc  += op->yes;
 713                 len -= op->yes;
 714         }
 715         return len == 0 ? 0 : -EINVAL;
 716 }
 717
 718 static int inet_csk_diag_dump(struct sock *sk,
 719                               struct sk_buff *skb,
 720                               struct netlink_callback *cb,
 721                               const struct inet_diag_req_v2 *r,
 722                               const struct nlattr *bc)
 723 {
 724         if (!inet_diag_bc_sk(bc, sk))
 725                 return 0;
 726
 727         return inet_csk_diag_fill(sk, skb, r,
 728                                   sk_user_ns(NETLINK_CB(cb->skb).sk),
 729                                   NETLINK_CB(cb->skb).portid,
 730                                   cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 731 }
 732
 733 static void twsk_build_assert(void)
 734 {
 735         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_family) !=
 736                      offsetof(struct sock, sk_family));
 737
 738         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_num) !=
 739                      offsetof(struct inet_sock, inet_num));
 740
 741         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_dport) !=
 742                      offsetof(struct inet_sock, inet_dport));
 743
 744         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_rcv_saddr) !=
 745                      offsetof(struct inet_sock, inet_rcv_saddr));
 746
 747         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_daddr) !=
 748                      offsetof(struct inet_sock, inet_daddr));
 749
 750 #if IS_ENABLED(CONFIG_IPV6)
 751         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_v6_rcv_saddr) !=
 752                      offsetof(struct sock, sk_v6_rcv_saddr));
 753
 754         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_v6_daddr) !=
 755                      offsetof(struct sock, sk_v6_daddr));
 756 #endif
 757 }
 758
 759 void inet_diag_dump_icsk(struct inet_hashinfo *hashinfo, struct sk_buff *skb,
 760                          struct netlink_callback *cb,
 761                          const struct inet_diag_req_v2 *r, struct nlattr *bc)
 762 {
 763         struct net *net = sock_net(skb->sk);
 764         int i, num, s_i, s_num;
 765         u32 idiag_states = r->idiag_states;
 766
 767         if (idiag_states & TCPF_SYN_RECV)
 768                 idiag_states |= TCPF_NEW_SYN_RECV;
 769         s_i = cb->args[1];
 770         s_num = num = cb->args[2];
 771
 772         if (cb->args[0] == 0) {
 773                 if (!(idiag_states & TCPF_LISTEN))
 774                         goto skip_listen_ht;
 775
 776                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 777                         struct inet_listen_hashbucket *ilb;
 778                         struct sock *sk;
 779
 780                         num = 0;
 781                         ilb = &hashinfo->listening_hash[i];
 782                         spin_lock_bh(&ilb->lock);
 783                         sk_for_each(sk, &ilb->head) {
 784                                 struct inet_sock *inet = inet_sk(sk);
 785
 786                                 if (!net_eq(sock_net(sk), net))
 787                                         continue;
 788
 789                                 if (num < s_num) {
 790                                         num++;
 791                                         continue;
 792                                 }
 793
 794                                 if (r->sdiag_family != AF_UNSPEC &&
 795                                     sk->sk_family != r->sdiag_family)
 796                                         goto next_listen;
 797
 798                                 if (r->id.idiag_sport != inet->inet_sport &&
 799                                     r->id.idiag_sport)
 800                                         goto next_listen;
 801
 802                                 if (r->id.idiag_dport ||
 803                                     cb->args[3] > 0)
 804                                         goto next_listen;
 805
 806                                 if (inet_csk_diag_dump(sk, skb, cb, r, bc) < 0) {
 807                                         spin_unlock_bh(&ilb->lock);
 808                                         goto done;
 809                                 }
 810
 811 next_listen:
 812                                 cb->args[3] = 0;
 813                                 cb->args[4] = 0;
 814                                 ++num;
 815                         }
 816                         spin_unlock_bh(&ilb->lock);
 817
 818                         s_num = 0;
 819                         cb->args[3] = 0;
 820                         cb->args[4] = 0;
 821                 }
 822 skip_listen_ht:
 823                 cb->args[0] = 1;
 824                 s_i = num = s_num = 0;
 825         }
 826
 827         if (!(idiag_states & ~TCPF_LISTEN))
 828                 goto out;
 829
 830         for (i = s_i; i <= hashinfo->ehash_mask; i++) {
 831                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 832                 spinlock_t *lock = inet_ehash_lockp(hashinfo, i);
 833                 struct hlist_nulls_node *node;
 834                 struct sock *sk;
 835
 836                 num = 0;
 837
 838                 if (hlist_nulls_empty(&head->chain))
 839                         continue;
 840
 841                 if (i > s_i)
 842                         s_num = 0;
 843
 844                 spin_lock_bh(lock);
 845                 sk_nulls_for_each(sk, node, &head->chain) {
 846                         int state, res;
 847
 848                         if (!net_eq(sock_net(sk), net))
 849                                 continue;
 850                         if (num < s_num)
 851                                 goto next_normal;
 852                         state = (sk->sk_state == TCP_TIME_WAIT) ?
 853                                 inet_twsk(sk)->tw_substate : sk->sk_state;
 854                         if (!(idiag_states & (1 << state)))
 855                                 goto next_normal;
 856                         if (r->sdiag_family != AF_UNSPEC &&
 857                             sk->sk_family != r->sdiag_family)
 858                                 goto next_normal;
 859                         if (r->id.idiag_sport != htons(sk->sk_num) &&
 860                             r->id.idiag_sport)
 861                                 goto next_normal;
 862                         if (r->id.idiag_dport != sk->sk_dport &&
 863                             r->id.idiag_dport)
 864                                 goto next_normal;
 865                         twsk_build_assert();
 866
 867                         if (!inet_diag_bc_sk(bc, sk))
 868                                 goto next_normal;
 869
 870                         res = sk_diag_fill(sk, skb, r,
 871                                            sk_user_ns(NETLINK_CB(cb->skb).sk),
 872                                            NETLINK_CB(cb->skb).portid,
 873                                            cb->nlh->nlmsg_seq, NLM_F_MULTI,
 874                                            cb->nlh);
 875                         if (res < 0) {
 876                                 spin_unlock_bh(lock);
 877                                 goto done;
 878                         }
 879 next_normal:
 880                         ++num;
 881                 }
 882
 883                 spin_unlock_bh(lock);
 884                 cond_resched();
 885         }
 886
 887 done:
 888         cb->args[1] = i;
 889         cb->args[2] = num;
 890 out:
 891         ;
 892 }
 893 EXPORT_SYMBOL_GPL(inet_diag_dump_icsk);
 894
 895 static int __inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb,
 896                             const struct inet_diag_req_v2 *r,
 897                             struct nlattr *bc)
 898 {
 899         const struct inet_diag_handler *handler;
 900         int err = 0;
 901
 902         handler = inet_diag_lock_handler(r->sdiag_protocol);
 903         if (!IS_ERR(handler))
 904                 handler->dump(skb, cb, r, bc);
 905         else
 906                 err = PTR_ERR(handler);
 907         inet_diag_unlock_handler(handler);
 908
 909         return err ? : skb->len;
 910 }
 911
 912 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 913 {
 914         int hdrlen = sizeof(struct inet_diag_req_v2);
 915         struct nlattr *bc = NULL;
 916
 917         if (nlmsg_attrlen(cb->nlh, hdrlen))
 918                 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
 919
 920         return __inet_diag_dump(skb, cb, nlmsg_data(cb->nlh), bc);
 921 }
 922
 923 static int inet_diag_type2proto(int type)
 924 {
 925         switch (type) {
 926         case TCPDIAG_GETSOCK:
 927                 return IPPROTO_TCP;
 928         case DCCPDIAG_GETSOCK:
 929                 return IPPROTO_DCCP;
 930         default:
 931                 return 0;
 932         }
 933 }
 934
 935 static int inet_diag_dump_compat(struct sk_buff *skb,
 936                                  struct netlink_callback *cb)
 937 {
 938         struct inet_diag_req *rc = nlmsg_data(cb->nlh);
 939         int hdrlen = sizeof(struct inet_diag_req);
 940         struct inet_diag_req_v2 req;
 941         struct nlattr *bc = NULL;
 942
 943         req.sdiag_family = AF_UNSPEC; /* compatibility */
 944         req.sdiag_protocol = inet_diag_type2proto(cb->nlh->nlmsg_type);
 945         req.idiag_ext = rc->idiag_ext;
 946         req.idiag_states = rc->idiag_states;
 947         req.id = rc->id;
 948
 949         if (nlmsg_attrlen(cb->nlh, hdrlen))
 950                 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
 951
 952         return __inet_diag_dump(skb, cb, &req, bc);
 953 }
 954
 955 static int inet_diag_get_exact_compat(struct sk_buff *in_skb,
 956                                       const struct nlmsghdr *nlh)
 957 {
 958         struct inet_diag_req *rc = nlmsg_data(nlh);
 959         struct inet_diag_req_v2 req;
 960
 961         req.sdiag_family = rc->idiag_family;
 962         req.sdiag_protocol = inet_diag_type2proto(nlh->nlmsg_type);
 963         req.idiag_ext = rc->idiag_ext;
 964         req.idiag_states = rc->idiag_states;
 965         req.id = rc->id;
 966
 967         return inet_diag_cmd_exact(SOCK_DIAG_BY_FAMILY, in_skb, nlh, &req);
 968 }
 969
 970 static int inet_diag_rcv_msg_compat(struct sk_buff *skb, struct nlmsghdr *nlh)
 971 {
 972         int hdrlen = sizeof(struct inet_diag_req);
 973         struct net *net = sock_net(skb->sk);
 974
 975         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
 976             nlmsg_len(nlh) < hdrlen)
 977                 return -EINVAL;
 978
 979         if (nlh->nlmsg_flags & NLM_F_DUMP) {
 980                 if (nlmsg_attrlen(nlh, hdrlen)) {
 981                         struct nlattr *attr;
 982
 983                         attr = nlmsg_find_attr(nlh, hdrlen,
 984                                                INET_DIAG_REQ_BYTECODE);
 985                         if (!attr ||
 986                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
 987                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
 988                                 return -EINVAL;
 989                 }
 990                 {
 991                         struct netlink_dump_control c = {
 992                                 .dump = inet_diag_dump_compat,
 993                         };
 994                         return netlink_dump_start(net->diag_nlsk, skb, nlh, &c);
 995                 }
 996         }
 997
 998         return inet_diag_get_exact_compat(skb, nlh);
 999 }
1000
1001 static int inet_diag_handler_cmd(struct sk_buff *skb, struct nlmsghdr *h)
1002 {
1003         int hdrlen = sizeof(struct inet_diag_req_v2);
1004         struct net *net = sock_net(skb->sk);
1005
1006         if (nlmsg_len(h) < hdrlen)
1007                 return -EINVAL;
1008
1009         if (h->nlmsg_type == SOCK_DIAG_BY_FAMILY &&
1010             h->nlmsg_flags & NLM_F_DUMP) {
1011                 if (nlmsg_attrlen(h, hdrlen)) {
1012                         struct nlattr *attr;
1013
1014                         attr = nlmsg_find_attr(h, hdrlen,
1015                                                INET_DIAG_REQ_BYTECODE);
1016                         if (!attr ||
1017                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
1018                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
1019                                 return -EINVAL;
1020                 }
1021                 {
1022                         struct netlink_dump_control c = {
1023                                 .dump = inet_diag_dump,
1024                         };
1025                         return netlink_dump_start(net->diag_nlsk, skb, h, &c);
1026                 }
1027         }
1028
1029         return inet_diag_cmd_exact(h->nlmsg_type, skb, h, nlmsg_data(h));
1030 }
1031
1032 static
1033 int inet_diag_handler_get_info(struct sk_buff *skb, struct sock *sk)
1034 {
1035         const struct inet_diag_handler *handler;
1036         struct nlmsghdr *nlh;
1037         struct nlattr *attr;
1038         struct inet_diag_msg *r;
1039         void *info = NULL;
1040         int err = 0;
1041
1042         nlh = nlmsg_put(skb, 0, 0, SOCK_DIAG_BY_FAMILY, sizeof(*r), 0);
1043         if (!nlh)
1044                 return -ENOMEM;
1045
1046         r = nlmsg_data(nlh);
1047         memset(r, 0, sizeof(*r));
1048         inet_diag_msg_common_fill(r, sk);
1049         if (sk->sk_type == SOCK_DGRAM || sk->sk_type == SOCK_STREAM)
1050                 r->id.idiag_sport = inet_sk(sk)->inet_sport;
1051         r->idiag_state = sk->sk_state;
1052
1053         if ((err = nla_put_u8(skb, INET_DIAG_PROTOCOL, sk->sk_protocol))) {
1054                 nlmsg_cancel(skb, nlh);
1055                 return err;
1056         }
1057
1058         handler = inet_diag_lock_handler(sk->sk_protocol);
1059         if (IS_ERR(handler)) {
1060                 inet_diag_unlock_handler(handler);
1061                 nlmsg_cancel(skb, nlh);
1062                 return PTR_ERR(handler);
1063         }
1064
1065         attr = handler->idiag_info_size
1066                 ? nla_reserve(skb, INET_DIAG_INFO, handler->idiag_info_size)
1067                 : NULL;
1068         if (attr)
1069                 info = nla_data(attr);
1070
1071         handler->idiag_get_info(sk, r, info);
1072         inet_diag_unlock_handler(handler);
1073
1074         nlmsg_end(skb, nlh);
1075         return 0;
1076 }
1077
1078 static const struct sock_diag_handler inet_diag_handler = {
1079         .family = AF_INET,
1080         .dump = inet_diag_handler_cmd,
1081         .get_info = inet_diag_handler_get_info,
1082         .destroy = inet_diag_handler_cmd,
1083 };
1084
1085 static const struct sock_diag_handler inet6_diag_handler = {
1086         .family = AF_INET6,
1087         .dump = inet_diag_handler_cmd,
1088         .get_info = inet_diag_handler_get_info,
1089         .destroy = inet_diag_handler_cmd,
1090 };
1091
1092 int inet_diag_register(const struct inet_diag_handler *h)
1093 {
1094         const __u16 type = h->idiag_type;
1095         int err = -EINVAL;
1096
1097         if (type >= IPPROTO_MAX)
1098                 goto out;
1099
1100         mutex_lock(&inet_diag_table_mutex);
1101         err = -EEXIST;
1102         if (!inet_diag_table[type]) {
1103                 inet_diag_table[type] = h;
1104                 err = 0;
1105         }
1106         mutex_unlock(&inet_diag_table_mutex);
1107 out:
1108         return err;
1109 }
1110 EXPORT_SYMBOL_GPL(inet_diag_register);
1111
1112 void inet_diag_unregister(const struct inet_diag_handler *h)
1113 {
1114         const __u16 type = h->idiag_type;
1115
1116         if (type >= IPPROTO_MAX)
1117                 return;
1118
1119         mutex_lock(&inet_diag_table_mutex);
1120         inet_diag_table[type] = NULL;
1121         mutex_unlock(&inet_diag_table_mutex);
1122 }
1123 EXPORT_SYMBOL_GPL(inet_diag_unregister);
1124
1125 static int __init inet_diag_init(void)
1126 {
1127         const int inet_diag_table_size = (IPPROTO_MAX *
1128                                           sizeof(struct inet_diag_handler *));
1129         int err = -ENOMEM;
1130
1131         inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
1132         if (!inet_diag_table)
1133                 goto out;
1134
1135         err = sock_diag_register(&inet_diag_handler);
1136         if (err)
1137                 goto out_free_nl;
1138
1139         err = sock_diag_register(&inet6_diag_handler);
1140         if (err)
1141                 goto out_free_inet;
1142
1143         sock_diag_register_inet_compat(inet_diag_rcv_msg_compat);
1144 out:
1145         return err;
1146
1147 out_free_inet:
1148         sock_diag_unregister(&inet_diag_handler);
1149 out_free_nl:
1150         kfree(inet_diag_table);
1151         goto out;
1152 }
1153
1154 static void __exit inet_diag_exit(void)
1155 {
1156         sock_diag_unregister(&inet6_diag_handler);
1157         sock_diag_unregister(&inet_diag_handler);
1158         sock_diag_unregister_inet_compat(inet_diag_rcv_msg_compat);
1159         kfree(inet_diag_table);
1160 }
1161
1162 module_init(inet_diag_init);
1163 module_exit(inet_diag_exit);
1164 MODULE_LICENSE("GPL");
1165 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 2 /* AF_INET */);
1166 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 10 /* AF_INET6 */);