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