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