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