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