net/ipv4/inet_diag.c

   1 /*
   2  * inet_diag.c  Module for monitoring INET transport protocols sockets.
   3  *
   4  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
   5  *
   6  *      This program is free software; you can redistribute it and/or
   7  *      modify it under the terms of the GNU General Public License
   8  *      as published by the Free Software Foundation; either version
   9  *      2 of the License, or (at your option) any later version.
  10  */
  11
  12 #include <linux/kernel.h>
  13 #include <linux/module.h>
  14 #include <linux/types.h>
  15 #include <linux/fcntl.h>
  16 #include <linux/random.h>
  17 #include <linux/slab.h>
  18 #include <linux/cache.h>
  19 #include <linux/init.h>
  20 #include <linux/time.h>
  21
  22 #include <net/icmp.h>
  23 #include <net/tcp.h>
  24 #include <net/ipv6.h>
  25 #include <net/inet_common.h>
  26 #include <net/inet_connection_sock.h>
  27 #include <net/inet_hashtables.h>
  28 #include <net/inet_timewait_sock.h>
  29 #include <net/inet6_hashtables.h>
  30 #include <net/netlink.h>
  31
  32 #include <linux/inet.h>
  33 #include <linux/stddef.h>
  34
  35 #include <linux/inet_diag.h>
  36 #include <linux/sock_diag.h>
  37
  38 static const struct inet_diag_handler **inet_diag_table;
  39
  40 struct inet_diag_entry {
  41         const __be32 *saddr;
  42         const __be32 *daddr;
  43         u16 sport;
  44         u16 dport;
  45         u16 family;
  46         u16 userlocks;
  47 };
  48
  49 static DEFINE_MUTEX(inet_diag_table_mutex);
  50
  51 static const struct inet_diag_handler *inet_diag_lock_handler(int proto)
  52 {
  53         if (!inet_diag_table[proto])
  54                 request_module("net-pf-%d-proto-%d-type-%d-%d", PF_NETLINK,
  55                                NETLINK_SOCK_DIAG, AF_INET, proto);
  56
  57         mutex_lock(&inet_diag_table_mutex);
  58         if (!inet_diag_table[proto])
  59                 return ERR_PTR(-ENOENT);
  60
  61         return inet_diag_table[proto];
  62 }
  63
  64 static void inet_diag_unlock_handler(const struct inet_diag_handler *handler)
  65 {
  66         mutex_unlock(&inet_diag_table_mutex);
  67 }
  68
  69 static void inet_diag_msg_common_fill(struct inet_diag_msg *r, struct sock *sk)
  70 {
  71         r->idiag_family = sk->sk_family;
  72
  73         r->id.idiag_sport = htons(sk->sk_num);
  74         r->id.idiag_dport = sk->sk_dport;
  75         r->id.idiag_if = sk->sk_bound_dev_if;
  76         sock_diag_save_cookie(sk, r->id.idiag_cookie);
  77
  78 #if IS_ENABLED(CONFIG_IPV6)
  79         if (sk->sk_family == AF_INET6) {
  80                 *(struct in6_addr *)r->id.idiag_src = sk->sk_v6_rcv_saddr;
  81                 *(struct in6_addr *)r->id.idiag_dst = sk->sk_v6_daddr;
  82         } else
  83 #endif
  84         {
  85         memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
  86         memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
  87
  88         r->id.idiag_src[0] = sk->sk_rcv_saddr;
  89         r->id.idiag_dst[0] = sk->sk_daddr;
  90         }
  91 }
  92
  93 int inet_sk_diag_fill(struct sock *sk, struct inet_connection_sock *icsk,
  94                       struct sk_buff *skb, const struct inet_diag_req_v2 *req,
  95                       struct user_namespace *user_ns,
  96                       u32 portid, u32 seq, u16 nlmsg_flags,
  97                       const struct nlmsghdr *unlh)
  98 {
  99         const struct inet_sock *inet = inet_sk(sk);
 100         const struct inet_diag_handler *handler;
 101         int ext = req->idiag_ext;
 102         struct inet_diag_msg *r;
 103         struct nlmsghdr  *nlh;
 104         struct nlattr *attr;
 105         void *info = NULL;
 106
 107         handler = inet_diag_table[req->sdiag_protocol];
 108         BUG_ON(!handler);
 109
 110         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
 111                         nlmsg_flags);
 112         if (!nlh)
 113                 return -EMSGSIZE;
 114
 115         r = nlmsg_data(nlh);
 116         BUG_ON(!sk_fullsock(sk));
 117
 118         inet_diag_msg_common_fill(r, sk);
 119         r->idiag_state = sk->sk_state;
 120         r->idiag_timer = 0;
 121         r->idiag_retrans = 0;
 122
 123         if (nla_put_u8(skb, INET_DIAG_SHUTDOWN, sk->sk_shutdown))
 124                 goto errout;
 125
 126         /* IPv6 dual-stack sockets use inet->tos for IPv4 connections,
 127          * hence this needs to be included regardless of socket family.
 128          */
 129         if (ext & (1 << (INET_DIAG_TOS - 1)))
 130                 if (nla_put_u8(skb, INET_DIAG_TOS, inet->tos) < 0)
 131                         goto errout;
 132
 133 #if IS_ENABLED(CONFIG_IPV6)
 134         if (r->idiag_family == AF_INET6) {
 135                 if (ext & (1 << (INET_DIAG_TCLASS - 1)))
 136                         if (nla_put_u8(skb, INET_DIAG_TCLASS,
 137                                        inet6_sk(sk)->tclass) < 0)
 138                                 goto errout;
 139         }
 140 #endif
 141
 142         r->idiag_uid = from_kuid_munged(user_ns, sock_i_uid(sk));
 143         r->idiag_inode = sock_i_ino(sk);
 144
 145         if (ext & (1 << (INET_DIAG_MEMINFO - 1))) {
 146                 struct inet_diag_meminfo minfo = {
 147                         .idiag_rmem = sk_rmem_alloc_get(sk),
 148                         .idiag_wmem = sk->sk_wmem_queued,
 149                         .idiag_fmem = sk->sk_forward_alloc,
 150                         .idiag_tmem = sk_wmem_alloc_get(sk),
 151                 };
 152
 153                 if (nla_put(skb, INET_DIAG_MEMINFO, sizeof(minfo), &minfo) < 0)
 154                         goto errout;
 155         }
 156
 157         if (ext & (1 << (INET_DIAG_SKMEMINFO - 1)))
 158                 if (sock_diag_put_meminfo(sk, skb, INET_DIAG_SKMEMINFO))
 159                         goto errout;
 160
 161         if (!icsk) {
 162                 handler->idiag_get_info(sk, r, NULL);
 163                 goto out;
 164         }
 165
 166 #define EXPIRES_IN_MS(tmo)  DIV_ROUND_UP((tmo - jiffies) * 1000, HZ)
 167
 168         if (icsk->icsk_pending == ICSK_TIME_RETRANS ||
 169             icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
 170             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
 171                 r->idiag_timer = 1;
 172                 r->idiag_retrans = icsk->icsk_retransmits;
 173                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 174         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 175                 r->idiag_timer = 4;
 176                 r->idiag_retrans = icsk->icsk_probes_out;
 177                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 178         } else if (timer_pending(&sk->sk_timer)) {
 179                 r->idiag_timer = 2;
 180                 r->idiag_retrans = icsk->icsk_probes_out;
 181                 r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 182         } else {
 183                 r->idiag_timer = 0;
 184                 r->idiag_expires = 0;
 185         }
 186 #undef EXPIRES_IN_MS
 187
 188         if (ext & (1 << (INET_DIAG_INFO - 1))) {
 189                 attr = nla_reserve(skb, INET_DIAG_INFO,
 190                                    sizeof(struct tcp_info));
 191                 if (!attr)
 192                         goto errout;
 193
 194                 info = nla_data(attr);
 195         }
 196
 197         if ((ext & (1 << (INET_DIAG_CONG - 1))) && icsk->icsk_ca_ops)
 198                 if (nla_put_string(skb, INET_DIAG_CONG,
 199                                    icsk->icsk_ca_ops->name) < 0)
 200                         goto errout;
 201
 202         handler->idiag_get_info(sk, r, info);
 203
 204         if (sk->sk_state < TCP_TIME_WAIT &&
 205             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 206                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 207
 208 out:
 209         nlmsg_end(skb, nlh);
 210         return 0;
 211
 212 errout:
 213         nlmsg_cancel(skb, nlh);
 214         return -EMSGSIZE;
 215 }
 216 EXPORT_SYMBOL_GPL(inet_sk_diag_fill);
 217
 218 static int inet_csk_diag_fill(struct sock *sk,
 219                               struct sk_buff *skb,
 220                               const struct inet_diag_req_v2 *req,
 221                               struct user_namespace *user_ns,
 222                               u32 portid, u32 seq, u16 nlmsg_flags,
 223                               const struct nlmsghdr *unlh)
 224 {
 225         return inet_sk_diag_fill(sk, inet_csk(sk), skb, req,
 226                                  user_ns, portid, seq, nlmsg_flags, unlh);
 227 }
 228
 229 static int inet_twsk_diag_fill(struct sock *sk,
 230                                struct sk_buff *skb,
 231                                u32 portid, u32 seq, u16 nlmsg_flags,
 232                                const struct nlmsghdr *unlh)
 233 {
 234         struct inet_timewait_sock *tw = inet_twsk(sk);
 235         struct inet_diag_msg *r;
 236         struct nlmsghdr *nlh;
 237         s32 tmo;
 238
 239         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
 240                         nlmsg_flags);
 241         if (!nlh)
 242                 return -EMSGSIZE;
 243
 244         r = nlmsg_data(nlh);
 245         BUG_ON(tw->tw_state != TCP_TIME_WAIT);
 246
 247         tmo = tw->tw_ttd - inet_tw_time_stamp();
 248         if (tmo < 0)
 249                 tmo = 0;
 250
 251         inet_diag_msg_common_fill(r, sk);
 252         r->idiag_retrans      = 0;
 253
 254         r->idiag_state        = tw->tw_substate;
 255         r->idiag_timer        = 3;
 256         r->idiag_expires      = jiffies_to_msecs(tmo);
 257         r->idiag_rqueue       = 0;
 258         r->idiag_wqueue       = 0;
 259         r->idiag_uid          = 0;
 260         r->idiag_inode        = 0;
 261
 262         nlmsg_end(skb, nlh);
 263         return 0;
 264 }
 265
 266 static int inet_req_diag_fill(struct sock *sk, struct sk_buff *skb,
 267                               u32 portid, u32 seq, u16 nlmsg_flags,
 268                               const struct nlmsghdr *unlh)
 269 {
 270         struct inet_diag_msg *r;
 271         struct nlmsghdr *nlh;
 272         long tmo;
 273
 274         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
 275                         nlmsg_flags);
 276         if (!nlh)
 277                 return -EMSGSIZE;
 278
 279         r = nlmsg_data(nlh);
 280         inet_diag_msg_common_fill(r, sk);
 281         r->idiag_state = TCP_SYN_RECV;
 282         r->idiag_timer = 1;
 283         r->idiag_retrans = inet_reqsk(sk)->num_retrans;
 284
 285         BUILD_BUG_ON(offsetof(struct inet_request_sock, ir_cookie) !=
 286                      offsetof(struct sock, sk_cookie));
 287
 288         tmo = inet_reqsk(sk)->rsk_timer.expires - jiffies;
 289         r->idiag_expires = (tmo >= 0) ? jiffies_to_msecs(tmo) : 0;
 290         r->idiag_rqueue = 0;
 291         r->idiag_wqueue = 0;
 292         r->idiag_uid    = 0;
 293         r->idiag_inode  = 0;
 294
 295         nlmsg_end(skb, nlh);
 296         return 0;
 297 }
 298
 299 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
 300                         const struct inet_diag_req_v2 *r,
 301                         struct user_namespace *user_ns,
 302                         u32 portid, u32 seq, u16 nlmsg_flags,
 303                         const struct nlmsghdr *unlh)
 304 {
 305         if (sk->sk_state == TCP_TIME_WAIT)
 306                 return inet_twsk_diag_fill(sk, skb, portid, seq,
 307                                            nlmsg_flags, unlh);
 308
 309         if (sk->sk_state == TCP_NEW_SYN_RECV)
 310                 return inet_req_diag_fill(sk, skb, portid, seq,
 311                                           nlmsg_flags, unlh);
 312
 313         return inet_csk_diag_fill(sk, skb, r, user_ns, portid, seq,
 314                                   nlmsg_flags, unlh);
 315 }
 316
 317 int inet_diag_dump_one_icsk(struct inet_hashinfo *hashinfo,
 318                             struct sk_buff *in_skb,
 319                             const struct nlmsghdr *nlh,
 320                             const struct inet_diag_req_v2 *req)
 321 {
 322         struct net *net = sock_net(in_skb->sk);
 323         struct sk_buff *rep;
 324         struct sock *sk;
 325         int err;
 326
 327         err = -EINVAL;
 328         if (req->sdiag_family == AF_INET)
 329                 sk = inet_lookup(net, hashinfo, req->id.idiag_dst[0],
 330                                  req->id.idiag_dport, req->id.idiag_src[0],
 331                                  req->id.idiag_sport, req->id.idiag_if);
 332 #if IS_ENABLED(CONFIG_IPV6)
 333         else if (req->sdiag_family == AF_INET6)
 334                 sk = inet6_lookup(net, hashinfo,
 335                                   (struct in6_addr *)req->id.idiag_dst,
 336                                   req->id.idiag_dport,
 337                                   (struct in6_addr *)req->id.idiag_src,
 338                                   req->id.idiag_sport,
 339                                   req->id.idiag_if);
 340 #endif
 341         else
 342                 goto out_nosk;
 343
 344         err = -ENOENT;
 345         if (!sk)
 346                 goto out_nosk;
 347
 348         err = sock_diag_check_cookie(sk, req->id.idiag_cookie);
 349         if (err)
 350                 goto out;
 351
 352         rep = nlmsg_new(sizeof(struct inet_diag_msg) +
 353                         sizeof(struct inet_diag_meminfo) +
 354                         sizeof(struct tcp_info) + 64, GFP_KERNEL);
 355         if (!rep) {
 356                 err = -ENOMEM;
 357                 goto out;
 358         }
 359
 360         err = sk_diag_fill(sk, rep, req,
 361                            sk_user_ns(NETLINK_CB(in_skb).sk),
 362                            NETLINK_CB(in_skb).portid,
 363                            nlh->nlmsg_seq, 0, nlh);
 364         if (err < 0) {
 365                 WARN_ON(err == -EMSGSIZE);
 366                 nlmsg_free(rep);
 367                 goto out;
 368         }
 369         err = netlink_unicast(net->diag_nlsk, rep, NETLINK_CB(in_skb).portid,
 370                               MSG_DONTWAIT);
 371         if (err > 0)
 372                 err = 0;
 373
 374 out:
 375         if (sk)
 376                 sock_gen_put(sk);
 377
 378 out_nosk:
 379         return err;
 380 }
 381 EXPORT_SYMBOL_GPL(inet_diag_dump_one_icsk);
 382
 383 static int inet_diag_get_exact(struct sk_buff *in_skb,
 384                                const struct nlmsghdr *nlh,
 385                                const struct inet_diag_req_v2 *req)
 386 {
 387         const struct inet_diag_handler *handler;
 388         int err;
 389
 390         handler = inet_diag_lock_handler(req->sdiag_protocol);
 391         if (IS_ERR(handler))
 392                 err = PTR_ERR(handler);
 393         else
 394                 err = handler->dump_one(in_skb, nlh, req);
 395         inet_diag_unlock_handler(handler);
 396
 397         return err;
 398 }
 399
 400 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
 401 {
 402         int words = bits >> 5;
 403
 404         bits &= 0x1f;
 405
 406         if (words) {
 407                 if (memcmp(a1, a2, words << 2))
 408                         return 0;
 409         }
 410         if (bits) {
 411                 __be32 w1, w2;
 412                 __be32 mask;
 413
 414                 w1 = a1[words];
 415                 w2 = a2[words];
 416
 417                 mask = htonl((0xffffffff) << (32 - bits));
 418
 419                 if ((w1 ^ w2) & mask)
 420                         return 0;
 421         }
 422
 423         return 1;
 424 }
 425
 426 static int inet_diag_bc_run(const struct nlattr *_bc,
 427                             const struct inet_diag_entry *entry)
 428 {
 429         const void *bc = nla_data(_bc);
 430         int len = nla_len(_bc);
 431
 432         while (len > 0) {
 433                 int yes = 1;
 434                 const struct inet_diag_bc_op *op = bc;
 435
 436                 switch (op->code) {
 437                 case INET_DIAG_BC_NOP:
 438                         break;
 439                 case INET_DIAG_BC_JMP:
 440                         yes = 0;
 441                         break;
 442                 case INET_DIAG_BC_S_GE:
 443                         yes = entry->sport >= op[1].no;
 444                         break;
 445                 case INET_DIAG_BC_S_LE:
 446                         yes = entry->sport <= op[1].no;
 447                         break;
 448                 case INET_DIAG_BC_D_GE:
 449                         yes = entry->dport >= op[1].no;
 450                         break;
 451                 case INET_DIAG_BC_D_LE:
 452                         yes = entry->dport <= op[1].no;
 453                         break;
 454                 case INET_DIAG_BC_AUTO:
 455                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 456                         break;
 457                 case INET_DIAG_BC_S_COND:
 458                 case INET_DIAG_BC_D_COND: {
 459                         const struct inet_diag_hostcond *cond;
 460                         const __be32 *addr;
 461
 462                         cond = (const struct inet_diag_hostcond *)(op + 1);
 463                         if (cond->port != -1 &&
 464                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 465                                              entry->sport : entry->dport)) {
 466                                 yes = 0;
 467                                 break;
 468                         }
 469
 470                         if (op->code == INET_DIAG_BC_S_COND)
 471                                 addr = entry->saddr;
 472                         else
 473                                 addr = entry->daddr;
 474
 475                         if (cond->family != AF_UNSPEC &&
 476                             cond->family != entry->family) {
 477                                 if (entry->family == AF_INET6 &&
 478                                     cond->family == AF_INET) {
 479                                         if (addr[0] == 0 && addr[1] == 0 &&
 480                                             addr[2] == htonl(0xffff) &&
 481                                             bitstring_match(addr + 3,
 482                                                             cond->addr,
 483                                                             cond->prefix_len))
 484                                                 break;
 485                                 }
 486                                 yes = 0;
 487                                 break;
 488                         }
 489
 490                         if (cond->prefix_len == 0)
 491                                 break;
 492                         if (bitstring_match(addr, cond->addr,
 493                                             cond->prefix_len))
 494                                 break;
 495                         yes = 0;
 496                         break;
 497                 }
 498                 }
 499
 500                 if (yes) {
 501                         len -= op->yes;
 502                         bc += op->yes;
 503                 } else {
 504                         len -= op->no;
 505                         bc += op->no;
 506                 }
 507         }
 508         return len == 0;
 509 }
 510
 511 /* This helper is available for all sockets (ESTABLISH, TIMEWAIT, SYN_RECV)
 512  */
 513 static void entry_fill_addrs(struct inet_diag_entry *entry,
 514                              const struct sock *sk)
 515 {
 516 #if IS_ENABLED(CONFIG_IPV6)
 517         if (sk->sk_family == AF_INET6) {
 518                 entry->saddr = sk->sk_v6_rcv_saddr.s6_addr32;
 519                 entry->daddr = sk->sk_v6_daddr.s6_addr32;
 520         } else
 521 #endif
 522         {
 523                 entry->saddr = &sk->sk_rcv_saddr;
 524                 entry->daddr = &sk->sk_daddr;
 525         }
 526 }
 527
 528 int inet_diag_bc_sk(const struct nlattr *bc, struct sock *sk)
 529 {
 530         struct inet_sock *inet = inet_sk(sk);
 531         struct inet_diag_entry entry;
 532
 533         if (!bc)
 534                 return 1;
 535
 536         entry.family = sk->sk_family;
 537         entry_fill_addrs(&entry, sk);
 538         entry.sport = inet->inet_num;
 539         entry.dport = ntohs(inet->inet_dport);
 540         entry.userlocks = sk_fullsock(sk) ? sk->sk_userlocks : 0;
 541
 542         return inet_diag_bc_run(bc, &entry);
 543 }
 544 EXPORT_SYMBOL_GPL(inet_diag_bc_sk);
 545
 546 static int valid_cc(const void *bc, int len, int cc)
 547 {
 548         while (len >= 0) {
 549                 const struct inet_diag_bc_op *op = bc;
 550
 551                 if (cc > len)
 552                         return 0;
 553                 if (cc == len)
 554                         return 1;
 555                 if (op->yes < 4 || op->yes & 3)
 556                         return 0;
 557                 len -= op->yes;
 558                 bc  += op->yes;
 559         }
 560         return 0;
 561 }
 562
 563 /* Validate an inet_diag_hostcond. */
 564 static bool valid_hostcond(const struct inet_diag_bc_op *op, int len,
 565                            int *min_len)
 566 {
 567         struct inet_diag_hostcond *cond;
 568         int addr_len;
 569
 570         /* Check hostcond space. */
 571         *min_len += sizeof(struct inet_diag_hostcond);
 572         if (len < *min_len)
 573                 return false;
 574         cond = (struct inet_diag_hostcond *)(op + 1);
 575
 576         /* Check address family and address length. */
 577         switch (cond->family) {
 578         case AF_UNSPEC:
 579                 addr_len = 0;
 580                 break;
 581         case AF_INET:
 582                 addr_len = sizeof(struct in_addr);
 583                 break;
 584         case AF_INET6:
 585                 addr_len = sizeof(struct in6_addr);
 586                 break;
 587         default:
 588                 return false;
 589         }
 590         *min_len += addr_len;
 591         if (len < *min_len)
 592                 return false;
 593
 594         /* Check prefix length (in bits) vs address length (in bytes). */
 595         if (cond->prefix_len > 8 * addr_len)
 596                 return false;
 597
 598         return true;
 599 }
 600
 601 /* Validate a port comparison operator. */
 602 static bool valid_port_comparison(const struct inet_diag_bc_op *op,
 603                                   int len, int *min_len)
 604 {
 605         /* Port comparisons put the port in a follow-on inet_diag_bc_op. */
 606         *min_len += sizeof(struct inet_diag_bc_op);
 607         if (len < *min_len)
 608                 return false;
 609         return true;
 610 }
 611
 612 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
 613 {
 614         const void *bc = bytecode;
 615         int  len = bytecode_len;
 616
 617         while (len > 0) {
 618                 int min_len = sizeof(struct inet_diag_bc_op);
 619                 const struct inet_diag_bc_op *op = bc;
 620
 621                 switch (op->code) {
 622                 case INET_DIAG_BC_S_COND:
 623                 case INET_DIAG_BC_D_COND:
 624                         if (!valid_hostcond(bc, len, &min_len))
 625                                 return -EINVAL;
 626                         break;
 627                 case INET_DIAG_BC_S_GE:
 628                 case INET_DIAG_BC_S_LE:
 629                 case INET_DIAG_BC_D_GE:
 630                 case INET_DIAG_BC_D_LE:
 631                         if (!valid_port_comparison(bc, len, &min_len))
 632                                 return -EINVAL;
 633                         break;
 634                 case INET_DIAG_BC_AUTO:
 635                 case INET_DIAG_BC_JMP:
 636                 case INET_DIAG_BC_NOP:
 637                         break;
 638                 default:
 639                         return -EINVAL;
 640                 }
 641
 642                 if (op->code != INET_DIAG_BC_NOP) {
 643                         if (op->no < min_len || op->no > len + 4 || op->no & 3)
 644                                 return -EINVAL;
 645                         if (op->no < len &&
 646                             !valid_cc(bytecode, bytecode_len, len - op->no))
 647                                 return -EINVAL;
 648                 }
 649
 650                 if (op->yes < min_len || op->yes > len + 4 || op->yes & 3)
 651                         return -EINVAL;
 652                 bc  += op->yes;
 653                 len -= op->yes;
 654         }
 655         return len == 0 ? 0 : -EINVAL;
 656 }
 657
 658 static int inet_csk_diag_dump(struct sock *sk,
 659                               struct sk_buff *skb,
 660                               struct netlink_callback *cb,
 661                               const struct inet_diag_req_v2 *r,
 662                               const struct nlattr *bc)
 663 {
 664         if (!inet_diag_bc_sk(bc, sk))
 665                 return 0;
 666
 667         return inet_csk_diag_fill(sk, skb, r,
 668                                   sk_user_ns(NETLINK_CB(cb->skb).sk),
 669                                   NETLINK_CB(cb->skb).portid,
 670                                   cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 671 }
 672
 673 static void twsk_build_assert(void)
 674 {
 675         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_family) !=
 676                      offsetof(struct sock, sk_family));
 677
 678         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_num) !=
 679                      offsetof(struct inet_sock, inet_num));
 680
 681         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_dport) !=
 682                      offsetof(struct inet_sock, inet_dport));
 683
 684         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_rcv_saddr) !=
 685                      offsetof(struct inet_sock, inet_rcv_saddr));
 686
 687         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_daddr) !=
 688                      offsetof(struct inet_sock, inet_daddr));
 689
 690 #if IS_ENABLED(CONFIG_IPV6)
 691         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_v6_rcv_saddr) !=
 692                      offsetof(struct sock, sk_v6_rcv_saddr));
 693
 694         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_v6_daddr) !=
 695                      offsetof(struct sock, sk_v6_daddr));
 696 #endif
 697 }
 698
 699 static int inet_diag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 700                                struct netlink_callback *cb,
 701                                const struct inet_diag_req_v2 *r,
 702                                const struct nlattr *bc)
 703 {
 704         struct inet_connection_sock *icsk = inet_csk(sk);
 705         struct inet_sock *inet = inet_sk(sk);
 706         struct inet_diag_entry entry;
 707         int j, s_j, reqnum, s_reqnum;
 708         struct listen_sock *lopt;
 709         int err = 0;
 710
 711         s_j = cb->args[3];
 712         s_reqnum = cb->args[4];
 713
 714         if (s_j > 0)
 715                 s_j--;
 716
 717         entry.family = sk->sk_family;
 718
 719         read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 720
 721         lopt = icsk->icsk_accept_queue.listen_opt;
 722         if (!lopt || !listen_sock_qlen(lopt))
 723                 goto out;
 724
 725         if (bc) {
 726                 entry.sport = inet->inet_num;
 727                 entry.userlocks = sk->sk_userlocks;
 728         }
 729
 730         for (j = s_j; j < lopt->nr_table_entries; j++) {
 731                 struct request_sock *req, *head = lopt->syn_table[j];
 732
 733                 reqnum = 0;
 734                 for (req = head; req; reqnum++, req = req->dl_next) {
 735                         struct inet_request_sock *ireq = inet_rsk(req);
 736
 737                         if (reqnum < s_reqnum)
 738                                 continue;
 739                         if (r->id.idiag_dport != ireq->ir_rmt_port &&
 740                             r->id.idiag_dport)
 741                                 continue;
 742
 743                         if (bc) {
 744                                 /* Note: entry.sport and entry.userlocks are already set */
 745                                 entry_fill_addrs(&entry, req_to_sk(req));
 746                                 entry.dport = ntohs(ireq->ir_rmt_port);
 747
 748                                 if (!inet_diag_bc_run(bc, &entry))
 749                                         continue;
 750                         }
 751
 752                         err = inet_req_diag_fill(req_to_sk(req), skb,
 753                                                  NETLINK_CB(cb->skb).portid,
 754                                                  cb->nlh->nlmsg_seq,
 755                                                  NLM_F_MULTI, cb->nlh);
 756                         if (err < 0) {
 757                                 cb->args[3] = j + 1;
 758                                 cb->args[4] = reqnum;
 759                                 goto out;
 760                         }
 761                 }
 762
 763                 s_reqnum = 0;
 764         }
 765
 766 out:
 767         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 768
 769         return err;
 770 }
 771
 772 void inet_diag_dump_icsk(struct inet_hashinfo *hashinfo, struct sk_buff *skb,
 773                          struct netlink_callback *cb,
 774                          const struct inet_diag_req_v2 *r, struct nlattr *bc)
 775 {
 776         struct net *net = sock_net(skb->sk);
 777         int i, num, s_i, s_num;
 778
 779         s_i = cb->args[1];
 780         s_num = num = cb->args[2];
 781
 782         if (cb->args[0] == 0) {
 783                 if (!(r->idiag_states & (TCPF_LISTEN | TCPF_SYN_RECV)))
 784                         goto skip_listen_ht;
 785
 786                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 787                         struct inet_listen_hashbucket *ilb;
 788                         struct hlist_nulls_node *node;
 789                         struct sock *sk;
 790
 791                         num = 0;
 792                         ilb = &hashinfo->listening_hash[i];
 793                         spin_lock_bh(&ilb->lock);
 794                         sk_nulls_for_each(sk, node, &ilb->head) {
 795                                 struct inet_sock *inet = inet_sk(sk);
 796
 797                                 if (!net_eq(sock_net(sk), net))
 798                                         continue;
 799
 800                                 if (num < s_num) {
 801                                         num++;
 802                                         continue;
 803                                 }
 804
 805                                 if (r->sdiag_family != AF_UNSPEC &&
 806                                     sk->sk_family != r->sdiag_family)
 807                                         goto next_listen;
 808
 809                                 if (r->id.idiag_sport != inet->inet_sport &&
 810                                     r->id.idiag_sport)
 811                                         goto next_listen;
 812
 813                                 if (!(r->idiag_states & TCPF_LISTEN) ||
 814                                     r->id.idiag_dport ||
 815                                     cb->args[3] > 0)
 816                                         goto syn_recv;
 817
 818                                 if (inet_csk_diag_dump(sk, skb, cb, r, bc) < 0) {
 819                                         spin_unlock_bh(&ilb->lock);
 820                                         goto done;
 821                                 }
 822
 823 syn_recv:
 824                                 if (!(r->idiag_states & TCPF_SYN_RECV))
 825                                         goto next_listen;
 826
 827                                 if (inet_diag_dump_reqs(skb, sk, cb, r, bc) < 0) {
 828                                         spin_unlock_bh(&ilb->lock);
 829                                         goto done;
 830                                 }
 831
 832 next_listen:
 833                                 cb->args[3] = 0;
 834                                 cb->args[4] = 0;
 835                                 ++num;
 836                         }
 837                         spin_unlock_bh(&ilb->lock);
 838
 839                         s_num = 0;
 840                         cb->args[3] = 0;
 841                         cb->args[4] = 0;
 842                 }
 843 skip_listen_ht:
 844                 cb->args[0] = 1;
 845                 s_i = num = s_num = 0;
 846         }
 847
 848         if (!(r->idiag_states & ~(TCPF_LISTEN | TCPF_SYN_RECV)))
 849                 goto out;
 850
 851         for (i = s_i; i <= hashinfo->ehash_mask; i++) {
 852                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 853                 spinlock_t *lock = inet_ehash_lockp(hashinfo, i);
 854                 struct hlist_nulls_node *node;
 855                 struct sock *sk;
 856
 857                 num = 0;
 858
 859                 if (hlist_nulls_empty(&head->chain))
 860                         continue;
 861
 862                 if (i > s_i)
 863                         s_num = 0;
 864
 865                 spin_lock_bh(lock);
 866                 sk_nulls_for_each(sk, node, &head->chain) {
 867                         int state, res;
 868
 869                         if (!net_eq(sock_net(sk), net))
 870                                 continue;
 871                         if (num < s_num)
 872                                 goto next_normal;
 873                         state = (sk->sk_state == TCP_TIME_WAIT) ?
 874                                 inet_twsk(sk)->tw_substate : sk->sk_state;
 875                         if (!(r->idiag_states & (1 << state)))
 876                                 goto next_normal;
 877                         if (r->sdiag_family != AF_UNSPEC &&
 878                             sk->sk_family != r->sdiag_family)
 879                                 goto next_normal;
 880                         if (r->id.idiag_sport != htons(sk->sk_num) &&
 881                             r->id.idiag_sport)
 882                                 goto next_normal;
 883                         if (r->id.idiag_dport != sk->sk_dport &&
 884                             r->id.idiag_dport)
 885                                 goto next_normal;
 886                         twsk_build_assert();
 887
 888                         if (!inet_diag_bc_sk(bc, sk))
 889                                 goto next_normal;
 890
 891                         res = sk_diag_fill(sk, skb, r,
 892                                            sk_user_ns(NETLINK_CB(cb->skb).sk),
 893                                            NETLINK_CB(cb->skb).portid,
 894                                            cb->nlh->nlmsg_seq, NLM_F_MULTI,
 895                                            cb->nlh);
 896                         if (res < 0) {
 897                                 spin_unlock_bh(lock);
 898                                 goto done;
 899                         }
 900 next_normal:
 901                         ++num;
 902                 }
 903
 904                 spin_unlock_bh(lock);
 905         }
 906
 907 done:
 908         cb->args[1] = i;
 909         cb->args[2] = num;
 910 out:
 911         ;
 912 }
 913 EXPORT_SYMBOL_GPL(inet_diag_dump_icsk);
 914
 915 static int __inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb,
 916                             const struct inet_diag_req_v2 *r,
 917                             struct nlattr *bc)
 918 {
 919         const struct inet_diag_handler *handler;
 920         int err = 0;
 921
 922         handler = inet_diag_lock_handler(r->sdiag_protocol);
 923         if (!IS_ERR(handler))
 924                 handler->dump(skb, cb, r, bc);
 925         else
 926                 err = PTR_ERR(handler);
 927         inet_diag_unlock_handler(handler);
 928
 929         return err ? : skb->len;
 930 }
 931
 932 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 933 {
 934         int hdrlen = sizeof(struct inet_diag_req_v2);
 935         struct nlattr *bc = NULL;
 936
 937         if (nlmsg_attrlen(cb->nlh, hdrlen))
 938                 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
 939
 940         return __inet_diag_dump(skb, cb, nlmsg_data(cb->nlh), bc);
 941 }
 942
 943 static int inet_diag_type2proto(int type)
 944 {
 945         switch (type) {
 946         case TCPDIAG_GETSOCK:
 947                 return IPPROTO_TCP;
 948         case DCCPDIAG_GETSOCK:
 949                 return IPPROTO_DCCP;
 950         default:
 951                 return 0;
 952         }
 953 }
 954
 955 static int inet_diag_dump_compat(struct sk_buff *skb,
 956                                  struct netlink_callback *cb)
 957 {
 958         struct inet_diag_req *rc = nlmsg_data(cb->nlh);
 959         int hdrlen = sizeof(struct inet_diag_req);
 960         struct inet_diag_req_v2 req;
 961         struct nlattr *bc = NULL;
 962
 963         req.sdiag_family = AF_UNSPEC; /* compatibility */
 964         req.sdiag_protocol = inet_diag_type2proto(cb->nlh->nlmsg_type);
 965         req.idiag_ext = rc->idiag_ext;
 966         req.idiag_states = rc->idiag_states;
 967         req.id = rc->id;
 968
 969         if (nlmsg_attrlen(cb->nlh, hdrlen))
 970                 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
 971
 972         return __inet_diag_dump(skb, cb, &req, bc);
 973 }
 974
 975 static int inet_diag_get_exact_compat(struct sk_buff *in_skb,
 976                                       const struct nlmsghdr *nlh)
 977 {
 978         struct inet_diag_req *rc = nlmsg_data(nlh);
 979         struct inet_diag_req_v2 req;
 980
 981         req.sdiag_family = rc->idiag_family;
 982         req.sdiag_protocol = inet_diag_type2proto(nlh->nlmsg_type);
 983         req.idiag_ext = rc->idiag_ext;
 984         req.idiag_states = rc->idiag_states;
 985         req.id = rc->id;
 986
 987         return inet_diag_get_exact(in_skb, nlh, &req);
 988 }
 989
 990 static int inet_diag_rcv_msg_compat(struct sk_buff *skb, struct nlmsghdr *nlh)
 991 {
 992         int hdrlen = sizeof(struct inet_diag_req);
 993         struct net *net = sock_net(skb->sk);
 994
 995         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
 996             nlmsg_len(nlh) < hdrlen)
 997                 return -EINVAL;
 998
 999         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1000                 if (nlmsg_attrlen(nlh, hdrlen)) {
1001                         struct nlattr *attr;
1002
1003                         attr = nlmsg_find_attr(nlh, hdrlen,
1004                                                INET_DIAG_REQ_BYTECODE);
1005                         if (!attr ||
1006                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
1007                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
1008                                 return -EINVAL;
1009                 }
1010                 {
1011                         struct netlink_dump_control c = {
1012                                 .dump = inet_diag_dump_compat,
1013                         };
1014                         return netlink_dump_start(net->diag_nlsk, skb, nlh, &c);
1015                 }
1016         }
1017
1018         return inet_diag_get_exact_compat(skb, nlh);
1019 }
1020
1021 static int inet_diag_handler_dump(struct sk_buff *skb, struct nlmsghdr *h)
1022 {
1023         int hdrlen = sizeof(struct inet_diag_req_v2);
1024         struct net *net = sock_net(skb->sk);
1025
1026         if (nlmsg_len(h) < hdrlen)
1027                 return -EINVAL;
1028
1029         if (h->nlmsg_flags & NLM_F_DUMP) {
1030                 if (nlmsg_attrlen(h, hdrlen)) {
1031                         struct nlattr *attr;
1032
1033                         attr = nlmsg_find_attr(h, hdrlen,
1034                                                INET_DIAG_REQ_BYTECODE);
1035                         if (!attr ||
1036                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
1037                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
1038                                 return -EINVAL;
1039                 }
1040                 {
1041                         struct netlink_dump_control c = {
1042                                 .dump = inet_diag_dump,
1043                         };
1044                         return netlink_dump_start(net->diag_nlsk, skb, h, &c);
1045                 }
1046         }
1047
1048         return inet_diag_get_exact(skb, h, nlmsg_data(h));
1049 }
1050
1051 static const struct sock_diag_handler inet_diag_handler = {
1052         .family = AF_INET,
1053         .dump = inet_diag_handler_dump,
1054 };
1055
1056 static const struct sock_diag_handler inet6_diag_handler = {
1057         .family = AF_INET6,
1058         .dump = inet_diag_handler_dump,
1059 };
1060
1061 int inet_diag_register(const struct inet_diag_handler *h)
1062 {
1063         const __u16 type = h->idiag_type;
1064         int err = -EINVAL;
1065
1066         if (type >= IPPROTO_MAX)
1067                 goto out;
1068
1069         mutex_lock(&inet_diag_table_mutex);
1070         err = -EEXIST;
1071         if (!inet_diag_table[type]) {
1072                 inet_diag_table[type] = h;
1073                 err = 0;
1074         }
1075         mutex_unlock(&inet_diag_table_mutex);
1076 out:
1077         return err;
1078 }
1079 EXPORT_SYMBOL_GPL(inet_diag_register);
1080
1081 void inet_diag_unregister(const struct inet_diag_handler *h)
1082 {
1083         const __u16 type = h->idiag_type;
1084
1085         if (type >= IPPROTO_MAX)
1086                 return;
1087
1088         mutex_lock(&inet_diag_table_mutex);
1089         inet_diag_table[type] = NULL;
1090         mutex_unlock(&inet_diag_table_mutex);
1091 }
1092 EXPORT_SYMBOL_GPL(inet_diag_unregister);
1093
1094 static int __init inet_diag_init(void)
1095 {
1096         const int inet_diag_table_size = (IPPROTO_MAX *
1097                                           sizeof(struct inet_diag_handler *));
1098         int err = -ENOMEM;
1099
1100         inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
1101         if (!inet_diag_table)
1102                 goto out;
1103
1104         err = sock_diag_register(&inet_diag_handler);
1105         if (err)
1106                 goto out_free_nl;
1107
1108         err = sock_diag_register(&inet6_diag_handler);
1109         if (err)
1110                 goto out_free_inet;
1111
1112         sock_diag_register_inet_compat(inet_diag_rcv_msg_compat);
1113 out:
1114         return err;
1115
1116 out_free_inet:
1117         sock_diag_unregister(&inet_diag_handler);
1118 out_free_nl:
1119         kfree(inet_diag_table);
1120         goto out;
1121 }
1122
1123 static void __exit inet_diag_exit(void)
1124 {
1125         sock_diag_unregister(&inet6_diag_handler);
1126         sock_diag_unregister(&inet_diag_handler);
1127         sock_diag_unregister_inet_compat(inet_diag_rcv_msg_compat);
1128         kfree(inet_diag_table);
1129 }
1130
1131 module_init(inet_diag_init);
1132 module_exit(inet_diag_exit);
1133 MODULE_LICENSE("GPL");
1134 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 2 /* AF_INET */);
1135 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 10 /* AF_INET6 */);