net/ipv4/inet_diag.c

   1 /*
   2  * inet_diag.c  Module for monitoring INET transport protocols sockets.
   3  *
   4  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
   5  *
   6  *      This program is free software; you can redistribute it and/or
   7  *      modify it under the terms of the GNU General Public License
   8  *      as published by the Free Software Foundation; either version
   9  *      2 of the License, or (at your option) any later version.
  10  */
  11
  12 #include <linux/kernel.h>
  13 #include <linux/module.h>
  14 #include <linux/types.h>
  15 #include <linux/fcntl.h>
  16 #include <linux/random.h>
  17 #include <linux/slab.h>
  18 #include <linux/cache.h>
  19 #include <linux/init.h>
  20 #include <linux/time.h>
  21
  22 #include <net/icmp.h>
  23 #include <net/tcp.h>
  24 #include <net/ipv6.h>
  25 #include <net/inet_common.h>
  26 #include <net/inet_connection_sock.h>
  27 #include <net/inet_hashtables.h>
  28 #include <net/inet_timewait_sock.h>
  29 #include <net/inet6_hashtables.h>
  30 #include <net/netlink.h>
  31
  32 #include <linux/inet.h>
  33 #include <linux/stddef.h>
  34
  35 #include <linux/inet_diag.h>
  36 #include <linux/sock_diag.h>
  37
  38 static const struct inet_diag_handler **inet_diag_table;
  39
  40 struct inet_diag_entry {
  41         const __be32 *saddr;
  42         const __be32 *daddr;
  43         u16 sport;
  44         u16 dport;
  45         u16 family;
  46         u16 userlocks;
  47 };
  48
  49 static DEFINE_MUTEX(inet_diag_table_mutex);
  50
  51 static const struct inet_diag_handler *inet_diag_lock_handler(int proto)
  52 {
  53         if (!inet_diag_table[proto])
  54                 request_module("net-pf-%d-proto-%d-type-%d-%d", PF_NETLINK,
  55                                NETLINK_SOCK_DIAG, AF_INET, proto);
  56
  57         mutex_lock(&inet_diag_table_mutex);
  58         if (!inet_diag_table[proto])
  59                 return ERR_PTR(-ENOENT);
  60
  61         return inet_diag_table[proto];
  62 }
  63
  64 static void inet_diag_unlock_handler(const struct inet_diag_handler *handler)
  65 {
  66         mutex_unlock(&inet_diag_table_mutex);
  67 }
  68
  69 static void inet_diag_msg_common_fill(struct inet_diag_msg *r, struct sock *sk)
  70 {
  71         r->idiag_family = sk->sk_family;
  72
  73         r->id.idiag_sport = htons(sk->sk_num);
  74         r->id.idiag_dport = sk->sk_dport;
  75         r->id.idiag_if = sk->sk_bound_dev_if;
  76         sock_diag_save_cookie(sk, r->id.idiag_cookie);
  77
  78 #if IS_ENABLED(CONFIG_IPV6)
  79         if (sk->sk_family == AF_INET6) {
  80                 *(struct in6_addr *)r->id.idiag_src = sk->sk_v6_rcv_saddr;
  81                 *(struct in6_addr *)r->id.idiag_dst = sk->sk_v6_daddr;
  82         } else
  83 #endif
  84         {
  85         memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
  86         memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
  87
  88         r->id.idiag_src[0] = sk->sk_rcv_saddr;
  89         r->id.idiag_dst[0] = sk->sk_daddr;
  90         }
  91 }
  92
  93 static size_t inet_sk_attr_size(void)
  94 {
  95         return    nla_total_size(sizeof(struct tcp_info))
  96                 + nla_total_size(1) /* INET_DIAG_SHUTDOWN */
  97                 + nla_total_size(1) /* INET_DIAG_TOS */
  98                 + nla_total_size(1) /* INET_DIAG_TCLASS */
  99                 + nla_total_size(sizeof(struct inet_diag_meminfo))
 100                 + nla_total_size(sizeof(struct inet_diag_msg))
 101                 + nla_total_size(SK_MEMINFO_VARS * sizeof(u32))
 102                 + nla_total_size(TCP_CA_NAME_MAX)
 103                 + nla_total_size(sizeof(struct tcpvegas_info))
 104                 + 64;
 105 }
 106
 107 int inet_sk_diag_fill(struct sock *sk, struct inet_connection_sock *icsk,
 108                       struct sk_buff *skb, const struct inet_diag_req_v2 *req,
 109                       struct user_namespace *user_ns,
 110                       u32 portid, u32 seq, u16 nlmsg_flags,
 111                       const struct nlmsghdr *unlh)
 112 {
 113         const struct inet_sock *inet = inet_sk(sk);
 114         const struct inet_diag_handler *handler;
 115         int ext = req->idiag_ext;
 116         struct inet_diag_msg *r;
 117         struct nlmsghdr  *nlh;
 118         struct nlattr *attr;
 119         void *info = NULL;
 120
 121         handler = inet_diag_table[req->sdiag_protocol];
 122         BUG_ON(!handler);
 123
 124         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
 125                         nlmsg_flags);
 126         if (!nlh)
 127                 return -EMSGSIZE;
 128
 129         r = nlmsg_data(nlh);
 130         BUG_ON(!sk_fullsock(sk));
 131
 132         inet_diag_msg_common_fill(r, sk);
 133         r->idiag_state = sk->sk_state;
 134         r->idiag_timer = 0;
 135         r->idiag_retrans = 0;
 136
 137         if (nla_put_u8(skb, INET_DIAG_SHUTDOWN, sk->sk_shutdown))
 138                 goto errout;
 139
 140         /* IPv6 dual-stack sockets use inet->tos for IPv4 connections,
 141          * hence this needs to be included regardless of socket family.
 142          */
 143         if (ext & (1 << (INET_DIAG_TOS - 1)))
 144                 if (nla_put_u8(skb, INET_DIAG_TOS, inet->tos) < 0)
 145                         goto errout;
 146
 147 #if IS_ENABLED(CONFIG_IPV6)
 148         if (r->idiag_family == AF_INET6) {
 149                 if (ext & (1 << (INET_DIAG_TCLASS - 1)))
 150                         if (nla_put_u8(skb, INET_DIAG_TCLASS,
 151                                        inet6_sk(sk)->tclass) < 0)
 152                                 goto errout;
 153         }
 154 #endif
 155
 156         r->idiag_uid = from_kuid_munged(user_ns, sock_i_uid(sk));
 157         r->idiag_inode = sock_i_ino(sk);
 158
 159         if (ext & (1 << (INET_DIAG_MEMINFO - 1))) {
 160                 struct inet_diag_meminfo minfo = {
 161                         .idiag_rmem = sk_rmem_alloc_get(sk),
 162                         .idiag_wmem = sk->sk_wmem_queued,
 163                         .idiag_fmem = sk->sk_forward_alloc,
 164                         .idiag_tmem = sk_wmem_alloc_get(sk),
 165                 };
 166
 167                 if (nla_put(skb, INET_DIAG_MEMINFO, sizeof(minfo), &minfo) < 0)
 168                         goto errout;
 169         }
 170
 171         if (ext & (1 << (INET_DIAG_SKMEMINFO - 1)))
 172                 if (sock_diag_put_meminfo(sk, skb, INET_DIAG_SKMEMINFO))
 173                         goto errout;
 174
 175         if (!icsk) {
 176                 handler->idiag_get_info(sk, r, NULL);
 177                 goto out;
 178         }
 179
 180 #define EXPIRES_IN_MS(tmo)  DIV_ROUND_UP((tmo - jiffies) * 1000, HZ)
 181
 182         if (icsk->icsk_pending == ICSK_TIME_RETRANS ||
 183             icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
 184             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
 185                 r->idiag_timer = 1;
 186                 r->idiag_retrans = icsk->icsk_retransmits;
 187                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 188         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 189                 r->idiag_timer = 4;
 190                 r->idiag_retrans = icsk->icsk_probes_out;
 191                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 192         } else if (timer_pending(&sk->sk_timer)) {
 193                 r->idiag_timer = 2;
 194                 r->idiag_retrans = icsk->icsk_probes_out;
 195                 r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 196         } else {
 197                 r->idiag_timer = 0;
 198                 r->idiag_expires = 0;
 199         }
 200 #undef EXPIRES_IN_MS
 201
 202         if (ext & (1 << (INET_DIAG_INFO - 1))) {
 203                 attr = nla_reserve(skb, INET_DIAG_INFO,
 204                                    sizeof(struct tcp_info));
 205                 if (!attr)
 206                         goto errout;
 207
 208                 info = nla_data(attr);
 209         }
 210
 211         if ((ext & (1 << (INET_DIAG_CONG - 1))) && icsk->icsk_ca_ops)
 212                 if (nla_put_string(skb, INET_DIAG_CONG,
 213                                    icsk->icsk_ca_ops->name) < 0)
 214                         goto errout;
 215
 216         handler->idiag_get_info(sk, r, info);
 217
 218         if (sk->sk_state < TCP_TIME_WAIT &&
 219             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 220                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 221
 222 out:
 223         nlmsg_end(skb, nlh);
 224         return 0;
 225
 226 errout:
 227         nlmsg_cancel(skb, nlh);
 228         return -EMSGSIZE;
 229 }
 230 EXPORT_SYMBOL_GPL(inet_sk_diag_fill);
 231
 232 static int inet_csk_diag_fill(struct sock *sk,
 233                               struct sk_buff *skb,
 234                               const struct inet_diag_req_v2 *req,
 235                               struct user_namespace *user_ns,
 236                               u32 portid, u32 seq, u16 nlmsg_flags,
 237                               const struct nlmsghdr *unlh)
 238 {
 239         return inet_sk_diag_fill(sk, inet_csk(sk), skb, req,
 240                                  user_ns, portid, seq, nlmsg_flags, unlh);
 241 }
 242
 243 static int inet_twsk_diag_fill(struct sock *sk,
 244                                struct sk_buff *skb,
 245                                u32 portid, u32 seq, u16 nlmsg_flags,
 246                                const struct nlmsghdr *unlh)
 247 {
 248         struct inet_timewait_sock *tw = inet_twsk(sk);
 249         struct inet_diag_msg *r;
 250         struct nlmsghdr *nlh;
 251         long tmo;
 252
 253         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
 254                         nlmsg_flags);
 255         if (!nlh)
 256                 return -EMSGSIZE;
 257
 258         r = nlmsg_data(nlh);
 259         BUG_ON(tw->tw_state != TCP_TIME_WAIT);
 260
 261         tmo = tw->tw_timer.expires - jiffies;
 262         if (tmo < 0)
 263                 tmo = 0;
 264
 265         inet_diag_msg_common_fill(r, sk);
 266         r->idiag_retrans      = 0;
 267
 268         r->idiag_state        = tw->tw_substate;
 269         r->idiag_timer        = 3;
 270         r->idiag_expires      = jiffies_to_msecs(tmo);
 271         r->idiag_rqueue       = 0;
 272         r->idiag_wqueue       = 0;
 273         r->idiag_uid          = 0;
 274         r->idiag_inode        = 0;
 275
 276         nlmsg_end(skb, nlh);
 277         return 0;
 278 }
 279
 280 static int inet_req_diag_fill(struct sock *sk, struct sk_buff *skb,
 281                               u32 portid, u32 seq, u16 nlmsg_flags,
 282                               const struct nlmsghdr *unlh)
 283 {
 284         struct inet_diag_msg *r;
 285         struct nlmsghdr *nlh;
 286         long tmo;
 287
 288         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
 289                         nlmsg_flags);
 290         if (!nlh)
 291                 return -EMSGSIZE;
 292
 293         r = nlmsg_data(nlh);
 294         inet_diag_msg_common_fill(r, sk);
 295         r->idiag_state = TCP_SYN_RECV;
 296         r->idiag_timer = 1;
 297         r->idiag_retrans = inet_reqsk(sk)->num_retrans;
 298
 299         BUILD_BUG_ON(offsetof(struct inet_request_sock, ir_cookie) !=
 300                      offsetof(struct sock, sk_cookie));
 301
 302         tmo = inet_reqsk(sk)->rsk_timer.expires - jiffies;
 303         r->idiag_expires = (tmo >= 0) ? jiffies_to_msecs(tmo) : 0;
 304         r->idiag_rqueue = 0;
 305         r->idiag_wqueue = 0;
 306         r->idiag_uid    = 0;
 307         r->idiag_inode  = 0;
 308
 309         nlmsg_end(skb, nlh);
 310         return 0;
 311 }
 312
 313 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
 314                         const struct inet_diag_req_v2 *r,
 315                         struct user_namespace *user_ns,
 316                         u32 portid, u32 seq, u16 nlmsg_flags,
 317                         const struct nlmsghdr *unlh)
 318 {
 319         if (sk->sk_state == TCP_TIME_WAIT)
 320                 return inet_twsk_diag_fill(sk, skb, portid, seq,
 321                                            nlmsg_flags, unlh);
 322
 323         if (sk->sk_state == TCP_NEW_SYN_RECV)
 324                 return inet_req_diag_fill(sk, skb, portid, seq,
 325                                           nlmsg_flags, unlh);
 326
 327         return inet_csk_diag_fill(sk, skb, r, user_ns, portid, seq,
 328                                   nlmsg_flags, unlh);
 329 }
 330
 331 int inet_diag_dump_one_icsk(struct inet_hashinfo *hashinfo,
 332                             struct sk_buff *in_skb,
 333                             const struct nlmsghdr *nlh,
 334                             const struct inet_diag_req_v2 *req)
 335 {
 336         struct net *net = sock_net(in_skb->sk);
 337         struct sk_buff *rep;
 338         struct sock *sk;
 339         int err;
 340
 341         err = -EINVAL;
 342         if (req->sdiag_family == AF_INET)
 343                 sk = inet_lookup(net, hashinfo, req->id.idiag_dst[0],
 344                                  req->id.idiag_dport, req->id.idiag_src[0],
 345                                  req->id.idiag_sport, req->id.idiag_if);
 346 #if IS_ENABLED(CONFIG_IPV6)
 347         else if (req->sdiag_family == AF_INET6)
 348                 sk = inet6_lookup(net, hashinfo,
 349                                   (struct in6_addr *)req->id.idiag_dst,
 350                                   req->id.idiag_dport,
 351                                   (struct in6_addr *)req->id.idiag_src,
 352                                   req->id.idiag_sport,
 353                                   req->id.idiag_if);
 354 #endif
 355         else
 356                 goto out_nosk;
 357
 358         err = -ENOENT;
 359         if (!sk)
 360                 goto out_nosk;
 361
 362         err = sock_diag_check_cookie(sk, req->id.idiag_cookie);
 363         if (err)
 364                 goto out;
 365
 366         rep = nlmsg_new(inet_sk_attr_size(), GFP_KERNEL);
 367         if (!rep) {
 368                 err = -ENOMEM;
 369                 goto out;
 370         }
 371
 372         err = sk_diag_fill(sk, rep, req,
 373                            sk_user_ns(NETLINK_CB(in_skb).sk),
 374                            NETLINK_CB(in_skb).portid,
 375                            nlh->nlmsg_seq, 0, nlh);
 376         if (err < 0) {
 377                 WARN_ON(err == -EMSGSIZE);
 378                 nlmsg_free(rep);
 379                 goto out;
 380         }
 381         err = netlink_unicast(net->diag_nlsk, rep, NETLINK_CB(in_skb).portid,
 382                               MSG_DONTWAIT);
 383         if (err > 0)
 384                 err = 0;
 385
 386 out:
 387         if (sk)
 388                 sock_gen_put(sk);
 389
 390 out_nosk:
 391         return err;
 392 }
 393 EXPORT_SYMBOL_GPL(inet_diag_dump_one_icsk);
 394
 395 static int inet_diag_get_exact(struct sk_buff *in_skb,
 396                                const struct nlmsghdr *nlh,
 397                                const struct inet_diag_req_v2 *req)
 398 {
 399         const struct inet_diag_handler *handler;
 400         int err;
 401
 402         handler = inet_diag_lock_handler(req->sdiag_protocol);
 403         if (IS_ERR(handler))
 404                 err = PTR_ERR(handler);
 405         else
 406                 err = handler->dump_one(in_skb, nlh, req);
 407         inet_diag_unlock_handler(handler);
 408
 409         return err;
 410 }
 411
 412 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
 413 {
 414         int words = bits >> 5;
 415
 416         bits &= 0x1f;
 417
 418         if (words) {
 419                 if (memcmp(a1, a2, words << 2))
 420                         return 0;
 421         }
 422         if (bits) {
 423                 __be32 w1, w2;
 424                 __be32 mask;
 425
 426                 w1 = a1[words];
 427                 w2 = a2[words];
 428
 429                 mask = htonl((0xffffffff) << (32 - bits));
 430
 431                 if ((w1 ^ w2) & mask)
 432                         return 0;
 433         }
 434
 435         return 1;
 436 }
 437
 438 static int inet_diag_bc_run(const struct nlattr *_bc,
 439                             const struct inet_diag_entry *entry)
 440 {
 441         const void *bc = nla_data(_bc);
 442         int len = nla_len(_bc);
 443
 444         while (len > 0) {
 445                 int yes = 1;
 446                 const struct inet_diag_bc_op *op = bc;
 447
 448                 switch (op->code) {
 449                 case INET_DIAG_BC_NOP:
 450                         break;
 451                 case INET_DIAG_BC_JMP:
 452                         yes = 0;
 453                         break;
 454                 case INET_DIAG_BC_S_GE:
 455                         yes = entry->sport >= op[1].no;
 456                         break;
 457                 case INET_DIAG_BC_S_LE:
 458                         yes = entry->sport <= op[1].no;
 459                         break;
 460                 case INET_DIAG_BC_D_GE:
 461                         yes = entry->dport >= op[1].no;
 462                         break;
 463                 case INET_DIAG_BC_D_LE:
 464                         yes = entry->dport <= op[1].no;
 465                         break;
 466                 case INET_DIAG_BC_AUTO:
 467                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 468                         break;
 469                 case INET_DIAG_BC_S_COND:
 470                 case INET_DIAG_BC_D_COND: {
 471                         const struct inet_diag_hostcond *cond;
 472                         const __be32 *addr;
 473
 474                         cond = (const struct inet_diag_hostcond *)(op + 1);
 475                         if (cond->port != -1 &&
 476                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 477                                              entry->sport : entry->dport)) {
 478                                 yes = 0;
 479                                 break;
 480                         }
 481
 482                         if (op->code == INET_DIAG_BC_S_COND)
 483                                 addr = entry->saddr;
 484                         else
 485                                 addr = entry->daddr;
 486
 487                         if (cond->family != AF_UNSPEC &&
 488                             cond->family != entry->family) {
 489                                 if (entry->family == AF_INET6 &&
 490                                     cond->family == AF_INET) {
 491                                         if (addr[0] == 0 && addr[1] == 0 &&
 492                                             addr[2] == htonl(0xffff) &&
 493                                             bitstring_match(addr + 3,
 494                                                             cond->addr,
 495                                                             cond->prefix_len))
 496                                                 break;
 497                                 }
 498                                 yes = 0;
 499                                 break;
 500                         }
 501
 502                         if (cond->prefix_len == 0)
 503                                 break;
 504                         if (bitstring_match(addr, cond->addr,
 505                                             cond->prefix_len))
 506                                 break;
 507                         yes = 0;
 508                         break;
 509                 }
 510                 }
 511
 512                 if (yes) {
 513                         len -= op->yes;
 514                         bc += op->yes;
 515                 } else {
 516                         len -= op->no;
 517                         bc += op->no;
 518                 }
 519         }
 520         return len == 0;
 521 }
 522
 523 /* This helper is available for all sockets (ESTABLISH, TIMEWAIT, SYN_RECV)
 524  */
 525 static void entry_fill_addrs(struct inet_diag_entry *entry,
 526                              const struct sock *sk)
 527 {
 528 #if IS_ENABLED(CONFIG_IPV6)
 529         if (sk->sk_family == AF_INET6) {
 530                 entry->saddr = sk->sk_v6_rcv_saddr.s6_addr32;
 531                 entry->daddr = sk->sk_v6_daddr.s6_addr32;
 532         } else
 533 #endif
 534         {
 535                 entry->saddr = &sk->sk_rcv_saddr;
 536                 entry->daddr = &sk->sk_daddr;
 537         }
 538 }
 539
 540 int inet_diag_bc_sk(const struct nlattr *bc, struct sock *sk)
 541 {
 542         struct inet_sock *inet = inet_sk(sk);
 543         struct inet_diag_entry entry;
 544
 545         if (!bc)
 546                 return 1;
 547
 548         entry.family = sk->sk_family;
 549         entry_fill_addrs(&entry, sk);
 550         entry.sport = inet->inet_num;
 551         entry.dport = ntohs(inet->inet_dport);
 552         entry.userlocks = sk_fullsock(sk) ? sk->sk_userlocks : 0;
 553
 554         return inet_diag_bc_run(bc, &entry);
 555 }
 556 EXPORT_SYMBOL_GPL(inet_diag_bc_sk);
 557
 558 static int valid_cc(const void *bc, int len, int cc)
 559 {
 560         while (len >= 0) {
 561                 const struct inet_diag_bc_op *op = bc;
 562
 563                 if (cc > len)
 564                         return 0;
 565                 if (cc == len)
 566                         return 1;
 567                 if (op->yes < 4 || op->yes & 3)
 568                         return 0;
 569                 len -= op->yes;
 570                 bc  += op->yes;
 571         }
 572         return 0;
 573 }
 574
 575 /* Validate an inet_diag_hostcond. */
 576 static bool valid_hostcond(const struct inet_diag_bc_op *op, int len,
 577                            int *min_len)
 578 {
 579         struct inet_diag_hostcond *cond;
 580         int addr_len;
 581
 582         /* Check hostcond space. */
 583         *min_len += sizeof(struct inet_diag_hostcond);
 584         if (len < *min_len)
 585                 return false;
 586         cond = (struct inet_diag_hostcond *)(op + 1);
 587
 588         /* Check address family and address length. */
 589         switch (cond->family) {
 590         case AF_UNSPEC:
 591                 addr_len = 0;
 592                 break;
 593         case AF_INET:
 594                 addr_len = sizeof(struct in_addr);
 595                 break;
 596         case AF_INET6:
 597                 addr_len = sizeof(struct in6_addr);
 598                 break;
 599         default:
 600                 return false;
 601         }
 602         *min_len += addr_len;
 603         if (len < *min_len)
 604                 return false;
 605
 606         /* Check prefix length (in bits) vs address length (in bytes). */
 607         if (cond->prefix_len > 8 * addr_len)
 608                 return false;
 609
 610         return true;
 611 }
 612
 613 /* Validate a port comparison operator. */
 614 static bool valid_port_comparison(const struct inet_diag_bc_op *op,
 615                                   int len, int *min_len)
 616 {
 617         /* Port comparisons put the port in a follow-on inet_diag_bc_op. */
 618         *min_len += sizeof(struct inet_diag_bc_op);
 619         if (len < *min_len)
 620                 return false;
 621         return true;
 622 }
 623
 624 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
 625 {
 626         const void *bc = bytecode;
 627         int  len = bytecode_len;
 628
 629         while (len > 0) {
 630                 int min_len = sizeof(struct inet_diag_bc_op);
 631                 const struct inet_diag_bc_op *op = bc;
 632
 633                 switch (op->code) {
 634                 case INET_DIAG_BC_S_COND:
 635                 case INET_DIAG_BC_D_COND:
 636                         if (!valid_hostcond(bc, len, &min_len))
 637                                 return -EINVAL;
 638                         break;
 639                 case INET_DIAG_BC_S_GE:
 640                 case INET_DIAG_BC_S_LE:
 641                 case INET_DIAG_BC_D_GE:
 642                 case INET_DIAG_BC_D_LE:
 643                         if (!valid_port_comparison(bc, len, &min_len))
 644                                 return -EINVAL;
 645                         break;
 646                 case INET_DIAG_BC_AUTO:
 647                 case INET_DIAG_BC_JMP:
 648                 case INET_DIAG_BC_NOP:
 649                         break;
 650                 default:
 651                         return -EINVAL;
 652                 }
 653
 654                 if (op->code != INET_DIAG_BC_NOP) {
 655                         if (op->no < min_len || op->no > len + 4 || op->no & 3)
 656                                 return -EINVAL;
 657                         if (op->no < len &&
 658                             !valid_cc(bytecode, bytecode_len, len - op->no))
 659                                 return -EINVAL;
 660                 }
 661
 662                 if (op->yes < min_len || op->yes > len + 4 || op->yes & 3)
 663                         return -EINVAL;
 664                 bc  += op->yes;
 665                 len -= op->yes;
 666         }
 667         return len == 0 ? 0 : -EINVAL;
 668 }
 669
 670 static int inet_csk_diag_dump(struct sock *sk,
 671                               struct sk_buff *skb,
 672                               struct netlink_callback *cb,
 673                               const struct inet_diag_req_v2 *r,
 674                               const struct nlattr *bc)
 675 {
 676         if (!inet_diag_bc_sk(bc, sk))
 677                 return 0;
 678
 679         return inet_csk_diag_fill(sk, skb, r,
 680                                   sk_user_ns(NETLINK_CB(cb->skb).sk),
 681                                   NETLINK_CB(cb->skb).portid,
 682                                   cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 683 }
 684
 685 static void twsk_build_assert(void)
 686 {
 687         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_family) !=
 688                      offsetof(struct sock, sk_family));
 689
 690         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_num) !=
 691                      offsetof(struct inet_sock, inet_num));
 692
 693         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_dport) !=
 694                      offsetof(struct inet_sock, inet_dport));
 695
 696         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_rcv_saddr) !=
 697                      offsetof(struct inet_sock, inet_rcv_saddr));
 698
 699         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_daddr) !=
 700                      offsetof(struct inet_sock, inet_daddr));
 701
 702 #if IS_ENABLED(CONFIG_IPV6)
 703         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_v6_rcv_saddr) !=
 704                      offsetof(struct sock, sk_v6_rcv_saddr));
 705
 706         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_v6_daddr) !=
 707                      offsetof(struct sock, sk_v6_daddr));
 708 #endif
 709 }
 710
 711 static int inet_diag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 712                                struct netlink_callback *cb,
 713                                const struct inet_diag_req_v2 *r,
 714                                const struct nlattr *bc)
 715 {
 716         struct inet_connection_sock *icsk = inet_csk(sk);
 717         struct inet_sock *inet = inet_sk(sk);
 718         struct inet_diag_entry entry;
 719         int j, s_j, reqnum, s_reqnum;
 720         struct listen_sock *lopt;
 721         int err = 0;
 722
 723         s_j = cb->args[3];
 724         s_reqnum = cb->args[4];
 725
 726         if (s_j > 0)
 727                 s_j--;
 728
 729         entry.family = sk->sk_family;
 730
 731         spin_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 732
 733         lopt = icsk->icsk_accept_queue.listen_opt;
 734         if (!lopt || !listen_sock_qlen(lopt))
 735                 goto out;
 736
 737         if (bc) {
 738                 entry.sport = inet->inet_num;
 739                 entry.userlocks = sk->sk_userlocks;
 740         }
 741
 742         for (j = s_j; j < lopt->nr_table_entries; j++) {
 743                 struct request_sock *req, *head = lopt->syn_table[j];
 744
 745                 reqnum = 0;
 746                 for (req = head; req; reqnum++, req = req->dl_next) {
 747                         struct inet_request_sock *ireq = inet_rsk(req);
 748
 749                         if (reqnum < s_reqnum)
 750                                 continue;
 751                         if (r->id.idiag_dport != ireq->ir_rmt_port &&
 752                             r->id.idiag_dport)
 753                                 continue;
 754
 755                         if (bc) {
 756                                 /* Note: entry.sport and entry.userlocks are already set */
 757                                 entry_fill_addrs(&entry, req_to_sk(req));
 758                                 entry.dport = ntohs(ireq->ir_rmt_port);
 759
 760                                 if (!inet_diag_bc_run(bc, &entry))
 761                                         continue;
 762                         }
 763
 764                         err = inet_req_diag_fill(req_to_sk(req), skb,
 765                                                  NETLINK_CB(cb->skb).portid,
 766                                                  cb->nlh->nlmsg_seq,
 767                                                  NLM_F_MULTI, cb->nlh);
 768                         if (err < 0) {
 769                                 cb->args[3] = j + 1;
 770                                 cb->args[4] = reqnum;
 771                                 goto out;
 772                         }
 773                 }
 774
 775                 s_reqnum = 0;
 776         }
 777
 778 out:
 779         spin_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 780
 781         return err;
 782 }
 783
 784 void inet_diag_dump_icsk(struct inet_hashinfo *hashinfo, struct sk_buff *skb,
 785                          struct netlink_callback *cb,
 786                          const struct inet_diag_req_v2 *r, struct nlattr *bc)
 787 {
 788         struct net *net = sock_net(skb->sk);
 789         int i, num, s_i, s_num;
 790
 791         s_i = cb->args[1];
 792         s_num = num = cb->args[2];
 793
 794         if (cb->args[0] == 0) {
 795                 if (!(r->idiag_states & (TCPF_LISTEN | TCPF_SYN_RECV)))
 796                         goto skip_listen_ht;
 797
 798                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 799                         struct inet_listen_hashbucket *ilb;
 800                         struct hlist_nulls_node *node;
 801                         struct sock *sk;
 802
 803                         num = 0;
 804                         ilb = &hashinfo->listening_hash[i];
 805                         spin_lock_bh(&ilb->lock);
 806                         sk_nulls_for_each(sk, node, &ilb->head) {
 807                                 struct inet_sock *inet = inet_sk(sk);
 808
 809                                 if (!net_eq(sock_net(sk), net))
 810                                         continue;
 811
 812                                 if (num < s_num) {
 813                                         num++;
 814                                         continue;
 815                                 }
 816
 817                                 if (r->sdiag_family != AF_UNSPEC &&
 818                                     sk->sk_family != r->sdiag_family)
 819                                         goto next_listen;
 820
 821                                 if (r->id.idiag_sport != inet->inet_sport &&
 822                                     r->id.idiag_sport)
 823                                         goto next_listen;
 824
 825                                 if (!(r->idiag_states & TCPF_LISTEN) ||
 826                                     r->id.idiag_dport ||
 827                                     cb->args[3] > 0)
 828                                         goto syn_recv;
 829
 830                                 if (inet_csk_diag_dump(sk, skb, cb, r, bc) < 0) {
 831                                         spin_unlock_bh(&ilb->lock);
 832                                         goto done;
 833                                 }
 834
 835 syn_recv:
 836                                 if (!(r->idiag_states & TCPF_SYN_RECV))
 837                                         goto next_listen;
 838
 839                                 if (inet_diag_dump_reqs(skb, sk, cb, r, bc) < 0) {
 840                                         spin_unlock_bh(&ilb->lock);
 841                                         goto done;
 842                                 }
 843
 844 next_listen:
 845                                 cb->args[3] = 0;
 846                                 cb->args[4] = 0;
 847                                 ++num;
 848                         }
 849                         spin_unlock_bh(&ilb->lock);
 850
 851                         s_num = 0;
 852                         cb->args[3] = 0;
 853                         cb->args[4] = 0;
 854                 }
 855 skip_listen_ht:
 856                 cb->args[0] = 1;
 857                 s_i = num = s_num = 0;
 858         }
 859
 860         if (!(r->idiag_states & ~(TCPF_LISTEN | TCPF_SYN_RECV)))
 861                 goto out;
 862
 863         for (i = s_i; i <= hashinfo->ehash_mask; i++) {
 864                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 865                 spinlock_t *lock = inet_ehash_lockp(hashinfo, i);
 866                 struct hlist_nulls_node *node;
 867                 struct sock *sk;
 868
 869                 num = 0;
 870
 871                 if (hlist_nulls_empty(&head->chain))
 872                         continue;
 873
 874                 if (i > s_i)
 875                         s_num = 0;
 876
 877                 spin_lock_bh(lock);
 878                 sk_nulls_for_each(sk, node, &head->chain) {
 879                         int state, res;
 880
 881                         if (!net_eq(sock_net(sk), net))
 882                                 continue;
 883                         if (num < s_num)
 884                                 goto next_normal;
 885                         state = (sk->sk_state == TCP_TIME_WAIT) ?
 886                                 inet_twsk(sk)->tw_substate : sk->sk_state;
 887                         if (!(r->idiag_states & (1 << state)))
 888                                 goto next_normal;
 889                         if (r->sdiag_family != AF_UNSPEC &&
 890                             sk->sk_family != r->sdiag_family)
 891                                 goto next_normal;
 892                         if (r->id.idiag_sport != htons(sk->sk_num) &&
 893                             r->id.idiag_sport)
 894                                 goto next_normal;
 895                         if (r->id.idiag_dport != sk->sk_dport &&
 896                             r->id.idiag_dport)
 897                                 goto next_normal;
 898                         twsk_build_assert();
 899
 900                         if (!inet_diag_bc_sk(bc, sk))
 901                                 goto next_normal;
 902
 903                         res = sk_diag_fill(sk, skb, r,
 904                                            sk_user_ns(NETLINK_CB(cb->skb).sk),
 905                                            NETLINK_CB(cb->skb).portid,
 906                                            cb->nlh->nlmsg_seq, NLM_F_MULTI,
 907                                            cb->nlh);
 908                         if (res < 0) {
 909                                 spin_unlock_bh(lock);
 910                                 goto done;
 911                         }
 912 next_normal:
 913                         ++num;
 914                 }
 915
 916                 spin_unlock_bh(lock);
 917         }
 918
 919 done:
 920         cb->args[1] = i;
 921         cb->args[2] = num;
 922 out:
 923         ;
 924 }
 925 EXPORT_SYMBOL_GPL(inet_diag_dump_icsk);
 926
 927 static int __inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb,
 928                             const struct inet_diag_req_v2 *r,
 929                             struct nlattr *bc)
 930 {
 931         const struct inet_diag_handler *handler;
 932         int err = 0;
 933
 934         handler = inet_diag_lock_handler(r->sdiag_protocol);
 935         if (!IS_ERR(handler))
 936                 handler->dump(skb, cb, r, bc);
 937         else
 938                 err = PTR_ERR(handler);
 939         inet_diag_unlock_handler(handler);
 940
 941         return err ? : skb->len;
 942 }
 943
 944 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 945 {
 946         int hdrlen = sizeof(struct inet_diag_req_v2);
 947         struct nlattr *bc = NULL;
 948
 949         if (nlmsg_attrlen(cb->nlh, hdrlen))
 950                 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
 951
 952         return __inet_diag_dump(skb, cb, nlmsg_data(cb->nlh), bc);
 953 }
 954
 955 static int inet_diag_type2proto(int type)
 956 {
 957         switch (type) {
 958         case TCPDIAG_GETSOCK:
 959                 return IPPROTO_TCP;
 960         case DCCPDIAG_GETSOCK:
 961                 return IPPROTO_DCCP;
 962         default:
 963                 return 0;
 964         }
 965 }
 966
 967 static int inet_diag_dump_compat(struct sk_buff *skb,
 968                                  struct netlink_callback *cb)
 969 {
 970         struct inet_diag_req *rc = nlmsg_data(cb->nlh);
 971         int hdrlen = sizeof(struct inet_diag_req);
 972         struct inet_diag_req_v2 req;
 973         struct nlattr *bc = NULL;
 974
 975         req.sdiag_family = AF_UNSPEC; /* compatibility */
 976         req.sdiag_protocol = inet_diag_type2proto(cb->nlh->nlmsg_type);
 977         req.idiag_ext = rc->idiag_ext;
 978         req.idiag_states = rc->idiag_states;
 979         req.id = rc->id;
 980
 981         if (nlmsg_attrlen(cb->nlh, hdrlen))
 982                 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
 983
 984         return __inet_diag_dump(skb, cb, &req, bc);
 985 }
 986
 987 static int inet_diag_get_exact_compat(struct sk_buff *in_skb,
 988                                       const struct nlmsghdr *nlh)
 989 {
 990         struct inet_diag_req *rc = nlmsg_data(nlh);
 991         struct inet_diag_req_v2 req;
 992
 993         req.sdiag_family = rc->idiag_family;
 994         req.sdiag_protocol = inet_diag_type2proto(nlh->nlmsg_type);
 995         req.idiag_ext = rc->idiag_ext;
 996         req.idiag_states = rc->idiag_states;
 997         req.id = rc->id;
 998
 999         return inet_diag_get_exact(in_skb, nlh, &req);
1000 }
1001
1002 static int inet_diag_rcv_msg_compat(struct sk_buff *skb, struct nlmsghdr *nlh)
1003 {
1004         int hdrlen = sizeof(struct inet_diag_req);
1005         struct net *net = sock_net(skb->sk);
1006
1007         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
1008             nlmsg_len(nlh) < hdrlen)
1009                 return -EINVAL;
1010
1011         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1012                 if (nlmsg_attrlen(nlh, hdrlen)) {
1013                         struct nlattr *attr;
1014
1015                         attr = nlmsg_find_attr(nlh, hdrlen,
1016                                                INET_DIAG_REQ_BYTECODE);
1017                         if (!attr ||
1018                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
1019                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
1020                                 return -EINVAL;
1021                 }
1022                 {
1023                         struct netlink_dump_control c = {
1024                                 .dump = inet_diag_dump_compat,
1025                         };
1026                         return netlink_dump_start(net->diag_nlsk, skb, nlh, &c);
1027                 }
1028         }
1029
1030         return inet_diag_get_exact_compat(skb, nlh);
1031 }
1032
1033 static int inet_diag_handler_dump(struct sk_buff *skb, struct nlmsghdr *h)
1034 {
1035         int hdrlen = sizeof(struct inet_diag_req_v2);
1036         struct net *net = sock_net(skb->sk);
1037
1038         if (nlmsg_len(h) < hdrlen)
1039                 return -EINVAL;
1040
1041         if (h->nlmsg_flags & NLM_F_DUMP) {
1042                 if (nlmsg_attrlen(h, hdrlen)) {
1043                         struct nlattr *attr;
1044
1045                         attr = nlmsg_find_attr(h, hdrlen,
1046                                                INET_DIAG_REQ_BYTECODE);
1047                         if (!attr ||
1048                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
1049                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
1050                                 return -EINVAL;
1051                 }
1052                 {
1053                         struct netlink_dump_control c = {
1054                                 .dump = inet_diag_dump,
1055                         };
1056                         return netlink_dump_start(net->diag_nlsk, skb, h, &c);
1057                 }
1058         }
1059
1060         return inet_diag_get_exact(skb, h, nlmsg_data(h));
1061 }
1062
1063 static const struct sock_diag_handler inet_diag_handler = {
1064         .family = AF_INET,
1065         .dump = inet_diag_handler_dump,
1066 };
1067
1068 static const struct sock_diag_handler inet6_diag_handler = {
1069         .family = AF_INET6,
1070         .dump = inet_diag_handler_dump,
1071 };
1072
1073 int inet_diag_register(const struct inet_diag_handler *h)
1074 {
1075         const __u16 type = h->idiag_type;
1076         int err = -EINVAL;
1077
1078         if (type >= IPPROTO_MAX)
1079                 goto out;
1080
1081         mutex_lock(&inet_diag_table_mutex);
1082         err = -EEXIST;
1083         if (!inet_diag_table[type]) {
1084                 inet_diag_table[type] = h;
1085                 err = 0;
1086         }
1087         mutex_unlock(&inet_diag_table_mutex);
1088 out:
1089         return err;
1090 }
1091 EXPORT_SYMBOL_GPL(inet_diag_register);
1092
1093 void inet_diag_unregister(const struct inet_diag_handler *h)
1094 {
1095         const __u16 type = h->idiag_type;
1096
1097         if (type >= IPPROTO_MAX)
1098                 return;
1099
1100         mutex_lock(&inet_diag_table_mutex);
1101         inet_diag_table[type] = NULL;
1102         mutex_unlock(&inet_diag_table_mutex);
1103 }
1104 EXPORT_SYMBOL_GPL(inet_diag_unregister);
1105
1106 static int __init inet_diag_init(void)
1107 {
1108         const int inet_diag_table_size = (IPPROTO_MAX *
1109                                           sizeof(struct inet_diag_handler *));
1110         int err = -ENOMEM;
1111
1112         inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
1113         if (!inet_diag_table)
1114                 goto out;
1115
1116         err = sock_diag_register(&inet_diag_handler);
1117         if (err)
1118                 goto out_free_nl;
1119
1120         err = sock_diag_register(&inet6_diag_handler);
1121         if (err)
1122                 goto out_free_inet;
1123
1124         sock_diag_register_inet_compat(inet_diag_rcv_msg_compat);
1125 out:
1126         return err;
1127
1128 out_free_inet:
1129         sock_diag_unregister(&inet_diag_handler);
1130 out_free_nl:
1131         kfree(inet_diag_table);
1132         goto out;
1133 }
1134
1135 static void __exit inet_diag_exit(void)
1136 {
1137         sock_diag_unregister(&inet6_diag_handler);
1138         sock_diag_unregister(&inet_diag_handler);
1139         sock_diag_unregister_inet_compat(inet_diag_rcv_msg_compat);
1140         kfree(inet_diag_table);
1141 }
1142
1143 module_init(inet_diag_init);
1144 module_exit(inet_diag_exit);
1145 MODULE_LICENSE("GPL");
1146 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 2 /* AF_INET */);
1147 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 10 /* AF_INET6 */);