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 #if IS_ENABLED(CONFIG_IPV6)
  48         struct in6_addr saddr_storage;  /* for IPv4-mapped-IPv6 addresses */
  49         struct in6_addr daddr_storage;  /* for IPv4-mapped-IPv6 addresses */
  50 #endif
  51 };
  52
  53 static DEFINE_MUTEX(inet_diag_table_mutex);
  54
  55 static const struct inet_diag_handler *inet_diag_lock_handler(int proto)
  56 {
  57         if (!inet_diag_table[proto])
  58                 request_module("net-pf-%d-proto-%d-type-%d-%d", PF_NETLINK,
  59                                NETLINK_SOCK_DIAG, AF_INET, proto);
  60
  61         mutex_lock(&inet_diag_table_mutex);
  62         if (!inet_diag_table[proto])
  63                 return ERR_PTR(-ENOENT);
  64
  65         return inet_diag_table[proto];
  66 }
  67
  68 static void inet_diag_unlock_handler(const struct inet_diag_handler *handler)
  69 {
  70         mutex_unlock(&inet_diag_table_mutex);
  71 }
  72
  73 int inet_sk_diag_fill(struct sock *sk, struct inet_connection_sock *icsk,
  74                       struct sk_buff *skb, const struct inet_diag_req_v2 *req,
  75                       struct user_namespace *user_ns,
  76                       u32 portid, u32 seq, u16 nlmsg_flags,
  77                       const struct nlmsghdr *unlh)
  78 {
  79         const struct inet_sock *inet = inet_sk(sk);
  80         const struct inet_diag_handler *handler;
  81         int ext = req->idiag_ext;
  82         struct inet_diag_msg *r;
  83         struct nlmsghdr  *nlh;
  84         struct nlattr *attr;
  85         void *info = NULL;
  86
  87         handler = inet_diag_table[req->sdiag_protocol];
  88         BUG_ON(!handler);
  89
  90         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
  91                         nlmsg_flags);
  92         if (!nlh)
  93                 return -EMSGSIZE;
  94
  95         r = nlmsg_data(nlh);
  96         BUG_ON(sk->sk_state == TCP_TIME_WAIT);
  97
  98         r->idiag_family = sk->sk_family;
  99         r->idiag_state = sk->sk_state;
 100         r->idiag_timer = 0;
 101         r->idiag_retrans = 0;
 102
 103         r->id.idiag_if = sk->sk_bound_dev_if;
 104         sock_diag_save_cookie(sk, r->id.idiag_cookie);
 105
 106         r->id.idiag_sport = inet->inet_sport;
 107         r->id.idiag_dport = inet->inet_dport;
 108
 109         memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
 110         memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
 111
 112         r->id.idiag_src[0] = inet->inet_rcv_saddr;
 113         r->id.idiag_dst[0] = inet->inet_daddr;
 114
 115         if (nla_put_u8(skb, INET_DIAG_SHUTDOWN, sk->sk_shutdown))
 116                 goto errout;
 117
 118         /* IPv6 dual-stack sockets use inet->tos for IPv4 connections,
 119          * hence this needs to be included regardless of socket family.
 120          */
 121         if (ext & (1 << (INET_DIAG_TOS - 1)))
 122                 if (nla_put_u8(skb, INET_DIAG_TOS, inet->tos) < 0)
 123                         goto errout;
 124
 125 #if IS_ENABLED(CONFIG_IPV6)
 126         if (r->idiag_family == AF_INET6) {
 127                 *(struct in6_addr *)r->id.idiag_src = sk->sk_v6_rcv_saddr;
 128                 *(struct in6_addr *)r->id.idiag_dst = sk->sk_v6_daddr;
 129
 130                 if (ext & (1 << (INET_DIAG_TCLASS - 1)))
 131                         if (nla_put_u8(skb, INET_DIAG_TCLASS,
 132                                        inet6_sk(sk)->tclass) < 0)
 133                                 goto errout;
 134         }
 135 #endif
 136
 137         r->idiag_uid = from_kuid_munged(user_ns, sock_i_uid(sk));
 138         r->idiag_inode = sock_i_ino(sk);
 139
 140         if (ext & (1 << (INET_DIAG_MEMINFO - 1))) {
 141                 struct inet_diag_meminfo minfo = {
 142                         .idiag_rmem = sk_rmem_alloc_get(sk),
 143                         .idiag_wmem = sk->sk_wmem_queued,
 144                         .idiag_fmem = sk->sk_forward_alloc,
 145                         .idiag_tmem = sk_wmem_alloc_get(sk),
 146                 };
 147
 148                 if (nla_put(skb, INET_DIAG_MEMINFO, sizeof(minfo), &minfo) < 0)
 149                         goto errout;
 150         }
 151
 152         if (ext & (1 << (INET_DIAG_SKMEMINFO - 1)))
 153                 if (sock_diag_put_meminfo(sk, skb, INET_DIAG_SKMEMINFO))
 154                         goto errout;
 155
 156         if (!icsk) {
 157                 handler->idiag_get_info(sk, r, NULL);
 158                 goto out;
 159         }
 160
 161 #define EXPIRES_IN_MS(tmo)  DIV_ROUND_UP((tmo - jiffies) * 1000, HZ)
 162
 163         if (icsk->icsk_pending == ICSK_TIME_RETRANS ||
 164             icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
 165             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
 166                 r->idiag_timer = 1;
 167                 r->idiag_retrans = icsk->icsk_retransmits;
 168                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 169         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 170                 r->idiag_timer = 4;
 171                 r->idiag_retrans = icsk->icsk_probes_out;
 172                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 173         } else if (timer_pending(&sk->sk_timer)) {
 174                 r->idiag_timer = 2;
 175                 r->idiag_retrans = icsk->icsk_probes_out;
 176                 r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 177         } else {
 178                 r->idiag_timer = 0;
 179                 r->idiag_expires = 0;
 180         }
 181 #undef EXPIRES_IN_MS
 182
 183         if (ext & (1 << (INET_DIAG_INFO - 1))) {
 184                 attr = nla_reserve(skb, INET_DIAG_INFO,
 185                                    sizeof(struct tcp_info));
 186                 if (!attr)
 187                         goto errout;
 188
 189                 info = nla_data(attr);
 190         }
 191
 192         if ((ext & (1 << (INET_DIAG_CONG - 1))) && icsk->icsk_ca_ops)
 193                 if (nla_put_string(skb, INET_DIAG_CONG,
 194                                    icsk->icsk_ca_ops->name) < 0)
 195                         goto errout;
 196
 197         handler->idiag_get_info(sk, r, info);
 198
 199         if (sk->sk_state < TCP_TIME_WAIT &&
 200             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 201                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 202
 203 out:
 204         nlmsg_end(skb, nlh);
 205         return 0;
 206
 207 errout:
 208         nlmsg_cancel(skb, nlh);
 209         return -EMSGSIZE;
 210 }
 211 EXPORT_SYMBOL_GPL(inet_sk_diag_fill);
 212
 213 static int inet_csk_diag_fill(struct sock *sk,
 214                               struct sk_buff *skb,
 215                               const struct inet_diag_req_v2 *req,
 216                               struct user_namespace *user_ns,
 217                               u32 portid, u32 seq, u16 nlmsg_flags,
 218                               const struct nlmsghdr *unlh)
 219 {
 220         return inet_sk_diag_fill(sk, inet_csk(sk), skb, req,
 221                                  user_ns, portid, seq, nlmsg_flags, unlh);
 222 }
 223
 224 static int inet_twsk_diag_fill(struct sock *sk,
 225                                struct sk_buff *skb,
 226                                const struct inet_diag_req_v2 *req,
 227                                u32 portid, u32 seq, u16 nlmsg_flags,
 228                                const struct nlmsghdr *unlh)
 229 {
 230         struct inet_timewait_sock *tw = inet_twsk(sk);
 231         struct inet_diag_msg *r;
 232         struct nlmsghdr *nlh;
 233         s32 tmo;
 234
 235         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
 236                         nlmsg_flags);
 237         if (!nlh)
 238                 return -EMSGSIZE;
 239
 240         r = nlmsg_data(nlh);
 241         BUG_ON(tw->tw_state != TCP_TIME_WAIT);
 242
 243         tmo = tw->tw_ttd - inet_tw_time_stamp();
 244         if (tmo < 0)
 245                 tmo = 0;
 246
 247         r->idiag_family       = tw->tw_family;
 248         r->idiag_retrans      = 0;
 249
 250         r->id.idiag_if        = tw->tw_bound_dev_if;
 251         sock_diag_save_cookie(sk, r->id.idiag_cookie);
 252
 253         r->id.idiag_sport     = tw->tw_sport;
 254         r->id.idiag_dport     = tw->tw_dport;
 255
 256         memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
 257         memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
 258
 259         r->id.idiag_src[0]    = tw->tw_rcv_saddr;
 260         r->id.idiag_dst[0]    = tw->tw_daddr;
 261
 262         r->idiag_state        = tw->tw_substate;
 263         r->idiag_timer        = 3;
 264         r->idiag_expires      = jiffies_to_msecs(tmo);
 265         r->idiag_rqueue       = 0;
 266         r->idiag_wqueue       = 0;
 267         r->idiag_uid          = 0;
 268         r->idiag_inode        = 0;
 269 #if IS_ENABLED(CONFIG_IPV6)
 270         if (tw->tw_family == AF_INET6) {
 271                 *(struct in6_addr *)r->id.idiag_src = tw->tw_v6_rcv_saddr;
 272                 *(struct in6_addr *)r->id.idiag_dst = tw->tw_v6_daddr;
 273         }
 274 #endif
 275
 276         nlmsg_end(skb, nlh);
 277         return 0;
 278 }
 279
 280 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
 281                         const struct inet_diag_req_v2 *r,
 282                         struct user_namespace *user_ns,
 283                         u32 portid, u32 seq, u16 nlmsg_flags,
 284                         const struct nlmsghdr *unlh)
 285 {
 286         if (sk->sk_state == TCP_TIME_WAIT)
 287                 return inet_twsk_diag_fill(sk, skb, r, portid, seq,
 288                                            nlmsg_flags, unlh);
 289
 290         return inet_csk_diag_fill(sk, skb, r, user_ns, portid, seq,
 291                                   nlmsg_flags, unlh);
 292 }
 293
 294 int inet_diag_dump_one_icsk(struct inet_hashinfo *hashinfo,
 295                             struct sk_buff *in_skb,
 296                             const struct nlmsghdr *nlh,
 297                             const struct inet_diag_req_v2 *req)
 298 {
 299         struct net *net = sock_net(in_skb->sk);
 300         struct sk_buff *rep;
 301         struct sock *sk;
 302         int err;
 303
 304         err = -EINVAL;
 305         if (req->sdiag_family == AF_INET)
 306                 sk = inet_lookup(net, hashinfo, req->id.idiag_dst[0],
 307                                  req->id.idiag_dport, req->id.idiag_src[0],
 308                                  req->id.idiag_sport, req->id.idiag_if);
 309 #if IS_ENABLED(CONFIG_IPV6)
 310         else if (req->sdiag_family == AF_INET6)
 311                 sk = inet6_lookup(net, hashinfo,
 312                                   (struct in6_addr *)req->id.idiag_dst,
 313                                   req->id.idiag_dport,
 314                                   (struct in6_addr *)req->id.idiag_src,
 315                                   req->id.idiag_sport,
 316                                   req->id.idiag_if);
 317 #endif
 318         else
 319                 goto out_nosk;
 320
 321         err = -ENOENT;
 322         if (!sk)
 323                 goto out_nosk;
 324
 325         err = sock_diag_check_cookie(sk, req->id.idiag_cookie);
 326         if (err)
 327                 goto out;
 328
 329         rep = nlmsg_new(sizeof(struct inet_diag_msg) +
 330                         sizeof(struct inet_diag_meminfo) +
 331                         sizeof(struct tcp_info) + 64, GFP_KERNEL);
 332         if (!rep) {
 333                 err = -ENOMEM;
 334                 goto out;
 335         }
 336
 337         err = sk_diag_fill(sk, rep, req,
 338                            sk_user_ns(NETLINK_CB(in_skb).sk),
 339                            NETLINK_CB(in_skb).portid,
 340                            nlh->nlmsg_seq, 0, nlh);
 341         if (err < 0) {
 342                 WARN_ON(err == -EMSGSIZE);
 343                 nlmsg_free(rep);
 344                 goto out;
 345         }
 346         err = netlink_unicast(net->diag_nlsk, rep, NETLINK_CB(in_skb).portid,
 347                               MSG_DONTWAIT);
 348         if (err > 0)
 349                 err = 0;
 350
 351 out:
 352         if (sk)
 353                 sock_gen_put(sk);
 354
 355 out_nosk:
 356         return err;
 357 }
 358 EXPORT_SYMBOL_GPL(inet_diag_dump_one_icsk);
 359
 360 static int inet_diag_get_exact(struct sk_buff *in_skb,
 361                                const struct nlmsghdr *nlh,
 362                                const struct inet_diag_req_v2 *req)
 363 {
 364         const struct inet_diag_handler *handler;
 365         int err;
 366
 367         handler = inet_diag_lock_handler(req->sdiag_protocol);
 368         if (IS_ERR(handler))
 369                 err = PTR_ERR(handler);
 370         else
 371                 err = handler->dump_one(in_skb, nlh, req);
 372         inet_diag_unlock_handler(handler);
 373
 374         return err;
 375 }
 376
 377 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
 378 {
 379         int words = bits >> 5;
 380
 381         bits &= 0x1f;
 382
 383         if (words) {
 384                 if (memcmp(a1, a2, words << 2))
 385                         return 0;
 386         }
 387         if (bits) {
 388                 __be32 w1, w2;
 389                 __be32 mask;
 390
 391                 w1 = a1[words];
 392                 w2 = a2[words];
 393
 394                 mask = htonl((0xffffffff) << (32 - bits));
 395
 396                 if ((w1 ^ w2) & mask)
 397                         return 0;
 398         }
 399
 400         return 1;
 401 }
 402
 403 static int inet_diag_bc_run(const struct nlattr *_bc,
 404                             const struct inet_diag_entry *entry)
 405 {
 406         const void *bc = nla_data(_bc);
 407         int len = nla_len(_bc);
 408
 409         while (len > 0) {
 410                 int yes = 1;
 411                 const struct inet_diag_bc_op *op = bc;
 412
 413                 switch (op->code) {
 414                 case INET_DIAG_BC_NOP:
 415                         break;
 416                 case INET_DIAG_BC_JMP:
 417                         yes = 0;
 418                         break;
 419                 case INET_DIAG_BC_S_GE:
 420                         yes = entry->sport >= op[1].no;
 421                         break;
 422                 case INET_DIAG_BC_S_LE:
 423                         yes = entry->sport <= op[1].no;
 424                         break;
 425                 case INET_DIAG_BC_D_GE:
 426                         yes = entry->dport >= op[1].no;
 427                         break;
 428                 case INET_DIAG_BC_D_LE:
 429                         yes = entry->dport <= op[1].no;
 430                         break;
 431                 case INET_DIAG_BC_AUTO:
 432                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 433                         break;
 434                 case INET_DIAG_BC_S_COND:
 435                 case INET_DIAG_BC_D_COND: {
 436                         const struct inet_diag_hostcond *cond;
 437                         const __be32 *addr;
 438
 439                         cond = (const struct inet_diag_hostcond *)(op + 1);
 440                         if (cond->port != -1 &&
 441                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 442                                              entry->sport : entry->dport)) {
 443                                 yes = 0;
 444                                 break;
 445                         }
 446
 447                         if (op->code == INET_DIAG_BC_S_COND)
 448                                 addr = entry->saddr;
 449                         else
 450                                 addr = entry->daddr;
 451
 452                         if (cond->family != AF_UNSPEC &&
 453                             cond->family != entry->family) {
 454                                 if (entry->family == AF_INET6 &&
 455                                     cond->family == AF_INET) {
 456                                         if (addr[0] == 0 && addr[1] == 0 &&
 457                                             addr[2] == htonl(0xffff) &&
 458                                             bitstring_match(addr + 3,
 459                                                             cond->addr,
 460                                                             cond->prefix_len))
 461                                                 break;
 462                                 }
 463                                 yes = 0;
 464                                 break;
 465                         }
 466
 467                         if (cond->prefix_len == 0)
 468                                 break;
 469                         if (bitstring_match(addr, cond->addr,
 470                                             cond->prefix_len))
 471                                 break;
 472                         yes = 0;
 473                         break;
 474                 }
 475                 }
 476
 477                 if (yes) {
 478                         len -= op->yes;
 479                         bc += op->yes;
 480                 } else {
 481                         len -= op->no;
 482                         bc += op->no;
 483                 }
 484         }
 485         return len == 0;
 486 }
 487
 488 int inet_diag_bc_sk(const struct nlattr *bc, struct sock *sk)
 489 {
 490         struct inet_sock *inet = inet_sk(sk);
 491         struct inet_diag_entry entry;
 492
 493         if (!bc)
 494                 return 1;
 495
 496         entry.family = sk->sk_family;
 497 #if IS_ENABLED(CONFIG_IPV6)
 498         if (entry.family == AF_INET6) {
 499                 entry.saddr = sk->sk_v6_rcv_saddr.s6_addr32;
 500                 entry.daddr = sk->sk_v6_daddr.s6_addr32;
 501         } else
 502 #endif
 503         {
 504                 entry.saddr = &inet->inet_rcv_saddr;
 505                 entry.daddr = &inet->inet_daddr;
 506         }
 507         entry.sport = inet->inet_num;
 508         entry.dport = ntohs(inet->inet_dport);
 509         entry.userlocks = (sk->sk_state != TCP_TIME_WAIT) ? sk->sk_userlocks : 0;
 510
 511         return inet_diag_bc_run(bc, &entry);
 512 }
 513 EXPORT_SYMBOL_GPL(inet_diag_bc_sk);
 514
 515 static int valid_cc(const void *bc, int len, int cc)
 516 {
 517         while (len >= 0) {
 518                 const struct inet_diag_bc_op *op = bc;
 519
 520                 if (cc > len)
 521                         return 0;
 522                 if (cc == len)
 523                         return 1;
 524                 if (op->yes < 4 || op->yes & 3)
 525                         return 0;
 526                 len -= op->yes;
 527                 bc  += op->yes;
 528         }
 529         return 0;
 530 }
 531
 532 /* Validate an inet_diag_hostcond. */
 533 static bool valid_hostcond(const struct inet_diag_bc_op *op, int len,
 534                            int *min_len)
 535 {
 536         struct inet_diag_hostcond *cond;
 537         int addr_len;
 538
 539         /* Check hostcond space. */
 540         *min_len += sizeof(struct inet_diag_hostcond);
 541         if (len < *min_len)
 542                 return false;
 543         cond = (struct inet_diag_hostcond *)(op + 1);
 544
 545         /* Check address family and address length. */
 546         switch (cond->family) {
 547         case AF_UNSPEC:
 548                 addr_len = 0;
 549                 break;
 550         case AF_INET:
 551                 addr_len = sizeof(struct in_addr);
 552                 break;
 553         case AF_INET6:
 554                 addr_len = sizeof(struct in6_addr);
 555                 break;
 556         default:
 557                 return false;
 558         }
 559         *min_len += addr_len;
 560         if (len < *min_len)
 561                 return false;
 562
 563         /* Check prefix length (in bits) vs address length (in bytes). */
 564         if (cond->prefix_len > 8 * addr_len)
 565                 return false;
 566
 567         return true;
 568 }
 569
 570 /* Validate a port comparison operator. */
 571 static bool valid_port_comparison(const struct inet_diag_bc_op *op,
 572                                   int len, int *min_len)
 573 {
 574         /* Port comparisons put the port in a follow-on inet_diag_bc_op. */
 575         *min_len += sizeof(struct inet_diag_bc_op);
 576         if (len < *min_len)
 577                 return false;
 578         return true;
 579 }
 580
 581 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
 582 {
 583         const void *bc = bytecode;
 584         int  len = bytecode_len;
 585
 586         while (len > 0) {
 587                 int min_len = sizeof(struct inet_diag_bc_op);
 588                 const struct inet_diag_bc_op *op = bc;
 589
 590                 switch (op->code) {
 591                 case INET_DIAG_BC_S_COND:
 592                 case INET_DIAG_BC_D_COND:
 593                         if (!valid_hostcond(bc, len, &min_len))
 594                                 return -EINVAL;
 595                         break;
 596                 case INET_DIAG_BC_S_GE:
 597                 case INET_DIAG_BC_S_LE:
 598                 case INET_DIAG_BC_D_GE:
 599                 case INET_DIAG_BC_D_LE:
 600                         if (!valid_port_comparison(bc, len, &min_len))
 601                                 return -EINVAL;
 602                         break;
 603                 case INET_DIAG_BC_AUTO:
 604                 case INET_DIAG_BC_JMP:
 605                 case INET_DIAG_BC_NOP:
 606                         break;
 607                 default:
 608                         return -EINVAL;
 609                 }
 610
 611                 if (op->code != INET_DIAG_BC_NOP) {
 612                         if (op->no < min_len || op->no > len + 4 || op->no & 3)
 613                                 return -EINVAL;
 614                         if (op->no < len &&
 615                             !valid_cc(bytecode, bytecode_len, len - op->no))
 616                                 return -EINVAL;
 617                 }
 618
 619                 if (op->yes < min_len || op->yes > len + 4 || op->yes & 3)
 620                         return -EINVAL;
 621                 bc  += op->yes;
 622                 len -= op->yes;
 623         }
 624         return len == 0 ? 0 : -EINVAL;
 625 }
 626
 627 static int inet_csk_diag_dump(struct sock *sk,
 628                               struct sk_buff *skb,
 629                               struct netlink_callback *cb,
 630                               const struct inet_diag_req_v2 *r,
 631                               const struct nlattr *bc)
 632 {
 633         if (!inet_diag_bc_sk(bc, sk))
 634                 return 0;
 635
 636         return inet_csk_diag_fill(sk, skb, r,
 637                                   sk_user_ns(NETLINK_CB(cb->skb).sk),
 638                                   NETLINK_CB(cb->skb).portid,
 639                                   cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 640 }
 641
 642 static void twsk_build_assert(void)
 643 {
 644         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_family) !=
 645                      offsetof(struct sock, sk_family));
 646
 647         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_num) !=
 648                      offsetof(struct inet_sock, inet_num));
 649
 650         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_dport) !=
 651                      offsetof(struct inet_sock, inet_dport));
 652
 653         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_rcv_saddr) !=
 654                      offsetof(struct inet_sock, inet_rcv_saddr));
 655
 656         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_daddr) !=
 657                      offsetof(struct inet_sock, inet_daddr));
 658
 659 #if IS_ENABLED(CONFIG_IPV6)
 660         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_v6_rcv_saddr) !=
 661                      offsetof(struct sock, sk_v6_rcv_saddr));
 662
 663         BUILD_BUG_ON(offsetof(struct inet_timewait_sock, tw_v6_daddr) !=
 664                      offsetof(struct sock, sk_v6_daddr));
 665 #endif
 666 }
 667
 668 static int inet_twsk_diag_dump(struct sock *sk,
 669                                struct sk_buff *skb,
 670                                struct netlink_callback *cb,
 671                                const struct inet_diag_req_v2 *r,
 672                                const struct nlattr *bc)
 673 {
 674         twsk_build_assert();
 675
 676         if (!inet_diag_bc_sk(bc, sk))
 677                 return 0;
 678
 679         return inet_twsk_diag_fill(sk, skb, r,
 680                                    NETLINK_CB(cb->skb).portid,
 681                                    cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 682 }
 683
 684 /* Get the IPv4, IPv6, or IPv4-mapped-IPv6 local and remote addresses
 685  * from a request_sock. For IPv4-mapped-IPv6 we must map IPv4 to IPv6.
 686  */
 687 static void inet_diag_req_addrs(const struct sock *sk,
 688                                 const struct request_sock *req,
 689                                 struct inet_diag_entry *entry)
 690 {
 691         const struct inet_request_sock *ireq = inet_rsk(req);
 692
 693 #if IS_ENABLED(CONFIG_IPV6)
 694         if (sk->sk_family == AF_INET6) {
 695                 if (req->rsk_ops->family == AF_INET6) {
 696                         entry->saddr = ireq->ir_v6_loc_addr.s6_addr32;
 697                         entry->daddr = ireq->ir_v6_rmt_addr.s6_addr32;
 698                 } else if (req->rsk_ops->family == AF_INET) {
 699                         ipv6_addr_set_v4mapped(ireq->ir_loc_addr,
 700                                                &entry->saddr_storage);
 701                         ipv6_addr_set_v4mapped(ireq->ir_rmt_addr,
 702                                                &entry->daddr_storage);
 703                         entry->saddr = entry->saddr_storage.s6_addr32;
 704                         entry->daddr = entry->daddr_storage.s6_addr32;
 705                 }
 706         } else
 707 #endif
 708         {
 709                 entry->saddr = &ireq->ir_loc_addr;
 710                 entry->daddr = &ireq->ir_rmt_addr;
 711         }
 712 }
 713
 714 static int inet_diag_fill_req(struct sk_buff *skb, struct sock *sk,
 715                               struct request_sock *req,
 716                               struct user_namespace *user_ns,
 717                               u32 portid, u32 seq,
 718                               const struct nlmsghdr *unlh)
 719 {
 720         const struct inet_request_sock *ireq = inet_rsk(req);
 721         struct inet_sock *inet = inet_sk(sk);
 722         struct inet_diag_msg *r;
 723         struct nlmsghdr *nlh;
 724         long tmo;
 725
 726         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
 727                         NLM_F_MULTI);
 728         if (!nlh)
 729                 return -EMSGSIZE;
 730
 731         r = nlmsg_data(nlh);
 732         r->idiag_family = sk->sk_family;
 733         r->idiag_state = TCP_SYN_RECV;
 734         r->idiag_timer = 1;
 735         r->idiag_retrans = req->num_retrans;
 736
 737         r->id.idiag_if = sk->sk_bound_dev_if;
 738
 739         BUILD_BUG_ON(offsetof(struct inet_request_sock, ir_cookie) !=
 740                      offsetof(struct sock, sk_cookie));
 741         sock_diag_save_cookie((struct sock *)ireq, r->id.idiag_cookie);
 742
 743         tmo = req->expires - jiffies;
 744         if (tmo < 0)
 745                 tmo = 0;
 746
 747         r->id.idiag_sport = inet->inet_sport;
 748         r->id.idiag_dport = ireq->ir_rmt_port;
 749
 750         memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
 751         memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
 752
 753         r->id.idiag_src[0] = ireq->ir_loc_addr;
 754         r->id.idiag_dst[0] = ireq->ir_rmt_addr;
 755
 756         r->idiag_expires = jiffies_to_msecs(tmo);
 757         r->idiag_rqueue = 0;
 758         r->idiag_wqueue = 0;
 759         r->idiag_uid = from_kuid_munged(user_ns, sock_i_uid(sk));
 760         r->idiag_inode = 0;
 761 #if IS_ENABLED(CONFIG_IPV6)
 762         if (r->idiag_family == AF_INET6) {
 763                 struct inet_diag_entry entry;
 764
 765                 inet_diag_req_addrs(sk, req, &entry);
 766                 memcpy(r->id.idiag_src, entry.saddr, sizeof(struct in6_addr));
 767                 memcpy(r->id.idiag_dst, entry.daddr, sizeof(struct in6_addr));
 768         }
 769 #endif
 770
 771         nlmsg_end(skb, nlh);
 772         return 0;
 773 }
 774
 775 static int inet_diag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 776                                struct netlink_callback *cb,
 777                                const struct inet_diag_req_v2 *r,
 778                                const struct nlattr *bc)
 779 {
 780         struct inet_connection_sock *icsk = inet_csk(sk);
 781         struct inet_sock *inet = inet_sk(sk);
 782         struct inet_diag_entry entry;
 783         int j, s_j, reqnum, s_reqnum;
 784         struct listen_sock *lopt;
 785         int err = 0;
 786
 787         s_j = cb->args[3];
 788         s_reqnum = cb->args[4];
 789
 790         if (s_j > 0)
 791                 s_j--;
 792
 793         entry.family = sk->sk_family;
 794
 795         read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 796
 797         lopt = icsk->icsk_accept_queue.listen_opt;
 798         if (!lopt || !lopt->qlen)
 799                 goto out;
 800
 801         if (bc) {
 802                 entry.sport = inet->inet_num;
 803                 entry.userlocks = sk->sk_userlocks;
 804         }
 805
 806         for (j = s_j; j < lopt->nr_table_entries; j++) {
 807                 struct request_sock *req, *head = lopt->syn_table[j];
 808
 809                 reqnum = 0;
 810                 for (req = head; req; reqnum++, req = req->dl_next) {
 811                         struct inet_request_sock *ireq = inet_rsk(req);
 812
 813                         if (reqnum < s_reqnum)
 814                                 continue;
 815                         if (r->id.idiag_dport != ireq->ir_rmt_port &&
 816                             r->id.idiag_dport)
 817                                 continue;
 818
 819                         if (bc) {
 820                                 inet_diag_req_addrs(sk, req, &entry);
 821                                 entry.dport = ntohs(ireq->ir_rmt_port);
 822
 823                                 if (!inet_diag_bc_run(bc, &entry))
 824                                         continue;
 825                         }
 826
 827                         err = inet_diag_fill_req(skb, sk, req,
 828                                                  sk_user_ns(NETLINK_CB(cb->skb).sk),
 829                                                  NETLINK_CB(cb->skb).portid,
 830                                                  cb->nlh->nlmsg_seq, cb->nlh);
 831                         if (err < 0) {
 832                                 cb->args[3] = j + 1;
 833                                 cb->args[4] = reqnum;
 834                                 goto out;
 835                         }
 836                 }
 837
 838                 s_reqnum = 0;
 839         }
 840
 841 out:
 842         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 843
 844         return err;
 845 }
 846
 847 void inet_diag_dump_icsk(struct inet_hashinfo *hashinfo, struct sk_buff *skb,
 848                          struct netlink_callback *cb,
 849                          const struct inet_diag_req_v2 *r, struct nlattr *bc)
 850 {
 851         struct net *net = sock_net(skb->sk);
 852         int i, num, s_i, s_num;
 853
 854         s_i = cb->args[1];
 855         s_num = num = cb->args[2];
 856
 857         if (cb->args[0] == 0) {
 858                 if (!(r->idiag_states & (TCPF_LISTEN | TCPF_SYN_RECV)))
 859                         goto skip_listen_ht;
 860
 861                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 862                         struct inet_listen_hashbucket *ilb;
 863                         struct hlist_nulls_node *node;
 864                         struct sock *sk;
 865
 866                         num = 0;
 867                         ilb = &hashinfo->listening_hash[i];
 868                         spin_lock_bh(&ilb->lock);
 869                         sk_nulls_for_each(sk, node, &ilb->head) {
 870                                 struct inet_sock *inet = inet_sk(sk);
 871
 872                                 if (!net_eq(sock_net(sk), net))
 873                                         continue;
 874
 875                                 if (num < s_num) {
 876                                         num++;
 877                                         continue;
 878                                 }
 879
 880                                 if (r->sdiag_family != AF_UNSPEC &&
 881                                     sk->sk_family != r->sdiag_family)
 882                                         goto next_listen;
 883
 884                                 if (r->id.idiag_sport != inet->inet_sport &&
 885                                     r->id.idiag_sport)
 886                                         goto next_listen;
 887
 888                                 if (!(r->idiag_states & TCPF_LISTEN) ||
 889                                     r->id.idiag_dport ||
 890                                     cb->args[3] > 0)
 891                                         goto syn_recv;
 892
 893                                 if (inet_csk_diag_dump(sk, skb, cb, r, bc) < 0) {
 894                                         spin_unlock_bh(&ilb->lock);
 895                                         goto done;
 896                                 }
 897
 898 syn_recv:
 899                                 if (!(r->idiag_states & TCPF_SYN_RECV))
 900                                         goto next_listen;
 901
 902                                 if (inet_diag_dump_reqs(skb, sk, cb, r, bc) < 0) {
 903                                         spin_unlock_bh(&ilb->lock);
 904                                         goto done;
 905                                 }
 906
 907 next_listen:
 908                                 cb->args[3] = 0;
 909                                 cb->args[4] = 0;
 910                                 ++num;
 911                         }
 912                         spin_unlock_bh(&ilb->lock);
 913
 914                         s_num = 0;
 915                         cb->args[3] = 0;
 916                         cb->args[4] = 0;
 917                 }
 918 skip_listen_ht:
 919                 cb->args[0] = 1;
 920                 s_i = num = s_num = 0;
 921         }
 922
 923         if (!(r->idiag_states & ~(TCPF_LISTEN | TCPF_SYN_RECV)))
 924                 goto out;
 925
 926         for (i = s_i; i <= hashinfo->ehash_mask; i++) {
 927                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 928                 spinlock_t *lock = inet_ehash_lockp(hashinfo, i);
 929                 struct hlist_nulls_node *node;
 930                 struct sock *sk;
 931
 932                 num = 0;
 933
 934                 if (hlist_nulls_empty(&head->chain))
 935                         continue;
 936
 937                 if (i > s_i)
 938                         s_num = 0;
 939
 940                 spin_lock_bh(lock);
 941                 sk_nulls_for_each(sk, node, &head->chain) {
 942                         int state, res;
 943
 944                         if (!net_eq(sock_net(sk), net))
 945                                 continue;
 946                         if (num < s_num)
 947                                 goto next_normal;
 948                         state = (sk->sk_state == TCP_TIME_WAIT) ?
 949                                 inet_twsk(sk)->tw_substate : sk->sk_state;
 950                         if (!(r->idiag_states & (1 << state)))
 951                                 goto next_normal;
 952                         if (r->sdiag_family != AF_UNSPEC &&
 953                             sk->sk_family != r->sdiag_family)
 954                                 goto next_normal;
 955                         if (r->id.idiag_sport != htons(sk->sk_num) &&
 956                             r->id.idiag_sport)
 957                                 goto next_normal;
 958                         if (r->id.idiag_dport != sk->sk_dport &&
 959                             r->id.idiag_dport)
 960                                 goto next_normal;
 961                         if (sk->sk_state == TCP_TIME_WAIT)
 962                                 res = inet_twsk_diag_dump(sk, skb, cb, r, bc);
 963                         else
 964                                 res = inet_csk_diag_dump(sk, skb, cb, r, bc);
 965                         if (res < 0) {
 966                                 spin_unlock_bh(lock);
 967                                 goto done;
 968                         }
 969 next_normal:
 970                         ++num;
 971                 }
 972
 973                 spin_unlock_bh(lock);
 974         }
 975
 976 done:
 977         cb->args[1] = i;
 978         cb->args[2] = num;
 979 out:
 980         ;
 981 }
 982 EXPORT_SYMBOL_GPL(inet_diag_dump_icsk);
 983
 984 static int __inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb,
 985                             const struct inet_diag_req_v2 *r,
 986                             struct nlattr *bc)
 987 {
 988         const struct inet_diag_handler *handler;
 989         int err = 0;
 990
 991         handler = inet_diag_lock_handler(r->sdiag_protocol);
 992         if (!IS_ERR(handler))
 993                 handler->dump(skb, cb, r, bc);
 994         else
 995                 err = PTR_ERR(handler);
 996         inet_diag_unlock_handler(handler);
 997
 998         return err ? : skb->len;
 999 }
1000
1001 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
1002 {
1003         int hdrlen = sizeof(struct inet_diag_req_v2);
1004         struct nlattr *bc = NULL;
1005
1006         if (nlmsg_attrlen(cb->nlh, hdrlen))
1007                 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
1008
1009         return __inet_diag_dump(skb, cb, nlmsg_data(cb->nlh), bc);
1010 }
1011
1012 static int inet_diag_type2proto(int type)
1013 {
1014         switch (type) {
1015         case TCPDIAG_GETSOCK:
1016                 return IPPROTO_TCP;
1017         case DCCPDIAG_GETSOCK:
1018                 return IPPROTO_DCCP;
1019         default:
1020                 return 0;
1021         }
1022 }
1023
1024 static int inet_diag_dump_compat(struct sk_buff *skb,
1025                                  struct netlink_callback *cb)
1026 {
1027         struct inet_diag_req *rc = nlmsg_data(cb->nlh);
1028         int hdrlen = sizeof(struct inet_diag_req);
1029         struct inet_diag_req_v2 req;
1030         struct nlattr *bc = NULL;
1031
1032         req.sdiag_family = AF_UNSPEC; /* compatibility */
1033         req.sdiag_protocol = inet_diag_type2proto(cb->nlh->nlmsg_type);
1034         req.idiag_ext = rc->idiag_ext;
1035         req.idiag_states = rc->idiag_states;
1036         req.id = rc->id;
1037
1038         if (nlmsg_attrlen(cb->nlh, hdrlen))
1039                 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
1040
1041         return __inet_diag_dump(skb, cb, &req, bc);
1042 }
1043
1044 static int inet_diag_get_exact_compat(struct sk_buff *in_skb,
1045                                       const struct nlmsghdr *nlh)
1046 {
1047         struct inet_diag_req *rc = nlmsg_data(nlh);
1048         struct inet_diag_req_v2 req;
1049
1050         req.sdiag_family = rc->idiag_family;
1051         req.sdiag_protocol = inet_diag_type2proto(nlh->nlmsg_type);
1052         req.idiag_ext = rc->idiag_ext;
1053         req.idiag_states = rc->idiag_states;
1054         req.id = rc->id;
1055
1056         return inet_diag_get_exact(in_skb, nlh, &req);
1057 }
1058
1059 static int inet_diag_rcv_msg_compat(struct sk_buff *skb, struct nlmsghdr *nlh)
1060 {
1061         int hdrlen = sizeof(struct inet_diag_req);
1062         struct net *net = sock_net(skb->sk);
1063
1064         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
1065             nlmsg_len(nlh) < hdrlen)
1066                 return -EINVAL;
1067
1068         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1069                 if (nlmsg_attrlen(nlh, hdrlen)) {
1070                         struct nlattr *attr;
1071
1072                         attr = nlmsg_find_attr(nlh, hdrlen,
1073                                                INET_DIAG_REQ_BYTECODE);
1074                         if (!attr ||
1075                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
1076                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
1077                                 return -EINVAL;
1078                 }
1079                 {
1080                         struct netlink_dump_control c = {
1081                                 .dump = inet_diag_dump_compat,
1082                         };
1083                         return netlink_dump_start(net->diag_nlsk, skb, nlh, &c);
1084                 }
1085         }
1086
1087         return inet_diag_get_exact_compat(skb, nlh);
1088 }
1089
1090 static int inet_diag_handler_dump(struct sk_buff *skb, struct nlmsghdr *h)
1091 {
1092         int hdrlen = sizeof(struct inet_diag_req_v2);
1093         struct net *net = sock_net(skb->sk);
1094
1095         if (nlmsg_len(h) < hdrlen)
1096                 return -EINVAL;
1097
1098         if (h->nlmsg_flags & NLM_F_DUMP) {
1099                 if (nlmsg_attrlen(h, hdrlen)) {
1100                         struct nlattr *attr;
1101
1102                         attr = nlmsg_find_attr(h, hdrlen,
1103                                                INET_DIAG_REQ_BYTECODE);
1104                         if (!attr ||
1105                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
1106                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
1107                                 return -EINVAL;
1108                 }
1109                 {
1110                         struct netlink_dump_control c = {
1111                                 .dump = inet_diag_dump,
1112                         };
1113                         return netlink_dump_start(net->diag_nlsk, skb, h, &c);
1114                 }
1115         }
1116
1117         return inet_diag_get_exact(skb, h, nlmsg_data(h));
1118 }
1119
1120 static const struct sock_diag_handler inet_diag_handler = {
1121         .family = AF_INET,
1122         .dump = inet_diag_handler_dump,
1123 };
1124
1125 static const struct sock_diag_handler inet6_diag_handler = {
1126         .family = AF_INET6,
1127         .dump = inet_diag_handler_dump,
1128 };
1129
1130 int inet_diag_register(const struct inet_diag_handler *h)
1131 {
1132         const __u16 type = h->idiag_type;
1133         int err = -EINVAL;
1134
1135         if (type >= IPPROTO_MAX)
1136                 goto out;
1137
1138         mutex_lock(&inet_diag_table_mutex);
1139         err = -EEXIST;
1140         if (!inet_diag_table[type]) {
1141                 inet_diag_table[type] = h;
1142                 err = 0;
1143         }
1144         mutex_unlock(&inet_diag_table_mutex);
1145 out:
1146         return err;
1147 }
1148 EXPORT_SYMBOL_GPL(inet_diag_register);
1149
1150 void inet_diag_unregister(const struct inet_diag_handler *h)
1151 {
1152         const __u16 type = h->idiag_type;
1153
1154         if (type >= IPPROTO_MAX)
1155                 return;
1156
1157         mutex_lock(&inet_diag_table_mutex);
1158         inet_diag_table[type] = NULL;
1159         mutex_unlock(&inet_diag_table_mutex);
1160 }
1161 EXPORT_SYMBOL_GPL(inet_diag_unregister);
1162
1163 static int __init inet_diag_init(void)
1164 {
1165         const int inet_diag_table_size = (IPPROTO_MAX *
1166                                           sizeof(struct inet_diag_handler *));
1167         int err = -ENOMEM;
1168
1169         inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
1170         if (!inet_diag_table)
1171                 goto out;
1172
1173         err = sock_diag_register(&inet_diag_handler);
1174         if (err)
1175                 goto out_free_nl;
1176
1177         err = sock_diag_register(&inet6_diag_handler);
1178         if (err)
1179                 goto out_free_inet;
1180
1181         sock_diag_register_inet_compat(inet_diag_rcv_msg_compat);
1182 out:
1183         return err;
1184
1185 out_free_inet:
1186         sock_diag_unregister(&inet_diag_handler);
1187 out_free_nl:
1188         kfree(inet_diag_table);
1189         goto out;
1190 }
1191
1192 static void __exit inet_diag_exit(void)
1193 {
1194         sock_diag_unregister(&inet6_diag_handler);
1195         sock_diag_unregister(&inet_diag_handler);
1196         sock_diag_unregister_inet_compat(inet_diag_rcv_msg_compat);
1197         kfree(inet_diag_table);
1198 }
1199
1200 module_init(inet_diag_init);
1201 module_exit(inet_diag_exit);
1202 MODULE_LICENSE("GPL");
1203 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 2 /* AF_INET */);
1204 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 10 /* AF_INET6 */);