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         __be32 *saddr;
  42         __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 inline void inet_diag_unlock_handler(
  69         const struct inet_diag_handler *handler)
  70 {
  71         mutex_unlock(&inet_diag_table_mutex);
  72 }
  73
  74 static size_t inet_sk_attr_size(void)
  75 {
  76         return    nla_total_size(sizeof(struct tcp_info))
  77                 + nla_total_size(1) /* INET_DIAG_SHUTDOWN */
  78                 + nla_total_size(1) /* INET_DIAG_TOS */
  79                 + nla_total_size(1) /* INET_DIAG_TCLASS */
  80                 + nla_total_size(sizeof(struct inet_diag_meminfo))
  81                 + nla_total_size(sizeof(struct inet_diag_msg))
  82                 + nla_total_size(SK_MEMINFO_VARS * sizeof(u32))
  83                 + nla_total_size(TCP_CA_NAME_MAX)
  84                 + nla_total_size(sizeof(struct tcpvegas_info))
  85                 + 64;
  86 }
  87
  88 int inet_sk_diag_fill(struct sock *sk, struct inet_connection_sock *icsk,
  89                               struct sk_buff *skb, struct inet_diag_req_v2 *req,
  90                               struct user_namespace *user_ns,
  91                               u32 portid, u32 seq, u16 nlmsg_flags,
  92                               const struct nlmsghdr *unlh)
  93 {
  94         const struct inet_sock *inet = inet_sk(sk);
  95         struct inet_diag_msg *r;
  96         struct nlmsghdr  *nlh;
  97         struct nlattr *attr;
  98         void *info = NULL;
  99         const struct inet_diag_handler *handler;
 100         int ext = req->idiag_ext;
 101
 102         handler = inet_diag_table[req->sdiag_protocol];
 103         BUG_ON(handler == NULL);
 104
 105         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
 106                         nlmsg_flags);
 107         if (!nlh)
 108                 return -EMSGSIZE;
 109
 110         r = nlmsg_data(nlh);
 111         BUG_ON(sk->sk_state == TCP_TIME_WAIT);
 112
 113         r->idiag_family = sk->sk_family;
 114         r->idiag_state = sk->sk_state;
 115         r->idiag_timer = 0;
 116         r->idiag_retrans = 0;
 117
 118         r->id.idiag_if = sk->sk_bound_dev_if;
 119         sock_diag_save_cookie(sk, r->id.idiag_cookie);
 120
 121         r->id.idiag_sport = inet->inet_sport;
 122         r->id.idiag_dport = inet->inet_dport;
 123
 124         memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
 125         memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
 126
 127         r->id.idiag_src[0] = inet->inet_rcv_saddr;
 128         r->id.idiag_dst[0] = inet->inet_daddr;
 129
 130         if (nla_put_u8(skb, INET_DIAG_SHUTDOWN, sk->sk_shutdown))
 131                 goto errout;
 132
 133         /* IPv6 dual-stack sockets use inet->tos for IPv4 connections,
 134          * hence this needs to be included regardless of socket family.
 135          */
 136         if (ext & (1 << (INET_DIAG_TOS - 1)))
 137                 if (nla_put_u8(skb, INET_DIAG_TOS, inet->tos) < 0)
 138                         goto errout;
 139
 140 #if IS_ENABLED(CONFIG_IPV6)
 141         if (r->idiag_family == AF_INET6) {
 142                 const struct ipv6_pinfo *np = inet6_sk(sk);
 143
 144                 *(struct in6_addr *)r->id.idiag_src = np->rcv_saddr;
 145                 *(struct in6_addr *)r->id.idiag_dst = np->daddr;
 146
 147                 if (ext & (1 << (INET_DIAG_TCLASS - 1)))
 148                         if (nla_put_u8(skb, INET_DIAG_TCLASS, np->tclass) < 0)
 149                                 goto errout;
 150         }
 151 #endif
 152
 153         r->idiag_uid = from_kuid_munged(user_ns, sock_i_uid(sk));
 154         r->idiag_inode = sock_i_ino(sk);
 155
 156         if (ext & (1 << (INET_DIAG_MEMINFO - 1))) {
 157                 struct inet_diag_meminfo minfo = {
 158                         .idiag_rmem = sk_rmem_alloc_get(sk),
 159                         .idiag_wmem = sk->sk_wmem_queued,
 160                         .idiag_fmem = sk->sk_forward_alloc,
 161                         .idiag_tmem = sk_wmem_alloc_get(sk),
 162                 };
 163
 164                 if (nla_put(skb, INET_DIAG_MEMINFO, sizeof(minfo), &minfo) < 0)
 165                         goto errout;
 166         }
 167
 168         if (ext & (1 << (INET_DIAG_SKMEMINFO - 1)))
 169                 if (sock_diag_put_meminfo(sk, skb, INET_DIAG_SKMEMINFO))
 170                         goto errout;
 171
 172         if (icsk == NULL) {
 173                 handler->idiag_get_info(sk, r, NULL);
 174                 goto out;
 175         }
 176
 177 #define EXPIRES_IN_MS(tmo)  DIV_ROUND_UP((tmo - jiffies) * 1000, HZ)
 178
 179         if (icsk->icsk_pending == ICSK_TIME_RETRANS ||
 180             icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
 181             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
 182                 r->idiag_timer = 1;
 183                 r->idiag_retrans = icsk->icsk_retransmits;
 184                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 185         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 186                 r->idiag_timer = 4;
 187                 r->idiag_retrans = icsk->icsk_probes_out;
 188                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 189         } else if (timer_pending(&sk->sk_timer)) {
 190                 r->idiag_timer = 2;
 191                 r->idiag_retrans = icsk->icsk_probes_out;
 192                 r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 193         } else {
 194                 r->idiag_timer = 0;
 195                 r->idiag_expires = 0;
 196         }
 197 #undef EXPIRES_IN_MS
 198
 199         if (ext & (1 << (INET_DIAG_INFO - 1))) {
 200                 attr = nla_reserve(skb, INET_DIAG_INFO,
 201                                    sizeof(struct tcp_info));
 202                 if (!attr)
 203                         goto errout;
 204
 205                 info = nla_data(attr);
 206         }
 207
 208         if ((ext & (1 << (INET_DIAG_CONG - 1))) && icsk->icsk_ca_ops)
 209                 if (nla_put_string(skb, INET_DIAG_CONG,
 210                                    icsk->icsk_ca_ops->name) < 0)
 211                         goto errout;
 212
 213         handler->idiag_get_info(sk, r, info);
 214
 215         if (sk->sk_state < TCP_TIME_WAIT &&
 216             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 217                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 218
 219 out:
 220         return nlmsg_end(skb, nlh);
 221
 222 errout:
 223         nlmsg_cancel(skb, nlh);
 224         return -EMSGSIZE;
 225 }
 226 EXPORT_SYMBOL_GPL(inet_sk_diag_fill);
 227
 228 static int inet_csk_diag_fill(struct sock *sk,
 229                               struct sk_buff *skb, struct inet_diag_req_v2 *req,
 230                               struct user_namespace *user_ns,
 231                               u32 portid, u32 seq, u16 nlmsg_flags,
 232                               const struct nlmsghdr *unlh)
 233 {
 234         return inet_sk_diag_fill(sk, inet_csk(sk),
 235                         skb, req, user_ns, portid, seq, nlmsg_flags, unlh);
 236 }
 237
 238 static int inet_twsk_diag_fill(struct inet_timewait_sock *tw,
 239                                struct sk_buff *skb, struct inet_diag_req_v2 *req,
 240                                u32 portid, u32 seq, u16 nlmsg_flags,
 241                                const struct nlmsghdr *unlh)
 242 {
 243         long tmo;
 244         struct inet_diag_msg *r;
 245         struct nlmsghdr *nlh;
 246
 247         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
 248                         nlmsg_flags);
 249         if (!nlh)
 250                 return -EMSGSIZE;
 251
 252         r = nlmsg_data(nlh);
 253         BUG_ON(tw->tw_state != TCP_TIME_WAIT);
 254
 255         tmo = tw->tw_ttd - jiffies;
 256         if (tmo < 0)
 257                 tmo = 0;
 258
 259         r->idiag_family       = tw->tw_family;
 260         r->idiag_retrans      = 0;
 261
 262         r->id.idiag_if        = tw->tw_bound_dev_if;
 263         sock_diag_save_cookie(tw, r->id.idiag_cookie);
 264
 265         r->id.idiag_sport     = tw->tw_sport;
 266         r->id.idiag_dport     = tw->tw_dport;
 267
 268         memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
 269         memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
 270
 271         r->id.idiag_src[0]    = tw->tw_rcv_saddr;
 272         r->id.idiag_dst[0]    = tw->tw_daddr;
 273
 274         r->idiag_state        = tw->tw_substate;
 275         r->idiag_timer        = 3;
 276         r->idiag_expires      = DIV_ROUND_UP(tmo * 1000, HZ);
 277         r->idiag_rqueue       = 0;
 278         r->idiag_wqueue       = 0;
 279         r->idiag_uid          = 0;
 280         r->idiag_inode        = 0;
 281 #if IS_ENABLED(CONFIG_IPV6)
 282         if (tw->tw_family == AF_INET6) {
 283                 const struct inet6_timewait_sock *tw6 =
 284                                                 inet6_twsk((struct sock *)tw);
 285
 286                 *(struct in6_addr *)r->id.idiag_src = tw6->tw_v6_rcv_saddr;
 287                 *(struct in6_addr *)r->id.idiag_dst = tw6->tw_v6_daddr;
 288         }
 289 #endif
 290
 291         return nlmsg_end(skb, nlh);
 292 }
 293
 294 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
 295                         struct inet_diag_req_v2 *r,
 296                         struct user_namespace *user_ns,
 297                         u32 portid, u32 seq, u16 nlmsg_flags,
 298                         const struct nlmsghdr *unlh)
 299 {
 300         if (sk->sk_state == TCP_TIME_WAIT)
 301                 return inet_twsk_diag_fill((struct inet_timewait_sock *)sk,
 302                                            skb, r, portid, seq, nlmsg_flags,
 303                                            unlh);
 304         return inet_csk_diag_fill(sk, skb, r, user_ns, portid, seq, nlmsg_flags, unlh);
 305 }
 306
 307 int inet_diag_dump_one_icsk(struct inet_hashinfo *hashinfo, struct sk_buff *in_skb,
 308                 const struct nlmsghdr *nlh, struct inet_diag_req_v2 *req)
 309 {
 310         int err;
 311         struct sock *sk;
 312         struct sk_buff *rep;
 313         struct net *net = sock_net(in_skb->sk);
 314
 315         err = -EINVAL;
 316         if (req->sdiag_family == AF_INET) {
 317                 sk = inet_lookup(net, hashinfo, req->id.idiag_dst[0],
 318                                  req->id.idiag_dport, req->id.idiag_src[0],
 319                                  req->id.idiag_sport, req->id.idiag_if);
 320         }
 321 #if IS_ENABLED(CONFIG_IPV6)
 322         else if (req->sdiag_family == AF_INET6) {
 323                 sk = inet6_lookup(net, hashinfo,
 324                                   (struct in6_addr *)req->id.idiag_dst,
 325                                   req->id.idiag_dport,
 326                                   (struct in6_addr *)req->id.idiag_src,
 327                                   req->id.idiag_sport,
 328                                   req->id.idiag_if);
 329         }
 330 #endif
 331         else {
 332                 goto out_nosk;
 333         }
 334
 335         err = -ENOENT;
 336         if (sk == NULL)
 337                 goto out_nosk;
 338
 339         err = sock_diag_check_cookie(sk, req->id.idiag_cookie);
 340         if (err)
 341                 goto out;
 342
 343         rep = nlmsg_new(inet_sk_attr_size(), GFP_KERNEL);
 344         if (!rep) {
 345                 err = -ENOMEM;
 346                 goto out;
 347         }
 348
 349         err = sk_diag_fill(sk, rep, req,
 350                            sk_user_ns(NETLINK_CB(in_skb).sk),
 351                            NETLINK_CB(in_skb).portid,
 352                            nlh->nlmsg_seq, 0, nlh);
 353         if (err < 0) {
 354                 WARN_ON(err == -EMSGSIZE);
 355                 nlmsg_free(rep);
 356                 goto out;
 357         }
 358         err = netlink_unicast(net->diag_nlsk, rep, NETLINK_CB(in_skb).portid,
 359                               MSG_DONTWAIT);
 360         if (err > 0)
 361                 err = 0;
 362
 363 out:
 364         if (sk) {
 365                 if (sk->sk_state == TCP_TIME_WAIT)
 366                         inet_twsk_put((struct inet_timewait_sock *)sk);
 367                 else
 368                         sock_put(sk);
 369         }
 370 out_nosk:
 371         return err;
 372 }
 373 EXPORT_SYMBOL_GPL(inet_diag_dump_one_icsk);
 374
 375 static int inet_diag_get_exact(struct sk_buff *in_skb,
 376                                const struct nlmsghdr *nlh,
 377                                struct inet_diag_req_v2 *req)
 378 {
 379         const struct inet_diag_handler *handler;
 380         int err;
 381
 382         handler = inet_diag_lock_handler(req->sdiag_protocol);
 383         if (IS_ERR(handler))
 384                 err = PTR_ERR(handler);
 385         else
 386                 err = handler->dump_one(in_skb, nlh, req);
 387         inet_diag_unlock_handler(handler);
 388
 389         return err;
 390 }
 391
 392 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
 393 {
 394         int words = bits >> 5;
 395
 396         bits &= 0x1f;
 397
 398         if (words) {
 399                 if (memcmp(a1, a2, words << 2))
 400                         return 0;
 401         }
 402         if (bits) {
 403                 __be32 w1, w2;
 404                 __be32 mask;
 405
 406                 w1 = a1[words];
 407                 w2 = a2[words];
 408
 409                 mask = htonl((0xffffffff) << (32 - bits));
 410
 411                 if ((w1 ^ w2) & mask)
 412                         return 0;
 413         }
 414
 415         return 1;
 416 }
 417
 418
 419 static int inet_diag_bc_run(const struct nlattr *_bc,
 420                 const struct inet_diag_entry *entry)
 421 {
 422         const void *bc = nla_data(_bc);
 423         int len = nla_len(_bc);
 424
 425         while (len > 0) {
 426                 int yes = 1;
 427                 const struct inet_diag_bc_op *op = bc;
 428
 429                 switch (op->code) {
 430                 case INET_DIAG_BC_NOP:
 431                         break;
 432                 case INET_DIAG_BC_JMP:
 433                         yes = 0;
 434                         break;
 435                 case INET_DIAG_BC_S_GE:
 436                         yes = entry->sport >= op[1].no;
 437                         break;
 438                 case INET_DIAG_BC_S_LE:
 439                         yes = entry->sport <= op[1].no;
 440                         break;
 441                 case INET_DIAG_BC_D_GE:
 442                         yes = entry->dport >= op[1].no;
 443                         break;
 444                 case INET_DIAG_BC_D_LE:
 445                         yes = entry->dport <= op[1].no;
 446                         break;
 447                 case INET_DIAG_BC_AUTO:
 448                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 449                         break;
 450                 case INET_DIAG_BC_S_COND:
 451                 case INET_DIAG_BC_D_COND: {
 452                         struct inet_diag_hostcond *cond;
 453                         __be32 *addr;
 454
 455                         cond = (struct inet_diag_hostcond *)(op + 1);
 456                         if (cond->port != -1 &&
 457                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 458                                              entry->sport : entry->dport)) {
 459                                 yes = 0;
 460                                 break;
 461                         }
 462
 463                         if (op->code == INET_DIAG_BC_S_COND)
 464                                 addr = entry->saddr;
 465                         else
 466                                 addr = entry->daddr;
 467
 468                         if (cond->family != AF_UNSPEC &&
 469                             cond->family != entry->family) {
 470                                 if (entry->family == AF_INET6 &&
 471                                     cond->family == AF_INET) {
 472                                         if (addr[0] == 0 && addr[1] == 0 &&
 473                                             addr[2] == htonl(0xffff) &&
 474                                             bitstring_match(addr + 3,
 475                                                             cond->addr,
 476                                                             cond->prefix_len))
 477                                                 break;
 478                                 }
 479                                 yes = 0;
 480                                 break;
 481                         }
 482
 483                         if (cond->prefix_len == 0)
 484                                 break;
 485                         if (bitstring_match(addr, cond->addr,
 486                                             cond->prefix_len))
 487                                 break;
 488                         yes = 0;
 489                         break;
 490                 }
 491                 }
 492
 493                 if (yes) {
 494                         len -= op->yes;
 495                         bc += op->yes;
 496                 } else {
 497                         len -= op->no;
 498                         bc += op->no;
 499                 }
 500         }
 501         return len == 0;
 502 }
 503
 504 int inet_diag_bc_sk(const struct nlattr *bc, struct sock *sk)
 505 {
 506         struct inet_diag_entry entry;
 507         struct inet_sock *inet = inet_sk(sk);
 508
 509         if (bc == NULL)
 510                 return 1;
 511
 512         entry.family = sk->sk_family;
 513 #if IS_ENABLED(CONFIG_IPV6)
 514         if (entry.family == AF_INET6) {
 515                 struct ipv6_pinfo *np = inet6_sk(sk);
 516
 517                 entry.saddr = np->rcv_saddr.s6_addr32;
 518                 entry.daddr = np->daddr.s6_addr32;
 519         } else
 520 #endif
 521         {
 522                 entry.saddr = &inet->inet_rcv_saddr;
 523                 entry.daddr = &inet->inet_daddr;
 524         }
 525         entry.sport = inet->inet_num;
 526         entry.dport = ntohs(inet->inet_dport);
 527         entry.userlocks = sk->sk_userlocks;
 528
 529         return inet_diag_bc_run(bc, &entry);
 530 }
 531 EXPORT_SYMBOL_GPL(inet_diag_bc_sk);
 532
 533 static int valid_cc(const void *bc, int len, int cc)
 534 {
 535         while (len >= 0) {
 536                 const struct inet_diag_bc_op *op = bc;
 537
 538                 if (cc > len)
 539                         return 0;
 540                 if (cc == len)
 541                         return 1;
 542                 if (op->yes < 4 || op->yes & 3)
 543                         return 0;
 544                 len -= op->yes;
 545                 bc  += op->yes;
 546         }
 547         return 0;
 548 }
 549
 550 /* Validate an inet_diag_hostcond. */
 551 static bool valid_hostcond(const struct inet_diag_bc_op *op, int len,
 552                            int *min_len)
 553 {
 554         int addr_len;
 555         struct inet_diag_hostcond *cond;
 556
 557         /* Check hostcond space. */
 558         *min_len += sizeof(struct inet_diag_hostcond);
 559         if (len < *min_len)
 560                 return false;
 561         cond = (struct inet_diag_hostcond *)(op + 1);
 562
 563         /* Check address family and address length. */
 564         switch (cond->family) {
 565         case AF_UNSPEC:
 566                 addr_len = 0;
 567                 break;
 568         case AF_INET:
 569                 addr_len = sizeof(struct in_addr);
 570                 break;
 571         case AF_INET6:
 572                 addr_len = sizeof(struct in6_addr);
 573                 break;
 574         default:
 575                 return false;
 576         }
 577         *min_len += addr_len;
 578         if (len < *min_len)
 579                 return false;
 580
 581         /* Check prefix length (in bits) vs address length (in bytes). */
 582         if (cond->prefix_len > 8 * addr_len)
 583                 return false;
 584
 585         return true;
 586 }
 587
 588 /* Validate a port comparison operator. */
 589 static inline bool valid_port_comparison(const struct inet_diag_bc_op *op,
 590                                          int len, int *min_len)
 591 {
 592         /* Port comparisons put the port in a follow-on inet_diag_bc_op. */
 593         *min_len += sizeof(struct inet_diag_bc_op);
 594         if (len < *min_len)
 595                 return false;
 596         return true;
 597 }
 598
 599 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
 600 {
 601         const void *bc = bytecode;
 602         int  len = bytecode_len;
 603
 604         while (len > 0) {
 605                 const struct inet_diag_bc_op *op = bc;
 606                 int min_len = sizeof(struct inet_diag_bc_op);
 607
 608 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 609                 switch (op->code) {
 610                 case INET_DIAG_BC_S_COND:
 611                 case INET_DIAG_BC_D_COND:
 612                         if (!valid_hostcond(bc, len, &min_len))
 613                                 return -EINVAL;
 614                         break;
 615                 case INET_DIAG_BC_S_GE:
 616                 case INET_DIAG_BC_S_LE:
 617                 case INET_DIAG_BC_D_GE:
 618                 case INET_DIAG_BC_D_LE:
 619                         if (!valid_port_comparison(bc, len, &min_len))
 620                                 return -EINVAL;
 621                         break;
 622                 case INET_DIAG_BC_AUTO:
 623                 case INET_DIAG_BC_JMP:
 624                 case INET_DIAG_BC_NOP:
 625                         break;
 626                 default:
 627                         return -EINVAL;
 628                 }
 629
 630                 if (op->code != INET_DIAG_BC_NOP) {
 631                         if (op->no < min_len || op->no > len + 4 || op->no & 3)
 632                                 return -EINVAL;
 633                         if (op->no < len &&
 634                             !valid_cc(bytecode, bytecode_len, len - op->no))
 635                                 return -EINVAL;
 636                 }
 637
 638                 if (op->yes < min_len || op->yes > len + 4 || op->yes & 3)
 639                         return -EINVAL;
 640                 bc  += op->yes;
 641                 len -= op->yes;
 642         }
 643         return len == 0 ? 0 : -EINVAL;
 644 }
 645
 646 static int inet_csk_diag_dump(struct sock *sk,
 647                               struct sk_buff *skb,
 648                               struct netlink_callback *cb,
 649                               struct inet_diag_req_v2 *r,
 650                               const struct nlattr *bc)
 651 {
 652         if (!inet_diag_bc_sk(bc, sk))
 653                 return 0;
 654
 655         return inet_csk_diag_fill(sk, skb, r,
 656                                   sk_user_ns(NETLINK_CB(cb->skb).sk),
 657                                   NETLINK_CB(cb->skb).portid,
 658                                   cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 659 }
 660
 661 static int inet_twsk_diag_dump(struct inet_timewait_sock *tw,
 662                                struct sk_buff *skb,
 663                                struct netlink_callback *cb,
 664                                struct inet_diag_req_v2 *r,
 665                                const struct nlattr *bc)
 666 {
 667         if (bc != NULL) {
 668                 struct inet_diag_entry entry;
 669
 670                 entry.family = tw->tw_family;
 671 #if IS_ENABLED(CONFIG_IPV6)
 672                 if (tw->tw_family == AF_INET6) {
 673                         struct inet6_timewait_sock *tw6 =
 674                                                 inet6_twsk((struct sock *)tw);
 675                         entry.saddr = tw6->tw_v6_rcv_saddr.s6_addr32;
 676                         entry.daddr = tw6->tw_v6_daddr.s6_addr32;
 677                 } else
 678 #endif
 679                 {
 680                         entry.saddr = &tw->tw_rcv_saddr;
 681                         entry.daddr = &tw->tw_daddr;
 682                 }
 683                 entry.sport = tw->tw_num;
 684                 entry.dport = ntohs(tw->tw_dport);
 685                 entry.userlocks = 0;
 686
 687                 if (!inet_diag_bc_run(bc, &entry))
 688                         return 0;
 689         }
 690
 691         return inet_twsk_diag_fill(tw, skb, r,
 692                                    NETLINK_CB(cb->skb).portid,
 693                                    cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 694 }
 695
 696 /* Get the IPv4, IPv6, or IPv4-mapped-IPv6 local and remote addresses
 697  * from a request_sock. For IPv4-mapped-IPv6 we must map IPv4 to IPv6.
 698  */
 699 static inline void inet_diag_req_addrs(const struct sock *sk,
 700                                        const struct request_sock *req,
 701                                        struct inet_diag_entry *entry)
 702 {
 703         struct inet_request_sock *ireq = inet_rsk(req);
 704
 705 #if IS_ENABLED(CONFIG_IPV6)
 706         if (sk->sk_family == AF_INET6) {
 707                 if (req->rsk_ops->family == AF_INET6) {
 708                         entry->saddr = inet6_rsk(req)->loc_addr.s6_addr32;
 709                         entry->daddr = inet6_rsk(req)->rmt_addr.s6_addr32;
 710                 } else if (req->rsk_ops->family == AF_INET) {
 711                         ipv6_addr_set_v4mapped(ireq->loc_addr,
 712                                                &entry->saddr_storage);
 713                         ipv6_addr_set_v4mapped(ireq->rmt_addr,
 714                                                &entry->daddr_storage);
 715                         entry->saddr = entry->saddr_storage.s6_addr32;
 716                         entry->daddr = entry->daddr_storage.s6_addr32;
 717                 }
 718         } else
 719 #endif
 720         {
 721                 entry->saddr = &ireq->loc_addr;
 722                 entry->daddr = &ireq->rmt_addr;
 723         }
 724 }
 725
 726 static int inet_diag_fill_req(struct sk_buff *skb, struct sock *sk,
 727                               struct request_sock *req,
 728                               struct user_namespace *user_ns,
 729                               u32 portid, u32 seq,
 730                               const struct nlmsghdr *unlh)
 731 {
 732         const struct inet_request_sock *ireq = inet_rsk(req);
 733         struct inet_sock *inet = inet_sk(sk);
 734         struct inet_diag_msg *r;
 735         struct nlmsghdr *nlh;
 736         long tmo;
 737
 738         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
 739                         NLM_F_MULTI);
 740         if (!nlh)
 741                 return -EMSGSIZE;
 742
 743         r = nlmsg_data(nlh);
 744         r->idiag_family = sk->sk_family;
 745         r->idiag_state = TCP_SYN_RECV;
 746         r->idiag_timer = 1;
 747         r->idiag_retrans = req->num_retrans;
 748
 749         r->id.idiag_if = sk->sk_bound_dev_if;
 750         sock_diag_save_cookie(req, r->id.idiag_cookie);
 751
 752         tmo = req->expires - jiffies;
 753         if (tmo < 0)
 754                 tmo = 0;
 755
 756         r->id.idiag_sport = inet->inet_sport;
 757         r->id.idiag_dport = ireq->rmt_port;
 758
 759         memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
 760         memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
 761
 762         r->id.idiag_src[0] = ireq->loc_addr;
 763         r->id.idiag_dst[0] = ireq->rmt_addr;
 764
 765         r->idiag_expires = jiffies_to_msecs(tmo);
 766         r->idiag_rqueue = 0;
 767         r->idiag_wqueue = 0;
 768         r->idiag_uid = from_kuid_munged(user_ns, sock_i_uid(sk));
 769         r->idiag_inode = 0;
 770 #if IS_ENABLED(CONFIG_IPV6)
 771         if (r->idiag_family == AF_INET6) {
 772                 struct inet_diag_entry entry;
 773                 inet_diag_req_addrs(sk, req, &entry);
 774                 memcpy(r->id.idiag_src, entry.saddr, sizeof(struct in6_addr));
 775                 memcpy(r->id.idiag_dst, entry.daddr, sizeof(struct in6_addr));
 776         }
 777 #endif
 778
 779         return nlmsg_end(skb, nlh);
 780 }
 781
 782 static int inet_diag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 783                                struct netlink_callback *cb,
 784                                struct inet_diag_req_v2 *r,
 785                                const struct nlattr *bc)
 786 {
 787         struct inet_diag_entry entry;
 788         struct inet_connection_sock *icsk = inet_csk(sk);
 789         struct listen_sock *lopt;
 790         struct inet_sock *inet = inet_sk(sk);
 791         int j, s_j;
 792         int reqnum, s_reqnum;
 793         int err = 0;
 794
 795         s_j = cb->args[3];
 796         s_reqnum = cb->args[4];
 797
 798         if (s_j > 0)
 799                 s_j--;
 800
 801         entry.family = sk->sk_family;
 802
 803         read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 804
 805         lopt = icsk->icsk_accept_queue.listen_opt;
 806         if (!lopt || !lopt->qlen)
 807                 goto out;
 808
 809         if (bc != NULL) {
 810                 entry.sport = inet->inet_num;
 811                 entry.userlocks = sk->sk_userlocks;
 812         }
 813
 814         for (j = s_j; j < lopt->nr_table_entries; j++) {
 815                 struct request_sock *req, *head = lopt->syn_table[j];
 816
 817                 reqnum = 0;
 818                 for (req = head; req; reqnum++, req = req->dl_next) {
 819                         struct inet_request_sock *ireq = inet_rsk(req);
 820
 821                         if (reqnum < s_reqnum)
 822                                 continue;
 823                         if (r->id.idiag_dport != ireq->rmt_port &&
 824                             r->id.idiag_dport)
 825                                 continue;
 826
 827                         if (bc) {
 828                                 inet_diag_req_addrs(sk, req, &entry);
 829                                 entry.dport = ntohs(ireq->rmt_port);
 830
 831                                 if (!inet_diag_bc_run(bc, &entry))
 832                                         continue;
 833                         }
 834
 835                         err = inet_diag_fill_req(skb, sk, req,
 836                                                sk_user_ns(NETLINK_CB(cb->skb).sk),
 837                                                NETLINK_CB(cb->skb).portid,
 838                                                cb->nlh->nlmsg_seq, cb->nlh);
 839                         if (err < 0) {
 840                                 cb->args[3] = j + 1;
 841                                 cb->args[4] = reqnum;
 842                                 goto out;
 843                         }
 844                 }
 845
 846                 s_reqnum = 0;
 847         }
 848
 849 out:
 850         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 851
 852         return err;
 853 }
 854
 855 void inet_diag_dump_icsk(struct inet_hashinfo *hashinfo, struct sk_buff *skb,
 856                 struct netlink_callback *cb, struct inet_diag_req_v2 *r, struct nlattr *bc)
 857 {
 858         int i, num;
 859         int s_i, s_num;
 860         struct net *net = sock_net(skb->sk);
 861
 862         s_i = cb->args[1];
 863         s_num = num = cb->args[2];
 864
 865         if (cb->args[0] == 0) {
 866                 if (!(r->idiag_states & (TCPF_LISTEN | TCPF_SYN_RECV)))
 867                         goto skip_listen_ht;
 868
 869                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 870                         struct sock *sk;
 871                         struct hlist_nulls_node *node;
 872                         struct inet_listen_hashbucket *ilb;
 873
 874                         num = 0;
 875                         ilb = &hashinfo->listening_hash[i];
 876                         spin_lock_bh(&ilb->lock);
 877                         sk_nulls_for_each(sk, node, &ilb->head) {
 878                                 struct inet_sock *inet = inet_sk(sk);
 879
 880                                 if (!net_eq(sock_net(sk), net))
 881                                         continue;
 882
 883                                 if (num < s_num) {
 884                                         num++;
 885                                         continue;
 886                                 }
 887
 888                                 if (r->sdiag_family != AF_UNSPEC &&
 889                                                 sk->sk_family != r->sdiag_family)
 890                                         goto next_listen;
 891
 892                                 if (r->id.idiag_sport != inet->inet_sport &&
 893                                     r->id.idiag_sport)
 894                                         goto next_listen;
 895
 896                                 if (!(r->idiag_states & TCPF_LISTEN) ||
 897                                     r->id.idiag_dport ||
 898                                     cb->args[3] > 0)
 899                                         goto syn_recv;
 900
 901                                 if (inet_csk_diag_dump(sk, skb, cb, r, bc) < 0) {
 902                                         spin_unlock_bh(&ilb->lock);
 903                                         goto done;
 904                                 }
 905
 906 syn_recv:
 907                                 if (!(r->idiag_states & TCPF_SYN_RECV))
 908                                         goto next_listen;
 909
 910                                 if (inet_diag_dump_reqs(skb, sk, cb, r, bc) < 0) {
 911                                         spin_unlock_bh(&ilb->lock);
 912                                         goto done;
 913                                 }
 914
 915 next_listen:
 916                                 cb->args[3] = 0;
 917                                 cb->args[4] = 0;
 918                                 ++num;
 919                         }
 920                         spin_unlock_bh(&ilb->lock);
 921
 922                         s_num = 0;
 923                         cb->args[3] = 0;
 924                         cb->args[4] = 0;
 925                 }
 926 skip_listen_ht:
 927                 cb->args[0] = 1;
 928                 s_i = num = s_num = 0;
 929         }
 930
 931         if (!(r->idiag_states & ~(TCPF_LISTEN | TCPF_SYN_RECV)))
 932                 goto out;
 933
 934         for (i = s_i; i <= hashinfo->ehash_mask; i++) {
 935                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 936                 spinlock_t *lock = inet_ehash_lockp(hashinfo, i);
 937                 struct sock *sk;
 938                 struct hlist_nulls_node *node;
 939
 940                 num = 0;
 941
 942                 if (hlist_nulls_empty(&head->chain) &&
 943                         hlist_nulls_empty(&head->twchain))
 944                         continue;
 945
 946                 if (i > s_i)
 947                         s_num = 0;
 948
 949                 spin_lock_bh(lock);
 950                 sk_nulls_for_each(sk, node, &head->chain) {
 951                         struct inet_sock *inet = inet_sk(sk);
 952
 953                         if (!net_eq(sock_net(sk), net))
 954                                 continue;
 955                         if (num < s_num)
 956                                 goto next_normal;
 957                         if (!(r->idiag_states & (1 << sk->sk_state)))
 958                                 goto next_normal;
 959                         if (r->sdiag_family != AF_UNSPEC &&
 960                                         sk->sk_family != r->sdiag_family)
 961                                 goto next_normal;
 962                         if (r->id.idiag_sport != inet->inet_sport &&
 963                             r->id.idiag_sport)
 964                                 goto next_normal;
 965                         if (r->id.idiag_dport != inet->inet_dport &&
 966                             r->id.idiag_dport)
 967                                 goto next_normal;
 968                         if (inet_csk_diag_dump(sk, skb, cb, r, bc) < 0) {
 969                                 spin_unlock_bh(lock);
 970                                 goto done;
 971                         }
 972 next_normal:
 973                         ++num;
 974                 }
 975
 976                 if (r->idiag_states & (TCPF_TIME_WAIT | TCPF_FIN_WAIT2)) {
 977                         struct inet_timewait_sock *tw;
 978
 979                         inet_twsk_for_each(tw, node,
 980                                     &head->twchain) {
 981                                 if (!net_eq(twsk_net(tw), net))
 982                                         continue;
 983
 984                                 if (num < s_num)
 985                                         goto next_dying;
 986                                 if (!(r->idiag_states & (1 << tw->tw_substate)))
 987                                         goto next_dying;
 988                                 if (r->sdiag_family != AF_UNSPEC &&
 989                                                 tw->tw_family != r->sdiag_family)
 990                                         goto next_dying;
 991                                 if (r->id.idiag_sport != tw->tw_sport &&
 992                                     r->id.idiag_sport)
 993                                         goto next_dying;
 994                                 if (r->id.idiag_dport != tw->tw_dport &&
 995                                     r->id.idiag_dport)
 996                                         goto next_dying;
 997                                 if (inet_twsk_diag_dump(tw, skb, cb, r, bc) < 0) {
 998                                         spin_unlock_bh(lock);
 999                                         goto done;
1000                                 }
1001 next_dying:
1002                                 ++num;
1003                         }
1004                 }
1005                 spin_unlock_bh(lock);
1006         }
1007
1008 done:
1009         cb->args[1] = i;
1010         cb->args[2] = num;
1011 out:
1012         ;
1013 }
1014 EXPORT_SYMBOL_GPL(inet_diag_dump_icsk);
1015
1016 static int __inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb,
1017                 struct inet_diag_req_v2 *r, struct nlattr *bc)
1018 {
1019         const struct inet_diag_handler *handler;
1020         int err = 0;
1021
1022         handler = inet_diag_lock_handler(r->sdiag_protocol);
1023         if (!IS_ERR(handler))
1024                 handler->dump(skb, cb, r, bc);
1025         else
1026                 err = PTR_ERR(handler);
1027         inet_diag_unlock_handler(handler);
1028
1029         return err ? : skb->len;
1030 }
1031
1032 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
1033 {
1034         struct nlattr *bc = NULL;
1035         int hdrlen = sizeof(struct inet_diag_req_v2);
1036
1037         if (nlmsg_attrlen(cb->nlh, hdrlen))
1038                 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
1039
1040         return __inet_diag_dump(skb, cb, nlmsg_data(cb->nlh), bc);
1041 }
1042
1043 static inline int inet_diag_type2proto(int type)
1044 {
1045         switch (type) {
1046         case TCPDIAG_GETSOCK:
1047                 return IPPROTO_TCP;
1048         case DCCPDIAG_GETSOCK:
1049                 return IPPROTO_DCCP;
1050         default:
1051                 return 0;
1052         }
1053 }
1054
1055 static int inet_diag_dump_compat(struct sk_buff *skb, struct netlink_callback *cb)
1056 {
1057         struct inet_diag_req *rc = nlmsg_data(cb->nlh);
1058         struct inet_diag_req_v2 req;
1059         struct nlattr *bc = NULL;
1060         int hdrlen = sizeof(struct inet_diag_req);
1061
1062         req.sdiag_family = AF_UNSPEC; /* compatibility */
1063         req.sdiag_protocol = inet_diag_type2proto(cb->nlh->nlmsg_type);
1064         req.idiag_ext = rc->idiag_ext;
1065         req.idiag_states = rc->idiag_states;
1066         req.id = rc->id;
1067
1068         if (nlmsg_attrlen(cb->nlh, hdrlen))
1069                 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
1070
1071         return __inet_diag_dump(skb, cb, &req, bc);
1072 }
1073
1074 static int inet_diag_get_exact_compat(struct sk_buff *in_skb,
1075                                const struct nlmsghdr *nlh)
1076 {
1077         struct inet_diag_req *rc = nlmsg_data(nlh);
1078         struct inet_diag_req_v2 req;
1079
1080         req.sdiag_family = rc->idiag_family;
1081         req.sdiag_protocol = inet_diag_type2proto(nlh->nlmsg_type);
1082         req.idiag_ext = rc->idiag_ext;
1083         req.idiag_states = rc->idiag_states;
1084         req.id = rc->id;
1085
1086         return inet_diag_get_exact(in_skb, nlh, &req);
1087 }
1088
1089 static int inet_diag_rcv_msg_compat(struct sk_buff *skb, struct nlmsghdr *nlh)
1090 {
1091         int hdrlen = sizeof(struct inet_diag_req);
1092         struct net *net = sock_net(skb->sk);
1093
1094         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
1095             nlmsg_len(nlh) < hdrlen)
1096                 return -EINVAL;
1097
1098         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1099                 if (nlmsg_attrlen(nlh, hdrlen)) {
1100                         struct nlattr *attr;
1101
1102                         attr = nlmsg_find_attr(nlh, hdrlen,
1103                                                INET_DIAG_REQ_BYTECODE);
1104                         if (attr == NULL ||
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_compat,
1112                         };
1113                         return netlink_dump_start(net->diag_nlsk, skb, nlh, &c);
1114                 }
1115         }
1116
1117         return inet_diag_get_exact_compat(skb, nlh);
1118 }
1119
1120 static int inet_diag_handler_dump(struct sk_buff *skb, struct nlmsghdr *h)
1121 {
1122         int hdrlen = sizeof(struct inet_diag_req_v2);
1123         struct net *net = sock_net(skb->sk);
1124
1125         if (nlmsg_len(h) < hdrlen)
1126                 return -EINVAL;
1127
1128         if (h->nlmsg_flags & NLM_F_DUMP) {
1129                 if (nlmsg_attrlen(h, hdrlen)) {
1130                         struct nlattr *attr;
1131                         attr = nlmsg_find_attr(h, hdrlen,
1132                                                INET_DIAG_REQ_BYTECODE);
1133                         if (attr == NULL ||
1134                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
1135                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
1136                                 return -EINVAL;
1137                 }
1138                 {
1139                         struct netlink_dump_control c = {
1140                                 .dump = inet_diag_dump,
1141                         };
1142                         return netlink_dump_start(net->diag_nlsk, skb, h, &c);
1143                 }
1144         }
1145
1146         return inet_diag_get_exact(skb, h, nlmsg_data(h));
1147 }
1148
1149 static const struct sock_diag_handler inet_diag_handler = {
1150         .family = AF_INET,
1151         .dump = inet_diag_handler_dump,
1152 };
1153
1154 static const struct sock_diag_handler inet6_diag_handler = {
1155         .family = AF_INET6,
1156         .dump = inet_diag_handler_dump,
1157 };
1158
1159 int inet_diag_register(const struct inet_diag_handler *h)
1160 {
1161         const __u16 type = h->idiag_type;
1162         int err = -EINVAL;
1163
1164         if (type >= IPPROTO_MAX)
1165                 goto out;
1166
1167         mutex_lock(&inet_diag_table_mutex);
1168         err = -EEXIST;
1169         if (inet_diag_table[type] == NULL) {
1170                 inet_diag_table[type] = h;
1171                 err = 0;
1172         }
1173         mutex_unlock(&inet_diag_table_mutex);
1174 out:
1175         return err;
1176 }
1177 EXPORT_SYMBOL_GPL(inet_diag_register);
1178
1179 void inet_diag_unregister(const struct inet_diag_handler *h)
1180 {
1181         const __u16 type = h->idiag_type;
1182
1183         if (type >= IPPROTO_MAX)
1184                 return;
1185
1186         mutex_lock(&inet_diag_table_mutex);
1187         inet_diag_table[type] = NULL;
1188         mutex_unlock(&inet_diag_table_mutex);
1189 }
1190 EXPORT_SYMBOL_GPL(inet_diag_unregister);
1191
1192 static int __init inet_diag_init(void)
1193 {
1194         const int inet_diag_table_size = (IPPROTO_MAX *
1195                                           sizeof(struct inet_diag_handler *));
1196         int err = -ENOMEM;
1197
1198         inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
1199         if (!inet_diag_table)
1200                 goto out;
1201
1202         err = sock_diag_register(&inet_diag_handler);
1203         if (err)
1204                 goto out_free_nl;
1205
1206         err = sock_diag_register(&inet6_diag_handler);
1207         if (err)
1208                 goto out_free_inet;
1209
1210         sock_diag_register_inet_compat(inet_diag_rcv_msg_compat);
1211 out:
1212         return err;
1213
1214 out_free_inet:
1215         sock_diag_unregister(&inet_diag_handler);
1216 out_free_nl:
1217         kfree(inet_diag_table);
1218         goto out;
1219 }
1220
1221 static void __exit inet_diag_exit(void)
1222 {
1223         sock_diag_unregister(&inet6_diag_handler);
1224         sock_diag_unregister(&inet_diag_handler);
1225         sock_diag_unregister_inet_compat(inet_diag_rcv_msg_compat);
1226         kfree(inet_diag_table);
1227 }
1228
1229 module_init(inet_diag_init);
1230 module_exit(inet_diag_exit);
1231 MODULE_LICENSE("GPL");
1232 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 2 /* AF_INET */);
1233 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 10 /* AF_INET6 */);