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