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 };
  48
  49 static struct sock *sdiagnl;
  50
  51 #define INET_DIAG_PUT(skb, attrtype, attrlen) \
  52         RTA_DATA(__RTA_PUT(skb, attrtype, attrlen))
  53
  54 static inline int inet_diag_type2proto(int type)
  55 {
  56         switch (type) {
  57         case TCPDIAG_GETSOCK:
  58                 return IPPROTO_TCP;
  59         case DCCPDIAG_GETSOCK:
  60                 return IPPROTO_DCCP;
  61         default:
  62                 return 0;
  63         }
  64 }
  65
  66 static DEFINE_MUTEX(inet_diag_table_mutex);
  67
  68 static const struct inet_diag_handler *inet_diag_lock_handler(int proto)
  69 {
  70         if (!inet_diag_table[proto])
  71                 request_module("net-pf-%d-proto-%d-type-%d", PF_NETLINK,
  72                                NETLINK_SOCK_DIAG, proto);
  73
  74         mutex_lock(&inet_diag_table_mutex);
  75         if (!inet_diag_table[proto])
  76                 return ERR_PTR(-ENOENT);
  77
  78         return inet_diag_table[proto];
  79 }
  80
  81 static inline void inet_diag_unlock_handler(
  82         const struct inet_diag_handler *handler)
  83 {
  84         mutex_unlock(&inet_diag_table_mutex);
  85 }
  86
  87 static int inet_csk_diag_fill(struct sock *sk,
  88                               struct sk_buff *skb,
  89                               int ext, u32 pid, u32 seq, u16 nlmsg_flags,
  90                               const struct nlmsghdr *unlh)
  91 {
  92         const struct inet_sock *inet = inet_sk(sk);
  93         const struct inet_connection_sock *icsk = inet_csk(sk);
  94         struct inet_diag_msg *r;
  95         struct nlmsghdr  *nlh;
  96         void *info = NULL;
  97         struct inet_diag_meminfo  *minfo = NULL;
  98         unsigned char    *b = skb_tail_pointer(skb);
  99         const struct inet_diag_handler *handler;
 100
 101         handler = inet_diag_table[inet_diag_type2proto(unlh->nlmsg_type)];
 102         BUG_ON(handler == NULL);
 103
 104         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
 105         nlh->nlmsg_flags = nlmsg_flags;
 106
 107         r = NLMSG_DATA(nlh);
 108         BUG_ON(sk->sk_state == TCP_TIME_WAIT);
 109
 110         if (ext & (1 << (INET_DIAG_MEMINFO - 1)))
 111                 minfo = INET_DIAG_PUT(skb, INET_DIAG_MEMINFO, sizeof(*minfo));
 112
 113         if (ext & (1 << (INET_DIAG_INFO - 1)))
 114                 info = INET_DIAG_PUT(skb, INET_DIAG_INFO,
 115                                      handler->idiag_info_size);
 116
 117         if ((ext & (1 << (INET_DIAG_CONG - 1))) && icsk->icsk_ca_ops) {
 118                 const size_t len = strlen(icsk->icsk_ca_ops->name);
 119
 120                 strcpy(INET_DIAG_PUT(skb, INET_DIAG_CONG, len + 1),
 121                        icsk->icsk_ca_ops->name);
 122         }
 123
 124         r->idiag_family = sk->sk_family;
 125         r->idiag_state = sk->sk_state;
 126         r->idiag_timer = 0;
 127         r->idiag_retrans = 0;
 128
 129         r->id.idiag_if = sk->sk_bound_dev_if;
 130         r->id.idiag_cookie[0] = (u32)(unsigned long)sk;
 131         r->id.idiag_cookie[1] = (u32)(((unsigned long)sk >> 31) >> 1);
 132
 133         r->id.idiag_sport = inet->inet_sport;
 134         r->id.idiag_dport = inet->inet_dport;
 135         r->id.idiag_src[0] = inet->inet_rcv_saddr;
 136         r->id.idiag_dst[0] = inet->inet_daddr;
 137
 138         /* IPv6 dual-stack sockets use inet->tos for IPv4 connections,
 139          * hence this needs to be included regardless of socket family.
 140          */
 141         if (ext & (1 << (INET_DIAG_TOS - 1)))
 142                 RTA_PUT_U8(skb, INET_DIAG_TOS, inet->tos);
 143
 144 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 145         if (r->idiag_family == AF_INET6) {
 146                 const struct ipv6_pinfo *np = inet6_sk(sk);
 147
 148                 *(struct in6_addr *)r->id.idiag_src = np->rcv_saddr;
 149                 *(struct in6_addr *)r->id.idiag_dst = np->daddr;
 150                 if (ext & (1 << (INET_DIAG_TCLASS - 1)))
 151                         RTA_PUT_U8(skb, INET_DIAG_TCLASS, np->tclass);
 152         }
 153 #endif
 154
 155 #define EXPIRES_IN_MS(tmo)  DIV_ROUND_UP((tmo - jiffies) * 1000, HZ)
 156
 157         if (icsk->icsk_pending == ICSK_TIME_RETRANS) {
 158                 r->idiag_timer = 1;
 159                 r->idiag_retrans = icsk->icsk_retransmits;
 160                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 161         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 162                 r->idiag_timer = 4;
 163                 r->idiag_retrans = icsk->icsk_probes_out;
 164                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 165         } else if (timer_pending(&sk->sk_timer)) {
 166                 r->idiag_timer = 2;
 167                 r->idiag_retrans = icsk->icsk_probes_out;
 168                 r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 169         } else {
 170                 r->idiag_timer = 0;
 171                 r->idiag_expires = 0;
 172         }
 173 #undef EXPIRES_IN_MS
 174
 175         r->idiag_uid = sock_i_uid(sk);
 176         r->idiag_inode = sock_i_ino(sk);
 177
 178         if (minfo) {
 179                 minfo->idiag_rmem = sk_rmem_alloc_get(sk);
 180                 minfo->idiag_wmem = sk->sk_wmem_queued;
 181                 minfo->idiag_fmem = sk->sk_forward_alloc;
 182                 minfo->idiag_tmem = sk_wmem_alloc_get(sk);
 183         }
 184
 185         handler->idiag_get_info(sk, r, info);
 186
 187         if (sk->sk_state < TCP_TIME_WAIT &&
 188             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 189                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 190
 191         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 192         return skb->len;
 193
 194 rtattr_failure:
 195 nlmsg_failure:
 196         nlmsg_trim(skb, b);
 197         return -EMSGSIZE;
 198 }
 199
 200 static int inet_twsk_diag_fill(struct inet_timewait_sock *tw,
 201                                struct sk_buff *skb, int ext, u32 pid,
 202                                u32 seq, u16 nlmsg_flags,
 203                                const struct nlmsghdr *unlh)
 204 {
 205         long tmo;
 206         struct inet_diag_msg *r;
 207         const unsigned char *previous_tail = skb_tail_pointer(skb);
 208         struct nlmsghdr *nlh = NLMSG_PUT(skb, pid, seq,
 209                                          unlh->nlmsg_type, sizeof(*r));
 210
 211         r = NLMSG_DATA(nlh);
 212         BUG_ON(tw->tw_state != TCP_TIME_WAIT);
 213
 214         nlh->nlmsg_flags = nlmsg_flags;
 215
 216         tmo = tw->tw_ttd - jiffies;
 217         if (tmo < 0)
 218                 tmo = 0;
 219
 220         r->idiag_family       = tw->tw_family;
 221         r->idiag_retrans      = 0;
 222         r->id.idiag_if        = tw->tw_bound_dev_if;
 223         r->id.idiag_cookie[0] = (u32)(unsigned long)tw;
 224         r->id.idiag_cookie[1] = (u32)(((unsigned long)tw >> 31) >> 1);
 225         r->id.idiag_sport     = tw->tw_sport;
 226         r->id.idiag_dport     = tw->tw_dport;
 227         r->id.idiag_src[0]    = tw->tw_rcv_saddr;
 228         r->id.idiag_dst[0]    = tw->tw_daddr;
 229         r->idiag_state        = tw->tw_substate;
 230         r->idiag_timer        = 3;
 231         r->idiag_expires      = DIV_ROUND_UP(tmo * 1000, HZ);
 232         r->idiag_rqueue       = 0;
 233         r->idiag_wqueue       = 0;
 234         r->idiag_uid          = 0;
 235         r->idiag_inode        = 0;
 236 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 237         if (tw->tw_family == AF_INET6) {
 238                 const struct inet6_timewait_sock *tw6 =
 239                                                 inet6_twsk((struct sock *)tw);
 240
 241                 *(struct in6_addr *)r->id.idiag_src = tw6->tw_v6_rcv_saddr;
 242                 *(struct in6_addr *)r->id.idiag_dst = tw6->tw_v6_daddr;
 243         }
 244 #endif
 245         nlh->nlmsg_len = skb_tail_pointer(skb) - previous_tail;
 246         return skb->len;
 247 nlmsg_failure:
 248         nlmsg_trim(skb, previous_tail);
 249         return -EMSGSIZE;
 250 }
 251
 252 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
 253                         int ext, u32 pid, u32 seq, u16 nlmsg_flags,
 254                         const struct nlmsghdr *unlh)
 255 {
 256         if (sk->sk_state == TCP_TIME_WAIT)
 257                 return inet_twsk_diag_fill((struct inet_timewait_sock *)sk,
 258                                            skb, ext, pid, seq, nlmsg_flags,
 259                                            unlh);
 260         return inet_csk_diag_fill(sk, skb, ext, pid, seq, nlmsg_flags, unlh);
 261 }
 262
 263 static int inet_diag_get_exact(struct sk_buff *in_skb,
 264                                const struct nlmsghdr *nlh)
 265 {
 266         int err;
 267         struct sock *sk;
 268         struct inet_diag_req *req = NLMSG_DATA(nlh);
 269         struct sk_buff *rep;
 270         struct inet_hashinfo *hashinfo;
 271         const struct inet_diag_handler *handler;
 272
 273         handler = inet_diag_lock_handler(inet_diag_type2proto(nlh->nlmsg_type));
 274         if (IS_ERR(handler)) {
 275                 err = PTR_ERR(handler);
 276                 goto unlock;
 277         }
 278
 279         hashinfo = handler->idiag_hashinfo;
 280         err = -EINVAL;
 281
 282         if (req->idiag_family == AF_INET) {
 283                 sk = inet_lookup(&init_net, hashinfo, req->id.idiag_dst[0],
 284                                  req->id.idiag_dport, req->id.idiag_src[0],
 285                                  req->id.idiag_sport, req->id.idiag_if);
 286         }
 287 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 288         else if (req->idiag_family == AF_INET6) {
 289                 sk = inet6_lookup(&init_net, hashinfo,
 290                                   (struct in6_addr *)req->id.idiag_dst,
 291                                   req->id.idiag_dport,
 292                                   (struct in6_addr *)req->id.idiag_src,
 293                                   req->id.idiag_sport,
 294                                   req->id.idiag_if);
 295         }
 296 #endif
 297         else {
 298                 goto unlock;
 299         }
 300
 301         err = -ENOENT;
 302         if (sk == NULL)
 303                 goto unlock;
 304
 305         err = -ESTALE;
 306         if ((req->id.idiag_cookie[0] != INET_DIAG_NOCOOKIE ||
 307              req->id.idiag_cookie[1] != INET_DIAG_NOCOOKIE) &&
 308             ((u32)(unsigned long)sk != req->id.idiag_cookie[0] ||
 309              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.idiag_cookie[1]))
 310                 goto out;
 311
 312         err = -ENOMEM;
 313         rep = alloc_skb(NLMSG_SPACE((sizeof(struct inet_diag_msg) +
 314                                      sizeof(struct inet_diag_meminfo) +
 315                                      handler->idiag_info_size + 64)),
 316                         GFP_KERNEL);
 317         if (!rep)
 318                 goto out;
 319
 320         err = sk_diag_fill(sk, rep, req->idiag_ext,
 321                            NETLINK_CB(in_skb).pid,
 322                            nlh->nlmsg_seq, 0, nlh);
 323         if (err < 0) {
 324                 WARN_ON(err == -EMSGSIZE);
 325                 kfree_skb(rep);
 326                 goto out;
 327         }
 328         err = netlink_unicast(sdiagnl, rep, NETLINK_CB(in_skb).pid,
 329                               MSG_DONTWAIT);
 330         if (err > 0)
 331                 err = 0;
 332
 333 out:
 334         if (sk) {
 335                 if (sk->sk_state == TCP_TIME_WAIT)
 336                         inet_twsk_put((struct inet_timewait_sock *)sk);
 337                 else
 338                         sock_put(sk);
 339         }
 340 unlock:
 341         inet_diag_unlock_handler(handler);
 342         return err;
 343 }
 344
 345 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
 346 {
 347         int words = bits >> 5;
 348
 349         bits &= 0x1f;
 350
 351         if (words) {
 352                 if (memcmp(a1, a2, words << 2))
 353                         return 0;
 354         }
 355         if (bits) {
 356                 __be32 w1, w2;
 357                 __be32 mask;
 358
 359                 w1 = a1[words];
 360                 w2 = a2[words];
 361
 362                 mask = htonl((0xffffffff) << (32 - bits));
 363
 364                 if ((w1 ^ w2) & mask)
 365                         return 0;
 366         }
 367
 368         return 1;
 369 }
 370
 371
 372 static int inet_diag_bc_run(const void *bc, int len,
 373                             const struct inet_diag_entry *entry)
 374 {
 375         while (len > 0) {
 376                 int yes = 1;
 377                 const struct inet_diag_bc_op *op = bc;
 378
 379                 switch (op->code) {
 380                 case INET_DIAG_BC_NOP:
 381                         break;
 382                 case INET_DIAG_BC_JMP:
 383                         yes = 0;
 384                         break;
 385                 case INET_DIAG_BC_S_GE:
 386                         yes = entry->sport >= op[1].no;
 387                         break;
 388                 case INET_DIAG_BC_S_LE:
 389                         yes = entry->sport <= op[1].no;
 390                         break;
 391                 case INET_DIAG_BC_D_GE:
 392                         yes = entry->dport >= op[1].no;
 393                         break;
 394                 case INET_DIAG_BC_D_LE:
 395                         yes = entry->dport <= op[1].no;
 396                         break;
 397                 case INET_DIAG_BC_AUTO:
 398                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 399                         break;
 400                 case INET_DIAG_BC_S_COND:
 401                 case INET_DIAG_BC_D_COND: {
 402                         struct inet_diag_hostcond *cond;
 403                         __be32 *addr;
 404
 405                         cond = (struct inet_diag_hostcond *)(op + 1);
 406                         if (cond->port != -1 &&
 407                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 408                                              entry->sport : entry->dport)) {
 409                                 yes = 0;
 410                                 break;
 411                         }
 412
 413                         if (cond->prefix_len == 0)
 414                                 break;
 415
 416                         if (op->code == INET_DIAG_BC_S_COND)
 417                                 addr = entry->saddr;
 418                         else
 419                                 addr = entry->daddr;
 420
 421                         if (bitstring_match(addr, cond->addr,
 422                                             cond->prefix_len))
 423                                 break;
 424                         if (entry->family == AF_INET6 &&
 425                             cond->family == AF_INET) {
 426                                 if (addr[0] == 0 && addr[1] == 0 &&
 427                                     addr[2] == htonl(0xffff) &&
 428                                     bitstring_match(addr + 3, cond->addr,
 429                                                     cond->prefix_len))
 430                                         break;
 431                         }
 432                         yes = 0;
 433                         break;
 434                 }
 435                 }
 436
 437                 if (yes) {
 438                         len -= op->yes;
 439                         bc += op->yes;
 440                 } else {
 441                         len -= op->no;
 442                         bc += op->no;
 443                 }
 444         }
 445         return len == 0;
 446 }
 447
 448 static int valid_cc(const void *bc, int len, int cc)
 449 {
 450         while (len >= 0) {
 451                 const struct inet_diag_bc_op *op = bc;
 452
 453                 if (cc > len)
 454                         return 0;
 455                 if (cc == len)
 456                         return 1;
 457                 if (op->yes < 4 || op->yes & 3)
 458                         return 0;
 459                 len -= op->yes;
 460                 bc  += op->yes;
 461         }
 462         return 0;
 463 }
 464
 465 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
 466 {
 467         const void *bc = bytecode;
 468         int  len = bytecode_len;
 469
 470         while (len > 0) {
 471                 const struct inet_diag_bc_op *op = bc;
 472
 473 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 474                 switch (op->code) {
 475                 case INET_DIAG_BC_AUTO:
 476                 case INET_DIAG_BC_S_COND:
 477                 case INET_DIAG_BC_D_COND:
 478                 case INET_DIAG_BC_S_GE:
 479                 case INET_DIAG_BC_S_LE:
 480                 case INET_DIAG_BC_D_GE:
 481                 case INET_DIAG_BC_D_LE:
 482                 case INET_DIAG_BC_JMP:
 483                         if (op->no < 4 || op->no > len + 4 || op->no & 3)
 484                                 return -EINVAL;
 485                         if (op->no < len &&
 486                             !valid_cc(bytecode, bytecode_len, len - op->no))
 487                                 return -EINVAL;
 488                         break;
 489                 case INET_DIAG_BC_NOP:
 490                         break;
 491                 default:
 492                         return -EINVAL;
 493                 }
 494                 if (op->yes < 4 || op->yes > len + 4 || op->yes & 3)
 495                         return -EINVAL;
 496                 bc  += op->yes;
 497                 len -= op->yes;
 498         }
 499         return len == 0 ? 0 : -EINVAL;
 500 }
 501
 502 static int inet_csk_diag_dump(struct sock *sk,
 503                               struct sk_buff *skb,
 504                               struct netlink_callback *cb,
 505                               const struct nlattr *bc)
 506 {
 507         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 508
 509         if (bc != NULL) {
 510                 struct inet_diag_entry entry;
 511                 struct inet_sock *inet = inet_sk(sk);
 512
 513                 entry.family = sk->sk_family;
 514 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 515                 if (entry.family == AF_INET6) {
 516                         struct ipv6_pinfo *np = inet6_sk(sk);
 517
 518                         entry.saddr = np->rcv_saddr.s6_addr32;
 519                         entry.daddr = np->daddr.s6_addr32;
 520                 } else
 521 #endif
 522                 {
 523                         entry.saddr = &inet->inet_rcv_saddr;
 524                         entry.daddr = &inet->inet_daddr;
 525                 }
 526                 entry.sport = inet->inet_num;
 527                 entry.dport = ntohs(inet->inet_dport);
 528                 entry.userlocks = sk->sk_userlocks;
 529
 530                 if (!inet_diag_bc_run(nla_data(bc), nla_len(bc), &entry))
 531                         return 0;
 532         }
 533
 534         return inet_csk_diag_fill(sk, skb, r->idiag_ext,
 535                                   NETLINK_CB(cb->skb).pid,
 536                                   cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 537 }
 538
 539 static int inet_twsk_diag_dump(struct inet_timewait_sock *tw,
 540                                struct sk_buff *skb,
 541                                struct netlink_callback *cb,
 542                                const struct nlattr *bc)
 543 {
 544         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 545
 546         if (bc != NULL) {
 547                 struct inet_diag_entry entry;
 548
 549                 entry.family = tw->tw_family;
 550 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 551                 if (tw->tw_family == AF_INET6) {
 552                         struct inet6_timewait_sock *tw6 =
 553                                                 inet6_twsk((struct sock *)tw);
 554                         entry.saddr = tw6->tw_v6_rcv_saddr.s6_addr32;
 555                         entry.daddr = tw6->tw_v6_daddr.s6_addr32;
 556                 } else
 557 #endif
 558                 {
 559                         entry.saddr = &tw->tw_rcv_saddr;
 560                         entry.daddr = &tw->tw_daddr;
 561                 }
 562                 entry.sport = tw->tw_num;
 563                 entry.dport = ntohs(tw->tw_dport);
 564                 entry.userlocks = 0;
 565
 566                 if (!inet_diag_bc_run(nla_data(bc), nla_len(bc), &entry))
 567                         return 0;
 568         }
 569
 570         return inet_twsk_diag_fill(tw, skb, r->idiag_ext,
 571                                    NETLINK_CB(cb->skb).pid,
 572                                    cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 573 }
 574
 575 static int inet_diag_fill_req(struct sk_buff *skb, struct sock *sk,
 576                               struct request_sock *req, u32 pid, u32 seq,
 577                               const struct nlmsghdr *unlh)
 578 {
 579         const struct inet_request_sock *ireq = inet_rsk(req);
 580         struct inet_sock *inet = inet_sk(sk);
 581         unsigned char *b = skb_tail_pointer(skb);
 582         struct inet_diag_msg *r;
 583         struct nlmsghdr *nlh;
 584         long tmo;
 585
 586         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
 587         nlh->nlmsg_flags = NLM_F_MULTI;
 588         r = NLMSG_DATA(nlh);
 589
 590         r->idiag_family = sk->sk_family;
 591         r->idiag_state = TCP_SYN_RECV;
 592         r->idiag_timer = 1;
 593         r->idiag_retrans = req->retrans;
 594
 595         r->id.idiag_if = sk->sk_bound_dev_if;
 596         r->id.idiag_cookie[0] = (u32)(unsigned long)req;
 597         r->id.idiag_cookie[1] = (u32)(((unsigned long)req >> 31) >> 1);
 598
 599         tmo = req->expires - jiffies;
 600         if (tmo < 0)
 601                 tmo = 0;
 602
 603         r->id.idiag_sport = inet->inet_sport;
 604         r->id.idiag_dport = ireq->rmt_port;
 605         r->id.idiag_src[0] = ireq->loc_addr;
 606         r->id.idiag_dst[0] = ireq->rmt_addr;
 607         r->idiag_expires = jiffies_to_msecs(tmo);
 608         r->idiag_rqueue = 0;
 609         r->idiag_wqueue = 0;
 610         r->idiag_uid = sock_i_uid(sk);
 611         r->idiag_inode = 0;
 612 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 613         if (r->idiag_family == AF_INET6) {
 614                 *(struct in6_addr *)r->id.idiag_src = inet6_rsk(req)->loc_addr;
 615                 *(struct in6_addr *)r->id.idiag_dst = inet6_rsk(req)->rmt_addr;
 616         }
 617 #endif
 618         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 619
 620         return skb->len;
 621
 622 nlmsg_failure:
 623         nlmsg_trim(skb, b);
 624         return -1;
 625 }
 626
 627 static int inet_diag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 628                                struct netlink_callback *cb,
 629                                const struct nlattr *bc)
 630 {
 631         struct inet_diag_entry entry;
 632         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 633         struct inet_connection_sock *icsk = inet_csk(sk);
 634         struct listen_sock *lopt;
 635         struct inet_sock *inet = inet_sk(sk);
 636         int j, s_j;
 637         int reqnum, s_reqnum;
 638         int err = 0;
 639
 640         s_j = cb->args[3];
 641         s_reqnum = cb->args[4];
 642
 643         if (s_j > 0)
 644                 s_j--;
 645
 646         entry.family = sk->sk_family;
 647
 648         read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 649
 650         lopt = icsk->icsk_accept_queue.listen_opt;
 651         if (!lopt || !lopt->qlen)
 652                 goto out;
 653
 654         if (bc != NULL) {
 655                 entry.sport = inet->inet_num;
 656                 entry.userlocks = sk->sk_userlocks;
 657         }
 658
 659         for (j = s_j; j < lopt->nr_table_entries; j++) {
 660                 struct request_sock *req, *head = lopt->syn_table[j];
 661
 662                 reqnum = 0;
 663                 for (req = head; req; reqnum++, req = req->dl_next) {
 664                         struct inet_request_sock *ireq = inet_rsk(req);
 665
 666                         if (reqnum < s_reqnum)
 667                                 continue;
 668                         if (r->id.idiag_dport != ireq->rmt_port &&
 669                             r->id.idiag_dport)
 670                                 continue;
 671
 672                         if (bc) {
 673                                 entry.saddr =
 674 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 675                                         (entry.family == AF_INET6) ?
 676                                         inet6_rsk(req)->loc_addr.s6_addr32 :
 677 #endif
 678                                         &ireq->loc_addr;
 679                                 entry.daddr =
 680 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 681                                         (entry.family == AF_INET6) ?
 682                                         inet6_rsk(req)->rmt_addr.s6_addr32 :
 683 #endif
 684                                         &ireq->rmt_addr;
 685                                 entry.dport = ntohs(ireq->rmt_port);
 686
 687                                 if (!inet_diag_bc_run(nla_data(bc),
 688                                                       nla_len(bc), &entry))
 689                                         continue;
 690                         }
 691
 692                         err = inet_diag_fill_req(skb, sk, req,
 693                                                NETLINK_CB(cb->skb).pid,
 694                                                cb->nlh->nlmsg_seq, cb->nlh);
 695                         if (err < 0) {
 696                                 cb->args[3] = j + 1;
 697                                 cb->args[4] = reqnum;
 698                                 goto out;
 699                         }
 700                 }
 701
 702                 s_reqnum = 0;
 703         }
 704
 705 out:
 706         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 707
 708         return err;
 709 }
 710
 711 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 712 {
 713         int i, num;
 714         int s_i, s_num;
 715         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 716         const struct inet_diag_handler *handler;
 717         struct inet_hashinfo *hashinfo;
 718         const struct nlattr *bc = NULL;
 719
 720         if (nlmsg_attrlen(cb->nlh, sizeof(struct inet_diag_req)))
 721                 bc = nlmsg_find_attr(cb->nlh, sizeof(*r), INET_DIAG_REQ_BYTECODE);
 722
 723         handler = inet_diag_lock_handler(inet_diag_type2proto(cb->nlh->nlmsg_type));
 724         if (IS_ERR(handler))
 725                 goto unlock;
 726
 727         hashinfo = handler->idiag_hashinfo;
 728
 729         s_i = cb->args[1];
 730         s_num = num = cb->args[2];
 731
 732         if (cb->args[0] == 0) {
 733                 if (!(r->idiag_states & (TCPF_LISTEN | TCPF_SYN_RECV)))
 734                         goto skip_listen_ht;
 735
 736                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 737                         struct sock *sk;
 738                         struct hlist_nulls_node *node;
 739                         struct inet_listen_hashbucket *ilb;
 740
 741                         num = 0;
 742                         ilb = &hashinfo->listening_hash[i];
 743                         spin_lock_bh(&ilb->lock);
 744                         sk_nulls_for_each(sk, node, &ilb->head) {
 745                                 struct inet_sock *inet = inet_sk(sk);
 746
 747                                 if (num < s_num) {
 748                                         num++;
 749                                         continue;
 750                                 }
 751
 752                                 if (r->id.idiag_sport != inet->inet_sport &&
 753                                     r->id.idiag_sport)
 754                                         goto next_listen;
 755
 756                                 if (!(r->idiag_states & TCPF_LISTEN) ||
 757                                     r->id.idiag_dport ||
 758                                     cb->args[3] > 0)
 759                                         goto syn_recv;
 760
 761                                 if (inet_csk_diag_dump(sk, skb, cb, bc) < 0) {
 762                                         spin_unlock_bh(&ilb->lock);
 763                                         goto done;
 764                                 }
 765
 766 syn_recv:
 767                                 if (!(r->idiag_states & TCPF_SYN_RECV))
 768                                         goto next_listen;
 769
 770                                 if (inet_diag_dump_reqs(skb, sk, cb, bc) < 0) {
 771                                         spin_unlock_bh(&ilb->lock);
 772                                         goto done;
 773                                 }
 774
 775 next_listen:
 776                                 cb->args[3] = 0;
 777                                 cb->args[4] = 0;
 778                                 ++num;
 779                         }
 780                         spin_unlock_bh(&ilb->lock);
 781
 782                         s_num = 0;
 783                         cb->args[3] = 0;
 784                         cb->args[4] = 0;
 785                 }
 786 skip_listen_ht:
 787                 cb->args[0] = 1;
 788                 s_i = num = s_num = 0;
 789         }
 790
 791         if (!(r->idiag_states & ~(TCPF_LISTEN | TCPF_SYN_RECV)))
 792                 goto unlock;
 793
 794         for (i = s_i; i <= hashinfo->ehash_mask; i++) {
 795                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 796                 spinlock_t *lock = inet_ehash_lockp(hashinfo, i);
 797                 struct sock *sk;
 798                 struct hlist_nulls_node *node;
 799
 800                 num = 0;
 801
 802                 if (hlist_nulls_empty(&head->chain) &&
 803                         hlist_nulls_empty(&head->twchain))
 804                         continue;
 805
 806                 if (i > s_i)
 807                         s_num = 0;
 808
 809                 spin_lock_bh(lock);
 810                 sk_nulls_for_each(sk, node, &head->chain) {
 811                         struct inet_sock *inet = inet_sk(sk);
 812
 813                         if (num < s_num)
 814                                 goto next_normal;
 815                         if (!(r->idiag_states & (1 << sk->sk_state)))
 816                                 goto next_normal;
 817                         if (r->id.idiag_sport != inet->inet_sport &&
 818                             r->id.idiag_sport)
 819                                 goto next_normal;
 820                         if (r->id.idiag_dport != inet->inet_dport &&
 821                             r->id.idiag_dport)
 822                                 goto next_normal;
 823                         if (inet_csk_diag_dump(sk, skb, cb, bc) < 0) {
 824                                 spin_unlock_bh(lock);
 825                                 goto done;
 826                         }
 827 next_normal:
 828                         ++num;
 829                 }
 830
 831                 if (r->idiag_states & TCPF_TIME_WAIT) {
 832                         struct inet_timewait_sock *tw;
 833
 834                         inet_twsk_for_each(tw, node,
 835                                     &head->twchain) {
 836
 837                                 if (num < s_num)
 838                                         goto next_dying;
 839                                 if (r->id.idiag_sport != tw->tw_sport &&
 840                                     r->id.idiag_sport)
 841                                         goto next_dying;
 842                                 if (r->id.idiag_dport != tw->tw_dport &&
 843                                     r->id.idiag_dport)
 844                                         goto next_dying;
 845                                 if (inet_twsk_diag_dump(tw, skb, cb, bc) < 0) {
 846                                         spin_unlock_bh(lock);
 847                                         goto done;
 848                                 }
 849 next_dying:
 850                                 ++num;
 851                         }
 852                 }
 853                 spin_unlock_bh(lock);
 854         }
 855
 856 done:
 857         cb->args[1] = i;
 858         cb->args[2] = num;
 859 unlock:
 860         inet_diag_unlock_handler(handler);
 861         return skb->len;
 862 }
 863
 864 static int inet_diag_rcv_msg_compat(struct sk_buff *skb, struct nlmsghdr *nlh)
 865 {
 866         int hdrlen = sizeof(struct inet_diag_req);
 867
 868         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
 869             nlmsg_len(nlh) < hdrlen)
 870                 return -EINVAL;
 871
 872         if (nlh->nlmsg_flags & NLM_F_DUMP) {
 873                 if (nlmsg_attrlen(nlh, hdrlen)) {
 874                         struct nlattr *attr;
 875
 876                         attr = nlmsg_find_attr(nlh, hdrlen,
 877                                                INET_DIAG_REQ_BYTECODE);
 878                         if (attr == NULL ||
 879                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
 880                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
 881                                 return -EINVAL;
 882                 }
 883
 884                 return netlink_dump_start(sdiagnl, skb, nlh,
 885                                           inet_diag_dump, NULL, 0);
 886         }
 887
 888         return inet_diag_get_exact(skb, nlh);
 889 }
 890
 891 static int inet_diag_handler_dump(struct sk_buff *skb, struct nlmsghdr *h)
 892 {
 893         int hdrlen = sizeof(struct inet_diag_req);
 894
 895         if (nlmsg_len(h) < hdrlen)
 896                 return -EINVAL;
 897
 898         if (h->nlmsg_flags & NLM_F_DUMP) {
 899                 return -EAFNOSUPPORT;
 900         }
 901
 902         return -EAFNOSUPPORT;
 903 }
 904
 905 static struct sock_diag_handler inet_diag_handler = {
 906         .family = AF_INET,
 907         .dump = inet_diag_handler_dump,
 908 };
 909
 910 static struct sock_diag_handler inet6_diag_handler = {
 911         .family = AF_INET6,
 912         .dump = inet_diag_handler_dump,
 913 };
 914
 915 static struct sock_diag_handler *sock_diag_handlers[AF_MAX];
 916 static DEFINE_MUTEX(sock_diag_table_mutex);
 917
 918 int sock_diag_register(struct sock_diag_handler *hndl)
 919 {
 920         int err = 0;
 921
 922         if (hndl->family > AF_MAX)
 923                 return -EINVAL;
 924
 925         mutex_lock(&sock_diag_table_mutex);
 926         if (sock_diag_handlers[hndl->family])
 927                 err = -EBUSY;
 928         else
 929                 sock_diag_handlers[hndl->family] = hndl;
 930         mutex_unlock(&sock_diag_table_mutex);
 931
 932         return err;
 933 }
 934
 935 void sock_diag_unregister(struct sock_diag_handler *hnld)
 936 {
 937         int family = hnld->family;
 938
 939         if (family > AF_MAX)
 940                 return;
 941
 942         mutex_lock(&sock_diag_table_mutex);
 943         BUG_ON(sock_diag_handlers[family] != hnld);
 944         sock_diag_handlers[family] = NULL;
 945         mutex_unlock(&sock_diag_table_mutex);
 946 }
 947
 948 static inline struct sock_diag_handler *sock_diag_lock_handler(int family)
 949 {
 950         mutex_lock(&sock_diag_table_mutex);
 951         return sock_diag_handlers[family];
 952 }
 953
 954 static inline void sock_diag_unlock_handler(struct sock_diag_handler *h)
 955 {
 956         mutex_unlock(&sock_diag_table_mutex);
 957 }
 958
 959 static int __sock_diag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 960 {
 961         int err;
 962         struct sock_diag_req *req = NLMSG_DATA(nlh);
 963         struct sock_diag_handler *hndl;
 964
 965         if (nlmsg_len(nlh) < sizeof(*req))
 966                 return -EINVAL;
 967
 968         hndl = sock_diag_lock_handler(req->sdiag_family);
 969         if (hndl == NULL)
 970                 err = -ENOENT;
 971         else
 972                 err = hndl->dump(skb, nlh);
 973         sock_diag_unlock_handler(hndl);
 974
 975         return err;
 976 }
 977
 978 static int sock_diag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 979 {
 980         switch (nlh->nlmsg_type) {
 981         case TCPDIAG_GETSOCK:
 982         case DCCPDIAG_GETSOCK:
 983                 return inet_diag_rcv_msg_compat(skb, nlh);
 984         case SOCK_DIAG_BY_FAMILY:
 985                 return __sock_diag_rcv_msg(skb, nlh);
 986         default:
 987                 return -EINVAL;
 988         }
 989 }
 990
 991 static DEFINE_MUTEX(sock_diag_mutex);
 992
 993 static void sock_diag_rcv(struct sk_buff *skb)
 994 {
 995         mutex_lock(&sock_diag_mutex);
 996         netlink_rcv_skb(skb, &sock_diag_rcv_msg);
 997         mutex_unlock(&sock_diag_mutex);
 998 }
 999
1000 int inet_diag_register(const struct inet_diag_handler *h)
1001 {
1002         const __u16 type = h->idiag_type;
1003         int err = -EINVAL;
1004
1005         if (type >= IPPROTO_MAX)
1006                 goto out;
1007
1008         mutex_lock(&inet_diag_table_mutex);
1009         err = -EEXIST;
1010         if (inet_diag_table[type] == NULL) {
1011                 inet_diag_table[type] = h;
1012                 err = 0;
1013         }
1014         mutex_unlock(&inet_diag_table_mutex);
1015 out:
1016         return err;
1017 }
1018 EXPORT_SYMBOL_GPL(inet_diag_register);
1019
1020 void inet_diag_unregister(const struct inet_diag_handler *h)
1021 {
1022         const __u16 type = h->idiag_type;
1023
1024         if (type >= IPPROTO_MAX)
1025                 return;
1026
1027         mutex_lock(&inet_diag_table_mutex);
1028         inet_diag_table[type] = NULL;
1029         mutex_unlock(&inet_diag_table_mutex);
1030 }
1031 EXPORT_SYMBOL_GPL(inet_diag_unregister);
1032
1033 static int __init inet_diag_init(void)
1034 {
1035         const int inet_diag_table_size = (IPPROTO_MAX *
1036                                           sizeof(struct inet_diag_handler *));
1037         int err = -ENOMEM;
1038
1039         inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
1040         if (!inet_diag_table)
1041                 goto out;
1042
1043         sdiagnl = netlink_kernel_create(&init_net, NETLINK_SOCK_DIAG, 0,
1044                                         sock_diag_rcv, NULL, THIS_MODULE);
1045         if (sdiagnl == NULL)
1046                 goto out_free_table;
1047
1048         err = sock_diag_register(&inet_diag_handler);
1049         if (err)
1050                 goto out_free_nl;
1051
1052         err = sock_diag_register(&inet6_diag_handler);
1053         if (err)
1054                 goto out_free_inet;
1055
1056 out:
1057         return err;
1058
1059 out_free_inet:
1060         sock_diag_unregister(&inet_diag_handler);
1061 out_free_nl:
1062         netlink_kernel_release(sdiagnl);
1063 out_free_table:
1064         kfree(inet_diag_table);
1065         goto out;
1066 }
1067
1068 static void __exit inet_diag_exit(void)
1069 {
1070         sock_diag_unregister(&inet6_diag_handler);
1071         sock_diag_unregister(&inet_diag_handler);
1072         netlink_kernel_release(sdiagnl);
1073         kfree(inet_diag_table);
1074 }
1075
1076 module_init(inet_diag_init);
1077 module_exit(inet_diag_exit);
1078 MODULE_LICENSE("GPL");
1079 MODULE_ALIAS_NET_PF_PROTO(PF_NETLINK, NETLINK_SOCK_DIAG);