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
  37 static const struct inet_diag_handler **inet_diag_table;
  38
  39 struct inet_diag_entry {
  40         __be32 *saddr;
  41         __be32 *daddr;
  42         u16 sport;
  43         u16 dport;
  44         u16 family;
  45         u16 userlocks;
  46 };
  47
  48 static struct sock *idiagnl;
  49
  50 #define INET_DIAG_PUT(skb, attrtype, attrlen) \
  51         RTA_DATA(__RTA_PUT(skb, attrtype, attrlen))
  52
  53 static DEFINE_MUTEX(inet_diag_table_mutex);
  54
  55 static const struct inet_diag_handler *inet_diag_lock_handler(int type)
  56 {
  57         if (!inet_diag_table[type])
  58                 request_module("net-pf-%d-proto-%d-type-%d", PF_NETLINK,
  59                                NETLINK_INET_DIAG, type);
  60
  61         mutex_lock(&inet_diag_table_mutex);
  62         if (!inet_diag_table[type])
  63                 return ERR_PTR(-ENOENT);
  64
  65         return inet_diag_table[type];
  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 int inet_csk_diag_fill(struct sock *sk,
  75                               struct sk_buff *skb,
  76                               int ext, u32 pid, u32 seq, u16 nlmsg_flags,
  77                               const struct nlmsghdr *unlh)
  78 {
  79         const struct inet_sock *inet = inet_sk(sk);
  80         const struct inet_connection_sock *icsk = inet_csk(sk);
  81         struct inet_diag_msg *r;
  82         struct nlmsghdr  *nlh;
  83         void *info = NULL;
  84         struct inet_diag_meminfo  *minfo = NULL;
  85         unsigned char    *b = skb_tail_pointer(skb);
  86         const struct inet_diag_handler *handler;
  87
  88         handler = inet_diag_table[unlh->nlmsg_type];
  89         BUG_ON(handler == NULL);
  90
  91         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
  92         nlh->nlmsg_flags = nlmsg_flags;
  93
  94         r = NLMSG_DATA(nlh);
  95         BUG_ON(sk->sk_state == TCP_TIME_WAIT);
  96
  97         if (ext & (1 << (INET_DIAG_MEMINFO - 1)))
  98                 minfo = INET_DIAG_PUT(skb, INET_DIAG_MEMINFO, sizeof(*minfo));
  99
 100         if (ext & (1 << (INET_DIAG_INFO - 1)))
 101                 info = INET_DIAG_PUT(skb, INET_DIAG_INFO,
 102                                      handler->idiag_info_size);
 103
 104         if ((ext & (1 << (INET_DIAG_CONG - 1))) && icsk->icsk_ca_ops) {
 105                 const size_t len = strlen(icsk->icsk_ca_ops->name);
 106
 107                 strcpy(INET_DIAG_PUT(skb, INET_DIAG_CONG, len + 1),
 108                        icsk->icsk_ca_ops->name);
 109         }
 110
 111         r->idiag_family = sk->sk_family;
 112         r->idiag_state = sk->sk_state;
 113         r->idiag_timer = 0;
 114         r->idiag_retrans = 0;
 115
 116         r->id.idiag_if = sk->sk_bound_dev_if;
 117         r->id.idiag_cookie[0] = (u32)(unsigned long)sk;
 118         r->id.idiag_cookie[1] = (u32)(((unsigned long)sk >> 31) >> 1);
 119
 120         r->id.idiag_sport = inet->inet_sport;
 121         r->id.idiag_dport = inet->inet_dport;
 122         r->id.idiag_src[0] = inet->inet_rcv_saddr;
 123         r->id.idiag_dst[0] = inet->inet_daddr;
 124
 125         /* IPv6 dual-stack sockets use inet->tos for IPv4 connections,
 126          * hence this needs to be included regardless of socket family.
 127          */
 128         if (ext & (1 << (INET_DIAG_TOS - 1)))
 129                 RTA_PUT_U8(skb, INET_DIAG_TOS, inet->tos);
 130
 131 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 132         if (r->idiag_family == AF_INET6) {
 133                 const struct ipv6_pinfo *np = inet6_sk(sk);
 134
 135                 if (ext & (1 << (INET_DIAG_TCLASS - 1)))
 136                         RTA_PUT_U8(skb, INET_DIAG_TCLASS, np->tclass);
 137
 138                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 139                                &np->rcv_saddr);
 140                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 141                                &np->daddr);
 142         }
 143 #endif
 144
 145 #define EXPIRES_IN_MS(tmo)  DIV_ROUND_UP((tmo - jiffies) * 1000, HZ)
 146
 147         if (icsk->icsk_pending == ICSK_TIME_RETRANS) {
 148                 r->idiag_timer = 1;
 149                 r->idiag_retrans = icsk->icsk_retransmits;
 150                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 151         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 152                 r->idiag_timer = 4;
 153                 r->idiag_retrans = icsk->icsk_probes_out;
 154                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 155         } else if (timer_pending(&sk->sk_timer)) {
 156                 r->idiag_timer = 2;
 157                 r->idiag_retrans = icsk->icsk_probes_out;
 158                 r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 159         } else {
 160                 r->idiag_timer = 0;
 161                 r->idiag_expires = 0;
 162         }
 163 #undef EXPIRES_IN_MS
 164
 165         r->idiag_uid = sock_i_uid(sk);
 166         r->idiag_inode = sock_i_ino(sk);
 167
 168         if (minfo) {
 169                 minfo->idiag_rmem = sk_rmem_alloc_get(sk);
 170                 minfo->idiag_wmem = sk->sk_wmem_queued;
 171                 minfo->idiag_fmem = sk->sk_forward_alloc;
 172                 minfo->idiag_tmem = sk_wmem_alloc_get(sk);
 173         }
 174
 175         handler->idiag_get_info(sk, r, info);
 176
 177         if (sk->sk_state < TCP_TIME_WAIT &&
 178             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 179                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 180
 181         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 182         return skb->len;
 183
 184 rtattr_failure:
 185 nlmsg_failure:
 186         nlmsg_trim(skb, b);
 187         return -EMSGSIZE;
 188 }
 189
 190 static int inet_twsk_diag_fill(struct inet_timewait_sock *tw,
 191                                struct sk_buff *skb, int ext, u32 pid,
 192                                u32 seq, u16 nlmsg_flags,
 193                                const struct nlmsghdr *unlh)
 194 {
 195         long tmo;
 196         struct inet_diag_msg *r;
 197         const unsigned char *previous_tail = skb_tail_pointer(skb);
 198         struct nlmsghdr *nlh = NLMSG_PUT(skb, pid, seq,
 199                                          unlh->nlmsg_type, sizeof(*r));
 200
 201         r = NLMSG_DATA(nlh);
 202         BUG_ON(tw->tw_state != TCP_TIME_WAIT);
 203
 204         nlh->nlmsg_flags = nlmsg_flags;
 205
 206         tmo = tw->tw_ttd - jiffies;
 207         if (tmo < 0)
 208                 tmo = 0;
 209
 210         r->idiag_family       = tw->tw_family;
 211         r->idiag_retrans      = 0;
 212         r->id.idiag_if        = tw->tw_bound_dev_if;
 213         r->id.idiag_cookie[0] = (u32)(unsigned long)tw;
 214         r->id.idiag_cookie[1] = (u32)(((unsigned long)tw >> 31) >> 1);
 215         r->id.idiag_sport     = tw->tw_sport;
 216         r->id.idiag_dport     = tw->tw_dport;
 217         r->id.idiag_src[0]    = tw->tw_rcv_saddr;
 218         r->id.idiag_dst[0]    = tw->tw_daddr;
 219         r->idiag_state        = tw->tw_substate;
 220         r->idiag_timer        = 3;
 221         r->idiag_expires      = DIV_ROUND_UP(tmo * 1000, HZ);
 222         r->idiag_rqueue       = 0;
 223         r->idiag_wqueue       = 0;
 224         r->idiag_uid          = 0;
 225         r->idiag_inode        = 0;
 226 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 227         if (tw->tw_family == AF_INET6) {
 228                 const struct inet6_timewait_sock *tw6 =
 229                                                 inet6_twsk((struct sock *)tw);
 230
 231                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 232                                &tw6->tw_v6_rcv_saddr);
 233                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 234                                &tw6->tw_v6_daddr);
 235         }
 236 #endif
 237         nlh->nlmsg_len = skb_tail_pointer(skb) - previous_tail;
 238         return skb->len;
 239 nlmsg_failure:
 240         nlmsg_trim(skb, previous_tail);
 241         return -EMSGSIZE;
 242 }
 243
 244 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
 245                         int ext, u32 pid, u32 seq, u16 nlmsg_flags,
 246                         const struct nlmsghdr *unlh)
 247 {
 248         if (sk->sk_state == TCP_TIME_WAIT)
 249                 return inet_twsk_diag_fill((struct inet_timewait_sock *)sk,
 250                                            skb, ext, pid, seq, nlmsg_flags,
 251                                            unlh);
 252         return inet_csk_diag_fill(sk, skb, ext, pid, seq, nlmsg_flags, unlh);
 253 }
 254
 255 static int inet_diag_get_exact(struct sk_buff *in_skb,
 256                                const struct nlmsghdr *nlh)
 257 {
 258         int err;
 259         struct sock *sk;
 260         struct inet_diag_req *req = NLMSG_DATA(nlh);
 261         struct sk_buff *rep;
 262         struct inet_hashinfo *hashinfo;
 263         const struct inet_diag_handler *handler;
 264
 265         handler = inet_diag_lock_handler(nlh->nlmsg_type);
 266         if (IS_ERR(handler)) {
 267                 err = PTR_ERR(handler);
 268                 goto unlock;
 269         }
 270
 271         hashinfo = handler->idiag_hashinfo;
 272         err = -EINVAL;
 273
 274         if (req->idiag_family == AF_INET) {
 275                 sk = inet_lookup(&init_net, hashinfo, req->id.idiag_dst[0],
 276                                  req->id.idiag_dport, req->id.idiag_src[0],
 277                                  req->id.idiag_sport, req->id.idiag_if);
 278         }
 279 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 280         else if (req->idiag_family == AF_INET6) {
 281                 sk = inet6_lookup(&init_net, hashinfo,
 282                                   (struct in6_addr *)req->id.idiag_dst,
 283                                   req->id.idiag_dport,
 284                                   (struct in6_addr *)req->id.idiag_src,
 285                                   req->id.idiag_sport,
 286                                   req->id.idiag_if);
 287         }
 288 #endif
 289         else {
 290                 goto unlock;
 291         }
 292
 293         err = -ENOENT;
 294         if (sk == NULL)
 295                 goto unlock;
 296
 297         err = -ESTALE;
 298         if ((req->id.idiag_cookie[0] != INET_DIAG_NOCOOKIE ||
 299              req->id.idiag_cookie[1] != INET_DIAG_NOCOOKIE) &&
 300             ((u32)(unsigned long)sk != req->id.idiag_cookie[0] ||
 301              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.idiag_cookie[1]))
 302                 goto out;
 303
 304         err = -ENOMEM;
 305         rep = alloc_skb(NLMSG_SPACE((sizeof(struct inet_diag_msg) +
 306                                      sizeof(struct inet_diag_meminfo) +
 307                                      handler->idiag_info_size + 64)),
 308                         GFP_KERNEL);
 309         if (!rep)
 310                 goto out;
 311
 312         err = sk_diag_fill(sk, rep, req->idiag_ext,
 313                            NETLINK_CB(in_skb).pid,
 314                            nlh->nlmsg_seq, 0, nlh);
 315         if (err < 0) {
 316                 WARN_ON(err == -EMSGSIZE);
 317                 kfree_skb(rep);
 318                 goto out;
 319         }
 320         err = netlink_unicast(idiagnl, rep, NETLINK_CB(in_skb).pid,
 321                               MSG_DONTWAIT);
 322         if (err > 0)
 323                 err = 0;
 324
 325 out:
 326         if (sk) {
 327                 if (sk->sk_state == TCP_TIME_WAIT)
 328                         inet_twsk_put((struct inet_timewait_sock *)sk);
 329                 else
 330                         sock_put(sk);
 331         }
 332 unlock:
 333         inet_diag_unlock_handler(handler);
 334         return err;
 335 }
 336
 337 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
 338 {
 339         int words = bits >> 5;
 340
 341         bits &= 0x1f;
 342
 343         if (words) {
 344                 if (memcmp(a1, a2, words << 2))
 345                         return 0;
 346         }
 347         if (bits) {
 348                 __be32 w1, w2;
 349                 __be32 mask;
 350
 351                 w1 = a1[words];
 352                 w2 = a2[words];
 353
 354                 mask = htonl((0xffffffff) << (32 - bits));
 355
 356                 if ((w1 ^ w2) & mask)
 357                         return 0;
 358         }
 359
 360         return 1;
 361 }
 362
 363
 364 static int inet_diag_bc_run(const void *bc, int len,
 365                             const struct inet_diag_entry *entry)
 366 {
 367         while (len > 0) {
 368                 int yes = 1;
 369                 const struct inet_diag_bc_op *op = bc;
 370
 371                 switch (op->code) {
 372                 case INET_DIAG_BC_NOP:
 373                         break;
 374                 case INET_DIAG_BC_JMP:
 375                         yes = 0;
 376                         break;
 377                 case INET_DIAG_BC_S_GE:
 378                         yes = entry->sport >= op[1].no;
 379                         break;
 380                 case INET_DIAG_BC_S_LE:
 381                         yes = entry->sport <= op[1].no;
 382                         break;
 383                 case INET_DIAG_BC_D_GE:
 384                         yes = entry->dport >= op[1].no;
 385                         break;
 386                 case INET_DIAG_BC_D_LE:
 387                         yes = entry->dport <= op[1].no;
 388                         break;
 389                 case INET_DIAG_BC_AUTO:
 390                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 391                         break;
 392                 case INET_DIAG_BC_S_COND:
 393                 case INET_DIAG_BC_D_COND: {
 394                         struct inet_diag_hostcond *cond;
 395                         __be32 *addr;
 396
 397                         cond = (struct inet_diag_hostcond *)(op + 1);
 398                         if (cond->port != -1 &&
 399                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 400                                              entry->sport : entry->dport)) {
 401                                 yes = 0;
 402                                 break;
 403                         }
 404
 405                         if (cond->prefix_len == 0)
 406                                 break;
 407
 408                         if (op->code == INET_DIAG_BC_S_COND)
 409                                 addr = entry->saddr;
 410                         else
 411                                 addr = entry->daddr;
 412
 413                         if (bitstring_match(addr, cond->addr,
 414                                             cond->prefix_len))
 415                                 break;
 416                         if (entry->family == AF_INET6 &&
 417                             cond->family == AF_INET) {
 418                                 if (addr[0] == 0 && addr[1] == 0 &&
 419                                     addr[2] == htonl(0xffff) &&
 420                                     bitstring_match(addr + 3, cond->addr,
 421                                                     cond->prefix_len))
 422                                         break;
 423                         }
 424                         yes = 0;
 425                         break;
 426                 }
 427                 }
 428
 429                 if (yes) {
 430                         len -= op->yes;
 431                         bc += op->yes;
 432                 } else {
 433                         len -= op->no;
 434                         bc += op->no;
 435                 }
 436         }
 437         return len == 0;
 438 }
 439
 440 static int valid_cc(const void *bc, int len, int cc)
 441 {
 442         while (len >= 0) {
 443                 const struct inet_diag_bc_op *op = bc;
 444
 445                 if (cc > len)
 446                         return 0;
 447                 if (cc == len)
 448                         return 1;
 449                 if (op->yes < 4 || op->yes & 3)
 450                         return 0;
 451                 len -= op->yes;
 452                 bc  += op->yes;
 453         }
 454         return 0;
 455 }
 456
 457 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
 458 {
 459         const void *bc = bytecode;
 460         int  len = bytecode_len;
 461
 462         while (len > 0) {
 463                 const struct inet_diag_bc_op *op = bc;
 464
 465 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 466                 switch (op->code) {
 467                 case INET_DIAG_BC_AUTO:
 468                 case INET_DIAG_BC_S_COND:
 469                 case INET_DIAG_BC_D_COND:
 470                 case INET_DIAG_BC_S_GE:
 471                 case INET_DIAG_BC_S_LE:
 472                 case INET_DIAG_BC_D_GE:
 473                 case INET_DIAG_BC_D_LE:
 474                 case INET_DIAG_BC_JMP:
 475                         if (op->no < 4 || op->no > len + 4 || op->no & 3)
 476                                 return -EINVAL;
 477                         if (op->no < len &&
 478                             !valid_cc(bytecode, bytecode_len, len - op->no))
 479                                 return -EINVAL;
 480                         break;
 481                 case INET_DIAG_BC_NOP:
 482                         break;
 483                 default:
 484                         return -EINVAL;
 485                 }
 486                 if (op->yes < 4 || op->yes > len + 4 || op->yes & 3)
 487                         return -EINVAL;
 488                 bc  += op->yes;
 489                 len -= op->yes;
 490         }
 491         return len == 0 ? 0 : -EINVAL;
 492 }
 493
 494 static int inet_csk_diag_dump(struct sock *sk,
 495                               struct sk_buff *skb,
 496                               struct netlink_callback *cb)
 497 {
 498         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 499
 500         if (nlmsg_attrlen(cb->nlh, sizeof(*r))) {
 501                 struct inet_diag_entry entry;
 502                 const struct nlattr *bc = nlmsg_find_attr(cb->nlh,
 503                                                           sizeof(*r),
 504                                                           INET_DIAG_REQ_BYTECODE);
 505                 struct inet_sock *inet = inet_sk(sk);
 506
 507                 entry.family = sk->sk_family;
 508 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 509                 if (entry.family == AF_INET6) {
 510                         struct ipv6_pinfo *np = inet6_sk(sk);
 511
 512                         entry.saddr = np->rcv_saddr.s6_addr32;
 513                         entry.daddr = np->daddr.s6_addr32;
 514                 } else
 515 #endif
 516                 {
 517                         entry.saddr = &inet->inet_rcv_saddr;
 518                         entry.daddr = &inet->inet_daddr;
 519                 }
 520                 entry.sport = inet->inet_num;
 521                 entry.dport = ntohs(inet->inet_dport);
 522                 entry.userlocks = sk->sk_userlocks;
 523
 524                 if (!inet_diag_bc_run(nla_data(bc), nla_len(bc), &entry))
 525                         return 0;
 526         }
 527
 528         return inet_csk_diag_fill(sk, skb, r->idiag_ext,
 529                                   NETLINK_CB(cb->skb).pid,
 530                                   cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 531 }
 532
 533 static int inet_twsk_diag_dump(struct inet_timewait_sock *tw,
 534                                struct sk_buff *skb,
 535                                struct netlink_callback *cb)
 536 {
 537         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 538
 539         if (nlmsg_attrlen(cb->nlh, sizeof(*r))) {
 540                 struct inet_diag_entry entry;
 541                 const struct nlattr *bc = nlmsg_find_attr(cb->nlh,
 542                                                           sizeof(*r),
 543                                                           INET_DIAG_REQ_BYTECODE);
 544
 545                 entry.family = tw->tw_family;
 546 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 547                 if (tw->tw_family == AF_INET6) {
 548                         struct inet6_timewait_sock *tw6 =
 549                                                 inet6_twsk((struct sock *)tw);
 550                         entry.saddr = tw6->tw_v6_rcv_saddr.s6_addr32;
 551                         entry.daddr = tw6->tw_v6_daddr.s6_addr32;
 552                 } else
 553 #endif
 554                 {
 555                         entry.saddr = &tw->tw_rcv_saddr;
 556                         entry.daddr = &tw->tw_daddr;
 557                 }
 558                 entry.sport = tw->tw_num;
 559                 entry.dport = ntohs(tw->tw_dport);
 560                 entry.userlocks = 0;
 561
 562                 if (!inet_diag_bc_run(nla_data(bc), nla_len(bc), &entry))
 563                         return 0;
 564         }
 565
 566         return inet_twsk_diag_fill(tw, skb, r->idiag_ext,
 567                                    NETLINK_CB(cb->skb).pid,
 568                                    cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 569 }
 570
 571 static int inet_diag_fill_req(struct sk_buff *skb, struct sock *sk,
 572                               struct request_sock *req, u32 pid, u32 seq,
 573                               const struct nlmsghdr *unlh)
 574 {
 575         const struct inet_request_sock *ireq = inet_rsk(req);
 576         struct inet_sock *inet = inet_sk(sk);
 577         unsigned char *b = skb_tail_pointer(skb);
 578         struct inet_diag_msg *r;
 579         struct nlmsghdr *nlh;
 580         long tmo;
 581
 582         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
 583         nlh->nlmsg_flags = NLM_F_MULTI;
 584         r = NLMSG_DATA(nlh);
 585
 586         r->idiag_family = sk->sk_family;
 587         r->idiag_state = TCP_SYN_RECV;
 588         r->idiag_timer = 1;
 589         r->idiag_retrans = req->retrans;
 590
 591         r->id.idiag_if = sk->sk_bound_dev_if;
 592         r->id.idiag_cookie[0] = (u32)(unsigned long)req;
 593         r->id.idiag_cookie[1] = (u32)(((unsigned long)req >> 31) >> 1);
 594
 595         tmo = req->expires - jiffies;
 596         if (tmo < 0)
 597                 tmo = 0;
 598
 599         r->id.idiag_sport = inet->inet_sport;
 600         r->id.idiag_dport = ireq->rmt_port;
 601         r->id.idiag_src[0] = ireq->loc_addr;
 602         r->id.idiag_dst[0] = ireq->rmt_addr;
 603         r->idiag_expires = jiffies_to_msecs(tmo);
 604         r->idiag_rqueue = 0;
 605         r->idiag_wqueue = 0;
 606         r->idiag_uid = sock_i_uid(sk);
 607         r->idiag_inode = 0;
 608 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 609         if (r->idiag_family == AF_INET6) {
 610                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 611                                &inet6_rsk(req)->loc_addr);
 612                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 613                                &inet6_rsk(req)->rmt_addr);
 614         }
 615 #endif
 616         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 617
 618         return skb->len;
 619
 620 nlmsg_failure:
 621         nlmsg_trim(skb, b);
 622         return -1;
 623 }
 624
 625 static int inet_diag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 626                                struct netlink_callback *cb)
 627 {
 628         struct inet_diag_entry entry;
 629         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 630         struct inet_connection_sock *icsk = inet_csk(sk);
 631         struct listen_sock *lopt;
 632         const struct nlattr *bc = NULL;
 633         struct inet_sock *inet = inet_sk(sk);
 634         int j, s_j;
 635         int reqnum, s_reqnum;
 636         int err = 0;
 637
 638         s_j = cb->args[3];
 639         s_reqnum = cb->args[4];
 640
 641         if (s_j > 0)
 642                 s_j--;
 643
 644         entry.family = sk->sk_family;
 645
 646         read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 647
 648         lopt = icsk->icsk_accept_queue.listen_opt;
 649         if (!lopt || !lopt->qlen)
 650                 goto out;
 651
 652         if (nlmsg_attrlen(cb->nlh, sizeof(*r))) {
 653                 bc = nlmsg_find_attr(cb->nlh, sizeof(*r),
 654                                      INET_DIAG_REQ_BYTECODE);
 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
 719         handler = inet_diag_lock_handler(cb->nlh->nlmsg_type);
 720         if (IS_ERR(handler))
 721                 goto unlock;
 722
 723         hashinfo = handler->idiag_hashinfo;
 724
 725         s_i = cb->args[1];
 726         s_num = num = cb->args[2];
 727
 728         if (cb->args[0] == 0) {
 729                 if (!(r->idiag_states & (TCPF_LISTEN | TCPF_SYN_RECV)))
 730                         goto skip_listen_ht;
 731
 732                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 733                         struct sock *sk;
 734                         struct hlist_nulls_node *node;
 735                         struct inet_listen_hashbucket *ilb;
 736
 737                         num = 0;
 738                         ilb = &hashinfo->listening_hash[i];
 739                         spin_lock_bh(&ilb->lock);
 740                         sk_nulls_for_each(sk, node, &ilb->head) {
 741                                 struct inet_sock *inet = inet_sk(sk);
 742
 743                                 if (num < s_num) {
 744                                         num++;
 745                                         continue;
 746                                 }
 747
 748                                 if (r->id.idiag_sport != inet->inet_sport &&
 749                                     r->id.idiag_sport)
 750                                         goto next_listen;
 751
 752                                 if (!(r->idiag_states & TCPF_LISTEN) ||
 753                                     r->id.idiag_dport ||
 754                                     cb->args[3] > 0)
 755                                         goto syn_recv;
 756
 757                                 if (inet_csk_diag_dump(sk, skb, cb) < 0) {
 758                                         spin_unlock_bh(&ilb->lock);
 759                                         goto done;
 760                                 }
 761
 762 syn_recv:
 763                                 if (!(r->idiag_states & TCPF_SYN_RECV))
 764                                         goto next_listen;
 765
 766                                 if (inet_diag_dump_reqs(skb, sk, cb) < 0) {
 767                                         spin_unlock_bh(&ilb->lock);
 768                                         goto done;
 769                                 }
 770
 771 next_listen:
 772                                 cb->args[3] = 0;
 773                                 cb->args[4] = 0;
 774                                 ++num;
 775                         }
 776                         spin_unlock_bh(&ilb->lock);
 777
 778                         s_num = 0;
 779                         cb->args[3] = 0;
 780                         cb->args[4] = 0;
 781                 }
 782 skip_listen_ht:
 783                 cb->args[0] = 1;
 784                 s_i = num = s_num = 0;
 785         }
 786
 787         if (!(r->idiag_states & ~(TCPF_LISTEN | TCPF_SYN_RECV)))
 788                 goto unlock;
 789
 790         for (i = s_i; i <= hashinfo->ehash_mask; i++) {
 791                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 792                 spinlock_t *lock = inet_ehash_lockp(hashinfo, i);
 793                 struct sock *sk;
 794                 struct hlist_nulls_node *node;
 795
 796                 num = 0;
 797
 798                 if (hlist_nulls_empty(&head->chain) &&
 799                         hlist_nulls_empty(&head->twchain))
 800                         continue;
 801
 802                 if (i > s_i)
 803                         s_num = 0;
 804
 805                 spin_lock_bh(lock);
 806                 sk_nulls_for_each(sk, node, &head->chain) {
 807                         struct inet_sock *inet = inet_sk(sk);
 808
 809                         if (num < s_num)
 810                                 goto next_normal;
 811                         if (!(r->idiag_states & (1 << sk->sk_state)))
 812                                 goto next_normal;
 813                         if (r->id.idiag_sport != inet->inet_sport &&
 814                             r->id.idiag_sport)
 815                                 goto next_normal;
 816                         if (r->id.idiag_dport != inet->inet_dport &&
 817                             r->id.idiag_dport)
 818                                 goto next_normal;
 819                         if (inet_csk_diag_dump(sk, skb, cb) < 0) {
 820                                 spin_unlock_bh(lock);
 821                                 goto done;
 822                         }
 823 next_normal:
 824                         ++num;
 825                 }
 826
 827                 if (r->idiag_states & TCPF_TIME_WAIT) {
 828                         struct inet_timewait_sock *tw;
 829
 830                         inet_twsk_for_each(tw, node,
 831                                     &head->twchain) {
 832
 833                                 if (num < s_num)
 834                                         goto next_dying;
 835                                 if (r->id.idiag_sport != tw->tw_sport &&
 836                                     r->id.idiag_sport)
 837                                         goto next_dying;
 838                                 if (r->id.idiag_dport != tw->tw_dport &&
 839                                     r->id.idiag_dport)
 840                                         goto next_dying;
 841                                 if (inet_twsk_diag_dump(tw, skb, cb) < 0) {
 842                                         spin_unlock_bh(lock);
 843                                         goto done;
 844                                 }
 845 next_dying:
 846                                 ++num;
 847                         }
 848                 }
 849                 spin_unlock_bh(lock);
 850         }
 851
 852 done:
 853         cb->args[1] = i;
 854         cb->args[2] = num;
 855 unlock:
 856         inet_diag_unlock_handler(handler);
 857         return skb->len;
 858 }
 859
 860 static int inet_diag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 861 {
 862         int hdrlen = sizeof(struct inet_diag_req);
 863
 864         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
 865             nlmsg_len(nlh) < hdrlen)
 866                 return -EINVAL;
 867
 868         if (nlh->nlmsg_flags & NLM_F_DUMP) {
 869                 if (nlmsg_attrlen(nlh, hdrlen)) {
 870                         struct nlattr *attr;
 871
 872                         attr = nlmsg_find_attr(nlh, hdrlen,
 873                                                INET_DIAG_REQ_BYTECODE);
 874                         if (attr == NULL ||
 875                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
 876                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
 877                                 return -EINVAL;
 878                 }
 879
 880                 return netlink_dump_start(idiagnl, skb, nlh,
 881                                           inet_diag_dump, NULL, 0);
 882         }
 883
 884         return inet_diag_get_exact(skb, nlh);
 885 }
 886
 887 static DEFINE_MUTEX(inet_diag_mutex);
 888
 889 static void inet_diag_rcv(struct sk_buff *skb)
 890 {
 891         mutex_lock(&inet_diag_mutex);
 892         netlink_rcv_skb(skb, &inet_diag_rcv_msg);
 893         mutex_unlock(&inet_diag_mutex);
 894 }
 895
 896 int inet_diag_register(const struct inet_diag_handler *h)
 897 {
 898         const __u16 type = h->idiag_type;
 899         int err = -EINVAL;
 900
 901         if (type >= INET_DIAG_GETSOCK_MAX)
 902                 goto out;
 903
 904         mutex_lock(&inet_diag_table_mutex);
 905         err = -EEXIST;
 906         if (inet_diag_table[type] == NULL) {
 907                 inet_diag_table[type] = h;
 908                 err = 0;
 909         }
 910         mutex_unlock(&inet_diag_table_mutex);
 911 out:
 912         return err;
 913 }
 914 EXPORT_SYMBOL_GPL(inet_diag_register);
 915
 916 void inet_diag_unregister(const struct inet_diag_handler *h)
 917 {
 918         const __u16 type = h->idiag_type;
 919
 920         if (type >= INET_DIAG_GETSOCK_MAX)
 921                 return;
 922
 923         mutex_lock(&inet_diag_table_mutex);
 924         inet_diag_table[type] = NULL;
 925         mutex_unlock(&inet_diag_table_mutex);
 926 }
 927 EXPORT_SYMBOL_GPL(inet_diag_unregister);
 928
 929 static int __init inet_diag_init(void)
 930 {
 931         const int inet_diag_table_size = (INET_DIAG_GETSOCK_MAX *
 932                                           sizeof(struct inet_diag_handler *));
 933         int err = -ENOMEM;
 934
 935         inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
 936         if (!inet_diag_table)
 937                 goto out;
 938
 939         idiagnl = netlink_kernel_create(&init_net, NETLINK_INET_DIAG, 0,
 940                                         inet_diag_rcv, NULL, THIS_MODULE);
 941         if (idiagnl == NULL)
 942                 goto out_free_table;
 943         err = 0;
 944 out:
 945         return err;
 946 out_free_table:
 947         kfree(inet_diag_table);
 948         goto out;
 949 }
 950
 951 static void __exit inet_diag_exit(void)
 952 {
 953         netlink_kernel_release(idiagnl);
 954         kfree(inet_diag_table);
 955 }
 956
 957 module_init(inet_diag_init);
 958 module_exit(inet_diag_exit);
 959 MODULE_LICENSE("GPL");
 960 MODULE_ALIAS_NET_PF_PROTO(PF_NETLINK, NETLINK_INET_DIAG);