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