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
 123         memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
 124         memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
 125
 126         r->id.idiag_src[0] = inet->inet_rcv_saddr;
 127         r->id.idiag_dst[0] = inet->inet_daddr;
 128
 129         /* IPv6 dual-stack sockets use inet->tos for IPv4 connections,
 130          * hence this needs to be included regardless of socket family.
 131          */
 132         if (ext & (1 << (INET_DIAG_TOS - 1)))
 133                 RTA_PUT_U8(skb, INET_DIAG_TOS, inet->tos);
 134
 135 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 136         if (r->idiag_family == AF_INET6) {
 137                 const struct ipv6_pinfo *np = inet6_sk(sk);
 138
 139                 if (ext & (1 << (INET_DIAG_TCLASS - 1)))
 140                         RTA_PUT_U8(skb, INET_DIAG_TCLASS, np->tclass);
 141
 142                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 143                                &np->rcv_saddr);
 144                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 145                                &np->daddr);
 146         }
 147 #endif
 148
 149 #define EXPIRES_IN_MS(tmo)  DIV_ROUND_UP((tmo - jiffies) * 1000, HZ)
 150
 151         if (icsk->icsk_pending == ICSK_TIME_RETRANS) {
 152                 r->idiag_timer = 1;
 153                 r->idiag_retrans = icsk->icsk_retransmits;
 154                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 155         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 156                 r->idiag_timer = 4;
 157                 r->idiag_retrans = icsk->icsk_probes_out;
 158                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 159         } else if (timer_pending(&sk->sk_timer)) {
 160                 r->idiag_timer = 2;
 161                 r->idiag_retrans = icsk->icsk_probes_out;
 162                 r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 163         } else {
 164                 r->idiag_timer = 0;
 165                 r->idiag_expires = 0;
 166         }
 167 #undef EXPIRES_IN_MS
 168
 169         r->idiag_uid = sock_i_uid(sk);
 170         r->idiag_inode = sock_i_ino(sk);
 171
 172         if (minfo) {
 173                 minfo->idiag_rmem = sk_rmem_alloc_get(sk);
 174                 minfo->idiag_wmem = sk->sk_wmem_queued;
 175                 minfo->idiag_fmem = sk->sk_forward_alloc;
 176                 minfo->idiag_tmem = sk_wmem_alloc_get(sk);
 177         }
 178
 179         handler->idiag_get_info(sk, r, info);
 180
 181         if (sk->sk_state < TCP_TIME_WAIT &&
 182             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 183                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 184
 185         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 186         return skb->len;
 187
 188 rtattr_failure:
 189 nlmsg_failure:
 190         nlmsg_trim(skb, b);
 191         return -EMSGSIZE;
 192 }
 193
 194 static int inet_twsk_diag_fill(struct inet_timewait_sock *tw,
 195                                struct sk_buff *skb, int ext, u32 pid,
 196                                u32 seq, u16 nlmsg_flags,
 197                                const struct nlmsghdr *unlh)
 198 {
 199         long tmo;
 200         struct inet_diag_msg *r;
 201         const unsigned char *previous_tail = skb_tail_pointer(skb);
 202         struct nlmsghdr *nlh = NLMSG_PUT(skb, pid, seq,
 203                                          unlh->nlmsg_type, sizeof(*r));
 204
 205         r = NLMSG_DATA(nlh);
 206         BUG_ON(tw->tw_state != TCP_TIME_WAIT);
 207
 208         nlh->nlmsg_flags = nlmsg_flags;
 209
 210         tmo = tw->tw_ttd - jiffies;
 211         if (tmo < 0)
 212                 tmo = 0;
 213
 214         r->idiag_family       = tw->tw_family;
 215         r->idiag_retrans      = 0;
 216
 217         r->id.idiag_if        = tw->tw_bound_dev_if;
 218         r->id.idiag_cookie[0] = (u32)(unsigned long)tw;
 219         r->id.idiag_cookie[1] = (u32)(((unsigned long)tw >> 31) >> 1);
 220
 221         r->id.idiag_sport     = tw->tw_sport;
 222         r->id.idiag_dport     = tw->tw_dport;
 223
 224         memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
 225         memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
 226
 227         r->id.idiag_src[0]    = tw->tw_rcv_saddr;
 228         r->id.idiag_dst[0]    = tw->tw_daddr;
 229
 230         r->idiag_state        = tw->tw_substate;
 231         r->idiag_timer        = 3;
 232         r->idiag_expires      = DIV_ROUND_UP(tmo * 1000, HZ);
 233         r->idiag_rqueue       = 0;
 234         r->idiag_wqueue       = 0;
 235         r->idiag_uid          = 0;
 236         r->idiag_inode        = 0;
 237 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 238         if (tw->tw_family == AF_INET6) {
 239                 const struct inet6_timewait_sock *tw6 =
 240                                                 inet6_twsk((struct sock *)tw);
 241
 242                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 243                                &tw6->tw_v6_rcv_saddr);
 244                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 245                                &tw6->tw_v6_daddr);
 246         }
 247 #endif
 248         nlh->nlmsg_len = skb_tail_pointer(skb) - previous_tail;
 249         return skb->len;
 250 nlmsg_failure:
 251         nlmsg_trim(skb, previous_tail);
 252         return -EMSGSIZE;
 253 }
 254
 255 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
 256                         int ext, u32 pid, u32 seq, u16 nlmsg_flags,
 257                         const struct nlmsghdr *unlh)
 258 {
 259         if (sk->sk_state == TCP_TIME_WAIT)
 260                 return inet_twsk_diag_fill((struct inet_timewait_sock *)sk,
 261                                            skb, ext, pid, seq, nlmsg_flags,
 262                                            unlh);
 263         return inet_csk_diag_fill(sk, skb, ext, pid, seq, nlmsg_flags, unlh);
 264 }
 265
 266 static int inet_diag_get_exact(struct sk_buff *in_skb,
 267                                const struct nlmsghdr *nlh)
 268 {
 269         int err;
 270         struct sock *sk;
 271         struct inet_diag_req *req = NLMSG_DATA(nlh);
 272         struct sk_buff *rep;
 273         struct inet_hashinfo *hashinfo;
 274         const struct inet_diag_handler *handler;
 275
 276         handler = inet_diag_lock_handler(nlh->nlmsg_type);
 277         if (IS_ERR(handler)) {
 278                 err = PTR_ERR(handler);
 279                 goto unlock;
 280         }
 281
 282         hashinfo = handler->idiag_hashinfo;
 283         err = -EINVAL;
 284
 285         if (req->idiag_family == AF_INET) {
 286                 sk = inet_lookup(&init_net, hashinfo, req->id.idiag_dst[0],
 287                                  req->id.idiag_dport, req->id.idiag_src[0],
 288                                  req->id.idiag_sport, req->id.idiag_if);
 289         }
 290 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 291         else if (req->idiag_family == AF_INET6) {
 292                 sk = inet6_lookup(&init_net, hashinfo,
 293                                   (struct in6_addr *)req->id.idiag_dst,
 294                                   req->id.idiag_dport,
 295                                   (struct in6_addr *)req->id.idiag_src,
 296                                   req->id.idiag_sport,
 297                                   req->id.idiag_if);
 298         }
 299 #endif
 300         else {
 301                 goto unlock;
 302         }
 303
 304         err = -ENOENT;
 305         if (sk == NULL)
 306                 goto unlock;
 307
 308         err = -ESTALE;
 309         if ((req->id.idiag_cookie[0] != INET_DIAG_NOCOOKIE ||
 310              req->id.idiag_cookie[1] != INET_DIAG_NOCOOKIE) &&
 311             ((u32)(unsigned long)sk != req->id.idiag_cookie[0] ||
 312              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.idiag_cookie[1]))
 313                 goto out;
 314
 315         err = -ENOMEM;
 316         rep = alloc_skb(NLMSG_SPACE((sizeof(struct inet_diag_msg) +
 317                                      sizeof(struct inet_diag_meminfo) +
 318                                      handler->idiag_info_size + 64)),
 319                         GFP_KERNEL);
 320         if (!rep)
 321                 goto out;
 322
 323         err = sk_diag_fill(sk, rep, req->idiag_ext,
 324                            NETLINK_CB(in_skb).pid,
 325                            nlh->nlmsg_seq, 0, nlh);
 326         if (err < 0) {
 327                 WARN_ON(err == -EMSGSIZE);
 328                 kfree_skb(rep);
 329                 goto out;
 330         }
 331         err = netlink_unicast(idiagnl, rep, NETLINK_CB(in_skb).pid,
 332                               MSG_DONTWAIT);
 333         if (err > 0)
 334                 err = 0;
 335
 336 out:
 337         if (sk) {
 338                 if (sk->sk_state == TCP_TIME_WAIT)
 339                         inet_twsk_put((struct inet_timewait_sock *)sk);
 340                 else
 341                         sock_put(sk);
 342         }
 343 unlock:
 344         inet_diag_unlock_handler(handler);
 345         return err;
 346 }
 347
 348 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
 349 {
 350         int words = bits >> 5;
 351
 352         bits &= 0x1f;
 353
 354         if (words) {
 355                 if (memcmp(a1, a2, words << 2))
 356                         return 0;
 357         }
 358         if (bits) {
 359                 __be32 w1, w2;
 360                 __be32 mask;
 361
 362                 w1 = a1[words];
 363                 w2 = a2[words];
 364
 365                 mask = htonl((0xffffffff) << (32 - bits));
 366
 367                 if ((w1 ^ w2) & mask)
 368                         return 0;
 369         }
 370
 371         return 1;
 372 }
 373
 374
 375 static int inet_diag_bc_run(const void *bc, int len,
 376                             const struct inet_diag_entry *entry)
 377 {
 378         while (len > 0) {
 379                 int yes = 1;
 380                 const struct inet_diag_bc_op *op = bc;
 381
 382                 switch (op->code) {
 383                 case INET_DIAG_BC_NOP:
 384                         break;
 385                 case INET_DIAG_BC_JMP:
 386                         yes = 0;
 387                         break;
 388                 case INET_DIAG_BC_S_GE:
 389                         yes = entry->sport >= op[1].no;
 390                         break;
 391                 case INET_DIAG_BC_S_LE:
 392                         yes = entry->sport <= op[1].no;
 393                         break;
 394                 case INET_DIAG_BC_D_GE:
 395                         yes = entry->dport >= op[1].no;
 396                         break;
 397                 case INET_DIAG_BC_D_LE:
 398                         yes = entry->dport <= op[1].no;
 399                         break;
 400                 case INET_DIAG_BC_AUTO:
 401                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 402                         break;
 403                 case INET_DIAG_BC_S_COND:
 404                 case INET_DIAG_BC_D_COND: {
 405                         struct inet_diag_hostcond *cond;
 406                         __be32 *addr;
 407
 408                         cond = (struct inet_diag_hostcond *)(op + 1);
 409                         if (cond->port != -1 &&
 410                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 411                                              entry->sport : entry->dport)) {
 412                                 yes = 0;
 413                                 break;
 414                         }
 415
 416                         if (cond->prefix_len == 0)
 417                                 break;
 418
 419                         if (op->code == INET_DIAG_BC_S_COND)
 420                                 addr = entry->saddr;
 421                         else
 422                                 addr = entry->daddr;
 423
 424                         if (bitstring_match(addr, cond->addr,
 425                                             cond->prefix_len))
 426                                 break;
 427                         if (entry->family == AF_INET6 &&
 428                             cond->family == AF_INET) {
 429                                 if (addr[0] == 0 && addr[1] == 0 &&
 430                                     addr[2] == htonl(0xffff) &&
 431                                     bitstring_match(addr + 3, cond->addr,
 432                                                     cond->prefix_len))
 433                                         break;
 434                         }
 435                         yes = 0;
 436                         break;
 437                 }
 438                 }
 439
 440                 if (yes) {
 441                         len -= op->yes;
 442                         bc += op->yes;
 443                 } else {
 444                         len -= op->no;
 445                         bc += op->no;
 446                 }
 447         }
 448         return len == 0;
 449 }
 450
 451 static int valid_cc(const void *bc, int len, int cc)
 452 {
 453         while (len >= 0) {
 454                 const struct inet_diag_bc_op *op = bc;
 455
 456                 if (cc > len)
 457                         return 0;
 458                 if (cc == len)
 459                         return 1;
 460                 if (op->yes < 4 || op->yes & 3)
 461                         return 0;
 462                 len -= op->yes;
 463                 bc  += op->yes;
 464         }
 465         return 0;
 466 }
 467
 468 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
 469 {
 470         const void *bc = bytecode;
 471         int  len = bytecode_len;
 472
 473         while (len > 0) {
 474                 const struct inet_diag_bc_op *op = bc;
 475
 476 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 477                 switch (op->code) {
 478                 case INET_DIAG_BC_AUTO:
 479                 case INET_DIAG_BC_S_COND:
 480                 case INET_DIAG_BC_D_COND:
 481                 case INET_DIAG_BC_S_GE:
 482                 case INET_DIAG_BC_S_LE:
 483                 case INET_DIAG_BC_D_GE:
 484                 case INET_DIAG_BC_D_LE:
 485                 case INET_DIAG_BC_JMP:
 486                         if (op->no < 4 || op->no > len + 4 || op->no & 3)
 487                                 return -EINVAL;
 488                         if (op->no < len &&
 489                             !valid_cc(bytecode, bytecode_len, len - op->no))
 490                                 return -EINVAL;
 491                         break;
 492                 case INET_DIAG_BC_NOP:
 493                         break;
 494                 default:
 495                         return -EINVAL;
 496                 }
 497                 if (op->yes < 4 || op->yes > len + 4 || op->yes & 3)
 498                         return -EINVAL;
 499                 bc  += op->yes;
 500                 len -= op->yes;
 501         }
 502         return len == 0 ? 0 : -EINVAL;
 503 }
 504
 505 static int inet_csk_diag_dump(struct sock *sk,
 506                               struct sk_buff *skb,
 507                               struct netlink_callback *cb)
 508 {
 509         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 510
 511         if (nlmsg_attrlen(cb->nlh, sizeof(*r))) {
 512                 struct inet_diag_entry entry;
 513                 const struct nlattr *bc = nlmsg_find_attr(cb->nlh,
 514                                                           sizeof(*r),
 515                                                           INET_DIAG_REQ_BYTECODE);
 516                 struct inet_sock *inet = inet_sk(sk);
 517
 518                 entry.family = sk->sk_family;
 519 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 520                 if (entry.family == AF_INET6) {
 521                         struct ipv6_pinfo *np = inet6_sk(sk);
 522
 523                         entry.saddr = np->rcv_saddr.s6_addr32;
 524                         entry.daddr = np->daddr.s6_addr32;
 525                 } else
 526 #endif
 527                 {
 528                         entry.saddr = &inet->inet_rcv_saddr;
 529                         entry.daddr = &inet->inet_daddr;
 530                 }
 531                 entry.sport = inet->inet_num;
 532                 entry.dport = ntohs(inet->inet_dport);
 533                 entry.userlocks = sk->sk_userlocks;
 534
 535                 if (!inet_diag_bc_run(nla_data(bc), nla_len(bc), &entry))
 536                         return 0;
 537         }
 538
 539         return inet_csk_diag_fill(sk, skb, r->idiag_ext,
 540                                   NETLINK_CB(cb->skb).pid,
 541                                   cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 542 }
 543
 544 static int inet_twsk_diag_dump(struct inet_timewait_sock *tw,
 545                                struct sk_buff *skb,
 546                                struct netlink_callback *cb)
 547 {
 548         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 549
 550         if (nlmsg_attrlen(cb->nlh, sizeof(*r))) {
 551                 struct inet_diag_entry entry;
 552                 const struct nlattr *bc = nlmsg_find_attr(cb->nlh,
 553                                                           sizeof(*r),
 554                                                           INET_DIAG_REQ_BYTECODE);
 555
 556                 entry.family = tw->tw_family;
 557 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 558                 if (tw->tw_family == AF_INET6) {
 559                         struct inet6_timewait_sock *tw6 =
 560                                                 inet6_twsk((struct sock *)tw);
 561                         entry.saddr = tw6->tw_v6_rcv_saddr.s6_addr32;
 562                         entry.daddr = tw6->tw_v6_daddr.s6_addr32;
 563                 } else
 564 #endif
 565                 {
 566                         entry.saddr = &tw->tw_rcv_saddr;
 567                         entry.daddr = &tw->tw_daddr;
 568                 }
 569                 entry.sport = tw->tw_num;
 570                 entry.dport = ntohs(tw->tw_dport);
 571                 entry.userlocks = 0;
 572
 573                 if (!inet_diag_bc_run(nla_data(bc), nla_len(bc), &entry))
 574                         return 0;
 575         }
 576
 577         return inet_twsk_diag_fill(tw, skb, r->idiag_ext,
 578                                    NETLINK_CB(cb->skb).pid,
 579                                    cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 580 }
 581
 582 static int inet_diag_fill_req(struct sk_buff *skb, struct sock *sk,
 583                               struct request_sock *req, u32 pid, u32 seq,
 584                               const struct nlmsghdr *unlh)
 585 {
 586         const struct inet_request_sock *ireq = inet_rsk(req);
 587         struct inet_sock *inet = inet_sk(sk);
 588         unsigned char *b = skb_tail_pointer(skb);
 589         struct inet_diag_msg *r;
 590         struct nlmsghdr *nlh;
 591         long tmo;
 592
 593         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
 594         nlh->nlmsg_flags = NLM_F_MULTI;
 595         r = NLMSG_DATA(nlh);
 596
 597         r->idiag_family = sk->sk_family;
 598         r->idiag_state = TCP_SYN_RECV;
 599         r->idiag_timer = 1;
 600         r->idiag_retrans = req->retrans;
 601
 602         r->id.idiag_if = sk->sk_bound_dev_if;
 603         r->id.idiag_cookie[0] = (u32)(unsigned long)req;
 604         r->id.idiag_cookie[1] = (u32)(((unsigned long)req >> 31) >> 1);
 605
 606         tmo = req->expires - jiffies;
 607         if (tmo < 0)
 608                 tmo = 0;
 609
 610         r->id.idiag_sport = inet->inet_sport;
 611         r->id.idiag_dport = ireq->rmt_port;
 612
 613         memset(&r->id.idiag_src, 0, sizeof(r->id.idiag_src));
 614         memset(&r->id.idiag_dst, 0, sizeof(r->id.idiag_dst));
 615
 616         r->id.idiag_src[0] = ireq->loc_addr;
 617         r->id.idiag_dst[0] = ireq->rmt_addr;
 618
 619         r->idiag_expires = jiffies_to_msecs(tmo);
 620         r->idiag_rqueue = 0;
 621         r->idiag_wqueue = 0;
 622         r->idiag_uid = sock_i_uid(sk);
 623         r->idiag_inode = 0;
 624 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 625         if (r->idiag_family == AF_INET6) {
 626                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 627                                &inet6_rsk(req)->loc_addr);
 628                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 629                                &inet6_rsk(req)->rmt_addr);
 630         }
 631 #endif
 632         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 633
 634         return skb->len;
 635
 636 nlmsg_failure:
 637         nlmsg_trim(skb, b);
 638         return -1;
 639 }
 640
 641 static int inet_diag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 642                                struct netlink_callback *cb)
 643 {
 644         struct inet_diag_entry entry;
 645         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 646         struct inet_connection_sock *icsk = inet_csk(sk);
 647         struct listen_sock *lopt;
 648         const struct nlattr *bc = NULL;
 649         struct inet_sock *inet = inet_sk(sk);
 650         int j, s_j;
 651         int reqnum, s_reqnum;
 652         int err = 0;
 653
 654         s_j = cb->args[3];
 655         s_reqnum = cb->args[4];
 656
 657         if (s_j > 0)
 658                 s_j--;
 659
 660         entry.family = sk->sk_family;
 661
 662         read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 663
 664         lopt = icsk->icsk_accept_queue.listen_opt;
 665         if (!lopt || !lopt->qlen)
 666                 goto out;
 667
 668         if (nlmsg_attrlen(cb->nlh, sizeof(*r))) {
 669                 bc = nlmsg_find_attr(cb->nlh, sizeof(*r),
 670                                      INET_DIAG_REQ_BYTECODE);
 671                 entry.sport = inet->inet_num;
 672                 entry.userlocks = sk->sk_userlocks;
 673         }
 674
 675         for (j = s_j; j < lopt->nr_table_entries; j++) {
 676                 struct request_sock *req, *head = lopt->syn_table[j];
 677
 678                 reqnum = 0;
 679                 for (req = head; req; reqnum++, req = req->dl_next) {
 680                         struct inet_request_sock *ireq = inet_rsk(req);
 681
 682                         if (reqnum < s_reqnum)
 683                                 continue;
 684                         if (r->id.idiag_dport != ireq->rmt_port &&
 685                             r->id.idiag_dport)
 686                                 continue;
 687
 688                         if (bc) {
 689                                 entry.saddr =
 690 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 691                                         (entry.family == AF_INET6) ?
 692                                         inet6_rsk(req)->loc_addr.s6_addr32 :
 693 #endif
 694                                         &ireq->loc_addr;
 695                                 entry.daddr =
 696 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 697                                         (entry.family == AF_INET6) ?
 698                                         inet6_rsk(req)->rmt_addr.s6_addr32 :
 699 #endif
 700                                         &ireq->rmt_addr;
 701                                 entry.dport = ntohs(ireq->rmt_port);
 702
 703                                 if (!inet_diag_bc_run(nla_data(bc),
 704                                                       nla_len(bc), &entry))
 705                                         continue;
 706                         }
 707
 708                         err = inet_diag_fill_req(skb, sk, req,
 709                                                NETLINK_CB(cb->skb).pid,
 710                                                cb->nlh->nlmsg_seq, cb->nlh);
 711                         if (err < 0) {
 712                                 cb->args[3] = j + 1;
 713                                 cb->args[4] = reqnum;
 714                                 goto out;
 715                         }
 716                 }
 717
 718                 s_reqnum = 0;
 719         }
 720
 721 out:
 722         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 723
 724         return err;
 725 }
 726
 727 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 728 {
 729         int i, num;
 730         int s_i, s_num;
 731         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 732         const struct inet_diag_handler *handler;
 733         struct inet_hashinfo *hashinfo;
 734
 735         handler = inet_diag_lock_handler(cb->nlh->nlmsg_type);
 736         if (IS_ERR(handler))
 737                 goto unlock;
 738
 739         hashinfo = handler->idiag_hashinfo;
 740
 741         s_i = cb->args[1];
 742         s_num = num = cb->args[2];
 743
 744         if (cb->args[0] == 0) {
 745                 if (!(r->idiag_states & (TCPF_LISTEN | TCPF_SYN_RECV)))
 746                         goto skip_listen_ht;
 747
 748                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 749                         struct sock *sk;
 750                         struct hlist_nulls_node *node;
 751                         struct inet_listen_hashbucket *ilb;
 752
 753                         num = 0;
 754                         ilb = &hashinfo->listening_hash[i];
 755                         spin_lock_bh(&ilb->lock);
 756                         sk_nulls_for_each(sk, node, &ilb->head) {
 757                                 struct inet_sock *inet = inet_sk(sk);
 758
 759                                 if (num < s_num) {
 760                                         num++;
 761                                         continue;
 762                                 }
 763
 764                                 if (r->id.idiag_sport != inet->inet_sport &&
 765                                     r->id.idiag_sport)
 766                                         goto next_listen;
 767
 768                                 if (!(r->idiag_states & TCPF_LISTEN) ||
 769                                     r->id.idiag_dport ||
 770                                     cb->args[3] > 0)
 771                                         goto syn_recv;
 772
 773                                 if (inet_csk_diag_dump(sk, skb, cb) < 0) {
 774                                         spin_unlock_bh(&ilb->lock);
 775                                         goto done;
 776                                 }
 777
 778 syn_recv:
 779                                 if (!(r->idiag_states & TCPF_SYN_RECV))
 780                                         goto next_listen;
 781
 782                                 if (inet_diag_dump_reqs(skb, sk, cb) < 0) {
 783                                         spin_unlock_bh(&ilb->lock);
 784                                         goto done;
 785                                 }
 786
 787 next_listen:
 788                                 cb->args[3] = 0;
 789                                 cb->args[4] = 0;
 790                                 ++num;
 791                         }
 792                         spin_unlock_bh(&ilb->lock);
 793
 794                         s_num = 0;
 795                         cb->args[3] = 0;
 796                         cb->args[4] = 0;
 797                 }
 798 skip_listen_ht:
 799                 cb->args[0] = 1;
 800                 s_i = num = s_num = 0;
 801         }
 802
 803         if (!(r->idiag_states & ~(TCPF_LISTEN | TCPF_SYN_RECV)))
 804                 goto unlock;
 805
 806         for (i = s_i; i <= hashinfo->ehash_mask; i++) {
 807                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 808                 spinlock_t *lock = inet_ehash_lockp(hashinfo, i);
 809                 struct sock *sk;
 810                 struct hlist_nulls_node *node;
 811
 812                 num = 0;
 813
 814                 if (hlist_nulls_empty(&head->chain) &&
 815                         hlist_nulls_empty(&head->twchain))
 816                         continue;
 817
 818                 if (i > s_i)
 819                         s_num = 0;
 820
 821                 spin_lock_bh(lock);
 822                 sk_nulls_for_each(sk, node, &head->chain) {
 823                         struct inet_sock *inet = inet_sk(sk);
 824
 825                         if (num < s_num)
 826                                 goto next_normal;
 827                         if (!(r->idiag_states & (1 << sk->sk_state)))
 828                                 goto next_normal;
 829                         if (r->id.idiag_sport != inet->inet_sport &&
 830                             r->id.idiag_sport)
 831                                 goto next_normal;
 832                         if (r->id.idiag_dport != inet->inet_dport &&
 833                             r->id.idiag_dport)
 834                                 goto next_normal;
 835                         if (inet_csk_diag_dump(sk, skb, cb) < 0) {
 836                                 spin_unlock_bh(lock);
 837                                 goto done;
 838                         }
 839 next_normal:
 840                         ++num;
 841                 }
 842
 843                 if (r->idiag_states & (TCPF_TIME_WAIT | TCPF_FIN_WAIT2)) {
 844                         struct inet_timewait_sock *tw;
 845
 846                         inet_twsk_for_each(tw, node,
 847                                     &head->twchain) {
 848
 849                                 if (num < s_num)
 850                                         goto next_dying;
 851                                 if (!(r->idiag_states & (1 << tw->tw_substate)))
 852                                         goto next_dying;
 853                                 if (r->id.idiag_sport != tw->tw_sport &&
 854                                     r->id.idiag_sport)
 855                                         goto next_dying;
 856                                 if (r->id.idiag_dport != tw->tw_dport &&
 857                                     r->id.idiag_dport)
 858                                         goto next_dying;
 859                                 if (inet_twsk_diag_dump(tw, skb, cb) < 0) {
 860                                         spin_unlock_bh(lock);
 861                                         goto done;
 862                                 }
 863 next_dying:
 864                                 ++num;
 865                         }
 866                 }
 867                 spin_unlock_bh(lock);
 868         }
 869
 870 done:
 871         cb->args[1] = i;
 872         cb->args[2] = num;
 873 unlock:
 874         inet_diag_unlock_handler(handler);
 875         return skb->len;
 876 }
 877
 878 static int inet_diag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 879 {
 880         int hdrlen = sizeof(struct inet_diag_req);
 881
 882         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
 883             nlmsg_len(nlh) < hdrlen)
 884                 return -EINVAL;
 885
 886         if (nlh->nlmsg_flags & NLM_F_DUMP) {
 887                 if (nlmsg_attrlen(nlh, hdrlen)) {
 888                         struct nlattr *attr;
 889
 890                         attr = nlmsg_find_attr(nlh, hdrlen,
 891                                                INET_DIAG_REQ_BYTECODE);
 892                         if (attr == NULL ||
 893                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
 894                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
 895                                 return -EINVAL;
 896                 }
 897
 898                 return netlink_dump_start(idiagnl, skb, nlh,
 899                                           inet_diag_dump, NULL, 0);
 900         }
 901
 902         return inet_diag_get_exact(skb, nlh);
 903 }
 904
 905 static DEFINE_MUTEX(inet_diag_mutex);
 906
 907 static void inet_diag_rcv(struct sk_buff *skb)
 908 {
 909         mutex_lock(&inet_diag_mutex);
 910         netlink_rcv_skb(skb, &inet_diag_rcv_msg);
 911         mutex_unlock(&inet_diag_mutex);
 912 }
 913
 914 int inet_diag_register(const struct inet_diag_handler *h)
 915 {
 916         const __u16 type = h->idiag_type;
 917         int err = -EINVAL;
 918
 919         if (type >= INET_DIAG_GETSOCK_MAX)
 920                 goto out;
 921
 922         mutex_lock(&inet_diag_table_mutex);
 923         err = -EEXIST;
 924         if (inet_diag_table[type] == NULL) {
 925                 inet_diag_table[type] = h;
 926                 err = 0;
 927         }
 928         mutex_unlock(&inet_diag_table_mutex);
 929 out:
 930         return err;
 931 }
 932 EXPORT_SYMBOL_GPL(inet_diag_register);
 933
 934 void inet_diag_unregister(const struct inet_diag_handler *h)
 935 {
 936         const __u16 type = h->idiag_type;
 937
 938         if (type >= INET_DIAG_GETSOCK_MAX)
 939                 return;
 940
 941         mutex_lock(&inet_diag_table_mutex);
 942         inet_diag_table[type] = NULL;
 943         mutex_unlock(&inet_diag_table_mutex);
 944 }
 945 EXPORT_SYMBOL_GPL(inet_diag_unregister);
 946
 947 static int __init inet_diag_init(void)
 948 {
 949         const int inet_diag_table_size = (INET_DIAG_GETSOCK_MAX *
 950                                           sizeof(struct inet_diag_handler *));
 951         int err = -ENOMEM;
 952
 953         inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
 954         if (!inet_diag_table)
 955                 goto out;
 956
 957         idiagnl = netlink_kernel_create(&init_net, NETLINK_INET_DIAG, 0,
 958                                         inet_diag_rcv, NULL, THIS_MODULE);
 959         if (idiagnl == NULL)
 960                 goto out_free_table;
 961         err = 0;
 962 out:
 963         return err;
 964 out_free_table:
 965         kfree(inet_diag_table);
 966         goto out;
 967 }
 968
 969 static void __exit inet_diag_exit(void)
 970 {
 971         netlink_kernel_release(idiagnl);
 972         kfree(inet_diag_table);
 973 }
 974
 975 module_init(inet_diag_init);
 976 module_exit(inet_diag_exit);
 977 MODULE_LICENSE("GPL");
 978 MODULE_ALIAS_NET_PF_PROTO(PF_NETLINK, NETLINK_INET_DIAG);