git.oblomov.eu Git - linux-2.6/blob - net/ipv4/inet_diag.c

   1 /*
   2  * inet_diag.c  Module for monitoring INET transport protocols sockets.
   3  *
   4  * Version:     $Id: inet_diag.c,v 1.3 2002/02/01 22:01:04 davem Exp $
   5  *
   6  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
   7  *
   8  *      This program is free software; you can redistribute it and/or
   9  *      modify it under the terms of the GNU General Public License
  10  *      as published by the Free Software Foundation; either version
  11  *      2 of the License, or (at your option) any later version.
  12  */
  13
  14 #include <linux/module.h>
  15 #include <linux/types.h>
  16 #include <linux/fcntl.h>
  17 #include <linux/random.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
  31 #include <linux/inet.h>
  32 #include <linux/stddef.h>
  33
  34 #include <linux/inet_diag.h>
  35
  36 static const struct inet_diag_handler **inet_diag_table;
  37
  38 struct inet_diag_entry {
  39         __be32 *saddr;
  40         __be32 *daddr;
  41         u16 sport;
  42         u16 dport;
  43         u16 family;
  44         u16 userlocks;
  45 };
  46
  47 static struct sock *idiagnl;
  48
  49 #define INET_DIAG_PUT(skb, attrtype, attrlen) \
  50         RTA_DATA(__RTA_PUT(skb, attrtype, attrlen))
  51
  52 static int inet_csk_diag_fill(struct sock *sk,
  53                               struct sk_buff *skb,
  54                               int ext, u32 pid, u32 seq, u16 nlmsg_flags,
  55                               const struct nlmsghdr *unlh)
  56 {
  57         const struct inet_sock *inet = inet_sk(sk);
  58         const struct inet_connection_sock *icsk = inet_csk(sk);
  59         struct inet_diag_msg *r;
  60         struct nlmsghdr  *nlh;
  61         void *info = NULL;
  62         struct inet_diag_meminfo  *minfo = NULL;
  63         unsigned char    *b = skb->tail;
  64         const struct inet_diag_handler *handler;
  65
  66         handler = inet_diag_table[unlh->nlmsg_type];
  67         BUG_ON(handler == NULL);
  68
  69         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
  70         nlh->nlmsg_flags = nlmsg_flags;
  71
  72         r = NLMSG_DATA(nlh);
  73         BUG_ON(sk->sk_state == TCP_TIME_WAIT);
  74
  75         if (ext & (1 << (INET_DIAG_MEMINFO - 1)))
  76                 minfo = INET_DIAG_PUT(skb, INET_DIAG_MEMINFO, sizeof(*minfo));
  77
  78         if (ext & (1 << (INET_DIAG_INFO - 1)))
  79                 info = INET_DIAG_PUT(skb, INET_DIAG_INFO,
  80                                      handler->idiag_info_size);
  81
  82         if ((ext & (1 << (INET_DIAG_CONG - 1))) && icsk->icsk_ca_ops) {
  83                 const size_t len = strlen(icsk->icsk_ca_ops->name);
  84
  85                 strcpy(INET_DIAG_PUT(skb, INET_DIAG_CONG, len + 1),
  86                        icsk->icsk_ca_ops->name);
  87         }
  88
  89         r->idiag_family = sk->sk_family;
  90         r->idiag_state = sk->sk_state;
  91         r->idiag_timer = 0;
  92         r->idiag_retrans = 0;
  93
  94         r->id.idiag_if = sk->sk_bound_dev_if;
  95         r->id.idiag_cookie[0] = (u32)(unsigned long)sk;
  96         r->id.idiag_cookie[1] = (u32)(((unsigned long)sk >> 31) >> 1);
  97
  98         r->id.idiag_sport = inet->sport;
  99         r->id.idiag_dport = inet->dport;
 100         r->id.idiag_src[0] = inet->rcv_saddr;
 101         r->id.idiag_dst[0] = inet->daddr;
 102
 103 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 104         if (r->idiag_family == AF_INET6) {
 105                 struct ipv6_pinfo *np = inet6_sk(sk);
 106
 107                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 108                                &np->rcv_saddr);
 109                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 110                                &np->daddr);
 111         }
 112 #endif
 113
 114 #define EXPIRES_IN_MS(tmo)  ((tmo - jiffies) * 1000 + HZ - 1) / HZ
 115
 116         if (icsk->icsk_pending == ICSK_TIME_RETRANS) {
 117                 r->idiag_timer = 1;
 118                 r->idiag_retrans = icsk->icsk_retransmits;
 119                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 120         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 121                 r->idiag_timer = 4;
 122                 r->idiag_retrans = icsk->icsk_probes_out;
 123                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 124         } else if (timer_pending(&sk->sk_timer)) {
 125                 r->idiag_timer = 2;
 126                 r->idiag_retrans = icsk->icsk_probes_out;
 127                 r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 128         } else {
 129                 r->idiag_timer = 0;
 130                 r->idiag_expires = 0;
 131         }
 132 #undef EXPIRES_IN_MS
 133
 134         r->idiag_uid = sock_i_uid(sk);
 135         r->idiag_inode = sock_i_ino(sk);
 136
 137         if (minfo) {
 138                 minfo->idiag_rmem = atomic_read(&sk->sk_rmem_alloc);
 139                 minfo->idiag_wmem = sk->sk_wmem_queued;
 140                 minfo->idiag_fmem = sk->sk_forward_alloc;
 141                 minfo->idiag_tmem = atomic_read(&sk->sk_wmem_alloc);
 142         }
 143
 144         handler->idiag_get_info(sk, r, info);
 145
 146         if (sk->sk_state < TCP_TIME_WAIT &&
 147             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 148                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 149
 150         nlh->nlmsg_len = skb->tail - b;
 151         return skb->len;
 152
 153 rtattr_failure:
 154 nlmsg_failure:
 155         skb_trim(skb, b - skb->data);
 156         return -EMSGSIZE;
 157 }
 158
 159 static int inet_twsk_diag_fill(struct inet_timewait_sock *tw,
 160                                struct sk_buff *skb, int ext, u32 pid,
 161                                u32 seq, u16 nlmsg_flags,
 162                                const struct nlmsghdr *unlh)
 163 {
 164         long tmo;
 165         struct inet_diag_msg *r;
 166         const unsigned char *previous_tail = skb->tail;
 167         struct nlmsghdr *nlh = NLMSG_PUT(skb, pid, seq,
 168                                          unlh->nlmsg_type, sizeof(*r));
 169
 170         r = NLMSG_DATA(nlh);
 171         BUG_ON(tw->tw_state != TCP_TIME_WAIT);
 172
 173         nlh->nlmsg_flags = nlmsg_flags;
 174
 175         tmo = tw->tw_ttd - jiffies;
 176         if (tmo < 0)
 177                 tmo = 0;
 178
 179         r->idiag_family       = tw->tw_family;
 180         r->idiag_state        = tw->tw_state;
 181         r->idiag_timer        = 0;
 182         r->idiag_retrans      = 0;
 183         r->id.idiag_if        = tw->tw_bound_dev_if;
 184         r->id.idiag_cookie[0] = (u32)(unsigned long)tw;
 185         r->id.idiag_cookie[1] = (u32)(((unsigned long)tw >> 31) >> 1);
 186         r->id.idiag_sport     = tw->tw_sport;
 187         r->id.idiag_dport     = tw->tw_dport;
 188         r->id.idiag_src[0]    = tw->tw_rcv_saddr;
 189         r->id.idiag_dst[0]    = tw->tw_daddr;
 190         r->idiag_state        = tw->tw_substate;
 191         r->idiag_timer        = 3;
 192         r->idiag_expires      = (tmo * 1000 + HZ - 1) / HZ;
 193         r->idiag_rqueue       = 0;
 194         r->idiag_wqueue       = 0;
 195         r->idiag_uid          = 0;
 196         r->idiag_inode        = 0;
 197 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 198         if (tw->tw_family == AF_INET6) {
 199                 const struct inet6_timewait_sock *tw6 =
 200                                                 inet6_twsk((struct sock *)tw);
 201
 202                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 203                                &tw6->tw_v6_rcv_saddr);
 204                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 205                                &tw6->tw_v6_daddr);
 206         }
 207 #endif
 208         nlh->nlmsg_len = skb->tail - previous_tail;
 209         return skb->len;
 210 nlmsg_failure:
 211         skb_trim(skb, previous_tail - skb->data);
 212         return -EMSGSIZE;
 213 }
 214
 215 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
 216                         int ext, u32 pid, u32 seq, u16 nlmsg_flags,
 217                         const struct nlmsghdr *unlh)
 218 {
 219         if (sk->sk_state == TCP_TIME_WAIT)
 220                 return inet_twsk_diag_fill((struct inet_timewait_sock *)sk,
 221                                            skb, ext, pid, seq, nlmsg_flags,
 222                                            unlh);
 223         return inet_csk_diag_fill(sk, skb, ext, pid, seq, nlmsg_flags, unlh);
 224 }
 225
 226 static int inet_diag_get_exact(struct sk_buff *in_skb,
 227                                const struct nlmsghdr *nlh)
 228 {
 229         int err;
 230         struct sock *sk;
 231         struct inet_diag_req *req = NLMSG_DATA(nlh);
 232         struct sk_buff *rep;
 233         struct inet_hashinfo *hashinfo;
 234         const struct inet_diag_handler *handler;
 235
 236         handler = inet_diag_table[nlh->nlmsg_type];
 237         BUG_ON(handler == NULL);
 238         hashinfo = handler->idiag_hashinfo;
 239
 240         if (req->idiag_family == AF_INET) {
 241                 sk = inet_lookup(hashinfo, req->id.idiag_dst[0],
 242                                  req->id.idiag_dport, req->id.idiag_src[0],
 243                                  req->id.idiag_sport, req->id.idiag_if);
 244         }
 245 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 246         else if (req->idiag_family == AF_INET6) {
 247                 sk = inet6_lookup(hashinfo,
 248                                   (struct in6_addr *)req->id.idiag_dst,
 249                                   req->id.idiag_dport,
 250                                   (struct in6_addr *)req->id.idiag_src,
 251                                   req->id.idiag_sport,
 252                                   req->id.idiag_if);
 253         }
 254 #endif
 255         else {
 256                 return -EINVAL;
 257         }
 258
 259         if (sk == NULL)
 260                 return -ENOENT;
 261
 262         err = -ESTALE;
 263         if ((req->id.idiag_cookie[0] != INET_DIAG_NOCOOKIE ||
 264              req->id.idiag_cookie[1] != INET_DIAG_NOCOOKIE) &&
 265             ((u32)(unsigned long)sk != req->id.idiag_cookie[0] ||
 266              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.idiag_cookie[1]))
 267                 goto out;
 268
 269         err = -ENOMEM;
 270         rep = alloc_skb(NLMSG_SPACE((sizeof(struct inet_diag_msg) +
 271                                      sizeof(struct inet_diag_meminfo) +
 272                                      handler->idiag_info_size + 64)),
 273                         GFP_KERNEL);
 274         if (!rep)
 275                 goto out;
 276
 277         err = sk_diag_fill(sk, rep, req->idiag_ext,
 278                            NETLINK_CB(in_skb).pid,
 279                            nlh->nlmsg_seq, 0, nlh);
 280         if (err < 0) {
 281                 WARN_ON(err == -EMSGSIZE);
 282                 kfree_skb(rep);
 283                 goto out;
 284         }
 285         err = netlink_unicast(idiagnl, rep, NETLINK_CB(in_skb).pid,
 286                               MSG_DONTWAIT);
 287         if (err > 0)
 288                 err = 0;
 289
 290 out:
 291         if (sk) {
 292                 if (sk->sk_state == TCP_TIME_WAIT)
 293                         inet_twsk_put((struct inet_timewait_sock *)sk);
 294                 else
 295                         sock_put(sk);
 296         }
 297         return err;
 298 }
 299
 300 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
 301 {
 302         int words = bits >> 5;
 303
 304         bits &= 0x1f;
 305
 306         if (words) {
 307                 if (memcmp(a1, a2, words << 2))
 308                         return 0;
 309         }
 310         if (bits) {
 311                 __be32 w1, w2;
 312                 __be32 mask;
 313
 314                 w1 = a1[words];
 315                 w2 = a2[words];
 316
 317                 mask = htonl((0xffffffff) << (32 - bits));
 318
 319                 if ((w1 ^ w2) & mask)
 320                         return 0;
 321         }
 322
 323         return 1;
 324 }
 325
 326
 327 static int inet_diag_bc_run(const void *bc, int len,
 328                             const struct inet_diag_entry *entry)
 329 {
 330         while (len > 0) {
 331                 int yes = 1;
 332                 const struct inet_diag_bc_op *op = bc;
 333
 334                 switch (op->code) {
 335                 case INET_DIAG_BC_NOP:
 336                         break;
 337                 case INET_DIAG_BC_JMP:
 338                         yes = 0;
 339                         break;
 340                 case INET_DIAG_BC_S_GE:
 341                         yes = entry->sport >= op[1].no;
 342                         break;
 343                 case INET_DIAG_BC_S_LE:
 344                         yes = entry->dport <= op[1].no;
 345                         break;
 346                 case INET_DIAG_BC_D_GE:
 347                         yes = entry->dport >= op[1].no;
 348                         break;
 349                 case INET_DIAG_BC_D_LE:
 350                         yes = entry->dport <= op[1].no;
 351                         break;
 352                 case INET_DIAG_BC_AUTO:
 353                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 354                         break;
 355                 case INET_DIAG_BC_S_COND:
 356                 case INET_DIAG_BC_D_COND: {
 357                         struct inet_diag_hostcond *cond;
 358                         __be32 *addr;
 359
 360                         cond = (struct inet_diag_hostcond *)(op + 1);
 361                         if (cond->port != -1 &&
 362                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 363                                              entry->sport : entry->dport)) {
 364                                 yes = 0;
 365                                 break;
 366                         }
 367
 368                         if (cond->prefix_len == 0)
 369                                 break;
 370
 371                         if (op->code == INET_DIAG_BC_S_COND)
 372                                 addr = entry->saddr;
 373                         else
 374                                 addr = entry->daddr;
 375
 376                         if (bitstring_match(addr, cond->addr,
 377                                             cond->prefix_len))
 378                                 break;
 379                         if (entry->family == AF_INET6 &&
 380                             cond->family == AF_INET) {
 381                                 if (addr[0] == 0 && addr[1] == 0 &&
 382                                     addr[2] == htonl(0xffff) &&
 383                                     bitstring_match(addr + 3, cond->addr,
 384                                                     cond->prefix_len))
 385                                         break;
 386                         }
 387                         yes = 0;
 388                         break;
 389                 }
 390                 }
 391
 392                 if (yes) {
 393                         len -= op->yes;
 394                         bc += op->yes;
 395                 } else {
 396                         len -= op->no;
 397                         bc += op->no;
 398                 }
 399         }
 400         return (len == 0);
 401 }
 402
 403 static int valid_cc(const void *bc, int len, int cc)
 404 {
 405         while (len >= 0) {
 406                 const struct inet_diag_bc_op *op = bc;
 407
 408                 if (cc > len)
 409                         return 0;
 410                 if (cc == len)
 411                         return 1;
 412                 if (op->yes < 4)
 413                         return 0;
 414                 len -= op->yes;
 415                 bc  += op->yes;
 416         }
 417         return 0;
 418 }
 419
 420 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
 421 {
 422         const unsigned char *bc = bytecode;
 423         int  len = bytecode_len;
 424
 425         while (len > 0) {
 426                 struct inet_diag_bc_op *op = (struct inet_diag_bc_op *)bc;
 427
 428 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 429                 switch (op->code) {
 430                 case INET_DIAG_BC_AUTO:
 431                 case INET_DIAG_BC_S_COND:
 432                 case INET_DIAG_BC_D_COND:
 433                 case INET_DIAG_BC_S_GE:
 434                 case INET_DIAG_BC_S_LE:
 435                 case INET_DIAG_BC_D_GE:
 436                 case INET_DIAG_BC_D_LE:
 437                         if (op->yes < 4 || op->yes > len + 4)
 438                                 return -EINVAL;
 439                 case INET_DIAG_BC_JMP:
 440                         if (op->no < 4 || op->no > len + 4)
 441                                 return -EINVAL;
 442                         if (op->no < len &&
 443                             !valid_cc(bytecode, bytecode_len, len - op->no))
 444                                 return -EINVAL;
 445                         break;
 446                 case INET_DIAG_BC_NOP:
 447                         if (op->yes < 4 || op->yes > len + 4)
 448                                 return -EINVAL;
 449                         break;
 450                 default:
 451                         return -EINVAL;
 452                 }
 453                 bc  += op->yes;
 454                 len -= op->yes;
 455         }
 456         return len == 0 ? 0 : -EINVAL;
 457 }
 458
 459 static int inet_csk_diag_dump(struct sock *sk,
 460                               struct sk_buff *skb,
 461                               struct netlink_callback *cb)
 462 {
 463         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 464
 465         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 466                 struct inet_diag_entry entry;
 467                 struct rtattr *bc = (struct rtattr *)(r + 1);
 468                 struct inet_sock *inet = inet_sk(sk);
 469
 470                 entry.family = sk->sk_family;
 471 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 472                 if (entry.family == AF_INET6) {
 473                         struct ipv6_pinfo *np = inet6_sk(sk);
 474
 475                         entry.saddr = np->rcv_saddr.s6_addr32;
 476                         entry.daddr = np->daddr.s6_addr32;
 477                 } else
 478 #endif
 479                 {
 480                         entry.saddr = &inet->rcv_saddr;
 481                         entry.daddr = &inet->daddr;
 482                 }
 483                 entry.sport = inet->num;
 484                 entry.dport = ntohs(inet->dport);
 485                 entry.userlocks = sk->sk_userlocks;
 486
 487                 if (!inet_diag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), &entry))
 488                         return 0;
 489         }
 490
 491         return inet_csk_diag_fill(sk, skb, r->idiag_ext,
 492                                   NETLINK_CB(cb->skb).pid,
 493                                   cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 494 }
 495
 496 static int inet_twsk_diag_dump(struct inet_timewait_sock *tw,
 497                                struct sk_buff *skb,
 498                                struct netlink_callback *cb)
 499 {
 500         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 501
 502         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 503                 struct inet_diag_entry entry;
 504                 struct rtattr *bc = (struct rtattr *)(r + 1);
 505
 506                 entry.family = tw->tw_family;
 507 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 508                 if (tw->tw_family == AF_INET6) {
 509                         struct inet6_timewait_sock *tw6 =
 510                                                 inet6_twsk((struct sock *)tw);
 511                         entry.saddr = tw6->tw_v6_rcv_saddr.s6_addr32;
 512                         entry.daddr = tw6->tw_v6_daddr.s6_addr32;
 513                 } else
 514 #endif
 515                 {
 516                         entry.saddr = &tw->tw_rcv_saddr;
 517                         entry.daddr = &tw->tw_daddr;
 518                 }
 519                 entry.sport = tw->tw_num;
 520                 entry.dport = ntohs(tw->tw_dport);
 521                 entry.userlocks = 0;
 522
 523                 if (!inet_diag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), &entry))
 524                         return 0;
 525         }
 526
 527         return inet_twsk_diag_fill(tw, skb, r->idiag_ext,
 528                                    NETLINK_CB(cb->skb).pid,
 529                                    cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 530 }
 531
 532 static int inet_diag_fill_req(struct sk_buff *skb, struct sock *sk,
 533                               struct request_sock *req, u32 pid, u32 seq,
 534                               const struct nlmsghdr *unlh)
 535 {
 536         const struct inet_request_sock *ireq = inet_rsk(req);
 537         struct inet_sock *inet = inet_sk(sk);
 538         unsigned char *b = skb->tail;
 539         struct inet_diag_msg *r;
 540         struct nlmsghdr *nlh;
 541         long tmo;
 542
 543         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
 544         nlh->nlmsg_flags = NLM_F_MULTI;
 545         r = NLMSG_DATA(nlh);
 546
 547         r->idiag_family = sk->sk_family;
 548         r->idiag_state = TCP_SYN_RECV;
 549         r->idiag_timer = 1;
 550         r->idiag_retrans = req->retrans;
 551
 552         r->id.idiag_if = sk->sk_bound_dev_if;
 553         r->id.idiag_cookie[0] = (u32)(unsigned long)req;
 554         r->id.idiag_cookie[1] = (u32)(((unsigned long)req >> 31) >> 1);
 555
 556         tmo = req->expires - jiffies;
 557         if (tmo < 0)
 558                 tmo = 0;
 559
 560         r->id.idiag_sport = inet->sport;
 561         r->id.idiag_dport = ireq->rmt_port;
 562         r->id.idiag_src[0] = ireq->loc_addr;
 563         r->id.idiag_dst[0] = ireq->rmt_addr;
 564         r->idiag_expires = jiffies_to_msecs(tmo);
 565         r->idiag_rqueue = 0;
 566         r->idiag_wqueue = 0;
 567         r->idiag_uid = sock_i_uid(sk);
 568         r->idiag_inode = 0;
 569 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 570         if (r->idiag_family == AF_INET6) {
 571                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 572                                &inet6_rsk(req)->loc_addr);
 573                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 574                                &inet6_rsk(req)->rmt_addr);
 575         }
 576 #endif
 577         nlh->nlmsg_len = skb->tail - b;
 578
 579         return skb->len;
 580
 581 nlmsg_failure:
 582         skb_trim(skb, b - skb->data);
 583         return -1;
 584 }
 585
 586 static int inet_diag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 587                                struct netlink_callback *cb)
 588 {
 589         struct inet_diag_entry entry;
 590         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 591         struct inet_connection_sock *icsk = inet_csk(sk);
 592         struct listen_sock *lopt;
 593         struct rtattr *bc = NULL;
 594         struct inet_sock *inet = inet_sk(sk);
 595         int j, s_j;
 596         int reqnum, s_reqnum;
 597         int err = 0;
 598
 599         s_j = cb->args[3];
 600         s_reqnum = cb->args[4];
 601
 602         if (s_j > 0)
 603                 s_j--;
 604
 605         entry.family = sk->sk_family;
 606
 607         read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 608
 609         lopt = icsk->icsk_accept_queue.listen_opt;
 610         if (!lopt || !lopt->qlen)
 611                 goto out;
 612
 613         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 614                 bc = (struct rtattr *)(r + 1);
 615                 entry.sport = inet->num;
 616                 entry.userlocks = sk->sk_userlocks;
 617         }
 618
 619         for (j = s_j; j < lopt->nr_table_entries; j++) {
 620                 struct request_sock *req, *head = lopt->syn_table[j];
 621
 622                 reqnum = 0;
 623                 for (req = head; req; reqnum++, req = req->dl_next) {
 624                         struct inet_request_sock *ireq = inet_rsk(req);
 625
 626                         if (reqnum < s_reqnum)
 627                                 continue;
 628                         if (r->id.idiag_dport != ireq->rmt_port &&
 629                             r->id.idiag_dport)
 630                                 continue;
 631
 632                         if (bc) {
 633                                 entry.saddr =
 634 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 635                                         (entry.family == AF_INET6) ?
 636                                         inet6_rsk(req)->loc_addr.s6_addr32 :
 637 #endif
 638                                         &ireq->loc_addr;
 639                                 entry.daddr =
 640 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 641                                         (entry.family == AF_INET6) ?
 642                                         inet6_rsk(req)->rmt_addr.s6_addr32 :
 643 #endif
 644                                         &ireq->rmt_addr;
 645                                 entry.dport = ntohs(ireq->rmt_port);
 646
 647                                 if (!inet_diag_bc_run(RTA_DATA(bc),
 648                                                     RTA_PAYLOAD(bc), &entry))
 649                                         continue;
 650                         }
 651
 652                         err = inet_diag_fill_req(skb, sk, req,
 653                                                NETLINK_CB(cb->skb).pid,
 654                                                cb->nlh->nlmsg_seq, cb->nlh);
 655                         if (err < 0) {
 656                                 cb->args[3] = j + 1;
 657                                 cb->args[4] = reqnum;
 658                                 goto out;
 659                         }
 660                 }
 661
 662                 s_reqnum = 0;
 663         }
 664
 665 out:
 666         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 667
 668         return err;
 669 }
 670
 671 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 672 {
 673         int i, num;
 674         int s_i, s_num;
 675         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 676         const struct inet_diag_handler *handler;
 677         struct inet_hashinfo *hashinfo;
 678
 679         handler = inet_diag_table[cb->nlh->nlmsg_type];
 680         BUG_ON(handler == NULL);
 681         hashinfo = handler->idiag_hashinfo;
 682
 683         s_i = cb->args[1];
 684         s_num = num = cb->args[2];
 685
 686         if (cb->args[0] == 0) {
 687                 if (!(r->idiag_states & (TCPF_LISTEN | TCPF_SYN_RECV)))
 688                         goto skip_listen_ht;
 689
 690                 inet_listen_lock(hashinfo);
 691                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 692                         struct sock *sk;
 693                         struct hlist_node *node;
 694
 695                         num = 0;
 696                         sk_for_each(sk, node, &hashinfo->listening_hash[i]) {
 697                                 struct inet_sock *inet = inet_sk(sk);
 698
 699                                 if (num < s_num) {
 700                                         num++;
 701                                         continue;
 702                                 }
 703
 704                                 if (r->id.idiag_sport != inet->sport &&
 705                                     r->id.idiag_sport)
 706                                         goto next_listen;
 707
 708                                 if (!(r->idiag_states & TCPF_LISTEN) ||
 709                                     r->id.idiag_dport ||
 710                                     cb->args[3] > 0)
 711                                         goto syn_recv;
 712
 713                                 if (inet_csk_diag_dump(sk, skb, cb) < 0) {
 714                                         inet_listen_unlock(hashinfo);
 715                                         goto done;
 716                                 }
 717
 718 syn_recv:
 719                                 if (!(r->idiag_states & TCPF_SYN_RECV))
 720                                         goto next_listen;
 721
 722                                 if (inet_diag_dump_reqs(skb, sk, cb) < 0) {
 723                                         inet_listen_unlock(hashinfo);
 724                                         goto done;
 725                                 }
 726
 727 next_listen:
 728                                 cb->args[3] = 0;
 729                                 cb->args[4] = 0;
 730                                 ++num;
 731                         }
 732
 733                         s_num = 0;
 734                         cb->args[3] = 0;
 735                         cb->args[4] = 0;
 736                 }
 737                 inet_listen_unlock(hashinfo);
 738 skip_listen_ht:
 739                 cb->args[0] = 1;
 740                 s_i = num = s_num = 0;
 741         }
 742
 743         if (!(r->idiag_states & ~(TCPF_LISTEN | TCPF_SYN_RECV)))
 744                 return skb->len;
 745
 746         for (i = s_i; i < hashinfo->ehash_size; i++) {
 747                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 748                 struct sock *sk;
 749                 struct hlist_node *node;
 750
 751                 if (i > s_i)
 752                         s_num = 0;
 753
 754                 read_lock_bh(&head->lock);
 755                 num = 0;
 756                 sk_for_each(sk, node, &head->chain) {
 757                         struct inet_sock *inet = inet_sk(sk);
 758
 759                         if (num < s_num)
 760                                 goto next_normal;
 761                         if (!(r->idiag_states & (1 << sk->sk_state)))
 762                                 goto next_normal;
 763                         if (r->id.idiag_sport != inet->sport &&
 764                             r->id.idiag_sport)
 765                                 goto next_normal;
 766                         if (r->id.idiag_dport != inet->dport &&
 767                             r->id.idiag_dport)
 768                                 goto next_normal;
 769                         if (inet_csk_diag_dump(sk, skb, cb) < 0) {
 770                                 read_unlock_bh(&head->lock);
 771                                 goto done;
 772                         }
 773 next_normal:
 774                         ++num;
 775                 }
 776
 777                 if (r->idiag_states & TCPF_TIME_WAIT) {
 778                         struct inet_timewait_sock *tw;
 779
 780                         inet_twsk_for_each(tw, node,
 781                                     &head->twchain) {
 782
 783                                 if (num < s_num)
 784                                         goto next_dying;
 785                                 if (r->id.idiag_sport != tw->tw_sport &&
 786                                     r->id.idiag_sport)
 787                                         goto next_dying;
 788                                 if (r->id.idiag_dport != tw->tw_dport &&
 789                                     r->id.idiag_dport)
 790                                         goto next_dying;
 791                                 if (inet_twsk_diag_dump(tw, skb, cb) < 0) {
 792                                         read_unlock_bh(&head->lock);
 793                                         goto done;
 794                                 }
 795 next_dying:
 796                                 ++num;
 797                         }
 798                 }
 799                 read_unlock_bh(&head->lock);
 800         }
 801
 802 done:
 803         cb->args[1] = i;
 804         cb->args[2] = num;
 805         return skb->len;
 806 }
 807
 808 static inline int inet_diag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 809 {
 810         if (!(nlh->nlmsg_flags&NLM_F_REQUEST))
 811                 return 0;
 812
 813         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX)
 814                 goto err_inval;
 815
 816         if (inet_diag_table[nlh->nlmsg_type] == NULL)
 817                 return -ENOENT;
 818
 819         if (NLMSG_LENGTH(sizeof(struct inet_diag_req)) > skb->len)
 820                 goto err_inval;
 821
 822         if (nlh->nlmsg_flags&NLM_F_DUMP) {
 823                 if (nlh->nlmsg_len >
 824                     (4 + NLMSG_SPACE(sizeof(struct inet_diag_req)))) {
 825                         struct rtattr *rta = (void *)(NLMSG_DATA(nlh) +
 826                                                  sizeof(struct inet_diag_req));
 827                         if (rta->rta_type != INET_DIAG_REQ_BYTECODE ||
 828                             rta->rta_len < 8 ||
 829                             rta->rta_len >
 830                             (nlh->nlmsg_len -
 831                              NLMSG_SPACE(sizeof(struct inet_diag_req))))
 832                                 goto err_inval;
 833                         if (inet_diag_bc_audit(RTA_DATA(rta), RTA_PAYLOAD(rta)))
 834                                 goto err_inval;
 835                 }
 836                 return netlink_dump_start(idiagnl, skb, nlh,
 837                                           inet_diag_dump, NULL);
 838         } else
 839                 return inet_diag_get_exact(skb, nlh);
 840
 841 err_inval:
 842         return -EINVAL;
 843 }
 844
 845
 846 static inline void inet_diag_rcv_skb(struct sk_buff *skb)
 847 {
 848         if (skb->len >= NLMSG_SPACE(0)) {
 849                 int err;
 850                 struct nlmsghdr *nlh = (struct nlmsghdr *)skb->data;
 851
 852                 if (nlh->nlmsg_len < sizeof(*nlh) ||
 853                     skb->len < nlh->nlmsg_len)
 854                         return;
 855                 err = inet_diag_rcv_msg(skb, nlh);
 856                 if (err || nlh->nlmsg_flags & NLM_F_ACK)
 857                         netlink_ack(skb, nlh, err);
 858         }
 859 }
 860
 861 static void inet_diag_rcv(struct sock *sk, int len)
 862 {
 863         struct sk_buff *skb;
 864         unsigned int qlen = skb_queue_len(&sk->sk_receive_queue);
 865
 866         while (qlen-- && (skb = skb_dequeue(&sk->sk_receive_queue))) {
 867                 inet_diag_rcv_skb(skb);
 868                 kfree_skb(skb);
 869         }
 870 }
 871
 872 static DEFINE_SPINLOCK(inet_diag_register_lock);
 873
 874 int inet_diag_register(const struct inet_diag_handler *h)
 875 {
 876         const __u16 type = h->idiag_type;
 877         int err = -EINVAL;
 878
 879         if (type >= INET_DIAG_GETSOCK_MAX)
 880                 goto out;
 881
 882         spin_lock(&inet_diag_register_lock);
 883         err = -EEXIST;
 884         if (inet_diag_table[type] == NULL) {
 885                 inet_diag_table[type] = h;
 886                 err = 0;
 887         }
 888         spin_unlock(&inet_diag_register_lock);
 889 out:
 890         return err;
 891 }
 892 EXPORT_SYMBOL_GPL(inet_diag_register);
 893
 894 void inet_diag_unregister(const struct inet_diag_handler *h)
 895 {
 896         const __u16 type = h->idiag_type;
 897
 898         if (type >= INET_DIAG_GETSOCK_MAX)
 899                 return;
 900
 901         spin_lock(&inet_diag_register_lock);
 902         inet_diag_table[type] = NULL;
 903         spin_unlock(&inet_diag_register_lock);
 904
 905         synchronize_rcu();
 906 }
 907 EXPORT_SYMBOL_GPL(inet_diag_unregister);
 908
 909 static int __init inet_diag_init(void)
 910 {
 911         const int inet_diag_table_size = (INET_DIAG_GETSOCK_MAX *
 912                                           sizeof(struct inet_diag_handler *));
 913         int err = -ENOMEM;
 914
 915         inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
 916         if (!inet_diag_table)
 917                 goto out;
 918
 919         idiagnl = netlink_kernel_create(NETLINK_INET_DIAG, 0, inet_diag_rcv,
 920                                         THIS_MODULE);
 921         if (idiagnl == NULL)
 922                 goto out_free_table;
 923         err = 0;
 924 out:
 925         return err;
 926 out_free_table:
 927         kfree(inet_diag_table);
 928         goto out;
 929 }
 930
 931 static void __exit inet_diag_exit(void)
 932 {
 933         sock_release(idiagnl->sk_socket);
 934         kfree(inet_diag_table);
 935 }
 936
 937 module_init(inet_diag_init);
 938 module_exit(inet_diag_exit);
 939 MODULE_LICENSE("GPL");