net/ipv4/tcp_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/config.h>
  15 #include <linux/module.h>
  16 #include <linux/types.h>
  17 #include <linux/fcntl.h>
  18 #include <linux/random.h>
  19 #include <linux/cache.h>
  20 #include <linux/init.h>
  21 #include <linux/time.h>
  22
  23 #include <net/icmp.h>
  24 #include <net/tcp.h>
  25 #include <net/ipv6.h>
  26 #include <net/inet_common.h>
  27 #include <net/inet_connection_sock.h>
  28 #include <net/inet_hashtables.h>
  29 #include <net/inet_timewait_sock.h>
  30 #include <net/inet6_hashtables.h>
  31
  32 #include <linux/inet.h>
  33 #include <linux/stddef.h>
  34
  35 #include <linux/tcp_diag.h>
  36
  37 static const struct inet_diag_handler **inet_diag_table;
  38
  39 struct inet_diag_entry {
  40         u32 *saddr;
  41         u32 *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 int inet_diag_fill(struct sk_buff *skb, struct sock *sk,
  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         if (sk->sk_state != TCP_TIME_WAIT) {
  74                 if (ext & (1 << (INET_DIAG_MEMINFO - 1)))
  75                         minfo = INET_DIAG_PUT(skb, INET_DIAG_MEMINFO,
  76                                               sizeof(*minfo));
  77                 if (ext & (1 << (INET_DIAG_INFO - 1)))
  78                         info = INET_DIAG_PUT(skb, INET_DIAG_INFO,
  79                                            handler->idiag_info_size);
  80
  81                 if ((ext & (1 << (INET_DIAG_CONG - 1))) && icsk->icsk_ca_ops) {
  82                         size_t len = strlen(icsk->icsk_ca_ops->name);
  83                         strcpy(INET_DIAG_PUT(skb, INET_DIAG_CONG, len + 1),
  84                                icsk->icsk_ca_ops->name);
  85                 }
  86         }
  87         r->idiag_family = sk->sk_family;
  88         r->idiag_state = sk->sk_state;
  89         r->idiag_timer = 0;
  90         r->idiag_retrans = 0;
  91
  92         r->id.idiag_if = sk->sk_bound_dev_if;
  93         r->id.idiag_cookie[0] = (u32)(unsigned long)sk;
  94         r->id.idiag_cookie[1] = (u32)(((unsigned long)sk >> 31) >> 1);
  95
  96         if (r->idiag_state == TCP_TIME_WAIT) {
  97                 const struct inet_timewait_sock *tw = inet_twsk(sk);
  98                 long tmo = tw->tw_ttd - jiffies;
  99                 if (tmo < 0)
 100                         tmo = 0;
 101
 102                 r->id.idiag_sport = tw->tw_sport;
 103                 r->id.idiag_dport = tw->tw_dport;
 104                 r->id.idiag_src[0] = tw->tw_rcv_saddr;
 105                 r->id.idiag_dst[0] = tw->tw_daddr;
 106                 r->idiag_state = tw->tw_substate;
 107                 r->idiag_timer = 3;
 108                 r->idiag_expires = (tmo * 1000 + HZ - 1) / HZ;
 109                 r->idiag_rqueue = 0;
 110                 r->idiag_wqueue = 0;
 111                 r->idiag_uid = 0;
 112                 r->idiag_inode = 0;
 113 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 114                 if (r->idiag_family == AF_INET6) {
 115                         const struct tcp6_timewait_sock *tcp6tw = tcp6_twsk(sk);
 116
 117                         ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 118                                        &tcp6tw->tw_v6_rcv_saddr);
 119                         ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 120                                        &tcp6tw->tw_v6_daddr);
 121                 }
 122 #endif
 123                 nlh->nlmsg_len = skb->tail - b;
 124                 return skb->len;
 125         }
 126
 127         r->id.idiag_sport = inet->sport;
 128         r->id.idiag_dport = inet->dport;
 129         r->id.idiag_src[0] = inet->rcv_saddr;
 130         r->id.idiag_dst[0] = inet->daddr;
 131
 132 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 133         if (r->idiag_family == AF_INET6) {
 134                 struct ipv6_pinfo *np = inet6_sk(sk);
 135
 136                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 137                                &np->rcv_saddr);
 138                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 139                                &np->daddr);
 140         }
 141 #endif
 142
 143 #define EXPIRES_IN_MS(tmo)  ((tmo-jiffies)*1000+HZ-1)/HZ
 144
 145         if (icsk->icsk_pending == ICSK_TIME_RETRANS) {
 146                 r->idiag_timer = 1;
 147                 r->idiag_retrans = icsk->icsk_retransmits;
 148                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 149         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 150                 r->idiag_timer = 4;
 151                 r->idiag_retrans = icsk->icsk_probes_out;
 152                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 153         } else if (timer_pending(&sk->sk_timer)) {
 154                 r->idiag_timer = 2;
 155                 r->idiag_retrans = icsk->icsk_probes_out;
 156                 r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 157         } else {
 158                 r->idiag_timer = 0;
 159                 r->idiag_expires = 0;
 160         }
 161 #undef EXPIRES_IN_MS
 162
 163         r->idiag_uid = sock_i_uid(sk);
 164         r->idiag_inode = sock_i_ino(sk);
 165
 166         if (minfo) {
 167                 minfo->idiag_rmem = atomic_read(&sk->sk_rmem_alloc);
 168                 minfo->idiag_wmem = sk->sk_wmem_queued;
 169                 minfo->idiag_fmem = sk->sk_forward_alloc;
 170                 minfo->idiag_tmem = atomic_read(&sk->sk_wmem_alloc);
 171         }
 172
 173         handler->idiag_get_info(sk, r, info);
 174
 175         if (sk->sk_state < TCP_TIME_WAIT &&
 176             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 177                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 178
 179         nlh->nlmsg_len = skb->tail - b;
 180         return skb->len;
 181
 182 rtattr_failure:
 183 nlmsg_failure:
 184         skb_trim(skb, b - skb->data);
 185         return -1;
 186 }
 187
 188 static int inet_diag_get_exact(struct sk_buff *in_skb, const struct nlmsghdr *nlh)
 189 {
 190         int err;
 191         struct sock *sk;
 192         struct inet_diag_req *req = NLMSG_DATA(nlh);
 193         struct sk_buff *rep;
 194         struct inet_hashinfo *hashinfo;
 195         const struct inet_diag_handler *handler;
 196
 197         handler = inet_diag_table[nlh->nlmsg_type];
 198         BUG_ON(handler == NULL);
 199         hashinfo = handler->idiag_hashinfo;
 200
 201         if (req->idiag_family == AF_INET) {
 202                 sk = inet_lookup(hashinfo, req->id.idiag_dst[0],
 203                                  req->id.idiag_dport, req->id.idiag_src[0],
 204                                  req->id.idiag_sport, req->id.idiag_if);
 205         }
 206 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 207         else if (req->idiag_family == AF_INET6) {
 208                 sk = inet6_lookup(hashinfo,
 209                                   (struct in6_addr *)req->id.idiag_dst,
 210                                   req->id.idiag_dport,
 211                                   (struct in6_addr *)req->id.idiag_src,
 212                                   req->id.idiag_sport,
 213                                   req->id.idiag_if);
 214         }
 215 #endif
 216         else {
 217                 return -EINVAL;
 218         }
 219
 220         if (sk == NULL)
 221                 return -ENOENT;
 222
 223         err = -ESTALE;
 224         if ((req->id.idiag_cookie[0] != INET_DIAG_NOCOOKIE ||
 225              req->id.idiag_cookie[1] != INET_DIAG_NOCOOKIE) &&
 226             ((u32)(unsigned long)sk != req->id.idiag_cookie[0] ||
 227              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.idiag_cookie[1]))
 228                 goto out;
 229
 230         err = -ENOMEM;
 231         rep = alloc_skb(NLMSG_SPACE((sizeof(struct inet_diag_msg) +
 232                                      sizeof(struct inet_diag_meminfo) +
 233                                      handler->idiag_info_size + 64)),
 234                         GFP_KERNEL);
 235         if (!rep)
 236                 goto out;
 237
 238         if (inet_diag_fill(rep, sk, req->idiag_ext,
 239                          NETLINK_CB(in_skb).pid,
 240                          nlh->nlmsg_seq, 0, nlh) <= 0)
 241                 BUG();
 242
 243         err = netlink_unicast(idiagnl, rep, NETLINK_CB(in_skb).pid,
 244                               MSG_DONTWAIT);
 245         if (err > 0)
 246                 err = 0;
 247
 248 out:
 249         if (sk) {
 250                 if (sk->sk_state == TCP_TIME_WAIT)
 251                         inet_twsk_put((struct inet_timewait_sock *)sk);
 252                 else
 253                         sock_put(sk);
 254         }
 255         return err;
 256 }
 257
 258 static int bitstring_match(const u32 *a1, const u32 *a2, int bits)
 259 {
 260         int words = bits >> 5;
 261
 262         bits &= 0x1f;
 263
 264         if (words) {
 265                 if (memcmp(a1, a2, words << 2))
 266                         return 0;
 267         }
 268         if (bits) {
 269                 __u32 w1, w2;
 270                 __u32 mask;
 271
 272                 w1 = a1[words];
 273                 w2 = a2[words];
 274
 275                 mask = htonl((0xffffffff) << (32 - bits));
 276
 277                 if ((w1 ^ w2) & mask)
 278                         return 0;
 279         }
 280
 281         return 1;
 282 }
 283
 284
 285 static int inet_diag_bc_run(const void *bc, int len,
 286                           const struct inet_diag_entry *entry)
 287 {
 288         while (len > 0) {
 289                 int yes = 1;
 290                 const struct inet_diag_bc_op *op = bc;
 291
 292                 switch (op->code) {
 293                 case INET_DIAG_BC_NOP:
 294                         break;
 295                 case INET_DIAG_BC_JMP:
 296                         yes = 0;
 297                         break;
 298                 case INET_DIAG_BC_S_GE:
 299                         yes = entry->sport >= op[1].no;
 300                         break;
 301                 case INET_DIAG_BC_S_LE:
 302                         yes = entry->dport <= op[1].no;
 303                         break;
 304                 case INET_DIAG_BC_D_GE:
 305                         yes = entry->dport >= op[1].no;
 306                         break;
 307                 case INET_DIAG_BC_D_LE:
 308                         yes = entry->dport <= op[1].no;
 309                         break;
 310                 case INET_DIAG_BC_AUTO:
 311                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 312                         break;
 313                 case INET_DIAG_BC_S_COND:
 314                 case INET_DIAG_BC_D_COND:
 315                 {
 316                         struct inet_diag_hostcond *cond = (struct inet_diag_hostcond*)(op+1);
 317                         u32 *addr;
 318
 319                         if (cond->port != -1 &&
 320                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 321                                              entry->sport : entry->dport)) {
 322                                 yes = 0;
 323                                 break;
 324                         }
 325
 326                         if (cond->prefix_len == 0)
 327                                 break;
 328
 329                         if (op->code == INET_DIAG_BC_S_COND)
 330                                 addr = entry->saddr;
 331                         else
 332                                 addr = entry->daddr;
 333
 334                         if (bitstring_match(addr, cond->addr, cond->prefix_len))
 335                                 break;
 336                         if (entry->family == AF_INET6 &&
 337                             cond->family == AF_INET) {
 338                                 if (addr[0] == 0 && addr[1] == 0 &&
 339                                     addr[2] == htonl(0xffff) &&
 340                                     bitstring_match(addr+3, cond->addr, cond->prefix_len))
 341                                         break;
 342                         }
 343                         yes = 0;
 344                         break;
 345                 }
 346                 }
 347
 348                 if (yes) {
 349                         len -= op->yes;
 350                         bc += op->yes;
 351                 } else {
 352                         len -= op->no;
 353                         bc += op->no;
 354                 }
 355         }
 356         return (len == 0);
 357 }
 358
 359 static int valid_cc(const void *bc, int len, int cc)
 360 {
 361         while (len >= 0) {
 362                 const struct inet_diag_bc_op *op = bc;
 363
 364                 if (cc > len)
 365                         return 0;
 366                 if (cc == len)
 367                         return 1;
 368                 if (op->yes < 4)
 369                         return 0;
 370                 len -= op->yes;
 371                 bc  += op->yes;
 372         }
 373         return 0;
 374 }
 375
 376 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
 377 {
 378         const unsigned char *bc = bytecode;
 379         int  len = bytecode_len;
 380
 381         while (len > 0) {
 382                 struct inet_diag_bc_op *op = (struct inet_diag_bc_op*)bc;
 383
 384 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 385                 switch (op->code) {
 386                 case INET_DIAG_BC_AUTO:
 387                 case INET_DIAG_BC_S_COND:
 388                 case INET_DIAG_BC_D_COND:
 389                 case INET_DIAG_BC_S_GE:
 390                 case INET_DIAG_BC_S_LE:
 391                 case INET_DIAG_BC_D_GE:
 392                 case INET_DIAG_BC_D_LE:
 393                         if (op->yes < 4 || op->yes > len+4)
 394                                 return -EINVAL;
 395                 case INET_DIAG_BC_JMP:
 396                         if (op->no < 4 || op->no > len+4)
 397                                 return -EINVAL;
 398                         if (op->no < len &&
 399                             !valid_cc(bytecode, bytecode_len, len-op->no))
 400                                 return -EINVAL;
 401                         break;
 402                 case INET_DIAG_BC_NOP:
 403                         if (op->yes < 4 || op->yes > len+4)
 404                                 return -EINVAL;
 405                         break;
 406                 default:
 407                         return -EINVAL;
 408                 }
 409                 bc += op->yes;
 410                 len -= op->yes;
 411         }
 412         return len == 0 ? 0 : -EINVAL;
 413 }
 414
 415 static int inet_diag_dump_sock(struct sk_buff *skb, struct sock *sk,
 416                              struct netlink_callback *cb)
 417 {
 418         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 419
 420         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 421                 struct inet_diag_entry entry;
 422                 struct rtattr *bc = (struct rtattr *)(r + 1);
 423                 struct inet_sock *inet = inet_sk(sk);
 424
 425                 entry.family = sk->sk_family;
 426 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 427                 if (entry.family == AF_INET6) {
 428                         struct ipv6_pinfo *np = inet6_sk(sk);
 429
 430                         entry.saddr = np->rcv_saddr.s6_addr32;
 431                         entry.daddr = np->daddr.s6_addr32;
 432                 } else
 433 #endif
 434                 {
 435                         entry.saddr = &inet->rcv_saddr;
 436                         entry.daddr = &inet->daddr;
 437                 }
 438                 entry.sport = inet->num;
 439                 entry.dport = ntohs(inet->dport);
 440                 entry.userlocks = sk->sk_userlocks;
 441
 442                 if (!inet_diag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), &entry))
 443                         return 0;
 444         }
 445
 446         return inet_diag_fill(skb, sk, r->idiag_ext, NETLINK_CB(cb->skb).pid,
 447                             cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 448 }
 449
 450 static int inet_diag_fill_req(struct sk_buff *skb, struct sock *sk,
 451                             struct request_sock *req,
 452                             u32 pid, u32 seq,
 453                             const struct nlmsghdr *unlh)
 454 {
 455         const struct inet_request_sock *ireq = inet_rsk(req);
 456         struct inet_sock *inet = inet_sk(sk);
 457         unsigned char *b = skb->tail;
 458         struct inet_diag_msg *r;
 459         struct nlmsghdr *nlh;
 460         long tmo;
 461
 462         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
 463         nlh->nlmsg_flags = NLM_F_MULTI;
 464         r = NLMSG_DATA(nlh);
 465
 466         r->idiag_family = sk->sk_family;
 467         r->idiag_state = TCP_SYN_RECV;
 468         r->idiag_timer = 1;
 469         r->idiag_retrans = req->retrans;
 470
 471         r->id.idiag_if = sk->sk_bound_dev_if;
 472         r->id.idiag_cookie[0] = (u32)(unsigned long)req;
 473         r->id.idiag_cookie[1] = (u32)(((unsigned long)req >> 31) >> 1);
 474
 475         tmo = req->expires - jiffies;
 476         if (tmo < 0)
 477                 tmo = 0;
 478
 479         r->id.idiag_sport = inet->sport;
 480         r->id.idiag_dport = ireq->rmt_port;
 481         r->id.idiag_src[0] = ireq->loc_addr;
 482         r->id.idiag_dst[0] = ireq->rmt_addr;
 483         r->idiag_expires = jiffies_to_msecs(tmo);
 484         r->idiag_rqueue = 0;
 485         r->idiag_wqueue = 0;
 486         r->idiag_uid = sock_i_uid(sk);
 487         r->idiag_inode = 0;
 488 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 489         if (r->idiag_family == AF_INET6) {
 490                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 491                                &tcp6_rsk(req)->loc_addr);
 492                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 493                                &tcp6_rsk(req)->rmt_addr);
 494         }
 495 #endif
 496         nlh->nlmsg_len = skb->tail - b;
 497
 498         return skb->len;
 499
 500 nlmsg_failure:
 501         skb_trim(skb, b - skb->data);
 502         return -1;
 503 }
 504
 505 static int inet_diag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 506                              struct netlink_callback *cb)
 507 {
 508         struct inet_diag_entry entry;
 509         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 510         struct inet_connection_sock *icsk = inet_csk(sk);
 511         struct listen_sock *lopt;
 512         struct rtattr *bc = NULL;
 513         struct inet_sock *inet = inet_sk(sk);
 514         int j, s_j;
 515         int reqnum, s_reqnum;
 516         int err = 0;
 517
 518         s_j = cb->args[3];
 519         s_reqnum = cb->args[4];
 520
 521         if (s_j > 0)
 522                 s_j--;
 523
 524         entry.family = sk->sk_family;
 525
 526         read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 527
 528         lopt = icsk->icsk_accept_queue.listen_opt;
 529         if (!lopt || !lopt->qlen)
 530                 goto out;
 531
 532         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 533                 bc = (struct rtattr *)(r + 1);
 534                 entry.sport = inet->num;
 535                 entry.userlocks = sk->sk_userlocks;
 536         }
 537
 538         for (j = s_j; j < lopt->nr_table_entries; j++) {
 539                 struct request_sock *req, *head = lopt->syn_table[j];
 540
 541                 reqnum = 0;
 542                 for (req = head; req; reqnum++, req = req->dl_next) {
 543                         struct inet_request_sock *ireq = inet_rsk(req);
 544
 545                         if (reqnum < s_reqnum)
 546                                 continue;
 547                         if (r->id.idiag_dport != ireq->rmt_port &&
 548                             r->id.idiag_dport)
 549                                 continue;
 550
 551                         if (bc) {
 552                                 entry.saddr =
 553 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 554                                         (entry.family == AF_INET6) ?
 555                                         tcp6_rsk(req)->loc_addr.s6_addr32 :
 556 #endif
 557                                         &ireq->loc_addr;
 558                                 entry.daddr =
 559 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 560                                         (entry.family == AF_INET6) ?
 561                                         tcp6_rsk(req)->rmt_addr.s6_addr32 :
 562 #endif
 563                                         &ireq->rmt_addr;
 564                                 entry.dport = ntohs(ireq->rmt_port);
 565
 566                                 if (!inet_diag_bc_run(RTA_DATA(bc),
 567                                                     RTA_PAYLOAD(bc), &entry))
 568                                         continue;
 569                         }
 570
 571                         err = inet_diag_fill_req(skb, sk, req,
 572                                                NETLINK_CB(cb->skb).pid,
 573                                                cb->nlh->nlmsg_seq, cb->nlh);
 574                         if (err < 0) {
 575                                 cb->args[3] = j + 1;
 576                                 cb->args[4] = reqnum;
 577                                 goto out;
 578                         }
 579                 }
 580
 581                 s_reqnum = 0;
 582         }
 583
 584 out:
 585         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 586
 587         return err;
 588 }
 589
 590 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 591 {
 592         int i, num;
 593         int s_i, s_num;
 594         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 595         const struct inet_diag_handler *handler;
 596         struct inet_hashinfo *hashinfo;
 597
 598         handler = inet_diag_table[cb->nlh->nlmsg_type];
 599         BUG_ON(handler == NULL);
 600         hashinfo = handler->idiag_hashinfo;
 601
 602         s_i = cb->args[1];
 603         s_num = num = cb->args[2];
 604
 605         if (cb->args[0] == 0) {
 606                 if (!(r->idiag_states & (TCPF_LISTEN | TCPF_SYN_RECV)))
 607                         goto skip_listen_ht;
 608
 609                 inet_listen_lock(hashinfo);
 610                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 611                         struct sock *sk;
 612                         struct hlist_node *node;
 613
 614                         num = 0;
 615                         sk_for_each(sk, node, &hashinfo->listening_hash[i]) {
 616                                 struct inet_sock *inet = inet_sk(sk);
 617
 618                                 if (num < s_num) {
 619                                         num++;
 620                                         continue;
 621                                 }
 622
 623                                 if (r->id.idiag_sport != inet->sport &&
 624                                     r->id.idiag_sport)
 625                                         goto next_listen;
 626
 627                                 if (!(r->idiag_states & TCPF_LISTEN) ||
 628                                     r->id.idiag_dport ||
 629                                     cb->args[3] > 0)
 630                                         goto syn_recv;
 631
 632                                 if (inet_diag_dump_sock(skb, sk, cb) < 0) {
 633                                         inet_listen_unlock(hashinfo);
 634                                         goto done;
 635                                 }
 636
 637 syn_recv:
 638                                 if (!(r->idiag_states & TCPF_SYN_RECV))
 639                                         goto next_listen;
 640
 641                                 if (inet_diag_dump_reqs(skb, sk, cb) < 0) {
 642                                         inet_listen_unlock(hashinfo);
 643                                         goto done;
 644                                 }
 645
 646 next_listen:
 647                                 cb->args[3] = 0;
 648                                 cb->args[4] = 0;
 649                                 ++num;
 650                         }
 651
 652                         s_num = 0;
 653                         cb->args[3] = 0;
 654                         cb->args[4] = 0;
 655                 }
 656                 inet_listen_unlock(hashinfo);
 657 skip_listen_ht:
 658                 cb->args[0] = 1;
 659                 s_i = num = s_num = 0;
 660         }
 661
 662         if (!(r->idiag_states & ~(TCPF_LISTEN | TCPF_SYN_RECV)))
 663                 return skb->len;
 664
 665         for (i = s_i; i < hashinfo->ehash_size; i++) {
 666                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 667                 struct sock *sk;
 668                 struct hlist_node *node;
 669
 670                 if (i > s_i)
 671                         s_num = 0;
 672
 673                 read_lock_bh(&head->lock);
 674
 675                 num = 0;
 676                 sk_for_each(sk, node, &head->chain) {
 677                         struct inet_sock *inet = inet_sk(sk);
 678
 679                         if (num < s_num)
 680                                 goto next_normal;
 681                         if (!(r->idiag_states & (1 << sk->sk_state)))
 682                                 goto next_normal;
 683                         if (r->id.idiag_sport != inet->sport &&
 684                             r->id.idiag_sport)
 685                                 goto next_normal;
 686                         if (r->id.idiag_dport != inet->dport && r->id.idiag_dport)
 687                                 goto next_normal;
 688                         if (inet_diag_dump_sock(skb, sk, cb) < 0) {
 689                                 read_unlock_bh(&head->lock);
 690                                 goto done;
 691                         }
 692 next_normal:
 693                         ++num;
 694                 }
 695
 696                 if (r->idiag_states & TCPF_TIME_WAIT) {
 697                         sk_for_each(sk, node,
 698                                     &hashinfo->ehash[i + hashinfo->ehash_size].chain) {
 699                                 struct inet_sock *inet = inet_sk(sk);
 700
 701                                 if (num < s_num)
 702                                         goto next_dying;
 703                                 if (r->id.idiag_sport != inet->sport &&
 704                                     r->id.idiag_sport)
 705                                         goto next_dying;
 706                                 if (r->id.idiag_dport != inet->dport &&
 707                                     r->id.idiag_dport)
 708                                         goto next_dying;
 709                                 if (inet_diag_dump_sock(skb, sk, cb) < 0) {
 710                                         read_unlock_bh(&head->lock);
 711                                         goto done;
 712                                 }
 713 next_dying:
 714                                 ++num;
 715                         }
 716                 }
 717                 read_unlock_bh(&head->lock);
 718         }
 719
 720 done:
 721         cb->args[1] = i;
 722         cb->args[2] = num;
 723         return skb->len;
 724 }
 725
 726 static int inet_diag_dump_done(struct netlink_callback *cb)
 727 {
 728         return 0;
 729 }
 730
 731
 732 static __inline__ int
 733 inet_diag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 734 {
 735         if (!(nlh->nlmsg_flags&NLM_F_REQUEST))
 736                 return 0;
 737
 738         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX)
 739                 goto err_inval;
 740
 741         if (inet_diag_table[nlh->nlmsg_type] == NULL)
 742                 return -ENOENT;
 743
 744         if (NLMSG_LENGTH(sizeof(struct inet_diag_req)) > skb->len)
 745                 goto err_inval;
 746
 747         if (nlh->nlmsg_flags&NLM_F_DUMP) {
 748                 if (nlh->nlmsg_len >
 749                     (4 + NLMSG_SPACE(sizeof(struct inet_diag_req)))) {
 750                         struct rtattr *rta = (void *)(NLMSG_DATA(nlh) +
 751                                                  sizeof(struct inet_diag_req));
 752                         if (rta->rta_type != INET_DIAG_REQ_BYTECODE ||
 753                             rta->rta_len < 8 ||
 754                             rta->rta_len >
 755                             (nlh->nlmsg_len -
 756                              NLMSG_SPACE(sizeof(struct inet_diag_req))))
 757                                 goto err_inval;
 758                         if (inet_diag_bc_audit(RTA_DATA(rta), RTA_PAYLOAD(rta)))
 759                                 goto err_inval;
 760                 }
 761                 return netlink_dump_start(idiagnl, skb, nlh,
 762                                           inet_diag_dump,
 763                                           inet_diag_dump_done);
 764         } else {
 765                 return inet_diag_get_exact(skb, nlh);
 766         }
 767
 768 err_inval:
 769         return -EINVAL;
 770 }
 771
 772
 773 static inline void inet_diag_rcv_skb(struct sk_buff *skb)
 774 {
 775         int err;
 776         struct nlmsghdr * nlh;
 777
 778         if (skb->len >= NLMSG_SPACE(0)) {
 779                 nlh = (struct nlmsghdr *)skb->data;
 780                 if (nlh->nlmsg_len < sizeof(*nlh) || skb->len < nlh->nlmsg_len)
 781                         return;
 782                 err = inet_diag_rcv_msg(skb, nlh);
 783                 if (err || nlh->nlmsg_flags & NLM_F_ACK)
 784                         netlink_ack(skb, nlh, err);
 785         }
 786 }
 787
 788 static void inet_diag_rcv(struct sock *sk, int len)
 789 {
 790         struct sk_buff *skb;
 791         unsigned int qlen = skb_queue_len(&sk->sk_receive_queue);
 792
 793         while (qlen-- && (skb = skb_dequeue(&sk->sk_receive_queue))) {
 794                 inet_diag_rcv_skb(skb);
 795                 kfree_skb(skb);
 796         }
 797 }
 798
 799 static void tcp_diag_get_info(struct sock *sk, struct inet_diag_msg *r,
 800                               void *_info)
 801 {
 802         const struct tcp_sock *tp = tcp_sk(sk);
 803         struct tcp_info *info = _info;
 804
 805         r->idiag_rqueue = tp->rcv_nxt - tp->copied_seq;
 806         r->idiag_wqueue = tp->write_seq - tp->snd_una;
 807         if (info != NULL)
 808                 tcp_get_info(sk, info);
 809 }
 810
 811 static struct inet_diag_handler tcp_diag_handler = {
 812         .idiag_hashinfo  = &tcp_hashinfo,
 813         .idiag_get_info  = tcp_diag_get_info,
 814         .idiag_type      = TCPDIAG_GETSOCK,
 815         .idiag_info_size = sizeof(struct tcp_info),
 816 };
 817
 818 static DEFINE_SPINLOCK(inet_diag_register_lock);
 819
 820 int inet_diag_register(const struct inet_diag_handler *h)
 821 {
 822         const __u16 type = h->idiag_type;
 823         int err = -EINVAL;
 824
 825         if (type >= INET_DIAG_GETSOCK_MAX)
 826                 goto out;
 827
 828         spin_lock(&inet_diag_register_lock);
 829         err = -EEXIST;
 830         if (inet_diag_table[type] == NULL) {
 831                 inet_diag_table[type] = h;
 832                 err = 0;
 833         }
 834         spin_unlock(&inet_diag_register_lock);
 835 out:
 836         return err;
 837 }
 838 EXPORT_SYMBOL_GPL(inet_diag_register);
 839
 840 void inet_diag_unregister(const struct inet_diag_handler *h)
 841 {
 842         const __u16 type = h->idiag_type;
 843
 844         if (type >= INET_DIAG_GETSOCK_MAX)
 845                 return;
 846
 847         spin_lock(&inet_diag_register_lock);
 848         inet_diag_table[type] = NULL;
 849         spin_unlock(&inet_diag_register_lock);
 850
 851         synchronize_rcu();
 852 }
 853 EXPORT_SYMBOL_GPL(inet_diag_unregister);
 854
 855 static int __init inet_diag_init(void)
 856 {
 857         const int inet_diag_table_size = (INET_DIAG_GETSOCK_MAX *
 858                                           sizeof(struct inet_diag_handler *));
 859         int err = -ENOMEM;
 860
 861         inet_diag_table = kmalloc(inet_diag_table_size, GFP_KERNEL);
 862         if (!inet_diag_table)
 863                 goto out;
 864
 865         memset(inet_diag_table, 0, inet_diag_table_size);
 866
 867         idiagnl = netlink_kernel_create(NETLINK_INET_DIAG, inet_diag_rcv,
 868                                         THIS_MODULE);
 869         if (idiagnl == NULL)
 870                 goto out_free_table;
 871
 872         err = inet_diag_register(&tcp_diag_handler);
 873         if (err)
 874                 goto out_sock_release;
 875 out:
 876         return err;
 877 out_sock_release:
 878         sock_release(idiagnl->sk_socket);
 879 out_free_table:
 880         kfree(inet_diag_table);
 881         goto out;
 882 }
 883
 884 static void __exit inet_diag_exit(void)
 885 {
 886         sock_release(idiagnl->sk_socket);
 887         kfree(inet_diag_table);
 888 }
 889
 890 module_init(inet_diag_init);
 891 module_exit(inet_diag_exit);
 892 MODULE_LICENSE("GPL");