net/ipv4/tcp_diag.c

   1 /*
   2  * tcp_diag.c   Module for monitoring TCP sockets.
   3  *
   4  * Version:     $Id: tcp_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 struct tcpdiag_entry
  38 {
  39         u32 *saddr;
  40         u32 *daddr;
  41         u16 sport;
  42         u16 dport;
  43         u16 family;
  44         u16 userlocks;
  45 };
  46
  47 static struct sock *tcpnl;
  48
  49 #define TCPDIAG_PUT(skb, attrtype, attrlen) \
  50         RTA_DATA(__RTA_PUT(skb, attrtype, attrlen))
  51
  52 #ifdef CONFIG_IP_TCPDIAG_DCCP
  53 extern struct inet_hashinfo dccp_hashinfo;
  54 #endif
  55
  56 static int tcpdiag_fill(struct sk_buff *skb, struct sock *sk,
  57                         int ext, u32 pid, u32 seq, u16 nlmsg_flags,
  58                         const struct nlmsghdr *unlh)
  59 {
  60         const struct inet_sock *inet = inet_sk(sk);
  61         const struct inet_connection_sock *icsk = inet_csk(sk);
  62         struct tcpdiagmsg *r;
  63         struct nlmsghdr  *nlh;
  64         struct tcp_info  *info = NULL;
  65         struct tcpdiag_meminfo  *minfo = NULL;
  66         unsigned char    *b = skb->tail;
  67
  68         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
  69         nlh->nlmsg_flags = nlmsg_flags;
  70         r = NLMSG_DATA(nlh);
  71         if (sk->sk_state != TCP_TIME_WAIT) {
  72                 if (ext & (1<<(TCPDIAG_MEMINFO-1)))
  73                         minfo = TCPDIAG_PUT(skb, TCPDIAG_MEMINFO, sizeof(*minfo));
  74                 if (ext & (1<<(TCPDIAG_INFO-1)))
  75                         info = TCPDIAG_PUT(skb, TCPDIAG_INFO, sizeof(*info));
  76
  77                 if ((ext & (1 << (TCPDIAG_CONG - 1))) && icsk->icsk_ca_ops) {
  78                         size_t len = strlen(icsk->icsk_ca_ops->name);
  79                         strcpy(TCPDIAG_PUT(skb, TCPDIAG_CONG, len+1),
  80                                icsk->icsk_ca_ops->name);
  81                 }
  82         }
  83         r->tcpdiag_family = sk->sk_family;
  84         r->tcpdiag_state = sk->sk_state;
  85         r->tcpdiag_timer = 0;
  86         r->tcpdiag_retrans = 0;
  87
  88         r->id.tcpdiag_if = sk->sk_bound_dev_if;
  89         r->id.tcpdiag_cookie[0] = (u32)(unsigned long)sk;
  90         r->id.tcpdiag_cookie[1] = (u32)(((unsigned long)sk >> 31) >> 1);
  91
  92         if (r->tcpdiag_state == TCP_TIME_WAIT) {
  93                 const struct inet_timewait_sock *tw = inet_twsk(sk);
  94                 long tmo = tw->tw_ttd - jiffies;
  95                 if (tmo < 0)
  96                         tmo = 0;
  97
  98                 r->id.tcpdiag_sport = tw->tw_sport;
  99                 r->id.tcpdiag_dport = tw->tw_dport;
 100                 r->id.tcpdiag_src[0] = tw->tw_rcv_saddr;
 101                 r->id.tcpdiag_dst[0] = tw->tw_daddr;
 102                 r->tcpdiag_state = tw->tw_substate;
 103                 r->tcpdiag_timer = 3;
 104                 r->tcpdiag_expires = (tmo*1000+HZ-1)/HZ;
 105                 r->tcpdiag_rqueue = 0;
 106                 r->tcpdiag_wqueue = 0;
 107                 r->tcpdiag_uid = 0;
 108                 r->tcpdiag_inode = 0;
 109 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 110                 if (r->tcpdiag_family == AF_INET6) {
 111                         const struct tcp6_timewait_sock *tcp6tw = tcp6_twsk(sk);
 112
 113                         ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
 114                                        &tcp6tw->tw_v6_rcv_saddr);
 115                         ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
 116                                        &tcp6tw->tw_v6_daddr);
 117                 }
 118 #endif
 119                 nlh->nlmsg_len = skb->tail - b;
 120                 return skb->len;
 121         }
 122
 123         r->id.tcpdiag_sport = inet->sport;
 124         r->id.tcpdiag_dport = inet->dport;
 125         r->id.tcpdiag_src[0] = inet->rcv_saddr;
 126         r->id.tcpdiag_dst[0] = inet->daddr;
 127
 128 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 129         if (r->tcpdiag_family == AF_INET6) {
 130                 struct ipv6_pinfo *np = inet6_sk(sk);
 131
 132                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
 133                                &np->rcv_saddr);
 134                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
 135                                &np->daddr);
 136         }
 137 #endif
 138
 139 #define EXPIRES_IN_MS(tmo)  ((tmo-jiffies)*1000+HZ-1)/HZ
 140
 141         if (icsk->icsk_pending == ICSK_TIME_RETRANS) {
 142                 r->tcpdiag_timer = 1;
 143                 r->tcpdiag_retrans = icsk->icsk_retransmits;
 144                 r->tcpdiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 145         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 146                 r->tcpdiag_timer = 4;
 147                 r->tcpdiag_retrans = icsk->icsk_probes_out;
 148                 r->tcpdiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 149         } else if (timer_pending(&sk->sk_timer)) {
 150                 r->tcpdiag_timer = 2;
 151                 r->tcpdiag_retrans = icsk->icsk_probes_out;
 152                 r->tcpdiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 153         } else {
 154                 r->tcpdiag_timer = 0;
 155                 r->tcpdiag_expires = 0;
 156         }
 157 #undef EXPIRES_IN_MS
 158         /*
 159          * Ahem... for now we'll have some knowledge about TCP -acme
 160          * But this is just one of two small exceptions, both in this
 161          * function, so lets close our eyes for some 15 lines or so... 8)
 162          * -acme
 163          */
 164         if (sk->sk_protocol == IPPROTO_TCP) {
 165                 const struct tcp_sock *tp = tcp_sk(sk);
 166
 167                 r->tcpdiag_rqueue = tp->rcv_nxt - tp->copied_seq;
 168                 r->tcpdiag_wqueue = tp->write_seq - tp->snd_una;
 169         } else
 170                 r->tcpdiag_rqueue = r->tcpdiag_wqueue = 0;
 171
 172         r->tcpdiag_uid = sock_i_uid(sk);
 173         r->tcpdiag_inode = sock_i_ino(sk);
 174
 175         if (minfo) {
 176                 minfo->tcpdiag_rmem = atomic_read(&sk->sk_rmem_alloc);
 177                 minfo->tcpdiag_wmem = sk->sk_wmem_queued;
 178                 minfo->tcpdiag_fmem = sk->sk_forward_alloc;
 179                 minfo->tcpdiag_tmem = atomic_read(&sk->sk_wmem_alloc);
 180         }
 181
 182         /* Ahem... for now we'll have some knowledge about TCP -acme */
 183         if (info) {
 184                 if (sk->sk_protocol == IPPROTO_TCP)
 185                         tcp_get_info(sk, info);
 186                 else
 187                         memset(info, 0, sizeof(*info));
 188         }
 189
 190         if (sk->sk_state < TCP_TIME_WAIT &&
 191             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 192                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 193
 194         nlh->nlmsg_len = skb->tail - b;
 195         return skb->len;
 196
 197 rtattr_failure:
 198 nlmsg_failure:
 199         skb_trim(skb, b - skb->data);
 200         return -1;
 201 }
 202
 203 static int tcpdiag_get_exact(struct sk_buff *in_skb, const struct nlmsghdr *nlh)
 204 {
 205         int err;
 206         struct sock *sk;
 207         struct tcpdiagreq *req = NLMSG_DATA(nlh);
 208         struct sk_buff *rep;
 209         struct inet_hashinfo *hashinfo = &tcp_hashinfo;
 210 #ifdef CONFIG_IP_TCPDIAG_DCCP
 211         if (nlh->nlmsg_type == DCCPDIAG_GETSOCK)
 212                 hashinfo = &dccp_hashinfo;
 213 #endif
 214         if (req->tcpdiag_family == AF_INET) {
 215                 sk = inet_lookup(hashinfo, req->id.tcpdiag_dst[0],
 216                                  req->id.tcpdiag_dport, req->id.tcpdiag_src[0],
 217                                  req->id.tcpdiag_sport, req->id.tcpdiag_if);
 218         }
 219 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 220         else if (req->tcpdiag_family == AF_INET6) {
 221                 sk = inet6_lookup(hashinfo,
 222                                   (struct in6_addr*)req->id.tcpdiag_dst,
 223                                   req->id.tcpdiag_dport,
 224                                   (struct in6_addr*)req->id.tcpdiag_src,
 225                                   req->id.tcpdiag_sport,
 226                                   req->id.tcpdiag_if);
 227         }
 228 #endif
 229         else {
 230                 return -EINVAL;
 231         }
 232
 233         if (sk == NULL)
 234                 return -ENOENT;
 235
 236         err = -ESTALE;
 237         if ((req->id.tcpdiag_cookie[0] != TCPDIAG_NOCOOKIE ||
 238              req->id.tcpdiag_cookie[1] != TCPDIAG_NOCOOKIE) &&
 239             ((u32)(unsigned long)sk != req->id.tcpdiag_cookie[0] ||
 240              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.tcpdiag_cookie[1]))
 241                 goto out;
 242
 243         err = -ENOMEM;
 244         rep = alloc_skb(NLMSG_SPACE(sizeof(struct tcpdiagmsg)+
 245                                     sizeof(struct tcpdiag_meminfo)+
 246                                     sizeof(struct tcp_info)+64), GFP_KERNEL);
 247         if (!rep)
 248                 goto out;
 249
 250         if (tcpdiag_fill(rep, sk, req->tcpdiag_ext,
 251                          NETLINK_CB(in_skb).pid,
 252                          nlh->nlmsg_seq, 0, nlh) <= 0)
 253                 BUG();
 254
 255         err = netlink_unicast(tcpnl, rep, NETLINK_CB(in_skb).pid, MSG_DONTWAIT);
 256         if (err > 0)
 257                 err = 0;
 258
 259 out:
 260         if (sk) {
 261                 if (sk->sk_state == TCP_TIME_WAIT)
 262                         inet_twsk_put((struct inet_timewait_sock *)sk);
 263                 else
 264                         sock_put(sk);
 265         }
 266         return err;
 267 }
 268
 269 static int bitstring_match(const u32 *a1, const u32 *a2, int bits)
 270 {
 271         int words = bits >> 5;
 272
 273         bits &= 0x1f;
 274
 275         if (words) {
 276                 if (memcmp(a1, a2, words << 2))
 277                         return 0;
 278         }
 279         if (bits) {
 280                 __u32 w1, w2;
 281                 __u32 mask;
 282
 283                 w1 = a1[words];
 284                 w2 = a2[words];
 285
 286                 mask = htonl((0xffffffff) << (32 - bits));
 287
 288                 if ((w1 ^ w2) & mask)
 289                         return 0;
 290         }
 291
 292         return 1;
 293 }
 294
 295
 296 static int tcpdiag_bc_run(const void *bc, int len,
 297                           const struct tcpdiag_entry *entry)
 298 {
 299         while (len > 0) {
 300                 int yes = 1;
 301                 const struct tcpdiag_bc_op *op = bc;
 302
 303                 switch (op->code) {
 304                 case TCPDIAG_BC_NOP:
 305                         break;
 306                 case TCPDIAG_BC_JMP:
 307                         yes = 0;
 308                         break;
 309                 case TCPDIAG_BC_S_GE:
 310                         yes = entry->sport >= op[1].no;
 311                         break;
 312                 case TCPDIAG_BC_S_LE:
 313                         yes = entry->dport <= op[1].no;
 314                         break;
 315                 case TCPDIAG_BC_D_GE:
 316                         yes = entry->dport >= op[1].no;
 317                         break;
 318                 case TCPDIAG_BC_D_LE:
 319                         yes = entry->dport <= op[1].no;
 320                         break;
 321                 case TCPDIAG_BC_AUTO:
 322                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 323                         break;
 324                 case TCPDIAG_BC_S_COND:
 325                 case TCPDIAG_BC_D_COND:
 326                 {
 327                         struct tcpdiag_hostcond *cond = (struct tcpdiag_hostcond*)(op+1);
 328                         u32 *addr;
 329
 330                         if (cond->port != -1 &&
 331                             cond->port != (op->code == TCPDIAG_BC_S_COND ?
 332                                              entry->sport : entry->dport)) {
 333                                 yes = 0;
 334                                 break;
 335                         }
 336
 337                         if (cond->prefix_len == 0)
 338                                 break;
 339
 340                         if (op->code == TCPDIAG_BC_S_COND)
 341                                 addr = entry->saddr;
 342                         else
 343                                 addr = entry->daddr;
 344
 345                         if (bitstring_match(addr, cond->addr, cond->prefix_len))
 346                                 break;
 347                         if (entry->family == AF_INET6 &&
 348                             cond->family == AF_INET) {
 349                                 if (addr[0] == 0 && addr[1] == 0 &&
 350                                     addr[2] == htonl(0xffff) &&
 351                                     bitstring_match(addr+3, cond->addr, cond->prefix_len))
 352                                         break;
 353                         }
 354                         yes = 0;
 355                         break;
 356                 }
 357                 }
 358
 359                 if (yes) {
 360                         len -= op->yes;
 361                         bc += op->yes;
 362                 } else {
 363                         len -= op->no;
 364                         bc += op->no;
 365                 }
 366         }
 367         return (len == 0);
 368 }
 369
 370 static int valid_cc(const void *bc, int len, int cc)
 371 {
 372         while (len >= 0) {
 373                 const struct tcpdiag_bc_op *op = bc;
 374
 375                 if (cc > len)
 376                         return 0;
 377                 if (cc == len)
 378                         return 1;
 379                 if (op->yes < 4)
 380                         return 0;
 381                 len -= op->yes;
 382                 bc  += op->yes;
 383         }
 384         return 0;
 385 }
 386
 387 static int tcpdiag_bc_audit(const void *bytecode, int bytecode_len)
 388 {
 389         const unsigned char *bc = bytecode;
 390         int  len = bytecode_len;
 391
 392         while (len > 0) {
 393                 struct tcpdiag_bc_op *op = (struct tcpdiag_bc_op*)bc;
 394
 395 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 396                 switch (op->code) {
 397                 case TCPDIAG_BC_AUTO:
 398                 case TCPDIAG_BC_S_COND:
 399                 case TCPDIAG_BC_D_COND:
 400                 case TCPDIAG_BC_S_GE:
 401                 case TCPDIAG_BC_S_LE:
 402                 case TCPDIAG_BC_D_GE:
 403                 case TCPDIAG_BC_D_LE:
 404                         if (op->yes < 4 || op->yes > len+4)
 405                                 return -EINVAL;
 406                 case TCPDIAG_BC_JMP:
 407                         if (op->no < 4 || op->no > len+4)
 408                                 return -EINVAL;
 409                         if (op->no < len &&
 410                             !valid_cc(bytecode, bytecode_len, len-op->no))
 411                                 return -EINVAL;
 412                         break;
 413                 case TCPDIAG_BC_NOP:
 414                         if (op->yes < 4 || op->yes > len+4)
 415                                 return -EINVAL;
 416                         break;
 417                 default:
 418                         return -EINVAL;
 419                 }
 420                 bc += op->yes;
 421                 len -= op->yes;
 422         }
 423         return len == 0 ? 0 : -EINVAL;
 424 }
 425
 426 static int tcpdiag_dump_sock(struct sk_buff *skb, struct sock *sk,
 427                              struct netlink_callback *cb)
 428 {
 429         struct tcpdiagreq *r = NLMSG_DATA(cb->nlh);
 430
 431         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 432                 struct tcpdiag_entry entry;
 433                 struct rtattr *bc = (struct rtattr *)(r + 1);
 434                 struct inet_sock *inet = inet_sk(sk);
 435
 436                 entry.family = sk->sk_family;
 437 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 438                 if (entry.family == AF_INET6) {
 439                         struct ipv6_pinfo *np = inet6_sk(sk);
 440
 441                         entry.saddr = np->rcv_saddr.s6_addr32;
 442                         entry.daddr = np->daddr.s6_addr32;
 443                 } else
 444 #endif
 445                 {
 446                         entry.saddr = &inet->rcv_saddr;
 447                         entry.daddr = &inet->daddr;
 448                 }
 449                 entry.sport = inet->num;
 450                 entry.dport = ntohs(inet->dport);
 451                 entry.userlocks = sk->sk_userlocks;
 452
 453                 if (!tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), &entry))
 454                         return 0;
 455         }
 456
 457         return tcpdiag_fill(skb, sk, r->tcpdiag_ext, NETLINK_CB(cb->skb).pid,
 458                             cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 459 }
 460
 461 static int tcpdiag_fill_req(struct sk_buff *skb, struct sock *sk,
 462                             struct request_sock *req,
 463                             u32 pid, u32 seq,
 464                             const struct nlmsghdr *unlh)
 465 {
 466         const struct inet_request_sock *ireq = inet_rsk(req);
 467         struct inet_sock *inet = inet_sk(sk);
 468         unsigned char *b = skb->tail;
 469         struct tcpdiagmsg *r;
 470         struct nlmsghdr *nlh;
 471         long tmo;
 472
 473         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
 474         nlh->nlmsg_flags = NLM_F_MULTI;
 475         r = NLMSG_DATA(nlh);
 476
 477         r->tcpdiag_family = sk->sk_family;
 478         r->tcpdiag_state = TCP_SYN_RECV;
 479         r->tcpdiag_timer = 1;
 480         r->tcpdiag_retrans = req->retrans;
 481
 482         r->id.tcpdiag_if = sk->sk_bound_dev_if;
 483         r->id.tcpdiag_cookie[0] = (u32)(unsigned long)req;
 484         r->id.tcpdiag_cookie[1] = (u32)(((unsigned long)req >> 31) >> 1);
 485
 486         tmo = req->expires - jiffies;
 487         if (tmo < 0)
 488                 tmo = 0;
 489
 490         r->id.tcpdiag_sport = inet->sport;
 491         r->id.tcpdiag_dport = ireq->rmt_port;
 492         r->id.tcpdiag_src[0] = ireq->loc_addr;
 493         r->id.tcpdiag_dst[0] = ireq->rmt_addr;
 494         r->tcpdiag_expires = jiffies_to_msecs(tmo),
 495         r->tcpdiag_rqueue = 0;
 496         r->tcpdiag_wqueue = 0;
 497         r->tcpdiag_uid = sock_i_uid(sk);
 498         r->tcpdiag_inode = 0;
 499 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 500         if (r->tcpdiag_family == AF_INET6) {
 501                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
 502                                &tcp6_rsk(req)->loc_addr);
 503                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
 504                                &tcp6_rsk(req)->rmt_addr);
 505         }
 506 #endif
 507         nlh->nlmsg_len = skb->tail - b;
 508
 509         return skb->len;
 510
 511 nlmsg_failure:
 512         skb_trim(skb, b - skb->data);
 513         return -1;
 514 }
 515
 516 static int tcpdiag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 517                              struct netlink_callback *cb)
 518 {
 519         struct tcpdiag_entry entry;
 520         struct tcpdiagreq *r = NLMSG_DATA(cb->nlh);
 521         struct inet_connection_sock *icsk = inet_csk(sk);
 522         struct listen_sock *lopt;
 523         struct rtattr *bc = NULL;
 524         struct inet_sock *inet = inet_sk(sk);
 525         int j, s_j;
 526         int reqnum, s_reqnum;
 527         int err = 0;
 528
 529         s_j = cb->args[3];
 530         s_reqnum = cb->args[4];
 531
 532         if (s_j > 0)
 533                 s_j--;
 534
 535         entry.family = sk->sk_family;
 536
 537         read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 538
 539         lopt = icsk->icsk_accept_queue.listen_opt;
 540         if (!lopt || !lopt->qlen)
 541                 goto out;
 542
 543         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 544                 bc = (struct rtattr *)(r + 1);
 545                 entry.sport = inet->num;
 546                 entry.userlocks = sk->sk_userlocks;
 547         }
 548
 549         for (j = s_j; j < lopt->nr_table_entries; j++) {
 550                 struct request_sock *req, *head = lopt->syn_table[j];
 551
 552                 reqnum = 0;
 553                 for (req = head; req; reqnum++, req = req->dl_next) {
 554                         struct inet_request_sock *ireq = inet_rsk(req);
 555
 556                         if (reqnum < s_reqnum)
 557                                 continue;
 558                         if (r->id.tcpdiag_dport != ireq->rmt_port &&
 559                             r->id.tcpdiag_dport)
 560                                 continue;
 561
 562                         if (bc) {
 563                                 entry.saddr =
 564 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 565                                         (entry.family == AF_INET6) ?
 566                                         tcp6_rsk(req)->loc_addr.s6_addr32 :
 567 #endif
 568                                         &ireq->loc_addr;
 569                                 entry.daddr =
 570 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 571                                         (entry.family == AF_INET6) ?
 572                                         tcp6_rsk(req)->rmt_addr.s6_addr32 :
 573 #endif
 574                                         &ireq->rmt_addr;
 575                                 entry.dport = ntohs(ireq->rmt_port);
 576
 577                                 if (!tcpdiag_bc_run(RTA_DATA(bc),
 578                                                     RTA_PAYLOAD(bc), &entry))
 579                                         continue;
 580                         }
 581
 582                         err = tcpdiag_fill_req(skb, sk, req,
 583                                                NETLINK_CB(cb->skb).pid,
 584                                                cb->nlh->nlmsg_seq, cb->nlh);
 585                         if (err < 0) {
 586                                 cb->args[3] = j + 1;
 587                                 cb->args[4] = reqnum;
 588                                 goto out;
 589                         }
 590                 }
 591
 592                 s_reqnum = 0;
 593         }
 594
 595 out:
 596         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 597
 598         return err;
 599 }
 600
 601 static int tcpdiag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 602 {
 603         int i, num;
 604         int s_i, s_num;
 605         struct tcpdiagreq *r = NLMSG_DATA(cb->nlh);
 606         struct inet_hashinfo *hashinfo;
 607
 608         s_i = cb->args[1];
 609         s_num = num = cb->args[2];
 610                 hashinfo = &tcp_hashinfo;
 611 #ifdef CONFIG_IP_TCPDIAG_DCCP
 612         if (cb->nlh->nlmsg_type == DCCPDIAG_GETSOCK)
 613                 hashinfo = &dccp_hashinfo;
 614 #endif
 615         if (cb->args[0] == 0) {
 616                 if (!(r->tcpdiag_states&(TCPF_LISTEN|TCPF_SYN_RECV)))
 617                         goto skip_listen_ht;
 618
 619                 inet_listen_lock(hashinfo);
 620                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 621                         struct sock *sk;
 622                         struct hlist_node *node;
 623
 624                         num = 0;
 625                         sk_for_each(sk, node, &hashinfo->listening_hash[i]) {
 626                                 struct inet_sock *inet = inet_sk(sk);
 627
 628                                 if (num < s_num) {
 629                                         num++;
 630                                         continue;
 631                                 }
 632
 633                                 if (r->id.tcpdiag_sport != inet->sport &&
 634                                     r->id.tcpdiag_sport)
 635                                         goto next_listen;
 636
 637                                 if (!(r->tcpdiag_states&TCPF_LISTEN) ||
 638                                     r->id.tcpdiag_dport ||
 639                                     cb->args[3] > 0)
 640                                         goto syn_recv;
 641
 642                                 if (tcpdiag_dump_sock(skb, sk, cb) < 0) {
 643                                         inet_listen_unlock(hashinfo);
 644                                         goto done;
 645                                 }
 646
 647 syn_recv:
 648                                 if (!(r->tcpdiag_states&TCPF_SYN_RECV))
 649                                         goto next_listen;
 650
 651                                 if (tcpdiag_dump_reqs(skb, sk, cb) < 0) {
 652                                         inet_listen_unlock(hashinfo);
 653                                         goto done;
 654                                 }
 655
 656 next_listen:
 657                                 cb->args[3] = 0;
 658                                 cb->args[4] = 0;
 659                                 ++num;
 660                         }
 661
 662                         s_num = 0;
 663                         cb->args[3] = 0;
 664                         cb->args[4] = 0;
 665                 }
 666                 inet_listen_unlock(hashinfo);
 667 skip_listen_ht:
 668                 cb->args[0] = 1;
 669                 s_i = num = s_num = 0;
 670         }
 671
 672         if (!(r->tcpdiag_states&~(TCPF_LISTEN|TCPF_SYN_RECV)))
 673                 return skb->len;
 674
 675         for (i = s_i; i < hashinfo->ehash_size; i++) {
 676                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 677                 struct sock *sk;
 678                 struct hlist_node *node;
 679
 680                 if (i > s_i)
 681                         s_num = 0;
 682
 683                 read_lock_bh(&head->lock);
 684
 685                 num = 0;
 686                 sk_for_each(sk, node, &head->chain) {
 687                         struct inet_sock *inet = inet_sk(sk);
 688
 689                         if (num < s_num)
 690                                 goto next_normal;
 691                         if (!(r->tcpdiag_states & (1 << sk->sk_state)))
 692                                 goto next_normal;
 693                         if (r->id.tcpdiag_sport != inet->sport &&
 694                             r->id.tcpdiag_sport)
 695                                 goto next_normal;
 696                         if (r->id.tcpdiag_dport != inet->dport && r->id.tcpdiag_dport)
 697                                 goto next_normal;
 698                         if (tcpdiag_dump_sock(skb, sk, cb) < 0) {
 699                                 read_unlock_bh(&head->lock);
 700                                 goto done;
 701                         }
 702 next_normal:
 703                         ++num;
 704                 }
 705
 706                 if (r->tcpdiag_states&TCPF_TIME_WAIT) {
 707                         sk_for_each(sk, node,
 708                                     &hashinfo->ehash[i + hashinfo->ehash_size].chain) {
 709                                 struct inet_sock *inet = inet_sk(sk);
 710
 711                                 if (num < s_num)
 712                                         goto next_dying;
 713                                 if (r->id.tcpdiag_sport != inet->sport &&
 714                                     r->id.tcpdiag_sport)
 715                                         goto next_dying;
 716                                 if (r->id.tcpdiag_dport != inet->dport &&
 717                                     r->id.tcpdiag_dport)
 718                                         goto next_dying;
 719                                 if (tcpdiag_dump_sock(skb, sk, cb) < 0) {
 720                                         read_unlock_bh(&head->lock);
 721                                         goto done;
 722                                 }
 723 next_dying:
 724                                 ++num;
 725                         }
 726                 }
 727                 read_unlock_bh(&head->lock);
 728         }
 729
 730 done:
 731         cb->args[1] = i;
 732         cb->args[2] = num;
 733         return skb->len;
 734 }
 735
 736 static int tcpdiag_dump_done(struct netlink_callback *cb)
 737 {
 738         return 0;
 739 }
 740
 741
 742 static __inline__ int
 743 tcpdiag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 744 {
 745         if (!(nlh->nlmsg_flags&NLM_F_REQUEST))
 746                 return 0;
 747
 748         if (nlh->nlmsg_type != TCPDIAG_GETSOCK
 749 #ifdef CONFIG_IP_TCPDIAG_DCCP
 750             && nlh->nlmsg_type != DCCPDIAG_GETSOCK
 751 #endif
 752            )
 753                 goto err_inval;
 754
 755         if (NLMSG_LENGTH(sizeof(struct tcpdiagreq)) > skb->len)
 756                 goto err_inval;
 757
 758         if (nlh->nlmsg_flags&NLM_F_DUMP) {
 759                 if (nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(struct tcpdiagreq))) {
 760                         struct rtattr *rta = (struct rtattr*)(NLMSG_DATA(nlh) + sizeof(struct tcpdiagreq));
 761                         if (rta->rta_type != TCPDIAG_REQ_BYTECODE ||
 762                             rta->rta_len < 8 ||
 763                             rta->rta_len > nlh->nlmsg_len - NLMSG_SPACE(sizeof(struct tcpdiagreq)))
 764                                 goto err_inval;
 765                         if (tcpdiag_bc_audit(RTA_DATA(rta), RTA_PAYLOAD(rta)))
 766                                 goto err_inval;
 767                 }
 768                 return netlink_dump_start(tcpnl, skb, nlh,
 769                                           tcpdiag_dump,
 770                                           tcpdiag_dump_done);
 771         } else {
 772                 return tcpdiag_get_exact(skb, nlh);
 773         }
 774
 775 err_inval:
 776         return -EINVAL;
 777 }
 778
 779
 780 static inline void tcpdiag_rcv_skb(struct sk_buff *skb)
 781 {
 782         int err;
 783         struct nlmsghdr * nlh;
 784
 785         if (skb->len >= NLMSG_SPACE(0)) {
 786                 nlh = (struct nlmsghdr *)skb->data;
 787                 if (nlh->nlmsg_len < sizeof(*nlh) || skb->len < nlh->nlmsg_len)
 788                         return;
 789                 err = tcpdiag_rcv_msg(skb, nlh);
 790                 if (err || nlh->nlmsg_flags & NLM_F_ACK)
 791                         netlink_ack(skb, nlh, err);
 792         }
 793 }
 794
 795 static void tcpdiag_rcv(struct sock *sk, int len)
 796 {
 797         struct sk_buff *skb;
 798         unsigned int qlen = skb_queue_len(&sk->sk_receive_queue);
 799
 800         while (qlen-- && (skb = skb_dequeue(&sk->sk_receive_queue))) {
 801                 tcpdiag_rcv_skb(skb);
 802                 kfree_skb(skb);
 803         }
 804 }
 805
 806 static int __init tcpdiag_init(void)
 807 {
 808         tcpnl = netlink_kernel_create(NETLINK_TCPDIAG, tcpdiag_rcv,
 809                                       THIS_MODULE);
 810         if (tcpnl == NULL)
 811                 return -ENOMEM;
 812         return 0;
 813 }
 814
 815 static void __exit tcpdiag_exit(void)
 816 {
 817         sock_release(tcpnl->sk_socket);
 818 }
 819
 820 module_init(tcpdiag_init);
 821 module_exit(tcpdiag_exit);
 822 MODULE_LICENSE("GPL");