SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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
  28 #include <linux/inet.h>
  29 #include <linux/stddef.h>
  30
  31 #include <linux/tcp_diag.h>
  32
  33 struct tcpdiag_entry
  34 {
  35         u32 *saddr;
  36         u32 *daddr;
  37         u16 sport;
  38         u16 dport;
  39         u16 family;
  40         u16 userlocks;
  41 };
  42
  43 static struct sock *tcpnl;
  44
  45 #define TCPDIAG_PUT(skb, attrtype, attrlen) \
  46         RTA_DATA(__RTA_PUT(skb, attrtype, attrlen))
  47
  48 static int tcpdiag_fill(struct sk_buff *skb, struct sock *sk,
  49                         int ext, u32 pid, u32 seq, u16 nlmsg_flags)
  50 {
  51         struct inet_sock *inet = inet_sk(sk);
  52         struct tcp_sock *tp = tcp_sk(sk);
  53         struct tcpdiagmsg *r;
  54         struct nlmsghdr  *nlh;
  55         struct tcp_info  *info = NULL;
  56         struct tcpdiag_meminfo  *minfo = NULL;
  57         unsigned char    *b = skb->tail;
  58
  59         nlh = NLMSG_PUT(skb, pid, seq, TCPDIAG_GETSOCK, sizeof(*r));
  60         nlh->nlmsg_flags = nlmsg_flags;
  61         r = NLMSG_DATA(nlh);
  62         if (sk->sk_state != TCP_TIME_WAIT) {
  63                 if (ext & (1<<(TCPDIAG_MEMINFO-1)))
  64                         minfo = TCPDIAG_PUT(skb, TCPDIAG_MEMINFO, sizeof(*minfo));
  65                 if (ext & (1<<(TCPDIAG_INFO-1)))
  66                         info = TCPDIAG_PUT(skb, TCPDIAG_INFO, sizeof(*info));
  67
  68                 if (ext & (1<<(TCPDIAG_CONG-1))) {
  69                         size_t len = strlen(tp->ca_ops->name);
  70                         strcpy(TCPDIAG_PUT(skb, TCPDIAG_CONG, len+1),
  71                                tp->ca_ops->name);
  72                 }
  73         }
  74         r->tcpdiag_family = sk->sk_family;
  75         r->tcpdiag_state = sk->sk_state;
  76         r->tcpdiag_timer = 0;
  77         r->tcpdiag_retrans = 0;
  78
  79         r->id.tcpdiag_if = sk->sk_bound_dev_if;
  80         r->id.tcpdiag_cookie[0] = (u32)(unsigned long)sk;
  81         r->id.tcpdiag_cookie[1] = (u32)(((unsigned long)sk >> 31) >> 1);
  82
  83         if (r->tcpdiag_state == TCP_TIME_WAIT) {
  84                 const struct inet_timewait_sock *tw = inet_twsk(sk);
  85                 long tmo = tw->tw_ttd - jiffies;
  86                 if (tmo < 0)
  87                         tmo = 0;
  88
  89                 r->id.tcpdiag_sport = tw->tw_sport;
  90                 r->id.tcpdiag_dport = tw->tw_dport;
  91                 r->id.tcpdiag_src[0] = tw->tw_rcv_saddr;
  92                 r->id.tcpdiag_dst[0] = tw->tw_daddr;
  93                 r->tcpdiag_state = tw->tw_substate;
  94                 r->tcpdiag_timer = 3;
  95                 r->tcpdiag_expires = (tmo*1000+HZ-1)/HZ;
  96                 r->tcpdiag_rqueue = 0;
  97                 r->tcpdiag_wqueue = 0;
  98                 r->tcpdiag_uid = 0;
  99                 r->tcpdiag_inode = 0;
 100 #ifdef CONFIG_IP_TCPDIAG_IPV6
 101                 if (r->tcpdiag_family == AF_INET6) {
 102                         const struct tcp6_timewait_sock *tcp6tw = tcp6_twsk(sk);
 103
 104                         ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
 105                                        &tcp6tw->tw_v6_rcv_saddr);
 106                         ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
 107                                        &tcp6tw->tw_v6_daddr);
 108                 }
 109 #endif
 110                 nlh->nlmsg_len = skb->tail - b;
 111                 return skb->len;
 112         }
 113
 114         r->id.tcpdiag_sport = inet->sport;
 115         r->id.tcpdiag_dport = inet->dport;
 116         r->id.tcpdiag_src[0] = inet->rcv_saddr;
 117         r->id.tcpdiag_dst[0] = inet->daddr;
 118
 119 #ifdef CONFIG_IP_TCPDIAG_IPV6
 120         if (r->tcpdiag_family == AF_INET6) {
 121                 struct ipv6_pinfo *np = inet6_sk(sk);
 122
 123                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
 124                                &np->rcv_saddr);
 125                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
 126                                &np->daddr);
 127         }
 128 #endif
 129
 130 #define EXPIRES_IN_MS(tmo)  ((tmo-jiffies)*1000+HZ-1)/HZ
 131
 132         if (tp->pending == TCP_TIME_RETRANS) {
 133                 r->tcpdiag_timer = 1;
 134                 r->tcpdiag_retrans = tp->retransmits;
 135                 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
 136         } else if (tp->pending == TCP_TIME_PROBE0) {
 137                 r->tcpdiag_timer = 4;
 138                 r->tcpdiag_retrans = tp->probes_out;
 139                 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
 140         } else if (timer_pending(&sk->sk_timer)) {
 141                 r->tcpdiag_timer = 2;
 142                 r->tcpdiag_retrans = tp->probes_out;
 143                 r->tcpdiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 144         } else {
 145                 r->tcpdiag_timer = 0;
 146                 r->tcpdiag_expires = 0;
 147         }
 148 #undef EXPIRES_IN_MS
 149
 150         r->tcpdiag_rqueue = tp->rcv_nxt - tp->copied_seq;
 151         r->tcpdiag_wqueue = tp->write_seq - tp->snd_una;
 152         r->tcpdiag_uid = sock_i_uid(sk);
 153         r->tcpdiag_inode = sock_i_ino(sk);
 154
 155         if (minfo) {
 156                 minfo->tcpdiag_rmem = atomic_read(&sk->sk_rmem_alloc);
 157                 minfo->tcpdiag_wmem = sk->sk_wmem_queued;
 158                 minfo->tcpdiag_fmem = sk->sk_forward_alloc;
 159                 minfo->tcpdiag_tmem = atomic_read(&sk->sk_wmem_alloc);
 160         }
 161
 162         if (info)
 163                 tcp_get_info(sk, info);
 164
 165         if (sk->sk_state < TCP_TIME_WAIT && tp->ca_ops->get_info)
 166                 tp->ca_ops->get_info(tp, ext, skb);
 167
 168         nlh->nlmsg_len = skb->tail - b;
 169         return skb->len;
 170
 171 rtattr_failure:
 172 nlmsg_failure:
 173         skb_trim(skb, b - skb->data);
 174         return -1;
 175 }
 176
 177 extern struct sock *tcp_v4_lookup(u32 saddr, u16 sport, u32 daddr, u16 dport,
 178                                   int dif);
 179 #ifdef CONFIG_IP_TCPDIAG_IPV6
 180 extern struct sock *tcp_v6_lookup(struct in6_addr *saddr, u16 sport,
 181                                   struct in6_addr *daddr, u16 dport,
 182                                   int dif);
 183 #else
 184 static inline struct sock *tcp_v6_lookup(struct in6_addr *saddr, u16 sport,
 185                                          struct in6_addr *daddr, u16 dport,
 186                                          int dif)
 187 {
 188         return NULL;
 189 }
 190 #endif
 191
 192 static int tcpdiag_get_exact(struct sk_buff *in_skb, const struct nlmsghdr *nlh)
 193 {
 194         int err;
 195         struct sock *sk;
 196         struct tcpdiagreq *req = NLMSG_DATA(nlh);
 197         struct sk_buff *rep;
 198
 199         if (req->tcpdiag_family == AF_INET) {
 200                 sk = tcp_v4_lookup(req->id.tcpdiag_dst[0], req->id.tcpdiag_dport,
 201                                    req->id.tcpdiag_src[0], req->id.tcpdiag_sport,
 202                                    req->id.tcpdiag_if);
 203         }
 204 #ifdef CONFIG_IP_TCPDIAG_IPV6
 205         else if (req->tcpdiag_family == AF_INET6) {
 206                 sk = tcp_v6_lookup((struct in6_addr*)req->id.tcpdiag_dst, req->id.tcpdiag_dport,
 207                                    (struct in6_addr*)req->id.tcpdiag_src, req->id.tcpdiag_sport,
 208                                    req->id.tcpdiag_if);
 209         }
 210 #endif
 211         else {
 212                 return -EINVAL;
 213         }
 214
 215         if (sk == NULL)
 216                 return -ENOENT;
 217
 218         err = -ESTALE;
 219         if ((req->id.tcpdiag_cookie[0] != TCPDIAG_NOCOOKIE ||
 220              req->id.tcpdiag_cookie[1] != TCPDIAG_NOCOOKIE) &&
 221             ((u32)(unsigned long)sk != req->id.tcpdiag_cookie[0] ||
 222              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.tcpdiag_cookie[1]))
 223                 goto out;
 224
 225         err = -ENOMEM;
 226         rep = alloc_skb(NLMSG_SPACE(sizeof(struct tcpdiagmsg)+
 227                                     sizeof(struct tcpdiag_meminfo)+
 228                                     sizeof(struct tcp_info)+64), GFP_KERNEL);
 229         if (!rep)
 230                 goto out;
 231
 232         if (tcpdiag_fill(rep, sk, req->tcpdiag_ext,
 233                          NETLINK_CB(in_skb).pid,
 234                          nlh->nlmsg_seq, 0) <= 0)
 235                 BUG();
 236
 237         err = netlink_unicast(tcpnl, rep, NETLINK_CB(in_skb).pid, MSG_DONTWAIT);
 238         if (err > 0)
 239                 err = 0;
 240
 241 out:
 242         if (sk) {
 243                 if (sk->sk_state == TCP_TIME_WAIT)
 244                         inet_twsk_put((struct inet_timewait_sock *)sk);
 245                 else
 246                         sock_put(sk);
 247         }
 248         return err;
 249 }
 250
 251 static int bitstring_match(const u32 *a1, const u32 *a2, int bits)
 252 {
 253         int words = bits >> 5;
 254
 255         bits &= 0x1f;
 256
 257         if (words) {
 258                 if (memcmp(a1, a2, words << 2))
 259                         return 0;
 260         }
 261         if (bits) {
 262                 __u32 w1, w2;
 263                 __u32 mask;
 264
 265                 w1 = a1[words];
 266                 w2 = a2[words];
 267
 268                 mask = htonl((0xffffffff) << (32 - bits));
 269
 270                 if ((w1 ^ w2) & mask)
 271                         return 0;
 272         }
 273
 274         return 1;
 275 }
 276
 277
 278 static int tcpdiag_bc_run(const void *bc, int len,
 279                           const struct tcpdiag_entry *entry)
 280 {
 281         while (len > 0) {
 282                 int yes = 1;
 283                 const struct tcpdiag_bc_op *op = bc;
 284
 285                 switch (op->code) {
 286                 case TCPDIAG_BC_NOP:
 287                         break;
 288                 case TCPDIAG_BC_JMP:
 289                         yes = 0;
 290                         break;
 291                 case TCPDIAG_BC_S_GE:
 292                         yes = entry->sport >= op[1].no;
 293                         break;
 294                 case TCPDIAG_BC_S_LE:
 295                         yes = entry->dport <= op[1].no;
 296                         break;
 297                 case TCPDIAG_BC_D_GE:
 298                         yes = entry->dport >= op[1].no;
 299                         break;
 300                 case TCPDIAG_BC_D_LE:
 301                         yes = entry->dport <= op[1].no;
 302                         break;
 303                 case TCPDIAG_BC_AUTO:
 304                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 305                         break;
 306                 case TCPDIAG_BC_S_COND:
 307                 case TCPDIAG_BC_D_COND:
 308                 {
 309                         struct tcpdiag_hostcond *cond = (struct tcpdiag_hostcond*)(op+1);
 310                         u32 *addr;
 311
 312                         if (cond->port != -1 &&
 313                             cond->port != (op->code == TCPDIAG_BC_S_COND ?
 314                                              entry->sport : entry->dport)) {
 315                                 yes = 0;
 316                                 break;
 317                         }
 318
 319                         if (cond->prefix_len == 0)
 320                                 break;
 321
 322                         if (op->code == TCPDIAG_BC_S_COND)
 323                                 addr = entry->saddr;
 324                         else
 325                                 addr = entry->daddr;
 326
 327                         if (bitstring_match(addr, cond->addr, cond->prefix_len))
 328                                 break;
 329                         if (entry->family == AF_INET6 &&
 330                             cond->family == AF_INET) {
 331                                 if (addr[0] == 0 && addr[1] == 0 &&
 332                                     addr[2] == htonl(0xffff) &&
 333                                     bitstring_match(addr+3, cond->addr, cond->prefix_len))
 334                                         break;
 335                         }
 336                         yes = 0;
 337                         break;
 338                 }
 339                 }
 340
 341                 if (yes) {
 342                         len -= op->yes;
 343                         bc += op->yes;
 344                 } else {
 345                         len -= op->no;
 346                         bc += op->no;
 347                 }
 348         }
 349         return (len == 0);
 350 }
 351
 352 static int valid_cc(const void *bc, int len, int cc)
 353 {
 354         while (len >= 0) {
 355                 const struct tcpdiag_bc_op *op = bc;
 356
 357                 if (cc > len)
 358                         return 0;
 359                 if (cc == len)
 360                         return 1;
 361                 if (op->yes < 4)
 362                         return 0;
 363                 len -= op->yes;
 364                 bc  += op->yes;
 365         }
 366         return 0;
 367 }
 368
 369 static int tcpdiag_bc_audit(const void *bytecode, int bytecode_len)
 370 {
 371         const unsigned char *bc = bytecode;
 372         int  len = bytecode_len;
 373
 374         while (len > 0) {
 375                 struct tcpdiag_bc_op *op = (struct tcpdiag_bc_op*)bc;
 376
 377 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 378                 switch (op->code) {
 379                 case TCPDIAG_BC_AUTO:
 380                 case TCPDIAG_BC_S_COND:
 381                 case TCPDIAG_BC_D_COND:
 382                 case TCPDIAG_BC_S_GE:
 383                 case TCPDIAG_BC_S_LE:
 384                 case TCPDIAG_BC_D_GE:
 385                 case TCPDIAG_BC_D_LE:
 386                         if (op->yes < 4 || op->yes > len+4)
 387                                 return -EINVAL;
 388                 case TCPDIAG_BC_JMP:
 389                         if (op->no < 4 || op->no > len+4)
 390                                 return -EINVAL;
 391                         if (op->no < len &&
 392                             !valid_cc(bytecode, bytecode_len, len-op->no))
 393                                 return -EINVAL;
 394                         break;
 395                 case TCPDIAG_BC_NOP:
 396                         if (op->yes < 4 || op->yes > len+4)
 397                                 return -EINVAL;
 398                         break;
 399                 default:
 400                         return -EINVAL;
 401                 }
 402                 bc += op->yes;
 403                 len -= op->yes;
 404         }
 405         return len == 0 ? 0 : -EINVAL;
 406 }
 407
 408 static int tcpdiag_dump_sock(struct sk_buff *skb, struct sock *sk,
 409                              struct netlink_callback *cb)
 410 {
 411         struct tcpdiagreq *r = NLMSG_DATA(cb->nlh);
 412
 413         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 414                 struct tcpdiag_entry entry;
 415                 struct rtattr *bc = (struct rtattr *)(r + 1);
 416                 struct inet_sock *inet = inet_sk(sk);
 417
 418                 entry.family = sk->sk_family;
 419 #ifdef CONFIG_IP_TCPDIAG_IPV6
 420                 if (entry.family == AF_INET6) {
 421                         struct ipv6_pinfo *np = inet6_sk(sk);
 422
 423                         entry.saddr = np->rcv_saddr.s6_addr32;
 424                         entry.daddr = np->daddr.s6_addr32;
 425                 } else
 426 #endif
 427                 {
 428                         entry.saddr = &inet->rcv_saddr;
 429                         entry.daddr = &inet->daddr;
 430                 }
 431                 entry.sport = inet->num;
 432                 entry.dport = ntohs(inet->dport);
 433                 entry.userlocks = sk->sk_userlocks;
 434
 435                 if (!tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), &entry))
 436                         return 0;
 437         }
 438
 439         return tcpdiag_fill(skb, sk, r->tcpdiag_ext, NETLINK_CB(cb->skb).pid,
 440                             cb->nlh->nlmsg_seq, NLM_F_MULTI);
 441 }
 442
 443 static int tcpdiag_fill_req(struct sk_buff *skb, struct sock *sk,
 444                             struct request_sock *req,
 445                             u32 pid, u32 seq)
 446 {
 447         const struct inet_request_sock *ireq = inet_rsk(req);
 448         struct inet_sock *inet = inet_sk(sk);
 449         unsigned char *b = skb->tail;
 450         struct tcpdiagmsg *r;
 451         struct nlmsghdr *nlh;
 452         long tmo;
 453
 454         nlh = NLMSG_PUT(skb, pid, seq, TCPDIAG_GETSOCK, sizeof(*r));
 455         nlh->nlmsg_flags = NLM_F_MULTI;
 456         r = NLMSG_DATA(nlh);
 457
 458         r->tcpdiag_family = sk->sk_family;
 459         r->tcpdiag_state = TCP_SYN_RECV;
 460         r->tcpdiag_timer = 1;
 461         r->tcpdiag_retrans = req->retrans;
 462
 463         r->id.tcpdiag_if = sk->sk_bound_dev_if;
 464         r->id.tcpdiag_cookie[0] = (u32)(unsigned long)req;
 465         r->id.tcpdiag_cookie[1] = (u32)(((unsigned long)req >> 31) >> 1);
 466
 467         tmo = req->expires - jiffies;
 468         if (tmo < 0)
 469                 tmo = 0;
 470
 471         r->id.tcpdiag_sport = inet->sport;
 472         r->id.tcpdiag_dport = ireq->rmt_port;
 473         r->id.tcpdiag_src[0] = ireq->loc_addr;
 474         r->id.tcpdiag_dst[0] = ireq->rmt_addr;
 475         r->tcpdiag_expires = jiffies_to_msecs(tmo),
 476         r->tcpdiag_rqueue = 0;
 477         r->tcpdiag_wqueue = 0;
 478         r->tcpdiag_uid = sock_i_uid(sk);
 479         r->tcpdiag_inode = 0;
 480 #ifdef CONFIG_IP_TCPDIAG_IPV6
 481         if (r->tcpdiag_family == AF_INET6) {
 482                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
 483                                &tcp6_rsk(req)->loc_addr);
 484                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
 485                                &tcp6_rsk(req)->rmt_addr);
 486         }
 487 #endif
 488         nlh->nlmsg_len = skb->tail - b;
 489
 490         return skb->len;
 491
 492 nlmsg_failure:
 493         skb_trim(skb, b - skb->data);
 494         return -1;
 495 }
 496
 497 static int tcpdiag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 498                              struct netlink_callback *cb)
 499 {
 500         struct tcpdiag_entry entry;
 501         struct tcpdiagreq *r = NLMSG_DATA(cb->nlh);
 502         struct tcp_sock *tp = tcp_sk(sk);
 503         struct listen_sock *lopt;
 504         struct rtattr *bc = NULL;
 505         struct inet_sock *inet = inet_sk(sk);
 506         int j, s_j;
 507         int reqnum, s_reqnum;
 508         int err = 0;
 509
 510         s_j = cb->args[3];
 511         s_reqnum = cb->args[4];
 512
 513         if (s_j > 0)
 514                 s_j--;
 515
 516         entry.family = sk->sk_family;
 517
 518         read_lock_bh(&tp->accept_queue.syn_wait_lock);
 519
 520         lopt = tp->accept_queue.listen_opt;
 521         if (!lopt || !lopt->qlen)
 522                 goto out;
 523
 524         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 525                 bc = (struct rtattr *)(r + 1);
 526                 entry.sport = inet->num;
 527                 entry.userlocks = sk->sk_userlocks;
 528         }
 529
 530         for (j = s_j; j < TCP_SYNQ_HSIZE; j++) {
 531                 struct request_sock *req, *head = lopt->syn_table[j];
 532
 533                 reqnum = 0;
 534                 for (req = head; req; reqnum++, req = req->dl_next) {
 535                         struct inet_request_sock *ireq = inet_rsk(req);
 536
 537                         if (reqnum < s_reqnum)
 538                                 continue;
 539                         if (r->id.tcpdiag_dport != ireq->rmt_port &&
 540                             r->id.tcpdiag_dport)
 541                                 continue;
 542
 543                         if (bc) {
 544                                 entry.saddr =
 545 #ifdef CONFIG_IP_TCPDIAG_IPV6
 546                                         (entry.family == AF_INET6) ?
 547                                         tcp6_rsk(req)->loc_addr.s6_addr32 :
 548 #endif
 549                                         &ireq->loc_addr;
 550                                 entry.daddr =
 551 #ifdef CONFIG_IP_TCPDIAG_IPV6
 552                                         (entry.family == AF_INET6) ?
 553                                         tcp6_rsk(req)->rmt_addr.s6_addr32 :
 554 #endif
 555                                         &ireq->rmt_addr;
 556                                 entry.dport = ntohs(ireq->rmt_port);
 557
 558                                 if (!tcpdiag_bc_run(RTA_DATA(bc),
 559                                                     RTA_PAYLOAD(bc), &entry))
 560                                         continue;
 561                         }
 562
 563                         err = tcpdiag_fill_req(skb, sk, req,
 564                                                NETLINK_CB(cb->skb).pid,
 565                                                cb->nlh->nlmsg_seq);
 566                         if (err < 0) {
 567                                 cb->args[3] = j + 1;
 568                                 cb->args[4] = reqnum;
 569                                 goto out;
 570                         }
 571                 }
 572
 573                 s_reqnum = 0;
 574         }
 575
 576 out:
 577         read_unlock_bh(&tp->accept_queue.syn_wait_lock);
 578
 579         return err;
 580 }
 581
 582 static int tcpdiag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 583 {
 584         int i, num;
 585         int s_i, s_num;
 586         struct tcpdiagreq *r = NLMSG_DATA(cb->nlh);
 587
 588         s_i = cb->args[1];
 589         s_num = num = cb->args[2];
 590
 591         if (cb->args[0] == 0) {
 592                 if (!(r->tcpdiag_states&(TCPF_LISTEN|TCPF_SYN_RECV)))
 593                         goto skip_listen_ht;
 594                 inet_listen_lock(&tcp_hashinfo);
 595                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 596                         struct sock *sk;
 597                         struct hlist_node *node;
 598
 599                         num = 0;
 600                         sk_for_each(sk, node, &tcp_hashinfo.listening_hash[i]) {
 601                                 struct inet_sock *inet = inet_sk(sk);
 602
 603                                 if (num < s_num) {
 604                                         num++;
 605                                         continue;
 606                                 }
 607
 608                                 if (r->id.tcpdiag_sport != inet->sport &&
 609                                     r->id.tcpdiag_sport)
 610                                         goto next_listen;
 611
 612                                 if (!(r->tcpdiag_states&TCPF_LISTEN) ||
 613                                     r->id.tcpdiag_dport ||
 614                                     cb->args[3] > 0)
 615                                         goto syn_recv;
 616
 617                                 if (tcpdiag_dump_sock(skb, sk, cb) < 0) {
 618                                         inet_listen_unlock(&tcp_hashinfo);
 619                                         goto done;
 620                                 }
 621
 622 syn_recv:
 623                                 if (!(r->tcpdiag_states&TCPF_SYN_RECV))
 624                                         goto next_listen;
 625
 626                                 if (tcpdiag_dump_reqs(skb, sk, cb) < 0) {
 627                                         inet_listen_unlock(&tcp_hashinfo);
 628                                         goto done;
 629                                 }
 630
 631 next_listen:
 632                                 cb->args[3] = 0;
 633                                 cb->args[4] = 0;
 634                                 ++num;
 635                         }
 636
 637                         s_num = 0;
 638                         cb->args[3] = 0;
 639                         cb->args[4] = 0;
 640                 }
 641                 inet_listen_unlock(&tcp_hashinfo);
 642 skip_listen_ht:
 643                 cb->args[0] = 1;
 644                 s_i = num = s_num = 0;
 645         }
 646
 647         if (!(r->tcpdiag_states&~(TCPF_LISTEN|TCPF_SYN_RECV)))
 648                 return skb->len;
 649
 650         for (i = s_i; i < tcp_hashinfo.ehash_size; i++) {
 651                 struct inet_ehash_bucket *head = &tcp_hashinfo.ehash[i];
 652                 struct sock *sk;
 653                 struct hlist_node *node;
 654
 655                 if (i > s_i)
 656                         s_num = 0;
 657
 658                 read_lock_bh(&head->lock);
 659
 660                 num = 0;
 661                 sk_for_each(sk, node, &head->chain) {
 662                         struct inet_sock *inet = inet_sk(sk);
 663
 664                         if (num < s_num)
 665                                 goto next_normal;
 666                         if (!(r->tcpdiag_states & (1 << sk->sk_state)))
 667                                 goto next_normal;
 668                         if (r->id.tcpdiag_sport != inet->sport &&
 669                             r->id.tcpdiag_sport)
 670                                 goto next_normal;
 671                         if (r->id.tcpdiag_dport != inet->dport && r->id.tcpdiag_dport)
 672                                 goto next_normal;
 673                         if (tcpdiag_dump_sock(skb, sk, cb) < 0) {
 674                                 read_unlock_bh(&head->lock);
 675                                 goto done;
 676                         }
 677 next_normal:
 678                         ++num;
 679                 }
 680
 681                 if (r->tcpdiag_states&TCPF_TIME_WAIT) {
 682                         sk_for_each(sk, node,
 683                                     &tcp_hashinfo.ehash[i + tcp_hashinfo.ehash_size].chain) {
 684                                 struct inet_sock *inet = inet_sk(sk);
 685
 686                                 if (num < s_num)
 687                                         goto next_dying;
 688                                 if (r->id.tcpdiag_sport != inet->sport &&
 689                                     r->id.tcpdiag_sport)
 690                                         goto next_dying;
 691                                 if (r->id.tcpdiag_dport != inet->dport &&
 692                                     r->id.tcpdiag_dport)
 693                                         goto next_dying;
 694                                 if (tcpdiag_dump_sock(skb, sk, cb) < 0) {
 695                                         read_unlock_bh(&head->lock);
 696                                         goto done;
 697                                 }
 698 next_dying:
 699                                 ++num;
 700                         }
 701                 }
 702                 read_unlock_bh(&head->lock);
 703         }
 704
 705 done:
 706         cb->args[1] = i;
 707         cb->args[2] = num;
 708         return skb->len;
 709 }
 710
 711 static int tcpdiag_dump_done(struct netlink_callback *cb)
 712 {
 713         return 0;
 714 }
 715
 716
 717 static __inline__ int
 718 tcpdiag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 719 {
 720         if (!(nlh->nlmsg_flags&NLM_F_REQUEST))
 721                 return 0;
 722
 723         if (nlh->nlmsg_type != TCPDIAG_GETSOCK)
 724                 goto err_inval;
 725
 726         if (NLMSG_LENGTH(sizeof(struct tcpdiagreq)) > skb->len)
 727                 goto err_inval;
 728
 729         if (nlh->nlmsg_flags&NLM_F_DUMP) {
 730                 if (nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(struct tcpdiagreq))) {
 731                         struct rtattr *rta = (struct rtattr*)(NLMSG_DATA(nlh) + sizeof(struct tcpdiagreq));
 732                         if (rta->rta_type != TCPDIAG_REQ_BYTECODE ||
 733                             rta->rta_len < 8 ||
 734                             rta->rta_len > nlh->nlmsg_len - NLMSG_SPACE(sizeof(struct tcpdiagreq)))
 735                                 goto err_inval;
 736                         if (tcpdiag_bc_audit(RTA_DATA(rta), RTA_PAYLOAD(rta)))
 737                                 goto err_inval;
 738                 }
 739                 return netlink_dump_start(tcpnl, skb, nlh,
 740                                           tcpdiag_dump,
 741                                           tcpdiag_dump_done);
 742         } else {
 743                 return tcpdiag_get_exact(skb, nlh);
 744         }
 745
 746 err_inval:
 747         return -EINVAL;
 748 }
 749
 750
 751 static inline void tcpdiag_rcv_skb(struct sk_buff *skb)
 752 {
 753         int err;
 754         struct nlmsghdr * nlh;
 755
 756         if (skb->len >= NLMSG_SPACE(0)) {
 757                 nlh = (struct nlmsghdr *)skb->data;
 758                 if (nlh->nlmsg_len < sizeof(*nlh) || skb->len < nlh->nlmsg_len)
 759                         return;
 760                 err = tcpdiag_rcv_msg(skb, nlh);
 761                 if (err || nlh->nlmsg_flags & NLM_F_ACK)
 762                         netlink_ack(skb, nlh, err);
 763         }
 764 }
 765
 766 static void tcpdiag_rcv(struct sock *sk, int len)
 767 {
 768         struct sk_buff *skb;
 769         unsigned int qlen = skb_queue_len(&sk->sk_receive_queue);
 770
 771         while (qlen-- && (skb = skb_dequeue(&sk->sk_receive_queue))) {
 772                 tcpdiag_rcv_skb(skb);
 773                 kfree_skb(skb);
 774         }
 775 }
 776
 777 static int __init tcpdiag_init(void)
 778 {
 779         tcpnl = netlink_kernel_create(NETLINK_TCPDIAG, tcpdiag_rcv,
 780                                       THIS_MODULE);
 781         if (tcpnl == NULL)
 782                 return -ENOMEM;
 783         return 0;
 784 }
 785
 786 static void __exit tcpdiag_exit(void)
 787 {
 788         sock_release(tcpnl->sk_socket);
 789 }
 790
 791 module_init(tcpdiag_init);
 792 module_exit(tcpdiag_exit);
 793 MODULE_LICENSE("GPL");