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
  22 #include <net/icmp.h>
  23 #include <net/tcp.h>
  24 #include <net/ipv6.h>
  25 #include <net/inet_common.h>
  26
  27 #include <linux/inet.h>
  28 #include <linux/stddef.h>
  29
  30 #include <linux/tcp_diag.h>
  31
  32 static struct sock *tcpnl;
  33
  34
  35 #define TCPDIAG_PUT(skb, attrtype, attrlen) \
  36 ({ int rtalen = RTA_LENGTH(attrlen);        \
  37    struct rtattr *rta;                      \
  38    if (skb_tailroom(skb) < RTA_ALIGN(rtalen)) goto nlmsg_failure; \
  39    rta = (void*)__skb_put(skb, RTA_ALIGN(rtalen)); \
  40    rta->rta_type = attrtype;                \
  41    rta->rta_len = rtalen;                   \
  42    RTA_DATA(rta); })
  43
  44 static int tcpdiag_fill(struct sk_buff *skb, struct sock *sk,
  45                         int ext, u32 pid, u32 seq)
  46 {
  47         struct inet_opt *inet = inet_sk(sk);
  48         struct tcp_opt *tp = tcp_sk(sk);
  49         struct tcpdiagmsg *r;
  50         struct nlmsghdr  *nlh;
  51         struct tcp_info  *info = NULL;
  52         struct tcpdiag_meminfo  *minfo = NULL;
  53         struct tcpvegas_info *vinfo = NULL;
  54         unsigned char    *b = skb->tail;
  55
  56         nlh = NLMSG_PUT(skb, pid, seq, TCPDIAG_GETSOCK, sizeof(*r));
  57         r = NLMSG_DATA(nlh);
  58         if (sk->sk_state != TCP_TIME_WAIT) {
  59                 if (ext & (1<<(TCPDIAG_MEMINFO-1)))
  60                         minfo = TCPDIAG_PUT(skb, TCPDIAG_MEMINFO, sizeof(*minfo));
  61                 if (ext & (1<<(TCPDIAG_INFO-1)))
  62                         info = TCPDIAG_PUT(skb, TCPDIAG_INFO, sizeof(*info));
  63
  64                 if (tcp_is_vegas(tp) && (ext & (1<<(TCPDIAG_VEGASINFO-1))))
  65                         vinfo = TCPDIAG_PUT(skb, TCPDIAG_VEGASINFO, sizeof(*vinfo));
  66         }
  67         r->tcpdiag_family = sk->sk_family;
  68         r->tcpdiag_state = sk->sk_state;
  69         r->tcpdiag_timer = 0;
  70         r->tcpdiag_retrans = 0;
  71
  72         r->id.tcpdiag_if = sk->sk_bound_dev_if;
  73         r->id.tcpdiag_cookie[0] = (u32)(unsigned long)sk;
  74         r->id.tcpdiag_cookie[1] = (u32)(((unsigned long)sk >> 31) >> 1);
  75
  76         if (r->tcpdiag_state == TCP_TIME_WAIT) {
  77                 struct tcp_tw_bucket *tw = (struct tcp_tw_bucket*)sk;
  78                 long tmo = tw->tw_ttd - jiffies;
  79                 if (tmo < 0)
  80                         tmo = 0;
  81
  82                 r->id.tcpdiag_sport = tw->tw_sport;
  83                 r->id.tcpdiag_dport = tw->tw_dport;
  84                 r->id.tcpdiag_src[0] = tw->tw_rcv_saddr;
  85                 r->id.tcpdiag_dst[0] = tw->tw_daddr;
  86                 r->tcpdiag_state = tw->tw_substate;
  87                 r->tcpdiag_timer = 3;
  88                 r->tcpdiag_expires = (tmo*1000+HZ-1)/HZ;
  89                 r->tcpdiag_rqueue = 0;
  90                 r->tcpdiag_wqueue = 0;
  91                 r->tcpdiag_uid = 0;
  92                 r->tcpdiag_inode = 0;
  93 #ifdef CONFIG_IPV6
  94                 if (r->tcpdiag_family == AF_INET6) {
  95                         ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
  96                                        &tw->tw_v6_rcv_saddr);
  97                         ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
  98                                        &tw->tw_v6_daddr);
  99                 }
 100 #endif
 101                 nlh->nlmsg_len = skb->tail - b;
 102                 return skb->len;
 103         }
 104
 105         r->id.tcpdiag_sport = inet->sport;
 106         r->id.tcpdiag_dport = inet->dport;
 107         r->id.tcpdiag_src[0] = inet->rcv_saddr;
 108         r->id.tcpdiag_dst[0] = inet->daddr;
 109
 110 #ifdef CONFIG_IPV6
 111         if (r->tcpdiag_family == AF_INET6) {
 112                 struct ipv6_pinfo *np = inet6_sk(sk);
 113
 114                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
 115                                &np->rcv_saddr);
 116                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
 117                                &np->daddr);
 118         }
 119 #endif
 120
 121 #define EXPIRES_IN_MS(tmo)  ((tmo-jiffies)*1000+HZ-1)/HZ
 122
 123         if (tp->pending == TCP_TIME_RETRANS) {
 124                 r->tcpdiag_timer = 1;
 125                 r->tcpdiag_retrans = tp->retransmits;
 126                 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
 127         } else if (tp->pending == TCP_TIME_PROBE0) {
 128                 r->tcpdiag_timer = 4;
 129                 r->tcpdiag_retrans = tp->probes_out;
 130                 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
 131         } else if (timer_pending(&sk->sk_timer)) {
 132                 r->tcpdiag_timer = 2;
 133                 r->tcpdiag_retrans = tp->probes_out;
 134                 r->tcpdiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 135         } else {
 136                 r->tcpdiag_timer = 0;
 137                 r->tcpdiag_expires = 0;
 138         }
 139 #undef EXPIRES_IN_MS
 140
 141         r->tcpdiag_rqueue = tp->rcv_nxt - tp->copied_seq;
 142         r->tcpdiag_wqueue = tp->write_seq - tp->snd_una;
 143         r->tcpdiag_uid = sock_i_uid(sk);
 144         r->tcpdiag_inode = sock_i_ino(sk);
 145
 146         if (minfo) {
 147                 minfo->tcpdiag_rmem = atomic_read(&sk->sk_rmem_alloc);
 148                 minfo->tcpdiag_wmem = sk->sk_wmem_queued;
 149                 minfo->tcpdiag_fmem = sk->sk_forward_alloc;
 150                 minfo->tcpdiag_tmem = atomic_read(&sk->sk_wmem_alloc);
 151         }
 152
 153         if (info) {
 154                 u32 now = tcp_time_stamp;
 155
 156                 info->tcpi_state = sk->sk_state;
 157                 info->tcpi_ca_state = tp->ca_state;
 158                 info->tcpi_retransmits = tp->retransmits;
 159                 info->tcpi_probes = tp->probes_out;
 160                 info->tcpi_backoff = tp->backoff;
 161                 info->tcpi_options = 0;
 162                 if (tp->tstamp_ok)
 163                         info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
 164                 if (tp->sack_ok)
 165                         info->tcpi_options |= TCPI_OPT_SACK;
 166                 if (tp->wscale_ok) {
 167                         info->tcpi_options |= TCPI_OPT_WSCALE;
 168                         info->tcpi_snd_wscale = tp->snd_wscale;
 169                         info->tcpi_rcv_wscale = tp->rcv_wscale;
 170                 } else {
 171                         info->tcpi_snd_wscale = 0;
 172                         info->tcpi_rcv_wscale = 0;
 173                 }
 174                 if (tp->ecn_flags&TCP_ECN_OK)
 175                         info->tcpi_options |= TCPI_OPT_ECN;
 176
 177                 info->tcpi_rto = (1000000*tp->rto)/HZ;
 178                 info->tcpi_ato = (1000000*tp->ack.ato)/HZ;
 179                 info->tcpi_snd_mss = tp->mss_cache;
 180                 info->tcpi_rcv_mss = tp->ack.rcv_mss;
 181
 182                 info->tcpi_unacked = tp->packets_out;
 183                 info->tcpi_sacked = tp->sacked_out;
 184                 info->tcpi_lost = tp->lost_out;
 185                 info->tcpi_retrans = tp->retrans_out;
 186                 info->tcpi_fackets = tp->fackets_out;
 187
 188                 info->tcpi_last_data_sent = ((now - tp->lsndtime)*1000)/HZ;
 189                 info->tcpi_last_ack_sent = 0;
 190                 info->tcpi_last_data_recv = ((now - tp->ack.lrcvtime)*1000)/HZ;
 191                 info->tcpi_last_ack_recv = ((now - tp->rcv_tstamp)*1000)/HZ;
 192
 193                 info->tcpi_pmtu = tp->pmtu_cookie;
 194                 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
 195                 info->tcpi_rtt = ((1000000*tp->srtt)/HZ)>>3;
 196                 info->tcpi_rttvar = ((1000000*tp->mdev)/HZ)>>2;
 197                 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
 198                 info->tcpi_snd_cwnd = tp->snd_cwnd;
 199                 info->tcpi_advmss = tp->advmss;
 200                 info->tcpi_reordering = tp->reordering;
 201         }
 202
 203         if (vinfo) {
 204                 vinfo->tcpv_enabled = tp->vegas.doing_vegas_now;
 205                 vinfo->tcpv_rttcnt = tp->vegas.cntRTT;
 206                 vinfo->tcpv_rtt = tp->vegas.baseRTT;
 207                 vinfo->tcpv_minrtt = tp->vegas.minRTT;
 208         }
 209
 210         nlh->nlmsg_len = skb->tail - b;
 211         return skb->len;
 212
 213 nlmsg_failure:
 214         skb_trim(skb, b - skb->data);
 215         return -1;
 216 }
 217
 218 extern struct sock *tcp_v4_lookup(u32 saddr, u16 sport, u32 daddr, u16 dport, int dif);
 219 #ifdef CONFIG_IPV6
 220 extern struct sock *tcp_v6_lookup(struct in6_addr *saddr, u16 sport,
 221                                   struct in6_addr *daddr, u16 dport,
 222                                   int dif);
 223 #endif
 224
 225 static int tcpdiag_get_exact(struct sk_buff *in_skb, const struct nlmsghdr *nlh)
 226 {
 227         int err;
 228         struct sock *sk;
 229         struct tcpdiagreq *req = NLMSG_DATA(nlh);
 230         struct sk_buff *rep;
 231
 232         if (req->tcpdiag_family == AF_INET) {
 233                 sk = tcp_v4_lookup(req->id.tcpdiag_dst[0], req->id.tcpdiag_dport,
 234                                    req->id.tcpdiag_src[0], req->id.tcpdiag_sport,
 235                                    req->id.tcpdiag_if);
 236         }
 237 #ifdef CONFIG_IPV6
 238         else if (req->tcpdiag_family == AF_INET6) {
 239                 sk = tcp_v6_lookup((struct in6_addr*)req->id.tcpdiag_dst, req->id.tcpdiag_dport,
 240                                    (struct in6_addr*)req->id.tcpdiag_src, req->id.tcpdiag_sport,
 241                                    req->id.tcpdiag_if);
 242         }
 243 #endif
 244         else {
 245                 return -EINVAL;
 246         }
 247
 248         if (sk == NULL)
 249                 return -ENOENT;
 250
 251         err = -ESTALE;
 252         if ((req->id.tcpdiag_cookie[0] != TCPDIAG_NOCOOKIE ||
 253              req->id.tcpdiag_cookie[1] != TCPDIAG_NOCOOKIE) &&
 254             ((u32)(unsigned long)sk != req->id.tcpdiag_cookie[0] ||
 255              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.tcpdiag_cookie[1]))
 256                 goto out;
 257
 258         err = -ENOMEM;
 259         rep = alloc_skb(NLMSG_SPACE(sizeof(struct tcpdiagmsg)+
 260                                     sizeof(struct tcpdiag_meminfo)+
 261                                     sizeof(struct tcp_info)+64), GFP_KERNEL);
 262         if (!rep)
 263                 goto out;
 264
 265         if (tcpdiag_fill(rep, sk, req->tcpdiag_ext,
 266                          NETLINK_CB(in_skb).pid,
 267                          nlh->nlmsg_seq) <= 0)
 268                 BUG();
 269
 270         err = netlink_unicast(tcpnl, rep, NETLINK_CB(in_skb).pid, MSG_DONTWAIT);
 271         if (err > 0)
 272                 err = 0;
 273
 274 out:
 275         if (sk) {
 276                 if (sk->sk_state == TCP_TIME_WAIT)
 277                         tcp_tw_put((struct tcp_tw_bucket*)sk);
 278                 else
 279                         sock_put(sk);
 280         }
 281         return err;
 282 }
 283
 284 static int bitstring_match(const u32 *a1, const u32 *a2, int bits)
 285 {
 286         int words = bits >> 5;
 287
 288         bits &= 0x1f;
 289
 290         if (words) {
 291                 if (memcmp(a1, a2, words << 2))
 292                         return 0;
 293         }
 294         if (bits) {
 295                 __u32 w1, w2;
 296                 __u32 mask;
 297
 298                 w1 = a1[words];
 299                 w2 = a2[words];
 300
 301                 mask = htonl((0xffffffff) << (32 - bits));
 302
 303                 if ((w1 ^ w2) & mask)
 304                         return 0;
 305         }
 306
 307         return 1;
 308 }
 309
 310
 311 static int tcpdiag_bc_run(const void *bc, int len, struct sock *sk)
 312 {
 313         while (len > 0) {
 314                 int yes = 1;
 315                 struct inet_opt *inet = inet_sk(sk);
 316                 const struct tcpdiag_bc_op *op = bc;
 317
 318                 switch (op->code) {
 319                 case TCPDIAG_BC_NOP:
 320                         break;
 321                 case TCPDIAG_BC_JMP:
 322                         yes = 0;
 323                         break;
 324                 case TCPDIAG_BC_S_GE:
 325                         yes = inet->num >= op[1].no;
 326                         break;
 327                 case TCPDIAG_BC_S_LE:
 328                         yes = inet->num <= op[1].no;
 329                         break;
 330                 case TCPDIAG_BC_D_GE:
 331                         yes = ntohs(inet->dport) >= op[1].no;
 332                         break;
 333                 case TCPDIAG_BC_D_LE:
 334                         yes = ntohs(inet->dport) <= op[1].no;
 335                         break;
 336                 case TCPDIAG_BC_AUTO:
 337                         yes = !(sk->sk_userlocks & SOCK_BINDPORT_LOCK);
 338                         break;
 339                 case TCPDIAG_BC_S_COND:
 340                 case TCPDIAG_BC_D_COND:
 341                 {
 342                         struct tcpdiag_hostcond *cond = (struct tcpdiag_hostcond*)(op+1);
 343                         u32 *addr;
 344
 345                         if (cond->port != -1 &&
 346                             cond->port != (op->code == TCPDIAG_BC_S_COND ?
 347                                              inet->num : ntohs(inet->dport))) {
 348                                 yes = 0;
 349                                 break;
 350                         }
 351
 352                         if (cond->prefix_len == 0)
 353                                 break;
 354
 355 #ifdef CONFIG_IPV6
 356                         if (sk->sk_family == AF_INET6) {
 357                                 struct ipv6_pinfo *np = inet6_sk(sk);
 358
 359                                 if (op->code == TCPDIAG_BC_S_COND)
 360                                         addr = (u32*)&np->rcv_saddr;
 361                                 else
 362                                         addr = (u32*)&np->daddr;
 363                         } else
 364 #endif
 365                         {
 366                                 if (op->code == TCPDIAG_BC_S_COND)
 367                                         addr = &inet->rcv_saddr;
 368                                 else
 369                                         addr = &inet->daddr;
 370                         }
 371
 372                         if (bitstring_match(addr, cond->addr, cond->prefix_len))
 373                                 break;
 374                         if (sk->sk_family == AF_INET6 &&
 375                             cond->family == AF_INET) {
 376                                 if (addr[0] == 0 && addr[1] == 0 &&
 377                                     addr[2] == htonl(0xffff) &&
 378                                     bitstring_match(addr+3, cond->addr, cond->prefix_len))
 379                                         break;
 380                         }
 381                         yes = 0;
 382                         break;
 383                 }
 384                 }
 385
 386                 if (yes) {
 387                         len -= op->yes;
 388                         bc += op->yes;
 389                 } else {
 390                         len -= op->no;
 391                         bc += op->no;
 392                 }
 393         }
 394         return (len == 0);
 395 }
 396
 397 static int valid_cc(const void *bc, int len, int cc)
 398 {
 399         while (len >= 0) {
 400                 const struct tcpdiag_bc_op *op = bc;
 401
 402                 if (cc > len)
 403                         return 0;
 404                 if (cc == len)
 405                         return 1;
 406                 if (op->yes < 4)
 407                         return 0;
 408                 len -= op->yes;
 409                 bc  += op->yes;
 410         }
 411         return 0;
 412 }
 413
 414 static int tcpdiag_bc_audit(const void *bytecode, int bytecode_len)
 415 {
 416         const unsigned char *bc = bytecode;
 417         int  len = bytecode_len;
 418
 419         while (len > 0) {
 420                 struct tcpdiag_bc_op *op = (struct tcpdiag_bc_op*)bc;
 421
 422 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 423                 switch (op->code) {
 424                 case TCPDIAG_BC_AUTO:
 425                 case TCPDIAG_BC_S_COND:
 426                 case TCPDIAG_BC_D_COND:
 427                 case TCPDIAG_BC_S_GE:
 428                 case TCPDIAG_BC_S_LE:
 429                 case TCPDIAG_BC_D_GE:
 430                 case TCPDIAG_BC_D_LE:
 431                         if (op->yes < 4 || op->yes > len+4)
 432                                 return -EINVAL;
 433                 case TCPDIAG_BC_JMP:
 434                         if (op->no < 4 || op->no > len+4)
 435                                 return -EINVAL;
 436                         if (op->no < len &&
 437                             !valid_cc(bytecode, bytecode_len, len-op->no))
 438                                 return -EINVAL;
 439                         break;
 440                 case TCPDIAG_BC_NOP:
 441                         if (op->yes < 4 || op->yes > len+4)
 442                                 return -EINVAL;
 443                         break;
 444                 default:
 445                         return -EINVAL;
 446                 }
 447                 bc += op->yes;
 448                 len -= op->yes;
 449         }
 450         return len == 0 ? 0 : -EINVAL;
 451 }
 452
 453
 454 static int tcpdiag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 455 {
 456         int i, num;
 457         int s_i, s_num;
 458         struct tcpdiagreq *r = NLMSG_DATA(cb->nlh);
 459         struct rtattr *bc = NULL;
 460
 461         if (cb->nlh->nlmsg_len > 4+NLMSG_SPACE(sizeof(struct tcpdiagreq)))
 462                 bc = (struct rtattr*)(r+1);
 463
 464         s_i = cb->args[1];
 465         s_num = num = cb->args[2];
 466
 467         if (cb->args[0] == 0) {
 468                 if (!(r->tcpdiag_states&(TCPF_LISTEN|TCPF_SYN_RECV)))
 469                         goto skip_listen_ht;
 470                 tcp_listen_lock();
 471                 for (i = s_i; i < TCP_LHTABLE_SIZE; i++) {
 472                         struct sock *sk;
 473                         struct hlist_node *node;
 474
 475                         if (i > s_i)
 476                                 s_num = 0;
 477
 478                         num = 0;
 479                         sk_for_each(sk, node, &tcp_listening_hash[i]) {
 480                                 struct inet_opt *inet = inet_sk(sk);
 481                                 if (num < s_num)
 482                                         continue;
 483                                 if (!(r->tcpdiag_states&TCPF_LISTEN) ||
 484                                     r->id.tcpdiag_dport)
 485                                         continue;
 486                                 if (r->id.tcpdiag_sport != inet->sport &&
 487                                     r->id.tcpdiag_sport)
 488                                         continue;
 489                                 if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
 490                                         continue;
 491                                 if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
 492                                                  NETLINK_CB(cb->skb).pid,
 493                                                  cb->nlh->nlmsg_seq) <= 0) {
 494                                         tcp_listen_unlock();
 495                                         goto done;
 496                                 }
 497                                 ++num;
 498                         }
 499                 }
 500                 tcp_listen_unlock();
 501 skip_listen_ht:
 502                 cb->args[0] = 1;
 503                 s_i = num = s_num = 0;
 504         }
 505
 506         if (!(r->tcpdiag_states&~(TCPF_LISTEN|TCPF_SYN_RECV)))
 507                 return skb->len;
 508
 509         for (i = s_i; i < tcp_ehash_size; i++) {
 510                 struct tcp_ehash_bucket *head = &tcp_ehash[i];
 511                 struct sock *sk;
 512                 struct hlist_node *node;
 513
 514                 if (i > s_i)
 515                         s_num = 0;
 516
 517                 read_lock_bh(&head->lock);
 518
 519                 num = 0;
 520                 sk_for_each(sk, node, &head->chain) {
 521                         struct inet_opt *inet = inet_sk(sk);
 522
 523                         if (num < s_num)
 524                                 continue;
 525                         if (!(r->tcpdiag_states & (1 << sk->sk_state)))
 526                                 continue;
 527                         if (r->id.tcpdiag_sport != inet->sport &&
 528                             r->id.tcpdiag_sport)
 529                                 continue;
 530                         if (r->id.tcpdiag_dport != inet->dport && r->id.tcpdiag_dport)
 531                                 continue;
 532                         if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
 533                                 continue;
 534                         if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
 535                                          NETLINK_CB(cb->skb).pid,
 536                                          cb->nlh->nlmsg_seq) <= 0) {
 537                                 read_unlock_bh(&head->lock);
 538                                 goto done;
 539                         }
 540                         ++num;
 541                 }
 542
 543                 if (r->tcpdiag_states&TCPF_TIME_WAIT) {
 544                         sk_for_each(sk, node,
 545                                     &tcp_ehash[i + tcp_ehash_size].chain) {
 546                                 struct inet_opt *inet = inet_sk(sk);
 547
 548                                 if (num < s_num)
 549                                         continue;
 550                                 if (!(r->tcpdiag_states & (1 << sk->sk_zapped)))
 551                                         continue;
 552                                 if (r->id.tcpdiag_sport != inet->sport &&
 553                                     r->id.tcpdiag_sport)
 554                                         continue;
 555                                 if (r->id.tcpdiag_dport != inet->dport &&
 556                                     r->id.tcpdiag_dport)
 557                                         continue;
 558                                 if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
 559                                         continue;
 560                                 if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
 561                                                  NETLINK_CB(cb->skb).pid,
 562                                                  cb->nlh->nlmsg_seq) <= 0) {
 563                                         read_unlock_bh(&head->lock);
 564                                         goto done;
 565                                 }
 566                                 ++num;
 567                         }
 568                 }
 569                 read_unlock_bh(&head->lock);
 570         }
 571
 572 done:
 573         cb->args[1] = i;
 574         cb->args[2] = num;
 575         return skb->len;
 576 }
 577
 578 static int tcpdiag_dump_done(struct netlink_callback *cb)
 579 {
 580         return 0;
 581 }
 582
 583
 584 static __inline__ int
 585 tcpdiag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 586 {
 587         if (!(nlh->nlmsg_flags&NLM_F_REQUEST))
 588                 return 0;
 589
 590         if (nlh->nlmsg_type != TCPDIAG_GETSOCK)
 591                 goto err_inval;
 592
 593         if (NLMSG_LENGTH(sizeof(struct tcpdiagreq)) > skb->len)
 594                 goto err_inval;
 595
 596         if (nlh->nlmsg_flags&NLM_F_DUMP) {
 597                 if (nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(struct tcpdiagreq))) {
 598                         struct rtattr *rta = (struct rtattr*)(NLMSG_DATA(nlh) + sizeof(struct tcpdiagreq));
 599                         if (rta->rta_type != TCPDIAG_REQ_BYTECODE ||
 600                             rta->rta_len < 8 ||
 601                             rta->rta_len > nlh->nlmsg_len - NLMSG_SPACE(sizeof(struct tcpdiagreq)))
 602                                 goto err_inval;
 603                         if (tcpdiag_bc_audit(RTA_DATA(rta), RTA_PAYLOAD(rta)))
 604                                 goto err_inval;
 605                 }
 606                 return netlink_dump_start(tcpnl, skb, nlh,
 607                                           tcpdiag_dump,
 608                                           tcpdiag_dump_done);
 609         } else {
 610                 return tcpdiag_get_exact(skb, nlh);
 611         }
 612
 613 err_inval:
 614         return -EINVAL;
 615 }
 616
 617
 618 static inline void tcpdiag_rcv_skb(struct sk_buff *skb)
 619 {
 620         int err;
 621         struct nlmsghdr * nlh;
 622
 623         if (skb->len >= NLMSG_SPACE(0)) {
 624                 nlh = (struct nlmsghdr *)skb->data;
 625                 if (nlh->nlmsg_len < sizeof(*nlh) || skb->len < nlh->nlmsg_len)
 626                         return;
 627                 err = tcpdiag_rcv_msg(skb, nlh);
 628                 if (err || nlh->nlmsg_flags & NLM_F_ACK)
 629                         netlink_ack(skb, nlh, err);
 630         }
 631 }
 632
 633 static void tcpdiag_rcv(struct sock *sk, int len)
 634 {
 635         struct sk_buff *skb;
 636
 637         while ((skb = skb_dequeue(&sk->sk_receive_queue)) != NULL) {
 638                 tcpdiag_rcv_skb(skb);
 639                 kfree_skb(skb);
 640         }
 641 }
 642
 643 void __init tcpdiag_init(void)
 644 {
 645         tcpnl = netlink_kernel_create(NETLINK_TCPDIAG, tcpdiag_rcv);
 646         if (tcpnl == NULL)
 647                 panic("tcpdiag_init: Cannot create netlink socket.");
 648 }