tcp_diag.c

   1 /*
   2  * tcp_diag.c   Module for monitoring TCP sockets.
   3  *
   4  * Version:     $Id: tcp_diag.c,v 1.2 2001/11/05 09:42:22 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 tcp_opt *tp = &sk->tp_pinfo.af_tcp;
  48         struct tcpdiagmsg *r;
  49         struct nlmsghdr  *nlh;
  50         struct tcp_info  *info = NULL;
  51         struct tcpdiag_meminfo  *minfo = NULL;
  52         struct tcpvegas_info *vinfo = NULL;
  53         unsigned char    *b = skb->tail;
  54
  55         nlh = NLMSG_PUT(skb, pid, seq, TCPDIAG_GETSOCK, sizeof(*r));
  56         r = NLMSG_DATA(nlh);
  57         if (sk->state != TCP_TIME_WAIT) {
  58                 if (ext & (1<<(TCPDIAG_MEMINFO-1)))
  59                         minfo = TCPDIAG_PUT(skb, TCPDIAG_MEMINFO, sizeof(*minfo));
  60                 if (ext & (1<<(TCPDIAG_INFO-1)))
  61                         info = TCPDIAG_PUT(skb, TCPDIAG_INFO, sizeof(*info));
  62
  63                 if ((tcp_is_westwood(tp) || tcp_is_vegas(tp))
  64                     && (ext & (1<<(TCPDIAG_VEGASINFO-1))))
  65                         vinfo = TCPDIAG_PUT(skb, TCPDIAG_VEGASINFO, sizeof(*vinfo));
  66         }
  67         r->tcpdiag_family = sk->family;
  68         r->tcpdiag_state = sk->state;
  69         r->tcpdiag_timer = 0;
  70         r->tcpdiag_retrans = 0;
  71
  72         r->id.tcpdiag_sport = sk->sport;
  73         r->id.tcpdiag_dport = sk->dport;
  74         r->id.tcpdiag_src[0] = sk->rcv_saddr;
  75         r->id.tcpdiag_dst[0] = sk->daddr;
  76         r->id.tcpdiag_if = sk->bound_dev_if;
  77         r->id.tcpdiag_cookie[0] = (u32)(unsigned long)sk;
  78         r->id.tcpdiag_cookie[1] = (u32)(((unsigned long)sk >> 31) >> 1);
  79
  80         if (r->tcpdiag_state == TCP_TIME_WAIT) {
  81                 struct tcp_tw_bucket *tw = (struct tcp_tw_bucket*)sk;
  82                 long tmo = tw->ttd - jiffies;
  83                 if (tmo < 0)
  84                         tmo = 0;
  85
  86                 r->tcpdiag_state = 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                         memcpy(r->id.tcpdiag_src, &tw->v6_rcv_saddr, 16);
  96                         memcpy(r->id.tcpdiag_dst, &tw->v6_daddr, 16);
  97                 }
  98 #endif
  99                 nlh->nlmsg_len = skb->tail - b;
 100                 return skb->len;
 101         }
 102
 103 #ifdef CONFIG_IPV6
 104         if (r->tcpdiag_family == AF_INET6) {
 105                 memcpy(r->id.tcpdiag_src, &sk->net_pinfo.af_inet6.rcv_saddr, 16);
 106                 memcpy(r->id.tcpdiag_dst, &sk->net_pinfo.af_inet6.daddr, 16);
 107         }
 108 #endif
 109
 110 #define EXPIRES_IN_MS(tmo)  ((tmo-jiffies)*1000+HZ-1)/HZ
 111
 112         if (tp->pending == TCP_TIME_RETRANS) {
 113                 r->tcpdiag_timer = 1;
 114                 r->tcpdiag_retrans = tp->retransmits;
 115                 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
 116         } else if (tp->pending == TCP_TIME_PROBE0) {
 117                 r->tcpdiag_timer = 4;
 118                 r->tcpdiag_retrans = tp->probes_out;
 119                 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
 120         } else if (timer_pending(&sk->timer)) {
 121                 r->tcpdiag_timer = 2;
 122                 r->tcpdiag_retrans = tp->probes_out;
 123                 r->tcpdiag_expires = EXPIRES_IN_MS(sk->timer.expires);
 124         } else {
 125                 r->tcpdiag_timer = 0;
 126                 r->tcpdiag_expires = 0;
 127         }
 128 #undef EXPIRES_IN_MS
 129
 130         r->tcpdiag_rqueue = tp->rcv_nxt - tp->copied_seq;
 131         r->tcpdiag_wqueue = tp->write_seq - tp->snd_una;
 132         r->tcpdiag_uid = sock_i_uid(sk);
 133         r->tcpdiag_inode = sock_i_ino(sk);
 134
 135         if (minfo) {
 136                 minfo->tcpdiag_rmem = atomic_read(&sk->rmem_alloc);
 137                 minfo->tcpdiag_wmem = sk->wmem_queued;
 138                 minfo->tcpdiag_fmem = sk->forward_alloc;
 139                 minfo->tcpdiag_tmem = atomic_read(&sk->wmem_alloc);
 140         }
 141
 142         if (info) {
 143                 u32 now = tcp_time_stamp;
 144
 145                 info->tcpi_state = sk->state;
 146                 info->tcpi_ca_state = tp->ca_state;
 147                 info->tcpi_retransmits = tp->retransmits;
 148                 info->tcpi_probes = tp->probes_out;
 149                 info->tcpi_backoff = tp->backoff;
 150                 info->tcpi_options = 0;
 151                 if (tp->tstamp_ok)
 152                         info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
 153                 if (tp->sack_ok)
 154                         info->tcpi_options |= TCPI_OPT_SACK;
 155                 if (tp->wscale_ok) {
 156                         info->tcpi_options |= TCPI_OPT_WSCALE;
 157                         info->tcpi_snd_wscale = tp->snd_wscale;
 158                         info->tcpi_rcv_wscale = tp->rcv_wscale;
 159                 } else {
 160                         info->tcpi_snd_wscale = 0;
 161                         info->tcpi_rcv_wscale = 0;
 162                 }
 163 #ifdef CONFIG_INET_ECN
 164                 if (tp->ecn_flags&TCP_ECN_OK)
 165                         info->tcpi_options |= TCPI_OPT_ECN;
 166 #endif
 167
 168                 info->tcpi_rto = (1000000*tp->rto)/HZ;
 169                 info->tcpi_ato = (1000000*tp->ack.ato)/HZ;
 170                 info->tcpi_snd_mss = tp->mss_cache;
 171                 info->tcpi_rcv_mss = tp->ack.rcv_mss;
 172
 173                 info->tcpi_unacked = tp->packets_out;
 174                 info->tcpi_sacked = tp->sacked_out;
 175                 info->tcpi_lost = tp->lost_out;
 176                 info->tcpi_retrans = tp->retrans_out;
 177                 info->tcpi_fackets = tp->fackets_out;
 178
 179                 info->tcpi_last_data_sent = ((now - tp->lsndtime)*1000)/HZ;
 180                 info->tcpi_last_ack_sent = 0;
 181                 info->tcpi_last_data_recv = ((now - tp->ack.lrcvtime)*1000)/HZ;
 182                 info->tcpi_last_ack_recv = ((now - tp->rcv_tstamp)*1000)/HZ;
 183
 184                 info->tcpi_pmtu = tp->pmtu_cookie;
 185                 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
 186                 info->tcpi_rtt = ((1000000*tp->srtt)/HZ)>>3;
 187                 info->tcpi_rttvar = ((1000000*tp->mdev)/HZ)>>2;
 188                 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
 189                 info->tcpi_snd_cwnd = tp->snd_cwnd;
 190                 info->tcpi_advmss = tp->advmss;
 191                 info->tcpi_reordering = tp->reordering;
 192         }
 193
 194         if (vinfo) {
 195                 if (tcp_is_vegas(tp)) {
 196                         vinfo->tcpv_enabled = tp->vegas.doing_vegas_now;
 197                         vinfo->tcpv_rttcnt = tp->vegas.cntRTT;
 198                         vinfo->tcpv_rtt = (1000000*tp->vegas.baseRTT)/HZ;
 199                         vinfo->tcpv_minrtt = (1000000*tp->vegas.minRTT)/HZ;
 200                 } else {
 201                         vinfo->tcpv_enabled = 0;
 202                         vinfo->tcpv_rttcnt = 0;
 203                         vinfo->tcpv_rtt = (1000000*tp->westwood.rtt)/HZ;
 204                         vinfo->tcpv_minrtt = (1000000*tp->westwood.rtt_min)/HZ;
 205                 }
 206         }
 207
 208         nlh->nlmsg_len = skb->tail - b;
 209         return skb->len;
 210
 211 nlmsg_failure:
 212         skb_trim(skb, b - skb->data);
 213         return -1;
 214 }
 215
 216 extern struct sock *tcp_v4_lookup(u32 saddr, u16 sport, u32 daddr, u16 dport, int dif);
 217 #ifdef CONFIG_IPV6
 218 extern struct sock *tcp_v6_lookup(struct in6_addr *saddr, u16 sport,
 219                                   struct in6_addr *daddr, u16 dport,
 220                                   int dif);
 221 #endif
 222
 223 static int tcpdiag_get_exact(struct sk_buff *in_skb, struct nlmsghdr *nlh)
 224 {
 225         int err;
 226         struct sock *sk;
 227         struct tcpdiagreq *req = NLMSG_DATA(nlh);
 228         struct sk_buff *rep;
 229
 230         if (req->tcpdiag_family == AF_INET) {
 231                 sk = tcp_v4_lookup(req->id.tcpdiag_dst[0], req->id.tcpdiag_dport,
 232                                    req->id.tcpdiag_src[0], req->id.tcpdiag_sport,
 233                                    req->id.tcpdiag_if);
 234         }
 235 #ifdef CONFIG_IPV6
 236         else if (req->tcpdiag_family == AF_INET6) {
 237                 sk = tcp_v6_lookup((struct in6_addr*)req->id.tcpdiag_dst, req->id.tcpdiag_dport,
 238                                    (struct in6_addr*)req->id.tcpdiag_src, req->id.tcpdiag_sport,
 239                                    req->id.tcpdiag_if);
 240         }
 241 #endif
 242         else {
 243                 return -EINVAL;
 244         }
 245
 246         if (sk == NULL)
 247                 return -ENOENT;
 248
 249         err = -ESTALE;
 250         if ((req->id.tcpdiag_cookie[0] != TCPDIAG_NOCOOKIE ||
 251              req->id.tcpdiag_cookie[1] != TCPDIAG_NOCOOKIE) &&
 252             ((u32)(unsigned long)sk != req->id.tcpdiag_cookie[0] ||
 253              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.tcpdiag_cookie[1]))
 254                 goto out;
 255
 256         err = -ENOMEM;
 257         rep = alloc_skb(NLMSG_SPACE(sizeof(struct tcpdiagmsg)+
 258                                     sizeof(struct tcpdiag_meminfo)+
 259                                     sizeof(struct tcp_info)+64), GFP_KERNEL);
 260         if (!rep)
 261                 goto out;
 262
 263         if (tcpdiag_fill(rep, sk, req->tcpdiag_ext,
 264                          NETLINK_CB(in_skb).pid,
 265                          nlh->nlmsg_seq) <= 0)
 266                 BUG();
 267
 268         err = netlink_unicast(tcpnl, rep, NETLINK_CB(in_skb).pid, MSG_DONTWAIT);
 269         if (err > 0)
 270                 err = 0;
 271
 272 out:
 273         if (sk) {
 274                 if (sk->state == TCP_TIME_WAIT)
 275                         tcp_tw_put((struct tcp_tw_bucket*)sk);
 276                 else
 277                         sock_put(sk);
 278         }
 279         return err;
 280 }
 281
 282 int bitstring_match(u32 *a1, u32 *a2, int bits)
 283 {
 284         int words = bits >> 5;
 285
 286         bits &= 0x1f;
 287
 288         if (words) {
 289                 if (memcmp(a1, a2, words << 2))
 290                         return 0;
 291         }
 292         if (bits) {
 293                 __u32 w1, w2;
 294                 __u32 mask;
 295
 296                 w1 = a1[words];
 297                 w2 = a2[words];
 298
 299                 mask = htonl((0xffffffff) << (32 - bits));
 300
 301                 if ((w1 ^ w2) & mask)
 302                         return 0;
 303         }
 304
 305         return 1;
 306 }
 307
 308
 309 int tcpdiag_bc_run(char *bc, int len, struct sock *sk)
 310 {
 311         while (len > 0) {
 312                 int yes = 1;
 313                 struct tcpdiag_bc_op *op = (struct tcpdiag_bc_op*)bc;
 314
 315                 switch (op->code) {
 316                 case TCPDIAG_BC_NOP:
 317                         break;
 318                 case TCPDIAG_BC_JMP:
 319                         yes = 0;
 320                         break;
 321                 case TCPDIAG_BC_S_GE:
 322                         yes = (sk->num >= op[1].no);
 323                         break;
 324                 case TCPDIAG_BC_S_LE:
 325                         yes = (sk->num <= op[1].no);
 326                         break;
 327                 case TCPDIAG_BC_D_GE:
 328                         yes = (ntohs(sk->dport) >= op[1].no);
 329                         break;
 330                 case TCPDIAG_BC_D_LE:
 331                         yes = (ntohs(sk->dport) <= op[1].no);
 332                         break;
 333                 case TCPDIAG_BC_AUTO:
 334                         yes = !(sk->userlocks&SOCK_BINDPORT_LOCK);
 335                         break;
 336                 case TCPDIAG_BC_S_COND:
 337                 case TCPDIAG_BC_D_COND:
 338                 {
 339                         struct tcpdiag_hostcond *cond = (struct tcpdiag_hostcond*)(op+1);
 340                         u32 *addr;
 341
 342                         if (cond->port != -1 &&
 343                             cond->port != (op->code == TCPDIAG_BC_S_COND ? sk->num : ntohs(sk->dport))) {
 344                                 yes = 0;
 345                                 break;
 346                         }
 347
 348                         if (cond->prefix_len == 0)
 349                                 break;
 350
 351 #ifdef CONFIG_IPV6
 352                         if (sk->family == AF_INET6) {
 353                                 if (op->code == TCPDIAG_BC_S_COND)
 354                                         addr = (u32*)&sk->net_pinfo.af_inet6.rcv_saddr;
 355                                 else
 356                                         addr = (u32*)&sk->net_pinfo.af_inet6.daddr;
 357                         } else
 358 #endif
 359                         {
 360                                 if (op->code == TCPDIAG_BC_S_COND)
 361                                         addr = &sk->rcv_saddr;
 362                                 else
 363                                         addr = &sk->daddr;
 364                         }
 365
 366                         if (bitstring_match(addr, cond->addr, cond->prefix_len))
 367                                 break;
 368                         if (sk->family == AF_INET6 && cond->family == AF_INET) {
 369                                 if (addr[0] == 0 && addr[1] == 0 &&
 370                                     addr[2] == htonl(0xffff) &&
 371                                     bitstring_match(addr+3, cond->addr, cond->prefix_len))
 372                                         break;
 373                         }
 374                         yes = 0;
 375                         break;
 376                 }
 377                 }
 378
 379                 if (yes) {
 380                         len -= op->yes;
 381                         bc += op->yes;
 382                 } else {
 383                         len -= op->no;
 384                         bc += op->no;
 385                 }
 386         }
 387         return (len == 0);
 388 }
 389
 390 int valid_cc(char *bc, int len, int cc)
 391 {
 392         while (len >= 0) {
 393                 struct tcpdiag_bc_op *op = (struct tcpdiag_bc_op*)bc;
 394
 395                 if (cc > len)
 396                         return 0;
 397                 if (cc == len)
 398                         return 1;
 399                 if (op->yes < 4)
 400                         return 0;
 401                 len -= op->yes;
 402                 bc  += op->yes;
 403         }
 404         return 0;
 405 }
 406
 407 int tcpdiag_bc_audit(char *bytecode, int bytecode_len)
 408 {
 409         char *bc = bytecode;
 410         int  len = bytecode_len;
 411
 412         while (len > 0) {
 413                 struct tcpdiag_bc_op *op = (struct tcpdiag_bc_op*)bc;
 414
 415 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 416                 switch (op->code) {
 417                 case TCPDIAG_BC_AUTO:
 418                 case TCPDIAG_BC_S_COND:
 419                 case TCPDIAG_BC_D_COND:
 420                 case TCPDIAG_BC_S_GE:
 421                 case TCPDIAG_BC_S_LE:
 422                 case TCPDIAG_BC_D_GE:
 423                 case TCPDIAG_BC_D_LE:
 424                         if (op->yes < 4 || op->yes > len+4)
 425                                 return -EINVAL;
 426                 case TCPDIAG_BC_JMP:
 427                         if (op->no < 4 || op->no > len+4)
 428                                 return -EINVAL;
 429                         if (op->no < len &&
 430                             !valid_cc(bytecode, bytecode_len, len-op->no))
 431                                 return -EINVAL;
 432                         break;
 433                 case TCPDIAG_BC_NOP:
 434                         if (op->yes < 4 || op->yes > len+4)
 435                                 return -EINVAL;
 436                         break;
 437                 default:
 438                         return -EINVAL;
 439                 }
 440                 bc += op->yes;
 441                 len -= op->yes;
 442         }
 443         return len == 0 ? 0 : -EINVAL;
 444 }
 445
 446
 447 int tcpdiag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 448 {
 449         int i, num;
 450         int s_i, s_num;
 451         struct tcpdiagreq *r = NLMSG_DATA(cb->nlh);
 452         struct rtattr *bc = NULL;
 453
 454         if (cb->nlh->nlmsg_len > 4+NLMSG_SPACE(sizeof(struct tcpdiagreq)))
 455                 bc = (struct rtattr*)(r+1);
 456
 457         s_i = cb->args[1];
 458         s_num = num = cb->args[2];
 459
 460         if (cb->args[0] == 0) {
 461                 if (!(r->tcpdiag_states&(TCPF_LISTEN|TCPF_SYN_RECV)))
 462                         goto skip_listen_ht;
 463                 tcp_listen_lock();
 464                 for (i = s_i; i < TCP_LHTABLE_SIZE; i++) {
 465                         struct sock *sk = tcp_listening_hash[i];
 466
 467                         if (i > s_i)
 468                                 s_num = 0;
 469
 470                         for (sk = tcp_listening_hash[i], num = 0;
 471                              sk != NULL;
 472                              sk = sk->next, num++) {
 473                                 if (num < s_num)
 474                                         continue;
 475                                 if (!(r->tcpdiag_states&TCPF_LISTEN) ||
 476                                     r->id.tcpdiag_dport)
 477                                         continue;
 478                                 if (r->id.tcpdiag_sport != sk->sport && r->id.tcpdiag_sport)
 479                                         continue;
 480                                 if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
 481                                         continue;
 482                                 if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
 483                                                  NETLINK_CB(cb->skb).pid,
 484                                                  cb->nlh->nlmsg_seq) <= 0) {
 485                                         tcp_listen_unlock();
 486                                         goto done;
 487                                 }
 488                         }
 489                 }
 490                 tcp_listen_unlock();
 491 skip_listen_ht:
 492                 cb->args[0] = 1;
 493                 s_i = num = s_num = 0;
 494         }
 495
 496         if (!(r->tcpdiag_states&~(TCPF_LISTEN|TCPF_SYN_RECV)))
 497                 return skb->len;
 498
 499         for (i = s_i; i < tcp_ehash_size; i++) {
 500                 struct tcp_ehash_bucket *head = &tcp_ehash[i];
 501                 struct sock *sk;
 502
 503                 if (i > s_i)
 504                         s_num = 0;
 505
 506                 read_lock_bh(&head->lock);
 507
 508                 for (sk = head->chain, num = 0;
 509                      sk != NULL;
 510                      sk = sk->next, num++) {
 511                         if (num < s_num)
 512                                 continue;
 513                         if (!(r->tcpdiag_states&(1<<sk->state)))
 514                                 continue;
 515                         if (r->id.tcpdiag_sport != sk->sport && r->id.tcpdiag_sport)
 516                                 continue;
 517                         if (r->id.tcpdiag_dport != sk->dport && r->id.tcpdiag_dport)
 518                                 continue;
 519                         if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
 520                                 continue;
 521                         if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
 522                                          NETLINK_CB(cb->skb).pid,
 523                                          cb->nlh->nlmsg_seq) <= 0) {
 524                                 read_unlock_bh(&head->lock);
 525                                 goto done;
 526                         }
 527                 }
 528
 529                 if (r->tcpdiag_states&TCPF_TIME_WAIT) {
 530                         for (sk = tcp_ehash[i+tcp_ehash_size].chain;
 531                              sk != NULL;
 532                              sk = sk->next, num++) {
 533                                 if (num < s_num)
 534                                         continue;
 535                                 if (!(r->tcpdiag_states&(1<<sk->zapped)))
 536                                         continue;
 537                                 if (r->id.tcpdiag_sport != sk->sport && r->id.tcpdiag_sport)
 538                                         continue;
 539                                 if (r->id.tcpdiag_dport != sk->dport && r->id.tcpdiag_dport)
 540                                         continue;
 541                                 if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
 542                                         continue;
 543                                 if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
 544                                                  NETLINK_CB(cb->skb).pid,
 545                                                  cb->nlh->nlmsg_seq) <= 0) {
 546                                         read_unlock_bh(&head->lock);
 547                                         goto done;
 548                                 }
 549                         }
 550                 }
 551                 read_unlock_bh(&head->lock);
 552         }
 553
 554 done:
 555         cb->args[1] = i;
 556         cb->args[2] = num;
 557         return skb->len;
 558 }
 559
 560 static int tcpdiag_dump_done(struct netlink_callback *cb)
 561 {
 562         return 0;
 563 }
 564
 565
 566 static __inline__ int
 567 tcpdiag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 568 {
 569         if (!(nlh->nlmsg_flags&NLM_F_REQUEST))
 570                 return 0;
 571
 572         if (nlh->nlmsg_type != TCPDIAG_GETSOCK)
 573                 goto err_inval;
 574
 575         if (NLMSG_LENGTH(sizeof(struct tcpdiagreq)) > skb->len)
 576                 goto err_inval;
 577
 578         if (nlh->nlmsg_flags&NLM_F_DUMP) {
 579                 if (nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(struct tcpdiagreq))) {
 580                         struct rtattr *rta = (struct rtattr*)(NLMSG_DATA(nlh) + sizeof(struct tcpdiagreq));
 581                         if (rta->rta_type != TCPDIAG_REQ_BYTECODE ||
 582                             rta->rta_len < 8 ||
 583                             rta->rta_len > nlh->nlmsg_len - NLMSG_SPACE(sizeof(struct tcpdiagreq)))
 584                                 goto err_inval;
 585                         if (tcpdiag_bc_audit(RTA_DATA(rta), RTA_PAYLOAD(rta)))
 586                                 goto err_inval;
 587                 }
 588                 return netlink_dump_start(tcpnl, skb, nlh,
 589                                           tcpdiag_dump,
 590                                           tcpdiag_dump_done);
 591         } else {
 592                 return tcpdiag_get_exact(skb, nlh);
 593         }
 594
 595 err_inval:
 596         return -EINVAL;
 597 }
 598
 599
 600 static inline void tcpdiag_rcv_skb(struct sk_buff *skb)
 601 {
 602         int err;
 603         struct nlmsghdr * nlh;
 604
 605         if (skb->len >= NLMSG_SPACE(0)) {
 606                 nlh = (struct nlmsghdr *)skb->data;
 607                 if (nlh->nlmsg_len < sizeof(*nlh) || skb->len < nlh->nlmsg_len)
 608                         return;
 609                 err = tcpdiag_rcv_msg(skb, nlh);
 610                 if (err || nlh->nlmsg_flags & NLM_F_ACK)
 611                         netlink_ack(skb, nlh, err);
 612         }
 613 }
 614
 615 static void tcpdiag_rcv(struct sock *sk, int len)
 616 {
 617         struct sk_buff *skb;
 618
 619         while ((skb = skb_dequeue(&sk->receive_queue)) != NULL) {
 620                 tcpdiag_rcv_skb(skb);
 621                 kfree_skb(skb);
 622         }
 623 }
 624
 625 void __init tcpdiag_init(void)
 626 {
 627         tcpnl = netlink_kernel_create(NETLINK_TCPDIAG, tcpdiag_rcv);
 628         if (tcpnl == NULL)
 629                 panic("tcpdiag_init: Cannot create netlink socket.");
 630 }