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 /* Return information about state of tcp endpoint in API format. */
  45 void tcp_get_info(struct sock *sk, struct tcp_info *info)
  46 {
  47         struct tcp_opt *tp = tcp_sk(sk);
  48         u32 now = tcp_time_stamp;
  49
  50         memset(info, 0, sizeof(*info));
  51
  52         info->tcpi_state = sk->sk_state;
  53         info->tcpi_ca_state = tp->ca_state;
  54         info->tcpi_retransmits = tp->retransmits;
  55         info->tcpi_probes = tp->probes_out;
  56         info->tcpi_backoff = tp->backoff;
  57
  58         if (tp->tstamp_ok)
  59                 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
  60         if (tp->sack_ok)
  61                 info->tcpi_options |= TCPI_OPT_SACK;
  62         if (tp->wscale_ok) {
  63                 info->tcpi_options |= TCPI_OPT_WSCALE;
  64                 info->tcpi_snd_wscale = tp->snd_wscale;
  65                 info->tcpi_rcv_wscale = tp->rcv_wscale;
  66         }
  67
  68         if (tp->ecn_flags&TCP_ECN_OK)
  69                 info->tcpi_options |= TCPI_OPT_ECN;
  70
  71         info->tcpi_rto = (1000000*tp->rto)/HZ;
  72         info->tcpi_ato = (1000000*tp->ack.ato)/HZ;
  73         info->tcpi_snd_mss = tp->mss_cache;
  74         info->tcpi_rcv_mss = tp->ack.rcv_mss;
  75
  76         info->tcpi_unacked = tp->packets_out;
  77         info->tcpi_sacked = tp->sacked_out;
  78         info->tcpi_lost = tp->lost_out;
  79         info->tcpi_retrans = tp->retrans_out;
  80         info->tcpi_fackets = tp->fackets_out;
  81
  82         info->tcpi_last_data_sent = ((now - tp->lsndtime)*1000)/HZ;
  83         info->tcpi_last_data_recv = ((now - tp->ack.lrcvtime)*1000)/HZ;
  84         info->tcpi_last_ack_recv = ((now - tp->rcv_tstamp)*1000)/HZ;
  85
  86         info->tcpi_pmtu = tp->pmtu_cookie;
  87         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
  88         info->tcpi_rtt = ((1000000*tp->srtt)/HZ)>>3;
  89         info->tcpi_rttvar = ((1000000*tp->mdev)/HZ)>>2;
  90         info->tcpi_snd_ssthresh = tp->snd_ssthresh;
  91         info->tcpi_snd_cwnd = tp->snd_cwnd;
  92         info->tcpi_advmss = tp->advmss;
  93         info->tcpi_reordering = tp->reordering;
  94
  95         info->tcpi_rcv_rtt = ((1000000*tp->rcv_rtt_est.rtt)/HZ)>>3;
  96         info->tcpi_rcv_space = tp->rcvq_space.space;
  97 }
  98
  99 static int tcpdiag_fill(struct sk_buff *skb, struct sock *sk,
 100                         int ext, u32 pid, u32 seq)
 101 {
 102         struct inet_opt *inet = inet_sk(sk);
 103         struct tcp_opt *tp = tcp_sk(sk);
 104         struct tcpdiagmsg *r;
 105         struct nlmsghdr  *nlh;
 106         struct tcp_info  *info = NULL;
 107         struct tcpdiag_meminfo  *minfo = NULL;
 108         struct tcpvegas_info *vinfo = NULL;
 109         unsigned char    *b = skb->tail;
 110
 111         nlh = NLMSG_PUT(skb, pid, seq, TCPDIAG_GETSOCK, sizeof(*r));
 112         r = NLMSG_DATA(nlh);
 113         if (sk->sk_state != TCP_TIME_WAIT) {
 114                 if (ext & (1<<(TCPDIAG_MEMINFO-1)))
 115                         minfo = TCPDIAG_PUT(skb, TCPDIAG_MEMINFO, sizeof(*minfo));
 116                 if (ext & (1<<(TCPDIAG_INFO-1)))
 117                         info = TCPDIAG_PUT(skb, TCPDIAG_INFO, sizeof(*info));
 118
 119                 if (tcp_is_vegas(tp) && (ext & (1<<(TCPDIAG_VEGASINFO-1))))
 120                         vinfo = TCPDIAG_PUT(skb, TCPDIAG_VEGASINFO, sizeof(*vinfo));
 121         }
 122         r->tcpdiag_family = sk->sk_family;
 123         r->tcpdiag_state = sk->sk_state;
 124         r->tcpdiag_timer = 0;
 125         r->tcpdiag_retrans = 0;
 126
 127         r->id.tcpdiag_if = sk->sk_bound_dev_if;
 128         r->id.tcpdiag_cookie[0] = (u32)(unsigned long)sk;
 129         r->id.tcpdiag_cookie[1] = (u32)(((unsigned long)sk >> 31) >> 1);
 130
 131         if (r->tcpdiag_state == TCP_TIME_WAIT) {
 132                 struct tcp_tw_bucket *tw = (struct tcp_tw_bucket*)sk;
 133                 long tmo = tw->tw_ttd - jiffies;
 134                 if (tmo < 0)
 135                         tmo = 0;
 136
 137                 r->id.tcpdiag_sport = tw->tw_sport;
 138                 r->id.tcpdiag_dport = tw->tw_dport;
 139                 r->id.tcpdiag_src[0] = tw->tw_rcv_saddr;
 140                 r->id.tcpdiag_dst[0] = tw->tw_daddr;
 141                 r->tcpdiag_state = tw->tw_substate;
 142                 r->tcpdiag_timer = 3;
 143                 r->tcpdiag_expires = (tmo*1000+HZ-1)/HZ;
 144                 r->tcpdiag_rqueue = 0;
 145                 r->tcpdiag_wqueue = 0;
 146                 r->tcpdiag_uid = 0;
 147                 r->tcpdiag_inode = 0;
 148 #ifdef CONFIG_IPV6
 149                 if (r->tcpdiag_family == AF_INET6) {
 150                         ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
 151                                        &tw->tw_v6_rcv_saddr);
 152                         ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
 153                                        &tw->tw_v6_daddr);
 154                 }
 155 #endif
 156                 nlh->nlmsg_len = skb->tail - b;
 157                 return skb->len;
 158         }
 159
 160         r->id.tcpdiag_sport = inet->sport;
 161         r->id.tcpdiag_dport = inet->dport;
 162         r->id.tcpdiag_src[0] = inet->rcv_saddr;
 163         r->id.tcpdiag_dst[0] = inet->daddr;
 164
 165 #ifdef CONFIG_IPV6
 166         if (r->tcpdiag_family == AF_INET6) {
 167                 struct ipv6_pinfo *np = inet6_sk(sk);
 168
 169                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
 170                                &np->rcv_saddr);
 171                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
 172                                &np->daddr);
 173         }
 174 #endif
 175
 176 #define EXPIRES_IN_MS(tmo)  ((tmo-jiffies)*1000+HZ-1)/HZ
 177
 178         if (tp->pending == TCP_TIME_RETRANS) {
 179                 r->tcpdiag_timer = 1;
 180                 r->tcpdiag_retrans = tp->retransmits;
 181                 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
 182         } else if (tp->pending == TCP_TIME_PROBE0) {
 183                 r->tcpdiag_timer = 4;
 184                 r->tcpdiag_retrans = tp->probes_out;
 185                 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
 186         } else if (timer_pending(&sk->sk_timer)) {
 187                 r->tcpdiag_timer = 2;
 188                 r->tcpdiag_retrans = tp->probes_out;
 189                 r->tcpdiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 190         } else {
 191                 r->tcpdiag_timer = 0;
 192                 r->tcpdiag_expires = 0;
 193         }
 194 #undef EXPIRES_IN_MS
 195
 196         r->tcpdiag_rqueue = tp->rcv_nxt - tp->copied_seq;
 197         r->tcpdiag_wqueue = tp->write_seq - tp->snd_una;
 198         r->tcpdiag_uid = sock_i_uid(sk);
 199         r->tcpdiag_inode = sock_i_ino(sk);
 200
 201         if (minfo) {
 202                 minfo->tcpdiag_rmem = atomic_read(&sk->sk_rmem_alloc);
 203                 minfo->tcpdiag_wmem = sk->sk_wmem_queued;
 204                 minfo->tcpdiag_fmem = sk->sk_forward_alloc;
 205                 minfo->tcpdiag_tmem = atomic_read(&sk->sk_wmem_alloc);
 206         }
 207
 208         if (info)
 209                 tcp_get_info(sk, info);
 210
 211         if (vinfo) {
 212                 vinfo->tcpv_enabled = tp->vegas.doing_vegas_now;
 213                 vinfo->tcpv_rttcnt = tp->vegas.cntRTT;
 214                 vinfo->tcpv_rtt = tp->vegas.baseRTT;
 215                 vinfo->tcpv_minrtt = tp->vegas.minRTT;
 216         }
 217
 218         nlh->nlmsg_len = skb->tail - b;
 219         return skb->len;
 220
 221 nlmsg_failure:
 222         skb_trim(skb, b - skb->data);
 223         return -1;
 224 }
 225
 226 extern struct sock *tcp_v4_lookup(u32 saddr, u16 sport, u32 daddr, u16 dport, int dif);
 227 #ifdef CONFIG_IPV6
 228 extern struct sock *tcp_v6_lookup(struct in6_addr *saddr, u16 sport,
 229                                   struct in6_addr *daddr, u16 dport,
 230                                   int dif);
 231 #endif
 232
 233 static int tcpdiag_get_exact(struct sk_buff *in_skb, const struct nlmsghdr *nlh)
 234 {
 235         int err;
 236         struct sock *sk;
 237         struct tcpdiagreq *req = NLMSG_DATA(nlh);
 238         struct sk_buff *rep;
 239
 240         if (req->tcpdiag_family == AF_INET) {
 241                 sk = tcp_v4_lookup(req->id.tcpdiag_dst[0], req->id.tcpdiag_dport,
 242                                    req->id.tcpdiag_src[0], req->id.tcpdiag_sport,
 243                                    req->id.tcpdiag_if);
 244         }
 245 #ifdef CONFIG_IPV6
 246         else if (req->tcpdiag_family == AF_INET6) {
 247                 sk = tcp_v6_lookup((struct in6_addr*)req->id.tcpdiag_dst, req->id.tcpdiag_dport,
 248                                    (struct in6_addr*)req->id.tcpdiag_src, req->id.tcpdiag_sport,
 249                                    req->id.tcpdiag_if);
 250         }
 251 #endif
 252         else {
 253                 return -EINVAL;
 254         }
 255
 256         if (sk == NULL)
 257                 return -ENOENT;
 258
 259         err = -ESTALE;
 260         if ((req->id.tcpdiag_cookie[0] != TCPDIAG_NOCOOKIE ||
 261              req->id.tcpdiag_cookie[1] != TCPDIAG_NOCOOKIE) &&
 262             ((u32)(unsigned long)sk != req->id.tcpdiag_cookie[0] ||
 263              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.tcpdiag_cookie[1]))
 264                 goto out;
 265
 266         err = -ENOMEM;
 267         rep = alloc_skb(NLMSG_SPACE(sizeof(struct tcpdiagmsg)+
 268                                     sizeof(struct tcpdiag_meminfo)+
 269                                     sizeof(struct tcp_info)+64), GFP_KERNEL);
 270         if (!rep)
 271                 goto out;
 272
 273         if (tcpdiag_fill(rep, sk, req->tcpdiag_ext,
 274                          NETLINK_CB(in_skb).pid,
 275                          nlh->nlmsg_seq) <= 0)
 276                 BUG();
 277
 278         err = netlink_unicast(tcpnl, rep, NETLINK_CB(in_skb).pid, MSG_DONTWAIT);
 279         if (err > 0)
 280                 err = 0;
 281
 282 out:
 283         if (sk) {
 284                 if (sk->sk_state == TCP_TIME_WAIT)
 285                         tcp_tw_put((struct tcp_tw_bucket*)sk);
 286                 else
 287                         sock_put(sk);
 288         }
 289         return err;
 290 }
 291
 292 static int bitstring_match(const u32 *a1, const u32 *a2, int bits)
 293 {
 294         int words = bits >> 5;
 295
 296         bits &= 0x1f;
 297
 298         if (words) {
 299                 if (memcmp(a1, a2, words << 2))
 300                         return 0;
 301         }
 302         if (bits) {
 303                 __u32 w1, w2;
 304                 __u32 mask;
 305
 306                 w1 = a1[words];
 307                 w2 = a2[words];
 308
 309                 mask = htonl((0xffffffff) << (32 - bits));
 310
 311                 if ((w1 ^ w2) & mask)
 312                         return 0;
 313         }
 314
 315         return 1;
 316 }
 317
 318
 319 static int tcpdiag_bc_run(const void *bc, int len, struct sock *sk)
 320 {
 321         while (len > 0) {
 322                 int yes = 1;
 323                 struct inet_opt *inet = inet_sk(sk);
 324                 const struct tcpdiag_bc_op *op = bc;
 325
 326                 switch (op->code) {
 327                 case TCPDIAG_BC_NOP:
 328                         break;
 329                 case TCPDIAG_BC_JMP:
 330                         yes = 0;
 331                         break;
 332                 case TCPDIAG_BC_S_GE:
 333                         yes = inet->num >= op[1].no;
 334                         break;
 335                 case TCPDIAG_BC_S_LE:
 336                         yes = inet->num <= op[1].no;
 337                         break;
 338                 case TCPDIAG_BC_D_GE:
 339                         yes = ntohs(inet->dport) >= op[1].no;
 340                         break;
 341                 case TCPDIAG_BC_D_LE:
 342                         yes = ntohs(inet->dport) <= op[1].no;
 343                         break;
 344                 case TCPDIAG_BC_AUTO:
 345                         yes = !(sk->sk_userlocks & SOCK_BINDPORT_LOCK);
 346                         break;
 347                 case TCPDIAG_BC_S_COND:
 348                 case TCPDIAG_BC_D_COND:
 349                 {
 350                         struct tcpdiag_hostcond *cond = (struct tcpdiag_hostcond*)(op+1);
 351                         u32 *addr;
 352
 353                         if (cond->port != -1 &&
 354                             cond->port != (op->code == TCPDIAG_BC_S_COND ?
 355                                              inet->num : ntohs(inet->dport))) {
 356                                 yes = 0;
 357                                 break;
 358                         }
 359
 360                         if (cond->prefix_len == 0)
 361                                 break;
 362
 363 #ifdef CONFIG_IPV6
 364                         if (sk->sk_family == AF_INET6) {
 365                                 struct ipv6_pinfo *np = inet6_sk(sk);
 366
 367                                 if (op->code == TCPDIAG_BC_S_COND)
 368                                         addr = (u32*)&np->rcv_saddr;
 369                                 else
 370                                         addr = (u32*)&np->daddr;
 371                         } else
 372 #endif
 373                         {
 374                                 if (op->code == TCPDIAG_BC_S_COND)
 375                                         addr = &inet->rcv_saddr;
 376                                 else
 377                                         addr = &inet->daddr;
 378                         }
 379
 380                         if (bitstring_match(addr, cond->addr, cond->prefix_len))
 381                                 break;
 382                         if (sk->sk_family == AF_INET6 &&
 383                             cond->family == AF_INET) {
 384                                 if (addr[0] == 0 && addr[1] == 0 &&
 385                                     addr[2] == htonl(0xffff) &&
 386                                     bitstring_match(addr+3, cond->addr, cond->prefix_len))
 387                                         break;
 388                         }
 389                         yes = 0;
 390                         break;
 391                 }
 392                 }
 393
 394                 if (yes) {
 395                         len -= op->yes;
 396                         bc += op->yes;
 397                 } else {
 398                         len -= op->no;
 399                         bc += op->no;
 400                 }
 401         }
 402         return (len == 0);
 403 }
 404
 405 static int valid_cc(const void *bc, int len, int cc)
 406 {
 407         while (len >= 0) {
 408                 const struct tcpdiag_bc_op *op = bc;
 409
 410                 if (cc > len)
 411                         return 0;
 412                 if (cc == len)
 413                         return 1;
 414                 if (op->yes < 4)
 415                         return 0;
 416                 len -= op->yes;
 417                 bc  += op->yes;
 418         }
 419         return 0;
 420 }
 421
 422 static int tcpdiag_bc_audit(const void *bytecode, int bytecode_len)
 423 {
 424         const unsigned char *bc = bytecode;
 425         int  len = bytecode_len;
 426
 427         while (len > 0) {
 428                 struct tcpdiag_bc_op *op = (struct tcpdiag_bc_op*)bc;
 429
 430 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 431                 switch (op->code) {
 432                 case TCPDIAG_BC_AUTO:
 433                 case TCPDIAG_BC_S_COND:
 434                 case TCPDIAG_BC_D_COND:
 435                 case TCPDIAG_BC_S_GE:
 436                 case TCPDIAG_BC_S_LE:
 437                 case TCPDIAG_BC_D_GE:
 438                 case TCPDIAG_BC_D_LE:
 439                         if (op->yes < 4 || op->yes > len+4)
 440                                 return -EINVAL;
 441                 case TCPDIAG_BC_JMP:
 442                         if (op->no < 4 || op->no > len+4)
 443                                 return -EINVAL;
 444                         if (op->no < len &&
 445                             !valid_cc(bytecode, bytecode_len, len-op->no))
 446                                 return -EINVAL;
 447                         break;
 448                 case TCPDIAG_BC_NOP:
 449                         if (op->yes < 4 || op->yes > len+4)
 450                                 return -EINVAL;
 451                         break;
 452                 default:
 453                         return -EINVAL;
 454                 }
 455                 bc += op->yes;
 456                 len -= op->yes;
 457         }
 458         return len == 0 ? 0 : -EINVAL;
 459 }
 460
 461
 462 static int tcpdiag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 463 {
 464         int i, num;
 465         int s_i, s_num;
 466         struct tcpdiagreq *r = NLMSG_DATA(cb->nlh);
 467         struct rtattr *bc = NULL;
 468
 469         if (cb->nlh->nlmsg_len > 4+NLMSG_SPACE(sizeof(struct tcpdiagreq)))
 470                 bc = (struct rtattr*)(r+1);
 471
 472         s_i = cb->args[1];
 473         s_num = num = cb->args[2];
 474
 475         if (cb->args[0] == 0) {
 476                 if (!(r->tcpdiag_states&(TCPF_LISTEN|TCPF_SYN_RECV)))
 477                         goto skip_listen_ht;
 478                 tcp_listen_lock();
 479                 for (i = s_i; i < TCP_LHTABLE_SIZE; i++) {
 480                         struct sock *sk;
 481                         struct hlist_node *node;
 482
 483                         if (i > s_i)
 484                                 s_num = 0;
 485
 486                         num = 0;
 487                         sk_for_each(sk, node, &tcp_listening_hash[i]) {
 488                                 struct inet_opt *inet = inet_sk(sk);
 489                                 if (num < s_num)
 490                                         continue;
 491                                 if (!(r->tcpdiag_states&TCPF_LISTEN) ||
 492                                     r->id.tcpdiag_dport)
 493                                         continue;
 494                                 if (r->id.tcpdiag_sport != inet->sport &&
 495                                     r->id.tcpdiag_sport)
 496                                         continue;
 497                                 if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
 498                                         continue;
 499                                 if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
 500                                                  NETLINK_CB(cb->skb).pid,
 501                                                  cb->nlh->nlmsg_seq) <= 0) {
 502                                         tcp_listen_unlock();
 503                                         goto done;
 504                                 }
 505                                 ++num;
 506                         }
 507                 }
 508                 tcp_listen_unlock();
 509 skip_listen_ht:
 510                 cb->args[0] = 1;
 511                 s_i = num = s_num = 0;
 512         }
 513
 514         if (!(r->tcpdiag_states&~(TCPF_LISTEN|TCPF_SYN_RECV)))
 515                 return skb->len;
 516
 517         for (i = s_i; i < tcp_ehash_size; i++) {
 518                 struct tcp_ehash_bucket *head = &tcp_ehash[i];
 519                 struct sock *sk;
 520                 struct hlist_node *node;
 521
 522                 if (i > s_i)
 523                         s_num = 0;
 524
 525                 read_lock_bh(&head->lock);
 526
 527                 num = 0;
 528                 sk_for_each(sk, node, &head->chain) {
 529                         struct inet_opt *inet = inet_sk(sk);
 530
 531                         if (num < s_num)
 532                                 continue;
 533                         if (!(r->tcpdiag_states & (1 << sk->sk_state)))
 534                                 continue;
 535                         if (r->id.tcpdiag_sport != inet->sport &&
 536                             r->id.tcpdiag_sport)
 537                                 continue;
 538                         if (r->id.tcpdiag_dport != inet->dport && r->id.tcpdiag_dport)
 539                                 continue;
 540                         if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
 541                                 continue;
 542                         if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
 543                                          NETLINK_CB(cb->skb).pid,
 544                                          cb->nlh->nlmsg_seq) <= 0) {
 545                                 read_unlock_bh(&head->lock);
 546                                 goto done;
 547                         }
 548                         ++num;
 549                 }
 550
 551                 if (r->tcpdiag_states&TCPF_TIME_WAIT) {
 552                         sk_for_each(sk, node,
 553                                     &tcp_ehash[i + tcp_ehash_size].chain) {
 554                                 struct inet_opt *inet = inet_sk(sk);
 555
 556                                 if (num < s_num)
 557                                         continue;
 558                                 if (!(r->tcpdiag_states & (1 << sk->sk_zapped)))
 559                                         continue;
 560                                 if (r->id.tcpdiag_sport != inet->sport &&
 561                                     r->id.tcpdiag_sport)
 562                                         continue;
 563                                 if (r->id.tcpdiag_dport != inet->dport &&
 564                                     r->id.tcpdiag_dport)
 565                                         continue;
 566                                 if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
 567                                         continue;
 568                                 if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
 569                                                  NETLINK_CB(cb->skb).pid,
 570                                                  cb->nlh->nlmsg_seq) <= 0) {
 571                                         read_unlock_bh(&head->lock);
 572                                         goto done;
 573                                 }
 574                                 ++num;
 575                         }
 576                 }
 577                 read_unlock_bh(&head->lock);
 578         }
 579
 580 done:
 581         cb->args[1] = i;
 582         cb->args[2] = num;
 583         return skb->len;
 584 }
 585
 586 static int tcpdiag_dump_done(struct netlink_callback *cb)
 587 {
 588         return 0;
 589 }
 590
 591
 592 static __inline__ int
 593 tcpdiag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 594 {
 595         if (!(nlh->nlmsg_flags&NLM_F_REQUEST))
 596                 return 0;
 597
 598         if (nlh->nlmsg_type != TCPDIAG_GETSOCK)
 599                 goto err_inval;
 600
 601         if (NLMSG_LENGTH(sizeof(struct tcpdiagreq)) > skb->len)
 602                 goto err_inval;
 603
 604         if (nlh->nlmsg_flags&NLM_F_DUMP) {
 605                 if (nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(struct tcpdiagreq))) {
 606                         struct rtattr *rta = (struct rtattr*)(NLMSG_DATA(nlh) + sizeof(struct tcpdiagreq));
 607                         if (rta->rta_type != TCPDIAG_REQ_BYTECODE ||
 608                             rta->rta_len < 8 ||
 609                             rta->rta_len > nlh->nlmsg_len - NLMSG_SPACE(sizeof(struct tcpdiagreq)))
 610                                 goto err_inval;
 611                         if (tcpdiag_bc_audit(RTA_DATA(rta), RTA_PAYLOAD(rta)))
 612                                 goto err_inval;
 613                 }
 614                 return netlink_dump_start(tcpnl, skb, nlh,
 615                                           tcpdiag_dump,
 616                                           tcpdiag_dump_done);
 617         } else {
 618                 return tcpdiag_get_exact(skb, nlh);
 619         }
 620
 621 err_inval:
 622         return -EINVAL;
 623 }
 624
 625
 626 static inline void tcpdiag_rcv_skb(struct sk_buff *skb)
 627 {
 628         int err;
 629         struct nlmsghdr * nlh;
 630
 631         if (skb->len >= NLMSG_SPACE(0)) {
 632                 nlh = (struct nlmsghdr *)skb->data;
 633                 if (nlh->nlmsg_len < sizeof(*nlh) || skb->len < nlh->nlmsg_len)
 634                         return;
 635                 err = tcpdiag_rcv_msg(skb, nlh);
 636                 if (err || nlh->nlmsg_flags & NLM_F_ACK)
 637                         netlink_ack(skb, nlh, err);
 638         }
 639 }
 640
 641 static void tcpdiag_rcv(struct sock *sk, int len)
 642 {
 643         struct sk_buff *skb;
 644
 645         while ((skb = skb_dequeue(&sk->sk_receive_queue)) != NULL) {
 646                 tcpdiag_rcv_skb(skb);
 647                 kfree_skb(skb);
 648         }
 649 }
 650
 651 void __init tcpdiag_init(void)
 652 {
 653         tcpnl = netlink_kernel_create(NETLINK_TCPDIAG, tcpdiag_rcv);
 654         if (tcpnl == NULL)
 655                 panic("tcpdiag_init: Cannot create netlink socket.");
 656 }