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 = jiffies_to_usecs(tp->rto);
  72         info->tcpi_ato = jiffies_to_usecs(tp->ack.ato);
  73         info->tcpi_snd_mss = tp->mss_cache_std;
  74         info->tcpi_rcv_mss = tp->ack.rcv_mss;
  75
  76         info->tcpi_unacked = tcp_get_pcount(&tp->packets_out);
  77         info->tcpi_sacked = tcp_get_pcount(&tp->sacked_out);
  78         info->tcpi_lost = tcp_get_pcount(&tp->lost_out);
  79         info->tcpi_retrans = tcp_get_pcount(&tp->retrans_out);
  80         info->tcpi_fackets = tcp_get_pcount(&tp->fackets_out);
  81
  82         info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
  83         info->tcpi_last_data_recv = jiffies_to_msecs(now - tp->ack.lrcvtime);
  84         info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
  85
  86         info->tcpi_pmtu = tp->pmtu_cookie;
  87         info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
  88         info->tcpi_rtt = jiffies_to_usecs(tp->srtt)>>3;
  89         info->tcpi_rttvar = jiffies_to_usecs(tp->mdev)>>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 = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>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_westwood(tp) || tcp_is_vegas(tp))
 120                     && (ext & (1<<(TCPDIAG_VEGASINFO-1))))
 121                         vinfo = TCPDIAG_PUT(skb, TCPDIAG_VEGASINFO, sizeof(*vinfo));
 122         }
 123         r->tcpdiag_family = sk->sk_family;
 124         r->tcpdiag_state = sk->sk_state;
 125         r->tcpdiag_timer = 0;
 126         r->tcpdiag_retrans = 0;
 127
 128         r->id.tcpdiag_if = sk->sk_bound_dev_if;
 129         r->id.tcpdiag_cookie[0] = (u32)(unsigned long)sk;
 130         r->id.tcpdiag_cookie[1] = (u32)(((unsigned long)sk >> 31) >> 1);
 131
 132         if (r->tcpdiag_state == TCP_TIME_WAIT) {
 133                 struct tcp_tw_bucket *tw = (struct tcp_tw_bucket*)sk;
 134                 long tmo = tw->tw_ttd - jiffies;
 135                 if (tmo < 0)
 136                         tmo = 0;
 137
 138                 r->id.tcpdiag_sport = tw->tw_sport;
 139                 r->id.tcpdiag_dport = tw->tw_dport;
 140                 r->id.tcpdiag_src[0] = tw->tw_rcv_saddr;
 141                 r->id.tcpdiag_dst[0] = tw->tw_daddr;
 142                 r->tcpdiag_state = tw->tw_substate;
 143                 r->tcpdiag_timer = 3;
 144                 r->tcpdiag_expires = (tmo*1000+HZ-1)/HZ;
 145                 r->tcpdiag_rqueue = 0;
 146                 r->tcpdiag_wqueue = 0;
 147                 r->tcpdiag_uid = 0;
 148                 r->tcpdiag_inode = 0;
 149 #ifdef CONFIG_IPV6
 150                 if (r->tcpdiag_family == AF_INET6) {
 151                         ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
 152                                        &tw->tw_v6_rcv_saddr);
 153                         ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
 154                                        &tw->tw_v6_daddr);
 155                 }
 156 #endif
 157                 nlh->nlmsg_len = skb->tail - b;
 158                 return skb->len;
 159         }
 160
 161         r->id.tcpdiag_sport = inet->sport;
 162         r->id.tcpdiag_dport = inet->dport;
 163         r->id.tcpdiag_src[0] = inet->rcv_saddr;
 164         r->id.tcpdiag_dst[0] = inet->daddr;
 165
 166 #ifdef CONFIG_IPV6
 167         if (r->tcpdiag_family == AF_INET6) {
 168                 struct ipv6_pinfo *np = inet6_sk(sk);
 169
 170                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
 171                                &np->rcv_saddr);
 172                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
 173                                &np->daddr);
 174         }
 175 #endif
 176
 177 #define EXPIRES_IN_MS(tmo)  ((tmo-jiffies)*1000+HZ-1)/HZ
 178
 179         if (tp->pending == TCP_TIME_RETRANS) {
 180                 r->tcpdiag_timer = 1;
 181                 r->tcpdiag_retrans = tp->retransmits;
 182                 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
 183         } else if (tp->pending == TCP_TIME_PROBE0) {
 184                 r->tcpdiag_timer = 4;
 185                 r->tcpdiag_retrans = tp->probes_out;
 186                 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
 187         } else if (timer_pending(&sk->sk_timer)) {
 188                 r->tcpdiag_timer = 2;
 189                 r->tcpdiag_retrans = tp->probes_out;
 190                 r->tcpdiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 191         } else {
 192                 r->tcpdiag_timer = 0;
 193                 r->tcpdiag_expires = 0;
 194         }
 195 #undef EXPIRES_IN_MS
 196
 197         r->tcpdiag_rqueue = tp->rcv_nxt - tp->copied_seq;
 198         r->tcpdiag_wqueue = tp->write_seq - tp->snd_una;
 199         r->tcpdiag_uid = sock_i_uid(sk);
 200         r->tcpdiag_inode = sock_i_ino(sk);
 201
 202         if (minfo) {
 203                 minfo->tcpdiag_rmem = atomic_read(&sk->sk_rmem_alloc);
 204                 minfo->tcpdiag_wmem = sk->sk_wmem_queued;
 205                 minfo->tcpdiag_fmem = sk->sk_forward_alloc;
 206                 minfo->tcpdiag_tmem = atomic_read(&sk->sk_wmem_alloc);
 207         }
 208
 209         if (info)
 210                 tcp_get_info(sk, info);
 211
 212         if (vinfo) {
 213                 if (tcp_is_vegas(tp)) {
 214                         vinfo->tcpv_enabled = tp->vegas.doing_vegas_now;
 215                         vinfo->tcpv_rttcnt = tp->vegas.cntRTT;
 216                         vinfo->tcpv_rtt = jiffies_to_usecs(tp->vegas.baseRTT);
 217                         vinfo->tcpv_minrtt = jiffies_to_usecs(tp->vegas.minRTT);
 218                 } else {
 219                         vinfo->tcpv_enabled = 0;
 220                         vinfo->tcpv_rttcnt = 0;
 221                         vinfo->tcpv_rtt = jiffies_to_usecs(tp->westwood.rtt);
 222                         vinfo->tcpv_minrtt = jiffies_to_usecs(tp->westwood.rtt_min);
 223                 }
 224         }
 225
 226         nlh->nlmsg_len = skb->tail - b;
 227         return skb->len;
 228
 229 nlmsg_failure:
 230         skb_trim(skb, b - skb->data);
 231         return -1;
 232 }
 233
 234 extern struct sock *tcp_v4_lookup(u32 saddr, u16 sport, u32 daddr, u16 dport, int dif);
 235 #ifdef CONFIG_IPV6
 236 extern struct sock *tcp_v6_lookup(struct in6_addr *saddr, u16 sport,
 237                                   struct in6_addr *daddr, u16 dport,
 238                                   int dif);
 239 #endif
 240
 241 static int tcpdiag_get_exact(struct sk_buff *in_skb, const struct nlmsghdr *nlh)
 242 {
 243         int err;
 244         struct sock *sk;
 245         struct tcpdiagreq *req = NLMSG_DATA(nlh);
 246         struct sk_buff *rep;
 247
 248         if (req->tcpdiag_family == AF_INET) {
 249                 sk = tcp_v4_lookup(req->id.tcpdiag_dst[0], req->id.tcpdiag_dport,
 250                                    req->id.tcpdiag_src[0], req->id.tcpdiag_sport,
 251                                    req->id.tcpdiag_if);
 252         }
 253 #ifdef CONFIG_IPV6
 254         else if (req->tcpdiag_family == AF_INET6) {
 255                 sk = tcp_v6_lookup((struct in6_addr*)req->id.tcpdiag_dst, req->id.tcpdiag_dport,
 256                                    (struct in6_addr*)req->id.tcpdiag_src, req->id.tcpdiag_sport,
 257                                    req->id.tcpdiag_if);
 258         }
 259 #endif
 260         else {
 261                 return -EINVAL;
 262         }
 263
 264         if (sk == NULL)
 265                 return -ENOENT;
 266
 267         err = -ESTALE;
 268         if ((req->id.tcpdiag_cookie[0] != TCPDIAG_NOCOOKIE ||
 269              req->id.tcpdiag_cookie[1] != TCPDIAG_NOCOOKIE) &&
 270             ((u32)(unsigned long)sk != req->id.tcpdiag_cookie[0] ||
 271              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.tcpdiag_cookie[1]))
 272                 goto out;
 273
 274         err = -ENOMEM;
 275         rep = alloc_skb(NLMSG_SPACE(sizeof(struct tcpdiagmsg)+
 276                                     sizeof(struct tcpdiag_meminfo)+
 277                                     sizeof(struct tcp_info)+64), GFP_KERNEL);
 278         if (!rep)
 279                 goto out;
 280
 281         if (tcpdiag_fill(rep, sk, req->tcpdiag_ext,
 282                          NETLINK_CB(in_skb).pid,
 283                          nlh->nlmsg_seq) <= 0)
 284                 BUG();
 285
 286         err = netlink_unicast(tcpnl, rep, NETLINK_CB(in_skb).pid, MSG_DONTWAIT);
 287         if (err > 0)
 288                 err = 0;
 289
 290 out:
 291         if (sk) {
 292                 if (sk->sk_state == TCP_TIME_WAIT)
 293                         tcp_tw_put((struct tcp_tw_bucket*)sk);
 294                 else
 295                         sock_put(sk);
 296         }
 297         return err;
 298 }
 299
 300 static int bitstring_match(const u32 *a1, const u32 *a2, int bits)
 301 {
 302         int words = bits >> 5;
 303
 304         bits &= 0x1f;
 305
 306         if (words) {
 307                 if (memcmp(a1, a2, words << 2))
 308                         return 0;
 309         }
 310         if (bits) {
 311                 __u32 w1, w2;
 312                 __u32 mask;
 313
 314                 w1 = a1[words];
 315                 w2 = a2[words];
 316
 317                 mask = htonl((0xffffffff) << (32 - bits));
 318
 319                 if ((w1 ^ w2) & mask)
 320                         return 0;
 321         }
 322
 323         return 1;
 324 }
 325
 326
 327 static int tcpdiag_bc_run(const void *bc, int len, struct sock *sk)
 328 {
 329         while (len > 0) {
 330                 int yes = 1;
 331                 struct inet_opt *inet = inet_sk(sk);
 332                 const struct tcpdiag_bc_op *op = bc;
 333
 334                 switch (op->code) {
 335                 case TCPDIAG_BC_NOP:
 336                         break;
 337                 case TCPDIAG_BC_JMP:
 338                         yes = 0;
 339                         break;
 340                 case TCPDIAG_BC_S_GE:
 341                         yes = inet->num >= op[1].no;
 342                         break;
 343                 case TCPDIAG_BC_S_LE:
 344                         yes = inet->num <= op[1].no;
 345                         break;
 346                 case TCPDIAG_BC_D_GE:
 347                         yes = ntohs(inet->dport) >= op[1].no;
 348                         break;
 349                 case TCPDIAG_BC_D_LE:
 350                         yes = ntohs(inet->dport) <= op[1].no;
 351                         break;
 352                 case TCPDIAG_BC_AUTO:
 353                         yes = !(sk->sk_userlocks & SOCK_BINDPORT_LOCK);
 354                         break;
 355                 case TCPDIAG_BC_S_COND:
 356                 case TCPDIAG_BC_D_COND:
 357                 {
 358                         struct tcpdiag_hostcond *cond = (struct tcpdiag_hostcond*)(op+1);
 359                         u32 *addr;
 360
 361                         if (cond->port != -1 &&
 362                             cond->port != (op->code == TCPDIAG_BC_S_COND ?
 363                                              inet->num : ntohs(inet->dport))) {
 364                                 yes = 0;
 365                                 break;
 366                         }
 367
 368                         if (cond->prefix_len == 0)
 369                                 break;
 370
 371 #ifdef CONFIG_IPV6
 372                         if (sk->sk_family == AF_INET6) {
 373                                 struct ipv6_pinfo *np = inet6_sk(sk);
 374
 375                                 if (op->code == TCPDIAG_BC_S_COND)
 376                                         addr = (u32*)&np->rcv_saddr;
 377                                 else
 378                                         addr = (u32*)&np->daddr;
 379                         } else
 380 #endif
 381                         {
 382                                 if (op->code == TCPDIAG_BC_S_COND)
 383                                         addr = &inet->rcv_saddr;
 384                                 else
 385                                         addr = &inet->daddr;
 386                         }
 387
 388                         if (bitstring_match(addr, cond->addr, cond->prefix_len))
 389                                 break;
 390                         if (sk->sk_family == AF_INET6 &&
 391                             cond->family == AF_INET) {
 392                                 if (addr[0] == 0 && addr[1] == 0 &&
 393                                     addr[2] == htonl(0xffff) &&
 394                                     bitstring_match(addr+3, cond->addr, cond->prefix_len))
 395                                         break;
 396                         }
 397                         yes = 0;
 398                         break;
 399                 }
 400                 }
 401
 402                 if (yes) {
 403                         len -= op->yes;
 404                         bc += op->yes;
 405                 } else {
 406                         len -= op->no;
 407                         bc += op->no;
 408                 }
 409         }
 410         return (len == 0);
 411 }
 412
 413 static int valid_cc(const void *bc, int len, int cc)
 414 {
 415         while (len >= 0) {
 416                 const struct tcpdiag_bc_op *op = bc;
 417
 418                 if (cc > len)
 419                         return 0;
 420                 if (cc == len)
 421                         return 1;
 422                 if (op->yes < 4)
 423                         return 0;
 424                 len -= op->yes;
 425                 bc  += op->yes;
 426         }
 427         return 0;
 428 }
 429
 430 static int tcpdiag_bc_audit(const void *bytecode, int bytecode_len)
 431 {
 432         const unsigned char *bc = bytecode;
 433         int  len = bytecode_len;
 434
 435         while (len > 0) {
 436                 struct tcpdiag_bc_op *op = (struct tcpdiag_bc_op*)bc;
 437
 438 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 439                 switch (op->code) {
 440                 case TCPDIAG_BC_AUTO:
 441                 case TCPDIAG_BC_S_COND:
 442                 case TCPDIAG_BC_D_COND:
 443                 case TCPDIAG_BC_S_GE:
 444                 case TCPDIAG_BC_S_LE:
 445                 case TCPDIAG_BC_D_GE:
 446                 case TCPDIAG_BC_D_LE:
 447                         if (op->yes < 4 || op->yes > len+4)
 448                                 return -EINVAL;
 449                 case TCPDIAG_BC_JMP:
 450                         if (op->no < 4 || op->no > len+4)
 451                                 return -EINVAL;
 452                         if (op->no < len &&
 453                             !valid_cc(bytecode, bytecode_len, len-op->no))
 454                                 return -EINVAL;
 455                         break;
 456                 case TCPDIAG_BC_NOP:
 457                         if (op->yes < 4 || op->yes > len+4)
 458                                 return -EINVAL;
 459                         break;
 460                 default:
 461                         return -EINVAL;
 462                 }
 463                 bc += op->yes;
 464                 len -= op->yes;
 465         }
 466         return len == 0 ? 0 : -EINVAL;
 467 }
 468
 469
 470 static int tcpdiag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 471 {
 472         int i, num;
 473         int s_i, s_num;
 474         struct tcpdiagreq *r = NLMSG_DATA(cb->nlh);
 475         struct rtattr *bc = NULL;
 476
 477         if (cb->nlh->nlmsg_len > 4+NLMSG_SPACE(sizeof(struct tcpdiagreq)))
 478                 bc = (struct rtattr*)(r+1);
 479
 480         s_i = cb->args[1];
 481         s_num = num = cb->args[2];
 482
 483         if (cb->args[0] == 0) {
 484                 if (!(r->tcpdiag_states&(TCPF_LISTEN|TCPF_SYN_RECV)))
 485                         goto skip_listen_ht;
 486                 tcp_listen_lock();
 487                 for (i = s_i; i < TCP_LHTABLE_SIZE; i++) {
 488                         struct sock *sk;
 489                         struct hlist_node *node;
 490
 491                         if (i > s_i)
 492                                 s_num = 0;
 493
 494                         num = 0;
 495                         sk_for_each(sk, node, &tcp_listening_hash[i]) {
 496                                 struct inet_opt *inet = inet_sk(sk);
 497                                 if (num < s_num)
 498                                         goto next_listen;
 499                                 if (!(r->tcpdiag_states&TCPF_LISTEN) ||
 500                                     r->id.tcpdiag_dport)
 501                                         goto next_listen;
 502                                 if (r->id.tcpdiag_sport != inet->sport &&
 503                                     r->id.tcpdiag_sport)
 504                                         goto next_listen;
 505                                 if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
 506                                         goto next_listen;
 507                                 if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
 508                                                  NETLINK_CB(cb->skb).pid,
 509                                                  cb->nlh->nlmsg_seq) <= 0) {
 510                                         tcp_listen_unlock();
 511                                         goto done;
 512                                 }
 513 next_listen:
 514                                 ++num;
 515                         }
 516                 }
 517                 tcp_listen_unlock();
 518 skip_listen_ht:
 519                 cb->args[0] = 1;
 520                 s_i = num = s_num = 0;
 521         }
 522
 523         if (!(r->tcpdiag_states&~(TCPF_LISTEN|TCPF_SYN_RECV)))
 524                 return skb->len;
 525
 526         for (i = s_i; i < tcp_ehash_size; i++) {
 527                 struct tcp_ehash_bucket *head = &tcp_ehash[i];
 528                 struct sock *sk;
 529                 struct hlist_node *node;
 530
 531                 if (i > s_i)
 532                         s_num = 0;
 533
 534                 read_lock_bh(&head->lock);
 535
 536                 num = 0;
 537                 sk_for_each(sk, node, &head->chain) {
 538                         struct inet_opt *inet = inet_sk(sk);
 539
 540                         if (num < s_num)
 541                                 goto next_normal;
 542                         if (!(r->tcpdiag_states & (1 << sk->sk_state)))
 543                                 goto next_normal;
 544                         if (r->id.tcpdiag_sport != inet->sport &&
 545                             r->id.tcpdiag_sport)
 546                                 goto next_normal;
 547                         if (r->id.tcpdiag_dport != inet->dport && r->id.tcpdiag_dport)
 548                                 goto next_normal;
 549                         if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
 550                                 goto next_normal;
 551                         if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
 552                                          NETLINK_CB(cb->skb).pid,
 553                                          cb->nlh->nlmsg_seq) <= 0) {
 554                                 read_unlock_bh(&head->lock);
 555                                 goto done;
 556                         }
 557 next_normal:
 558                         ++num;
 559                 }
 560
 561                 if (r->tcpdiag_states&TCPF_TIME_WAIT) {
 562                         sk_for_each(sk, node,
 563                                     &tcp_ehash[i + tcp_ehash_size].chain) {
 564                                 struct inet_opt *inet = inet_sk(sk);
 565
 566                                 if (num < s_num)
 567                                         goto next_dying;
 568                                 if (r->id.tcpdiag_sport != inet->sport &&
 569                                     r->id.tcpdiag_sport)
 570                                         goto next_dying;
 571                                 if (r->id.tcpdiag_dport != inet->dport &&
 572                                     r->id.tcpdiag_dport)
 573                                         goto next_dying;
 574                                 if (bc && !tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), sk))
 575                                         goto next_dying;
 576                                 if (tcpdiag_fill(skb, sk, r->tcpdiag_ext,
 577                                                  NETLINK_CB(cb->skb).pid,
 578                                                  cb->nlh->nlmsg_seq) <= 0) {
 579                                         read_unlock_bh(&head->lock);
 580                                         goto done;
 581                                 }
 582 next_dying:
 583                                 ++num;
 584                         }
 585                 }
 586                 read_unlock_bh(&head->lock);
 587         }
 588
 589 done:
 590         cb->args[1] = i;
 591         cb->args[2] = num;
 592         return skb->len;
 593 }
 594
 595 static int tcpdiag_dump_done(struct netlink_callback *cb)
 596 {
 597         return 0;
 598 }
 599
 600
 601 static __inline__ int
 602 tcpdiag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 603 {
 604         if (!(nlh->nlmsg_flags&NLM_F_REQUEST))
 605                 return 0;
 606
 607         if (nlh->nlmsg_type != TCPDIAG_GETSOCK)
 608                 goto err_inval;
 609
 610         if (NLMSG_LENGTH(sizeof(struct tcpdiagreq)) > skb->len)
 611                 goto err_inval;
 612
 613         if (nlh->nlmsg_flags&NLM_F_DUMP) {
 614                 if (nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(struct tcpdiagreq))) {
 615                         struct rtattr *rta = (struct rtattr*)(NLMSG_DATA(nlh) + sizeof(struct tcpdiagreq));
 616                         if (rta->rta_type != TCPDIAG_REQ_BYTECODE ||
 617                             rta->rta_len < 8 ||
 618                             rta->rta_len > nlh->nlmsg_len - NLMSG_SPACE(sizeof(struct tcpdiagreq)))
 619                                 goto err_inval;
 620                         if (tcpdiag_bc_audit(RTA_DATA(rta), RTA_PAYLOAD(rta)))
 621                                 goto err_inval;
 622                 }
 623                 return netlink_dump_start(tcpnl, skb, nlh,
 624                                           tcpdiag_dump,
 625                                           tcpdiag_dump_done);
 626         } else {
 627                 return tcpdiag_get_exact(skb, nlh);
 628         }
 629
 630 err_inval:
 631         return -EINVAL;
 632 }
 633
 634
 635 static inline void tcpdiag_rcv_skb(struct sk_buff *skb)
 636 {
 637         int err;
 638         struct nlmsghdr * nlh;
 639
 640         if (skb->len >= NLMSG_SPACE(0)) {
 641                 nlh = (struct nlmsghdr *)skb->data;
 642                 if (nlh->nlmsg_len < sizeof(*nlh) || skb->len < nlh->nlmsg_len)
 643                         return;
 644                 err = tcpdiag_rcv_msg(skb, nlh);
 645                 if (err || nlh->nlmsg_flags & NLM_F_ACK)
 646                         netlink_ack(skb, nlh, err);
 647         }
 648 }
 649
 650 static void tcpdiag_rcv(struct sock *sk, int len)
 651 {
 652         struct sk_buff *skb;
 653
 654         while ((skb = skb_dequeue(&sk->sk_receive_queue)) != NULL) {
 655                 tcpdiag_rcv_skb(skb);
 656                 kfree_skb(skb);
 657         }
 658 }
 659
 660 void __init tcpdiag_init(void)
 661 {
 662         tcpnl = netlink_kernel_create(NETLINK_TCPDIAG, tcpdiag_rcv);
 663         if (tcpnl == NULL)
 664                 panic("tcpdiag_init: Cannot create netlink socket.");
 665 }