net/ipv4/tcp_diag.c

   1 /*
   2  * tcp_diag.c   Module for monitoring TCP sockets.
   3  *
   4  * Version:     $Id: tcp_diag.c,v 1.3 2002/02/01 22:01:04 davem Exp $
   5  *
   6  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
   7  *
   8  *      This program is free software; you can redistribute it and/or
   9  *      modify it under the terms of the GNU General Public License
  10  *      as published by the Free Software Foundation; either version
  11  *      2 of the License, or (at your option) any later version.
  12  */
  13
  14 #include <linux/config.h>
  15 #include <linux/module.h>
  16 #include <linux/types.h>
  17 #include <linux/fcntl.h>
  18 #include <linux/random.h>
  19 #include <linux/cache.h>
  20 #include <linux/init.h>
  21 #include <linux/time.h>
  22
  23 #include <net/icmp.h>
  24 #include <net/tcp.h>
  25 #include <net/ipv6.h>
  26 #include <net/inet_common.h>
  27
  28 #include <linux/inet.h>
  29 #include <linux/stddef.h>
  30
  31 #include <linux/tcp_diag.h>
  32
  33 struct tcpdiag_entry
  34 {
  35         u32 *saddr;
  36         u32 *daddr;
  37         u16 sport;
  38         u16 dport;
  39         u16 family;
  40         u16 userlocks;
  41 };
  42
  43 static struct sock *tcpnl;
  44
  45 #define TCPDIAG_PUT(skb, attrtype, attrlen) \
  46         RTA_DATA(__RTA_PUT(skb, attrtype, attrlen))
  47
  48 static int tcpdiag_fill(struct sk_buff *skb, struct sock *sk,
  49                         int ext, u32 pid, u32 seq, u16 nlmsg_flags)
  50 {
  51         struct inet_sock *inet = inet_sk(sk);
  52         struct tcp_sock *tp = tcp_sk(sk);
  53         struct tcpdiagmsg *r;
  54         struct nlmsghdr  *nlh;
  55         struct tcp_info  *info = NULL;
  56         struct tcpdiag_meminfo  *minfo = NULL;
  57         unsigned char    *b = skb->tail;
  58
  59         nlh = NLMSG_PUT(skb, pid, seq, TCPDIAG_GETSOCK, sizeof(*r));
  60         nlh->nlmsg_flags = nlmsg_flags;
  61         r = NLMSG_DATA(nlh);
  62         if (sk->sk_state != TCP_TIME_WAIT) {
  63                 if (ext & (1<<(TCPDIAG_MEMINFO-1)))
  64                         minfo = TCPDIAG_PUT(skb, TCPDIAG_MEMINFO, sizeof(*minfo));
  65                 if (ext & (1<<(TCPDIAG_INFO-1)))
  66                         info = TCPDIAG_PUT(skb, TCPDIAG_INFO, sizeof(*info));
  67
  68         }
  69         r->tcpdiag_family = sk->sk_family;
  70         r->tcpdiag_state = sk->sk_state;
  71         r->tcpdiag_timer = 0;
  72         r->tcpdiag_retrans = 0;
  73
  74         r->id.tcpdiag_if = sk->sk_bound_dev_if;
  75         r->id.tcpdiag_cookie[0] = (u32)(unsigned long)sk;
  76         r->id.tcpdiag_cookie[1] = (u32)(((unsigned long)sk >> 31) >> 1);
  77
  78         if (r->tcpdiag_state == TCP_TIME_WAIT) {
  79                 struct tcp_tw_bucket *tw = (struct tcp_tw_bucket*)sk;
  80                 long tmo = tw->tw_ttd - jiffies;
  81                 if (tmo < 0)
  82                         tmo = 0;
  83
  84                 r->id.tcpdiag_sport = tw->tw_sport;
  85                 r->id.tcpdiag_dport = tw->tw_dport;
  86                 r->id.tcpdiag_src[0] = tw->tw_rcv_saddr;
  87                 r->id.tcpdiag_dst[0] = tw->tw_daddr;
  88                 r->tcpdiag_state = tw->tw_substate;
  89                 r->tcpdiag_timer = 3;
  90                 r->tcpdiag_expires = (tmo*1000+HZ-1)/HZ;
  91                 r->tcpdiag_rqueue = 0;
  92                 r->tcpdiag_wqueue = 0;
  93                 r->tcpdiag_uid = 0;
  94                 r->tcpdiag_inode = 0;
  95 #ifdef CONFIG_IP_TCPDIAG_IPV6
  96                 if (r->tcpdiag_family == AF_INET6) {
  97                         ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
  98                                        &tw->tw_v6_rcv_saddr);
  99                         ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
 100                                        &tw->tw_v6_daddr);
 101                 }
 102 #endif
 103                 nlh->nlmsg_len = skb->tail - b;
 104                 return skb->len;
 105         }
 106
 107         r->id.tcpdiag_sport = inet->sport;
 108         r->id.tcpdiag_dport = inet->dport;
 109         r->id.tcpdiag_src[0] = inet->rcv_saddr;
 110         r->id.tcpdiag_dst[0] = inet->daddr;
 111
 112 #ifdef CONFIG_IP_TCPDIAG_IPV6
 113         if (r->tcpdiag_family == AF_INET6) {
 114                 struct ipv6_pinfo *np = inet6_sk(sk);
 115
 116                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
 117                                &np->rcv_saddr);
 118                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
 119                                &np->daddr);
 120         }
 121 #endif
 122
 123 #define EXPIRES_IN_MS(tmo)  ((tmo-jiffies)*1000+HZ-1)/HZ
 124
 125         if (tp->pending == TCP_TIME_RETRANS) {
 126                 r->tcpdiag_timer = 1;
 127                 r->tcpdiag_retrans = tp->retransmits;
 128                 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
 129         } else if (tp->pending == TCP_TIME_PROBE0) {
 130                 r->tcpdiag_timer = 4;
 131                 r->tcpdiag_retrans = tp->probes_out;
 132                 r->tcpdiag_expires = EXPIRES_IN_MS(tp->timeout);
 133         } else if (timer_pending(&sk->sk_timer)) {
 134                 r->tcpdiag_timer = 2;
 135                 r->tcpdiag_retrans = tp->probes_out;
 136                 r->tcpdiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 137         } else {
 138                 r->tcpdiag_timer = 0;
 139                 r->tcpdiag_expires = 0;
 140         }
 141 #undef EXPIRES_IN_MS
 142
 143         r->tcpdiag_rqueue = tp->rcv_nxt - tp->copied_seq;
 144         r->tcpdiag_wqueue = tp->write_seq - tp->snd_una;
 145         r->tcpdiag_uid = sock_i_uid(sk);
 146         r->tcpdiag_inode = sock_i_ino(sk);
 147
 148         if (minfo) {
 149                 minfo->tcpdiag_rmem = atomic_read(&sk->sk_rmem_alloc);
 150                 minfo->tcpdiag_wmem = sk->sk_wmem_queued;
 151                 minfo->tcpdiag_fmem = sk->sk_forward_alloc;
 152                 minfo->tcpdiag_tmem = atomic_read(&sk->sk_wmem_alloc);
 153         }
 154
 155         if (info)
 156                 tcp_get_info(sk, info);
 157
 158         if (sk->sk_state < TCP_TIME_WAIT && tp->ca_ops->get_info)
 159                 tp->ca_ops->get_info(tp, ext, skb);
 160
 161         nlh->nlmsg_len = skb->tail - b;
 162         return skb->len;
 163
 164 rtattr_failure:
 165 nlmsg_failure:
 166         skb_trim(skb, b - skb->data);
 167         return -1;
 168 }
 169
 170 extern struct sock *tcp_v4_lookup(u32 saddr, u16 sport, u32 daddr, u16 dport,
 171                                   int dif);
 172 #ifdef CONFIG_IP_TCPDIAG_IPV6
 173 extern struct sock *tcp_v6_lookup(struct in6_addr *saddr, u16 sport,
 174                                   struct in6_addr *daddr, u16 dport,
 175                                   int dif);
 176 #else
 177 static inline struct sock *tcp_v6_lookup(struct in6_addr *saddr, u16 sport,
 178                                          struct in6_addr *daddr, u16 dport,
 179                                          int dif)
 180 {
 181         return NULL;
 182 }
 183 #endif
 184
 185 static int tcpdiag_get_exact(struct sk_buff *in_skb, const struct nlmsghdr *nlh)
 186 {
 187         int err;
 188         struct sock *sk;
 189         struct tcpdiagreq *req = NLMSG_DATA(nlh);
 190         struct sk_buff *rep;
 191
 192         if (req->tcpdiag_family == AF_INET) {
 193                 sk = tcp_v4_lookup(req->id.tcpdiag_dst[0], req->id.tcpdiag_dport,
 194                                    req->id.tcpdiag_src[0], req->id.tcpdiag_sport,
 195                                    req->id.tcpdiag_if);
 196         }
 197 #ifdef CONFIG_IP_TCPDIAG_IPV6
 198         else if (req->tcpdiag_family == AF_INET6) {
 199                 sk = tcp_v6_lookup((struct in6_addr*)req->id.tcpdiag_dst, req->id.tcpdiag_dport,
 200                                    (struct in6_addr*)req->id.tcpdiag_src, req->id.tcpdiag_sport,
 201                                    req->id.tcpdiag_if);
 202         }
 203 #endif
 204         else {
 205                 return -EINVAL;
 206         }
 207
 208         if (sk == NULL)
 209                 return -ENOENT;
 210
 211         err = -ESTALE;
 212         if ((req->id.tcpdiag_cookie[0] != TCPDIAG_NOCOOKIE ||
 213              req->id.tcpdiag_cookie[1] != TCPDIAG_NOCOOKIE) &&
 214             ((u32)(unsigned long)sk != req->id.tcpdiag_cookie[0] ||
 215              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.tcpdiag_cookie[1]))
 216                 goto out;
 217
 218         err = -ENOMEM;
 219         rep = alloc_skb(NLMSG_SPACE(sizeof(struct tcpdiagmsg)+
 220                                     sizeof(struct tcpdiag_meminfo)+
 221                                     sizeof(struct tcp_info)+64), GFP_KERNEL);
 222         if (!rep)
 223                 goto out;
 224
 225         if (tcpdiag_fill(rep, sk, req->tcpdiag_ext,
 226                          NETLINK_CB(in_skb).pid,
 227                          nlh->nlmsg_seq, 0) <= 0)
 228                 BUG();
 229
 230         err = netlink_unicast(tcpnl, rep, NETLINK_CB(in_skb).pid, MSG_DONTWAIT);
 231         if (err > 0)
 232                 err = 0;
 233
 234 out:
 235         if (sk) {
 236                 if (sk->sk_state == TCP_TIME_WAIT)
 237                         tcp_tw_put((struct tcp_tw_bucket*)sk);
 238                 else
 239                         sock_put(sk);
 240         }
 241         return err;
 242 }
 243
 244 static int bitstring_match(const u32 *a1, const u32 *a2, int bits)
 245 {
 246         int words = bits >> 5;
 247
 248         bits &= 0x1f;
 249
 250         if (words) {
 251                 if (memcmp(a1, a2, words << 2))
 252                         return 0;
 253         }
 254         if (bits) {
 255                 __u32 w1, w2;
 256                 __u32 mask;
 257
 258                 w1 = a1[words];
 259                 w2 = a2[words];
 260
 261                 mask = htonl((0xffffffff) << (32 - bits));
 262
 263                 if ((w1 ^ w2) & mask)
 264                         return 0;
 265         }
 266
 267         return 1;
 268 }
 269
 270
 271 static int tcpdiag_bc_run(const void *bc, int len,
 272                           const struct tcpdiag_entry *entry)
 273 {
 274         while (len > 0) {
 275                 int yes = 1;
 276                 const struct tcpdiag_bc_op *op = bc;
 277
 278                 switch (op->code) {
 279                 case TCPDIAG_BC_NOP:
 280                         break;
 281                 case TCPDIAG_BC_JMP:
 282                         yes = 0;
 283                         break;
 284                 case TCPDIAG_BC_S_GE:
 285                         yes = entry->sport >= op[1].no;
 286                         break;
 287                 case TCPDIAG_BC_S_LE:
 288                         yes = entry->dport <= op[1].no;
 289                         break;
 290                 case TCPDIAG_BC_D_GE:
 291                         yes = entry->dport >= op[1].no;
 292                         break;
 293                 case TCPDIAG_BC_D_LE:
 294                         yes = entry->dport <= op[1].no;
 295                         break;
 296                 case TCPDIAG_BC_AUTO:
 297                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 298                         break;
 299                 case TCPDIAG_BC_S_COND:
 300                 case TCPDIAG_BC_D_COND:
 301                 {
 302                         struct tcpdiag_hostcond *cond = (struct tcpdiag_hostcond*)(op+1);
 303                         u32 *addr;
 304
 305                         if (cond->port != -1 &&
 306                             cond->port != (op->code == TCPDIAG_BC_S_COND ?
 307                                              entry->sport : entry->dport)) {
 308                                 yes = 0;
 309                                 break;
 310                         }
 311
 312                         if (cond->prefix_len == 0)
 313                                 break;
 314
 315                         if (op->code == TCPDIAG_BC_S_COND)
 316                                 addr = entry->saddr;
 317                         else
 318                                 addr = entry->daddr;
 319
 320                         if (bitstring_match(addr, cond->addr, cond->prefix_len))
 321                                 break;
 322                         if (entry->family == AF_INET6 &&
 323                             cond->family == AF_INET) {
 324                                 if (addr[0] == 0 && addr[1] == 0 &&
 325                                     addr[2] == htonl(0xffff) &&
 326                                     bitstring_match(addr+3, cond->addr, cond->prefix_len))
 327                                         break;
 328                         }
 329                         yes = 0;
 330                         break;
 331                 }
 332                 }
 333
 334                 if (yes) {
 335                         len -= op->yes;
 336                         bc += op->yes;
 337                 } else {
 338                         len -= op->no;
 339                         bc += op->no;
 340                 }
 341         }
 342         return (len == 0);
 343 }
 344
 345 static int valid_cc(const void *bc, int len, int cc)
 346 {
 347         while (len >= 0) {
 348                 const struct tcpdiag_bc_op *op = bc;
 349
 350                 if (cc > len)
 351                         return 0;
 352                 if (cc == len)
 353                         return 1;
 354                 if (op->yes < 4)
 355                         return 0;
 356                 len -= op->yes;
 357                 bc  += op->yes;
 358         }
 359         return 0;
 360 }
 361
 362 static int tcpdiag_bc_audit(const void *bytecode, int bytecode_len)
 363 {
 364         const unsigned char *bc = bytecode;
 365         int  len = bytecode_len;
 366
 367         while (len > 0) {
 368                 struct tcpdiag_bc_op *op = (struct tcpdiag_bc_op*)bc;
 369
 370 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 371                 switch (op->code) {
 372                 case TCPDIAG_BC_AUTO:
 373                 case TCPDIAG_BC_S_COND:
 374                 case TCPDIAG_BC_D_COND:
 375                 case TCPDIAG_BC_S_GE:
 376                 case TCPDIAG_BC_S_LE:
 377                 case TCPDIAG_BC_D_GE:
 378                 case TCPDIAG_BC_D_LE:
 379                         if (op->yes < 4 || op->yes > len+4)
 380                                 return -EINVAL;
 381                 case TCPDIAG_BC_JMP:
 382                         if (op->no < 4 || op->no > len+4)
 383                                 return -EINVAL;
 384                         if (op->no < len &&
 385                             !valid_cc(bytecode, bytecode_len, len-op->no))
 386                                 return -EINVAL;
 387                         break;
 388                 case TCPDIAG_BC_NOP:
 389                         if (op->yes < 4 || op->yes > len+4)
 390                                 return -EINVAL;
 391                         break;
 392                 default:
 393                         return -EINVAL;
 394                 }
 395                 bc += op->yes;
 396                 len -= op->yes;
 397         }
 398         return len == 0 ? 0 : -EINVAL;
 399 }
 400
 401 static int tcpdiag_dump_sock(struct sk_buff *skb, struct sock *sk,
 402                              struct netlink_callback *cb)
 403 {
 404         struct tcpdiagreq *r = NLMSG_DATA(cb->nlh);
 405
 406         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 407                 struct tcpdiag_entry entry;
 408                 struct rtattr *bc = (struct rtattr *)(r + 1);
 409                 struct inet_sock *inet = inet_sk(sk);
 410
 411                 entry.family = sk->sk_family;
 412 #ifdef CONFIG_IP_TCPDIAG_IPV6
 413                 if (entry.family == AF_INET6) {
 414                         struct ipv6_pinfo *np = inet6_sk(sk);
 415
 416                         entry.saddr = np->rcv_saddr.s6_addr32;
 417                         entry.daddr = np->daddr.s6_addr32;
 418                 } else
 419 #endif
 420                 {
 421                         entry.saddr = &inet->rcv_saddr;
 422                         entry.daddr = &inet->daddr;
 423                 }
 424                 entry.sport = inet->num;
 425                 entry.dport = ntohs(inet->dport);
 426                 entry.userlocks = sk->sk_userlocks;
 427
 428                 if (!tcpdiag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), &entry))
 429                         return 0;
 430         }
 431
 432         return tcpdiag_fill(skb, sk, r->tcpdiag_ext, NETLINK_CB(cb->skb).pid,
 433                             cb->nlh->nlmsg_seq, NLM_F_MULTI);
 434 }
 435
 436 static int tcpdiag_fill_req(struct sk_buff *skb, struct sock *sk,
 437                             struct request_sock *req,
 438                             u32 pid, u32 seq)
 439 {
 440         const struct inet_request_sock *ireq = inet_rsk(req);
 441         struct inet_sock *inet = inet_sk(sk);
 442         unsigned char *b = skb->tail;
 443         struct tcpdiagmsg *r;
 444         struct nlmsghdr *nlh;
 445         long tmo;
 446
 447         nlh = NLMSG_PUT(skb, pid, seq, TCPDIAG_GETSOCK, sizeof(*r));
 448         nlh->nlmsg_flags = NLM_F_MULTI;
 449         r = NLMSG_DATA(nlh);
 450
 451         r->tcpdiag_family = sk->sk_family;
 452         r->tcpdiag_state = TCP_SYN_RECV;
 453         r->tcpdiag_timer = 1;
 454         r->tcpdiag_retrans = req->retrans;
 455
 456         r->id.tcpdiag_if = sk->sk_bound_dev_if;
 457         r->id.tcpdiag_cookie[0] = (u32)(unsigned long)req;
 458         r->id.tcpdiag_cookie[1] = (u32)(((unsigned long)req >> 31) >> 1);
 459
 460         tmo = req->expires - jiffies;
 461         if (tmo < 0)
 462                 tmo = 0;
 463
 464         r->id.tcpdiag_sport = inet->sport;
 465         r->id.tcpdiag_dport = ireq->rmt_port;
 466         r->id.tcpdiag_src[0] = ireq->loc_addr;
 467         r->id.tcpdiag_dst[0] = ireq->rmt_addr;
 468         r->tcpdiag_expires = jiffies_to_msecs(tmo),
 469         r->tcpdiag_rqueue = 0;
 470         r->tcpdiag_wqueue = 0;
 471         r->tcpdiag_uid = sock_i_uid(sk);
 472         r->tcpdiag_inode = 0;
 473 #ifdef CONFIG_IP_TCPDIAG_IPV6
 474         if (r->tcpdiag_family == AF_INET6) {
 475                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_src,
 476                                &tcp6_rsk(req)->loc_addr);
 477                 ipv6_addr_copy((struct in6_addr *)r->id.tcpdiag_dst,
 478                                &tcp6_rsk(req)->rmt_addr);
 479         }
 480 #endif
 481         nlh->nlmsg_len = skb->tail - b;
 482
 483         return skb->len;
 484
 485 nlmsg_failure:
 486         skb_trim(skb, b - skb->data);
 487         return -1;
 488 }
 489
 490 static int tcpdiag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 491                              struct netlink_callback *cb)
 492 {
 493         struct tcpdiag_entry entry;
 494         struct tcpdiagreq *r = NLMSG_DATA(cb->nlh);
 495         struct tcp_sock *tp = tcp_sk(sk);
 496         struct listen_sock *lopt;
 497         struct rtattr *bc = NULL;
 498         struct inet_sock *inet = inet_sk(sk);
 499         int j, s_j;
 500         int reqnum, s_reqnum;
 501         int err = 0;
 502
 503         s_j = cb->args[3];
 504         s_reqnum = cb->args[4];
 505
 506         if (s_j > 0)
 507                 s_j--;
 508
 509         entry.family = sk->sk_family;
 510
 511         read_lock_bh(&tp->accept_queue.syn_wait_lock);
 512
 513         lopt = tp->accept_queue.listen_opt;
 514         if (!lopt || !lopt->qlen)
 515                 goto out;
 516
 517         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 518                 bc = (struct rtattr *)(r + 1);
 519                 entry.sport = inet->num;
 520                 entry.userlocks = sk->sk_userlocks;
 521         }
 522
 523         for (j = s_j; j < TCP_SYNQ_HSIZE; j++) {
 524                 struct request_sock *req, *head = lopt->syn_table[j];
 525
 526                 reqnum = 0;
 527                 for (req = head; req; reqnum++, req = req->dl_next) {
 528                         struct inet_request_sock *ireq = inet_rsk(req);
 529
 530                         if (reqnum < s_reqnum)
 531                                 continue;
 532                         if (r->id.tcpdiag_dport != ireq->rmt_port &&
 533                             r->id.tcpdiag_dport)
 534                                 continue;
 535
 536                         if (bc) {
 537                                 entry.saddr =
 538 #ifdef CONFIG_IP_TCPDIAG_IPV6
 539                                         (entry.family == AF_INET6) ?
 540                                         tcp6_rsk(req)->loc_addr.s6_addr32 :
 541 #endif
 542                                         &ireq->loc_addr;
 543                                 entry.daddr =
 544 #ifdef CONFIG_IP_TCPDIAG_IPV6
 545                                         (entry.family == AF_INET6) ?
 546                                         tcp6_rsk(req)->rmt_addr.s6_addr32 :
 547 #endif
 548                                         &ireq->rmt_addr;
 549                                 entry.dport = ntohs(ireq->rmt_port);
 550
 551                                 if (!tcpdiag_bc_run(RTA_DATA(bc),
 552                                                     RTA_PAYLOAD(bc), &entry))
 553                                         continue;
 554                         }
 555
 556                         err = tcpdiag_fill_req(skb, sk, req,
 557                                                NETLINK_CB(cb->skb).pid,
 558                                                cb->nlh->nlmsg_seq);
 559                         if (err < 0) {
 560                                 cb->args[3] = j + 1;
 561                                 cb->args[4] = reqnum;
 562                                 goto out;
 563                         }
 564                 }
 565
 566                 s_reqnum = 0;
 567         }
 568
 569 out:
 570         read_unlock_bh(&tp->accept_queue.syn_wait_lock);
 571
 572         return err;
 573 }
 574
 575 static int tcpdiag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 576 {
 577         int i, num;
 578         int s_i, s_num;
 579         struct tcpdiagreq *r = NLMSG_DATA(cb->nlh);
 580
 581         s_i = cb->args[1];
 582         s_num = num = cb->args[2];
 583
 584         if (cb->args[0] == 0) {
 585                 if (!(r->tcpdiag_states&(TCPF_LISTEN|TCPF_SYN_RECV)))
 586                         goto skip_listen_ht;
 587                 tcp_listen_lock();
 588                 for (i = s_i; i < TCP_LHTABLE_SIZE; i++) {
 589                         struct sock *sk;
 590                         struct hlist_node *node;
 591
 592                         num = 0;
 593                         sk_for_each(sk, node, &tcp_listening_hash[i]) {
 594                                 struct inet_sock *inet = inet_sk(sk);
 595
 596                                 if (num < s_num) {
 597                                         num++;
 598                                         continue;
 599                                 }
 600
 601                                 if (r->id.tcpdiag_sport != inet->sport &&
 602                                     r->id.tcpdiag_sport)
 603                                         goto next_listen;
 604
 605                                 if (!(r->tcpdiag_states&TCPF_LISTEN) ||
 606                                     r->id.tcpdiag_dport ||
 607                                     cb->args[3] > 0)
 608                                         goto syn_recv;
 609
 610                                 if (tcpdiag_dump_sock(skb, sk, cb) < 0) {
 611                                         tcp_listen_unlock();
 612                                         goto done;
 613                                 }
 614
 615 syn_recv:
 616                                 if (!(r->tcpdiag_states&TCPF_SYN_RECV))
 617                                         goto next_listen;
 618
 619                                 if (tcpdiag_dump_reqs(skb, sk, cb) < 0) {
 620                                         tcp_listen_unlock();
 621                                         goto done;
 622                                 }
 623
 624 next_listen:
 625                                 cb->args[3] = 0;
 626                                 cb->args[4] = 0;
 627                                 ++num;
 628                         }
 629
 630                         s_num = 0;
 631                         cb->args[3] = 0;
 632                         cb->args[4] = 0;
 633                 }
 634                 tcp_listen_unlock();
 635 skip_listen_ht:
 636                 cb->args[0] = 1;
 637                 s_i = num = s_num = 0;
 638         }
 639
 640         if (!(r->tcpdiag_states&~(TCPF_LISTEN|TCPF_SYN_RECV)))
 641                 return skb->len;
 642
 643         for (i = s_i; i < tcp_ehash_size; i++) {
 644                 struct tcp_ehash_bucket *head = &tcp_ehash[i];
 645                 struct sock *sk;
 646                 struct hlist_node *node;
 647
 648                 if (i > s_i)
 649                         s_num = 0;
 650
 651                 read_lock_bh(&head->lock);
 652
 653                 num = 0;
 654                 sk_for_each(sk, node, &head->chain) {
 655                         struct inet_sock *inet = inet_sk(sk);
 656
 657                         if (num < s_num)
 658                                 goto next_normal;
 659                         if (!(r->tcpdiag_states & (1 << sk->sk_state)))
 660                                 goto next_normal;
 661                         if (r->id.tcpdiag_sport != inet->sport &&
 662                             r->id.tcpdiag_sport)
 663                                 goto next_normal;
 664                         if (r->id.tcpdiag_dport != inet->dport && r->id.tcpdiag_dport)
 665                                 goto next_normal;
 666                         if (tcpdiag_dump_sock(skb, sk, cb) < 0) {
 667                                 read_unlock_bh(&head->lock);
 668                                 goto done;
 669                         }
 670 next_normal:
 671                         ++num;
 672                 }
 673
 674                 if (r->tcpdiag_states&TCPF_TIME_WAIT) {
 675                         sk_for_each(sk, node,
 676                                     &tcp_ehash[i + tcp_ehash_size].chain) {
 677                                 struct inet_sock *inet = inet_sk(sk);
 678
 679                                 if (num < s_num)
 680                                         goto next_dying;
 681                                 if (r->id.tcpdiag_sport != inet->sport &&
 682                                     r->id.tcpdiag_sport)
 683                                         goto next_dying;
 684                                 if (r->id.tcpdiag_dport != inet->dport &&
 685                                     r->id.tcpdiag_dport)
 686                                         goto next_dying;
 687                                 if (tcpdiag_dump_sock(skb, sk, cb) < 0) {
 688                                         read_unlock_bh(&head->lock);
 689                                         goto done;
 690                                 }
 691 next_dying:
 692                                 ++num;
 693                         }
 694                 }
 695                 read_unlock_bh(&head->lock);
 696         }
 697
 698 done:
 699         cb->args[1] = i;
 700         cb->args[2] = num;
 701         return skb->len;
 702 }
 703
 704 static int tcpdiag_dump_done(struct netlink_callback *cb)
 705 {
 706         return 0;
 707 }
 708
 709
 710 static __inline__ int
 711 tcpdiag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 712 {
 713         if (!(nlh->nlmsg_flags&NLM_F_REQUEST))
 714                 return 0;
 715
 716         if (nlh->nlmsg_type != TCPDIAG_GETSOCK)
 717                 goto err_inval;
 718
 719         if (NLMSG_LENGTH(sizeof(struct tcpdiagreq)) > skb->len)
 720                 goto err_inval;
 721
 722         if (nlh->nlmsg_flags&NLM_F_DUMP) {
 723                 if (nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(struct tcpdiagreq))) {
 724                         struct rtattr *rta = (struct rtattr*)(NLMSG_DATA(nlh) + sizeof(struct tcpdiagreq));
 725                         if (rta->rta_type != TCPDIAG_REQ_BYTECODE ||
 726                             rta->rta_len < 8 ||
 727                             rta->rta_len > nlh->nlmsg_len - NLMSG_SPACE(sizeof(struct tcpdiagreq)))
 728                                 goto err_inval;
 729                         if (tcpdiag_bc_audit(RTA_DATA(rta), RTA_PAYLOAD(rta)))
 730                                 goto err_inval;
 731                 }
 732                 return netlink_dump_start(tcpnl, skb, nlh,
 733                                           tcpdiag_dump,
 734                                           tcpdiag_dump_done);
 735         } else {
 736                 return tcpdiag_get_exact(skb, nlh);
 737         }
 738
 739 err_inval:
 740         return -EINVAL;
 741 }
 742
 743
 744 static inline void tcpdiag_rcv_skb(struct sk_buff *skb)
 745 {
 746         int err;
 747         struct nlmsghdr * nlh;
 748
 749         if (skb->len >= NLMSG_SPACE(0)) {
 750                 nlh = (struct nlmsghdr *)skb->data;
 751                 if (nlh->nlmsg_len < sizeof(*nlh) || skb->len < nlh->nlmsg_len)
 752                         return;
 753                 err = tcpdiag_rcv_msg(skb, nlh);
 754                 if (err || nlh->nlmsg_flags & NLM_F_ACK)
 755                         netlink_ack(skb, nlh, err);
 756         }
 757 }
 758
 759 static void tcpdiag_rcv(struct sock *sk, int len)
 760 {
 761         struct sk_buff *skb;
 762         unsigned int qlen = skb_queue_len(&sk->sk_receive_queue);
 763
 764         while (qlen-- && (skb = skb_dequeue(&sk->sk_receive_queue))) {
 765                 tcpdiag_rcv_skb(skb);
 766                 kfree_skb(skb);
 767         }
 768 }
 769
 770 static int __init tcpdiag_init(void)
 771 {
 772         tcpnl = netlink_kernel_create(NETLINK_TCPDIAG, tcpdiag_rcv);
 773         if (tcpnl == NULL)
 774                 return -ENOMEM;
 775         return 0;
 776 }
 777
 778 static void __exit tcpdiag_exit(void)
 779 {
 780         sock_release(tcpnl->sk_socket);
 781 }
 782
 783 module_init(tcpdiag_init);
 784 module_exit(tcpdiag_exit);
 785 MODULE_LICENSE("GPL");