net/ipv4/inet_diag.c

   1 /*
   2  * inet_diag.c  Module for monitoring INET transport protocols sockets.
   3  *
   4  * Version:     $Id: inet_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/kernel.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 #include <net/inet_connection_sock.h>
  28 #include <net/inet_hashtables.h>
  29 #include <net/inet_timewait_sock.h>
  30 #include <net/inet6_hashtables.h>
  31 #include <net/netlink.h>
  32
  33 #include <linux/inet.h>
  34 #include <linux/stddef.h>
  35
  36 #include <linux/inet_diag.h>
  37
  38 static const struct inet_diag_handler **inet_diag_table;
  39
  40 struct inet_diag_entry {
  41         __be32 *saddr;
  42         __be32 *daddr;
  43         u16 sport;
  44         u16 dport;
  45         u16 family;
  46         u16 userlocks;
  47 };
  48
  49 static struct sock *idiagnl;
  50
  51 #define INET_DIAG_PUT(skb, attrtype, attrlen) \
  52         RTA_DATA(__RTA_PUT(skb, attrtype, attrlen))
  53
  54 static int inet_csk_diag_fill(struct sock *sk,
  55                               struct sk_buff *skb,
  56                               int ext, u32 pid, u32 seq, u16 nlmsg_flags,
  57                               const struct nlmsghdr *unlh)
  58 {
  59         const struct inet_sock *inet = inet_sk(sk);
  60         const struct inet_connection_sock *icsk = inet_csk(sk);
  61         struct inet_diag_msg *r;
  62         struct nlmsghdr  *nlh;
  63         void *info = NULL;
  64         struct inet_diag_meminfo  *minfo = NULL;
  65         unsigned char    *b = skb_tail_pointer(skb);
  66         const struct inet_diag_handler *handler;
  67
  68         handler = inet_diag_table[unlh->nlmsg_type];
  69         BUG_ON(handler == NULL);
  70
  71         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
  72         nlh->nlmsg_flags = nlmsg_flags;
  73
  74         r = NLMSG_DATA(nlh);
  75         BUG_ON(sk->sk_state == TCP_TIME_WAIT);
  76
  77         if (ext & (1 << (INET_DIAG_MEMINFO - 1)))
  78                 minfo = INET_DIAG_PUT(skb, INET_DIAG_MEMINFO, sizeof(*minfo));
  79
  80         if (ext & (1 << (INET_DIAG_INFO - 1)))
  81                 info = INET_DIAG_PUT(skb, INET_DIAG_INFO,
  82                                      handler->idiag_info_size);
  83
  84         if ((ext & (1 << (INET_DIAG_CONG - 1))) && icsk->icsk_ca_ops) {
  85                 const size_t len = strlen(icsk->icsk_ca_ops->name);
  86
  87                 strcpy(INET_DIAG_PUT(skb, INET_DIAG_CONG, len + 1),
  88                        icsk->icsk_ca_ops->name);
  89         }
  90
  91         r->idiag_family = sk->sk_family;
  92         r->idiag_state = sk->sk_state;
  93         r->idiag_timer = 0;
  94         r->idiag_retrans = 0;
  95
  96         r->id.idiag_if = sk->sk_bound_dev_if;
  97         r->id.idiag_cookie[0] = (u32)(unsigned long)sk;
  98         r->id.idiag_cookie[1] = (u32)(((unsigned long)sk >> 31) >> 1);
  99
 100         r->id.idiag_sport = inet->sport;
 101         r->id.idiag_dport = inet->dport;
 102         r->id.idiag_src[0] = inet->rcv_saddr;
 103         r->id.idiag_dst[0] = inet->daddr;
 104
 105 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 106         if (r->idiag_family == AF_INET6) {
 107                 struct ipv6_pinfo *np = inet6_sk(sk);
 108
 109                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 110                                &np->rcv_saddr);
 111                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 112                                &np->daddr);
 113         }
 114 #endif
 115
 116 #define EXPIRES_IN_MS(tmo)  DIV_ROUND_UP((tmo - jiffies) * 1000, HZ)
 117
 118         if (icsk->icsk_pending == ICSK_TIME_RETRANS) {
 119                 r->idiag_timer = 1;
 120                 r->idiag_retrans = icsk->icsk_retransmits;
 121                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 122         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 123                 r->idiag_timer = 4;
 124                 r->idiag_retrans = icsk->icsk_probes_out;
 125                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 126         } else if (timer_pending(&sk->sk_timer)) {
 127                 r->idiag_timer = 2;
 128                 r->idiag_retrans = icsk->icsk_probes_out;
 129                 r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 130         } else {
 131                 r->idiag_timer = 0;
 132                 r->idiag_expires = 0;
 133         }
 134 #undef EXPIRES_IN_MS
 135
 136         r->idiag_uid = sock_i_uid(sk);
 137         r->idiag_inode = sock_i_ino(sk);
 138
 139         if (minfo) {
 140                 minfo->idiag_rmem = atomic_read(&sk->sk_rmem_alloc);
 141                 minfo->idiag_wmem = sk->sk_wmem_queued;
 142                 minfo->idiag_fmem = sk->sk_forward_alloc;
 143                 minfo->idiag_tmem = atomic_read(&sk->sk_wmem_alloc);
 144         }
 145
 146         handler->idiag_get_info(sk, r, info);
 147
 148         if (sk->sk_state < TCP_TIME_WAIT &&
 149             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 150                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 151
 152         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 153         return skb->len;
 154
 155 rtattr_failure:
 156 nlmsg_failure:
 157         nlmsg_trim(skb, b);
 158         return -EMSGSIZE;
 159 }
 160
 161 static int inet_twsk_diag_fill(struct inet_timewait_sock *tw,
 162                                struct sk_buff *skb, int ext, u32 pid,
 163                                u32 seq, u16 nlmsg_flags,
 164                                const struct nlmsghdr *unlh)
 165 {
 166         long tmo;
 167         struct inet_diag_msg *r;
 168         const unsigned char *previous_tail = skb_tail_pointer(skb);
 169         struct nlmsghdr *nlh = NLMSG_PUT(skb, pid, seq,
 170                                          unlh->nlmsg_type, sizeof(*r));
 171
 172         r = NLMSG_DATA(nlh);
 173         BUG_ON(tw->tw_state != TCP_TIME_WAIT);
 174
 175         nlh->nlmsg_flags = nlmsg_flags;
 176
 177         tmo = tw->tw_ttd - jiffies;
 178         if (tmo < 0)
 179                 tmo = 0;
 180
 181         r->idiag_family       = tw->tw_family;
 182         r->idiag_state        = tw->tw_state;
 183         r->idiag_timer        = 0;
 184         r->idiag_retrans      = 0;
 185         r->id.idiag_if        = tw->tw_bound_dev_if;
 186         r->id.idiag_cookie[0] = (u32)(unsigned long)tw;
 187         r->id.idiag_cookie[1] = (u32)(((unsigned long)tw >> 31) >> 1);
 188         r->id.idiag_sport     = tw->tw_sport;
 189         r->id.idiag_dport     = tw->tw_dport;
 190         r->id.idiag_src[0]    = tw->tw_rcv_saddr;
 191         r->id.idiag_dst[0]    = tw->tw_daddr;
 192         r->idiag_state        = tw->tw_substate;
 193         r->idiag_timer        = 3;
 194         r->idiag_expires      = DIV_ROUND_UP(tmo * 1000, HZ);
 195         r->idiag_rqueue       = 0;
 196         r->idiag_wqueue       = 0;
 197         r->idiag_uid          = 0;
 198         r->idiag_inode        = 0;
 199 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 200         if (tw->tw_family == AF_INET6) {
 201                 const struct inet6_timewait_sock *tw6 =
 202                                                 inet6_twsk((struct sock *)tw);
 203
 204                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 205                                &tw6->tw_v6_rcv_saddr);
 206                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 207                                &tw6->tw_v6_daddr);
 208         }
 209 #endif
 210         nlh->nlmsg_len = skb_tail_pointer(skb) - previous_tail;
 211         return skb->len;
 212 nlmsg_failure:
 213         nlmsg_trim(skb, previous_tail);
 214         return -EMSGSIZE;
 215 }
 216
 217 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
 218                         int ext, u32 pid, u32 seq, u16 nlmsg_flags,
 219                         const struct nlmsghdr *unlh)
 220 {
 221         if (sk->sk_state == TCP_TIME_WAIT)
 222                 return inet_twsk_diag_fill((struct inet_timewait_sock *)sk,
 223                                            skb, ext, pid, seq, nlmsg_flags,
 224                                            unlh);
 225         return inet_csk_diag_fill(sk, skb, ext, pid, seq, nlmsg_flags, unlh);
 226 }
 227
 228 static int inet_diag_get_exact(struct sk_buff *in_skb,
 229                                const struct nlmsghdr *nlh)
 230 {
 231         int err;
 232         struct sock *sk;
 233         struct inet_diag_req *req = NLMSG_DATA(nlh);
 234         struct sk_buff *rep;
 235         struct inet_hashinfo *hashinfo;
 236         const struct inet_diag_handler *handler;
 237
 238         handler = inet_diag_table[nlh->nlmsg_type];
 239         BUG_ON(handler == NULL);
 240         hashinfo = handler->idiag_hashinfo;
 241
 242         if (req->idiag_family == AF_INET) {
 243                 sk = inet_lookup(hashinfo, req->id.idiag_dst[0],
 244                                  req->id.idiag_dport, req->id.idiag_src[0],
 245                                  req->id.idiag_sport, req->id.idiag_if);
 246         }
 247 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 248         else if (req->idiag_family == AF_INET6) {
 249                 sk = inet6_lookup(hashinfo,
 250                                   (struct in6_addr *)req->id.idiag_dst,
 251                                   req->id.idiag_dport,
 252                                   (struct in6_addr *)req->id.idiag_src,
 253                                   req->id.idiag_sport,
 254                                   req->id.idiag_if);
 255         }
 256 #endif
 257         else {
 258                 return -EINVAL;
 259         }
 260
 261         if (sk == NULL)
 262                 return -ENOENT;
 263
 264         err = -ESTALE;
 265         if ((req->id.idiag_cookie[0] != INET_DIAG_NOCOOKIE ||
 266              req->id.idiag_cookie[1] != INET_DIAG_NOCOOKIE) &&
 267             ((u32)(unsigned long)sk != req->id.idiag_cookie[0] ||
 268              (u32)((((unsigned long)sk) >> 31) >> 1) != req->id.idiag_cookie[1]))
 269                 goto out;
 270
 271         err = -ENOMEM;
 272         rep = alloc_skb(NLMSG_SPACE((sizeof(struct inet_diag_msg) +
 273                                      sizeof(struct inet_diag_meminfo) +
 274                                      handler->idiag_info_size + 64)),
 275                         GFP_KERNEL);
 276         if (!rep)
 277                 goto out;
 278
 279         err = sk_diag_fill(sk, rep, req->idiag_ext,
 280                            NETLINK_CB(in_skb).pid,
 281                            nlh->nlmsg_seq, 0, nlh);
 282         if (err < 0) {
 283                 WARN_ON(err == -EMSGSIZE);
 284                 kfree_skb(rep);
 285                 goto out;
 286         }
 287         err = netlink_unicast(idiagnl, rep, NETLINK_CB(in_skb).pid,
 288                               MSG_DONTWAIT);
 289         if (err > 0)
 290                 err = 0;
 291
 292 out:
 293         if (sk) {
 294                 if (sk->sk_state == TCP_TIME_WAIT)
 295                         inet_twsk_put((struct inet_timewait_sock *)sk);
 296                 else
 297                         sock_put(sk);
 298         }
 299         return err;
 300 }
 301
 302 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
 303 {
 304         int words = bits >> 5;
 305
 306         bits &= 0x1f;
 307
 308         if (words) {
 309                 if (memcmp(a1, a2, words << 2))
 310                         return 0;
 311         }
 312         if (bits) {
 313                 __be32 w1, w2;
 314                 __be32 mask;
 315
 316                 w1 = a1[words];
 317                 w2 = a2[words];
 318
 319                 mask = htonl((0xffffffff) << (32 - bits));
 320
 321                 if ((w1 ^ w2) & mask)
 322                         return 0;
 323         }
 324
 325         return 1;
 326 }
 327
 328
 329 static int inet_diag_bc_run(const void *bc, int len,
 330                             const struct inet_diag_entry *entry)
 331 {
 332         while (len > 0) {
 333                 int yes = 1;
 334                 const struct inet_diag_bc_op *op = bc;
 335
 336                 switch (op->code) {
 337                 case INET_DIAG_BC_NOP:
 338                         break;
 339                 case INET_DIAG_BC_JMP:
 340                         yes = 0;
 341                         break;
 342                 case INET_DIAG_BC_S_GE:
 343                         yes = entry->sport >= op[1].no;
 344                         break;
 345                 case INET_DIAG_BC_S_LE:
 346                         yes = entry->dport <= op[1].no;
 347                         break;
 348                 case INET_DIAG_BC_D_GE:
 349                         yes = entry->dport >= op[1].no;
 350                         break;
 351                 case INET_DIAG_BC_D_LE:
 352                         yes = entry->dport <= op[1].no;
 353                         break;
 354                 case INET_DIAG_BC_AUTO:
 355                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 356                         break;
 357                 case INET_DIAG_BC_S_COND:
 358                 case INET_DIAG_BC_D_COND: {
 359                         struct inet_diag_hostcond *cond;
 360                         __be32 *addr;
 361
 362                         cond = (struct inet_diag_hostcond *)(op + 1);
 363                         if (cond->port != -1 &&
 364                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 365                                              entry->sport : entry->dport)) {
 366                                 yes = 0;
 367                                 break;
 368                         }
 369
 370                         if (cond->prefix_len == 0)
 371                                 break;
 372
 373                         if (op->code == INET_DIAG_BC_S_COND)
 374                                 addr = entry->saddr;
 375                         else
 376                                 addr = entry->daddr;
 377
 378                         if (bitstring_match(addr, cond->addr,
 379                                             cond->prefix_len))
 380                                 break;
 381                         if (entry->family == AF_INET6 &&
 382                             cond->family == AF_INET) {
 383                                 if (addr[0] == 0 && addr[1] == 0 &&
 384                                     addr[2] == htonl(0xffff) &&
 385                                     bitstring_match(addr + 3, cond->addr,
 386                                                     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 inet_diag_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 inet_diag_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 inet_diag_bc_op *op = (struct inet_diag_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 INET_DIAG_BC_AUTO:
 433                 case INET_DIAG_BC_S_COND:
 434                 case INET_DIAG_BC_D_COND:
 435                 case INET_DIAG_BC_S_GE:
 436                 case INET_DIAG_BC_S_LE:
 437                 case INET_DIAG_BC_D_GE:
 438                 case INET_DIAG_BC_D_LE:
 439                         if (op->yes < 4 || op->yes > len + 4)
 440                                 return -EINVAL;
 441                 case INET_DIAG_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 INET_DIAG_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 static int inet_csk_diag_dump(struct sock *sk,
 462                               struct sk_buff *skb,
 463                               struct netlink_callback *cb)
 464 {
 465         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 466
 467         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 468                 struct inet_diag_entry entry;
 469                 struct rtattr *bc = (struct rtattr *)(r + 1);
 470                 struct inet_sock *inet = inet_sk(sk);
 471
 472                 entry.family = sk->sk_family;
 473 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 474                 if (entry.family == AF_INET6) {
 475                         struct ipv6_pinfo *np = inet6_sk(sk);
 476
 477                         entry.saddr = np->rcv_saddr.s6_addr32;
 478                         entry.daddr = np->daddr.s6_addr32;
 479                 } else
 480 #endif
 481                 {
 482                         entry.saddr = &inet->rcv_saddr;
 483                         entry.daddr = &inet->daddr;
 484                 }
 485                 entry.sport = inet->num;
 486                 entry.dport = ntohs(inet->dport);
 487                 entry.userlocks = sk->sk_userlocks;
 488
 489                 if (!inet_diag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), &entry))
 490                         return 0;
 491         }
 492
 493         return inet_csk_diag_fill(sk, skb, r->idiag_ext,
 494                                   NETLINK_CB(cb->skb).pid,
 495                                   cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 496 }
 497
 498 static int inet_twsk_diag_dump(struct inet_timewait_sock *tw,
 499                                struct sk_buff *skb,
 500                                struct netlink_callback *cb)
 501 {
 502         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 503
 504         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 505                 struct inet_diag_entry entry;
 506                 struct rtattr *bc = (struct rtattr *)(r + 1);
 507
 508                 entry.family = tw->tw_family;
 509 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 510                 if (tw->tw_family == AF_INET6) {
 511                         struct inet6_timewait_sock *tw6 =
 512                                                 inet6_twsk((struct sock *)tw);
 513                         entry.saddr = tw6->tw_v6_rcv_saddr.s6_addr32;
 514                         entry.daddr = tw6->tw_v6_daddr.s6_addr32;
 515                 } else
 516 #endif
 517                 {
 518                         entry.saddr = &tw->tw_rcv_saddr;
 519                         entry.daddr = &tw->tw_daddr;
 520                 }
 521                 entry.sport = tw->tw_num;
 522                 entry.dport = ntohs(tw->tw_dport);
 523                 entry.userlocks = 0;
 524
 525                 if (!inet_diag_bc_run(RTA_DATA(bc), RTA_PAYLOAD(bc), &entry))
 526                         return 0;
 527         }
 528
 529         return inet_twsk_diag_fill(tw, skb, r->idiag_ext,
 530                                    NETLINK_CB(cb->skb).pid,
 531                                    cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 532 }
 533
 534 static int inet_diag_fill_req(struct sk_buff *skb, struct sock *sk,
 535                               struct request_sock *req, u32 pid, u32 seq,
 536                               const struct nlmsghdr *unlh)
 537 {
 538         const struct inet_request_sock *ireq = inet_rsk(req);
 539         struct inet_sock *inet = inet_sk(sk);
 540         unsigned char *b = skb_tail_pointer(skb);
 541         struct inet_diag_msg *r;
 542         struct nlmsghdr *nlh;
 543         long tmo;
 544
 545         nlh = NLMSG_PUT(skb, pid, seq, unlh->nlmsg_type, sizeof(*r));
 546         nlh->nlmsg_flags = NLM_F_MULTI;
 547         r = NLMSG_DATA(nlh);
 548
 549         r->idiag_family = sk->sk_family;
 550         r->idiag_state = TCP_SYN_RECV;
 551         r->idiag_timer = 1;
 552         r->idiag_retrans = req->retrans;
 553
 554         r->id.idiag_if = sk->sk_bound_dev_if;
 555         r->id.idiag_cookie[0] = (u32)(unsigned long)req;
 556         r->id.idiag_cookie[1] = (u32)(((unsigned long)req >> 31) >> 1);
 557
 558         tmo = req->expires - jiffies;
 559         if (tmo < 0)
 560                 tmo = 0;
 561
 562         r->id.idiag_sport = inet->sport;
 563         r->id.idiag_dport = ireq->rmt_port;
 564         r->id.idiag_src[0] = ireq->loc_addr;
 565         r->id.idiag_dst[0] = ireq->rmt_addr;
 566         r->idiag_expires = jiffies_to_msecs(tmo);
 567         r->idiag_rqueue = 0;
 568         r->idiag_wqueue = 0;
 569         r->idiag_uid = sock_i_uid(sk);
 570         r->idiag_inode = 0;
 571 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 572         if (r->idiag_family == AF_INET6) {
 573                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_src,
 574                                &inet6_rsk(req)->loc_addr);
 575                 ipv6_addr_copy((struct in6_addr *)r->id.idiag_dst,
 576                                &inet6_rsk(req)->rmt_addr);
 577         }
 578 #endif
 579         nlh->nlmsg_len = skb_tail_pointer(skb) - b;
 580
 581         return skb->len;
 582
 583 nlmsg_failure:
 584         nlmsg_trim(skb, b);
 585         return -1;
 586 }
 587
 588 static int inet_diag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 589                                struct netlink_callback *cb)
 590 {
 591         struct inet_diag_entry entry;
 592         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 593         struct inet_connection_sock *icsk = inet_csk(sk);
 594         struct listen_sock *lopt;
 595         struct rtattr *bc = NULL;
 596         struct inet_sock *inet = inet_sk(sk);
 597         int j, s_j;
 598         int reqnum, s_reqnum;
 599         int err = 0;
 600
 601         s_j = cb->args[3];
 602         s_reqnum = cb->args[4];
 603
 604         if (s_j > 0)
 605                 s_j--;
 606
 607         entry.family = sk->sk_family;
 608
 609         read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 610
 611         lopt = icsk->icsk_accept_queue.listen_opt;
 612         if (!lopt || !lopt->qlen)
 613                 goto out;
 614
 615         if (cb->nlh->nlmsg_len > 4 + NLMSG_SPACE(sizeof(*r))) {
 616                 bc = (struct rtattr *)(r + 1);
 617                 entry.sport = inet->num;
 618                 entry.userlocks = sk->sk_userlocks;
 619         }
 620
 621         for (j = s_j; j < lopt->nr_table_entries; j++) {
 622                 struct request_sock *req, *head = lopt->syn_table[j];
 623
 624                 reqnum = 0;
 625                 for (req = head; req; reqnum++, req = req->dl_next) {
 626                         struct inet_request_sock *ireq = inet_rsk(req);
 627
 628                         if (reqnum < s_reqnum)
 629                                 continue;
 630                         if (r->id.idiag_dport != ireq->rmt_port &&
 631                             r->id.idiag_dport)
 632                                 continue;
 633
 634                         if (bc) {
 635                                 entry.saddr =
 636 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 637                                         (entry.family == AF_INET6) ?
 638                                         inet6_rsk(req)->loc_addr.s6_addr32 :
 639 #endif
 640                                         &ireq->loc_addr;
 641                                 entry.daddr =
 642 #if defined(CONFIG_IPV6) || defined (CONFIG_IPV6_MODULE)
 643                                         (entry.family == AF_INET6) ?
 644                                         inet6_rsk(req)->rmt_addr.s6_addr32 :
 645 #endif
 646                                         &ireq->rmt_addr;
 647                                 entry.dport = ntohs(ireq->rmt_port);
 648
 649                                 if (!inet_diag_bc_run(RTA_DATA(bc),
 650                                                     RTA_PAYLOAD(bc), &entry))
 651                                         continue;
 652                         }
 653
 654                         err = inet_diag_fill_req(skb, sk, req,
 655                                                NETLINK_CB(cb->skb).pid,
 656                                                cb->nlh->nlmsg_seq, cb->nlh);
 657                         if (err < 0) {
 658                                 cb->args[3] = j + 1;
 659                                 cb->args[4] = reqnum;
 660                                 goto out;
 661                         }
 662                 }
 663
 664                 s_reqnum = 0;
 665         }
 666
 667 out:
 668         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 669
 670         return err;
 671 }
 672
 673 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 674 {
 675         int i, num;
 676         int s_i, s_num;
 677         struct inet_diag_req *r = NLMSG_DATA(cb->nlh);
 678         const struct inet_diag_handler *handler;
 679         struct inet_hashinfo *hashinfo;
 680
 681         handler = inet_diag_table[cb->nlh->nlmsg_type];
 682         BUG_ON(handler == NULL);
 683         hashinfo = handler->idiag_hashinfo;
 684
 685         s_i = cb->args[1];
 686         s_num = num = cb->args[2];
 687
 688         if (cb->args[0] == 0) {
 689                 if (!(r->idiag_states & (TCPF_LISTEN | TCPF_SYN_RECV)))
 690                         goto skip_listen_ht;
 691
 692                 inet_listen_lock(hashinfo);
 693                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 694                         struct sock *sk;
 695                         struct hlist_node *node;
 696
 697                         num = 0;
 698                         sk_for_each(sk, node, &hashinfo->listening_hash[i]) {
 699                                 struct inet_sock *inet = inet_sk(sk);
 700
 701                                 if (num < s_num) {
 702                                         num++;
 703                                         continue;
 704                                 }
 705
 706                                 if (r->id.idiag_sport != inet->sport &&
 707                                     r->id.idiag_sport)
 708                                         goto next_listen;
 709
 710                                 if (!(r->idiag_states & TCPF_LISTEN) ||
 711                                     r->id.idiag_dport ||
 712                                     cb->args[3] > 0)
 713                                         goto syn_recv;
 714
 715                                 if (inet_csk_diag_dump(sk, skb, cb) < 0) {
 716                                         inet_listen_unlock(hashinfo);
 717                                         goto done;
 718                                 }
 719
 720 syn_recv:
 721                                 if (!(r->idiag_states & TCPF_SYN_RECV))
 722                                         goto next_listen;
 723
 724                                 if (inet_diag_dump_reqs(skb, sk, cb) < 0) {
 725                                         inet_listen_unlock(hashinfo);
 726                                         goto done;
 727                                 }
 728
 729 next_listen:
 730                                 cb->args[3] = 0;
 731                                 cb->args[4] = 0;
 732                                 ++num;
 733                         }
 734
 735                         s_num = 0;
 736                         cb->args[3] = 0;
 737                         cb->args[4] = 0;
 738                 }
 739                 inet_listen_unlock(hashinfo);
 740 skip_listen_ht:
 741                 cb->args[0] = 1;
 742                 s_i = num = s_num = 0;
 743         }
 744
 745         if (!(r->idiag_states & ~(TCPF_LISTEN | TCPF_SYN_RECV)))
 746                 return skb->len;
 747
 748         for (i = s_i; i < hashinfo->ehash_size; i++) {
 749                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 750                 struct sock *sk;
 751                 struct hlist_node *node;
 752
 753                 if (i > s_i)
 754                         s_num = 0;
 755
 756                 read_lock_bh(&head->lock);
 757                 num = 0;
 758                 sk_for_each(sk, node, &head->chain) {
 759                         struct inet_sock *inet = inet_sk(sk);
 760
 761                         if (num < s_num)
 762                                 goto next_normal;
 763                         if (!(r->idiag_states & (1 << sk->sk_state)))
 764                                 goto next_normal;
 765                         if (r->id.idiag_sport != inet->sport &&
 766                             r->id.idiag_sport)
 767                                 goto next_normal;
 768                         if (r->id.idiag_dport != inet->dport &&
 769                             r->id.idiag_dport)
 770                                 goto next_normal;
 771                         if (inet_csk_diag_dump(sk, skb, cb) < 0) {
 772                                 read_unlock_bh(&head->lock);
 773                                 goto done;
 774                         }
 775 next_normal:
 776                         ++num;
 777                 }
 778
 779                 if (r->idiag_states & TCPF_TIME_WAIT) {
 780                         struct inet_timewait_sock *tw;
 781
 782                         inet_twsk_for_each(tw, node,
 783                                     &head->twchain) {
 784
 785                                 if (num < s_num)
 786                                         goto next_dying;
 787                                 if (r->id.idiag_sport != tw->tw_sport &&
 788                                     r->id.idiag_sport)
 789                                         goto next_dying;
 790                                 if (r->id.idiag_dport != tw->tw_dport &&
 791                                     r->id.idiag_dport)
 792                                         goto next_dying;
 793                                 if (inet_twsk_diag_dump(tw, skb, cb) < 0) {
 794                                         read_unlock_bh(&head->lock);
 795                                         goto done;
 796                                 }
 797 next_dying:
 798                                 ++num;
 799                         }
 800                 }
 801                 read_unlock_bh(&head->lock);
 802         }
 803
 804 done:
 805         cb->args[1] = i;
 806         cb->args[2] = num;
 807         return skb->len;
 808 }
 809
 810 static int inet_diag_rcv_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 811 {
 812         int hdrlen = sizeof(struct inet_diag_req);
 813
 814         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
 815             nlmsg_len(nlh) < hdrlen)
 816                 return -EINVAL;
 817
 818 #ifdef CONFIG_KMOD
 819         if (inet_diag_table[nlh->nlmsg_type] == NULL)
 820                 request_module("net-pf-%d-proto-%d-type-%d", PF_NETLINK,
 821                                NETLINK_INET_DIAG, nlh->nlmsg_type);
 822 #endif
 823
 824         if (inet_diag_table[nlh->nlmsg_type] == NULL)
 825                 return -ENOENT;
 826
 827         if (nlh->nlmsg_flags & NLM_F_DUMP) {
 828                 if (nlmsg_attrlen(nlh, hdrlen)) {
 829                         struct nlattr *attr;
 830
 831                         attr = nlmsg_find_attr(nlh, hdrlen,
 832                                                INET_DIAG_REQ_BYTECODE);
 833                         if (attr == NULL ||
 834                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
 835                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
 836                                 return -EINVAL;
 837                 }
 838
 839                 return netlink_dump_start(idiagnl, skb, nlh,
 840                                           inet_diag_dump, NULL);
 841         }
 842
 843         return inet_diag_get_exact(skb, nlh);
 844 }
 845
 846 static DEFINE_MUTEX(inet_diag_mutex);
 847
 848 static void inet_diag_rcv(struct sk_buff *skb)
 849 {
 850         mutex_lock(&inet_diag_mutex);
 851         netlink_rcv_skb(skb, &inet_diag_rcv_msg);
 852         mutex_unlock(&inet_diag_mutex);
 853 }
 854
 855 static DEFINE_SPINLOCK(inet_diag_register_lock);
 856
 857 int inet_diag_register(const struct inet_diag_handler *h)
 858 {
 859         const __u16 type = h->idiag_type;
 860         int err = -EINVAL;
 861
 862         if (type >= INET_DIAG_GETSOCK_MAX)
 863                 goto out;
 864
 865         spin_lock(&inet_diag_register_lock);
 866         err = -EEXIST;
 867         if (inet_diag_table[type] == NULL) {
 868                 inet_diag_table[type] = h;
 869                 err = 0;
 870         }
 871         spin_unlock(&inet_diag_register_lock);
 872 out:
 873         return err;
 874 }
 875 EXPORT_SYMBOL_GPL(inet_diag_register);
 876
 877 void inet_diag_unregister(const struct inet_diag_handler *h)
 878 {
 879         const __u16 type = h->idiag_type;
 880
 881         if (type >= INET_DIAG_GETSOCK_MAX)
 882                 return;
 883
 884         spin_lock(&inet_diag_register_lock);
 885         inet_diag_table[type] = NULL;
 886         spin_unlock(&inet_diag_register_lock);
 887
 888         synchronize_rcu();
 889 }
 890 EXPORT_SYMBOL_GPL(inet_diag_unregister);
 891
 892 static int __init inet_diag_init(void)
 893 {
 894         const int inet_diag_table_size = (INET_DIAG_GETSOCK_MAX *
 895                                           sizeof(struct inet_diag_handler *));
 896         int err = -ENOMEM;
 897
 898         inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
 899         if (!inet_diag_table)
 900                 goto out;
 901
 902         idiagnl = netlink_kernel_create(&init_net, NETLINK_INET_DIAG, 0,
 903                                         inet_diag_rcv, NULL, THIS_MODULE);
 904         if (idiagnl == NULL)
 905                 goto out_free_table;
 906         err = 0;
 907 out:
 908         return err;
 909 out_free_table:
 910         kfree(inet_diag_table);
 911         goto out;
 912 }
 913
 914 static void __exit inet_diag_exit(void)
 915 {
 916         sock_release(idiagnl->sk_socket);
 917         kfree(inet_diag_table);
 918 }
 919
 920 module_init(inet_diag_init);
 921 module_exit(inet_diag_exit);
 922 MODULE_LICENSE("GPL");
 923 MODULE_ALIAS_NET_PF_PROTO(PF_NETLINK, NETLINK_INET_DIAG);