net/ipv4/inet_diag.c

   1 /*
   2  * inet_diag.c  Module for monitoring INET transport protocols sockets.
   3  *
   4  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
   5  *
   6  *      This program is free software; you can redistribute it and/or
   7  *      modify it under the terms of the GNU General Public License
   8  *      as published by the Free Software Foundation; either version
   9  *      2 of the License, or (at your option) any later version.
  10  */
  11
  12 #include <linux/kernel.h>
  13 #include <linux/module.h>
  14 #include <linux/types.h>
  15 #include <linux/fcntl.h>
  16 #include <linux/random.h>
  17 #include <linux/slab.h>
  18 #include <linux/cache.h>
  19 #include <linux/init.h>
  20 #include <linux/time.h>
  21
  22 #include <net/icmp.h>
  23 #include <net/tcp.h>
  24 #include <net/ipv6.h>
  25 #include <net/inet_common.h>
  26 #include <net/inet_connection_sock.h>
  27 #include <net/inet_hashtables.h>
  28 #include <net/inet_timewait_sock.h>
  29 #include <net/inet6_hashtables.h>
  30 #include <net/netlink.h>
  31
  32 #include <linux/inet.h>
  33 #include <linux/stddef.h>
  34
  35 #include <linux/inet_diag.h>
  36 #include <linux/sock_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 #if IS_ENABLED(CONFIG_IPV6)
  48         struct in6_addr saddr_storage;  /* for IPv4-mapped-IPv6 addresses */
  49         struct in6_addr daddr_storage;  /* for IPv4-mapped-IPv6 addresses */
  50 #endif
  51 };
  52
  53 static DEFINE_MUTEX(inet_diag_table_mutex);
  54
  55 static const struct inet_diag_handler *inet_diag_lock_handler(int proto)
  56 {
  57         if (!inet_diag_table[proto])
  58                 request_module("net-pf-%d-proto-%d-type-%d-%d", PF_NETLINK,
  59                                NETLINK_SOCK_DIAG, AF_INET, proto);
  60
  61         mutex_lock(&inet_diag_table_mutex);
  62         if (!inet_diag_table[proto])
  63                 return ERR_PTR(-ENOENT);
  64
  65         return inet_diag_table[proto];
  66 }
  67
  68 static inline void inet_diag_unlock_handler(
  69         const struct inet_diag_handler *handler)
  70 {
  71         mutex_unlock(&inet_diag_table_mutex);
  72 }
  73
  74 int inet_sk_diag_fill(struct sock *sk, struct inet_connection_sock *icsk,
  75                               struct sk_buff *skb, struct inet_diag_req_v2 *req,
  76                               struct user_namespace *user_ns,
  77                               u32 portid, u32 seq, u16 nlmsg_flags,
  78                               const struct nlmsghdr *unlh)
  79 {
  80         const struct inet_sock *inet = inet_sk(sk);
  81         struct inet_diag_msg *r;
  82         struct nlmsghdr  *nlh;
  83         struct nlattr *attr;
  84         void *info = NULL;
  85         const struct inet_diag_handler *handler;
  86         int ext = req->idiag_ext;
  87
  88         handler = inet_diag_table[req->sdiag_protocol];
  89         BUG_ON(handler == NULL);
  90
  91         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
  92                         nlmsg_flags);
  93         if (!nlh)
  94                 return -EMSGSIZE;
  95
  96         r = nlmsg_data(nlh);
  97         BUG_ON(sk->sk_state == TCP_TIME_WAIT);
  98
  99         r->idiag_family = sk->sk_family;
 100         r->idiag_state = sk->sk_state;
 101         r->idiag_timer = 0;
 102         r->idiag_retrans = 0;
 103
 104         r->id.idiag_if = sk->sk_bound_dev_if;
 105         sock_diag_save_cookie(sk, r->id.idiag_cookie);
 106
 107         r->id.idiag_sport = inet->inet_sport;
 108         r->id.idiag_dport = inet->inet_dport;
 109         r->id.idiag_src[0] = inet->inet_rcv_saddr;
 110         r->id.idiag_dst[0] = inet->inet_daddr;
 111
 112         /* IPv6 dual-stack sockets use inet->tos for IPv4 connections,
 113          * hence this needs to be included regardless of socket family.
 114          */
 115         if (ext & (1 << (INET_DIAG_TOS - 1)))
 116                 if (nla_put_u8(skb, INET_DIAG_TOS, inet->tos) < 0)
 117                         goto errout;
 118
 119 #if IS_ENABLED(CONFIG_IPV6)
 120         if (r->idiag_family == AF_INET6) {
 121                 const struct ipv6_pinfo *np = inet6_sk(sk);
 122
 123                 *(struct in6_addr *)r->id.idiag_src = np->rcv_saddr;
 124                 *(struct in6_addr *)r->id.idiag_dst = np->daddr;
 125
 126                 if (ext & (1 << (INET_DIAG_TCLASS - 1)))
 127                         if (nla_put_u8(skb, INET_DIAG_TCLASS, np->tclass) < 0)
 128                                 goto errout;
 129         }
 130 #endif
 131
 132         r->idiag_uid = from_kuid_munged(user_ns, sock_i_uid(sk));
 133         r->idiag_inode = sock_i_ino(sk);
 134
 135         if (ext & (1 << (INET_DIAG_MEMINFO - 1))) {
 136                 struct inet_diag_meminfo minfo = {
 137                         .idiag_rmem = sk_rmem_alloc_get(sk),
 138                         .idiag_wmem = sk->sk_wmem_queued,
 139                         .idiag_fmem = sk->sk_forward_alloc,
 140                         .idiag_tmem = sk_wmem_alloc_get(sk),
 141                 };
 142
 143                 if (nla_put(skb, INET_DIAG_MEMINFO, sizeof(minfo), &minfo) < 0)
 144                         goto errout;
 145         }
 146
 147         if (ext & (1 << (INET_DIAG_SKMEMINFO - 1)))
 148                 if (sock_diag_put_meminfo(sk, skb, INET_DIAG_SKMEMINFO))
 149                         goto errout;
 150
 151         if (icsk == NULL) {
 152                 handler->idiag_get_info(sk, r, NULL);
 153                 goto out;
 154         }
 155
 156 #define EXPIRES_IN_MS(tmo)  DIV_ROUND_UP((tmo - jiffies) * 1000, HZ)
 157
 158         if (icsk->icsk_pending == ICSK_TIME_RETRANS) {
 159                 r->idiag_timer = 1;
 160                 r->idiag_retrans = icsk->icsk_retransmits;
 161                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 162         } else if (icsk->icsk_pending == ICSK_TIME_PROBE0) {
 163                 r->idiag_timer = 4;
 164                 r->idiag_retrans = icsk->icsk_probes_out;
 165                 r->idiag_expires = EXPIRES_IN_MS(icsk->icsk_timeout);
 166         } else if (timer_pending(&sk->sk_timer)) {
 167                 r->idiag_timer = 2;
 168                 r->idiag_retrans = icsk->icsk_probes_out;
 169                 r->idiag_expires = EXPIRES_IN_MS(sk->sk_timer.expires);
 170         } else {
 171                 r->idiag_timer = 0;
 172                 r->idiag_expires = 0;
 173         }
 174 #undef EXPIRES_IN_MS
 175
 176         if (ext & (1 << (INET_DIAG_INFO - 1))) {
 177                 attr = nla_reserve(skb, INET_DIAG_INFO,
 178                                    sizeof(struct tcp_info));
 179                 if (!attr)
 180                         goto errout;
 181
 182                 info = nla_data(attr);
 183         }
 184
 185         if ((ext & (1 << (INET_DIAG_CONG - 1))) && icsk->icsk_ca_ops)
 186                 if (nla_put_string(skb, INET_DIAG_CONG,
 187                                    icsk->icsk_ca_ops->name) < 0)
 188                         goto errout;
 189
 190         handler->idiag_get_info(sk, r, info);
 191
 192         if (sk->sk_state < TCP_TIME_WAIT &&
 193             icsk->icsk_ca_ops && icsk->icsk_ca_ops->get_info)
 194                 icsk->icsk_ca_ops->get_info(sk, ext, skb);
 195
 196 out:
 197         return nlmsg_end(skb, nlh);
 198
 199 errout:
 200         nlmsg_cancel(skb, nlh);
 201         return -EMSGSIZE;
 202 }
 203 EXPORT_SYMBOL_GPL(inet_sk_diag_fill);
 204
 205 static int inet_csk_diag_fill(struct sock *sk,
 206                               struct sk_buff *skb, struct inet_diag_req_v2 *req,
 207                               struct user_namespace *user_ns,
 208                               u32 portid, u32 seq, u16 nlmsg_flags,
 209                               const struct nlmsghdr *unlh)
 210 {
 211         return inet_sk_diag_fill(sk, inet_csk(sk),
 212                         skb, req, user_ns, portid, seq, nlmsg_flags, unlh);
 213 }
 214
 215 static int inet_twsk_diag_fill(struct inet_timewait_sock *tw,
 216                                struct sk_buff *skb, struct inet_diag_req_v2 *req,
 217                                u32 portid, u32 seq, u16 nlmsg_flags,
 218                                const struct nlmsghdr *unlh)
 219 {
 220         long tmo;
 221         struct inet_diag_msg *r;
 222         struct nlmsghdr *nlh;
 223
 224         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
 225                         nlmsg_flags);
 226         if (!nlh)
 227                 return -EMSGSIZE;
 228
 229         r = nlmsg_data(nlh);
 230         BUG_ON(tw->tw_state != TCP_TIME_WAIT);
 231
 232         tmo = tw->tw_ttd - jiffies;
 233         if (tmo < 0)
 234                 tmo = 0;
 235
 236         r->idiag_family       = tw->tw_family;
 237         r->idiag_retrans      = 0;
 238         r->id.idiag_if        = tw->tw_bound_dev_if;
 239         sock_diag_save_cookie(tw, r->id.idiag_cookie);
 240         r->id.idiag_sport     = tw->tw_sport;
 241         r->id.idiag_dport     = tw->tw_dport;
 242         r->id.idiag_src[0]    = tw->tw_rcv_saddr;
 243         r->id.idiag_dst[0]    = tw->tw_daddr;
 244         r->idiag_state        = tw->tw_substate;
 245         r->idiag_timer        = 3;
 246         r->idiag_expires      = DIV_ROUND_UP(tmo * 1000, HZ);
 247         r->idiag_rqueue       = 0;
 248         r->idiag_wqueue       = 0;
 249         r->idiag_uid          = 0;
 250         r->idiag_inode        = 0;
 251 #if IS_ENABLED(CONFIG_IPV6)
 252         if (tw->tw_family == AF_INET6) {
 253                 const struct inet6_timewait_sock *tw6 =
 254                                                 inet6_twsk((struct sock *)tw);
 255
 256                 *(struct in6_addr *)r->id.idiag_src = tw6->tw_v6_rcv_saddr;
 257                 *(struct in6_addr *)r->id.idiag_dst = tw6->tw_v6_daddr;
 258         }
 259 #endif
 260
 261         return nlmsg_end(skb, nlh);
 262 }
 263
 264 static int sk_diag_fill(struct sock *sk, struct sk_buff *skb,
 265                         struct inet_diag_req_v2 *r,
 266                         struct user_namespace *user_ns,
 267                         u32 portid, u32 seq, u16 nlmsg_flags,
 268                         const struct nlmsghdr *unlh)
 269 {
 270         if (sk->sk_state == TCP_TIME_WAIT)
 271                 return inet_twsk_diag_fill((struct inet_timewait_sock *)sk,
 272                                            skb, r, portid, seq, nlmsg_flags,
 273                                            unlh);
 274         return inet_csk_diag_fill(sk, skb, r, user_ns, portid, seq, nlmsg_flags, unlh);
 275 }
 276
 277 int inet_diag_dump_one_icsk(struct inet_hashinfo *hashinfo, struct sk_buff *in_skb,
 278                 const struct nlmsghdr *nlh, struct inet_diag_req_v2 *req)
 279 {
 280         int err;
 281         struct sock *sk;
 282         struct sk_buff *rep;
 283         struct net *net = sock_net(in_skb->sk);
 284
 285         err = -EINVAL;
 286         if (req->sdiag_family == AF_INET) {
 287                 sk = inet_lookup(net, hashinfo, req->id.idiag_dst[0],
 288                                  req->id.idiag_dport, req->id.idiag_src[0],
 289                                  req->id.idiag_sport, req->id.idiag_if);
 290         }
 291 #if IS_ENABLED(CONFIG_IPV6)
 292         else if (req->sdiag_family == AF_INET6) {
 293                 sk = inet6_lookup(net, hashinfo,
 294                                   (struct in6_addr *)req->id.idiag_dst,
 295                                   req->id.idiag_dport,
 296                                   (struct in6_addr *)req->id.idiag_src,
 297                                   req->id.idiag_sport,
 298                                   req->id.idiag_if);
 299         }
 300 #endif
 301         else {
 302                 goto out_nosk;
 303         }
 304
 305         err = -ENOENT;
 306         if (sk == NULL)
 307                 goto out_nosk;
 308
 309         err = sock_diag_check_cookie(sk, req->id.idiag_cookie);
 310         if (err)
 311                 goto out;
 312
 313         rep = nlmsg_new(sizeof(struct inet_diag_msg) +
 314                         sizeof(struct inet_diag_meminfo) +
 315                         sizeof(struct tcp_info) + 64, GFP_KERNEL);
 316         if (!rep) {
 317                 err = -ENOMEM;
 318                 goto out;
 319         }
 320
 321         err = sk_diag_fill(sk, rep, req,
 322                            sk_user_ns(NETLINK_CB(in_skb).ssk),
 323                            NETLINK_CB(in_skb).portid,
 324                            nlh->nlmsg_seq, 0, nlh);
 325         if (err < 0) {
 326                 WARN_ON(err == -EMSGSIZE);
 327                 nlmsg_free(rep);
 328                 goto out;
 329         }
 330         err = netlink_unicast(net->diag_nlsk, rep, NETLINK_CB(in_skb).portid,
 331                               MSG_DONTWAIT);
 332         if (err > 0)
 333                 err = 0;
 334
 335 out:
 336         if (sk) {
 337                 if (sk->sk_state == TCP_TIME_WAIT)
 338                         inet_twsk_put((struct inet_timewait_sock *)sk);
 339                 else
 340                         sock_put(sk);
 341         }
 342 out_nosk:
 343         return err;
 344 }
 345 EXPORT_SYMBOL_GPL(inet_diag_dump_one_icsk);
 346
 347 static int inet_diag_get_exact(struct sk_buff *in_skb,
 348                                const struct nlmsghdr *nlh,
 349                                struct inet_diag_req_v2 *req)
 350 {
 351         const struct inet_diag_handler *handler;
 352         int err;
 353
 354         handler = inet_diag_lock_handler(req->sdiag_protocol);
 355         if (IS_ERR(handler))
 356                 err = PTR_ERR(handler);
 357         else
 358                 err = handler->dump_one(in_skb, nlh, req);
 359         inet_diag_unlock_handler(handler);
 360
 361         return err;
 362 }
 363
 364 static int bitstring_match(const __be32 *a1, const __be32 *a2, int bits)
 365 {
 366         int words = bits >> 5;
 367
 368         bits &= 0x1f;
 369
 370         if (words) {
 371                 if (memcmp(a1, a2, words << 2))
 372                         return 0;
 373         }
 374         if (bits) {
 375                 __be32 w1, w2;
 376                 __be32 mask;
 377
 378                 w1 = a1[words];
 379                 w2 = a2[words];
 380
 381                 mask = htonl((0xffffffff) << (32 - bits));
 382
 383                 if ((w1 ^ w2) & mask)
 384                         return 0;
 385         }
 386
 387         return 1;
 388 }
 389
 390
 391 static int inet_diag_bc_run(const struct nlattr *_bc,
 392                 const struct inet_diag_entry *entry)
 393 {
 394         const void *bc = nla_data(_bc);
 395         int len = nla_len(_bc);
 396
 397         while (len > 0) {
 398                 int yes = 1;
 399                 const struct inet_diag_bc_op *op = bc;
 400
 401                 switch (op->code) {
 402                 case INET_DIAG_BC_NOP:
 403                         break;
 404                 case INET_DIAG_BC_JMP:
 405                         yes = 0;
 406                         break;
 407                 case INET_DIAG_BC_S_GE:
 408                         yes = entry->sport >= op[1].no;
 409                         break;
 410                 case INET_DIAG_BC_S_LE:
 411                         yes = entry->sport <= op[1].no;
 412                         break;
 413                 case INET_DIAG_BC_D_GE:
 414                         yes = entry->dport >= op[1].no;
 415                         break;
 416                 case INET_DIAG_BC_D_LE:
 417                         yes = entry->dport <= op[1].no;
 418                         break;
 419                 case INET_DIAG_BC_AUTO:
 420                         yes = !(entry->userlocks & SOCK_BINDPORT_LOCK);
 421                         break;
 422                 case INET_DIAG_BC_S_COND:
 423                 case INET_DIAG_BC_D_COND: {
 424                         struct inet_diag_hostcond *cond;
 425                         __be32 *addr;
 426
 427                         cond = (struct inet_diag_hostcond *)(op + 1);
 428                         if (cond->port != -1 &&
 429                             cond->port != (op->code == INET_DIAG_BC_S_COND ?
 430                                              entry->sport : entry->dport)) {
 431                                 yes = 0;
 432                                 break;
 433                         }
 434
 435                         if (cond->prefix_len == 0)
 436                                 break;
 437
 438                         if (op->code == INET_DIAG_BC_S_COND)
 439                                 addr = entry->saddr;
 440                         else
 441                                 addr = entry->daddr;
 442
 443                         if (bitstring_match(addr, cond->addr,
 444                                             cond->prefix_len))
 445                                 break;
 446                         if (entry->family == AF_INET6 &&
 447                             cond->family == AF_INET) {
 448                                 if (addr[0] == 0 && addr[1] == 0 &&
 449                                     addr[2] == htonl(0xffff) &&
 450                                     bitstring_match(addr + 3, cond->addr,
 451                                                     cond->prefix_len))
 452                                         break;
 453                         }
 454                         yes = 0;
 455                         break;
 456                 }
 457                 }
 458
 459                 if (yes) {
 460                         len -= op->yes;
 461                         bc += op->yes;
 462                 } else {
 463                         len -= op->no;
 464                         bc += op->no;
 465                 }
 466         }
 467         return len == 0;
 468 }
 469
 470 int inet_diag_bc_sk(const struct nlattr *bc, struct sock *sk)
 471 {
 472         struct inet_diag_entry entry;
 473         struct inet_sock *inet = inet_sk(sk);
 474
 475         if (bc == NULL)
 476                 return 1;
 477
 478         entry.family = sk->sk_family;
 479 #if IS_ENABLED(CONFIG_IPV6)
 480         if (entry.family == AF_INET6) {
 481                 struct ipv6_pinfo *np = inet6_sk(sk);
 482
 483                 entry.saddr = np->rcv_saddr.s6_addr32;
 484                 entry.daddr = np->daddr.s6_addr32;
 485         } else
 486 #endif
 487         {
 488                 entry.saddr = &inet->inet_rcv_saddr;
 489                 entry.daddr = &inet->inet_daddr;
 490         }
 491         entry.sport = inet->inet_num;
 492         entry.dport = ntohs(inet->inet_dport);
 493         entry.userlocks = sk->sk_userlocks;
 494
 495         return inet_diag_bc_run(bc, &entry);
 496 }
 497 EXPORT_SYMBOL_GPL(inet_diag_bc_sk);
 498
 499 static int valid_cc(const void *bc, int len, int cc)
 500 {
 501         while (len >= 0) {
 502                 const struct inet_diag_bc_op *op = bc;
 503
 504                 if (cc > len)
 505                         return 0;
 506                 if (cc == len)
 507                         return 1;
 508                 if (op->yes < 4 || op->yes & 3)
 509                         return 0;
 510                 len -= op->yes;
 511                 bc  += op->yes;
 512         }
 513         return 0;
 514 }
 515
 516 /* Validate an inet_diag_hostcond. */
 517 static bool valid_hostcond(const struct inet_diag_bc_op *op, int len,
 518                            int *min_len)
 519 {
 520         int addr_len;
 521         struct inet_diag_hostcond *cond;
 522
 523         /* Check hostcond space. */
 524         *min_len += sizeof(struct inet_diag_hostcond);
 525         if (len < *min_len)
 526                 return false;
 527         cond = (struct inet_diag_hostcond *)(op + 1);
 528
 529         /* Check address family and address length. */
 530         switch (cond->family) {
 531         case AF_UNSPEC:
 532                 addr_len = 0;
 533                 break;
 534         case AF_INET:
 535                 addr_len = sizeof(struct in_addr);
 536                 break;
 537         case AF_INET6:
 538                 addr_len = sizeof(struct in6_addr);
 539                 break;
 540         default:
 541                 return false;
 542         }
 543         *min_len += addr_len;
 544         if (len < *min_len)
 545                 return false;
 546
 547         /* Check prefix length (in bits) vs address length (in bytes). */
 548         if (cond->prefix_len > 8 * addr_len)
 549                 return false;
 550
 551         return true;
 552 }
 553
 554 static int inet_diag_bc_audit(const void *bytecode, int bytecode_len)
 555 {
 556         const void *bc = bytecode;
 557         int  len = bytecode_len;
 558
 559         while (len > 0) {
 560                 const struct inet_diag_bc_op *op = bc;
 561                 int min_len = sizeof(struct inet_diag_bc_op);
 562
 563 //printk("BC: %d %d %d {%d} / %d\n", op->code, op->yes, op->no, op[1].no, len);
 564                 switch (op->code) {
 565                 case INET_DIAG_BC_S_COND:
 566                 case INET_DIAG_BC_D_COND:
 567                         if (!valid_hostcond(bc, len, &min_len))
 568                                 return -EINVAL;
 569                         /* fall through */
 570                 case INET_DIAG_BC_AUTO:
 571                 case INET_DIAG_BC_S_GE:
 572                 case INET_DIAG_BC_S_LE:
 573                 case INET_DIAG_BC_D_GE:
 574                 case INET_DIAG_BC_D_LE:
 575                 case INET_DIAG_BC_JMP:
 576                         if (op->no < min_len || op->no > len + 4 || op->no & 3)
 577                                 return -EINVAL;
 578                         if (op->no < len &&
 579                             !valid_cc(bytecode, bytecode_len, len - op->no))
 580                                 return -EINVAL;
 581                         break;
 582                 case INET_DIAG_BC_NOP:
 583                         break;
 584                 default:
 585                         return -EINVAL;
 586                 }
 587                 if (op->yes < min_len || op->yes > len + 4 || op->yes & 3)
 588                         return -EINVAL;
 589                 bc  += op->yes;
 590                 len -= op->yes;
 591         }
 592         return len == 0 ? 0 : -EINVAL;
 593 }
 594
 595 static int inet_csk_diag_dump(struct sock *sk,
 596                               struct sk_buff *skb,
 597                               struct netlink_callback *cb,
 598                               struct inet_diag_req_v2 *r,
 599                               const struct nlattr *bc)
 600 {
 601         if (!inet_diag_bc_sk(bc, sk))
 602                 return 0;
 603
 604         return inet_csk_diag_fill(sk, skb, r,
 605                                   sk_user_ns(NETLINK_CB(cb->skb).ssk),
 606                                   NETLINK_CB(cb->skb).portid,
 607                                   cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 608 }
 609
 610 static int inet_twsk_diag_dump(struct inet_timewait_sock *tw,
 611                                struct sk_buff *skb,
 612                                struct netlink_callback *cb,
 613                                struct inet_diag_req_v2 *r,
 614                                const struct nlattr *bc)
 615 {
 616         if (bc != NULL) {
 617                 struct inet_diag_entry entry;
 618
 619                 entry.family = tw->tw_family;
 620 #if IS_ENABLED(CONFIG_IPV6)
 621                 if (tw->tw_family == AF_INET6) {
 622                         struct inet6_timewait_sock *tw6 =
 623                                                 inet6_twsk((struct sock *)tw);
 624                         entry.saddr = tw6->tw_v6_rcv_saddr.s6_addr32;
 625                         entry.daddr = tw6->tw_v6_daddr.s6_addr32;
 626                 } else
 627 #endif
 628                 {
 629                         entry.saddr = &tw->tw_rcv_saddr;
 630                         entry.daddr = &tw->tw_daddr;
 631                 }
 632                 entry.sport = tw->tw_num;
 633                 entry.dport = ntohs(tw->tw_dport);
 634                 entry.userlocks = 0;
 635
 636                 if (!inet_diag_bc_run(bc, &entry))
 637                         return 0;
 638         }
 639
 640         return inet_twsk_diag_fill(tw, skb, r,
 641                                    NETLINK_CB(cb->skb).portid,
 642                                    cb->nlh->nlmsg_seq, NLM_F_MULTI, cb->nlh);
 643 }
 644
 645 /* Get the IPv4, IPv6, or IPv4-mapped-IPv6 local and remote addresses
 646  * from a request_sock. For IPv4-mapped-IPv6 we must map IPv4 to IPv6.
 647  */
 648 static inline void inet_diag_req_addrs(const struct sock *sk,
 649                                        const struct request_sock *req,
 650                                        struct inet_diag_entry *entry)
 651 {
 652         struct inet_request_sock *ireq = inet_rsk(req);
 653
 654 #if IS_ENABLED(CONFIG_IPV6)
 655         if (sk->sk_family == AF_INET6) {
 656                 if (req->rsk_ops->family == AF_INET6) {
 657                         entry->saddr = inet6_rsk(req)->loc_addr.s6_addr32;
 658                         entry->daddr = inet6_rsk(req)->rmt_addr.s6_addr32;
 659                 } else if (req->rsk_ops->family == AF_INET) {
 660                         ipv6_addr_set_v4mapped(ireq->loc_addr,
 661                                                &entry->saddr_storage);
 662                         ipv6_addr_set_v4mapped(ireq->rmt_addr,
 663                                                &entry->daddr_storage);
 664                         entry->saddr = entry->saddr_storage.s6_addr32;
 665                         entry->daddr = entry->daddr_storage.s6_addr32;
 666                 }
 667         } else
 668 #endif
 669         {
 670                 entry->saddr = &ireq->loc_addr;
 671                 entry->daddr = &ireq->rmt_addr;
 672         }
 673 }
 674
 675 static int inet_diag_fill_req(struct sk_buff *skb, struct sock *sk,
 676                               struct request_sock *req,
 677                               struct user_namespace *user_ns,
 678                               u32 portid, u32 seq,
 679                               const struct nlmsghdr *unlh)
 680 {
 681         const struct inet_request_sock *ireq = inet_rsk(req);
 682         struct inet_sock *inet = inet_sk(sk);
 683         struct inet_diag_msg *r;
 684         struct nlmsghdr *nlh;
 685         long tmo;
 686
 687         nlh = nlmsg_put(skb, portid, seq, unlh->nlmsg_type, sizeof(*r),
 688                         NLM_F_MULTI);
 689         if (!nlh)
 690                 return -EMSGSIZE;
 691
 692         r = nlmsg_data(nlh);
 693         r->idiag_family = sk->sk_family;
 694         r->idiag_state = TCP_SYN_RECV;
 695         r->idiag_timer = 1;
 696         r->idiag_retrans = req->retrans;
 697
 698         r->id.idiag_if = sk->sk_bound_dev_if;
 699         sock_diag_save_cookie(req, r->id.idiag_cookie);
 700
 701         tmo = req->expires - jiffies;
 702         if (tmo < 0)
 703                 tmo = 0;
 704
 705         r->id.idiag_sport = inet->inet_sport;
 706         r->id.idiag_dport = ireq->rmt_port;
 707         r->id.idiag_src[0] = ireq->loc_addr;
 708         r->id.idiag_dst[0] = ireq->rmt_addr;
 709         r->idiag_expires = jiffies_to_msecs(tmo);
 710         r->idiag_rqueue = 0;
 711         r->idiag_wqueue = 0;
 712         r->idiag_uid = from_kuid_munged(user_ns, sock_i_uid(sk));
 713         r->idiag_inode = 0;
 714 #if IS_ENABLED(CONFIG_IPV6)
 715         if (r->idiag_family == AF_INET6) {
 716                 struct inet_diag_entry entry;
 717                 inet_diag_req_addrs(sk, req, &entry);
 718                 memcpy(r->id.idiag_src, entry.saddr, sizeof(struct in6_addr));
 719                 memcpy(r->id.idiag_dst, entry.daddr, sizeof(struct in6_addr));
 720         }
 721 #endif
 722
 723         return nlmsg_end(skb, nlh);
 724 }
 725
 726 static int inet_diag_dump_reqs(struct sk_buff *skb, struct sock *sk,
 727                                struct netlink_callback *cb,
 728                                struct inet_diag_req_v2 *r,
 729                                const struct nlattr *bc)
 730 {
 731         struct inet_diag_entry entry;
 732         struct inet_connection_sock *icsk = inet_csk(sk);
 733         struct listen_sock *lopt;
 734         struct inet_sock *inet = inet_sk(sk);
 735         int j, s_j;
 736         int reqnum, s_reqnum;
 737         int err = 0;
 738
 739         s_j = cb->args[3];
 740         s_reqnum = cb->args[4];
 741
 742         if (s_j > 0)
 743                 s_j--;
 744
 745         entry.family = sk->sk_family;
 746
 747         read_lock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 748
 749         lopt = icsk->icsk_accept_queue.listen_opt;
 750         if (!lopt || !lopt->qlen)
 751                 goto out;
 752
 753         if (bc != NULL) {
 754                 entry.sport = inet->inet_num;
 755                 entry.userlocks = sk->sk_userlocks;
 756         }
 757
 758         for (j = s_j; j < lopt->nr_table_entries; j++) {
 759                 struct request_sock *req, *head = lopt->syn_table[j];
 760
 761                 reqnum = 0;
 762                 for (req = head; req; reqnum++, req = req->dl_next) {
 763                         struct inet_request_sock *ireq = inet_rsk(req);
 764
 765                         if (reqnum < s_reqnum)
 766                                 continue;
 767                         if (r->id.idiag_dport != ireq->rmt_port &&
 768                             r->id.idiag_dport)
 769                                 continue;
 770
 771                         if (bc) {
 772                                 inet_diag_req_addrs(sk, req, &entry);
 773                                 entry.dport = ntohs(ireq->rmt_port);
 774
 775                                 if (!inet_diag_bc_run(bc, &entry))
 776                                         continue;
 777                         }
 778
 779                         err = inet_diag_fill_req(skb, sk, req,
 780                                                sk_user_ns(NETLINK_CB(cb->skb).ssk),
 781                                                NETLINK_CB(cb->skb).portid,
 782                                                cb->nlh->nlmsg_seq, cb->nlh);
 783                         if (err < 0) {
 784                                 cb->args[3] = j + 1;
 785                                 cb->args[4] = reqnum;
 786                                 goto out;
 787                         }
 788                 }
 789
 790                 s_reqnum = 0;
 791         }
 792
 793 out:
 794         read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 795
 796         return err;
 797 }
 798
 799 void inet_diag_dump_icsk(struct inet_hashinfo *hashinfo, struct sk_buff *skb,
 800                 struct netlink_callback *cb, struct inet_diag_req_v2 *r, struct nlattr *bc)
 801 {
 802         int i, num;
 803         int s_i, s_num;
 804         struct net *net = sock_net(skb->sk);
 805
 806         s_i = cb->args[1];
 807         s_num = num = cb->args[2];
 808
 809         if (cb->args[0] == 0) {
 810                 if (!(r->idiag_states & (TCPF_LISTEN | TCPF_SYN_RECV)))
 811                         goto skip_listen_ht;
 812
 813                 for (i = s_i; i < INET_LHTABLE_SIZE; i++) {
 814                         struct sock *sk;
 815                         struct hlist_nulls_node *node;
 816                         struct inet_listen_hashbucket *ilb;
 817
 818                         num = 0;
 819                         ilb = &hashinfo->listening_hash[i];
 820                         spin_lock_bh(&ilb->lock);
 821                         sk_nulls_for_each(sk, node, &ilb->head) {
 822                                 struct inet_sock *inet = inet_sk(sk);
 823
 824                                 if (!net_eq(sock_net(sk), net))
 825                                         continue;
 826
 827                                 if (num < s_num) {
 828                                         num++;
 829                                         continue;
 830                                 }
 831
 832                                 if (r->sdiag_family != AF_UNSPEC &&
 833                                                 sk->sk_family != r->sdiag_family)
 834                                         goto next_listen;
 835
 836                                 if (r->id.idiag_sport != inet->inet_sport &&
 837                                     r->id.idiag_sport)
 838                                         goto next_listen;
 839
 840                                 if (!(r->idiag_states & TCPF_LISTEN) ||
 841                                     r->id.idiag_dport ||
 842                                     cb->args[3] > 0)
 843                                         goto syn_recv;
 844
 845                                 if (inet_csk_diag_dump(sk, skb, cb, r, bc) < 0) {
 846                                         spin_unlock_bh(&ilb->lock);
 847                                         goto done;
 848                                 }
 849
 850 syn_recv:
 851                                 if (!(r->idiag_states & TCPF_SYN_RECV))
 852                                         goto next_listen;
 853
 854                                 if (inet_diag_dump_reqs(skb, sk, cb, r, bc) < 0) {
 855                                         spin_unlock_bh(&ilb->lock);
 856                                         goto done;
 857                                 }
 858
 859 next_listen:
 860                                 cb->args[3] = 0;
 861                                 cb->args[4] = 0;
 862                                 ++num;
 863                         }
 864                         spin_unlock_bh(&ilb->lock);
 865
 866                         s_num = 0;
 867                         cb->args[3] = 0;
 868                         cb->args[4] = 0;
 869                 }
 870 skip_listen_ht:
 871                 cb->args[0] = 1;
 872                 s_i = num = s_num = 0;
 873         }
 874
 875         if (!(r->idiag_states & ~(TCPF_LISTEN | TCPF_SYN_RECV)))
 876                 goto out;
 877
 878         for (i = s_i; i <= hashinfo->ehash_mask; i++) {
 879                 struct inet_ehash_bucket *head = &hashinfo->ehash[i];
 880                 spinlock_t *lock = inet_ehash_lockp(hashinfo, i);
 881                 struct sock *sk;
 882                 struct hlist_nulls_node *node;
 883
 884                 num = 0;
 885
 886                 if (hlist_nulls_empty(&head->chain) &&
 887                         hlist_nulls_empty(&head->twchain))
 888                         continue;
 889
 890                 if (i > s_i)
 891                         s_num = 0;
 892
 893                 spin_lock_bh(lock);
 894                 sk_nulls_for_each(sk, node, &head->chain) {
 895                         struct inet_sock *inet = inet_sk(sk);
 896
 897                         if (!net_eq(sock_net(sk), net))
 898                                 continue;
 899                         if (num < s_num)
 900                                 goto next_normal;
 901                         if (!(r->idiag_states & (1 << sk->sk_state)))
 902                                 goto next_normal;
 903                         if (r->sdiag_family != AF_UNSPEC &&
 904                                         sk->sk_family != r->sdiag_family)
 905                                 goto next_normal;
 906                         if (r->id.idiag_sport != inet->inet_sport &&
 907                             r->id.idiag_sport)
 908                                 goto next_normal;
 909                         if (r->id.idiag_dport != inet->inet_dport &&
 910                             r->id.idiag_dport)
 911                                 goto next_normal;
 912                         if (inet_csk_diag_dump(sk, skb, cb, r, bc) < 0) {
 913                                 spin_unlock_bh(lock);
 914                                 goto done;
 915                         }
 916 next_normal:
 917                         ++num;
 918                 }
 919
 920                 if (r->idiag_states & TCPF_TIME_WAIT) {
 921                         struct inet_timewait_sock *tw;
 922
 923                         inet_twsk_for_each(tw, node,
 924                                     &head->twchain) {
 925                                 if (!net_eq(twsk_net(tw), net))
 926                                         continue;
 927
 928                                 if (num < s_num)
 929                                         goto next_dying;
 930                                 if (r->sdiag_family != AF_UNSPEC &&
 931                                                 tw->tw_family != r->sdiag_family)
 932                                         goto next_dying;
 933                                 if (r->id.idiag_sport != tw->tw_sport &&
 934                                     r->id.idiag_sport)
 935                                         goto next_dying;
 936                                 if (r->id.idiag_dport != tw->tw_dport &&
 937                                     r->id.idiag_dport)
 938                                         goto next_dying;
 939                                 if (inet_twsk_diag_dump(tw, skb, cb, r, bc) < 0) {
 940                                         spin_unlock_bh(lock);
 941                                         goto done;
 942                                 }
 943 next_dying:
 944                                 ++num;
 945                         }
 946                 }
 947                 spin_unlock_bh(lock);
 948         }
 949
 950 done:
 951         cb->args[1] = i;
 952         cb->args[2] = num;
 953 out:
 954         ;
 955 }
 956 EXPORT_SYMBOL_GPL(inet_diag_dump_icsk);
 957
 958 static int __inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb,
 959                 struct inet_diag_req_v2 *r, struct nlattr *bc)
 960 {
 961         const struct inet_diag_handler *handler;
 962         int err = 0;
 963
 964         handler = inet_diag_lock_handler(r->sdiag_protocol);
 965         if (!IS_ERR(handler))
 966                 handler->dump(skb, cb, r, bc);
 967         else
 968                 err = PTR_ERR(handler);
 969         inet_diag_unlock_handler(handler);
 970
 971         return err ? : skb->len;
 972 }
 973
 974 static int inet_diag_dump(struct sk_buff *skb, struct netlink_callback *cb)
 975 {
 976         struct nlattr *bc = NULL;
 977         int hdrlen = sizeof(struct inet_diag_req_v2);
 978
 979         if (nlmsg_attrlen(cb->nlh, hdrlen))
 980                 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
 981
 982         return __inet_diag_dump(skb, cb, nlmsg_data(cb->nlh), bc);
 983 }
 984
 985 static inline int inet_diag_type2proto(int type)
 986 {
 987         switch (type) {
 988         case TCPDIAG_GETSOCK:
 989                 return IPPROTO_TCP;
 990         case DCCPDIAG_GETSOCK:
 991                 return IPPROTO_DCCP;
 992         default:
 993                 return 0;
 994         }
 995 }
 996
 997 static int inet_diag_dump_compat(struct sk_buff *skb, struct netlink_callback *cb)
 998 {
 999         struct inet_diag_req *rc = nlmsg_data(cb->nlh);
1000         struct inet_diag_req_v2 req;
1001         struct nlattr *bc = NULL;
1002         int hdrlen = sizeof(struct inet_diag_req);
1003
1004         req.sdiag_family = AF_UNSPEC; /* compatibility */
1005         req.sdiag_protocol = inet_diag_type2proto(cb->nlh->nlmsg_type);
1006         req.idiag_ext = rc->idiag_ext;
1007         req.idiag_states = rc->idiag_states;
1008         req.id = rc->id;
1009
1010         if (nlmsg_attrlen(cb->nlh, hdrlen))
1011                 bc = nlmsg_find_attr(cb->nlh, hdrlen, INET_DIAG_REQ_BYTECODE);
1012
1013         return __inet_diag_dump(skb, cb, &req, bc);
1014 }
1015
1016 static int inet_diag_get_exact_compat(struct sk_buff *in_skb,
1017                                const struct nlmsghdr *nlh)
1018 {
1019         struct inet_diag_req *rc = nlmsg_data(nlh);
1020         struct inet_diag_req_v2 req;
1021
1022         req.sdiag_family = rc->idiag_family;
1023         req.sdiag_protocol = inet_diag_type2proto(nlh->nlmsg_type);
1024         req.idiag_ext = rc->idiag_ext;
1025         req.idiag_states = rc->idiag_states;
1026         req.id = rc->id;
1027
1028         return inet_diag_get_exact(in_skb, nlh, &req);
1029 }
1030
1031 static int inet_diag_rcv_msg_compat(struct sk_buff *skb, struct nlmsghdr *nlh)
1032 {
1033         int hdrlen = sizeof(struct inet_diag_req);
1034         struct net *net = sock_net(skb->sk);
1035
1036         if (nlh->nlmsg_type >= INET_DIAG_GETSOCK_MAX ||
1037             nlmsg_len(nlh) < hdrlen)
1038                 return -EINVAL;
1039
1040         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1041                 if (nlmsg_attrlen(nlh, hdrlen)) {
1042                         struct nlattr *attr;
1043
1044                         attr = nlmsg_find_attr(nlh, hdrlen,
1045                                                INET_DIAG_REQ_BYTECODE);
1046                         if (attr == NULL ||
1047                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
1048                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
1049                                 return -EINVAL;
1050                 }
1051                 {
1052                         struct netlink_dump_control c = {
1053                                 .dump = inet_diag_dump_compat,
1054                         };
1055                         return netlink_dump_start(net->diag_nlsk, skb, nlh, &c);
1056                 }
1057         }
1058
1059         return inet_diag_get_exact_compat(skb, nlh);
1060 }
1061
1062 static int inet_diag_handler_dump(struct sk_buff *skb, struct nlmsghdr *h)
1063 {
1064         int hdrlen = sizeof(struct inet_diag_req_v2);
1065         struct net *net = sock_net(skb->sk);
1066
1067         if (nlmsg_len(h) < hdrlen)
1068                 return -EINVAL;
1069
1070         if (h->nlmsg_flags & NLM_F_DUMP) {
1071                 if (nlmsg_attrlen(h, hdrlen)) {
1072                         struct nlattr *attr;
1073                         attr = nlmsg_find_attr(h, hdrlen,
1074                                                INET_DIAG_REQ_BYTECODE);
1075                         if (attr == NULL ||
1076                             nla_len(attr) < sizeof(struct inet_diag_bc_op) ||
1077                             inet_diag_bc_audit(nla_data(attr), nla_len(attr)))
1078                                 return -EINVAL;
1079                 }
1080                 {
1081                         struct netlink_dump_control c = {
1082                                 .dump = inet_diag_dump,
1083                         };
1084                         return netlink_dump_start(net->diag_nlsk, skb, h, &c);
1085                 }
1086         }
1087
1088         return inet_diag_get_exact(skb, h, nlmsg_data(h));
1089 }
1090
1091 static const struct sock_diag_handler inet_diag_handler = {
1092         .family = AF_INET,
1093         .dump = inet_diag_handler_dump,
1094 };
1095
1096 static const struct sock_diag_handler inet6_diag_handler = {
1097         .family = AF_INET6,
1098         .dump = inet_diag_handler_dump,
1099 };
1100
1101 int inet_diag_register(const struct inet_diag_handler *h)
1102 {
1103         const __u16 type = h->idiag_type;
1104         int err = -EINVAL;
1105
1106         if (type >= IPPROTO_MAX)
1107                 goto out;
1108
1109         mutex_lock(&inet_diag_table_mutex);
1110         err = -EEXIST;
1111         if (inet_diag_table[type] == NULL) {
1112                 inet_diag_table[type] = h;
1113                 err = 0;
1114         }
1115         mutex_unlock(&inet_diag_table_mutex);
1116 out:
1117         return err;
1118 }
1119 EXPORT_SYMBOL_GPL(inet_diag_register);
1120
1121 void inet_diag_unregister(const struct inet_diag_handler *h)
1122 {
1123         const __u16 type = h->idiag_type;
1124
1125         if (type >= IPPROTO_MAX)
1126                 return;
1127
1128         mutex_lock(&inet_diag_table_mutex);
1129         inet_diag_table[type] = NULL;
1130         mutex_unlock(&inet_diag_table_mutex);
1131 }
1132 EXPORT_SYMBOL_GPL(inet_diag_unregister);
1133
1134 static int __init inet_diag_init(void)
1135 {
1136         const int inet_diag_table_size = (IPPROTO_MAX *
1137                                           sizeof(struct inet_diag_handler *));
1138         int err = -ENOMEM;
1139
1140         inet_diag_table = kzalloc(inet_diag_table_size, GFP_KERNEL);
1141         if (!inet_diag_table)
1142                 goto out;
1143
1144         err = sock_diag_register(&inet_diag_handler);
1145         if (err)
1146                 goto out_free_nl;
1147
1148         err = sock_diag_register(&inet6_diag_handler);
1149         if (err)
1150                 goto out_free_inet;
1151
1152         sock_diag_register_inet_compat(inet_diag_rcv_msg_compat);
1153 out:
1154         return err;
1155
1156 out_free_inet:
1157         sock_diag_unregister(&inet_diag_handler);
1158 out_free_nl:
1159         kfree(inet_diag_table);
1160         goto out;
1161 }
1162
1163 static void __exit inet_diag_exit(void)
1164 {
1165         sock_diag_unregister(&inet6_diag_handler);
1166         sock_diag_unregister(&inet_diag_handler);
1167         sock_diag_unregister_inet_compat(inet_diag_rcv_msg_compat);
1168         kfree(inet_diag_table);
1169 }
1170
1171 module_init(inet_diag_init);
1172 module_exit(inet_diag_exit);
1173 MODULE_LICENSE("GPL");
1174 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 2 /* AF_INET */);
1175 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_NETLINK, NETLINK_SOCK_DIAG, 10 /* AF_INET6 */);