Pileus Git - ~andy/linux/blob - net/ipv4/sysctl_net_ipv4.c

   1 /*
   2  * sysctl_net_ipv4.c: sysctl interface to net IPV4 subsystem.
   3  *
   4  * Begun April 1, 1996, Mike Shaver.
   5  * Added /proc/sys/net/ipv4 directory entry (empty =) ). [MS]
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/module.h>
  10 #include <linux/sysctl.h>
  11 #include <linux/igmp.h>
  12 #include <linux/inetdevice.h>
  13 #include <linux/seqlock.h>
  14 #include <linux/init.h>
  15 #include <linux/slab.h>
  16 #include <linux/nsproxy.h>
  17 #include <linux/swap.h>
  18 #include <net/snmp.h>
  19 #include <net/icmp.h>
  20 #include <net/ip.h>
  21 #include <net/route.h>
  22 #include <net/tcp.h>
  23 #include <net/udp.h>
  24 #include <net/cipso_ipv4.h>
  25 #include <net/inet_frag.h>
  26 #include <net/ping.h>
  27 #include <net/tcp_memcontrol.h>
  28
  29 static int zero;
  30 static int two = 2;
  31 static int tcp_retr1_max = 255;
  32 static int ip_local_port_range_min[] = { 1, 1 };
  33 static int ip_local_port_range_max[] = { 65535, 65535 };
  34 static int tcp_adv_win_scale_min = -31;
  35 static int tcp_adv_win_scale_max = 31;
  36 static int ip_ttl_min = 1;
  37 static int ip_ttl_max = 255;
  38 static int ip_ping_group_range_min[] = { 0, 0 };
  39 static int ip_ping_group_range_max[] = { GID_T_MAX, GID_T_MAX };
  40
  41 /* Update system visible IP port range */
  42 static void set_local_port_range(int range[2])
  43 {
  44         write_seqlock(&sysctl_local_ports.lock);
  45         sysctl_local_ports.range[0] = range[0];
  46         sysctl_local_ports.range[1] = range[1];
  47         write_sequnlock(&sysctl_local_ports.lock);
  48 }
  49
  50 /* Validate changes from /proc interface. */
  51 static int ipv4_local_port_range(ctl_table *table, int write,
  52                                  void __user *buffer,
  53                                  size_t *lenp, loff_t *ppos)
  54 {
  55         int ret;
  56         int range[2];
  57         ctl_table tmp = {
  58                 .data = &range,
  59                 .maxlen = sizeof(range),
  60                 .mode = table->mode,
  61                 .extra1 = &ip_local_port_range_min,
  62                 .extra2 = &ip_local_port_range_max,
  63         };
  64
  65         inet_get_local_port_range(range, range + 1);
  66         ret = proc_dointvec_minmax(&tmp, write, buffer, lenp, ppos);
  67
  68         if (write && ret == 0) {
  69                 if (range[1] < range[0])
  70                         ret = -EINVAL;
  71                 else
  72                         set_local_port_range(range);
  73         }
  74
  75         return ret;
  76 }
  77
  78
  79 static void inet_get_ping_group_range_table(struct ctl_table *table, kgid_t *low, kgid_t *high)
  80 {
  81         kgid_t *data = table->data;
  82         unsigned int seq;
  83         do {
  84                 seq = read_seqbegin(&sysctl_local_ports.lock);
  85
  86                 *low = data[0];
  87                 *high = data[1];
  88         } while (read_seqretry(&sysctl_local_ports.lock, seq));
  89 }
  90
  91 /* Update system visible IP port range */
  92 static void set_ping_group_range(struct ctl_table *table, kgid_t low, kgid_t high)
  93 {
  94         kgid_t *data = table->data;
  95         write_seqlock(&sysctl_local_ports.lock);
  96         data[0] = low;
  97         data[1] = high;
  98         write_sequnlock(&sysctl_local_ports.lock);
  99 }
 100
 101 /* Validate changes from /proc interface. */
 102 static int ipv4_ping_group_range(ctl_table *table, int write,
 103                                  void __user *buffer,
 104                                  size_t *lenp, loff_t *ppos)
 105 {
 106         struct user_namespace *user_ns = current_user_ns();
 107         int ret;
 108         gid_t urange[2];
 109         kgid_t low, high;
 110         ctl_table tmp = {
 111                 .data = &urange,
 112                 .maxlen = sizeof(urange),
 113                 .mode = table->mode,
 114                 .extra1 = &ip_ping_group_range_min,
 115                 .extra2 = &ip_ping_group_range_max,
 116         };
 117
 118         inet_get_ping_group_range_table(table, &low, &high);
 119         urange[0] = from_kgid_munged(user_ns, low);
 120         urange[1] = from_kgid_munged(user_ns, high);
 121         ret = proc_dointvec_minmax(&tmp, write, buffer, lenp, ppos);
 122
 123         if (write && ret == 0) {
 124                 low = make_kgid(user_ns, urange[0]);
 125                 high = make_kgid(user_ns, urange[1]);
 126                 if (!gid_valid(low) || !gid_valid(high) ||
 127                     (urange[1] < urange[0]) || gid_lt(high, low)) {
 128                         low = make_kgid(&init_user_ns, 1);
 129                         high = make_kgid(&init_user_ns, 0);
 130                 }
 131                 set_ping_group_range(table, low, high);
 132         }
 133
 134         return ret;
 135 }
 136
 137 static int proc_tcp_congestion_control(ctl_table *ctl, int write,
 138                                        void __user *buffer, size_t *lenp, loff_t *ppos)
 139 {
 140         char val[TCP_CA_NAME_MAX];
 141         ctl_table tbl = {
 142                 .data = val,
 143                 .maxlen = TCP_CA_NAME_MAX,
 144         };
 145         int ret;
 146
 147         tcp_get_default_congestion_control(val);
 148
 149         ret = proc_dostring(&tbl, write, buffer, lenp, ppos);
 150         if (write && ret == 0)
 151                 ret = tcp_set_default_congestion_control(val);
 152         return ret;
 153 }
 154
 155 static int proc_tcp_available_congestion_control(ctl_table *ctl,
 156                                                  int write,
 157                                                  void __user *buffer, size_t *lenp,
 158                                                  loff_t *ppos)
 159 {
 160         ctl_table tbl = { .maxlen = TCP_CA_BUF_MAX, };
 161         int ret;
 162
 163         tbl.data = kmalloc(tbl.maxlen, GFP_USER);
 164         if (!tbl.data)
 165                 return -ENOMEM;
 166         tcp_get_available_congestion_control(tbl.data, TCP_CA_BUF_MAX);
 167         ret = proc_dostring(&tbl, write, buffer, lenp, ppos);
 168         kfree(tbl.data);
 169         return ret;
 170 }
 171
 172 static int proc_allowed_congestion_control(ctl_table *ctl,
 173                                            int write,
 174                                            void __user *buffer, size_t *lenp,
 175                                            loff_t *ppos)
 176 {
 177         ctl_table tbl = { .maxlen = TCP_CA_BUF_MAX };
 178         int ret;
 179
 180         tbl.data = kmalloc(tbl.maxlen, GFP_USER);
 181         if (!tbl.data)
 182                 return -ENOMEM;
 183
 184         tcp_get_allowed_congestion_control(tbl.data, tbl.maxlen);
 185         ret = proc_dostring(&tbl, write, buffer, lenp, ppos);
 186         if (write && ret == 0)
 187                 ret = tcp_set_allowed_congestion_control(tbl.data);
 188         kfree(tbl.data);
 189         return ret;
 190 }
 191
 192 static int ipv4_tcp_mem(ctl_table *ctl, int write,
 193                            void __user *buffer, size_t *lenp,
 194                            loff_t *ppos)
 195 {
 196         int ret;
 197         unsigned long vec[3];
 198         struct net *net = current->nsproxy->net_ns;
 199 #ifdef CONFIG_MEMCG_KMEM
 200         struct mem_cgroup *memcg;
 201 #endif
 202
 203         ctl_table tmp = {
 204                 .data = &vec,
 205                 .maxlen = sizeof(vec),
 206                 .mode = ctl->mode,
 207         };
 208
 209         if (!write) {
 210                 ctl->data = &net->ipv4.sysctl_tcp_mem;
 211                 return proc_doulongvec_minmax(ctl, write, buffer, lenp, ppos);
 212         }
 213
 214         ret = proc_doulongvec_minmax(&tmp, write, buffer, lenp, ppos);
 215         if (ret)
 216                 return ret;
 217
 218 #ifdef CONFIG_MEMCG_KMEM
 219         rcu_read_lock();
 220         memcg = mem_cgroup_from_task(current);
 221
 222         tcp_prot_mem(memcg, vec[0], 0);
 223         tcp_prot_mem(memcg, vec[1], 1);
 224         tcp_prot_mem(memcg, vec[2], 2);
 225         rcu_read_unlock();
 226 #endif
 227
 228         net->ipv4.sysctl_tcp_mem[0] = vec[0];
 229         net->ipv4.sysctl_tcp_mem[1] = vec[1];
 230         net->ipv4.sysctl_tcp_mem[2] = vec[2];
 231
 232         return 0;
 233 }
 234
 235 int proc_tcp_fastopen_key(ctl_table *ctl, int write, void __user *buffer,
 236                           size_t *lenp, loff_t *ppos)
 237 {
 238         ctl_table tbl = { .maxlen = (TCP_FASTOPEN_KEY_LENGTH * 2 + 10) };
 239         struct tcp_fastopen_context *ctxt;
 240         int ret;
 241         u32  user_key[4]; /* 16 bytes, matching TCP_FASTOPEN_KEY_LENGTH */
 242
 243         tbl.data = kmalloc(tbl.maxlen, GFP_KERNEL);
 244         if (!tbl.data)
 245                 return -ENOMEM;
 246
 247         rcu_read_lock();
 248         ctxt = rcu_dereference(tcp_fastopen_ctx);
 249         if (ctxt)
 250                 memcpy(user_key, ctxt->key, TCP_FASTOPEN_KEY_LENGTH);
 251         rcu_read_unlock();
 252
 253         snprintf(tbl.data, tbl.maxlen, "%08x-%08x-%08x-%08x",
 254                 user_key[0], user_key[1], user_key[2], user_key[3]);
 255         ret = proc_dostring(&tbl, write, buffer, lenp, ppos);
 256
 257         if (write && ret == 0) {
 258                 if (sscanf(tbl.data, "%x-%x-%x-%x", user_key, user_key + 1,
 259                            user_key + 2, user_key + 3) != 4) {
 260                         ret = -EINVAL;
 261                         goto bad_key;
 262                 }
 263                 tcp_fastopen_reset_cipher(user_key, TCP_FASTOPEN_KEY_LENGTH);
 264         }
 265
 266 bad_key:
 267         pr_debug("proc FO key set 0x%x-%x-%x-%x <- 0x%s: %u\n",
 268                user_key[0], user_key[1], user_key[2], user_key[3],
 269                (char *)tbl.data, ret);
 270         kfree(tbl.data);
 271         return ret;
 272 }
 273
 274 static struct ctl_table ipv4_table[] = {
 275         {
 276                 .procname       = "tcp_timestamps",
 277                 .data           = &sysctl_tcp_timestamps,
 278                 .maxlen         = sizeof(int),
 279                 .mode           = 0644,
 280                 .proc_handler   = proc_dointvec
 281         },
 282         {
 283                 .procname       = "tcp_window_scaling",
 284                 .data           = &sysctl_tcp_window_scaling,
 285                 .maxlen         = sizeof(int),
 286                 .mode           = 0644,
 287                 .proc_handler   = proc_dointvec
 288         },
 289         {
 290                 .procname       = "tcp_sack",
 291                 .data           = &sysctl_tcp_sack,
 292                 .maxlen         = sizeof(int),
 293                 .mode           = 0644,
 294                 .proc_handler   = proc_dointvec
 295         },
 296         {
 297                 .procname       = "tcp_retrans_collapse",
 298                 .data           = &sysctl_tcp_retrans_collapse,
 299                 .maxlen         = sizeof(int),
 300                 .mode           = 0644,
 301                 .proc_handler   = proc_dointvec
 302         },
 303         {
 304                 .procname       = "ip_default_ttl",
 305                 .data           = &sysctl_ip_default_ttl,
 306                 .maxlen         = sizeof(int),
 307                 .mode           = 0644,
 308                 .proc_handler   = proc_dointvec_minmax,
 309                 .extra1         = &ip_ttl_min,
 310                 .extra2         = &ip_ttl_max,
 311         },
 312         {
 313                 .procname       = "ip_no_pmtu_disc",
 314                 .data           = &ipv4_config.no_pmtu_disc,
 315                 .maxlen         = sizeof(int),
 316                 .mode           = 0644,
 317                 .proc_handler   = proc_dointvec
 318         },
 319         {
 320                 .procname       = "ip_nonlocal_bind",
 321                 .data           = &sysctl_ip_nonlocal_bind,
 322                 .maxlen         = sizeof(int),
 323                 .mode           = 0644,
 324                 .proc_handler   = proc_dointvec
 325         },
 326         {
 327                 .procname       = "tcp_syn_retries",
 328                 .data           = &sysctl_tcp_syn_retries,
 329                 .maxlen         = sizeof(int),
 330                 .mode           = 0644,
 331                 .proc_handler   = proc_dointvec
 332         },
 333         {
 334                 .procname       = "tcp_synack_retries",
 335                 .data           = &sysctl_tcp_synack_retries,
 336                 .maxlen         = sizeof(int),
 337                 .mode           = 0644,
 338                 .proc_handler   = proc_dointvec
 339         },
 340         {
 341                 .procname       = "tcp_max_orphans",
 342                 .data           = &sysctl_tcp_max_orphans,
 343                 .maxlen         = sizeof(int),
 344                 .mode           = 0644,
 345                 .proc_handler   = proc_dointvec
 346         },
 347         {
 348                 .procname       = "tcp_max_tw_buckets",
 349                 .data           = &tcp_death_row.sysctl_max_tw_buckets,
 350                 .maxlen         = sizeof(int),
 351                 .mode           = 0644,
 352                 .proc_handler   = proc_dointvec
 353         },
 354         {
 355                 .procname       = "ip_early_demux",
 356                 .data           = &sysctl_ip_early_demux,
 357                 .maxlen         = sizeof(int),
 358                 .mode           = 0644,
 359                 .proc_handler   = proc_dointvec
 360         },
 361         {
 362                 .procname       = "ip_dynaddr",
 363                 .data           = &sysctl_ip_dynaddr,
 364                 .maxlen         = sizeof(int),
 365                 .mode           = 0644,
 366                 .proc_handler   = proc_dointvec
 367         },
 368         {
 369                 .procname       = "tcp_keepalive_time",
 370                 .data           = &sysctl_tcp_keepalive_time,
 371                 .maxlen         = sizeof(int),
 372                 .mode           = 0644,
 373                 .proc_handler   = proc_dointvec_jiffies,
 374         },
 375         {
 376                 .procname       = "tcp_keepalive_probes",
 377                 .data           = &sysctl_tcp_keepalive_probes,
 378                 .maxlen         = sizeof(int),
 379                 .mode           = 0644,
 380                 .proc_handler   = proc_dointvec
 381         },
 382         {
 383                 .procname       = "tcp_keepalive_intvl",
 384                 .data           = &sysctl_tcp_keepalive_intvl,
 385                 .maxlen         = sizeof(int),
 386                 .mode           = 0644,
 387                 .proc_handler   = proc_dointvec_jiffies,
 388         },
 389         {
 390                 .procname       = "tcp_retries1",
 391                 .data           = &sysctl_tcp_retries1,
 392                 .maxlen         = sizeof(int),
 393                 .mode           = 0644,
 394                 .proc_handler   = proc_dointvec_minmax,
 395                 .extra2         = &tcp_retr1_max
 396         },
 397         {
 398                 .procname       = "tcp_retries2",
 399                 .data           = &sysctl_tcp_retries2,
 400                 .maxlen         = sizeof(int),
 401                 .mode           = 0644,
 402                 .proc_handler   = proc_dointvec
 403         },
 404         {
 405                 .procname       = "tcp_fin_timeout",
 406                 .data           = &sysctl_tcp_fin_timeout,
 407                 .maxlen         = sizeof(int),
 408                 .mode           = 0644,
 409                 .proc_handler   = proc_dointvec_jiffies,
 410         },
 411 #ifdef CONFIG_SYN_COOKIES
 412         {
 413                 .procname       = "tcp_syncookies",
 414                 .data           = &sysctl_tcp_syncookies,
 415                 .maxlen         = sizeof(int),
 416                 .mode           = 0644,
 417                 .proc_handler   = proc_dointvec
 418         },
 419 #endif
 420         {
 421                 .procname       = "tcp_fastopen",
 422                 .data           = &sysctl_tcp_fastopen,
 423                 .maxlen         = sizeof(int),
 424                 .mode           = 0644,
 425                 .proc_handler   = proc_dointvec,
 426         },
 427         {
 428                 .procname       = "tcp_fastopen_key",
 429                 .mode           = 0600,
 430                 .maxlen         = ((TCP_FASTOPEN_KEY_LENGTH * 2) + 10),
 431                 .proc_handler   = proc_tcp_fastopen_key,
 432         },
 433         {
 434                 .procname       = "tcp_tw_recycle",
 435                 .data           = &tcp_death_row.sysctl_tw_recycle,
 436                 .maxlen         = sizeof(int),
 437                 .mode           = 0644,
 438                 .proc_handler   = proc_dointvec
 439         },
 440         {
 441                 .procname       = "tcp_abort_on_overflow",
 442                 .data           = &sysctl_tcp_abort_on_overflow,
 443                 .maxlen         = sizeof(int),
 444                 .mode           = 0644,
 445                 .proc_handler   = proc_dointvec
 446         },
 447         {
 448                 .procname       = "tcp_stdurg",
 449                 .data           = &sysctl_tcp_stdurg,
 450                 .maxlen         = sizeof(int),
 451                 .mode           = 0644,
 452                 .proc_handler   = proc_dointvec
 453         },
 454         {
 455                 .procname       = "tcp_rfc1337",
 456                 .data           = &sysctl_tcp_rfc1337,
 457                 .maxlen         = sizeof(int),
 458                 .mode           = 0644,
 459                 .proc_handler   = proc_dointvec
 460         },
 461         {
 462                 .procname       = "tcp_max_syn_backlog",
 463                 .data           = &sysctl_max_syn_backlog,
 464                 .maxlen         = sizeof(int),
 465                 .mode           = 0644,
 466                 .proc_handler   = proc_dointvec
 467         },
 468         {
 469                 .procname       = "ip_local_port_range",
 470                 .data           = &sysctl_local_ports.range,
 471                 .maxlen         = sizeof(sysctl_local_ports.range),
 472                 .mode           = 0644,
 473                 .proc_handler   = ipv4_local_port_range,
 474         },
 475         {
 476                 .procname       = "ip_local_reserved_ports",
 477                 .data           = NULL, /* initialized in sysctl_ipv4_init */
 478                 .maxlen         = 65536,
 479                 .mode           = 0644,
 480                 .proc_handler   = proc_do_large_bitmap,
 481         },
 482         {
 483                 .procname       = "igmp_max_memberships",
 484                 .data           = &sysctl_igmp_max_memberships,
 485                 .maxlen         = sizeof(int),
 486                 .mode           = 0644,
 487                 .proc_handler   = proc_dointvec
 488         },
 489         {
 490                 .procname       = "igmp_max_msf",
 491                 .data           = &sysctl_igmp_max_msf,
 492                 .maxlen         = sizeof(int),
 493                 .mode           = 0644,
 494                 .proc_handler   = proc_dointvec
 495         },
 496         {
 497                 .procname       = "inet_peer_threshold",
 498                 .data           = &inet_peer_threshold,
 499                 .maxlen         = sizeof(int),
 500                 .mode           = 0644,
 501                 .proc_handler   = proc_dointvec
 502         },
 503         {
 504                 .procname       = "inet_peer_minttl",
 505                 .data           = &inet_peer_minttl,
 506                 .maxlen         = sizeof(int),
 507                 .mode           = 0644,
 508                 .proc_handler   = proc_dointvec_jiffies,
 509         },
 510         {
 511                 .procname       = "inet_peer_maxttl",
 512                 .data           = &inet_peer_maxttl,
 513                 .maxlen         = sizeof(int),
 514                 .mode           = 0644,
 515                 .proc_handler   = proc_dointvec_jiffies,
 516         },
 517         {
 518                 .procname       = "tcp_orphan_retries",
 519                 .data           = &sysctl_tcp_orphan_retries,
 520                 .maxlen         = sizeof(int),
 521                 .mode           = 0644,
 522                 .proc_handler   = proc_dointvec
 523         },
 524         {
 525                 .procname       = "tcp_fack",
 526                 .data           = &sysctl_tcp_fack,
 527                 .maxlen         = sizeof(int),
 528                 .mode           = 0644,
 529                 .proc_handler   = proc_dointvec
 530         },
 531         {
 532                 .procname       = "tcp_reordering",
 533                 .data           = &sysctl_tcp_reordering,
 534                 .maxlen         = sizeof(int),
 535                 .mode           = 0644,
 536                 .proc_handler   = proc_dointvec
 537         },
 538         {
 539                 .procname       = "tcp_ecn",
 540                 .data           = &sysctl_tcp_ecn,
 541                 .maxlen         = sizeof(int),
 542                 .mode           = 0644,
 543                 .proc_handler   = proc_dointvec
 544         },
 545         {
 546                 .procname       = "tcp_dsack",
 547                 .data           = &sysctl_tcp_dsack,
 548                 .maxlen         = sizeof(int),
 549                 .mode           = 0644,
 550                 .proc_handler   = proc_dointvec
 551         },
 552         {
 553                 .procname       = "tcp_wmem",
 554                 .data           = &sysctl_tcp_wmem,
 555                 .maxlen         = sizeof(sysctl_tcp_wmem),
 556                 .mode           = 0644,
 557                 .proc_handler   = proc_dointvec
 558         },
 559         {
 560                 .procname       = "tcp_rmem",
 561                 .data           = &sysctl_tcp_rmem,
 562                 .maxlen         = sizeof(sysctl_tcp_rmem),
 563                 .mode           = 0644,
 564                 .proc_handler   = proc_dointvec
 565         },
 566         {
 567                 .procname       = "tcp_app_win",
 568                 .data           = &sysctl_tcp_app_win,
 569                 .maxlen         = sizeof(int),
 570                 .mode           = 0644,
 571                 .proc_handler   = proc_dointvec
 572         },
 573         {
 574                 .procname       = "tcp_adv_win_scale",
 575                 .data           = &sysctl_tcp_adv_win_scale,
 576                 .maxlen         = sizeof(int),
 577                 .mode           = 0644,
 578                 .proc_handler   = proc_dointvec_minmax,
 579                 .extra1         = &tcp_adv_win_scale_min,
 580                 .extra2         = &tcp_adv_win_scale_max,
 581         },
 582         {
 583                 .procname       = "tcp_tw_reuse",
 584                 .data           = &sysctl_tcp_tw_reuse,
 585                 .maxlen         = sizeof(int),
 586                 .mode           = 0644,
 587                 .proc_handler   = proc_dointvec
 588         },
 589         {
 590                 .procname       = "tcp_frto",
 591                 .data           = &sysctl_tcp_frto,
 592                 .maxlen         = sizeof(int),
 593                 .mode           = 0644,
 594                 .proc_handler   = proc_dointvec
 595         },
 596         {
 597                 .procname       = "tcp_frto_response",
 598                 .data           = &sysctl_tcp_frto_response,
 599                 .maxlen         = sizeof(int),
 600                 .mode           = 0644,
 601                 .proc_handler   = proc_dointvec
 602         },
 603         {
 604                 .procname       = "tcp_low_latency",
 605                 .data           = &sysctl_tcp_low_latency,
 606                 .maxlen         = sizeof(int),
 607                 .mode           = 0644,
 608                 .proc_handler   = proc_dointvec
 609         },
 610         {
 611                 .procname       = "tcp_no_metrics_save",
 612                 .data           = &sysctl_tcp_nometrics_save,
 613                 .maxlen         = sizeof(int),
 614                 .mode           = 0644,
 615                 .proc_handler   = proc_dointvec,
 616         },
 617         {
 618                 .procname       = "tcp_moderate_rcvbuf",
 619                 .data           = &sysctl_tcp_moderate_rcvbuf,
 620                 .maxlen         = sizeof(int),
 621                 .mode           = 0644,
 622                 .proc_handler   = proc_dointvec,
 623         },
 624         {
 625                 .procname       = "tcp_tso_win_divisor",
 626                 .data           = &sysctl_tcp_tso_win_divisor,
 627                 .maxlen         = sizeof(int),
 628                 .mode           = 0644,
 629                 .proc_handler   = proc_dointvec,
 630         },
 631         {
 632                 .procname       = "tcp_congestion_control",
 633                 .mode           = 0644,
 634                 .maxlen         = TCP_CA_NAME_MAX,
 635                 .proc_handler   = proc_tcp_congestion_control,
 636         },
 637         {
 638                 .procname       = "tcp_abc",
 639                 .data           = &sysctl_tcp_abc,
 640                 .maxlen         = sizeof(int),
 641                 .mode           = 0644,
 642                 .proc_handler   = proc_dointvec,
 643         },
 644         {
 645                 .procname       = "tcp_mtu_probing",
 646                 .data           = &sysctl_tcp_mtu_probing,
 647                 .maxlen         = sizeof(int),
 648                 .mode           = 0644,
 649                 .proc_handler   = proc_dointvec,
 650         },
 651         {
 652                 .procname       = "tcp_base_mss",
 653                 .data           = &sysctl_tcp_base_mss,
 654                 .maxlen         = sizeof(int),
 655                 .mode           = 0644,
 656                 .proc_handler   = proc_dointvec,
 657         },
 658         {
 659                 .procname       = "tcp_workaround_signed_windows",
 660                 .data           = &sysctl_tcp_workaround_signed_windows,
 661                 .maxlen         = sizeof(int),
 662                 .mode           = 0644,
 663                 .proc_handler   = proc_dointvec
 664         },
 665         {
 666                 .procname       = "tcp_limit_output_bytes",
 667                 .data           = &sysctl_tcp_limit_output_bytes,
 668                 .maxlen         = sizeof(int),
 669                 .mode           = 0644,
 670                 .proc_handler   = proc_dointvec
 671         },
 672         {
 673                 .procname       = "tcp_challenge_ack_limit",
 674                 .data           = &sysctl_tcp_challenge_ack_limit,
 675                 .maxlen         = sizeof(int),
 676                 .mode           = 0644,
 677                 .proc_handler   = proc_dointvec
 678         },
 679 #ifdef CONFIG_NET_DMA
 680         {
 681                 .procname       = "tcp_dma_copybreak",
 682                 .data           = &sysctl_tcp_dma_copybreak,
 683                 .maxlen         = sizeof(int),
 684                 .mode           = 0644,
 685                 .proc_handler   = proc_dointvec
 686         },
 687 #endif
 688         {
 689                 .procname       = "tcp_slow_start_after_idle",
 690                 .data           = &sysctl_tcp_slow_start_after_idle,
 691                 .maxlen         = sizeof(int),
 692                 .mode           = 0644,
 693                 .proc_handler   = proc_dointvec
 694         },
 695 #ifdef CONFIG_NETLABEL
 696         {
 697                 .procname       = "cipso_cache_enable",
 698                 .data           = &cipso_v4_cache_enabled,
 699                 .maxlen         = sizeof(int),
 700                 .mode           = 0644,
 701                 .proc_handler   = proc_dointvec,
 702         },
 703         {
 704                 .procname       = "cipso_cache_bucket_size",
 705                 .data           = &cipso_v4_cache_bucketsize,
 706                 .maxlen         = sizeof(int),
 707                 .mode           = 0644,
 708                 .proc_handler   = proc_dointvec,
 709         },
 710         {
 711                 .procname       = "cipso_rbm_optfmt",
 712                 .data           = &cipso_v4_rbm_optfmt,
 713                 .maxlen         = sizeof(int),
 714                 .mode           = 0644,
 715                 .proc_handler   = proc_dointvec,
 716         },
 717         {
 718                 .procname       = "cipso_rbm_strictvalid",
 719                 .data           = &cipso_v4_rbm_strictvalid,
 720                 .maxlen         = sizeof(int),
 721                 .mode           = 0644,
 722                 .proc_handler   = proc_dointvec,
 723         },
 724 #endif /* CONFIG_NETLABEL */
 725         {
 726                 .procname       = "tcp_available_congestion_control",
 727                 .maxlen         = TCP_CA_BUF_MAX,
 728                 .mode           = 0444,
 729                 .proc_handler   = proc_tcp_available_congestion_control,
 730         },
 731         {
 732                 .procname       = "tcp_allowed_congestion_control",
 733                 .maxlen         = TCP_CA_BUF_MAX,
 734                 .mode           = 0644,
 735                 .proc_handler   = proc_allowed_congestion_control,
 736         },
 737         {
 738                 .procname       = "tcp_max_ssthresh",
 739                 .data           = &sysctl_tcp_max_ssthresh,
 740                 .maxlen         = sizeof(int),
 741                 .mode           = 0644,
 742                 .proc_handler   = proc_dointvec,
 743         },
 744         {
 745                 .procname       = "tcp_cookie_size",
 746                 .data           = &sysctl_tcp_cookie_size,
 747                 .maxlen         = sizeof(int),
 748                 .mode           = 0644,
 749                 .proc_handler   = proc_dointvec
 750         },
 751         {
 752                 .procname       = "tcp_thin_linear_timeouts",
 753                 .data           = &sysctl_tcp_thin_linear_timeouts,
 754                 .maxlen         = sizeof(int),
 755                 .mode           = 0644,
 756                 .proc_handler   = proc_dointvec
 757         },
 758         {
 759                 .procname       = "tcp_thin_dupack",
 760                 .data           = &sysctl_tcp_thin_dupack,
 761                 .maxlen         = sizeof(int),
 762                 .mode           = 0644,
 763                 .proc_handler   = proc_dointvec
 764         },
 765         {
 766                 .procname       = "tcp_early_retrans",
 767                 .data           = &sysctl_tcp_early_retrans,
 768                 .maxlen         = sizeof(int),
 769                 .mode           = 0644,
 770                 .proc_handler   = proc_dointvec_minmax,
 771                 .extra1         = &zero,
 772                 .extra2         = &two,
 773         },
 774         {
 775                 .procname       = "udp_mem",
 776                 .data           = &sysctl_udp_mem,
 777                 .maxlen         = sizeof(sysctl_udp_mem),
 778                 .mode           = 0644,
 779                 .proc_handler   = proc_doulongvec_minmax,
 780         },
 781         {
 782                 .procname       = "udp_rmem_min",
 783                 .data           = &sysctl_udp_rmem_min,
 784                 .maxlen         = sizeof(sysctl_udp_rmem_min),
 785                 .mode           = 0644,
 786                 .proc_handler   = proc_dointvec_minmax,
 787                 .extra1         = &zero
 788         },
 789         {
 790                 .procname       = "udp_wmem_min",
 791                 .data           = &sysctl_udp_wmem_min,
 792                 .maxlen         = sizeof(sysctl_udp_wmem_min),
 793                 .mode           = 0644,
 794                 .proc_handler   = proc_dointvec_minmax,
 795                 .extra1         = &zero
 796         },
 797         { }
 798 };
 799
 800 static struct ctl_table ipv4_net_table[] = {
 801         {
 802                 .procname       = "icmp_echo_ignore_all",
 803                 .data           = &init_net.ipv4.sysctl_icmp_echo_ignore_all,
 804                 .maxlen         = sizeof(int),
 805                 .mode           = 0644,
 806                 .proc_handler   = proc_dointvec
 807         },
 808         {
 809                 .procname       = "icmp_echo_ignore_broadcasts",
 810                 .data           = &init_net.ipv4.sysctl_icmp_echo_ignore_broadcasts,
 811                 .maxlen         = sizeof(int),
 812                 .mode           = 0644,
 813                 .proc_handler   = proc_dointvec
 814         },
 815         {
 816                 .procname       = "icmp_ignore_bogus_error_responses",
 817                 .data           = &init_net.ipv4.sysctl_icmp_ignore_bogus_error_responses,
 818                 .maxlen         = sizeof(int),
 819                 .mode           = 0644,
 820                 .proc_handler   = proc_dointvec
 821         },
 822         {
 823                 .procname       = "icmp_errors_use_inbound_ifaddr",
 824                 .data           = &init_net.ipv4.sysctl_icmp_errors_use_inbound_ifaddr,
 825                 .maxlen         = sizeof(int),
 826                 .mode           = 0644,
 827                 .proc_handler   = proc_dointvec
 828         },
 829         {
 830                 .procname       = "icmp_ratelimit",
 831                 .data           = &init_net.ipv4.sysctl_icmp_ratelimit,
 832                 .maxlen         = sizeof(int),
 833                 .mode           = 0644,
 834                 .proc_handler   = proc_dointvec_ms_jiffies,
 835         },
 836         {
 837                 .procname       = "icmp_ratemask",
 838                 .data           = &init_net.ipv4.sysctl_icmp_ratemask,
 839                 .maxlen         = sizeof(int),
 840                 .mode           = 0644,
 841                 .proc_handler   = proc_dointvec
 842         },
 843         {
 844                 .procname       = "ping_group_range",
 845                 .data           = &init_net.ipv4.sysctl_ping_group_range,
 846                 .maxlen         = sizeof(gid_t)*2,
 847                 .mode           = 0644,
 848                 .proc_handler   = ipv4_ping_group_range,
 849         },
 850         {
 851                 .procname       = "tcp_mem",
 852                 .maxlen         = sizeof(init_net.ipv4.sysctl_tcp_mem),
 853                 .mode           = 0644,
 854                 .proc_handler   = ipv4_tcp_mem,
 855         },
 856         { }
 857 };
 858
 859 static __net_init int ipv4_sysctl_init_net(struct net *net)
 860 {
 861         struct ctl_table *table;
 862
 863         table = ipv4_net_table;
 864         if (!net_eq(net, &init_net)) {
 865                 table = kmemdup(table, sizeof(ipv4_net_table), GFP_KERNEL);
 866                 if (table == NULL)
 867                         goto err_alloc;
 868
 869                 table[0].data =
 870                         &net->ipv4.sysctl_icmp_echo_ignore_all;
 871                 table[1].data =
 872                         &net->ipv4.sysctl_icmp_echo_ignore_broadcasts;
 873                 table[2].data =
 874                         &net->ipv4.sysctl_icmp_ignore_bogus_error_responses;
 875                 table[3].data =
 876                         &net->ipv4.sysctl_icmp_errors_use_inbound_ifaddr;
 877                 table[4].data =
 878                         &net->ipv4.sysctl_icmp_ratelimit;
 879                 table[5].data =
 880                         &net->ipv4.sysctl_icmp_ratemask;
 881                 table[6].data =
 882                         &net->ipv4.sysctl_ping_group_range;
 883
 884         }
 885
 886         /*
 887          * Sane defaults - nobody may create ping sockets.
 888          * Boot scripts should set this to distro-specific group.
 889          */
 890         net->ipv4.sysctl_ping_group_range[0] = make_kgid(&init_user_ns, 1);
 891         net->ipv4.sysctl_ping_group_range[1] = make_kgid(&init_user_ns, 0);
 892
 893         tcp_init_mem(net);
 894
 895         net->ipv4.ipv4_hdr = register_net_sysctl(net, "net/ipv4", table);
 896         if (net->ipv4.ipv4_hdr == NULL)
 897                 goto err_reg;
 898
 899         return 0;
 900
 901 err_reg:
 902         if (!net_eq(net, &init_net))
 903                 kfree(table);
 904 err_alloc:
 905         return -ENOMEM;
 906 }
 907
 908 static __net_exit void ipv4_sysctl_exit_net(struct net *net)
 909 {
 910         struct ctl_table *table;
 911
 912         table = net->ipv4.ipv4_hdr->ctl_table_arg;
 913         unregister_net_sysctl_table(net->ipv4.ipv4_hdr);
 914         kfree(table);
 915 }
 916
 917 static __net_initdata struct pernet_operations ipv4_sysctl_ops = {
 918         .init = ipv4_sysctl_init_net,
 919         .exit = ipv4_sysctl_exit_net,
 920 };
 921
 922 static __init int sysctl_ipv4_init(void)
 923 {
 924         struct ctl_table_header *hdr;
 925         struct ctl_table *i;
 926
 927         for (i = ipv4_table; i->procname; i++) {
 928                 if (strcmp(i->procname, "ip_local_reserved_ports") == 0) {
 929                         i->data = sysctl_local_reserved_ports;
 930                         break;
 931                 }
 932         }
 933         if (!i->procname)
 934                 return -EINVAL;
 935
 936         hdr = register_net_sysctl(&init_net, "net/ipv4", ipv4_table);
 937         if (hdr == NULL)
 938                 return -ENOMEM;
 939
 940         if (register_pernet_subsys(&ipv4_sysctl_ops)) {
 941                 unregister_net_sysctl_table(hdr);
 942                 return -ENOMEM;
 943         }
 944
 945         return 0;
 946 }
 947
 948 __initcall(sysctl_ipv4_init);