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 static struct ctl_table ipv4_table[] = {
 236         {
 237                 .procname       = "tcp_timestamps",
 238                 .data           = &sysctl_tcp_timestamps,
 239                 .maxlen         = sizeof(int),
 240                 .mode           = 0644,
 241                 .proc_handler   = proc_dointvec
 242         },
 243         {
 244                 .procname       = "tcp_window_scaling",
 245                 .data           = &sysctl_tcp_window_scaling,
 246                 .maxlen         = sizeof(int),
 247                 .mode           = 0644,
 248                 .proc_handler   = proc_dointvec
 249         },
 250         {
 251                 .procname       = "tcp_sack",
 252                 .data           = &sysctl_tcp_sack,
 253                 .maxlen         = sizeof(int),
 254                 .mode           = 0644,
 255                 .proc_handler   = proc_dointvec
 256         },
 257         {
 258                 .procname       = "tcp_retrans_collapse",
 259                 .data           = &sysctl_tcp_retrans_collapse,
 260                 .maxlen         = sizeof(int),
 261                 .mode           = 0644,
 262                 .proc_handler   = proc_dointvec
 263         },
 264         {
 265                 .procname       = "ip_default_ttl",
 266                 .data           = &sysctl_ip_default_ttl,
 267                 .maxlen         = sizeof(int),
 268                 .mode           = 0644,
 269                 .proc_handler   = proc_dointvec_minmax,
 270                 .extra1         = &ip_ttl_min,
 271                 .extra2         = &ip_ttl_max,
 272         },
 273         {
 274                 .procname       = "ip_no_pmtu_disc",
 275                 .data           = &ipv4_config.no_pmtu_disc,
 276                 .maxlen         = sizeof(int),
 277                 .mode           = 0644,
 278                 .proc_handler   = proc_dointvec
 279         },
 280         {
 281                 .procname       = "ip_nonlocal_bind",
 282                 .data           = &sysctl_ip_nonlocal_bind,
 283                 .maxlen         = sizeof(int),
 284                 .mode           = 0644,
 285                 .proc_handler   = proc_dointvec
 286         },
 287         {
 288                 .procname       = "tcp_syn_retries",
 289                 .data           = &sysctl_tcp_syn_retries,
 290                 .maxlen         = sizeof(int),
 291                 .mode           = 0644,
 292                 .proc_handler   = proc_dointvec
 293         },
 294         {
 295                 .procname       = "tcp_synack_retries",
 296                 .data           = &sysctl_tcp_synack_retries,
 297                 .maxlen         = sizeof(int),
 298                 .mode           = 0644,
 299                 .proc_handler   = proc_dointvec
 300         },
 301         {
 302                 .procname       = "tcp_max_orphans",
 303                 .data           = &sysctl_tcp_max_orphans,
 304                 .maxlen         = sizeof(int),
 305                 .mode           = 0644,
 306                 .proc_handler   = proc_dointvec
 307         },
 308         {
 309                 .procname       = "tcp_max_tw_buckets",
 310                 .data           = &tcp_death_row.sysctl_max_tw_buckets,
 311                 .maxlen         = sizeof(int),
 312                 .mode           = 0644,
 313                 .proc_handler   = proc_dointvec
 314         },
 315         {
 316                 .procname       = "ip_early_demux",
 317                 .data           = &sysctl_ip_early_demux,
 318                 .maxlen         = sizeof(int),
 319                 .mode           = 0644,
 320                 .proc_handler   = proc_dointvec
 321         },
 322         {
 323                 .procname       = "ip_dynaddr",
 324                 .data           = &sysctl_ip_dynaddr,
 325                 .maxlen         = sizeof(int),
 326                 .mode           = 0644,
 327                 .proc_handler   = proc_dointvec
 328         },
 329         {
 330                 .procname       = "tcp_keepalive_time",
 331                 .data           = &sysctl_tcp_keepalive_time,
 332                 .maxlen         = sizeof(int),
 333                 .mode           = 0644,
 334                 .proc_handler   = proc_dointvec_jiffies,
 335         },
 336         {
 337                 .procname       = "tcp_keepalive_probes",
 338                 .data           = &sysctl_tcp_keepalive_probes,
 339                 .maxlen         = sizeof(int),
 340                 .mode           = 0644,
 341                 .proc_handler   = proc_dointvec
 342         },
 343         {
 344                 .procname       = "tcp_keepalive_intvl",
 345                 .data           = &sysctl_tcp_keepalive_intvl,
 346                 .maxlen         = sizeof(int),
 347                 .mode           = 0644,
 348                 .proc_handler   = proc_dointvec_jiffies,
 349         },
 350         {
 351                 .procname       = "tcp_retries1",
 352                 .data           = &sysctl_tcp_retries1,
 353                 .maxlen         = sizeof(int),
 354                 .mode           = 0644,
 355                 .proc_handler   = proc_dointvec_minmax,
 356                 .extra2         = &tcp_retr1_max
 357         },
 358         {
 359                 .procname       = "tcp_retries2",
 360                 .data           = &sysctl_tcp_retries2,
 361                 .maxlen         = sizeof(int),
 362                 .mode           = 0644,
 363                 .proc_handler   = proc_dointvec
 364         },
 365         {
 366                 .procname       = "tcp_fin_timeout",
 367                 .data           = &sysctl_tcp_fin_timeout,
 368                 .maxlen         = sizeof(int),
 369                 .mode           = 0644,
 370                 .proc_handler   = proc_dointvec_jiffies,
 371         },
 372 #ifdef CONFIG_SYN_COOKIES
 373         {
 374                 .procname       = "tcp_syncookies",
 375                 .data           = &sysctl_tcp_syncookies,
 376                 .maxlen         = sizeof(int),
 377                 .mode           = 0644,
 378                 .proc_handler   = proc_dointvec
 379         },
 380 #endif
 381         {
 382                 .procname       = "tcp_fastopen",
 383                 .data           = &sysctl_tcp_fastopen,
 384                 .maxlen         = sizeof(int),
 385                 .mode           = 0644,
 386                 .proc_handler   = proc_dointvec,
 387         },
 388         {
 389                 .procname       = "tcp_tw_recycle",
 390                 .data           = &tcp_death_row.sysctl_tw_recycle,
 391                 .maxlen         = sizeof(int),
 392                 .mode           = 0644,
 393                 .proc_handler   = proc_dointvec
 394         },
 395         {
 396                 .procname       = "tcp_abort_on_overflow",
 397                 .data           = &sysctl_tcp_abort_on_overflow,
 398                 .maxlen         = sizeof(int),
 399                 .mode           = 0644,
 400                 .proc_handler   = proc_dointvec
 401         },
 402         {
 403                 .procname       = "tcp_stdurg",
 404                 .data           = &sysctl_tcp_stdurg,
 405                 .maxlen         = sizeof(int),
 406                 .mode           = 0644,
 407                 .proc_handler   = proc_dointvec
 408         },
 409         {
 410                 .procname       = "tcp_rfc1337",
 411                 .data           = &sysctl_tcp_rfc1337,
 412                 .maxlen         = sizeof(int),
 413                 .mode           = 0644,
 414                 .proc_handler   = proc_dointvec
 415         },
 416         {
 417                 .procname       = "tcp_max_syn_backlog",
 418                 .data           = &sysctl_max_syn_backlog,
 419                 .maxlen         = sizeof(int),
 420                 .mode           = 0644,
 421                 .proc_handler   = proc_dointvec
 422         },
 423         {
 424                 .procname       = "ip_local_port_range",
 425                 .data           = &sysctl_local_ports.range,
 426                 .maxlen         = sizeof(sysctl_local_ports.range),
 427                 .mode           = 0644,
 428                 .proc_handler   = ipv4_local_port_range,
 429         },
 430         {
 431                 .procname       = "ip_local_reserved_ports",
 432                 .data           = NULL, /* initialized in sysctl_ipv4_init */
 433                 .maxlen         = 65536,
 434                 .mode           = 0644,
 435                 .proc_handler   = proc_do_large_bitmap,
 436         },
 437         {
 438                 .procname       = "igmp_max_memberships",
 439                 .data           = &sysctl_igmp_max_memberships,
 440                 .maxlen         = sizeof(int),
 441                 .mode           = 0644,
 442                 .proc_handler   = proc_dointvec
 443         },
 444         {
 445                 .procname       = "igmp_max_msf",
 446                 .data           = &sysctl_igmp_max_msf,
 447                 .maxlen         = sizeof(int),
 448                 .mode           = 0644,
 449                 .proc_handler   = proc_dointvec
 450         },
 451         {
 452                 .procname       = "inet_peer_threshold",
 453                 .data           = &inet_peer_threshold,
 454                 .maxlen         = sizeof(int),
 455                 .mode           = 0644,
 456                 .proc_handler   = proc_dointvec
 457         },
 458         {
 459                 .procname       = "inet_peer_minttl",
 460                 .data           = &inet_peer_minttl,
 461                 .maxlen         = sizeof(int),
 462                 .mode           = 0644,
 463                 .proc_handler   = proc_dointvec_jiffies,
 464         },
 465         {
 466                 .procname       = "inet_peer_maxttl",
 467                 .data           = &inet_peer_maxttl,
 468                 .maxlen         = sizeof(int),
 469                 .mode           = 0644,
 470                 .proc_handler   = proc_dointvec_jiffies,
 471         },
 472         {
 473                 .procname       = "tcp_orphan_retries",
 474                 .data           = &sysctl_tcp_orphan_retries,
 475                 .maxlen         = sizeof(int),
 476                 .mode           = 0644,
 477                 .proc_handler   = proc_dointvec
 478         },
 479         {
 480                 .procname       = "tcp_fack",
 481                 .data           = &sysctl_tcp_fack,
 482                 .maxlen         = sizeof(int),
 483                 .mode           = 0644,
 484                 .proc_handler   = proc_dointvec
 485         },
 486         {
 487                 .procname       = "tcp_reordering",
 488                 .data           = &sysctl_tcp_reordering,
 489                 .maxlen         = sizeof(int),
 490                 .mode           = 0644,
 491                 .proc_handler   = proc_dointvec
 492         },
 493         {
 494                 .procname       = "tcp_ecn",
 495                 .data           = &sysctl_tcp_ecn,
 496                 .maxlen         = sizeof(int),
 497                 .mode           = 0644,
 498                 .proc_handler   = proc_dointvec
 499         },
 500         {
 501                 .procname       = "tcp_dsack",
 502                 .data           = &sysctl_tcp_dsack,
 503                 .maxlen         = sizeof(int),
 504                 .mode           = 0644,
 505                 .proc_handler   = proc_dointvec
 506         },
 507         {
 508                 .procname       = "tcp_wmem",
 509                 .data           = &sysctl_tcp_wmem,
 510                 .maxlen         = sizeof(sysctl_tcp_wmem),
 511                 .mode           = 0644,
 512                 .proc_handler   = proc_dointvec
 513         },
 514         {
 515                 .procname       = "tcp_rmem",
 516                 .data           = &sysctl_tcp_rmem,
 517                 .maxlen         = sizeof(sysctl_tcp_rmem),
 518                 .mode           = 0644,
 519                 .proc_handler   = proc_dointvec
 520         },
 521         {
 522                 .procname       = "tcp_app_win",
 523                 .data           = &sysctl_tcp_app_win,
 524                 .maxlen         = sizeof(int),
 525                 .mode           = 0644,
 526                 .proc_handler   = proc_dointvec
 527         },
 528         {
 529                 .procname       = "tcp_adv_win_scale",
 530                 .data           = &sysctl_tcp_adv_win_scale,
 531                 .maxlen         = sizeof(int),
 532                 .mode           = 0644,
 533                 .proc_handler   = proc_dointvec_minmax,
 534                 .extra1         = &tcp_adv_win_scale_min,
 535                 .extra2         = &tcp_adv_win_scale_max,
 536         },
 537         {
 538                 .procname       = "tcp_tw_reuse",
 539                 .data           = &sysctl_tcp_tw_reuse,
 540                 .maxlen         = sizeof(int),
 541                 .mode           = 0644,
 542                 .proc_handler   = proc_dointvec
 543         },
 544         {
 545                 .procname       = "tcp_frto",
 546                 .data           = &sysctl_tcp_frto,
 547                 .maxlen         = sizeof(int),
 548                 .mode           = 0644,
 549                 .proc_handler   = proc_dointvec
 550         },
 551         {
 552                 .procname       = "tcp_frto_response",
 553                 .data           = &sysctl_tcp_frto_response,
 554                 .maxlen         = sizeof(int),
 555                 .mode           = 0644,
 556                 .proc_handler   = proc_dointvec
 557         },
 558         {
 559                 .procname       = "tcp_low_latency",
 560                 .data           = &sysctl_tcp_low_latency,
 561                 .maxlen         = sizeof(int),
 562                 .mode           = 0644,
 563                 .proc_handler   = proc_dointvec
 564         },
 565         {
 566                 .procname       = "tcp_no_metrics_save",
 567                 .data           = &sysctl_tcp_nometrics_save,
 568                 .maxlen         = sizeof(int),
 569                 .mode           = 0644,
 570                 .proc_handler   = proc_dointvec,
 571         },
 572         {
 573                 .procname       = "tcp_moderate_rcvbuf",
 574                 .data           = &sysctl_tcp_moderate_rcvbuf,
 575                 .maxlen         = sizeof(int),
 576                 .mode           = 0644,
 577                 .proc_handler   = proc_dointvec,
 578         },
 579         {
 580                 .procname       = "tcp_tso_win_divisor",
 581                 .data           = &sysctl_tcp_tso_win_divisor,
 582                 .maxlen         = sizeof(int),
 583                 .mode           = 0644,
 584                 .proc_handler   = proc_dointvec,
 585         },
 586         {
 587                 .procname       = "tcp_congestion_control",
 588                 .mode           = 0644,
 589                 .maxlen         = TCP_CA_NAME_MAX,
 590                 .proc_handler   = proc_tcp_congestion_control,
 591         },
 592         {
 593                 .procname       = "tcp_abc",
 594                 .data           = &sysctl_tcp_abc,
 595                 .maxlen         = sizeof(int),
 596                 .mode           = 0644,
 597                 .proc_handler   = proc_dointvec,
 598         },
 599         {
 600                 .procname       = "tcp_mtu_probing",
 601                 .data           = &sysctl_tcp_mtu_probing,
 602                 .maxlen         = sizeof(int),
 603                 .mode           = 0644,
 604                 .proc_handler   = proc_dointvec,
 605         },
 606         {
 607                 .procname       = "tcp_base_mss",
 608                 .data           = &sysctl_tcp_base_mss,
 609                 .maxlen         = sizeof(int),
 610                 .mode           = 0644,
 611                 .proc_handler   = proc_dointvec,
 612         },
 613         {
 614                 .procname       = "tcp_workaround_signed_windows",
 615                 .data           = &sysctl_tcp_workaround_signed_windows,
 616                 .maxlen         = sizeof(int),
 617                 .mode           = 0644,
 618                 .proc_handler   = proc_dointvec
 619         },
 620         {
 621                 .procname       = "tcp_limit_output_bytes",
 622                 .data           = &sysctl_tcp_limit_output_bytes,
 623                 .maxlen         = sizeof(int),
 624                 .mode           = 0644,
 625                 .proc_handler   = proc_dointvec
 626         },
 627         {
 628                 .procname       = "tcp_challenge_ack_limit",
 629                 .data           = &sysctl_tcp_challenge_ack_limit,
 630                 .maxlen         = sizeof(int),
 631                 .mode           = 0644,
 632                 .proc_handler   = proc_dointvec
 633         },
 634 #ifdef CONFIG_NET_DMA
 635         {
 636                 .procname       = "tcp_dma_copybreak",
 637                 .data           = &sysctl_tcp_dma_copybreak,
 638                 .maxlen         = sizeof(int),
 639                 .mode           = 0644,
 640                 .proc_handler   = proc_dointvec
 641         },
 642 #endif
 643         {
 644                 .procname       = "tcp_slow_start_after_idle",
 645                 .data           = &sysctl_tcp_slow_start_after_idle,
 646                 .maxlen         = sizeof(int),
 647                 .mode           = 0644,
 648                 .proc_handler   = proc_dointvec
 649         },
 650 #ifdef CONFIG_NETLABEL
 651         {
 652                 .procname       = "cipso_cache_enable",
 653                 .data           = &cipso_v4_cache_enabled,
 654                 .maxlen         = sizeof(int),
 655                 .mode           = 0644,
 656                 .proc_handler   = proc_dointvec,
 657         },
 658         {
 659                 .procname       = "cipso_cache_bucket_size",
 660                 .data           = &cipso_v4_cache_bucketsize,
 661                 .maxlen         = sizeof(int),
 662                 .mode           = 0644,
 663                 .proc_handler   = proc_dointvec,
 664         },
 665         {
 666                 .procname       = "cipso_rbm_optfmt",
 667                 .data           = &cipso_v4_rbm_optfmt,
 668                 .maxlen         = sizeof(int),
 669                 .mode           = 0644,
 670                 .proc_handler   = proc_dointvec,
 671         },
 672         {
 673                 .procname       = "cipso_rbm_strictvalid",
 674                 .data           = &cipso_v4_rbm_strictvalid,
 675                 .maxlen         = sizeof(int),
 676                 .mode           = 0644,
 677                 .proc_handler   = proc_dointvec,
 678         },
 679 #endif /* CONFIG_NETLABEL */
 680         {
 681                 .procname       = "tcp_available_congestion_control",
 682                 .maxlen         = TCP_CA_BUF_MAX,
 683                 .mode           = 0444,
 684                 .proc_handler   = proc_tcp_available_congestion_control,
 685         },
 686         {
 687                 .procname       = "tcp_allowed_congestion_control",
 688                 .maxlen         = TCP_CA_BUF_MAX,
 689                 .mode           = 0644,
 690                 .proc_handler   = proc_allowed_congestion_control,
 691         },
 692         {
 693                 .procname       = "tcp_max_ssthresh",
 694                 .data           = &sysctl_tcp_max_ssthresh,
 695                 .maxlen         = sizeof(int),
 696                 .mode           = 0644,
 697                 .proc_handler   = proc_dointvec,
 698         },
 699         {
 700                 .procname       = "tcp_cookie_size",
 701                 .data           = &sysctl_tcp_cookie_size,
 702                 .maxlen         = sizeof(int),
 703                 .mode           = 0644,
 704                 .proc_handler   = proc_dointvec
 705         },
 706         {
 707                 .procname       = "tcp_thin_linear_timeouts",
 708                 .data           = &sysctl_tcp_thin_linear_timeouts,
 709                 .maxlen         = sizeof(int),
 710                 .mode           = 0644,
 711                 .proc_handler   = proc_dointvec
 712         },
 713         {
 714                 .procname       = "tcp_thin_dupack",
 715                 .data           = &sysctl_tcp_thin_dupack,
 716                 .maxlen         = sizeof(int),
 717                 .mode           = 0644,
 718                 .proc_handler   = proc_dointvec
 719         },
 720         {
 721                 .procname       = "tcp_early_retrans",
 722                 .data           = &sysctl_tcp_early_retrans,
 723                 .maxlen         = sizeof(int),
 724                 .mode           = 0644,
 725                 .proc_handler   = proc_dointvec_minmax,
 726                 .extra1         = &zero,
 727                 .extra2         = &two,
 728         },
 729         {
 730                 .procname       = "udp_mem",
 731                 .data           = &sysctl_udp_mem,
 732                 .maxlen         = sizeof(sysctl_udp_mem),
 733                 .mode           = 0644,
 734                 .proc_handler   = proc_doulongvec_minmax,
 735         },
 736         {
 737                 .procname       = "udp_rmem_min",
 738                 .data           = &sysctl_udp_rmem_min,
 739                 .maxlen         = sizeof(sysctl_udp_rmem_min),
 740                 .mode           = 0644,
 741                 .proc_handler   = proc_dointvec_minmax,
 742                 .extra1         = &zero
 743         },
 744         {
 745                 .procname       = "udp_wmem_min",
 746                 .data           = &sysctl_udp_wmem_min,
 747                 .maxlen         = sizeof(sysctl_udp_wmem_min),
 748                 .mode           = 0644,
 749                 .proc_handler   = proc_dointvec_minmax,
 750                 .extra1         = &zero
 751         },
 752         { }
 753 };
 754
 755 static struct ctl_table ipv4_net_table[] = {
 756         {
 757                 .procname       = "icmp_echo_ignore_all",
 758                 .data           = &init_net.ipv4.sysctl_icmp_echo_ignore_all,
 759                 .maxlen         = sizeof(int),
 760                 .mode           = 0644,
 761                 .proc_handler   = proc_dointvec
 762         },
 763         {
 764                 .procname       = "icmp_echo_ignore_broadcasts",
 765                 .data           = &init_net.ipv4.sysctl_icmp_echo_ignore_broadcasts,
 766                 .maxlen         = sizeof(int),
 767                 .mode           = 0644,
 768                 .proc_handler   = proc_dointvec
 769         },
 770         {
 771                 .procname       = "icmp_ignore_bogus_error_responses",
 772                 .data           = &init_net.ipv4.sysctl_icmp_ignore_bogus_error_responses,
 773                 .maxlen         = sizeof(int),
 774                 .mode           = 0644,
 775                 .proc_handler   = proc_dointvec
 776         },
 777         {
 778                 .procname       = "icmp_errors_use_inbound_ifaddr",
 779                 .data           = &init_net.ipv4.sysctl_icmp_errors_use_inbound_ifaddr,
 780                 .maxlen         = sizeof(int),
 781                 .mode           = 0644,
 782                 .proc_handler   = proc_dointvec
 783         },
 784         {
 785                 .procname       = "icmp_ratelimit",
 786                 .data           = &init_net.ipv4.sysctl_icmp_ratelimit,
 787                 .maxlen         = sizeof(int),
 788                 .mode           = 0644,
 789                 .proc_handler   = proc_dointvec_ms_jiffies,
 790         },
 791         {
 792                 .procname       = "icmp_ratemask",
 793                 .data           = &init_net.ipv4.sysctl_icmp_ratemask,
 794                 .maxlen         = sizeof(int),
 795                 .mode           = 0644,
 796                 .proc_handler   = proc_dointvec
 797         },
 798         {
 799                 .procname       = "ping_group_range",
 800                 .data           = &init_net.ipv4.sysctl_ping_group_range,
 801                 .maxlen         = sizeof(gid_t)*2,
 802                 .mode           = 0644,
 803                 .proc_handler   = ipv4_ping_group_range,
 804         },
 805         {
 806                 .procname       = "tcp_mem",
 807                 .maxlen         = sizeof(init_net.ipv4.sysctl_tcp_mem),
 808                 .mode           = 0644,
 809                 .proc_handler   = ipv4_tcp_mem,
 810         },
 811         { }
 812 };
 813
 814 static __net_init int ipv4_sysctl_init_net(struct net *net)
 815 {
 816         struct ctl_table *table;
 817
 818         table = ipv4_net_table;
 819         if (!net_eq(net, &init_net)) {
 820                 table = kmemdup(table, sizeof(ipv4_net_table), GFP_KERNEL);
 821                 if (table == NULL)
 822                         goto err_alloc;
 823
 824                 table[0].data =
 825                         &net->ipv4.sysctl_icmp_echo_ignore_all;
 826                 table[1].data =
 827                         &net->ipv4.sysctl_icmp_echo_ignore_broadcasts;
 828                 table[2].data =
 829                         &net->ipv4.sysctl_icmp_ignore_bogus_error_responses;
 830                 table[3].data =
 831                         &net->ipv4.sysctl_icmp_errors_use_inbound_ifaddr;
 832                 table[4].data =
 833                         &net->ipv4.sysctl_icmp_ratelimit;
 834                 table[5].data =
 835                         &net->ipv4.sysctl_icmp_ratemask;
 836                 table[6].data =
 837                         &net->ipv4.sysctl_ping_group_range;
 838
 839         }
 840
 841         /*
 842          * Sane defaults - nobody may create ping sockets.
 843          * Boot scripts should set this to distro-specific group.
 844          */
 845         net->ipv4.sysctl_ping_group_range[0] = make_kgid(&init_user_ns, 1);
 846         net->ipv4.sysctl_ping_group_range[1] = make_kgid(&init_user_ns, 0);
 847
 848         tcp_init_mem(net);
 849
 850         net->ipv4.ipv4_hdr = register_net_sysctl(net, "net/ipv4", table);
 851         if (net->ipv4.ipv4_hdr == NULL)
 852                 goto err_reg;
 853
 854         return 0;
 855
 856 err_reg:
 857         if (!net_eq(net, &init_net))
 858                 kfree(table);
 859 err_alloc:
 860         return -ENOMEM;
 861 }
 862
 863 static __net_exit void ipv4_sysctl_exit_net(struct net *net)
 864 {
 865         struct ctl_table *table;
 866
 867         table = net->ipv4.ipv4_hdr->ctl_table_arg;
 868         unregister_net_sysctl_table(net->ipv4.ipv4_hdr);
 869         kfree(table);
 870 }
 871
 872 static __net_initdata struct pernet_operations ipv4_sysctl_ops = {
 873         .init = ipv4_sysctl_init_net,
 874         .exit = ipv4_sysctl_exit_net,
 875 };
 876
 877 static __init int sysctl_ipv4_init(void)
 878 {
 879         struct ctl_table_header *hdr;
 880         struct ctl_table *i;
 881
 882         for (i = ipv4_table; i->procname; i++) {
 883                 if (strcmp(i->procname, "ip_local_reserved_ports") == 0) {
 884                         i->data = sysctl_local_reserved_ports;
 885                         break;
 886                 }
 887         }
 888         if (!i->procname)
 889                 return -EINVAL;
 890
 891         hdr = register_net_sysctl(&init_net, "net/ipv4", ipv4_table);
 892         if (hdr == NULL)
 893                 return -ENOMEM;
 894
 895         if (register_pernet_subsys(&ipv4_sysctl_ops)) {
 896                 unregister_net_sysctl_table(hdr);
 897                 return -ENOMEM;
 898         }
 899
 900         return 0;
 901 }
 902
 903 __initcall(sysctl_ipv4_init);