kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/mm.h>
  23 #include <linux/swap.h>
  24 #include <linux/slab.h>
  25 #include <linux/sysctl.h>
  26 #include <linux/proc_fs.h>
  27 #include <linux/security.h>
  28 #include <linux/ctype.h>
  29 #include <linux/utsname.h>
  30 #include <linux/smp_lock.h>
  31 #include <linux/fs.h>
  32 #include <linux/init.h>
  33 #include <linux/kernel.h>
  34 #include <linux/kobject.h>
  35 #include <linux/net.h>
  36 #include <linux/sysrq.h>
  37 #include <linux/highuid.h>
  38 #include <linux/writeback.h>
  39 #include <linux/hugetlb.h>
  40 #include <linux/initrd.h>
  41 #include <linux/key.h>
  42 #include <linux/times.h>
  43 #include <linux/limits.h>
  44 #include <linux/dcache.h>
  45 #include <linux/syscalls.h>
  46 #include <linux/vmstat.h>
  47 #include <linux/nfs_fs.h>
  48 #include <linux/acpi.h>
  49 #include <linux/reboot.h>
  50 #include <linux/ftrace.h>
  51
  52 #include <asm/uaccess.h>
  53 #include <asm/processor.h>
  54
  55 #ifdef CONFIG_X86
  56 #include <asm/nmi.h>
  57 #include <asm/stacktrace.h>
  58 #include <asm/io.h>
  59 #endif
  60
  61 static int deprecated_sysctl_warning(struct __sysctl_args *args);
  62
  63 #if defined(CONFIG_SYSCTL)
  64
  65 /* External variables not in a header file. */
  66 extern int C_A_D;
  67 extern int print_fatal_signals;
  68 extern int sysctl_overcommit_memory;
  69 extern int sysctl_overcommit_ratio;
  70 extern int sysctl_panic_on_oom;
  71 extern int sysctl_oom_kill_allocating_task;
  72 extern int sysctl_oom_dump_tasks;
  73 extern int max_threads;
  74 extern int core_uses_pid;
  75 extern int suid_dumpable;
  76 extern char core_pattern[];
  77 extern int pid_max;
  78 extern int min_free_kbytes;
  79 extern int pid_max_min, pid_max_max;
  80 extern int sysctl_drop_caches;
  81 extern int percpu_pagelist_fraction;
  82 extern int compat_log;
  83 extern int maps_protect;
  84 extern int latencytop_enabled;
  85 extern int sysctl_nr_open_min, sysctl_nr_open_max;
  86 #ifdef CONFIG_RCU_TORTURE_TEST
  87 extern int rcutorture_runnable;
  88 #endif /* #ifdef CONFIG_RCU_TORTURE_TEST */
  89
  90 /* Constants used for minimum and  maximum */
  91 #if defined(CONFIG_HIGHMEM) || defined(CONFIG_DETECT_SOFTLOCKUP)
  92 static int one = 1;
  93 #endif
  94
  95 #ifdef CONFIG_DETECT_SOFTLOCKUP
  96 static int sixty = 60;
  97 static int neg_one = -1;
  98 #endif
  99
 100 #ifdef CONFIG_MMU
 101 static int two = 2;
 102 #endif
 103
 104 static int zero;
 105 static int one_hundred = 100;
 106
 107 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 108 static int maxolduid = 65535;
 109 static int minolduid;
 110 static int min_percpu_pagelist_fract = 8;
 111
 112 static int ngroups_max = NGROUPS_MAX;
 113
 114 #ifdef CONFIG_MODULES
 115 extern char modprobe_path[];
 116 #endif
 117 #ifdef CONFIG_CHR_DEV_SG
 118 extern int sg_big_buff;
 119 #endif
 120
 121 #ifdef __sparc__
 122 extern char reboot_command [];
 123 extern int stop_a_enabled;
 124 extern int scons_pwroff;
 125 #endif
 126
 127 #ifdef __hppa__
 128 extern int pwrsw_enabled;
 129 extern int unaligned_enabled;
 130 #endif
 131
 132 #ifdef CONFIG_S390
 133 #ifdef CONFIG_MATHEMU
 134 extern int sysctl_ieee_emulation_warnings;
 135 #endif
 136 extern int sysctl_userprocess_debug;
 137 extern int spin_retry;
 138 #endif
 139
 140 #ifdef CONFIG_BSD_PROCESS_ACCT
 141 extern int acct_parm[];
 142 #endif
 143
 144 #ifdef CONFIG_IA64
 145 extern int no_unaligned_warning;
 146 #endif
 147
 148 #ifdef CONFIG_RT_MUTEXES
 149 extern int max_lock_depth;
 150 #endif
 151
 152 #ifdef CONFIG_PROC_SYSCTL
 153 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
 154                   void __user *buffer, size_t *lenp, loff_t *ppos);
 155 static int proc_dointvec_taint(struct ctl_table *table, int write, struct file *filp,
 156                                void __user *buffer, size_t *lenp, loff_t *ppos);
 157 #endif
 158
 159 static struct ctl_table root_table[];
 160 static struct ctl_table_root sysctl_table_root;
 161 static struct ctl_table_header root_table_header = {
 162         .count = 1,
 163         .ctl_table = root_table,
 164         .ctl_entry = LIST_HEAD_INIT(sysctl_table_root.default_set.list),
 165         .root = &sysctl_table_root,
 166         .set = &sysctl_table_root.default_set,
 167 };
 168 static struct ctl_table_root sysctl_table_root = {
 169         .root_list = LIST_HEAD_INIT(sysctl_table_root.root_list),
 170         .default_set.list = LIST_HEAD_INIT(root_table_header.ctl_entry),
 171 };
 172
 173 static struct ctl_table kern_table[];
 174 static struct ctl_table vm_table[];
 175 static struct ctl_table fs_table[];
 176 static struct ctl_table debug_table[];
 177 static struct ctl_table dev_table[];
 178 extern struct ctl_table random_table[];
 179 #ifdef CONFIG_INOTIFY_USER
 180 extern struct ctl_table inotify_table[];
 181 #endif
 182
 183 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 184 int sysctl_legacy_va_layout;
 185 #endif
 186
 187 extern int prove_locking;
 188 extern int lock_stat;
 189
 190 /* The default sysctl tables: */
 191
 192 static struct ctl_table root_table[] = {
 193         {
 194                 .ctl_name       = CTL_KERN,
 195                 .procname       = "kernel",
 196                 .mode           = 0555,
 197                 .child          = kern_table,
 198         },
 199         {
 200                 .ctl_name       = CTL_VM,
 201                 .procname       = "vm",
 202                 .mode           = 0555,
 203                 .child          = vm_table,
 204         },
 205         {
 206                 .ctl_name       = CTL_FS,
 207                 .procname       = "fs",
 208                 .mode           = 0555,
 209                 .child          = fs_table,
 210         },
 211         {
 212                 .ctl_name       = CTL_DEBUG,
 213                 .procname       = "debug",
 214                 .mode           = 0555,
 215                 .child          = debug_table,
 216         },
 217         {
 218                 .ctl_name       = CTL_DEV,
 219                 .procname       = "dev",
 220                 .mode           = 0555,
 221                 .child          = dev_table,
 222         },
 223 /*
 224  * NOTE: do not add new entries to this table unless you have read
 225  * Documentation/sysctl/ctl_unnumbered.txt
 226  */
 227         { .ctl_name = 0 }
 228 };
 229
 230 #ifdef CONFIG_SCHED_DEBUG
 231 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 232 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 233 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 234 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 235 #endif
 236
 237 static struct ctl_table kern_table[] = {
 238 #ifdef CONFIG_SCHED_DEBUG
 239         {
 240                 .ctl_name       = CTL_UNNUMBERED,
 241                 .procname       = "sched_min_granularity_ns",
 242                 .data           = &sysctl_sched_min_granularity,
 243                 .maxlen         = sizeof(unsigned int),
 244                 .mode           = 0644,
 245                 .proc_handler   = &sched_nr_latency_handler,
 246                 .strategy       = &sysctl_intvec,
 247                 .extra1         = &min_sched_granularity_ns,
 248                 .extra2         = &max_sched_granularity_ns,
 249         },
 250         {
 251                 .ctl_name       = CTL_UNNUMBERED,
 252                 .procname       = "sched_latency_ns",
 253                 .data           = &sysctl_sched_latency,
 254                 .maxlen         = sizeof(unsigned int),
 255                 .mode           = 0644,
 256                 .proc_handler   = &sched_nr_latency_handler,
 257                 .strategy       = &sysctl_intvec,
 258                 .extra1         = &min_sched_granularity_ns,
 259                 .extra2         = &max_sched_granularity_ns,
 260         },
 261         {
 262                 .ctl_name       = CTL_UNNUMBERED,
 263                 .procname       = "sched_wakeup_granularity_ns",
 264                 .data           = &sysctl_sched_wakeup_granularity,
 265                 .maxlen         = sizeof(unsigned int),
 266                 .mode           = 0644,
 267                 .proc_handler   = &proc_dointvec_minmax,
 268                 .strategy       = &sysctl_intvec,
 269                 .extra1         = &min_wakeup_granularity_ns,
 270                 .extra2         = &max_wakeup_granularity_ns,
 271         },
 272         {
 273                 .ctl_name       = CTL_UNNUMBERED,
 274                 .procname       = "sched_shares_ratelimit",
 275                 .data           = &sysctl_sched_shares_ratelimit,
 276                 .maxlen         = sizeof(unsigned int),
 277                 .mode           = 0644,
 278                 .proc_handler   = &proc_dointvec,
 279         },
 280         {
 281                 .ctl_name       = CTL_UNNUMBERED,
 282                 .procname       = "sched_child_runs_first",
 283                 .data           = &sysctl_sched_child_runs_first,
 284                 .maxlen         = sizeof(unsigned int),
 285                 .mode           = 0644,
 286                 .proc_handler   = &proc_dointvec,
 287         },
 288         {
 289                 .ctl_name       = CTL_UNNUMBERED,
 290                 .procname       = "sched_features",
 291                 .data           = &sysctl_sched_features,
 292                 .maxlen         = sizeof(unsigned int),
 293                 .mode           = 0644,
 294                 .proc_handler   = &proc_dointvec,
 295         },
 296         {
 297                 .ctl_name       = CTL_UNNUMBERED,
 298                 .procname       = "sched_migration_cost",
 299                 .data           = &sysctl_sched_migration_cost,
 300                 .maxlen         = sizeof(unsigned int),
 301                 .mode           = 0644,
 302                 .proc_handler   = &proc_dointvec,
 303         },
 304         {
 305                 .ctl_name       = CTL_UNNUMBERED,
 306                 .procname       = "sched_nr_migrate",
 307                 .data           = &sysctl_sched_nr_migrate,
 308                 .maxlen         = sizeof(unsigned int),
 309                 .mode           = 0644,
 310                 .proc_handler   = &proc_dointvec,
 311         },
 312 #endif
 313         {
 314                 .ctl_name       = CTL_UNNUMBERED,
 315                 .procname       = "sched_rt_period_us",
 316                 .data           = &sysctl_sched_rt_period,
 317                 .maxlen         = sizeof(unsigned int),
 318                 .mode           = 0644,
 319                 .proc_handler   = &sched_rt_handler,
 320         },
 321         {
 322                 .ctl_name       = CTL_UNNUMBERED,
 323                 .procname       = "sched_rt_runtime_us",
 324                 .data           = &sysctl_sched_rt_runtime,
 325                 .maxlen         = sizeof(int),
 326                 .mode           = 0644,
 327                 .proc_handler   = &sched_rt_handler,
 328         },
 329         {
 330                 .ctl_name       = CTL_UNNUMBERED,
 331                 .procname       = "sched_compat_yield",
 332                 .data           = &sysctl_sched_compat_yield,
 333                 .maxlen         = sizeof(unsigned int),
 334                 .mode           = 0644,
 335                 .proc_handler   = &proc_dointvec,
 336         },
 337 #ifdef CONFIG_PROVE_LOCKING
 338         {
 339                 .ctl_name       = CTL_UNNUMBERED,
 340                 .procname       = "prove_locking",
 341                 .data           = &prove_locking,
 342                 .maxlen         = sizeof(int),
 343                 .mode           = 0644,
 344                 .proc_handler   = &proc_dointvec,
 345         },
 346 #endif
 347 #ifdef CONFIG_LOCK_STAT
 348         {
 349                 .ctl_name       = CTL_UNNUMBERED,
 350                 .procname       = "lock_stat",
 351                 .data           = &lock_stat,
 352                 .maxlen         = sizeof(int),
 353                 .mode           = 0644,
 354                 .proc_handler   = &proc_dointvec,
 355         },
 356 #endif
 357         {
 358                 .ctl_name       = KERN_PANIC,
 359                 .procname       = "panic",
 360                 .data           = &panic_timeout,
 361                 .maxlen         = sizeof(int),
 362                 .mode           = 0644,
 363                 .proc_handler   = &proc_dointvec,
 364         },
 365         {
 366                 .ctl_name       = KERN_CORE_USES_PID,
 367                 .procname       = "core_uses_pid",
 368                 .data           = &core_uses_pid,
 369                 .maxlen         = sizeof(int),
 370                 .mode           = 0644,
 371                 .proc_handler   = &proc_dointvec,
 372         },
 373         {
 374                 .ctl_name       = KERN_CORE_PATTERN,
 375                 .procname       = "core_pattern",
 376                 .data           = core_pattern,
 377                 .maxlen         = CORENAME_MAX_SIZE,
 378                 .mode           = 0644,
 379                 .proc_handler   = &proc_dostring,
 380                 .strategy       = &sysctl_string,
 381         },
 382 #ifdef CONFIG_PROC_SYSCTL
 383         {
 384                 .procname       = "tainted",
 385                 .data           = &tainted,
 386                 .maxlen         = sizeof(int),
 387                 .mode           = 0644,
 388                 .proc_handler   = &proc_dointvec_taint,
 389         },
 390 #endif
 391 #ifdef CONFIG_LATENCYTOP
 392         {
 393                 .procname       = "latencytop",
 394                 .data           = &latencytop_enabled,
 395                 .maxlen         = sizeof(int),
 396                 .mode           = 0644,
 397                 .proc_handler   = &proc_dointvec,
 398         },
 399 #endif
 400 #ifdef CONFIG_BLK_DEV_INITRD
 401         {
 402                 .ctl_name       = KERN_REALROOTDEV,
 403                 .procname       = "real-root-dev",
 404                 .data           = &real_root_dev,
 405                 .maxlen         = sizeof(int),
 406                 .mode           = 0644,
 407                 .proc_handler   = &proc_dointvec,
 408         },
 409 #endif
 410         {
 411                 .ctl_name       = CTL_UNNUMBERED,
 412                 .procname       = "print-fatal-signals",
 413                 .data           = &print_fatal_signals,
 414                 .maxlen         = sizeof(int),
 415                 .mode           = 0644,
 416                 .proc_handler   = &proc_dointvec,
 417         },
 418 #ifdef __sparc__
 419         {
 420                 .ctl_name       = KERN_SPARC_REBOOT,
 421                 .procname       = "reboot-cmd",
 422                 .data           = reboot_command,
 423                 .maxlen         = 256,
 424                 .mode           = 0644,
 425                 .proc_handler   = &proc_dostring,
 426                 .strategy       = &sysctl_string,
 427         },
 428         {
 429                 .ctl_name       = KERN_SPARC_STOP_A,
 430                 .procname       = "stop-a",
 431                 .data           = &stop_a_enabled,
 432                 .maxlen         = sizeof (int),
 433                 .mode           = 0644,
 434                 .proc_handler   = &proc_dointvec,
 435         },
 436         {
 437                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 438                 .procname       = "scons-poweroff",
 439                 .data           = &scons_pwroff,
 440                 .maxlen         = sizeof (int),
 441                 .mode           = 0644,
 442                 .proc_handler   = &proc_dointvec,
 443         },
 444 #endif
 445 #ifdef __hppa__
 446         {
 447                 .ctl_name       = KERN_HPPA_PWRSW,
 448                 .procname       = "soft-power",
 449                 .data           = &pwrsw_enabled,
 450                 .maxlen         = sizeof (int),
 451                 .mode           = 0644,
 452                 .proc_handler   = &proc_dointvec,
 453         },
 454         {
 455                 .ctl_name       = KERN_HPPA_UNALIGNED,
 456                 .procname       = "unaligned-trap",
 457                 .data           = &unaligned_enabled,
 458                 .maxlen         = sizeof (int),
 459                 .mode           = 0644,
 460                 .proc_handler   = &proc_dointvec,
 461         },
 462 #endif
 463         {
 464                 .ctl_name       = KERN_CTLALTDEL,
 465                 .procname       = "ctrl-alt-del",
 466                 .data           = &C_A_D,
 467                 .maxlen         = sizeof(int),
 468                 .mode           = 0644,
 469                 .proc_handler   = &proc_dointvec,
 470         },
 471 #ifdef CONFIG_FTRACE
 472         {
 473                 .ctl_name       = CTL_UNNUMBERED,
 474                 .procname       = "ftrace_enabled",
 475                 .data           = &ftrace_enabled,
 476                 .maxlen         = sizeof(int),
 477                 .mode           = 0644,
 478                 .proc_handler   = &ftrace_enable_sysctl,
 479         },
 480 #endif
 481 #ifdef CONFIG_MODULES
 482         {
 483                 .ctl_name       = KERN_MODPROBE,
 484                 .procname       = "modprobe",
 485                 .data           = &modprobe_path,
 486                 .maxlen         = KMOD_PATH_LEN,
 487                 .mode           = 0644,
 488                 .proc_handler   = &proc_dostring,
 489                 .strategy       = &sysctl_string,
 490         },
 491 #endif
 492 #if defined(CONFIG_HOTPLUG) && defined(CONFIG_NET)
 493         {
 494                 .ctl_name       = KERN_HOTPLUG,
 495                 .procname       = "hotplug",
 496                 .data           = &uevent_helper,
 497                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 498                 .mode           = 0644,
 499                 .proc_handler   = &proc_dostring,
 500                 .strategy       = &sysctl_string,
 501         },
 502 #endif
 503 #ifdef CONFIG_CHR_DEV_SG
 504         {
 505                 .ctl_name       = KERN_SG_BIG_BUFF,
 506                 .procname       = "sg-big-buff",
 507                 .data           = &sg_big_buff,
 508                 .maxlen         = sizeof (int),
 509                 .mode           = 0444,
 510                 .proc_handler   = &proc_dointvec,
 511         },
 512 #endif
 513 #ifdef CONFIG_BSD_PROCESS_ACCT
 514         {
 515                 .ctl_name       = KERN_ACCT,
 516                 .procname       = "acct",
 517                 .data           = &acct_parm,
 518                 .maxlen         = 3*sizeof(int),
 519                 .mode           = 0644,
 520                 .proc_handler   = &proc_dointvec,
 521         },
 522 #endif
 523 #ifdef CONFIG_MAGIC_SYSRQ
 524         {
 525                 .ctl_name       = KERN_SYSRQ,
 526                 .procname       = "sysrq",
 527                 .data           = &__sysrq_enabled,
 528                 .maxlen         = sizeof (int),
 529                 .mode           = 0644,
 530                 .proc_handler   = &proc_dointvec,
 531         },
 532 #endif
 533 #ifdef CONFIG_PROC_SYSCTL
 534         {
 535                 .procname       = "cad_pid",
 536                 .data           = NULL,
 537                 .maxlen         = sizeof (int),
 538                 .mode           = 0600,
 539                 .proc_handler   = &proc_do_cad_pid,
 540         },
 541 #endif
 542         {
 543                 .ctl_name       = KERN_MAX_THREADS,
 544                 .procname       = "threads-max",
 545                 .data           = &max_threads,
 546                 .maxlen         = sizeof(int),
 547                 .mode           = 0644,
 548                 .proc_handler   = &proc_dointvec,
 549         },
 550         {
 551                 .ctl_name       = KERN_RANDOM,
 552                 .procname       = "random",
 553                 .mode           = 0555,
 554                 .child          = random_table,
 555         },
 556         {
 557                 .ctl_name       = KERN_OVERFLOWUID,
 558                 .procname       = "overflowuid",
 559                 .data           = &overflowuid,
 560                 .maxlen         = sizeof(int),
 561                 .mode           = 0644,
 562                 .proc_handler   = &proc_dointvec_minmax,
 563                 .strategy       = &sysctl_intvec,
 564                 .extra1         = &minolduid,
 565                 .extra2         = &maxolduid,
 566         },
 567         {
 568                 .ctl_name       = KERN_OVERFLOWGID,
 569                 .procname       = "overflowgid",
 570                 .data           = &overflowgid,
 571                 .maxlen         = sizeof(int),
 572                 .mode           = 0644,
 573                 .proc_handler   = &proc_dointvec_minmax,
 574                 .strategy       = &sysctl_intvec,
 575                 .extra1         = &minolduid,
 576                 .extra2         = &maxolduid,
 577         },
 578 #ifdef CONFIG_S390
 579 #ifdef CONFIG_MATHEMU
 580         {
 581                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 582                 .procname       = "ieee_emulation_warnings",
 583                 .data           = &sysctl_ieee_emulation_warnings,
 584                 .maxlen         = sizeof(int),
 585                 .mode           = 0644,
 586                 .proc_handler   = &proc_dointvec,
 587         },
 588 #endif
 589         {
 590                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 591                 .procname       = "userprocess_debug",
 592                 .data           = &sysctl_userprocess_debug,
 593                 .maxlen         = sizeof(int),
 594                 .mode           = 0644,
 595                 .proc_handler   = &proc_dointvec,
 596         },
 597 #endif
 598         {
 599                 .ctl_name       = KERN_PIDMAX,
 600                 .procname       = "pid_max",
 601                 .data           = &pid_max,
 602                 .maxlen         = sizeof (int),
 603                 .mode           = 0644,
 604                 .proc_handler   = &proc_dointvec_minmax,
 605                 .strategy       = sysctl_intvec,
 606                 .extra1         = &pid_max_min,
 607                 .extra2         = &pid_max_max,
 608         },
 609         {
 610                 .ctl_name       = KERN_PANIC_ON_OOPS,
 611                 .procname       = "panic_on_oops",
 612                 .data           = &panic_on_oops,
 613                 .maxlen         = sizeof(int),
 614                 .mode           = 0644,
 615                 .proc_handler   = &proc_dointvec,
 616         },
 617 #if defined CONFIG_PRINTK
 618         {
 619                 .ctl_name       = KERN_PRINTK,
 620                 .procname       = "printk",
 621                 .data           = &console_loglevel,
 622                 .maxlen         = 4*sizeof(int),
 623                 .mode           = 0644,
 624                 .proc_handler   = &proc_dointvec,
 625         },
 626         {
 627                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 628                 .procname       = "printk_ratelimit",
 629                 .data           = &printk_ratelimit_state.interval,
 630                 .maxlen         = sizeof(int),
 631                 .mode           = 0644,
 632                 .proc_handler   = &proc_dointvec_jiffies,
 633                 .strategy       = &sysctl_jiffies,
 634         },
 635         {
 636                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 637                 .procname       = "printk_ratelimit_burst",
 638                 .data           = &printk_ratelimit_state.burst,
 639                 .maxlen         = sizeof(int),
 640                 .mode           = 0644,
 641                 .proc_handler   = &proc_dointvec,
 642         },
 643 #endif
 644         {
 645                 .ctl_name       = KERN_NGROUPS_MAX,
 646                 .procname       = "ngroups_max",
 647                 .data           = &ngroups_max,
 648                 .maxlen         = sizeof (int),
 649                 .mode           = 0444,
 650                 .proc_handler   = &proc_dointvec,
 651         },
 652 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 653         {
 654                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 655                 .procname       = "unknown_nmi_panic",
 656                 .data           = &unknown_nmi_panic,
 657                 .maxlen         = sizeof (int),
 658                 .mode           = 0644,
 659                 .proc_handler   = &proc_dointvec,
 660         },
 661         {
 662                 .procname       = "nmi_watchdog",
 663                 .data           = &nmi_watchdog_enabled,
 664                 .maxlen         = sizeof (int),
 665                 .mode           = 0644,
 666                 .proc_handler   = &proc_nmi_enabled,
 667         },
 668 #endif
 669 #if defined(CONFIG_X86)
 670         {
 671                 .ctl_name       = KERN_PANIC_ON_NMI,
 672                 .procname       = "panic_on_unrecovered_nmi",
 673                 .data           = &panic_on_unrecovered_nmi,
 674                 .maxlen         = sizeof(int),
 675                 .mode           = 0644,
 676                 .proc_handler   = &proc_dointvec,
 677         },
 678         {
 679                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 680                 .procname       = "bootloader_type",
 681                 .data           = &bootloader_type,
 682                 .maxlen         = sizeof (int),
 683                 .mode           = 0444,
 684                 .proc_handler   = &proc_dointvec,
 685         },
 686         {
 687                 .ctl_name       = CTL_UNNUMBERED,
 688                 .procname       = "kstack_depth_to_print",
 689                 .data           = &kstack_depth_to_print,
 690                 .maxlen         = sizeof(int),
 691                 .mode           = 0644,
 692                 .proc_handler   = &proc_dointvec,
 693         },
 694         {
 695                 .ctl_name       = CTL_UNNUMBERED,
 696                 .procname       = "io_delay_type",
 697                 .data           = &io_delay_type,
 698                 .maxlen         = sizeof(int),
 699                 .mode           = 0644,
 700                 .proc_handler   = &proc_dointvec,
 701         },
 702 #endif
 703 #if defined(CONFIG_MMU)
 704         {
 705                 .ctl_name       = KERN_RANDOMIZE,
 706                 .procname       = "randomize_va_space",
 707                 .data           = &randomize_va_space,
 708                 .maxlen         = sizeof(int),
 709                 .mode           = 0644,
 710                 .proc_handler   = &proc_dointvec,
 711         },
 712 #endif
 713 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 714         {
 715                 .ctl_name       = KERN_SPIN_RETRY,
 716                 .procname       = "spin_retry",
 717                 .data           = &spin_retry,
 718                 .maxlen         = sizeof (int),
 719                 .mode           = 0644,
 720                 .proc_handler   = &proc_dointvec,
 721         },
 722 #endif
 723 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 724         {
 725                 .procname       = "acpi_video_flags",
 726                 .data           = &acpi_realmode_flags,
 727                 .maxlen         = sizeof (unsigned long),
 728                 .mode           = 0644,
 729                 .proc_handler   = &proc_doulongvec_minmax,
 730         },
 731 #endif
 732 #ifdef CONFIG_IA64
 733         {
 734                 .ctl_name       = KERN_IA64_UNALIGNED,
 735                 .procname       = "ignore-unaligned-usertrap",
 736                 .data           = &no_unaligned_warning,
 737                 .maxlen         = sizeof (int),
 738                 .mode           = 0644,
 739                 .proc_handler   = &proc_dointvec,
 740         },
 741 #endif
 742 #ifdef CONFIG_DETECT_SOFTLOCKUP
 743         {
 744                 .ctl_name       = CTL_UNNUMBERED,
 745                 .procname       = "softlockup_panic",
 746                 .data           = &softlockup_panic,
 747                 .maxlen         = sizeof(int),
 748                 .mode           = 0644,
 749                 .proc_handler   = &proc_dointvec_minmax,
 750                 .strategy       = &sysctl_intvec,
 751                 .extra1         = &zero,
 752                 .extra2         = &one,
 753         },
 754         {
 755                 .ctl_name       = CTL_UNNUMBERED,
 756                 .procname       = "softlockup_thresh",
 757                 .data           = &softlockup_thresh,
 758                 .maxlen         = sizeof(int),
 759                 .mode           = 0644,
 760                 .proc_handler   = &proc_dointvec_minmax,
 761                 .strategy       = &sysctl_intvec,
 762                 .extra1         = &neg_one,
 763                 .extra2         = &sixty,
 764         },
 765         {
 766                 .ctl_name       = CTL_UNNUMBERED,
 767                 .procname       = "hung_task_check_count",
 768                 .data           = &sysctl_hung_task_check_count,
 769                 .maxlen         = sizeof(unsigned long),
 770                 .mode           = 0644,
 771                 .proc_handler   = &proc_doulongvec_minmax,
 772                 .strategy       = &sysctl_intvec,
 773         },
 774         {
 775                 .ctl_name       = CTL_UNNUMBERED,
 776                 .procname       = "hung_task_timeout_secs",
 777                 .data           = &sysctl_hung_task_timeout_secs,
 778                 .maxlen         = sizeof(unsigned long),
 779                 .mode           = 0644,
 780                 .proc_handler   = &proc_doulongvec_minmax,
 781                 .strategy       = &sysctl_intvec,
 782         },
 783         {
 784                 .ctl_name       = CTL_UNNUMBERED,
 785                 .procname       = "hung_task_warnings",
 786                 .data           = &sysctl_hung_task_warnings,
 787                 .maxlen         = sizeof(unsigned long),
 788                 .mode           = 0644,
 789                 .proc_handler   = &proc_doulongvec_minmax,
 790                 .strategy       = &sysctl_intvec,
 791         },
 792 #endif
 793 #ifdef CONFIG_COMPAT
 794         {
 795                 .ctl_name       = KERN_COMPAT_LOG,
 796                 .procname       = "compat-log",
 797                 .data           = &compat_log,
 798                 .maxlen         = sizeof (int),
 799                 .mode           = 0644,
 800                 .proc_handler   = &proc_dointvec,
 801         },
 802 #endif
 803 #ifdef CONFIG_RT_MUTEXES
 804         {
 805                 .ctl_name       = KERN_MAX_LOCK_DEPTH,
 806                 .procname       = "max_lock_depth",
 807                 .data           = &max_lock_depth,
 808                 .maxlen         = sizeof(int),
 809                 .mode           = 0644,
 810                 .proc_handler   = &proc_dointvec,
 811         },
 812 #endif
 813 #ifdef CONFIG_PROC_FS
 814         {
 815                 .ctl_name       = CTL_UNNUMBERED,
 816                 .procname       = "maps_protect",
 817                 .data           = &maps_protect,
 818                 .maxlen         = sizeof(int),
 819                 .mode           = 0644,
 820                 .proc_handler   = &proc_dointvec,
 821         },
 822 #endif
 823         {
 824                 .ctl_name       = CTL_UNNUMBERED,
 825                 .procname       = "poweroff_cmd",
 826                 .data           = &poweroff_cmd,
 827                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 828                 .mode           = 0644,
 829                 .proc_handler   = &proc_dostring,
 830                 .strategy       = &sysctl_string,
 831         },
 832 #ifdef CONFIG_KEYS
 833         {
 834                 .ctl_name       = CTL_UNNUMBERED,
 835                 .procname       = "keys",
 836                 .mode           = 0555,
 837                 .child          = key_sysctls,
 838         },
 839 #endif
 840 #ifdef CONFIG_RCU_TORTURE_TEST
 841         {
 842                 .ctl_name       = CTL_UNNUMBERED,
 843                 .procname       = "rcutorture_runnable",
 844                 .data           = &rcutorture_runnable,
 845                 .maxlen         = sizeof(int),
 846                 .mode           = 0644,
 847                 .proc_handler   = &proc_dointvec,
 848         },
 849 #endif
 850 /*
 851  * NOTE: do not add new entries to this table unless you have read
 852  * Documentation/sysctl/ctl_unnumbered.txt
 853  */
 854         { .ctl_name = 0 }
 855 };
 856
 857 static struct ctl_table vm_table[] = {
 858         {
 859                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 860                 .procname       = "overcommit_memory",
 861                 .data           = &sysctl_overcommit_memory,
 862                 .maxlen         = sizeof(sysctl_overcommit_memory),
 863                 .mode           = 0644,
 864                 .proc_handler   = &proc_dointvec,
 865         },
 866         {
 867                 .ctl_name       = VM_PANIC_ON_OOM,
 868                 .procname       = "panic_on_oom",
 869                 .data           = &sysctl_panic_on_oom,
 870                 .maxlen         = sizeof(sysctl_panic_on_oom),
 871                 .mode           = 0644,
 872                 .proc_handler   = &proc_dointvec,
 873         },
 874         {
 875                 .ctl_name       = CTL_UNNUMBERED,
 876                 .procname       = "oom_kill_allocating_task",
 877                 .data           = &sysctl_oom_kill_allocating_task,
 878                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
 879                 .mode           = 0644,
 880                 .proc_handler   = &proc_dointvec,
 881         },
 882         {
 883                 .ctl_name       = CTL_UNNUMBERED,
 884                 .procname       = "oom_dump_tasks",
 885                 .data           = &sysctl_oom_dump_tasks,
 886                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
 887                 .mode           = 0644,
 888                 .proc_handler   = &proc_dointvec,
 889         },
 890         {
 891                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 892                 .procname       = "overcommit_ratio",
 893                 .data           = &sysctl_overcommit_ratio,
 894                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 895                 .mode           = 0644,
 896                 .proc_handler   = &proc_dointvec,
 897         },
 898         {
 899                 .ctl_name       = VM_PAGE_CLUSTER,
 900                 .procname       = "page-cluster",
 901                 .data           = &page_cluster,
 902                 .maxlen         = sizeof(int),
 903                 .mode           = 0644,
 904                 .proc_handler   = &proc_dointvec,
 905         },
 906         {
 907                 .ctl_name       = VM_DIRTY_BACKGROUND,
 908                 .procname       = "dirty_background_ratio",
 909                 .data           = &dirty_background_ratio,
 910                 .maxlen         = sizeof(dirty_background_ratio),
 911                 .mode           = 0644,
 912                 .proc_handler   = &proc_dointvec_minmax,
 913                 .strategy       = &sysctl_intvec,
 914                 .extra1         = &zero,
 915                 .extra2         = &one_hundred,
 916         },
 917         {
 918                 .ctl_name       = VM_DIRTY_RATIO,
 919                 .procname       = "dirty_ratio",
 920                 .data           = &vm_dirty_ratio,
 921                 .maxlen         = sizeof(vm_dirty_ratio),
 922                 .mode           = 0644,
 923                 .proc_handler   = &dirty_ratio_handler,
 924                 .strategy       = &sysctl_intvec,
 925                 .extra1         = &zero,
 926                 .extra2         = &one_hundred,
 927         },
 928         {
 929                 .procname       = "dirty_writeback_centisecs",
 930                 .data           = &dirty_writeback_interval,
 931                 .maxlen         = sizeof(dirty_writeback_interval),
 932                 .mode           = 0644,
 933                 .proc_handler   = &dirty_writeback_centisecs_handler,
 934         },
 935         {
 936                 .procname       = "dirty_expire_centisecs",
 937                 .data           = &dirty_expire_interval,
 938                 .maxlen         = sizeof(dirty_expire_interval),
 939                 .mode           = 0644,
 940                 .proc_handler   = &proc_dointvec_userhz_jiffies,
 941         },
 942         {
 943                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
 944                 .procname       = "nr_pdflush_threads",
 945                 .data           = &nr_pdflush_threads,
 946                 .maxlen         = sizeof nr_pdflush_threads,
 947                 .mode           = 0444 /* read-only*/,
 948                 .proc_handler   = &proc_dointvec,
 949         },
 950         {
 951                 .ctl_name       = VM_SWAPPINESS,
 952                 .procname       = "swappiness",
 953                 .data           = &vm_swappiness,
 954                 .maxlen         = sizeof(vm_swappiness),
 955                 .mode           = 0644,
 956                 .proc_handler   = &proc_dointvec_minmax,
 957                 .strategy       = &sysctl_intvec,
 958                 .extra1         = &zero,
 959                 .extra2         = &one_hundred,
 960         },
 961 #ifdef CONFIG_HUGETLB_PAGE
 962          {
 963                 .procname       = "nr_hugepages",
 964                 .data           = NULL,
 965                 .maxlen         = sizeof(unsigned long),
 966                 .mode           = 0644,
 967                 .proc_handler   = &hugetlb_sysctl_handler,
 968                 .extra1         = (void *)&hugetlb_zero,
 969                 .extra2         = (void *)&hugetlb_infinity,
 970          },
 971          {
 972                 .ctl_name       = VM_HUGETLB_GROUP,
 973                 .procname       = "hugetlb_shm_group",
 974                 .data           = &sysctl_hugetlb_shm_group,
 975                 .maxlen         = sizeof(gid_t),
 976                 .mode           = 0644,
 977                 .proc_handler   = &proc_dointvec,
 978          },
 979          {
 980                 .ctl_name       = CTL_UNNUMBERED,
 981                 .procname       = "hugepages_treat_as_movable",
 982                 .data           = &hugepages_treat_as_movable,
 983                 .maxlen         = sizeof(int),
 984                 .mode           = 0644,
 985                 .proc_handler   = &hugetlb_treat_movable_handler,
 986         },
 987         {
 988                 .ctl_name       = CTL_UNNUMBERED,
 989                 .procname       = "nr_overcommit_hugepages",
 990                 .data           = NULL,
 991                 .maxlen         = sizeof(unsigned long),
 992                 .mode           = 0644,
 993                 .proc_handler   = &hugetlb_overcommit_handler,
 994                 .extra1         = (void *)&hugetlb_zero,
 995                 .extra2         = (void *)&hugetlb_infinity,
 996         },
 997 #endif
 998         {
 999                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
1000                 .procname       = "lowmem_reserve_ratio",
1001                 .data           = &sysctl_lowmem_reserve_ratio,
1002                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1003                 .mode           = 0644,
1004                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
1005                 .strategy       = &sysctl_intvec,
1006         },
1007         {
1008                 .ctl_name       = VM_DROP_PAGECACHE,
1009                 .procname       = "drop_caches",
1010                 .data           = &sysctl_drop_caches,
1011                 .maxlen         = sizeof(int),
1012                 .mode           = 0644,
1013                 .proc_handler   = drop_caches_sysctl_handler,
1014                 .strategy       = &sysctl_intvec,
1015         },
1016         {
1017                 .ctl_name       = VM_MIN_FREE_KBYTES,
1018                 .procname       = "min_free_kbytes",
1019                 .data           = &min_free_kbytes,
1020                 .maxlen         = sizeof(min_free_kbytes),
1021                 .mode           = 0644,
1022                 .proc_handler   = &min_free_kbytes_sysctl_handler,
1023                 .strategy       = &sysctl_intvec,
1024                 .extra1         = &zero,
1025         },
1026         {
1027                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
1028                 .procname       = "percpu_pagelist_fraction",
1029                 .data           = &percpu_pagelist_fraction,
1030                 .maxlen         = sizeof(percpu_pagelist_fraction),
1031                 .mode           = 0644,
1032                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
1033                 .strategy       = &sysctl_intvec,
1034                 .extra1         = &min_percpu_pagelist_fract,
1035         },
1036 #ifdef CONFIG_MMU
1037         {
1038                 .ctl_name       = VM_MAX_MAP_COUNT,
1039                 .procname       = "max_map_count",
1040                 .data           = &sysctl_max_map_count,
1041                 .maxlen         = sizeof(sysctl_max_map_count),
1042                 .mode           = 0644,
1043                 .proc_handler   = &proc_dointvec
1044         },
1045 #endif
1046         {
1047                 .ctl_name       = VM_LAPTOP_MODE,
1048                 .procname       = "laptop_mode",
1049                 .data           = &laptop_mode,
1050                 .maxlen         = sizeof(laptop_mode),
1051                 .mode           = 0644,
1052                 .proc_handler   = &proc_dointvec_jiffies,
1053                 .strategy       = &sysctl_jiffies,
1054         },
1055         {
1056                 .ctl_name       = VM_BLOCK_DUMP,
1057                 .procname       = "block_dump",
1058                 .data           = &block_dump,
1059                 .maxlen         = sizeof(block_dump),
1060                 .mode           = 0644,
1061                 .proc_handler   = &proc_dointvec,
1062                 .strategy       = &sysctl_intvec,
1063                 .extra1         = &zero,
1064         },
1065         {
1066                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
1067                 .procname       = "vfs_cache_pressure",
1068                 .data           = &sysctl_vfs_cache_pressure,
1069                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1070                 .mode           = 0644,
1071                 .proc_handler   = &proc_dointvec,
1072                 .strategy       = &sysctl_intvec,
1073                 .extra1         = &zero,
1074         },
1075 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1076         {
1077                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
1078                 .procname       = "legacy_va_layout",
1079                 .data           = &sysctl_legacy_va_layout,
1080                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1081                 .mode           = 0644,
1082                 .proc_handler   = &proc_dointvec,
1083                 .strategy       = &sysctl_intvec,
1084                 .extra1         = &zero,
1085         },
1086 #endif
1087 #ifdef CONFIG_NUMA
1088         {
1089                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
1090                 .procname       = "zone_reclaim_mode",
1091                 .data           = &zone_reclaim_mode,
1092                 .maxlen         = sizeof(zone_reclaim_mode),
1093                 .mode           = 0644,
1094                 .proc_handler   = &proc_dointvec,
1095                 .strategy       = &sysctl_intvec,
1096                 .extra1         = &zero,
1097         },
1098         {
1099                 .ctl_name       = VM_MIN_UNMAPPED,
1100                 .procname       = "min_unmapped_ratio",
1101                 .data           = &sysctl_min_unmapped_ratio,
1102                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1103                 .mode           = 0644,
1104                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
1105                 .strategy       = &sysctl_intvec,
1106                 .extra1         = &zero,
1107                 .extra2         = &one_hundred,
1108         },
1109         {
1110                 .ctl_name       = VM_MIN_SLAB,
1111                 .procname       = "min_slab_ratio",
1112                 .data           = &sysctl_min_slab_ratio,
1113                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1114                 .mode           = 0644,
1115                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
1116                 .strategy       = &sysctl_intvec,
1117                 .extra1         = &zero,
1118                 .extra2         = &one_hundred,
1119         },
1120 #endif
1121 #ifdef CONFIG_SMP
1122         {
1123                 .ctl_name       = CTL_UNNUMBERED,
1124                 .procname       = "stat_interval",
1125                 .data           = &sysctl_stat_interval,
1126                 .maxlen         = sizeof(sysctl_stat_interval),
1127                 .mode           = 0644,
1128                 .proc_handler   = &proc_dointvec_jiffies,
1129                 .strategy       = &sysctl_jiffies,
1130         },
1131 #endif
1132 #ifdef CONFIG_SECURITY
1133         {
1134                 .ctl_name       = CTL_UNNUMBERED,
1135                 .procname       = "mmap_min_addr",
1136                 .data           = &mmap_min_addr,
1137                 .maxlen         = sizeof(unsigned long),
1138                 .mode           = 0644,
1139                 .proc_handler   = &proc_doulongvec_minmax,
1140         },
1141 #endif
1142 #ifdef CONFIG_NUMA
1143         {
1144                 .ctl_name       = CTL_UNNUMBERED,
1145                 .procname       = "numa_zonelist_order",
1146                 .data           = &numa_zonelist_order,
1147                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1148                 .mode           = 0644,
1149                 .proc_handler   = &numa_zonelist_order_handler,
1150                 .strategy       = &sysctl_string,
1151         },
1152 #endif
1153 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1154    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1155         {
1156                 .ctl_name       = VM_VDSO_ENABLED,
1157                 .procname       = "vdso_enabled",
1158                 .data           = &vdso_enabled,
1159                 .maxlen         = sizeof(vdso_enabled),
1160                 .mode           = 0644,
1161                 .proc_handler   = &proc_dointvec,
1162                 .strategy       = &sysctl_intvec,
1163                 .extra1         = &zero,
1164         },
1165 #endif
1166 #ifdef CONFIG_HIGHMEM
1167         {
1168                 .ctl_name       = CTL_UNNUMBERED,
1169                 .procname       = "highmem_is_dirtyable",
1170                 .data           = &vm_highmem_is_dirtyable,
1171                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1172                 .mode           = 0644,
1173                 .proc_handler   = &proc_dointvec_minmax,
1174                 .strategy       = &sysctl_intvec,
1175                 .extra1         = &zero,
1176                 .extra2         = &one,
1177         },
1178 #endif
1179 /*
1180  * NOTE: do not add new entries to this table unless you have read
1181  * Documentation/sysctl/ctl_unnumbered.txt
1182  */
1183         { .ctl_name = 0 }
1184 };
1185
1186 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1187 static struct ctl_table binfmt_misc_table[] = {
1188         { .ctl_name = 0 }
1189 };
1190 #endif
1191
1192 static struct ctl_table fs_table[] = {
1193         {
1194                 .ctl_name       = FS_NRINODE,
1195                 .procname       = "inode-nr",
1196                 .data           = &inodes_stat,
1197                 .maxlen         = 2*sizeof(int),
1198                 .mode           = 0444,
1199                 .proc_handler   = &proc_dointvec,
1200         },
1201         {
1202                 .ctl_name       = FS_STATINODE,
1203                 .procname       = "inode-state",
1204                 .data           = &inodes_stat,
1205                 .maxlen         = 7*sizeof(int),
1206                 .mode           = 0444,
1207                 .proc_handler   = &proc_dointvec,
1208         },
1209         {
1210                 .procname       = "file-nr",
1211                 .data           = &files_stat,
1212                 .maxlen         = 3*sizeof(int),
1213                 .mode           = 0444,
1214                 .proc_handler   = &proc_nr_files,
1215         },
1216         {
1217                 .ctl_name       = FS_MAXFILE,
1218                 .procname       = "file-max",
1219                 .data           = &files_stat.max_files,
1220                 .maxlen         = sizeof(int),
1221                 .mode           = 0644,
1222                 .proc_handler   = &proc_dointvec,
1223         },
1224         {
1225                 .ctl_name       = CTL_UNNUMBERED,
1226                 .procname       = "nr_open",
1227                 .data           = &sysctl_nr_open,
1228                 .maxlen         = sizeof(int),
1229                 .mode           = 0644,
1230                 .proc_handler   = &proc_dointvec_minmax,
1231                 .extra1         = &sysctl_nr_open_min,
1232                 .extra2         = &sysctl_nr_open_max,
1233         },
1234         {
1235                 .ctl_name       = FS_DENTRY,
1236                 .procname       = "dentry-state",
1237                 .data           = &dentry_stat,
1238                 .maxlen         = 6*sizeof(int),
1239                 .mode           = 0444,
1240                 .proc_handler   = &proc_dointvec,
1241         },
1242         {
1243                 .ctl_name       = FS_OVERFLOWUID,
1244                 .procname       = "overflowuid",
1245                 .data           = &fs_overflowuid,
1246                 .maxlen         = sizeof(int),
1247                 .mode           = 0644,
1248                 .proc_handler   = &proc_dointvec_minmax,
1249                 .strategy       = &sysctl_intvec,
1250                 .extra1         = &minolduid,
1251                 .extra2         = &maxolduid,
1252         },
1253         {
1254                 .ctl_name       = FS_OVERFLOWGID,
1255                 .procname       = "overflowgid",
1256                 .data           = &fs_overflowgid,
1257                 .maxlen         = sizeof(int),
1258                 .mode           = 0644,
1259                 .proc_handler   = &proc_dointvec_minmax,
1260                 .strategy       = &sysctl_intvec,
1261                 .extra1         = &minolduid,
1262                 .extra2         = &maxolduid,
1263         },
1264         {
1265                 .ctl_name       = FS_LEASES,
1266                 .procname       = "leases-enable",
1267                 .data           = &leases_enable,
1268                 .maxlen         = sizeof(int),
1269                 .mode           = 0644,
1270                 .proc_handler   = &proc_dointvec,
1271         },
1272 #ifdef CONFIG_DNOTIFY
1273         {
1274                 .ctl_name       = FS_DIR_NOTIFY,
1275                 .procname       = "dir-notify-enable",
1276                 .data           = &dir_notify_enable,
1277                 .maxlen         = sizeof(int),
1278                 .mode           = 0644,
1279                 .proc_handler   = &proc_dointvec,
1280         },
1281 #endif
1282 #ifdef CONFIG_MMU
1283         {
1284                 .ctl_name       = FS_LEASE_TIME,
1285                 .procname       = "lease-break-time",
1286                 .data           = &lease_break_time,
1287                 .maxlen         = sizeof(int),
1288                 .mode           = 0644,
1289                 .proc_handler   = &proc_dointvec_minmax,
1290                 .strategy       = &sysctl_intvec,
1291                 .extra1         = &zero,
1292                 .extra2         = &two,
1293         },
1294         {
1295                 .procname       = "aio-nr",
1296                 .data           = &aio_nr,
1297                 .maxlen         = sizeof(aio_nr),
1298                 .mode           = 0444,
1299                 .proc_handler   = &proc_doulongvec_minmax,
1300         },
1301         {
1302                 .procname       = "aio-max-nr",
1303                 .data           = &aio_max_nr,
1304                 .maxlen         = sizeof(aio_max_nr),
1305                 .mode           = 0644,
1306                 .proc_handler   = &proc_doulongvec_minmax,
1307         },
1308 #ifdef CONFIG_INOTIFY_USER
1309         {
1310                 .ctl_name       = FS_INOTIFY,
1311                 .procname       = "inotify",
1312                 .mode           = 0555,
1313                 .child          = inotify_table,
1314         },
1315 #endif
1316 #endif
1317         {
1318                 .ctl_name       = KERN_SETUID_DUMPABLE,
1319                 .procname       = "suid_dumpable",
1320                 .data           = &suid_dumpable,
1321                 .maxlen         = sizeof(int),
1322                 .mode           = 0644,
1323                 .proc_handler   = &proc_dointvec,
1324         },
1325 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1326         {
1327                 .ctl_name       = CTL_UNNUMBERED,
1328                 .procname       = "binfmt_misc",
1329                 .mode           = 0555,
1330                 .child          = binfmt_misc_table,
1331         },
1332 #endif
1333 /*
1334  * NOTE: do not add new entries to this table unless you have read
1335  * Documentation/sysctl/ctl_unnumbered.txt
1336  */
1337         { .ctl_name = 0 }
1338 };
1339
1340 static struct ctl_table debug_table[] = {
1341 #if defined(CONFIG_X86) || defined(CONFIG_PPC)
1342         {
1343                 .ctl_name       = CTL_UNNUMBERED,
1344                 .procname       = "exception-trace",
1345                 .data           = &show_unhandled_signals,
1346                 .maxlen         = sizeof(int),
1347                 .mode           = 0644,
1348                 .proc_handler   = proc_dointvec
1349         },
1350 #endif
1351         { .ctl_name = 0 }
1352 };
1353
1354 static struct ctl_table dev_table[] = {
1355         { .ctl_name = 0 }
1356 };
1357
1358 static DEFINE_SPINLOCK(sysctl_lock);
1359
1360 /* called under sysctl_lock */
1361 static int use_table(struct ctl_table_header *p)
1362 {
1363         if (unlikely(p->unregistering))
1364                 return 0;
1365         p->used++;
1366         return 1;
1367 }
1368
1369 /* called under sysctl_lock */
1370 static void unuse_table(struct ctl_table_header *p)
1371 {
1372         if (!--p->used)
1373                 if (unlikely(p->unregistering))
1374                         complete(p->unregistering);
1375 }
1376
1377 /* called under sysctl_lock, will reacquire if has to wait */
1378 static void start_unregistering(struct ctl_table_header *p)
1379 {
1380         /*
1381          * if p->used is 0, nobody will ever touch that entry again;
1382          * we'll eliminate all paths to it before dropping sysctl_lock
1383          */
1384         if (unlikely(p->used)) {
1385                 struct completion wait;
1386                 init_completion(&wait);
1387                 p->unregistering = &wait;
1388                 spin_unlock(&sysctl_lock);
1389                 wait_for_completion(&wait);
1390                 spin_lock(&sysctl_lock);
1391         } else {
1392                 /* anything non-NULL; we'll never dereference it */
1393                 p->unregistering = ERR_PTR(-EINVAL);
1394         }
1395         /*
1396          * do not remove from the list until nobody holds it; walking the
1397          * list in do_sysctl() relies on that.
1398          */
1399         list_del_init(&p->ctl_entry);
1400 }
1401
1402 void sysctl_head_get(struct ctl_table_header *head)
1403 {
1404         spin_lock(&sysctl_lock);
1405         head->count++;
1406         spin_unlock(&sysctl_lock);
1407 }
1408
1409 void sysctl_head_put(struct ctl_table_header *head)
1410 {
1411         spin_lock(&sysctl_lock);
1412         if (!--head->count)
1413                 kfree(head);
1414         spin_unlock(&sysctl_lock);
1415 }
1416
1417 struct ctl_table_header *sysctl_head_grab(struct ctl_table_header *head)
1418 {
1419         if (!head)
1420                 BUG();
1421         spin_lock(&sysctl_lock);
1422         if (!use_table(head))
1423                 head = ERR_PTR(-ENOENT);
1424         spin_unlock(&sysctl_lock);
1425         return head;
1426 }
1427
1428 void sysctl_head_finish(struct ctl_table_header *head)
1429 {
1430         if (!head)
1431                 return;
1432         spin_lock(&sysctl_lock);
1433         unuse_table(head);
1434         spin_unlock(&sysctl_lock);
1435 }
1436
1437 static struct ctl_table_set *
1438 lookup_header_set(struct ctl_table_root *root, struct nsproxy *namespaces)
1439 {
1440         struct ctl_table_set *set = &root->default_set;
1441         if (root->lookup)
1442                 set = root->lookup(root, namespaces);
1443         return set;
1444 }
1445
1446 static struct list_head *
1447 lookup_header_list(struct ctl_table_root *root, struct nsproxy *namespaces)
1448 {
1449         struct ctl_table_set *set = lookup_header_set(root, namespaces);
1450         return &set->list;
1451 }
1452
1453 struct ctl_table_header *__sysctl_head_next(struct nsproxy *namespaces,
1454                                             struct ctl_table_header *prev)
1455 {
1456         struct ctl_table_root *root;
1457         struct list_head *header_list;
1458         struct ctl_table_header *head;
1459         struct list_head *tmp;
1460
1461         spin_lock(&sysctl_lock);
1462         if (prev) {
1463                 head = prev;
1464                 tmp = &prev->ctl_entry;
1465                 unuse_table(prev);
1466                 goto next;
1467         }
1468         tmp = &root_table_header.ctl_entry;
1469         for (;;) {
1470                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1471
1472                 if (!use_table(head))
1473                         goto next;
1474                 spin_unlock(&sysctl_lock);
1475                 return head;
1476         next:
1477                 root = head->root;
1478                 tmp = tmp->next;
1479                 header_list = lookup_header_list(root, namespaces);
1480                 if (tmp != header_list)
1481                         continue;
1482
1483                 do {
1484                         root = list_entry(root->root_list.next,
1485                                         struct ctl_table_root, root_list);
1486                         if (root == &sysctl_table_root)
1487                                 goto out;
1488                         header_list = lookup_header_list(root, namespaces);
1489                 } while (list_empty(header_list));
1490                 tmp = header_list->next;
1491         }
1492 out:
1493         spin_unlock(&sysctl_lock);
1494         return NULL;
1495 }
1496
1497 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1498 {
1499         return __sysctl_head_next(current->nsproxy, prev);
1500 }
1501
1502 void register_sysctl_root(struct ctl_table_root *root)
1503 {
1504         spin_lock(&sysctl_lock);
1505         list_add_tail(&root->root_list, &sysctl_table_root.root_list);
1506         spin_unlock(&sysctl_lock);
1507 }
1508
1509 #ifdef CONFIG_SYSCTL_SYSCALL
1510 /* Perform the actual read/write of a sysctl table entry. */
1511 static int do_sysctl_strategy(struct ctl_table_root *root,
1512                         struct ctl_table *table,
1513                         int __user *name, int nlen,
1514                         void __user *oldval, size_t __user *oldlenp,
1515                         void __user *newval, size_t newlen)
1516 {
1517         int op = 0, rc;
1518
1519         if (oldval)
1520                 op |= MAY_READ;
1521         if (newval)
1522                 op |= MAY_WRITE;
1523         if (sysctl_perm(root, table, op))
1524                 return -EPERM;
1525
1526         if (table->strategy) {
1527                 rc = table->strategy(table, name, nlen, oldval, oldlenp,
1528                                      newval, newlen);
1529                 if (rc < 0)
1530                         return rc;
1531                 if (rc > 0)
1532                         return 0;
1533         }
1534
1535         /* If there is no strategy routine, or if the strategy returns
1536          * zero, proceed with automatic r/w */
1537         if (table->data && table->maxlen) {
1538                 rc = sysctl_data(table, name, nlen, oldval, oldlenp,
1539                                  newval, newlen);
1540                 if (rc < 0)
1541                         return rc;
1542         }
1543         return 0;
1544 }
1545
1546 static int parse_table(int __user *name, int nlen,
1547                        void __user *oldval, size_t __user *oldlenp,
1548                        void __user *newval, size_t newlen,
1549                        struct ctl_table_root *root,
1550                        struct ctl_table *table)
1551 {
1552         int n;
1553 repeat:
1554         if (!nlen)
1555                 return -ENOTDIR;
1556         if (get_user(n, name))
1557                 return -EFAULT;
1558         for ( ; table->ctl_name || table->procname; table++) {
1559                 if (!table->ctl_name)
1560                         continue;
1561                 if (n == table->ctl_name) {
1562                         int error;
1563                         if (table->child) {
1564                                 if (sysctl_perm(root, table, MAY_EXEC))
1565                                         return -EPERM;
1566                                 name++;
1567                                 nlen--;
1568                                 table = table->child;
1569                                 goto repeat;
1570                         }
1571                         error = do_sysctl_strategy(root, table, name, nlen,
1572                                                    oldval, oldlenp,
1573                                                    newval, newlen);
1574                         return error;
1575                 }
1576         }
1577         return -ENOTDIR;
1578 }
1579
1580 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1581                void __user *newval, size_t newlen)
1582 {
1583         struct ctl_table_header *head;
1584         int error = -ENOTDIR;
1585
1586         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1587                 return -ENOTDIR;
1588         if (oldval) {
1589                 int old_len;
1590                 if (!oldlenp || get_user(old_len, oldlenp))
1591                         return -EFAULT;
1592         }
1593
1594         for (head = sysctl_head_next(NULL); head;
1595                         head = sysctl_head_next(head)) {
1596                 error = parse_table(name, nlen, oldval, oldlenp,
1597                                         newval, newlen,
1598                                         head->root, head->ctl_table);
1599                 if (error != -ENOTDIR) {
1600                         sysctl_head_finish(head);
1601                         break;
1602                 }
1603         }
1604         return error;
1605 }
1606
1607 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1608 {
1609         struct __sysctl_args tmp;
1610         int error;
1611
1612         if (copy_from_user(&tmp, args, sizeof(tmp)))
1613                 return -EFAULT;
1614
1615         error = deprecated_sysctl_warning(&tmp);
1616         if (error)
1617                 goto out;
1618
1619         lock_kernel();
1620         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1621                           tmp.newval, tmp.newlen);
1622         unlock_kernel();
1623 out:
1624         return error;
1625 }
1626 #endif /* CONFIG_SYSCTL_SYSCALL */
1627
1628 /*
1629  * sysctl_perm does NOT grant the superuser all rights automatically, because
1630  * some sysctl variables are readonly even to root.
1631  */
1632
1633 static int test_perm(int mode, int op)
1634 {
1635         if (!current->euid)
1636                 mode >>= 6;
1637         else if (in_egroup_p(0))
1638                 mode >>= 3;
1639         if ((op & ~mode & (MAY_READ|MAY_WRITE|MAY_EXEC)) == 0)
1640                 return 0;
1641         return -EACCES;
1642 }
1643
1644 int sysctl_perm(struct ctl_table_root *root, struct ctl_table *table, int op)
1645 {
1646         int error;
1647         int mode;
1648
1649         error = security_sysctl(table, op & (MAY_READ | MAY_WRITE | MAY_EXEC));
1650         if (error)
1651                 return error;
1652
1653         if (root->permissions)
1654                 mode = root->permissions(root, current->nsproxy, table);
1655         else
1656                 mode = table->mode;
1657
1658         return test_perm(mode, op);
1659 }
1660
1661 static void sysctl_set_parent(struct ctl_table *parent, struct ctl_table *table)
1662 {
1663         for (; table->ctl_name || table->procname; table++) {
1664                 table->parent = parent;
1665                 if (table->child)
1666                         sysctl_set_parent(table, table->child);
1667         }
1668 }
1669
1670 static __init int sysctl_init(void)
1671 {
1672         sysctl_set_parent(NULL, root_table);
1673 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1674         {
1675                 int err;
1676                 err = sysctl_check_table(current->nsproxy, root_table);
1677         }
1678 #endif
1679         return 0;
1680 }
1681
1682 core_initcall(sysctl_init);
1683
1684 static struct ctl_table *is_branch_in(struct ctl_table *branch,
1685                                       struct ctl_table *table)
1686 {
1687         struct ctl_table *p;
1688         const char *s = branch->procname;
1689
1690         /* branch should have named subdirectory as its first element */
1691         if (!s || !branch->child)
1692                 return NULL;
1693
1694         /* ... and nothing else */
1695         if (branch[1].procname || branch[1].ctl_name)
1696                 return NULL;
1697
1698         /* table should contain subdirectory with the same name */
1699         for (p = table; p->procname || p->ctl_name; p++) {
1700                 if (!p->child)
1701                         continue;
1702                 if (p->procname && strcmp(p->procname, s) == 0)
1703                         return p;
1704         }
1705         return NULL;
1706 }
1707
1708 /* see if attaching q to p would be an improvement */
1709 static void try_attach(struct ctl_table_header *p, struct ctl_table_header *q)
1710 {
1711         struct ctl_table *to = p->ctl_table, *by = q->ctl_table;
1712         struct ctl_table *next;
1713         int is_better = 0;
1714         int not_in_parent = !p->attached_by;
1715
1716         while ((next = is_branch_in(by, to)) != NULL) {
1717                 if (by == q->attached_by)
1718                         is_better = 1;
1719                 if (to == p->attached_by)
1720                         not_in_parent = 1;
1721                 by = by->child;
1722                 to = next->child;
1723         }
1724
1725         if (is_better && not_in_parent) {
1726                 q->attached_by = by;
1727                 q->attached_to = to;
1728                 q->parent = p;
1729         }
1730 }
1731
1732 /**
1733  * __register_sysctl_paths - register a sysctl hierarchy
1734  * @root: List of sysctl headers to register on
1735  * @namespaces: Data to compute which lists of sysctl entries are visible
1736  * @path: The path to the directory the sysctl table is in.
1737  * @table: the top-level table structure
1738  *
1739  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1740  * array. A completely 0 filled entry terminates the table.
1741  *
1742  * The members of the &struct ctl_table structure are used as follows:
1743  *
1744  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1745  *            must be unique within that level of sysctl
1746  *
1747  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1748  *            enter a sysctl file
1749  *
1750  * data - a pointer to data for use by proc_handler
1751  *
1752  * maxlen - the maximum size in bytes of the data
1753  *
1754  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1755  *
1756  * child - a pointer to the child sysctl table if this entry is a directory, or
1757  *         %NULL.
1758  *
1759  * proc_handler - the text handler routine (described below)
1760  *
1761  * strategy - the strategy routine (described below)
1762  *
1763  * de - for internal use by the sysctl routines
1764  *
1765  * extra1, extra2 - extra pointers usable by the proc handler routines
1766  *
1767  * Leaf nodes in the sysctl tree will be represented by a single file
1768  * under /proc; non-leaf nodes will be represented by directories.
1769  *
1770  * sysctl(2) can automatically manage read and write requests through
1771  * the sysctl table.  The data and maxlen fields of the ctl_table
1772  * struct enable minimal validation of the values being written to be
1773  * performed, and the mode field allows minimal authentication.
1774  *
1775  * More sophisticated management can be enabled by the provision of a
1776  * strategy routine with the table entry.  This will be called before
1777  * any automatic read or write of the data is performed.
1778  *
1779  * The strategy routine may return
1780  *
1781  * < 0 - Error occurred (error is passed to user process)
1782  *
1783  * 0   - OK - proceed with automatic read or write.
1784  *
1785  * > 0 - OK - read or write has been done by the strategy routine, so
1786  *       return immediately.
1787  *
1788  * There must be a proc_handler routine for any terminal nodes
1789  * mirrored under /proc/sys (non-terminals are handled by a built-in
1790  * directory handler).  Several default handlers are available to
1791  * cover common cases -
1792  *
1793  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1794  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1795  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1796  *
1797  * It is the handler's job to read the input buffer from user memory
1798  * and process it. The handler should return 0 on success.
1799  *
1800  * This routine returns %NULL on a failure to register, and a pointer
1801  * to the table header on success.
1802  */
1803 struct ctl_table_header *__register_sysctl_paths(
1804         struct ctl_table_root *root,
1805         struct nsproxy *namespaces,
1806         const struct ctl_path *path, struct ctl_table *table)
1807 {
1808         struct ctl_table_header *header;
1809         struct ctl_table *new, **prevp;
1810         unsigned int n, npath;
1811         struct ctl_table_set *set;
1812
1813         /* Count the path components */
1814         for (npath = 0; path[npath].ctl_name || path[npath].procname; ++npath)
1815                 ;
1816
1817         /*
1818          * For each path component, allocate a 2-element ctl_table array.
1819          * The first array element will be filled with the sysctl entry
1820          * for this, the second will be the sentinel (ctl_name == 0).
1821          *
1822          * We allocate everything in one go so that we don't have to
1823          * worry about freeing additional memory in unregister_sysctl_table.
1824          */
1825         header = kzalloc(sizeof(struct ctl_table_header) +
1826                          (2 * npath * sizeof(struct ctl_table)), GFP_KERNEL);
1827         if (!header)
1828                 return NULL;
1829
1830         new = (struct ctl_table *) (header + 1);
1831
1832         /* Now connect the dots */
1833         prevp = &header->ctl_table;
1834         for (n = 0; n < npath; ++n, ++path) {
1835                 /* Copy the procname */
1836                 new->procname = path->procname;
1837                 new->ctl_name = path->ctl_name;
1838                 new->mode     = 0555;
1839
1840                 *prevp = new;
1841                 prevp = &new->child;
1842
1843                 new += 2;
1844         }
1845         *prevp = table;
1846         header->ctl_table_arg = table;
1847
1848         INIT_LIST_HEAD(&header->ctl_entry);
1849         header->used = 0;
1850         header->unregistering = NULL;
1851         header->root = root;
1852         sysctl_set_parent(NULL, header->ctl_table);
1853         header->count = 1;
1854 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1855         if (sysctl_check_table(namespaces, header->ctl_table)) {
1856                 kfree(header);
1857                 return NULL;
1858         }
1859 #endif
1860         spin_lock(&sysctl_lock);
1861         header->set = lookup_header_set(root, namespaces);
1862         header->attached_by = header->ctl_table;
1863         header->attached_to = root_table;
1864         header->parent = &root_table_header;
1865         for (set = header->set; set; set = set->parent) {
1866                 struct ctl_table_header *p;
1867                 list_for_each_entry(p, &set->list, ctl_entry) {
1868                         if (p->unregistering)
1869                                 continue;
1870                         try_attach(p, header);
1871                 }
1872         }
1873         header->parent->count++;
1874         list_add_tail(&header->ctl_entry, &header->set->list);
1875         spin_unlock(&sysctl_lock);
1876
1877         return header;
1878 }
1879
1880 /**
1881  * register_sysctl_table_path - register a sysctl table hierarchy
1882  * @path: The path to the directory the sysctl table is in.
1883  * @table: the top-level table structure
1884  *
1885  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1886  * array. A completely 0 filled entry terminates the table.
1887  *
1888  * See __register_sysctl_paths for more details.
1889  */
1890 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
1891                                                 struct ctl_table *table)
1892 {
1893         return __register_sysctl_paths(&sysctl_table_root, current->nsproxy,
1894                                         path, table);
1895 }
1896
1897 /**
1898  * register_sysctl_table - register a sysctl table hierarchy
1899  * @table: the top-level table structure
1900  *
1901  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1902  * array. A completely 0 filled entry terminates the table.
1903  *
1904  * See register_sysctl_paths for more details.
1905  */
1906 struct ctl_table_header *register_sysctl_table(struct ctl_table *table)
1907 {
1908         static const struct ctl_path null_path[] = { {} };
1909
1910         return register_sysctl_paths(null_path, table);
1911 }
1912
1913 /**
1914  * unregister_sysctl_table - unregister a sysctl table hierarchy
1915  * @header: the header returned from register_sysctl_table
1916  *
1917  * Unregisters the sysctl table and all children. proc entries may not
1918  * actually be removed until they are no longer used by anyone.
1919  */
1920 void unregister_sysctl_table(struct ctl_table_header * header)
1921 {
1922         might_sleep();
1923
1924         if (header == NULL)
1925                 return;
1926
1927         spin_lock(&sysctl_lock);
1928         start_unregistering(header);
1929         if (!--header->parent->count) {
1930                 WARN_ON(1);
1931                 kfree(header->parent);
1932         }
1933         if (!--header->count)
1934                 kfree(header);
1935         spin_unlock(&sysctl_lock);
1936 }
1937
1938 int sysctl_is_seen(struct ctl_table_header *p)
1939 {
1940         struct ctl_table_set *set = p->set;
1941         int res;
1942         spin_lock(&sysctl_lock);
1943         if (p->unregistering)
1944                 res = 0;
1945         else if (!set->is_seen)
1946                 res = 1;
1947         else
1948                 res = set->is_seen(set);
1949         spin_unlock(&sysctl_lock);
1950         return res;
1951 }
1952
1953 void setup_sysctl_set(struct ctl_table_set *p,
1954         struct ctl_table_set *parent,
1955         int (*is_seen)(struct ctl_table_set *))
1956 {
1957         INIT_LIST_HEAD(&p->list);
1958         p->parent = parent ? parent : &sysctl_table_root.default_set;
1959         p->is_seen = is_seen;
1960 }
1961
1962 #else /* !CONFIG_SYSCTL */
1963 struct ctl_table_header *register_sysctl_table(struct ctl_table * table)
1964 {
1965         return NULL;
1966 }
1967
1968 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
1969                                                     struct ctl_table *table)
1970 {
1971         return NULL;
1972 }
1973
1974 void unregister_sysctl_table(struct ctl_table_header * table)
1975 {
1976 }
1977
1978 void setup_sysctl_set(struct ctl_table_set *p,
1979         struct ctl_table_set *parent,
1980         int (*is_seen)(struct ctl_table_set *))
1981 {
1982 }
1983
1984 void sysctl_head_put(struct ctl_table_header *head)
1985 {
1986 }
1987
1988 #endif /* CONFIG_SYSCTL */
1989
1990 /*
1991  * /proc/sys support
1992  */
1993
1994 #ifdef CONFIG_PROC_SYSCTL
1995
1996 static int _proc_do_string(void* data, int maxlen, int write,
1997                            struct file *filp, void __user *buffer,
1998                            size_t *lenp, loff_t *ppos)
1999 {
2000         size_t len;
2001         char __user *p;
2002         char c;
2003
2004         if (!data || !maxlen || !*lenp) {
2005                 *lenp = 0;
2006                 return 0;
2007         }
2008
2009         if (write) {
2010                 len = 0;
2011                 p = buffer;
2012                 while (len < *lenp) {
2013                         if (get_user(c, p++))
2014                                 return -EFAULT;
2015                         if (c == 0 || c == '\n')
2016                                 break;
2017                         len++;
2018                 }
2019                 if (len >= maxlen)
2020                         len = maxlen-1;
2021                 if(copy_from_user(data, buffer, len))
2022                         return -EFAULT;
2023                 ((char *) data)[len] = 0;
2024                 *ppos += *lenp;
2025         } else {
2026                 len = strlen(data);
2027                 if (len > maxlen)
2028                         len = maxlen;
2029
2030                 if (*ppos > len) {
2031                         *lenp = 0;
2032                         return 0;
2033                 }
2034
2035                 data += *ppos;
2036                 len  -= *ppos;
2037
2038                 if (len > *lenp)
2039                         len = *lenp;
2040                 if (len)
2041                         if(copy_to_user(buffer, data, len))
2042                                 return -EFAULT;
2043                 if (len < *lenp) {
2044                         if(put_user('\n', ((char __user *) buffer) + len))
2045                                 return -EFAULT;
2046                         len++;
2047                 }
2048                 *lenp = len;
2049                 *ppos += len;
2050         }
2051         return 0;
2052 }
2053
2054 /**
2055  * proc_dostring - read a string sysctl
2056  * @table: the sysctl table
2057  * @write: %TRUE if this is a write to the sysctl file
2058  * @filp: the file structure
2059  * @buffer: the user buffer
2060  * @lenp: the size of the user buffer
2061  * @ppos: file position
2062  *
2063  * Reads/writes a string from/to the user buffer. If the kernel
2064  * buffer provided is not large enough to hold the string, the
2065  * string is truncated. The copied string is %NULL-terminated.
2066  * If the string is being read by the user process, it is copied
2067  * and a newline '\n' is added. It is truncated if the buffer is
2068  * not large enough.
2069  *
2070  * Returns 0 on success.
2071  */
2072 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2073                   void __user *buffer, size_t *lenp, loff_t *ppos)
2074 {
2075         return _proc_do_string(table->data, table->maxlen, write, filp,
2076                                buffer, lenp, ppos);
2077 }
2078
2079
2080 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
2081                                  int *valp,
2082                                  int write, void *data)
2083 {
2084         if (write) {
2085                 *valp = *negp ? -*lvalp : *lvalp;
2086         } else {
2087                 int val = *valp;
2088                 if (val < 0) {
2089                         *negp = -1;
2090                         *lvalp = (unsigned long)-val;
2091                 } else {
2092                         *negp = 0;
2093                         *lvalp = (unsigned long)val;
2094                 }
2095         }
2096         return 0;
2097 }
2098
2099 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2100                   int write, struct file *filp, void __user *buffer,
2101                   size_t *lenp, loff_t *ppos,
2102                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2103                               int write, void *data),
2104                   void *data)
2105 {
2106 #define TMPBUFLEN 21
2107         int *i, vleft, first=1, neg, val;
2108         unsigned long lval;
2109         size_t left, len;
2110
2111         char buf[TMPBUFLEN], *p;
2112         char __user *s = buffer;
2113
2114         if (!tbl_data || !table->maxlen || !*lenp ||
2115             (*ppos && !write)) {
2116                 *lenp = 0;
2117                 return 0;
2118         }
2119
2120         i = (int *) tbl_data;
2121         vleft = table->maxlen / sizeof(*i);
2122         left = *lenp;
2123
2124         if (!conv)
2125                 conv = do_proc_dointvec_conv;
2126
2127         for (; left && vleft--; i++, first=0) {
2128                 if (write) {
2129                         while (left) {
2130                                 char c;
2131                                 if (get_user(c, s))
2132                                         return -EFAULT;
2133                                 if (!isspace(c))
2134                                         break;
2135                                 left--;
2136                                 s++;
2137                         }
2138                         if (!left)
2139                                 break;
2140                         neg = 0;
2141                         len = left;
2142                         if (len > sizeof(buf) - 1)
2143                                 len = sizeof(buf) - 1;
2144                         if (copy_from_user(buf, s, len))
2145                                 return -EFAULT;
2146                         buf[len] = 0;
2147                         p = buf;
2148                         if (*p == '-' && left > 1) {
2149                                 neg = 1;
2150                                 p++;
2151                         }
2152                         if (*p < '0' || *p > '9')
2153                                 break;
2154
2155                         lval = simple_strtoul(p, &p, 0);
2156
2157                         len = p-buf;
2158                         if ((len < left) && *p && !isspace(*p))
2159                                 break;
2160                         if (neg)
2161                                 val = -val;
2162                         s += len;
2163                         left -= len;
2164
2165                         if (conv(&neg, &lval, i, 1, data))
2166                                 break;
2167                 } else {
2168                         p = buf;
2169                         if (!first)
2170                                 *p++ = '\t';
2171
2172                         if (conv(&neg, &lval, i, 0, data))
2173                                 break;
2174
2175                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
2176                         len = strlen(buf);
2177                         if (len > left)
2178                                 len = left;
2179                         if(copy_to_user(s, buf, len))
2180                                 return -EFAULT;
2181                         left -= len;
2182                         s += len;
2183                 }
2184         }
2185
2186         if (!write && !first && left) {
2187                 if(put_user('\n', s))
2188                         return -EFAULT;
2189                 left--, s++;
2190         }
2191         if (write) {
2192                 while (left) {
2193                         char c;
2194                         if (get_user(c, s++))
2195                                 return -EFAULT;
2196                         if (!isspace(c))
2197                                 break;
2198                         left--;
2199                 }
2200         }
2201         if (write && first)
2202                 return -EINVAL;
2203         *lenp -= left;
2204         *ppos += *lenp;
2205         return 0;
2206 #undef TMPBUFLEN
2207 }
2208
2209 static int do_proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2210                   void __user *buffer, size_t *lenp, loff_t *ppos,
2211                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2212                               int write, void *data),
2213                   void *data)
2214 {
2215         return __do_proc_dointvec(table->data, table, write, filp,
2216                         buffer, lenp, ppos, conv, data);
2217 }
2218
2219 /**
2220  * proc_dointvec - read a vector of integers
2221  * @table: the sysctl table
2222  * @write: %TRUE if this is a write to the sysctl file
2223  * @filp: the file structure
2224  * @buffer: the user buffer
2225  * @lenp: the size of the user buffer
2226  * @ppos: file position
2227  *
2228  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2229  * values from/to the user buffer, treated as an ASCII string.
2230  *
2231  * Returns 0 on success.
2232  */
2233 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2234                      void __user *buffer, size_t *lenp, loff_t *ppos)
2235 {
2236     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2237                             NULL,NULL);
2238 }
2239
2240 #define OP_SET  0
2241 #define OP_AND  1
2242 #define OP_OR   2
2243
2244 static int do_proc_dointvec_bset_conv(int *negp, unsigned long *lvalp,
2245                                       int *valp,
2246                                       int write, void *data)
2247 {
2248         int op = *(int *)data;
2249         if (write) {
2250                 int val = *negp ? -*lvalp : *lvalp;
2251                 switch(op) {
2252                 case OP_SET:    *valp = val; break;
2253                 case OP_AND:    *valp &= val; break;
2254                 case OP_OR:     *valp |= val; break;
2255                 }
2256         } else {
2257                 int val = *valp;
2258                 if (val < 0) {
2259                         *negp = -1;
2260                         *lvalp = (unsigned long)-val;
2261                 } else {
2262                         *negp = 0;
2263                         *lvalp = (unsigned long)val;
2264                 }
2265         }
2266         return 0;
2267 }
2268
2269 /*
2270  *      Taint values can only be increased
2271  */
2272 static int proc_dointvec_taint(struct ctl_table *table, int write, struct file *filp,
2273                                void __user *buffer, size_t *lenp, loff_t *ppos)
2274 {
2275         int op;
2276
2277         if (write && !capable(CAP_SYS_ADMIN))
2278                 return -EPERM;
2279
2280         op = OP_OR;
2281         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2282                                 do_proc_dointvec_bset_conv,&op);
2283 }
2284
2285 struct do_proc_dointvec_minmax_conv_param {
2286         int *min;
2287         int *max;
2288 };
2289
2290 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
2291                                         int *valp,
2292                                         int write, void *data)
2293 {
2294         struct do_proc_dointvec_minmax_conv_param *param = data;
2295         if (write) {
2296                 int val = *negp ? -*lvalp : *lvalp;
2297                 if ((param->min && *param->min > val) ||
2298                     (param->max && *param->max < val))
2299                         return -EINVAL;
2300                 *valp = val;
2301         } else {
2302                 int val = *valp;
2303                 if (val < 0) {
2304                         *negp = -1;
2305                         *lvalp = (unsigned long)-val;
2306                 } else {
2307                         *negp = 0;
2308                         *lvalp = (unsigned long)val;
2309                 }
2310         }
2311         return 0;
2312 }
2313
2314 /**
2315  * proc_dointvec_minmax - read a vector of integers with min/max values
2316  * @table: the sysctl table
2317  * @write: %TRUE if this is a write to the sysctl file
2318  * @filp: the file structure
2319  * @buffer: the user buffer
2320  * @lenp: the size of the user buffer
2321  * @ppos: file position
2322  *
2323  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2324  * values from/to the user buffer, treated as an ASCII string.
2325  *
2326  * This routine will ensure the values are within the range specified by
2327  * table->extra1 (min) and table->extra2 (max).
2328  *
2329  * Returns 0 on success.
2330  */
2331 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2332                   void __user *buffer, size_t *lenp, loff_t *ppos)
2333 {
2334         struct do_proc_dointvec_minmax_conv_param param = {
2335                 .min = (int *) table->extra1,
2336                 .max = (int *) table->extra2,
2337         };
2338         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2339                                 do_proc_dointvec_minmax_conv, &param);
2340 }
2341
2342 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2343                                      struct file *filp,
2344                                      void __user *buffer,
2345                                      size_t *lenp, loff_t *ppos,
2346                                      unsigned long convmul,
2347                                      unsigned long convdiv)
2348 {
2349 #define TMPBUFLEN 21
2350         unsigned long *i, *min, *max, val;
2351         int vleft, first=1, neg;
2352         size_t len, left;
2353         char buf[TMPBUFLEN], *p;
2354         char __user *s = buffer;
2355
2356         if (!data || !table->maxlen || !*lenp ||
2357             (*ppos && !write)) {
2358                 *lenp = 0;
2359                 return 0;
2360         }
2361
2362         i = (unsigned long *) data;
2363         min = (unsigned long *) table->extra1;
2364         max = (unsigned long *) table->extra2;
2365         vleft = table->maxlen / sizeof(unsigned long);
2366         left = *lenp;
2367
2368         for (; left && vleft--; i++, min++, max++, first=0) {
2369                 if (write) {
2370                         while (left) {
2371                                 char c;
2372                                 if (get_user(c, s))
2373                                         return -EFAULT;
2374                                 if (!isspace(c))
2375                                         break;
2376                                 left--;
2377                                 s++;
2378                         }
2379                         if (!left)
2380                                 break;
2381                         neg = 0;
2382                         len = left;
2383                         if (len > TMPBUFLEN-1)
2384                                 len = TMPBUFLEN-1;
2385                         if (copy_from_user(buf, s, len))
2386                                 return -EFAULT;
2387                         buf[len] = 0;
2388                         p = buf;
2389                         if (*p == '-' && left > 1) {
2390                                 neg = 1;
2391                                 p++;
2392                         }
2393                         if (*p < '0' || *p > '9')
2394                                 break;
2395                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
2396                         len = p-buf;
2397                         if ((len < left) && *p && !isspace(*p))
2398                                 break;
2399                         if (neg)
2400                                 val = -val;
2401                         s += len;
2402                         left -= len;
2403
2404                         if(neg)
2405                                 continue;
2406                         if ((min && val < *min) || (max && val > *max))
2407                                 continue;
2408                         *i = val;
2409                 } else {
2410                         p = buf;
2411                         if (!first)
2412                                 *p++ = '\t';
2413                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2414                         len = strlen(buf);
2415                         if (len > left)
2416                                 len = left;
2417                         if(copy_to_user(s, buf, len))
2418                                 return -EFAULT;
2419                         left -= len;
2420                         s += len;
2421                 }
2422         }
2423
2424         if (!write && !first && left) {
2425                 if(put_user('\n', s))
2426                         return -EFAULT;
2427                 left--, s++;
2428         }
2429         if (write) {
2430                 while (left) {
2431                         char c;
2432                         if (get_user(c, s++))
2433                                 return -EFAULT;
2434                         if (!isspace(c))
2435                                 break;
2436                         left--;
2437                 }
2438         }
2439         if (write && first)
2440                 return -EINVAL;
2441         *lenp -= left;
2442         *ppos += *lenp;
2443         return 0;
2444 #undef TMPBUFLEN
2445 }
2446
2447 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2448                                      struct file *filp,
2449                                      void __user *buffer,
2450                                      size_t *lenp, loff_t *ppos,
2451                                      unsigned long convmul,
2452                                      unsigned long convdiv)
2453 {
2454         return __do_proc_doulongvec_minmax(table->data, table, write,
2455                         filp, buffer, lenp, ppos, convmul, convdiv);
2456 }
2457
2458 /**
2459  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2460  * @table: the sysctl table
2461  * @write: %TRUE if this is a write to the sysctl file
2462  * @filp: the file structure
2463  * @buffer: the user buffer
2464  * @lenp: the size of the user buffer
2465  * @ppos: file position
2466  *
2467  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2468  * values from/to the user buffer, treated as an ASCII string.
2469  *
2470  * This routine will ensure the values are within the range specified by
2471  * table->extra1 (min) and table->extra2 (max).
2472  *
2473  * Returns 0 on success.
2474  */
2475 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2476                            void __user *buffer, size_t *lenp, loff_t *ppos)
2477 {
2478     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2479 }
2480
2481 /**
2482  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2483  * @table: the sysctl table
2484  * @write: %TRUE if this is a write to the sysctl file
2485  * @filp: the file structure
2486  * @buffer: the user buffer
2487  * @lenp: the size of the user buffer
2488  * @ppos: file position
2489  *
2490  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2491  * values from/to the user buffer, treated as an ASCII string. The values
2492  * are treated as milliseconds, and converted to jiffies when they are stored.
2493  *
2494  * This routine will ensure the values are within the range specified by
2495  * table->extra1 (min) and table->extra2 (max).
2496  *
2497  * Returns 0 on success.
2498  */
2499 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2500                                       struct file *filp,
2501                                       void __user *buffer,
2502                                       size_t *lenp, loff_t *ppos)
2503 {
2504     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2505                                      lenp, ppos, HZ, 1000l);
2506 }
2507
2508
2509 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2510                                          int *valp,
2511                                          int write, void *data)
2512 {
2513         if (write) {
2514                 if (*lvalp > LONG_MAX / HZ)
2515                         return 1;
2516                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2517         } else {
2518                 int val = *valp;
2519                 unsigned long lval;
2520                 if (val < 0) {
2521                         *negp = -1;
2522                         lval = (unsigned long)-val;
2523                 } else {
2524                         *negp = 0;
2525                         lval = (unsigned long)val;
2526                 }
2527                 *lvalp = lval / HZ;
2528         }
2529         return 0;
2530 }
2531
2532 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2533                                                 int *valp,
2534                                                 int write, void *data)
2535 {
2536         if (write) {
2537                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2538                         return 1;
2539                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2540         } else {
2541                 int val = *valp;
2542                 unsigned long lval;
2543                 if (val < 0) {
2544                         *negp = -1;
2545                         lval = (unsigned long)-val;
2546                 } else {
2547                         *negp = 0;
2548                         lval = (unsigned long)val;
2549                 }
2550                 *lvalp = jiffies_to_clock_t(lval);
2551         }
2552         return 0;
2553 }
2554
2555 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2556                                             int *valp,
2557                                             int write, void *data)
2558 {
2559         if (write) {
2560                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2561         } else {
2562                 int val = *valp;
2563                 unsigned long lval;
2564                 if (val < 0) {
2565                         *negp = -1;
2566                         lval = (unsigned long)-val;
2567                 } else {
2568                         *negp = 0;
2569                         lval = (unsigned long)val;
2570                 }
2571                 *lvalp = jiffies_to_msecs(lval);
2572         }
2573         return 0;
2574 }
2575
2576 /**
2577  * proc_dointvec_jiffies - read a vector of integers as seconds
2578  * @table: the sysctl table
2579  * @write: %TRUE if this is a write to the sysctl file
2580  * @filp: the file structure
2581  * @buffer: the user buffer
2582  * @lenp: the size of the user buffer
2583  * @ppos: file position
2584  *
2585  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2586  * values from/to the user buffer, treated as an ASCII string.
2587  * The values read are assumed to be in seconds, and are converted into
2588  * jiffies.
2589  *
2590  * Returns 0 on success.
2591  */
2592 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2593                           void __user *buffer, size_t *lenp, loff_t *ppos)
2594 {
2595     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2596                             do_proc_dointvec_jiffies_conv,NULL);
2597 }
2598
2599 /**
2600  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2601  * @table: the sysctl table
2602  * @write: %TRUE if this is a write to the sysctl file
2603  * @filp: the file structure
2604  * @buffer: the user buffer
2605  * @lenp: the size of the user buffer
2606  * @ppos: pointer to the file position
2607  *
2608  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2609  * values from/to the user buffer, treated as an ASCII string.
2610  * The values read are assumed to be in 1/USER_HZ seconds, and
2611  * are converted into jiffies.
2612  *
2613  * Returns 0 on success.
2614  */
2615 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2616                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2617 {
2618     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2619                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2620 }
2621
2622 /**
2623  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2624  * @table: the sysctl table
2625  * @write: %TRUE if this is a write to the sysctl file
2626  * @filp: the file structure
2627  * @buffer: the user buffer
2628  * @lenp: the size of the user buffer
2629  * @ppos: file position
2630  * @ppos: the current position in the file
2631  *
2632  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2633  * values from/to the user buffer, treated as an ASCII string.
2634  * The values read are assumed to be in 1/1000 seconds, and
2635  * are converted into jiffies.
2636  *
2637  * Returns 0 on success.
2638  */
2639 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2640                              void __user *buffer, size_t *lenp, loff_t *ppos)
2641 {
2642         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2643                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2644 }
2645
2646 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
2647                            void __user *buffer, size_t *lenp, loff_t *ppos)
2648 {
2649         struct pid *new_pid;
2650         pid_t tmp;
2651         int r;
2652
2653         tmp = pid_vnr(cad_pid);
2654
2655         r = __do_proc_dointvec(&tmp, table, write, filp, buffer,
2656                                lenp, ppos, NULL, NULL);
2657         if (r || !write)
2658                 return r;
2659
2660         new_pid = find_get_pid(tmp);
2661         if (!new_pid)
2662                 return -ESRCH;
2663
2664         put_pid(xchg(&cad_pid, new_pid));
2665         return 0;
2666 }
2667
2668 #else /* CONFIG_PROC_FS */
2669
2670 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2671                   void __user *buffer, size_t *lenp, loff_t *ppos)
2672 {
2673         return -ENOSYS;
2674 }
2675
2676 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2677                   void __user *buffer, size_t *lenp, loff_t *ppos)
2678 {
2679         return -ENOSYS;
2680 }
2681
2682 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2683                     void __user *buffer, size_t *lenp, loff_t *ppos)
2684 {
2685         return -ENOSYS;
2686 }
2687
2688 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2689                     void __user *buffer, size_t *lenp, loff_t *ppos)
2690 {
2691         return -ENOSYS;
2692 }
2693
2694 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2695                     void __user *buffer, size_t *lenp, loff_t *ppos)
2696 {
2697         return -ENOSYS;
2698 }
2699
2700 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2701                              void __user *buffer, size_t *lenp, loff_t *ppos)
2702 {
2703         return -ENOSYS;
2704 }
2705
2706 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2707                     void __user *buffer, size_t *lenp, loff_t *ppos)
2708 {
2709         return -ENOSYS;
2710 }
2711
2712 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2713                                       struct file *filp,
2714                                       void __user *buffer,
2715                                       size_t *lenp, loff_t *ppos)
2716 {
2717     return -ENOSYS;
2718 }
2719
2720
2721 #endif /* CONFIG_PROC_FS */
2722
2723
2724 #ifdef CONFIG_SYSCTL_SYSCALL
2725 /*
2726  * General sysctl support routines
2727  */
2728
2729 /* The generic sysctl data routine (used if no strategy routine supplied) */
2730 int sysctl_data(struct ctl_table *table, int __user *name, int nlen,
2731                 void __user *oldval, size_t __user *oldlenp,
2732                 void __user *newval, size_t newlen)
2733 {
2734         size_t len;
2735
2736         /* Get out of I don't have a variable */
2737         if (!table->data || !table->maxlen)
2738                 return -ENOTDIR;
2739
2740         if (oldval && oldlenp) {
2741                 if (get_user(len, oldlenp))
2742                         return -EFAULT;
2743                 if (len) {
2744                         if (len > table->maxlen)
2745                                 len = table->maxlen;
2746                         if (copy_to_user(oldval, table->data, len))
2747                                 return -EFAULT;
2748                         if (put_user(len, oldlenp))
2749                                 return -EFAULT;
2750                 }
2751         }
2752
2753         if (newval && newlen) {
2754                 if (newlen > table->maxlen)
2755                         newlen = table->maxlen;
2756
2757                 if (copy_from_user(table->data, newval, newlen))
2758                         return -EFAULT;
2759         }
2760         return 1;
2761 }
2762
2763 /* The generic string strategy routine: */
2764 int sysctl_string(struct ctl_table *table, int __user *name, int nlen,
2765                   void __user *oldval, size_t __user *oldlenp,
2766                   void __user *newval, size_t newlen)
2767 {
2768         if (!table->data || !table->maxlen)
2769                 return -ENOTDIR;
2770
2771         if (oldval && oldlenp) {
2772                 size_t bufsize;
2773                 if (get_user(bufsize, oldlenp))
2774                         return -EFAULT;
2775                 if (bufsize) {
2776                         size_t len = strlen(table->data), copied;
2777
2778                         /* This shouldn't trigger for a well-formed sysctl */
2779                         if (len > table->maxlen)
2780                                 len = table->maxlen;
2781
2782                         /* Copy up to a max of bufsize-1 bytes of the string */
2783                         copied = (len >= bufsize) ? bufsize - 1 : len;
2784
2785                         if (copy_to_user(oldval, table->data, copied) ||
2786                             put_user(0, (char __user *)(oldval + copied)))
2787                                 return -EFAULT;
2788                         if (put_user(len, oldlenp))
2789                                 return -EFAULT;
2790                 }
2791         }
2792         if (newval && newlen) {
2793                 size_t len = newlen;
2794                 if (len > table->maxlen)
2795                         len = table->maxlen;
2796                 if(copy_from_user(table->data, newval, len))
2797                         return -EFAULT;
2798                 if (len == table->maxlen)
2799                         len--;
2800                 ((char *) table->data)[len] = 0;
2801         }
2802         return 1;
2803 }
2804
2805 /*
2806  * This function makes sure that all of the integers in the vector
2807  * are between the minimum and maximum values given in the arrays
2808  * table->extra1 and table->extra2, respectively.
2809  */
2810 int sysctl_intvec(struct ctl_table *table, int __user *name, int nlen,
2811                 void __user *oldval, size_t __user *oldlenp,
2812                 void __user *newval, size_t newlen)
2813 {
2814
2815         if (newval && newlen) {
2816                 int __user *vec = (int __user *) newval;
2817                 int *min = (int *) table->extra1;
2818                 int *max = (int *) table->extra2;
2819                 size_t length;
2820                 int i;
2821
2822                 if (newlen % sizeof(int) != 0)
2823                         return -EINVAL;
2824
2825                 if (!table->extra1 && !table->extra2)
2826                         return 0;
2827
2828                 if (newlen > table->maxlen)
2829                         newlen = table->maxlen;
2830                 length = newlen / sizeof(int);
2831
2832                 for (i = 0; i < length; i++) {
2833                         int value;
2834                         if (get_user(value, vec + i))
2835                                 return -EFAULT;
2836                         if (min && value < min[i])
2837                                 return -EINVAL;
2838                         if (max && value > max[i])
2839                                 return -EINVAL;
2840                 }
2841         }
2842         return 0;
2843 }
2844
2845 /* Strategy function to convert jiffies to seconds */
2846 int sysctl_jiffies(struct ctl_table *table, int __user *name, int nlen,
2847                 void __user *oldval, size_t __user *oldlenp,
2848                 void __user *newval, size_t newlen)
2849 {
2850         if (oldval && oldlenp) {
2851                 size_t olen;
2852
2853                 if (get_user(olen, oldlenp))
2854                         return -EFAULT;
2855                 if (olen) {
2856                         int val;
2857
2858                         if (olen < sizeof(int))
2859                                 return -EINVAL;
2860
2861                         val = *(int *)(table->data) / HZ;
2862                         if (put_user(val, (int __user *)oldval))
2863                                 return -EFAULT;
2864                         if (put_user(sizeof(int), oldlenp))
2865                                 return -EFAULT;
2866                 }
2867         }
2868         if (newval && newlen) {
2869                 int new;
2870                 if (newlen != sizeof(int))
2871                         return -EINVAL;
2872                 if (get_user(new, (int __user *)newval))
2873                         return -EFAULT;
2874                 *(int *)(table->data) = new*HZ;
2875         }
2876         return 1;
2877 }
2878
2879 /* Strategy function to convert jiffies to seconds */
2880 int sysctl_ms_jiffies(struct ctl_table *table, int __user *name, int nlen,
2881                 void __user *oldval, size_t __user *oldlenp,
2882                 void __user *newval, size_t newlen)
2883 {
2884         if (oldval && oldlenp) {
2885                 size_t olen;
2886
2887                 if (get_user(olen, oldlenp))
2888                         return -EFAULT;
2889                 if (olen) {
2890                         int val;
2891
2892                         if (olen < sizeof(int))
2893                                 return -EINVAL;
2894
2895                         val = jiffies_to_msecs(*(int *)(table->data));
2896                         if (put_user(val, (int __user *)oldval))
2897                                 return -EFAULT;
2898                         if (put_user(sizeof(int), oldlenp))
2899                                 return -EFAULT;
2900                 }
2901         }
2902         if (newval && newlen) {
2903                 int new;
2904                 if (newlen != sizeof(int))
2905                         return -EINVAL;
2906                 if (get_user(new, (int __user *)newval))
2907                         return -EFAULT;
2908                 *(int *)(table->data) = msecs_to_jiffies(new);
2909         }
2910         return 1;
2911 }
2912
2913
2914
2915 #else /* CONFIG_SYSCTL_SYSCALL */
2916
2917
2918 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2919 {
2920         struct __sysctl_args tmp;
2921         int error;
2922
2923         if (copy_from_user(&tmp, args, sizeof(tmp)))
2924                 return -EFAULT;
2925
2926         error = deprecated_sysctl_warning(&tmp);
2927
2928         /* If no error reading the parameters then just -ENOSYS ... */
2929         if (!error)
2930                 error = -ENOSYS;
2931
2932         return error;
2933 }
2934
2935 int sysctl_data(struct ctl_table *table, int __user *name, int nlen,
2936                   void __user *oldval, size_t __user *oldlenp,
2937                   void __user *newval, size_t newlen)
2938 {
2939         return -ENOSYS;
2940 }
2941
2942 int sysctl_string(struct ctl_table *table, int __user *name, int nlen,
2943                   void __user *oldval, size_t __user *oldlenp,
2944                   void __user *newval, size_t newlen)
2945 {
2946         return -ENOSYS;
2947 }
2948
2949 int sysctl_intvec(struct ctl_table *table, int __user *name, int nlen,
2950                 void __user *oldval, size_t __user *oldlenp,
2951                 void __user *newval, size_t newlen)
2952 {
2953         return -ENOSYS;
2954 }
2955
2956 int sysctl_jiffies(struct ctl_table *table, int __user *name, int nlen,
2957                 void __user *oldval, size_t __user *oldlenp,
2958                 void __user *newval, size_t newlen)
2959 {
2960         return -ENOSYS;
2961 }
2962
2963 int sysctl_ms_jiffies(struct ctl_table *table, int __user *name, int nlen,
2964                 void __user *oldval, size_t __user *oldlenp,
2965                 void __user *newval, size_t newlen)
2966 {
2967         return -ENOSYS;
2968 }
2969
2970 #endif /* CONFIG_SYSCTL_SYSCALL */
2971
2972 static int deprecated_sysctl_warning(struct __sysctl_args *args)
2973 {
2974         static int msg_count;
2975         int name[CTL_MAXNAME];
2976         int i;
2977
2978         /* Check args->nlen. */
2979         if (args->nlen < 0 || args->nlen > CTL_MAXNAME)
2980                 return -ENOTDIR;
2981
2982         /* Read in the sysctl name for better debug message logging */
2983         for (i = 0; i < args->nlen; i++)
2984                 if (get_user(name[i], args->name + i))
2985                         return -EFAULT;
2986
2987         /* Ignore accesses to kernel.version */
2988         if ((args->nlen == 2) && (name[0] == CTL_KERN) && (name[1] == KERN_VERSION))
2989                 return 0;
2990
2991         if (msg_count < 5) {
2992                 msg_count++;
2993                 printk(KERN_INFO
2994                         "warning: process `%s' used the deprecated sysctl "
2995                         "system call with ", current->comm);
2996                 for (i = 0; i < args->nlen; i++)
2997                         printk("%d.", name[i]);
2998                 printk("\n");
2999         }
3000         return 0;
3001 }
3002
3003 /*
3004  * No sense putting this after each symbol definition, twice,
3005  * exception granted :-)
3006  */
3007 EXPORT_SYMBOL(proc_dointvec);
3008 EXPORT_SYMBOL(proc_dointvec_jiffies);
3009 EXPORT_SYMBOL(proc_dointvec_minmax);
3010 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
3011 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
3012 EXPORT_SYMBOL(proc_dostring);
3013 EXPORT_SYMBOL(proc_doulongvec_minmax);
3014 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
3015 EXPORT_SYMBOL(register_sysctl_table);
3016 EXPORT_SYMBOL(register_sysctl_paths);
3017 EXPORT_SYMBOL(sysctl_intvec);
3018 EXPORT_SYMBOL(sysctl_jiffies);
3019 EXPORT_SYMBOL(sysctl_ms_jiffies);
3020 EXPORT_SYMBOL(sysctl_string);
3021 EXPORT_SYMBOL(sysctl_data);
3022 EXPORT_SYMBOL(unregister_sysctl_table);