SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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/signal.h>
  27 #include <linux/proc_fs.h>
  28 #include <linux/security.h>
  29 #include <linux/ctype.h>
  30 #include <linux/kmemcheck.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/ratelimit.h>
  40 #include <linux/hugetlb.h>
  41 #include <linux/initrd.h>
  42 #include <linux/key.h>
  43 #include <linux/times.h>
  44 #include <linux/limits.h>
  45 #include <linux/dcache.h>
  46 #include <linux/syscalls.h>
  47 #include <linux/vmstat.h>
  48 #include <linux/nfs_fs.h>
  49 #include <linux/acpi.h>
  50 #include <linux/reboot.h>
  51 #include <linux/ftrace.h>
  52 #include <linux/slow-work.h>
  53 #include <linux/perf_event.h>
  54 #include <linux/kprobes.h>
  55
  56 #include <asm/uaccess.h>
  57 #include <asm/processor.h>
  58
  59 #ifdef CONFIG_X86
  60 #include <asm/nmi.h>
  61 #include <asm/stacktrace.h>
  62 #include <asm/io.h>
  63 #endif
  64 #ifdef CONFIG_CHR_DEV_SG
  65 #include <scsi/sg.h>
  66 #endif
  67
  68
  69 #if defined(CONFIG_SYSCTL)
  70
  71 /* External variables not in a header file. */
  72 extern int sysctl_overcommit_memory;
  73 extern int sysctl_overcommit_ratio;
  74 extern int sysctl_panic_on_oom;
  75 extern int sysctl_oom_kill_allocating_task;
  76 extern int sysctl_oom_dump_tasks;
  77 extern int max_threads;
  78 extern int core_uses_pid;
  79 extern int suid_dumpable;
  80 extern char core_pattern[];
  81 extern unsigned int core_pipe_limit;
  82 extern int pid_max;
  83 extern int min_free_kbytes;
  84 extern int pid_max_min, pid_max_max;
  85 extern int sysctl_drop_caches;
  86 extern int percpu_pagelist_fraction;
  87 extern int compat_log;
  88 extern int latencytop_enabled;
  89 extern int sysctl_nr_open_min, sysctl_nr_open_max;
  90 #ifndef CONFIG_MMU
  91 extern int sysctl_nr_trim_pages;
  92 #endif
  93 #ifdef CONFIG_BLOCK
  94 extern int blk_iopoll_enabled;
  95 #endif
  96
  97 /* Constants used for minimum and  maximum */
  98 #ifdef CONFIG_DETECT_SOFTLOCKUP
  99 static int sixty = 60;
 100 static int neg_one = -1;
 101 #endif
 102
 103 static int zero;
 104 static int __maybe_unused one = 1;
 105 static int __maybe_unused two = 2;
 106 static unsigned long one_ul = 1;
 107 static int one_hundred = 100;
 108 #ifdef CONFIG_PRINTK
 109 static int ten_thousand = 10000;
 110 #endif
 111
 112 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 113 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 114
 115 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 116 static int maxolduid = 65535;
 117 static int minolduid;
 118 static int min_percpu_pagelist_fract = 8;
 119
 120 static int ngroups_max = NGROUPS_MAX;
 121
 122 #ifdef CONFIG_SPARC
 123 #include <asm/system.h>
 124 #endif
 125
 126 #ifdef CONFIG_SPARC64
 127 extern int sysctl_tsb_ratio;
 128 #endif
 129
 130 #ifdef __hppa__
 131 extern int pwrsw_enabled;
 132 extern int unaligned_enabled;
 133 #endif
 134
 135 #ifdef CONFIG_S390
 136 #ifdef CONFIG_MATHEMU
 137 extern int sysctl_ieee_emulation_warnings;
 138 #endif
 139 extern int sysctl_userprocess_debug;
 140 extern int spin_retry;
 141 #endif
 142
 143 #ifdef CONFIG_BSD_PROCESS_ACCT
 144 extern int acct_parm[];
 145 #endif
 146
 147 #ifdef CONFIG_IA64
 148 extern int no_unaligned_warning;
 149 extern int unaligned_dump_stack;
 150 #endif
 151
 152 extern struct ratelimit_state printk_ratelimit_state;
 153
 154 #ifdef CONFIG_RT_MUTEXES
 155 extern int max_lock_depth;
 156 #endif
 157
 158 #ifdef CONFIG_PROC_SYSCTL
 159 static int proc_do_cad_pid(struct ctl_table *table, int write,
 160                   void __user *buffer, size_t *lenp, loff_t *ppos);
 161 static int proc_taint(struct ctl_table *table, int write,
 162                                void __user *buffer, size_t *lenp, loff_t *ppos);
 163 #endif
 164
 165 static struct ctl_table root_table[];
 166 static struct ctl_table_root sysctl_table_root;
 167 static struct ctl_table_header root_table_header = {
 168         .count = 1,
 169         .ctl_table = root_table,
 170         .ctl_entry = LIST_HEAD_INIT(sysctl_table_root.default_set.list),
 171         .root = &sysctl_table_root,
 172         .set = &sysctl_table_root.default_set,
 173 };
 174 static struct ctl_table_root sysctl_table_root = {
 175         .root_list = LIST_HEAD_INIT(sysctl_table_root.root_list),
 176         .default_set.list = LIST_HEAD_INIT(root_table_header.ctl_entry),
 177 };
 178
 179 static struct ctl_table kern_table[];
 180 static struct ctl_table vm_table[];
 181 static struct ctl_table fs_table[];
 182 static struct ctl_table debug_table[];
 183 static struct ctl_table dev_table[];
 184 extern struct ctl_table random_table[];
 185 #ifdef CONFIG_INOTIFY_USER
 186 extern struct ctl_table inotify_table[];
 187 #endif
 188 #ifdef CONFIG_EPOLL
 189 extern struct ctl_table epoll_table[];
 190 #endif
 191
 192 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 193 int sysctl_legacy_va_layout;
 194 #endif
 195
 196 extern int prove_locking;
 197 extern int lock_stat;
 198
 199 /* The default sysctl tables: */
 200
 201 static struct ctl_table root_table[] = {
 202         {
 203                 .procname       = "kernel",
 204                 .mode           = 0555,
 205                 .child          = kern_table,
 206         },
 207         {
 208                 .procname       = "vm",
 209                 .mode           = 0555,
 210                 .child          = vm_table,
 211         },
 212         {
 213                 .procname       = "fs",
 214                 .mode           = 0555,
 215                 .child          = fs_table,
 216         },
 217         {
 218                 .procname       = "debug",
 219                 .mode           = 0555,
 220                 .child          = debug_table,
 221         },
 222         {
 223                 .procname       = "dev",
 224                 .mode           = 0555,
 225                 .child          = dev_table,
 226         },
 227 /*
 228  * NOTE: do not add new entries to this table unless you have read
 229  * Documentation/sysctl/ctl_unnumbered.txt
 230  */
 231         { }
 232 };
 233
 234 #ifdef CONFIG_SCHED_DEBUG
 235 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 236 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 237 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 238 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 239 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 240 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 241 static int min_sched_shares_ratelimit = 100000; /* 100 usec */
 242 static int max_sched_shares_ratelimit = NSEC_PER_SEC; /* 1 second */
 243 #endif
 244
 245 static struct ctl_table kern_table[] = {
 246         {
 247                 .procname       = "sched_child_runs_first",
 248                 .data           = &sysctl_sched_child_runs_first,
 249                 .maxlen         = sizeof(unsigned int),
 250                 .mode           = 0644,
 251                 .proc_handler   = proc_dointvec,
 252         },
 253 #ifdef CONFIG_SCHED_DEBUG
 254         {
 255                 .procname       = "sched_min_granularity_ns",
 256                 .data           = &sysctl_sched_min_granularity,
 257                 .maxlen         = sizeof(unsigned int),
 258                 .mode           = 0644,
 259                 .proc_handler   = sched_proc_update_handler,
 260                 .extra1         = &min_sched_granularity_ns,
 261                 .extra2         = &max_sched_granularity_ns,
 262         },
 263         {
 264                 .procname       = "sched_latency_ns",
 265                 .data           = &sysctl_sched_latency,
 266                 .maxlen         = sizeof(unsigned int),
 267                 .mode           = 0644,
 268                 .proc_handler   = sched_proc_update_handler,
 269                 .extra1         = &min_sched_granularity_ns,
 270                 .extra2         = &max_sched_granularity_ns,
 271         },
 272         {
 273                 .procname       = "sched_wakeup_granularity_ns",
 274                 .data           = &sysctl_sched_wakeup_granularity,
 275                 .maxlen         = sizeof(unsigned int),
 276                 .mode           = 0644,
 277                 .proc_handler   = sched_proc_update_handler,
 278                 .extra1         = &min_wakeup_granularity_ns,
 279                 .extra2         = &max_wakeup_granularity_ns,
 280         },
 281         {
 282                 .procname       = "sched_shares_ratelimit",
 283                 .data           = &sysctl_sched_shares_ratelimit,
 284                 .maxlen         = sizeof(unsigned int),
 285                 .mode           = 0644,
 286                 .proc_handler   = sched_proc_update_handler,
 287                 .extra1         = &min_sched_shares_ratelimit,
 288                 .extra2         = &max_sched_shares_ratelimit,
 289         },
 290         {
 291                 .procname       = "sched_tunable_scaling",
 292                 .data           = &sysctl_sched_tunable_scaling,
 293                 .maxlen         = sizeof(enum sched_tunable_scaling),
 294                 .mode           = 0644,
 295                 .proc_handler   = sched_proc_update_handler,
 296                 .extra1         = &min_sched_tunable_scaling,
 297                 .extra2         = &max_sched_tunable_scaling,
 298         },
 299         {
 300                 .procname       = "sched_shares_thresh",
 301                 .data           = &sysctl_sched_shares_thresh,
 302                 .maxlen         = sizeof(unsigned int),
 303                 .mode           = 0644,
 304                 .proc_handler   = proc_dointvec_minmax,
 305                 .extra1         = &zero,
 306         },
 307         {
 308                 .procname       = "sched_migration_cost",
 309                 .data           = &sysctl_sched_migration_cost,
 310                 .maxlen         = sizeof(unsigned int),
 311                 .mode           = 0644,
 312                 .proc_handler   = proc_dointvec,
 313         },
 314         {
 315                 .procname       = "sched_nr_migrate",
 316                 .data           = &sysctl_sched_nr_migrate,
 317                 .maxlen         = sizeof(unsigned int),
 318                 .mode           = 0644,
 319                 .proc_handler   = proc_dointvec,
 320         },
 321         {
 322                 .procname       = "sched_time_avg",
 323                 .data           = &sysctl_sched_time_avg,
 324                 .maxlen         = sizeof(unsigned int),
 325                 .mode           = 0644,
 326                 .proc_handler   = proc_dointvec,
 327         },
 328         {
 329                 .procname       = "timer_migration",
 330                 .data           = &sysctl_timer_migration,
 331                 .maxlen         = sizeof(unsigned int),
 332                 .mode           = 0644,
 333                 .proc_handler   = proc_dointvec_minmax,
 334                 .extra1         = &zero,
 335                 .extra2         = &one,
 336         },
 337 #endif
 338         {
 339                 .procname       = "sched_rt_period_us",
 340                 .data           = &sysctl_sched_rt_period,
 341                 .maxlen         = sizeof(unsigned int),
 342                 .mode           = 0644,
 343                 .proc_handler   = sched_rt_handler,
 344         },
 345         {
 346                 .procname       = "sched_rt_runtime_us",
 347                 .data           = &sysctl_sched_rt_runtime,
 348                 .maxlen         = sizeof(int),
 349                 .mode           = 0644,
 350                 .proc_handler   = sched_rt_handler,
 351         },
 352         {
 353                 .procname       = "sched_compat_yield",
 354                 .data           = &sysctl_sched_compat_yield,
 355                 .maxlen         = sizeof(unsigned int),
 356                 .mode           = 0644,
 357                 .proc_handler   = proc_dointvec,
 358         },
 359 #ifdef CONFIG_PROVE_LOCKING
 360         {
 361                 .procname       = "prove_locking",
 362                 .data           = &prove_locking,
 363                 .maxlen         = sizeof(int),
 364                 .mode           = 0644,
 365                 .proc_handler   = proc_dointvec,
 366         },
 367 #endif
 368 #ifdef CONFIG_LOCK_STAT
 369         {
 370                 .procname       = "lock_stat",
 371                 .data           = &lock_stat,
 372                 .maxlen         = sizeof(int),
 373                 .mode           = 0644,
 374                 .proc_handler   = proc_dointvec,
 375         },
 376 #endif
 377         {
 378                 .procname       = "panic",
 379                 .data           = &panic_timeout,
 380                 .maxlen         = sizeof(int),
 381                 .mode           = 0644,
 382                 .proc_handler   = proc_dointvec,
 383         },
 384         {
 385                 .procname       = "core_uses_pid",
 386                 .data           = &core_uses_pid,
 387                 .maxlen         = sizeof(int),
 388                 .mode           = 0644,
 389                 .proc_handler   = proc_dointvec,
 390         },
 391         {
 392                 .procname       = "core_pattern",
 393                 .data           = core_pattern,
 394                 .maxlen         = CORENAME_MAX_SIZE,
 395                 .mode           = 0644,
 396                 .proc_handler   = proc_dostring,
 397         },
 398         {
 399                 .procname       = "core_pipe_limit",
 400                 .data           = &core_pipe_limit,
 401                 .maxlen         = sizeof(unsigned int),
 402                 .mode           = 0644,
 403                 .proc_handler   = proc_dointvec,
 404         },
 405 #ifdef CONFIG_PROC_SYSCTL
 406         {
 407                 .procname       = "tainted",
 408                 .maxlen         = sizeof(long),
 409                 .mode           = 0644,
 410                 .proc_handler   = proc_taint,
 411         },
 412 #endif
 413 #ifdef CONFIG_LATENCYTOP
 414         {
 415                 .procname       = "latencytop",
 416                 .data           = &latencytop_enabled,
 417                 .maxlen         = sizeof(int),
 418                 .mode           = 0644,
 419                 .proc_handler   = proc_dointvec,
 420         },
 421 #endif
 422 #ifdef CONFIG_BLK_DEV_INITRD
 423         {
 424                 .procname       = "real-root-dev",
 425                 .data           = &real_root_dev,
 426                 .maxlen         = sizeof(int),
 427                 .mode           = 0644,
 428                 .proc_handler   = proc_dointvec,
 429         },
 430 #endif
 431         {
 432                 .procname       = "print-fatal-signals",
 433                 .data           = &print_fatal_signals,
 434                 .maxlen         = sizeof(int),
 435                 .mode           = 0644,
 436                 .proc_handler   = proc_dointvec,
 437         },
 438 #ifdef CONFIG_SPARC
 439         {
 440                 .procname       = "reboot-cmd",
 441                 .data           = reboot_command,
 442                 .maxlen         = 256,
 443                 .mode           = 0644,
 444                 .proc_handler   = proc_dostring,
 445         },
 446         {
 447                 .procname       = "stop-a",
 448                 .data           = &stop_a_enabled,
 449                 .maxlen         = sizeof (int),
 450                 .mode           = 0644,
 451                 .proc_handler   = proc_dointvec,
 452         },
 453         {
 454                 .procname       = "scons-poweroff",
 455                 .data           = &scons_pwroff,
 456                 .maxlen         = sizeof (int),
 457                 .mode           = 0644,
 458                 .proc_handler   = proc_dointvec,
 459         },
 460 #endif
 461 #ifdef CONFIG_SPARC64
 462         {
 463                 .procname       = "tsb-ratio",
 464                 .data           = &sysctl_tsb_ratio,
 465                 .maxlen         = sizeof (int),
 466                 .mode           = 0644,
 467                 .proc_handler   = proc_dointvec,
 468         },
 469 #endif
 470 #ifdef __hppa__
 471         {
 472                 .procname       = "soft-power",
 473                 .data           = &pwrsw_enabled,
 474                 .maxlen         = sizeof (int),
 475                 .mode           = 0644,
 476                 .proc_handler   = proc_dointvec,
 477         },
 478         {
 479                 .procname       = "unaligned-trap",
 480                 .data           = &unaligned_enabled,
 481                 .maxlen         = sizeof (int),
 482                 .mode           = 0644,
 483                 .proc_handler   = proc_dointvec,
 484         },
 485 #endif
 486         {
 487                 .procname       = "ctrl-alt-del",
 488                 .data           = &C_A_D,
 489                 .maxlen         = sizeof(int),
 490                 .mode           = 0644,
 491                 .proc_handler   = proc_dointvec,
 492         },
 493 #ifdef CONFIG_FUNCTION_TRACER
 494         {
 495                 .procname       = "ftrace_enabled",
 496                 .data           = &ftrace_enabled,
 497                 .maxlen         = sizeof(int),
 498                 .mode           = 0644,
 499                 .proc_handler   = ftrace_enable_sysctl,
 500         },
 501 #endif
 502 #ifdef CONFIG_STACK_TRACER
 503         {
 504                 .procname       = "stack_tracer_enabled",
 505                 .data           = &stack_tracer_enabled,
 506                 .maxlen         = sizeof(int),
 507                 .mode           = 0644,
 508                 .proc_handler   = stack_trace_sysctl,
 509         },
 510 #endif
 511 #ifdef CONFIG_TRACING
 512         {
 513                 .procname       = "ftrace_dump_on_oops",
 514                 .data           = &ftrace_dump_on_oops,
 515                 .maxlen         = sizeof(int),
 516                 .mode           = 0644,
 517                 .proc_handler   = proc_dointvec,
 518         },
 519 #endif
 520 #ifdef CONFIG_MODULES
 521         {
 522                 .procname       = "modprobe",
 523                 .data           = &modprobe_path,
 524                 .maxlen         = KMOD_PATH_LEN,
 525                 .mode           = 0644,
 526                 .proc_handler   = proc_dostring,
 527         },
 528         {
 529                 .procname       = "modules_disabled",
 530                 .data           = &modules_disabled,
 531                 .maxlen         = sizeof(int),
 532                 .mode           = 0644,
 533                 /* only handle a transition from default "0" to "1" */
 534                 .proc_handler   = proc_dointvec_minmax,
 535                 .extra1         = &one,
 536                 .extra2         = &one,
 537         },
 538 #endif
 539 #if defined(CONFIG_HOTPLUG) && defined(CONFIG_NET)
 540         {
 541                 .procname       = "hotplug",
 542                 .data           = &uevent_helper,
 543                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 544                 .mode           = 0644,
 545                 .proc_handler   = proc_dostring,
 546         },
 547 #endif
 548 #ifdef CONFIG_CHR_DEV_SG
 549         {
 550                 .procname       = "sg-big-buff",
 551                 .data           = &sg_big_buff,
 552                 .maxlen         = sizeof (int),
 553                 .mode           = 0444,
 554                 .proc_handler   = proc_dointvec,
 555         },
 556 #endif
 557 #ifdef CONFIG_BSD_PROCESS_ACCT
 558         {
 559                 .procname       = "acct",
 560                 .data           = &acct_parm,
 561                 .maxlen         = 3*sizeof(int),
 562                 .mode           = 0644,
 563                 .proc_handler   = proc_dointvec,
 564         },
 565 #endif
 566 #ifdef CONFIG_MAGIC_SYSRQ
 567         {
 568                 .procname       = "sysrq",
 569                 .data           = &__sysrq_enabled,
 570                 .maxlen         = sizeof (int),
 571                 .mode           = 0644,
 572                 .proc_handler   = proc_dointvec,
 573         },
 574 #endif
 575 #ifdef CONFIG_PROC_SYSCTL
 576         {
 577                 .procname       = "cad_pid",
 578                 .data           = NULL,
 579                 .maxlen         = sizeof (int),
 580                 .mode           = 0600,
 581                 .proc_handler   = proc_do_cad_pid,
 582         },
 583 #endif
 584         {
 585                 .procname       = "threads-max",
 586                 .data           = &max_threads,
 587                 .maxlen         = sizeof(int),
 588                 .mode           = 0644,
 589                 .proc_handler   = proc_dointvec,
 590         },
 591         {
 592                 .procname       = "random",
 593                 .mode           = 0555,
 594                 .child          = random_table,
 595         },
 596         {
 597                 .procname       = "overflowuid",
 598                 .data           = &overflowuid,
 599                 .maxlen         = sizeof(int),
 600                 .mode           = 0644,
 601                 .proc_handler   = proc_dointvec_minmax,
 602                 .extra1         = &minolduid,
 603                 .extra2         = &maxolduid,
 604         },
 605         {
 606                 .procname       = "overflowgid",
 607                 .data           = &overflowgid,
 608                 .maxlen         = sizeof(int),
 609                 .mode           = 0644,
 610                 .proc_handler   = proc_dointvec_minmax,
 611                 .extra1         = &minolduid,
 612                 .extra2         = &maxolduid,
 613         },
 614 #ifdef CONFIG_S390
 615 #ifdef CONFIG_MATHEMU
 616         {
 617                 .procname       = "ieee_emulation_warnings",
 618                 .data           = &sysctl_ieee_emulation_warnings,
 619                 .maxlen         = sizeof(int),
 620                 .mode           = 0644,
 621                 .proc_handler   = proc_dointvec,
 622         },
 623 #endif
 624         {
 625                 .procname       = "userprocess_debug",
 626                 .data           = &sysctl_userprocess_debug,
 627                 .maxlen         = sizeof(int),
 628                 .mode           = 0644,
 629                 .proc_handler   = proc_dointvec,
 630         },
 631 #endif
 632         {
 633                 .procname       = "pid_max",
 634                 .data           = &pid_max,
 635                 .maxlen         = sizeof (int),
 636                 .mode           = 0644,
 637                 .proc_handler   = proc_dointvec_minmax,
 638                 .extra1         = &pid_max_min,
 639                 .extra2         = &pid_max_max,
 640         },
 641         {
 642                 .procname       = "panic_on_oops",
 643                 .data           = &panic_on_oops,
 644                 .maxlen         = sizeof(int),
 645                 .mode           = 0644,
 646                 .proc_handler   = proc_dointvec,
 647         },
 648 #if defined CONFIG_PRINTK
 649         {
 650                 .procname       = "printk",
 651                 .data           = &console_loglevel,
 652                 .maxlen         = 4*sizeof(int),
 653                 .mode           = 0644,
 654                 .proc_handler   = proc_dointvec,
 655         },
 656         {
 657                 .procname       = "printk_ratelimit",
 658                 .data           = &printk_ratelimit_state.interval,
 659                 .maxlen         = sizeof(int),
 660                 .mode           = 0644,
 661                 .proc_handler   = proc_dointvec_jiffies,
 662         },
 663         {
 664                 .procname       = "printk_ratelimit_burst",
 665                 .data           = &printk_ratelimit_state.burst,
 666                 .maxlen         = sizeof(int),
 667                 .mode           = 0644,
 668                 .proc_handler   = proc_dointvec,
 669         },
 670         {
 671                 .procname       = "printk_delay",
 672                 .data           = &printk_delay_msec,
 673                 .maxlen         = sizeof(int),
 674                 .mode           = 0644,
 675                 .proc_handler   = proc_dointvec_minmax,
 676                 .extra1         = &zero,
 677                 .extra2         = &ten_thousand,
 678         },
 679 #endif
 680         {
 681                 .procname       = "ngroups_max",
 682                 .data           = &ngroups_max,
 683                 .maxlen         = sizeof (int),
 684                 .mode           = 0444,
 685                 .proc_handler   = proc_dointvec,
 686         },
 687 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 688         {
 689                 .procname       = "unknown_nmi_panic",
 690                 .data           = &unknown_nmi_panic,
 691                 .maxlen         = sizeof (int),
 692                 .mode           = 0644,
 693                 .proc_handler   = proc_dointvec,
 694         },
 695         {
 696                 .procname       = "nmi_watchdog",
 697                 .data           = &nmi_watchdog_enabled,
 698                 .maxlen         = sizeof (int),
 699                 .mode           = 0644,
 700                 .proc_handler   = proc_nmi_enabled,
 701         },
 702 #endif
 703 #if defined(CONFIG_X86)
 704         {
 705                 .procname       = "panic_on_unrecovered_nmi",
 706                 .data           = &panic_on_unrecovered_nmi,
 707                 .maxlen         = sizeof(int),
 708                 .mode           = 0644,
 709                 .proc_handler   = proc_dointvec,
 710         },
 711         {
 712                 .procname       = "panic_on_io_nmi",
 713                 .data           = &panic_on_io_nmi,
 714                 .maxlen         = sizeof(int),
 715                 .mode           = 0644,
 716                 .proc_handler   = proc_dointvec,
 717         },
 718         {
 719                 .procname       = "bootloader_type",
 720                 .data           = &bootloader_type,
 721                 .maxlen         = sizeof (int),
 722                 .mode           = 0444,
 723                 .proc_handler   = proc_dointvec,
 724         },
 725         {
 726                 .procname       = "bootloader_version",
 727                 .data           = &bootloader_version,
 728                 .maxlen         = sizeof (int),
 729                 .mode           = 0444,
 730                 .proc_handler   = proc_dointvec,
 731         },
 732         {
 733                 .procname       = "kstack_depth_to_print",
 734                 .data           = &kstack_depth_to_print,
 735                 .maxlen         = sizeof(int),
 736                 .mode           = 0644,
 737                 .proc_handler   = proc_dointvec,
 738         },
 739         {
 740                 .procname       = "io_delay_type",
 741                 .data           = &io_delay_type,
 742                 .maxlen         = sizeof(int),
 743                 .mode           = 0644,
 744                 .proc_handler   = proc_dointvec,
 745         },
 746 #endif
 747 #if defined(CONFIG_MMU)
 748         {
 749                 .procname       = "randomize_va_space",
 750                 .data           = &randomize_va_space,
 751                 .maxlen         = sizeof(int),
 752                 .mode           = 0644,
 753                 .proc_handler   = proc_dointvec,
 754         },
 755 #endif
 756 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 757         {
 758                 .procname       = "spin_retry",
 759                 .data           = &spin_retry,
 760                 .maxlen         = sizeof (int),
 761                 .mode           = 0644,
 762                 .proc_handler   = proc_dointvec,
 763         },
 764 #endif
 765 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 766         {
 767                 .procname       = "acpi_video_flags",
 768                 .data           = &acpi_realmode_flags,
 769                 .maxlen         = sizeof (unsigned long),
 770                 .mode           = 0644,
 771                 .proc_handler   = proc_doulongvec_minmax,
 772         },
 773 #endif
 774 #ifdef CONFIG_IA64
 775         {
 776                 .procname       = "ignore-unaligned-usertrap",
 777                 .data           = &no_unaligned_warning,
 778                 .maxlen         = sizeof (int),
 779                 .mode           = 0644,
 780                 .proc_handler   = proc_dointvec,
 781         },
 782         {
 783                 .procname       = "unaligned-dump-stack",
 784                 .data           = &unaligned_dump_stack,
 785                 .maxlen         = sizeof (int),
 786                 .mode           = 0644,
 787                 .proc_handler   = proc_dointvec,
 788         },
 789 #endif
 790 #ifdef CONFIG_DETECT_SOFTLOCKUP
 791         {
 792                 .procname       = "softlockup_panic",
 793                 .data           = &softlockup_panic,
 794                 .maxlen         = sizeof(int),
 795                 .mode           = 0644,
 796                 .proc_handler   = proc_dointvec_minmax,
 797                 .extra1         = &zero,
 798                 .extra2         = &one,
 799         },
 800         {
 801                 .procname       = "softlockup_thresh",
 802                 .data           = &softlockup_thresh,
 803                 .maxlen         = sizeof(int),
 804                 .mode           = 0644,
 805                 .proc_handler   = proc_dosoftlockup_thresh,
 806                 .extra1         = &neg_one,
 807                 .extra2         = &sixty,
 808         },
 809 #endif
 810 #ifdef CONFIG_DETECT_HUNG_TASK
 811         {
 812                 .procname       = "hung_task_panic",
 813                 .data           = &sysctl_hung_task_panic,
 814                 .maxlen         = sizeof(int),
 815                 .mode           = 0644,
 816                 .proc_handler   = proc_dointvec_minmax,
 817                 .extra1         = &zero,
 818                 .extra2         = &one,
 819         },
 820         {
 821                 .procname       = "hung_task_check_count",
 822                 .data           = &sysctl_hung_task_check_count,
 823                 .maxlen         = sizeof(unsigned long),
 824                 .mode           = 0644,
 825                 .proc_handler   = proc_doulongvec_minmax,
 826         },
 827         {
 828                 .procname       = "hung_task_timeout_secs",
 829                 .data           = &sysctl_hung_task_timeout_secs,
 830                 .maxlen         = sizeof(unsigned long),
 831                 .mode           = 0644,
 832                 .proc_handler   = proc_dohung_task_timeout_secs,
 833         },
 834         {
 835                 .procname       = "hung_task_warnings",
 836                 .data           = &sysctl_hung_task_warnings,
 837                 .maxlen         = sizeof(unsigned long),
 838                 .mode           = 0644,
 839                 .proc_handler   = proc_doulongvec_minmax,
 840         },
 841 #endif
 842 #ifdef CONFIG_COMPAT
 843         {
 844                 .procname       = "compat-log",
 845                 .data           = &compat_log,
 846                 .maxlen         = sizeof (int),
 847                 .mode           = 0644,
 848                 .proc_handler   = proc_dointvec,
 849         },
 850 #endif
 851 #ifdef CONFIG_RT_MUTEXES
 852         {
 853                 .procname       = "max_lock_depth",
 854                 .data           = &max_lock_depth,
 855                 .maxlen         = sizeof(int),
 856                 .mode           = 0644,
 857                 .proc_handler   = proc_dointvec,
 858         },
 859 #endif
 860         {
 861                 .procname       = "poweroff_cmd",
 862                 .data           = &poweroff_cmd,
 863                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 864                 .mode           = 0644,
 865                 .proc_handler   = proc_dostring,
 866         },
 867 #ifdef CONFIG_KEYS
 868         {
 869                 .procname       = "keys",
 870                 .mode           = 0555,
 871                 .child          = key_sysctls,
 872         },
 873 #endif
 874 #ifdef CONFIG_RCU_TORTURE_TEST
 875         {
 876                 .procname       = "rcutorture_runnable",
 877                 .data           = &rcutorture_runnable,
 878                 .maxlen         = sizeof(int),
 879                 .mode           = 0644,
 880                 .proc_handler   = proc_dointvec,
 881         },
 882 #endif
 883 #ifdef CONFIG_SLOW_WORK
 884         {
 885                 .procname       = "slow-work",
 886                 .mode           = 0555,
 887                 .child          = slow_work_sysctls,
 888         },
 889 #endif
 890 #ifdef CONFIG_PERF_EVENTS
 891         {
 892                 .procname       = "perf_event_paranoid",
 893                 .data           = &sysctl_perf_event_paranoid,
 894                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
 895                 .mode           = 0644,
 896                 .proc_handler   = proc_dointvec,
 897         },
 898         {
 899                 .procname       = "perf_event_mlock_kb",
 900                 .data           = &sysctl_perf_event_mlock,
 901                 .maxlen         = sizeof(sysctl_perf_event_mlock),
 902                 .mode           = 0644,
 903                 .proc_handler   = proc_dointvec,
 904         },
 905         {
 906                 .procname       = "perf_event_max_sample_rate",
 907                 .data           = &sysctl_perf_event_sample_rate,
 908                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
 909                 .mode           = 0644,
 910                 .proc_handler   = proc_dointvec,
 911         },
 912 #endif
 913 #ifdef CONFIG_KMEMCHECK
 914         {
 915                 .procname       = "kmemcheck",
 916                 .data           = &kmemcheck_enabled,
 917                 .maxlen         = sizeof(int),
 918                 .mode           = 0644,
 919                 .proc_handler   = proc_dointvec,
 920         },
 921 #endif
 922 #ifdef CONFIG_BLOCK
 923         {
 924                 .procname       = "blk_iopoll",
 925                 .data           = &blk_iopoll_enabled,
 926                 .maxlen         = sizeof(int),
 927                 .mode           = 0644,
 928                 .proc_handler   = proc_dointvec,
 929         },
 930 #endif
 931 /*
 932  * NOTE: do not add new entries to this table unless you have read
 933  * Documentation/sysctl/ctl_unnumbered.txt
 934  */
 935         { }
 936 };
 937
 938 static struct ctl_table vm_table[] = {
 939         {
 940                 .procname       = "overcommit_memory",
 941                 .data           = &sysctl_overcommit_memory,
 942                 .maxlen         = sizeof(sysctl_overcommit_memory),
 943                 .mode           = 0644,
 944                 .proc_handler   = proc_dointvec,
 945         },
 946         {
 947                 .procname       = "panic_on_oom",
 948                 .data           = &sysctl_panic_on_oom,
 949                 .maxlen         = sizeof(sysctl_panic_on_oom),
 950                 .mode           = 0644,
 951                 .proc_handler   = proc_dointvec,
 952         },
 953         {
 954                 .procname       = "oom_kill_allocating_task",
 955                 .data           = &sysctl_oom_kill_allocating_task,
 956                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
 957                 .mode           = 0644,
 958                 .proc_handler   = proc_dointvec,
 959         },
 960         {
 961                 .procname       = "oom_dump_tasks",
 962                 .data           = &sysctl_oom_dump_tasks,
 963                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
 964                 .mode           = 0644,
 965                 .proc_handler   = proc_dointvec,
 966         },
 967         {
 968                 .procname       = "overcommit_ratio",
 969                 .data           = &sysctl_overcommit_ratio,
 970                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 971                 .mode           = 0644,
 972                 .proc_handler   = proc_dointvec,
 973         },
 974         {
 975                 .procname       = "page-cluster",
 976                 .data           = &page_cluster,
 977                 .maxlen         = sizeof(int),
 978                 .mode           = 0644,
 979                 .proc_handler   = proc_dointvec,
 980         },
 981         {
 982                 .procname       = "dirty_background_ratio",
 983                 .data           = &dirty_background_ratio,
 984                 .maxlen         = sizeof(dirty_background_ratio),
 985                 .mode           = 0644,
 986                 .proc_handler   = dirty_background_ratio_handler,
 987                 .extra1         = &zero,
 988                 .extra2         = &one_hundred,
 989         },
 990         {
 991                 .procname       = "dirty_background_bytes",
 992                 .data           = &dirty_background_bytes,
 993                 .maxlen         = sizeof(dirty_background_bytes),
 994                 .mode           = 0644,
 995                 .proc_handler   = dirty_background_bytes_handler,
 996                 .extra1         = &one_ul,
 997         },
 998         {
 999                 .procname       = "dirty_ratio",
1000                 .data           = &vm_dirty_ratio,
1001                 .maxlen         = sizeof(vm_dirty_ratio),
1002                 .mode           = 0644,
1003                 .proc_handler   = dirty_ratio_handler,
1004                 .extra1         = &zero,
1005                 .extra2         = &one_hundred,
1006         },
1007         {
1008                 .procname       = "dirty_bytes",
1009                 .data           = &vm_dirty_bytes,
1010                 .maxlen         = sizeof(vm_dirty_bytes),
1011                 .mode           = 0644,
1012                 .proc_handler   = dirty_bytes_handler,
1013                 .extra1         = &dirty_bytes_min,
1014         },
1015         {
1016                 .procname       = "dirty_writeback_centisecs",
1017                 .data           = &dirty_writeback_interval,
1018                 .maxlen         = sizeof(dirty_writeback_interval),
1019                 .mode           = 0644,
1020                 .proc_handler   = dirty_writeback_centisecs_handler,
1021         },
1022         {
1023                 .procname       = "dirty_expire_centisecs",
1024                 .data           = &dirty_expire_interval,
1025                 .maxlen         = sizeof(dirty_expire_interval),
1026                 .mode           = 0644,
1027                 .proc_handler   = proc_dointvec,
1028         },
1029         {
1030                 .procname       = "nr_pdflush_threads",
1031                 .data           = &nr_pdflush_threads,
1032                 .maxlen         = sizeof nr_pdflush_threads,
1033                 .mode           = 0444 /* read-only*/,
1034                 .proc_handler   = proc_dointvec,
1035         },
1036         {
1037                 .procname       = "swappiness",
1038                 .data           = &vm_swappiness,
1039                 .maxlen         = sizeof(vm_swappiness),
1040                 .mode           = 0644,
1041                 .proc_handler   = proc_dointvec_minmax,
1042                 .extra1         = &zero,
1043                 .extra2         = &one_hundred,
1044         },
1045 #ifdef CONFIG_HUGETLB_PAGE
1046         {
1047                 .procname       = "nr_hugepages",
1048                 .data           = NULL,
1049                 .maxlen         = sizeof(unsigned long),
1050                 .mode           = 0644,
1051                 .proc_handler   = hugetlb_sysctl_handler,
1052                 .extra1         = (void *)&hugetlb_zero,
1053                 .extra2         = (void *)&hugetlb_infinity,
1054         },
1055 #ifdef CONFIG_NUMA
1056         {
1057                 .procname       = "nr_hugepages_mempolicy",
1058                 .data           = NULL,
1059                 .maxlen         = sizeof(unsigned long),
1060                 .mode           = 0644,
1061                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1062                 .extra1         = (void *)&hugetlb_zero,
1063                 .extra2         = (void *)&hugetlb_infinity,
1064         },
1065 #endif
1066          {
1067                 .procname       = "hugetlb_shm_group",
1068                 .data           = &sysctl_hugetlb_shm_group,
1069                 .maxlen         = sizeof(gid_t),
1070                 .mode           = 0644,
1071                 .proc_handler   = proc_dointvec,
1072          },
1073          {
1074                 .procname       = "hugepages_treat_as_movable",
1075                 .data           = &hugepages_treat_as_movable,
1076                 .maxlen         = sizeof(int),
1077                 .mode           = 0644,
1078                 .proc_handler   = hugetlb_treat_movable_handler,
1079         },
1080         {
1081                 .procname       = "nr_overcommit_hugepages",
1082                 .data           = NULL,
1083                 .maxlen         = sizeof(unsigned long),
1084                 .mode           = 0644,
1085                 .proc_handler   = hugetlb_overcommit_handler,
1086                 .extra1         = (void *)&hugetlb_zero,
1087                 .extra2         = (void *)&hugetlb_infinity,
1088         },
1089 #endif
1090         {
1091                 .procname       = "lowmem_reserve_ratio",
1092                 .data           = &sysctl_lowmem_reserve_ratio,
1093                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1094                 .mode           = 0644,
1095                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1096         },
1097         {
1098                 .procname       = "drop_caches",
1099                 .data           = &sysctl_drop_caches,
1100                 .maxlen         = sizeof(int),
1101                 .mode           = 0644,
1102                 .proc_handler   = drop_caches_sysctl_handler,
1103         },
1104         {
1105                 .procname       = "min_free_kbytes",
1106                 .data           = &min_free_kbytes,
1107                 .maxlen         = sizeof(min_free_kbytes),
1108                 .mode           = 0644,
1109                 .proc_handler   = min_free_kbytes_sysctl_handler,
1110                 .extra1         = &zero,
1111         },
1112         {
1113                 .procname       = "percpu_pagelist_fraction",
1114                 .data           = &percpu_pagelist_fraction,
1115                 .maxlen         = sizeof(percpu_pagelist_fraction),
1116                 .mode           = 0644,
1117                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1118                 .extra1         = &min_percpu_pagelist_fract,
1119         },
1120 #ifdef CONFIG_MMU
1121         {
1122                 .procname       = "max_map_count",
1123                 .data           = &sysctl_max_map_count,
1124                 .maxlen         = sizeof(sysctl_max_map_count),
1125                 .mode           = 0644,
1126                 .proc_handler   = proc_dointvec_minmax,
1127                 .extra1         = &zero,
1128         },
1129 #else
1130         {
1131                 .procname       = "nr_trim_pages",
1132                 .data           = &sysctl_nr_trim_pages,
1133                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1134                 .mode           = 0644,
1135                 .proc_handler   = proc_dointvec_minmax,
1136                 .extra1         = &zero,
1137         },
1138 #endif
1139         {
1140                 .procname       = "laptop_mode",
1141                 .data           = &laptop_mode,
1142                 .maxlen         = sizeof(laptop_mode),
1143                 .mode           = 0644,
1144                 .proc_handler   = proc_dointvec_jiffies,
1145         },
1146         {
1147                 .procname       = "block_dump",
1148                 .data           = &block_dump,
1149                 .maxlen         = sizeof(block_dump),
1150                 .mode           = 0644,
1151                 .proc_handler   = proc_dointvec,
1152                 .extra1         = &zero,
1153         },
1154         {
1155                 .procname       = "vfs_cache_pressure",
1156                 .data           = &sysctl_vfs_cache_pressure,
1157                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1158                 .mode           = 0644,
1159                 .proc_handler   = proc_dointvec,
1160                 .extra1         = &zero,
1161         },
1162 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1163         {
1164                 .procname       = "legacy_va_layout",
1165                 .data           = &sysctl_legacy_va_layout,
1166                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1167                 .mode           = 0644,
1168                 .proc_handler   = proc_dointvec,
1169                 .extra1         = &zero,
1170         },
1171 #endif
1172 #ifdef CONFIG_NUMA
1173         {
1174                 .procname       = "zone_reclaim_mode",
1175                 .data           = &zone_reclaim_mode,
1176                 .maxlen         = sizeof(zone_reclaim_mode),
1177                 .mode           = 0644,
1178                 .proc_handler   = proc_dointvec,
1179                 .extra1         = &zero,
1180         },
1181         {
1182                 .procname       = "min_unmapped_ratio",
1183                 .data           = &sysctl_min_unmapped_ratio,
1184                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1185                 .mode           = 0644,
1186                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1187                 .extra1         = &zero,
1188                 .extra2         = &one_hundred,
1189         },
1190         {
1191                 .procname       = "min_slab_ratio",
1192                 .data           = &sysctl_min_slab_ratio,
1193                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1194                 .mode           = 0644,
1195                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1196                 .extra1         = &zero,
1197                 .extra2         = &one_hundred,
1198         },
1199 #endif
1200 #ifdef CONFIG_SMP
1201         {
1202                 .procname       = "stat_interval",
1203                 .data           = &sysctl_stat_interval,
1204                 .maxlen         = sizeof(sysctl_stat_interval),
1205                 .mode           = 0644,
1206                 .proc_handler   = proc_dointvec_jiffies,
1207         },
1208 #endif
1209 #ifdef CONFIG_MMU
1210         {
1211                 .procname       = "mmap_min_addr",
1212                 .data           = &dac_mmap_min_addr,
1213                 .maxlen         = sizeof(unsigned long),
1214                 .mode           = 0644,
1215                 .proc_handler   = mmap_min_addr_handler,
1216         },
1217 #endif
1218 #ifdef CONFIG_NUMA
1219         {
1220                 .procname       = "numa_zonelist_order",
1221                 .data           = &numa_zonelist_order,
1222                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1223                 .mode           = 0644,
1224                 .proc_handler   = numa_zonelist_order_handler,
1225         },
1226 #endif
1227 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1228    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1229         {
1230                 .procname       = "vdso_enabled",
1231                 .data           = &vdso_enabled,
1232                 .maxlen         = sizeof(vdso_enabled),
1233                 .mode           = 0644,
1234                 .proc_handler   = proc_dointvec,
1235                 .extra1         = &zero,
1236         },
1237 #endif
1238 #ifdef CONFIG_HIGHMEM
1239         {
1240                 .procname       = "highmem_is_dirtyable",
1241                 .data           = &vm_highmem_is_dirtyable,
1242                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1243                 .mode           = 0644,
1244                 .proc_handler   = proc_dointvec_minmax,
1245                 .extra1         = &zero,
1246                 .extra2         = &one,
1247         },
1248 #endif
1249         {
1250                 .procname       = "scan_unevictable_pages",
1251                 .data           = &scan_unevictable_pages,
1252                 .maxlen         = sizeof(scan_unevictable_pages),
1253                 .mode           = 0644,
1254                 .proc_handler   = scan_unevictable_handler,
1255         },
1256 #ifdef CONFIG_MEMORY_FAILURE
1257         {
1258                 .procname       = "memory_failure_early_kill",
1259                 .data           = &sysctl_memory_failure_early_kill,
1260                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1261                 .mode           = 0644,
1262                 .proc_handler   = proc_dointvec_minmax,
1263                 .extra1         = &zero,
1264                 .extra2         = &one,
1265         },
1266         {
1267                 .procname       = "memory_failure_recovery",
1268                 .data           = &sysctl_memory_failure_recovery,
1269                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1270                 .mode           = 0644,
1271                 .proc_handler   = proc_dointvec_minmax,
1272                 .extra1         = &zero,
1273                 .extra2         = &one,
1274         },
1275 #endif
1276
1277 /*
1278  * NOTE: do not add new entries to this table unless you have read
1279  * Documentation/sysctl/ctl_unnumbered.txt
1280  */
1281         { }
1282 };
1283
1284 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1285 static struct ctl_table binfmt_misc_table[] = {
1286         { }
1287 };
1288 #endif
1289
1290 static struct ctl_table fs_table[] = {
1291         {
1292                 .procname       = "inode-nr",
1293                 .data           = &inodes_stat,
1294                 .maxlen         = 2*sizeof(int),
1295                 .mode           = 0444,
1296                 .proc_handler   = proc_dointvec,
1297         },
1298         {
1299                 .procname       = "inode-state",
1300                 .data           = &inodes_stat,
1301                 .maxlen         = 7*sizeof(int),
1302                 .mode           = 0444,
1303                 .proc_handler   = proc_dointvec,
1304         },
1305         {
1306                 .procname       = "file-nr",
1307                 .data           = &files_stat,
1308                 .maxlen         = 3*sizeof(int),
1309                 .mode           = 0444,
1310                 .proc_handler   = proc_nr_files,
1311         },
1312         {
1313                 .procname       = "file-max",
1314                 .data           = &files_stat.max_files,
1315                 .maxlen         = sizeof(int),
1316                 .mode           = 0644,
1317                 .proc_handler   = proc_dointvec,
1318         },
1319         {
1320                 .procname       = "nr_open",
1321                 .data           = &sysctl_nr_open,
1322                 .maxlen         = sizeof(int),
1323                 .mode           = 0644,
1324                 .proc_handler   = proc_dointvec_minmax,
1325                 .extra1         = &sysctl_nr_open_min,
1326                 .extra2         = &sysctl_nr_open_max,
1327         },
1328         {
1329                 .procname       = "dentry-state",
1330                 .data           = &dentry_stat,
1331                 .maxlen         = 6*sizeof(int),
1332                 .mode           = 0444,
1333                 .proc_handler   = proc_dointvec,
1334         },
1335         {
1336                 .procname       = "overflowuid",
1337                 .data           = &fs_overflowuid,
1338                 .maxlen         = sizeof(int),
1339                 .mode           = 0644,
1340                 .proc_handler   = proc_dointvec_minmax,
1341                 .extra1         = &minolduid,
1342                 .extra2         = &maxolduid,
1343         },
1344         {
1345                 .procname       = "overflowgid",
1346                 .data           = &fs_overflowgid,
1347                 .maxlen         = sizeof(int),
1348                 .mode           = 0644,
1349                 .proc_handler   = proc_dointvec_minmax,
1350                 .extra1         = &minolduid,
1351                 .extra2         = &maxolduid,
1352         },
1353 #ifdef CONFIG_FILE_LOCKING
1354         {
1355                 .procname       = "leases-enable",
1356                 .data           = &leases_enable,
1357                 .maxlen         = sizeof(int),
1358                 .mode           = 0644,
1359                 .proc_handler   = proc_dointvec,
1360         },
1361 #endif
1362 #ifdef CONFIG_DNOTIFY
1363         {
1364                 .procname       = "dir-notify-enable",
1365                 .data           = &dir_notify_enable,
1366                 .maxlen         = sizeof(int),
1367                 .mode           = 0644,
1368                 .proc_handler   = proc_dointvec,
1369         },
1370 #endif
1371 #ifdef CONFIG_MMU
1372 #ifdef CONFIG_FILE_LOCKING
1373         {
1374                 .procname       = "lease-break-time",
1375                 .data           = &lease_break_time,
1376                 .maxlen         = sizeof(int),
1377                 .mode           = 0644,
1378                 .proc_handler   = proc_dointvec,
1379         },
1380 #endif
1381 #ifdef CONFIG_AIO
1382         {
1383                 .procname       = "aio-nr",
1384                 .data           = &aio_nr,
1385                 .maxlen         = sizeof(aio_nr),
1386                 .mode           = 0444,
1387                 .proc_handler   = proc_doulongvec_minmax,
1388         },
1389         {
1390                 .procname       = "aio-max-nr",
1391                 .data           = &aio_max_nr,
1392                 .maxlen         = sizeof(aio_max_nr),
1393                 .mode           = 0644,
1394                 .proc_handler   = proc_doulongvec_minmax,
1395         },
1396 #endif /* CONFIG_AIO */
1397 #ifdef CONFIG_INOTIFY_USER
1398         {
1399                 .procname       = "inotify",
1400                 .mode           = 0555,
1401                 .child          = inotify_table,
1402         },
1403 #endif
1404 #ifdef CONFIG_EPOLL
1405         {
1406                 .procname       = "epoll",
1407                 .mode           = 0555,
1408                 .child          = epoll_table,
1409         },
1410 #endif
1411 #endif
1412         {
1413                 .procname       = "suid_dumpable",
1414                 .data           = &suid_dumpable,
1415                 .maxlen         = sizeof(int),
1416                 .mode           = 0644,
1417                 .proc_handler   = proc_dointvec_minmax,
1418                 .extra1         = &zero,
1419                 .extra2         = &two,
1420         },
1421 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1422         {
1423                 .procname       = "binfmt_misc",
1424                 .mode           = 0555,
1425                 .child          = binfmt_misc_table,
1426         },
1427 #endif
1428 /*
1429  * NOTE: do not add new entries to this table unless you have read
1430  * Documentation/sysctl/ctl_unnumbered.txt
1431  */
1432         { }
1433 };
1434
1435 static struct ctl_table debug_table[] = {
1436 #if defined(CONFIG_X86) || defined(CONFIG_PPC) || defined(CONFIG_SPARC)
1437         {
1438                 .procname       = "exception-trace",
1439                 .data           = &show_unhandled_signals,
1440                 .maxlen         = sizeof(int),
1441                 .mode           = 0644,
1442                 .proc_handler   = proc_dointvec
1443         },
1444 #endif
1445 #if defined(CONFIG_OPTPROBES)
1446         {
1447                 .procname       = "kprobes-optimization",
1448                 .data           = &sysctl_kprobes_optimization,
1449                 .maxlen         = sizeof(int),
1450                 .mode           = 0644,
1451                 .proc_handler   = proc_kprobes_optimization_handler,
1452                 .extra1         = &zero,
1453                 .extra2         = &one,
1454         },
1455 #endif
1456         { }
1457 };
1458
1459 static struct ctl_table dev_table[] = {
1460         { }
1461 };
1462
1463 static DEFINE_SPINLOCK(sysctl_lock);
1464
1465 /* called under sysctl_lock */
1466 static int use_table(struct ctl_table_header *p)
1467 {
1468         if (unlikely(p->unregistering))
1469                 return 0;
1470         p->used++;
1471         return 1;
1472 }
1473
1474 /* called under sysctl_lock */
1475 static void unuse_table(struct ctl_table_header *p)
1476 {
1477         if (!--p->used)
1478                 if (unlikely(p->unregistering))
1479                         complete(p->unregistering);
1480 }
1481
1482 /* called under sysctl_lock, will reacquire if has to wait */
1483 static void start_unregistering(struct ctl_table_header *p)
1484 {
1485         /*
1486          * if p->used is 0, nobody will ever touch that entry again;
1487          * we'll eliminate all paths to it before dropping sysctl_lock
1488          */
1489         if (unlikely(p->used)) {
1490                 struct completion wait;
1491                 init_completion(&wait);
1492                 p->unregistering = &wait;
1493                 spin_unlock(&sysctl_lock);
1494                 wait_for_completion(&wait);
1495                 spin_lock(&sysctl_lock);
1496         } else {
1497                 /* anything non-NULL; we'll never dereference it */
1498                 p->unregistering = ERR_PTR(-EINVAL);
1499         }
1500         /*
1501          * do not remove from the list until nobody holds it; walking the
1502          * list in do_sysctl() relies on that.
1503          */
1504         list_del_init(&p->ctl_entry);
1505 }
1506
1507 void sysctl_head_get(struct ctl_table_header *head)
1508 {
1509         spin_lock(&sysctl_lock);
1510         head->count++;
1511         spin_unlock(&sysctl_lock);
1512 }
1513
1514 void sysctl_head_put(struct ctl_table_header *head)
1515 {
1516         spin_lock(&sysctl_lock);
1517         if (!--head->count)
1518                 kfree(head);
1519         spin_unlock(&sysctl_lock);
1520 }
1521
1522 struct ctl_table_header *sysctl_head_grab(struct ctl_table_header *head)
1523 {
1524         if (!head)
1525                 BUG();
1526         spin_lock(&sysctl_lock);
1527         if (!use_table(head))
1528                 head = ERR_PTR(-ENOENT);
1529         spin_unlock(&sysctl_lock);
1530         return head;
1531 }
1532
1533 void sysctl_head_finish(struct ctl_table_header *head)
1534 {
1535         if (!head)
1536                 return;
1537         spin_lock(&sysctl_lock);
1538         unuse_table(head);
1539         spin_unlock(&sysctl_lock);
1540 }
1541
1542 static struct ctl_table_set *
1543 lookup_header_set(struct ctl_table_root *root, struct nsproxy *namespaces)
1544 {
1545         struct ctl_table_set *set = &root->default_set;
1546         if (root->lookup)
1547                 set = root->lookup(root, namespaces);
1548         return set;
1549 }
1550
1551 static struct list_head *
1552 lookup_header_list(struct ctl_table_root *root, struct nsproxy *namespaces)
1553 {
1554         struct ctl_table_set *set = lookup_header_set(root, namespaces);
1555         return &set->list;
1556 }
1557
1558 struct ctl_table_header *__sysctl_head_next(struct nsproxy *namespaces,
1559                                             struct ctl_table_header *prev)
1560 {
1561         struct ctl_table_root *root;
1562         struct list_head *header_list;
1563         struct ctl_table_header *head;
1564         struct list_head *tmp;
1565
1566         spin_lock(&sysctl_lock);
1567         if (prev) {
1568                 head = prev;
1569                 tmp = &prev->ctl_entry;
1570                 unuse_table(prev);
1571                 goto next;
1572         }
1573         tmp = &root_table_header.ctl_entry;
1574         for (;;) {
1575                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1576
1577                 if (!use_table(head))
1578                         goto next;
1579                 spin_unlock(&sysctl_lock);
1580                 return head;
1581         next:
1582                 root = head->root;
1583                 tmp = tmp->next;
1584                 header_list = lookup_header_list(root, namespaces);
1585                 if (tmp != header_list)
1586                         continue;
1587
1588                 do {
1589                         root = list_entry(root->root_list.next,
1590                                         struct ctl_table_root, root_list);
1591                         if (root == &sysctl_table_root)
1592                                 goto out;
1593                         header_list = lookup_header_list(root, namespaces);
1594                 } while (list_empty(header_list));
1595                 tmp = header_list->next;
1596         }
1597 out:
1598         spin_unlock(&sysctl_lock);
1599         return NULL;
1600 }
1601
1602 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1603 {
1604         return __sysctl_head_next(current->nsproxy, prev);
1605 }
1606
1607 void register_sysctl_root(struct ctl_table_root *root)
1608 {
1609         spin_lock(&sysctl_lock);
1610         list_add_tail(&root->root_list, &sysctl_table_root.root_list);
1611         spin_unlock(&sysctl_lock);
1612 }
1613
1614 /*
1615  * sysctl_perm does NOT grant the superuser all rights automatically, because
1616  * some sysctl variables are readonly even to root.
1617  */
1618
1619 static int test_perm(int mode, int op)
1620 {
1621         if (!current_euid())
1622                 mode >>= 6;
1623         else if (in_egroup_p(0))
1624                 mode >>= 3;
1625         if ((op & ~mode & (MAY_READ|MAY_WRITE|MAY_EXEC)) == 0)
1626                 return 0;
1627         return -EACCES;
1628 }
1629
1630 int sysctl_perm(struct ctl_table_root *root, struct ctl_table *table, int op)
1631 {
1632         int error;
1633         int mode;
1634
1635         error = security_sysctl(table, op & (MAY_READ | MAY_WRITE | MAY_EXEC));
1636         if (error)
1637                 return error;
1638
1639         if (root->permissions)
1640                 mode = root->permissions(root, current->nsproxy, table);
1641         else
1642                 mode = table->mode;
1643
1644         return test_perm(mode, op);
1645 }
1646
1647 static void sysctl_set_parent(struct ctl_table *parent, struct ctl_table *table)
1648 {
1649         for (; table->procname; table++) {
1650                 table->parent = parent;
1651                 if (table->child)
1652                         sysctl_set_parent(table, table->child);
1653         }
1654 }
1655
1656 static __init int sysctl_init(void)
1657 {
1658         sysctl_set_parent(NULL, root_table);
1659 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1660         {
1661                 int err;
1662                 err = sysctl_check_table(current->nsproxy, root_table);
1663         }
1664 #endif
1665         return 0;
1666 }
1667
1668 core_initcall(sysctl_init);
1669
1670 static struct ctl_table *is_branch_in(struct ctl_table *branch,
1671                                       struct ctl_table *table)
1672 {
1673         struct ctl_table *p;
1674         const char *s = branch->procname;
1675
1676         /* branch should have named subdirectory as its first element */
1677         if (!s || !branch->child)
1678                 return NULL;
1679
1680         /* ... and nothing else */
1681         if (branch[1].procname)
1682                 return NULL;
1683
1684         /* table should contain subdirectory with the same name */
1685         for (p = table; p->procname; p++) {
1686                 if (!p->child)
1687                         continue;
1688                 if (p->procname && strcmp(p->procname, s) == 0)
1689                         return p;
1690         }
1691         return NULL;
1692 }
1693
1694 /* see if attaching q to p would be an improvement */
1695 static void try_attach(struct ctl_table_header *p, struct ctl_table_header *q)
1696 {
1697         struct ctl_table *to = p->ctl_table, *by = q->ctl_table;
1698         struct ctl_table *next;
1699         int is_better = 0;
1700         int not_in_parent = !p->attached_by;
1701
1702         while ((next = is_branch_in(by, to)) != NULL) {
1703                 if (by == q->attached_by)
1704                         is_better = 1;
1705                 if (to == p->attached_by)
1706                         not_in_parent = 1;
1707                 by = by->child;
1708                 to = next->child;
1709         }
1710
1711         if (is_better && not_in_parent) {
1712                 q->attached_by = by;
1713                 q->attached_to = to;
1714                 q->parent = p;
1715         }
1716 }
1717
1718 /**
1719  * __register_sysctl_paths - register a sysctl hierarchy
1720  * @root: List of sysctl headers to register on
1721  * @namespaces: Data to compute which lists of sysctl entries are visible
1722  * @path: The path to the directory the sysctl table is in.
1723  * @table: the top-level table structure
1724  *
1725  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1726  * array. A completely 0 filled entry terminates the table.
1727  *
1728  * The members of the &struct ctl_table structure are used as follows:
1729  *
1730  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1731  *            enter a sysctl file
1732  *
1733  * data - a pointer to data for use by proc_handler
1734  *
1735  * maxlen - the maximum size in bytes of the data
1736  *
1737  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1738  *
1739  * child - a pointer to the child sysctl table if this entry is a directory, or
1740  *         %NULL.
1741  *
1742  * proc_handler - the text handler routine (described below)
1743  *
1744  * de - for internal use by the sysctl routines
1745  *
1746  * extra1, extra2 - extra pointers usable by the proc handler routines
1747  *
1748  * Leaf nodes in the sysctl tree will be represented by a single file
1749  * under /proc; non-leaf nodes will be represented by directories.
1750  *
1751  * sysctl(2) can automatically manage read and write requests through
1752  * the sysctl table.  The data and maxlen fields of the ctl_table
1753  * struct enable minimal validation of the values being written to be
1754  * performed, and the mode field allows minimal authentication.
1755  *
1756  * There must be a proc_handler routine for any terminal nodes
1757  * mirrored under /proc/sys (non-terminals are handled by a built-in
1758  * directory handler).  Several default handlers are available to
1759  * cover common cases -
1760  *
1761  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1762  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1763  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1764  *
1765  * It is the handler's job to read the input buffer from user memory
1766  * and process it. The handler should return 0 on success.
1767  *
1768  * This routine returns %NULL on a failure to register, and a pointer
1769  * to the table header on success.
1770  */
1771 struct ctl_table_header *__register_sysctl_paths(
1772         struct ctl_table_root *root,
1773         struct nsproxy *namespaces,
1774         const struct ctl_path *path, struct ctl_table *table)
1775 {
1776         struct ctl_table_header *header;
1777         struct ctl_table *new, **prevp;
1778         unsigned int n, npath;
1779         struct ctl_table_set *set;
1780
1781         /* Count the path components */
1782         for (npath = 0; path[npath].procname; ++npath)
1783                 ;
1784
1785         /*
1786          * For each path component, allocate a 2-element ctl_table array.
1787          * The first array element will be filled with the sysctl entry
1788          * for this, the second will be the sentinel (procname == 0).
1789          *
1790          * We allocate everything in one go so that we don't have to
1791          * worry about freeing additional memory in unregister_sysctl_table.
1792          */
1793         header = kzalloc(sizeof(struct ctl_table_header) +
1794                          (2 * npath * sizeof(struct ctl_table)), GFP_KERNEL);
1795         if (!header)
1796                 return NULL;
1797
1798         new = (struct ctl_table *) (header + 1);
1799
1800         /* Now connect the dots */
1801         prevp = &header->ctl_table;
1802         for (n = 0; n < npath; ++n, ++path) {
1803                 /* Copy the procname */
1804                 new->procname = path->procname;
1805                 new->mode     = 0555;
1806
1807                 *prevp = new;
1808                 prevp = &new->child;
1809
1810                 new += 2;
1811         }
1812         *prevp = table;
1813         header->ctl_table_arg = table;
1814
1815         INIT_LIST_HEAD(&header->ctl_entry);
1816         header->used = 0;
1817         header->unregistering = NULL;
1818         header->root = root;
1819         sysctl_set_parent(NULL, header->ctl_table);
1820         header->count = 1;
1821 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1822         if (sysctl_check_table(namespaces, header->ctl_table)) {
1823                 kfree(header);
1824                 return NULL;
1825         }
1826 #endif
1827         spin_lock(&sysctl_lock);
1828         header->set = lookup_header_set(root, namespaces);
1829         header->attached_by = header->ctl_table;
1830         header->attached_to = root_table;
1831         header->parent = &root_table_header;
1832         for (set = header->set; set; set = set->parent) {
1833                 struct ctl_table_header *p;
1834                 list_for_each_entry(p, &set->list, ctl_entry) {
1835                         if (p->unregistering)
1836                                 continue;
1837                         try_attach(p, header);
1838                 }
1839         }
1840         header->parent->count++;
1841         list_add_tail(&header->ctl_entry, &header->set->list);
1842         spin_unlock(&sysctl_lock);
1843
1844         return header;
1845 }
1846
1847 /**
1848  * register_sysctl_table_path - register a sysctl table hierarchy
1849  * @path: The path to the directory the sysctl table is in.
1850  * @table: the top-level table structure
1851  *
1852  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1853  * array. A completely 0 filled entry terminates the table.
1854  *
1855  * See __register_sysctl_paths for more details.
1856  */
1857 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
1858                                                 struct ctl_table *table)
1859 {
1860         return __register_sysctl_paths(&sysctl_table_root, current->nsproxy,
1861                                         path, table);
1862 }
1863
1864 /**
1865  * register_sysctl_table - register a sysctl table hierarchy
1866  * @table: the top-level table structure
1867  *
1868  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1869  * array. A completely 0 filled entry terminates the table.
1870  *
1871  * See register_sysctl_paths for more details.
1872  */
1873 struct ctl_table_header *register_sysctl_table(struct ctl_table *table)
1874 {
1875         static const struct ctl_path null_path[] = { {} };
1876
1877         return register_sysctl_paths(null_path, table);
1878 }
1879
1880 /**
1881  * unregister_sysctl_table - unregister a sysctl table hierarchy
1882  * @header: the header returned from register_sysctl_table
1883  *
1884  * Unregisters the sysctl table and all children. proc entries may not
1885  * actually be removed until they are no longer used by anyone.
1886  */
1887 void unregister_sysctl_table(struct ctl_table_header * header)
1888 {
1889         might_sleep();
1890
1891         if (header == NULL)
1892                 return;
1893
1894         spin_lock(&sysctl_lock);
1895         start_unregistering(header);
1896         if (!--header->parent->count) {
1897                 WARN_ON(1);
1898                 kfree(header->parent);
1899         }
1900         if (!--header->count)
1901                 kfree(header);
1902         spin_unlock(&sysctl_lock);
1903 }
1904
1905 int sysctl_is_seen(struct ctl_table_header *p)
1906 {
1907         struct ctl_table_set *set = p->set;
1908         int res;
1909         spin_lock(&sysctl_lock);
1910         if (p->unregistering)
1911                 res = 0;
1912         else if (!set->is_seen)
1913                 res = 1;
1914         else
1915                 res = set->is_seen(set);
1916         spin_unlock(&sysctl_lock);
1917         return res;
1918 }
1919
1920 void setup_sysctl_set(struct ctl_table_set *p,
1921         struct ctl_table_set *parent,
1922         int (*is_seen)(struct ctl_table_set *))
1923 {
1924         INIT_LIST_HEAD(&p->list);
1925         p->parent = parent ? parent : &sysctl_table_root.default_set;
1926         p->is_seen = is_seen;
1927 }
1928
1929 #else /* !CONFIG_SYSCTL */
1930 struct ctl_table_header *register_sysctl_table(struct ctl_table * table)
1931 {
1932         return NULL;
1933 }
1934
1935 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
1936                                                     struct ctl_table *table)
1937 {
1938         return NULL;
1939 }
1940
1941 void unregister_sysctl_table(struct ctl_table_header * table)
1942 {
1943 }
1944
1945 void setup_sysctl_set(struct ctl_table_set *p,
1946         struct ctl_table_set *parent,
1947         int (*is_seen)(struct ctl_table_set *))
1948 {
1949 }
1950
1951 void sysctl_head_put(struct ctl_table_header *head)
1952 {
1953 }
1954
1955 #endif /* CONFIG_SYSCTL */
1956
1957 /*
1958  * /proc/sys support
1959  */
1960
1961 #ifdef CONFIG_PROC_SYSCTL
1962
1963 static int _proc_do_string(void* data, int maxlen, int write,
1964                            void __user *buffer,
1965                            size_t *lenp, loff_t *ppos)
1966 {
1967         size_t len;
1968         char __user *p;
1969         char c;
1970
1971         if (!data || !maxlen || !*lenp) {
1972                 *lenp = 0;
1973                 return 0;
1974         }
1975
1976         if (write) {
1977                 len = 0;
1978                 p = buffer;
1979                 while (len < *lenp) {
1980                         if (get_user(c, p++))
1981                                 return -EFAULT;
1982                         if (c == 0 || c == '\n')
1983                                 break;
1984                         len++;
1985                 }
1986                 if (len >= maxlen)
1987                         len = maxlen-1;
1988                 if(copy_from_user(data, buffer, len))
1989                         return -EFAULT;
1990                 ((char *) data)[len] = 0;
1991                 *ppos += *lenp;
1992         } else {
1993                 len = strlen(data);
1994                 if (len > maxlen)
1995                         len = maxlen;
1996
1997                 if (*ppos > len) {
1998                         *lenp = 0;
1999                         return 0;
2000                 }
2001
2002                 data += *ppos;
2003                 len  -= *ppos;
2004
2005                 if (len > *lenp)
2006                         len = *lenp;
2007                 if (len)
2008                         if(copy_to_user(buffer, data, len))
2009                                 return -EFAULT;
2010                 if (len < *lenp) {
2011                         if(put_user('\n', ((char __user *) buffer) + len))
2012                                 return -EFAULT;
2013                         len++;
2014                 }
2015                 *lenp = len;
2016                 *ppos += len;
2017         }
2018         return 0;
2019 }
2020
2021 /**
2022  * proc_dostring - read a string sysctl
2023  * @table: the sysctl table
2024  * @write: %TRUE if this is a write to the sysctl file
2025  * @buffer: the user buffer
2026  * @lenp: the size of the user buffer
2027  * @ppos: file position
2028  *
2029  * Reads/writes a string from/to the user buffer. If the kernel
2030  * buffer provided is not large enough to hold the string, the
2031  * string is truncated. The copied string is %NULL-terminated.
2032  * If the string is being read by the user process, it is copied
2033  * and a newline '\n' is added. It is truncated if the buffer is
2034  * not large enough.
2035  *
2036  * Returns 0 on success.
2037  */
2038 int proc_dostring(struct ctl_table *table, int write,
2039                   void __user *buffer, size_t *lenp, loff_t *ppos)
2040 {
2041         return _proc_do_string(table->data, table->maxlen, write,
2042                                buffer, lenp, ppos);
2043 }
2044
2045
2046 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
2047                                  int *valp,
2048                                  int write, void *data)
2049 {
2050         if (write) {
2051                 *valp = *negp ? -*lvalp : *lvalp;
2052         } else {
2053                 int val = *valp;
2054                 if (val < 0) {
2055                         *negp = -1;
2056                         *lvalp = (unsigned long)-val;
2057                 } else {
2058                         *negp = 0;
2059                         *lvalp = (unsigned long)val;
2060                 }
2061         }
2062         return 0;
2063 }
2064
2065 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2066                   int write, void __user *buffer,
2067                   size_t *lenp, loff_t *ppos,
2068                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2069                               int write, void *data),
2070                   void *data)
2071 {
2072 #define TMPBUFLEN 21
2073         int *i, vleft, first = 1, neg;
2074         unsigned long lval;
2075         size_t left, len;
2076
2077         char buf[TMPBUFLEN], *p;
2078         char __user *s = buffer;
2079
2080         if (!tbl_data || !table->maxlen || !*lenp ||
2081             (*ppos && !write)) {
2082                 *lenp = 0;
2083                 return 0;
2084         }
2085
2086         i = (int *) tbl_data;
2087         vleft = table->maxlen / sizeof(*i);
2088         left = *lenp;
2089
2090         if (!conv)
2091                 conv = do_proc_dointvec_conv;
2092
2093         for (; left && vleft--; i++, first=0) {
2094                 if (write) {
2095                         while (left) {
2096                                 char c;
2097                                 if (get_user(c, s))
2098                                         return -EFAULT;
2099                                 if (!isspace(c))
2100                                         break;
2101                                 left--;
2102                                 s++;
2103                         }
2104                         if (!left)
2105                                 break;
2106                         neg = 0;
2107                         len = left;
2108                         if (len > sizeof(buf) - 1)
2109                                 len = sizeof(buf) - 1;
2110                         if (copy_from_user(buf, s, len))
2111                                 return -EFAULT;
2112                         buf[len] = 0;
2113                         p = buf;
2114                         if (*p == '-' && left > 1) {
2115                                 neg = 1;
2116                                 p++;
2117                         }
2118                         if (*p < '0' || *p > '9')
2119                                 break;
2120
2121                         lval = simple_strtoul(p, &p, 0);
2122
2123                         len = p-buf;
2124                         if ((len < left) && *p && !isspace(*p))
2125                                 break;
2126                         s += len;
2127                         left -= len;
2128
2129                         if (conv(&neg, &lval, i, 1, data))
2130                                 break;
2131                 } else {
2132                         p = buf;
2133                         if (!first)
2134                                 *p++ = '\t';
2135
2136                         if (conv(&neg, &lval, i, 0, data))
2137                                 break;
2138
2139                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
2140                         len = strlen(buf);
2141                         if (len > left)
2142                                 len = left;
2143                         if(copy_to_user(s, buf, len))
2144                                 return -EFAULT;
2145                         left -= len;
2146                         s += len;
2147                 }
2148         }
2149
2150         if (!write && !first && left) {
2151                 if(put_user('\n', s))
2152                         return -EFAULT;
2153                 left--, s++;
2154         }
2155         if (write) {
2156                 while (left) {
2157                         char c;
2158                         if (get_user(c, s++))
2159                                 return -EFAULT;
2160                         if (!isspace(c))
2161                                 break;
2162                         left--;
2163                 }
2164         }
2165         if (write && first)
2166                 return -EINVAL;
2167         *lenp -= left;
2168         *ppos += *lenp;
2169         return 0;
2170 #undef TMPBUFLEN
2171 }
2172
2173 static int do_proc_dointvec(struct ctl_table *table, int write,
2174                   void __user *buffer, size_t *lenp, loff_t *ppos,
2175                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2176                               int write, void *data),
2177                   void *data)
2178 {
2179         return __do_proc_dointvec(table->data, table, write,
2180                         buffer, lenp, ppos, conv, data);
2181 }
2182
2183 /**
2184  * proc_dointvec - read a vector of integers
2185  * @table: the sysctl table
2186  * @write: %TRUE if this is a write to the sysctl file
2187  * @buffer: the user buffer
2188  * @lenp: the size of the user buffer
2189  * @ppos: file position
2190  *
2191  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2192  * values from/to the user buffer, treated as an ASCII string.
2193  *
2194  * Returns 0 on success.
2195  */
2196 int proc_dointvec(struct ctl_table *table, int write,
2197                      void __user *buffer, size_t *lenp, loff_t *ppos)
2198 {
2199     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2200                             NULL,NULL);
2201 }
2202
2203 /*
2204  * Taint values can only be increased
2205  * This means we can safely use a temporary.
2206  */
2207 static int proc_taint(struct ctl_table *table, int write,
2208                                void __user *buffer, size_t *lenp, loff_t *ppos)
2209 {
2210         struct ctl_table t;
2211         unsigned long tmptaint = get_taint();
2212         int err;
2213
2214         if (write && !capable(CAP_SYS_ADMIN))
2215                 return -EPERM;
2216
2217         t = *table;
2218         t.data = &tmptaint;
2219         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2220         if (err < 0)
2221                 return err;
2222
2223         if (write) {
2224                 /*
2225                  * Poor man's atomic or. Not worth adding a primitive
2226                  * to everyone's atomic.h for this
2227                  */
2228                 int i;
2229                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2230                         if ((tmptaint >> i) & 1)
2231                                 add_taint(i);
2232                 }
2233         }
2234
2235         return err;
2236 }
2237
2238 struct do_proc_dointvec_minmax_conv_param {
2239         int *min;
2240         int *max;
2241 };
2242
2243 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
2244                                         int *valp,
2245                                         int write, void *data)
2246 {
2247         struct do_proc_dointvec_minmax_conv_param *param = data;
2248         if (write) {
2249                 int val = *negp ? -*lvalp : *lvalp;
2250                 if ((param->min && *param->min > val) ||
2251                     (param->max && *param->max < val))
2252                         return -EINVAL;
2253                 *valp = val;
2254         } else {
2255                 int val = *valp;
2256                 if (val < 0) {
2257                         *negp = -1;
2258                         *lvalp = (unsigned long)-val;
2259                 } else {
2260                         *negp = 0;
2261                         *lvalp = (unsigned long)val;
2262                 }
2263         }
2264         return 0;
2265 }
2266
2267 /**
2268  * proc_dointvec_minmax - read a vector of integers with min/max values
2269  * @table: the sysctl table
2270  * @write: %TRUE if this is a write to the sysctl file
2271  * @buffer: the user buffer
2272  * @lenp: the size of the user buffer
2273  * @ppos: file position
2274  *
2275  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2276  * values from/to the user buffer, treated as an ASCII string.
2277  *
2278  * This routine will ensure the values are within the range specified by
2279  * table->extra1 (min) and table->extra2 (max).
2280  *
2281  * Returns 0 on success.
2282  */
2283 int proc_dointvec_minmax(struct ctl_table *table, int write,
2284                   void __user *buffer, size_t *lenp, loff_t *ppos)
2285 {
2286         struct do_proc_dointvec_minmax_conv_param param = {
2287                 .min = (int *) table->extra1,
2288                 .max = (int *) table->extra2,
2289         };
2290         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2291                                 do_proc_dointvec_minmax_conv, &param);
2292 }
2293
2294 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2295                                      void __user *buffer,
2296                                      size_t *lenp, loff_t *ppos,
2297                                      unsigned long convmul,
2298                                      unsigned long convdiv)
2299 {
2300 #define TMPBUFLEN 21
2301         unsigned long *i, *min, *max, val;
2302         int vleft, first=1, neg;
2303         size_t len, left;
2304         char buf[TMPBUFLEN], *p;
2305         char __user *s = buffer;
2306
2307         if (!data || !table->maxlen || !*lenp ||
2308             (*ppos && !write)) {
2309                 *lenp = 0;
2310                 return 0;
2311         }
2312
2313         i = (unsigned long *) data;
2314         min = (unsigned long *) table->extra1;
2315         max = (unsigned long *) table->extra2;
2316         vleft = table->maxlen / sizeof(unsigned long);
2317         left = *lenp;
2318
2319         for (; left && vleft--; i++, min++, max++, first=0) {
2320                 if (write) {
2321                         while (left) {
2322                                 char c;
2323                                 if (get_user(c, s))
2324                                         return -EFAULT;
2325                                 if (!isspace(c))
2326                                         break;
2327                                 left--;
2328                                 s++;
2329                         }
2330                         if (!left)
2331                                 break;
2332                         neg = 0;
2333                         len = left;
2334                         if (len > TMPBUFLEN-1)
2335                                 len = TMPBUFLEN-1;
2336                         if (copy_from_user(buf, s, len))
2337                                 return -EFAULT;
2338                         buf[len] = 0;
2339                         p = buf;
2340                         if (*p == '-' && left > 1) {
2341                                 neg = 1;
2342                                 p++;
2343                         }
2344                         if (*p < '0' || *p > '9')
2345                                 break;
2346                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
2347                         len = p-buf;
2348                         if ((len < left) && *p && !isspace(*p))
2349                                 break;
2350                         if (neg)
2351                                 val = -val;
2352                         s += len;
2353                         left -= len;
2354
2355                         if(neg)
2356                                 continue;
2357                         if ((min && val < *min) || (max && val > *max))
2358                                 continue;
2359                         *i = val;
2360                 } else {
2361                         p = buf;
2362                         if (!first)
2363                                 *p++ = '\t';
2364                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2365                         len = strlen(buf);
2366                         if (len > left)
2367                                 len = left;
2368                         if(copy_to_user(s, buf, len))
2369                                 return -EFAULT;
2370                         left -= len;
2371                         s += len;
2372                 }
2373         }
2374
2375         if (!write && !first && left) {
2376                 if(put_user('\n', s))
2377                         return -EFAULT;
2378                 left--, s++;
2379         }
2380         if (write) {
2381                 while (left) {
2382                         char c;
2383                         if (get_user(c, s++))
2384                                 return -EFAULT;
2385                         if (!isspace(c))
2386                                 break;
2387                         left--;
2388                 }
2389         }
2390         if (write && first)
2391                 return -EINVAL;
2392         *lenp -= left;
2393         *ppos += *lenp;
2394         return 0;
2395 #undef TMPBUFLEN
2396 }
2397
2398 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2399                                      void __user *buffer,
2400                                      size_t *lenp, loff_t *ppos,
2401                                      unsigned long convmul,
2402                                      unsigned long convdiv)
2403 {
2404         return __do_proc_doulongvec_minmax(table->data, table, write,
2405                         buffer, lenp, ppos, convmul, convdiv);
2406 }
2407
2408 /**
2409  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2410  * @table: the sysctl table
2411  * @write: %TRUE if this is a write to the sysctl file
2412  * @buffer: the user buffer
2413  * @lenp: the size of the user buffer
2414  * @ppos: file position
2415  *
2416  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2417  * values from/to the user buffer, treated as an ASCII string.
2418  *
2419  * This routine will ensure the values are within the range specified by
2420  * table->extra1 (min) and table->extra2 (max).
2421  *
2422  * Returns 0 on success.
2423  */
2424 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2425                            void __user *buffer, size_t *lenp, loff_t *ppos)
2426 {
2427     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2428 }
2429
2430 /**
2431  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2432  * @table: the sysctl table
2433  * @write: %TRUE if this is a write to the sysctl file
2434  * @buffer: the user buffer
2435  * @lenp: the size of the user buffer
2436  * @ppos: file position
2437  *
2438  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2439  * values from/to the user buffer, treated as an ASCII string. The values
2440  * are treated as milliseconds, and converted to jiffies when they are stored.
2441  *
2442  * This routine will ensure the values are within the range specified by
2443  * table->extra1 (min) and table->extra2 (max).
2444  *
2445  * Returns 0 on success.
2446  */
2447 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2448                                       void __user *buffer,
2449                                       size_t *lenp, loff_t *ppos)
2450 {
2451     return do_proc_doulongvec_minmax(table, write, buffer,
2452                                      lenp, ppos, HZ, 1000l);
2453 }
2454
2455
2456 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2457                                          int *valp,
2458                                          int write, void *data)
2459 {
2460         if (write) {
2461                 if (*lvalp > LONG_MAX / HZ)
2462                         return 1;
2463                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2464         } else {
2465                 int val = *valp;
2466                 unsigned long lval;
2467                 if (val < 0) {
2468                         *negp = -1;
2469                         lval = (unsigned long)-val;
2470                 } else {
2471                         *negp = 0;
2472                         lval = (unsigned long)val;
2473                 }
2474                 *lvalp = lval / HZ;
2475         }
2476         return 0;
2477 }
2478
2479 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2480                                                 int *valp,
2481                                                 int write, void *data)
2482 {
2483         if (write) {
2484                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2485                         return 1;
2486                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2487         } else {
2488                 int val = *valp;
2489                 unsigned long lval;
2490                 if (val < 0) {
2491                         *negp = -1;
2492                         lval = (unsigned long)-val;
2493                 } else {
2494                         *negp = 0;
2495                         lval = (unsigned long)val;
2496                 }
2497                 *lvalp = jiffies_to_clock_t(lval);
2498         }
2499         return 0;
2500 }
2501
2502 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2503                                             int *valp,
2504                                             int write, void *data)
2505 {
2506         if (write) {
2507                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2508         } else {
2509                 int val = *valp;
2510                 unsigned long lval;
2511                 if (val < 0) {
2512                         *negp = -1;
2513                         lval = (unsigned long)-val;
2514                 } else {
2515                         *negp = 0;
2516                         lval = (unsigned long)val;
2517                 }
2518                 *lvalp = jiffies_to_msecs(lval);
2519         }
2520         return 0;
2521 }
2522
2523 /**
2524  * proc_dointvec_jiffies - read a vector of integers as seconds
2525  * @table: the sysctl table
2526  * @write: %TRUE if this is a write to the sysctl file
2527  * @buffer: the user buffer
2528  * @lenp: the size of the user buffer
2529  * @ppos: file position
2530  *
2531  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2532  * values from/to the user buffer, treated as an ASCII string.
2533  * The values read are assumed to be in seconds, and are converted into
2534  * jiffies.
2535  *
2536  * Returns 0 on success.
2537  */
2538 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2539                           void __user *buffer, size_t *lenp, loff_t *ppos)
2540 {
2541     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2542                             do_proc_dointvec_jiffies_conv,NULL);
2543 }
2544
2545 /**
2546  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2547  * @table: the sysctl table
2548  * @write: %TRUE if this is a write to the sysctl file
2549  * @buffer: the user buffer
2550  * @lenp: the size of the user buffer
2551  * @ppos: pointer to the file position
2552  *
2553  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2554  * values from/to the user buffer, treated as an ASCII string.
2555  * The values read are assumed to be in 1/USER_HZ seconds, and
2556  * are converted into jiffies.
2557  *
2558  * Returns 0 on success.
2559  */
2560 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2561                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2562 {
2563     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2564                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2565 }
2566
2567 /**
2568  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2569  * @table: the sysctl table
2570  * @write: %TRUE if this is a write to the sysctl file
2571  * @buffer: the user buffer
2572  * @lenp: the size of the user buffer
2573  * @ppos: file position
2574  * @ppos: the current position in the file
2575  *
2576  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2577  * values from/to the user buffer, treated as an ASCII string.
2578  * The values read are assumed to be in 1/1000 seconds, and
2579  * are converted into jiffies.
2580  *
2581  * Returns 0 on success.
2582  */
2583 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2584                              void __user *buffer, size_t *lenp, loff_t *ppos)
2585 {
2586         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2587                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2588 }
2589
2590 static int proc_do_cad_pid(struct ctl_table *table, int write,
2591                            void __user *buffer, size_t *lenp, loff_t *ppos)
2592 {
2593         struct pid *new_pid;
2594         pid_t tmp;
2595         int r;
2596
2597         tmp = pid_vnr(cad_pid);
2598
2599         r = __do_proc_dointvec(&tmp, table, write, buffer,
2600                                lenp, ppos, NULL, NULL);
2601         if (r || !write)
2602                 return r;
2603
2604         new_pid = find_get_pid(tmp);
2605         if (!new_pid)
2606                 return -ESRCH;
2607
2608         put_pid(xchg(&cad_pid, new_pid));
2609         return 0;
2610 }
2611
2612 #else /* CONFIG_PROC_FS */
2613
2614 int proc_dostring(struct ctl_table *table, int write,
2615                   void __user *buffer, size_t *lenp, loff_t *ppos)
2616 {
2617         return -ENOSYS;
2618 }
2619
2620 int proc_dointvec(struct ctl_table *table, int write,
2621                   void __user *buffer, size_t *lenp, loff_t *ppos)
2622 {
2623         return -ENOSYS;
2624 }
2625
2626 int proc_dointvec_minmax(struct ctl_table *table, int write,
2627                     void __user *buffer, size_t *lenp, loff_t *ppos)
2628 {
2629         return -ENOSYS;
2630 }
2631
2632 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2633                     void __user *buffer, size_t *lenp, loff_t *ppos)
2634 {
2635         return -ENOSYS;
2636 }
2637
2638 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2639                     void __user *buffer, size_t *lenp, loff_t *ppos)
2640 {
2641         return -ENOSYS;
2642 }
2643
2644 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2645                              void __user *buffer, size_t *lenp, loff_t *ppos)
2646 {
2647         return -ENOSYS;
2648 }
2649
2650 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2651                     void __user *buffer, size_t *lenp, loff_t *ppos)
2652 {
2653         return -ENOSYS;
2654 }
2655
2656 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2657                                       void __user *buffer,
2658                                       size_t *lenp, loff_t *ppos)
2659 {
2660     return -ENOSYS;
2661 }
2662
2663
2664 #endif /* CONFIG_PROC_FS */
2665
2666 /*
2667  * No sense putting this after each symbol definition, twice,
2668  * exception granted :-)
2669  */
2670 EXPORT_SYMBOL(proc_dointvec);
2671 EXPORT_SYMBOL(proc_dointvec_jiffies);
2672 EXPORT_SYMBOL(proc_dointvec_minmax);
2673 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2674 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2675 EXPORT_SYMBOL(proc_dostring);
2676 EXPORT_SYMBOL(proc_doulongvec_minmax);
2677 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2678 EXPORT_SYMBOL(register_sysctl_table);
2679 EXPORT_SYMBOL(register_sysctl_paths);
2680 EXPORT_SYMBOL(unregister_sysctl_table);