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_BSD_PROCESS_ACCT
  65 #include <linux/acct.h>
  66 #endif
  67 #ifdef CONFIG_RT_MUTEXES
  68 #include <linux/rtmutex.h>
  69 #endif
  70 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  71 #include <linux/lockdep.h>
  72 #endif
  73 #ifdef CONFIG_CHR_DEV_SG
  74 #include <scsi/sg.h>
  75 #endif
  76
  77 #ifdef CONFIG_NMI_WATCHDOG
  78 #include <linux/nmi.h>
  79 #endif
  80
  81
  82 #if defined(CONFIG_SYSCTL)
  83
  84 /* External variables not in a header file. */
  85 extern int sysctl_overcommit_memory;
  86 extern int sysctl_overcommit_ratio;
  87 extern int sysctl_panic_on_oom;
  88 extern int sysctl_oom_kill_allocating_task;
  89 extern int sysctl_oom_dump_tasks;
  90 extern int max_threads;
  91 extern int core_uses_pid;
  92 extern int suid_dumpable;
  93 extern char core_pattern[];
  94 extern unsigned int core_pipe_limit;
  95 extern int pid_max;
  96 extern int min_free_kbytes;
  97 extern int pid_max_min, pid_max_max;
  98 extern int sysctl_drop_caches;
  99 extern int percpu_pagelist_fraction;
 100 extern int compat_log;
 101 extern int latencytop_enabled;
 102 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 103 #ifndef CONFIG_MMU
 104 extern int sysctl_nr_trim_pages;
 105 #endif
 106 #ifdef CONFIG_BLOCK
 107 extern int blk_iopoll_enabled;
 108 #endif
 109
 110 /* Constants used for minimum and  maximum */
 111 #ifdef CONFIG_DETECT_SOFTLOCKUP
 112 static int sixty = 60;
 113 static int neg_one = -1;
 114 #endif
 115
 116 static int zero;
 117 static int __maybe_unused one = 1;
 118 static int __maybe_unused two = 2;
 119 static unsigned long one_ul = 1;
 120 static int one_hundred = 100;
 121 #ifdef CONFIG_PRINTK
 122 static int ten_thousand = 10000;
 123 #endif
 124
 125 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 126 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 127
 128 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 129 static int maxolduid = 65535;
 130 static int minolduid;
 131 static int min_percpu_pagelist_fract = 8;
 132
 133 static int ngroups_max = NGROUPS_MAX;
 134
 135 #ifdef CONFIG_SPARC
 136 #include <asm/system.h>
 137 #endif
 138
 139 #ifdef CONFIG_SPARC64
 140 extern int sysctl_tsb_ratio;
 141 #endif
 142
 143 #ifdef __hppa__
 144 extern int pwrsw_enabled;
 145 extern int unaligned_enabled;
 146 #endif
 147
 148 #ifdef CONFIG_S390
 149 #ifdef CONFIG_MATHEMU
 150 extern int sysctl_ieee_emulation_warnings;
 151 #endif
 152 extern int sysctl_userprocess_debug;
 153 extern int spin_retry;
 154 #endif
 155
 156 #ifdef CONFIG_IA64
 157 extern int no_unaligned_warning;
 158 extern int unaligned_dump_stack;
 159 #endif
 160
 161 extern struct ratelimit_state printk_ratelimit_state;
 162
 163 #ifdef CONFIG_PROC_SYSCTL
 164 static int proc_do_cad_pid(struct ctl_table *table, int write,
 165                   void __user *buffer, size_t *lenp, loff_t *ppos);
 166 static int proc_taint(struct ctl_table *table, int write,
 167                                void __user *buffer, size_t *lenp, loff_t *ppos);
 168 #endif
 169
 170 static struct ctl_table root_table[];
 171 static struct ctl_table_root sysctl_table_root;
 172 static struct ctl_table_header root_table_header = {
 173         .count = 1,
 174         .ctl_table = root_table,
 175         .ctl_entry = LIST_HEAD_INIT(sysctl_table_root.default_set.list),
 176         .root = &sysctl_table_root,
 177         .set = &sysctl_table_root.default_set,
 178 };
 179 static struct ctl_table_root sysctl_table_root = {
 180         .root_list = LIST_HEAD_INIT(sysctl_table_root.root_list),
 181         .default_set.list = LIST_HEAD_INIT(root_table_header.ctl_entry),
 182 };
 183
 184 static struct ctl_table kern_table[];
 185 static struct ctl_table vm_table[];
 186 static struct ctl_table fs_table[];
 187 static struct ctl_table debug_table[];
 188 static struct ctl_table dev_table[];
 189 extern struct ctl_table random_table[];
 190 #ifdef CONFIG_INOTIFY_USER
 191 extern struct ctl_table inotify_table[];
 192 #endif
 193 #ifdef CONFIG_EPOLL
 194 extern struct ctl_table epoll_table[];
 195 #endif
 196
 197 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 198 int sysctl_legacy_va_layout;
 199 #endif
 200
 201 /* The default sysctl tables: */
 202
 203 static struct ctl_table root_table[] = {
 204         {
 205                 .procname       = "kernel",
 206                 .mode           = 0555,
 207                 .child          = kern_table,
 208         },
 209         {
 210                 .procname       = "vm",
 211                 .mode           = 0555,
 212                 .child          = vm_table,
 213         },
 214         {
 215                 .procname       = "fs",
 216                 .mode           = 0555,
 217                 .child          = fs_table,
 218         },
 219         {
 220                 .procname       = "debug",
 221                 .mode           = 0555,
 222                 .child          = debug_table,
 223         },
 224         {
 225                 .procname       = "dev",
 226                 .mode           = 0555,
 227                 .child          = dev_table,
 228         },
 229 /*
 230  * NOTE: do not add new entries to this table unless you have read
 231  * Documentation/sysctl/ctl_unnumbered.txt
 232  */
 233         { }
 234 };
 235
 236 #ifdef CONFIG_SCHED_DEBUG
 237 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 238 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 239 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 240 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 241 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 242 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 243 static int min_sched_shares_ratelimit = 100000; /* 100 usec */
 244 static int max_sched_shares_ratelimit = NSEC_PER_SEC; /* 1 second */
 245 #endif
 246
 247 static struct ctl_table kern_table[] = {
 248         {
 249                 .procname       = "sched_child_runs_first",
 250                 .data           = &sysctl_sched_child_runs_first,
 251                 .maxlen         = sizeof(unsigned int),
 252                 .mode           = 0644,
 253                 .proc_handler   = proc_dointvec,
 254         },
 255 #ifdef CONFIG_SCHED_DEBUG
 256         {
 257                 .procname       = "sched_min_granularity_ns",
 258                 .data           = &sysctl_sched_min_granularity,
 259                 .maxlen         = sizeof(unsigned int),
 260                 .mode           = 0644,
 261                 .proc_handler   = sched_proc_update_handler,
 262                 .extra1         = &min_sched_granularity_ns,
 263                 .extra2         = &max_sched_granularity_ns,
 264         },
 265         {
 266                 .procname       = "sched_latency_ns",
 267                 .data           = &sysctl_sched_latency,
 268                 .maxlen         = sizeof(unsigned int),
 269                 .mode           = 0644,
 270                 .proc_handler   = sched_proc_update_handler,
 271                 .extra1         = &min_sched_granularity_ns,
 272                 .extra2         = &max_sched_granularity_ns,
 273         },
 274         {
 275                 .procname       = "sched_wakeup_granularity_ns",
 276                 .data           = &sysctl_sched_wakeup_granularity,
 277                 .maxlen         = sizeof(unsigned int),
 278                 .mode           = 0644,
 279                 .proc_handler   = sched_proc_update_handler,
 280                 .extra1         = &min_wakeup_granularity_ns,
 281                 .extra2         = &max_wakeup_granularity_ns,
 282         },
 283         {
 284                 .procname       = "sched_shares_ratelimit",
 285                 .data           = &sysctl_sched_shares_ratelimit,
 286                 .maxlen         = sizeof(unsigned int),
 287                 .mode           = 0644,
 288                 .proc_handler   = sched_proc_update_handler,
 289                 .extra1         = &min_sched_shares_ratelimit,
 290                 .extra2         = &max_sched_shares_ratelimit,
 291         },
 292         {
 293                 .procname       = "sched_tunable_scaling",
 294                 .data           = &sysctl_sched_tunable_scaling,
 295                 .maxlen         = sizeof(enum sched_tunable_scaling),
 296                 .mode           = 0644,
 297                 .proc_handler   = sched_proc_update_handler,
 298                 .extra1         = &min_sched_tunable_scaling,
 299                 .extra2         = &max_sched_tunable_scaling,
 300         },
 301         {
 302                 .procname       = "sched_shares_thresh",
 303                 .data           = &sysctl_sched_shares_thresh,
 304                 .maxlen         = sizeof(unsigned int),
 305                 .mode           = 0644,
 306                 .proc_handler   = proc_dointvec_minmax,
 307                 .extra1         = &zero,
 308         },
 309         {
 310                 .procname       = "sched_migration_cost",
 311                 .data           = &sysctl_sched_migration_cost,
 312                 .maxlen         = sizeof(unsigned int),
 313                 .mode           = 0644,
 314                 .proc_handler   = proc_dointvec,
 315         },
 316         {
 317                 .procname       = "sched_nr_migrate",
 318                 .data           = &sysctl_sched_nr_migrate,
 319                 .maxlen         = sizeof(unsigned int),
 320                 .mode           = 0644,
 321                 .proc_handler   = proc_dointvec,
 322         },
 323         {
 324                 .procname       = "sched_time_avg",
 325                 .data           = &sysctl_sched_time_avg,
 326                 .maxlen         = sizeof(unsigned int),
 327                 .mode           = 0644,
 328                 .proc_handler   = proc_dointvec,
 329         },
 330         {
 331                 .procname       = "timer_migration",
 332                 .data           = &sysctl_timer_migration,
 333                 .maxlen         = sizeof(unsigned int),
 334                 .mode           = 0644,
 335                 .proc_handler   = proc_dointvec_minmax,
 336                 .extra1         = &zero,
 337                 .extra2         = &one,
 338         },
 339 #endif
 340         {
 341                 .procname       = "sched_rt_period_us",
 342                 .data           = &sysctl_sched_rt_period,
 343                 .maxlen         = sizeof(unsigned int),
 344                 .mode           = 0644,
 345                 .proc_handler   = sched_rt_handler,
 346         },
 347         {
 348                 .procname       = "sched_rt_runtime_us",
 349                 .data           = &sysctl_sched_rt_runtime,
 350                 .maxlen         = sizeof(int),
 351                 .mode           = 0644,
 352                 .proc_handler   = sched_rt_handler,
 353         },
 354         {
 355                 .procname       = "sched_compat_yield",
 356                 .data           = &sysctl_sched_compat_yield,
 357                 .maxlen         = sizeof(unsigned int),
 358                 .mode           = 0644,
 359                 .proc_handler   = proc_dointvec,
 360         },
 361 #ifdef CONFIG_PROVE_LOCKING
 362         {
 363                 .procname       = "prove_locking",
 364                 .data           = &prove_locking,
 365                 .maxlen         = sizeof(int),
 366                 .mode           = 0644,
 367                 .proc_handler   = proc_dointvec,
 368         },
 369 #endif
 370 #ifdef CONFIG_LOCK_STAT
 371         {
 372                 .procname       = "lock_stat",
 373                 .data           = &lock_stat,
 374                 .maxlen         = sizeof(int),
 375                 .mode           = 0644,
 376                 .proc_handler   = proc_dointvec,
 377         },
 378 #endif
 379         {
 380                 .procname       = "panic",
 381                 .data           = &panic_timeout,
 382                 .maxlen         = sizeof(int),
 383                 .mode           = 0644,
 384                 .proc_handler   = proc_dointvec,
 385         },
 386         {
 387                 .procname       = "core_uses_pid",
 388                 .data           = &core_uses_pid,
 389                 .maxlen         = sizeof(int),
 390                 .mode           = 0644,
 391                 .proc_handler   = proc_dointvec,
 392         },
 393         {
 394                 .procname       = "core_pattern",
 395                 .data           = core_pattern,
 396                 .maxlen         = CORENAME_MAX_SIZE,
 397                 .mode           = 0644,
 398                 .proc_handler   = proc_dostring,
 399         },
 400         {
 401                 .procname       = "core_pipe_limit",
 402                 .data           = &core_pipe_limit,
 403                 .maxlen         = sizeof(unsigned int),
 404                 .mode           = 0644,
 405                 .proc_handler   = proc_dointvec,
 406         },
 407 #ifdef CONFIG_PROC_SYSCTL
 408         {
 409                 .procname       = "tainted",
 410                 .maxlen         = sizeof(long),
 411                 .mode           = 0644,
 412                 .proc_handler   = proc_taint,
 413         },
 414 #endif
 415 #ifdef CONFIG_LATENCYTOP
 416         {
 417                 .procname       = "latencytop",
 418                 .data           = &latencytop_enabled,
 419                 .maxlen         = sizeof(int),
 420                 .mode           = 0644,
 421                 .proc_handler   = proc_dointvec,
 422         },
 423 #endif
 424 #ifdef CONFIG_BLK_DEV_INITRD
 425         {
 426                 .procname       = "real-root-dev",
 427                 .data           = &real_root_dev,
 428                 .maxlen         = sizeof(int),
 429                 .mode           = 0644,
 430                 .proc_handler   = proc_dointvec,
 431         },
 432 #endif
 433         {
 434                 .procname       = "print-fatal-signals",
 435                 .data           = &print_fatal_signals,
 436                 .maxlen         = sizeof(int),
 437                 .mode           = 0644,
 438                 .proc_handler   = proc_dointvec,
 439         },
 440 #ifdef CONFIG_SPARC
 441         {
 442                 .procname       = "reboot-cmd",
 443                 .data           = reboot_command,
 444                 .maxlen         = 256,
 445                 .mode           = 0644,
 446                 .proc_handler   = proc_dostring,
 447         },
 448         {
 449                 .procname       = "stop-a",
 450                 .data           = &stop_a_enabled,
 451                 .maxlen         = sizeof (int),
 452                 .mode           = 0644,
 453                 .proc_handler   = proc_dointvec,
 454         },
 455         {
 456                 .procname       = "scons-poweroff",
 457                 .data           = &scons_pwroff,
 458                 .maxlen         = sizeof (int),
 459                 .mode           = 0644,
 460                 .proc_handler   = proc_dointvec,
 461         },
 462 #endif
 463 #ifdef CONFIG_SPARC64
 464         {
 465                 .procname       = "tsb-ratio",
 466                 .data           = &sysctl_tsb_ratio,
 467                 .maxlen         = sizeof (int),
 468                 .mode           = 0644,
 469                 .proc_handler   = proc_dointvec,
 470         },
 471 #endif
 472 #ifdef __hppa__
 473         {
 474                 .procname       = "soft-power",
 475                 .data           = &pwrsw_enabled,
 476                 .maxlen         = sizeof (int),
 477                 .mode           = 0644,
 478                 .proc_handler   = proc_dointvec,
 479         },
 480         {
 481                 .procname       = "unaligned-trap",
 482                 .data           = &unaligned_enabled,
 483                 .maxlen         = sizeof (int),
 484                 .mode           = 0644,
 485                 .proc_handler   = proc_dointvec,
 486         },
 487 #endif
 488         {
 489                 .procname       = "ctrl-alt-del",
 490                 .data           = &C_A_D,
 491                 .maxlen         = sizeof(int),
 492                 .mode           = 0644,
 493                 .proc_handler   = proc_dointvec,
 494         },
 495 #ifdef CONFIG_FUNCTION_TRACER
 496         {
 497                 .procname       = "ftrace_enabled",
 498                 .data           = &ftrace_enabled,
 499                 .maxlen         = sizeof(int),
 500                 .mode           = 0644,
 501                 .proc_handler   = ftrace_enable_sysctl,
 502         },
 503 #endif
 504 #ifdef CONFIG_STACK_TRACER
 505         {
 506                 .procname       = "stack_tracer_enabled",
 507                 .data           = &stack_tracer_enabled,
 508                 .maxlen         = sizeof(int),
 509                 .mode           = 0644,
 510                 .proc_handler   = stack_trace_sysctl,
 511         },
 512 #endif
 513 #ifdef CONFIG_TRACING
 514         {
 515                 .procname       = "ftrace_dump_on_oops",
 516                 .data           = &ftrace_dump_on_oops,
 517                 .maxlen         = sizeof(int),
 518                 .mode           = 0644,
 519                 .proc_handler   = proc_dointvec,
 520         },
 521 #endif
 522 #ifdef CONFIG_MODULES
 523         {
 524                 .procname       = "modprobe",
 525                 .data           = &modprobe_path,
 526                 .maxlen         = KMOD_PATH_LEN,
 527                 .mode           = 0644,
 528                 .proc_handler   = proc_dostring,
 529         },
 530         {
 531                 .procname       = "modules_disabled",
 532                 .data           = &modules_disabled,
 533                 .maxlen         = sizeof(int),
 534                 .mode           = 0644,
 535                 /* only handle a transition from default "0" to "1" */
 536                 .proc_handler   = proc_dointvec_minmax,
 537                 .extra1         = &one,
 538                 .extra2         = &one,
 539         },
 540 #endif
 541 #if defined(CONFIG_HOTPLUG) && defined(CONFIG_NET)
 542         {
 543                 .procname       = "hotplug",
 544                 .data           = &uevent_helper,
 545                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 546                 .mode           = 0644,
 547                 .proc_handler   = proc_dostring,
 548         },
 549 #endif
 550 #ifdef CONFIG_CHR_DEV_SG
 551         {
 552                 .procname       = "sg-big-buff",
 553                 .data           = &sg_big_buff,
 554                 .maxlen         = sizeof (int),
 555                 .mode           = 0444,
 556                 .proc_handler   = proc_dointvec,
 557         },
 558 #endif
 559 #ifdef CONFIG_BSD_PROCESS_ACCT
 560         {
 561                 .procname       = "acct",
 562                 .data           = &acct_parm,
 563                 .maxlen         = 3*sizeof(int),
 564                 .mode           = 0644,
 565                 .proc_handler   = proc_dointvec,
 566         },
 567 #endif
 568 #ifdef CONFIG_MAGIC_SYSRQ
 569         {
 570                 .procname       = "sysrq",
 571                 .data           = &__sysrq_enabled,
 572                 .maxlen         = sizeof (int),
 573                 .mode           = 0644,
 574                 .proc_handler   = proc_dointvec,
 575         },
 576 #endif
 577 #ifdef CONFIG_PROC_SYSCTL
 578         {
 579                 .procname       = "cad_pid",
 580                 .data           = NULL,
 581                 .maxlen         = sizeof (int),
 582                 .mode           = 0600,
 583                 .proc_handler   = proc_do_cad_pid,
 584         },
 585 #endif
 586         {
 587                 .procname       = "threads-max",
 588                 .data           = &max_threads,
 589                 .maxlen         = sizeof(int),
 590                 .mode           = 0644,
 591                 .proc_handler   = proc_dointvec,
 592         },
 593         {
 594                 .procname       = "random",
 595                 .mode           = 0555,
 596                 .child          = random_table,
 597         },
 598         {
 599                 .procname       = "overflowuid",
 600                 .data           = &overflowuid,
 601                 .maxlen         = sizeof(int),
 602                 .mode           = 0644,
 603                 .proc_handler   = proc_dointvec_minmax,
 604                 .extra1         = &minolduid,
 605                 .extra2         = &maxolduid,
 606         },
 607         {
 608                 .procname       = "overflowgid",
 609                 .data           = &overflowgid,
 610                 .maxlen         = sizeof(int),
 611                 .mode           = 0644,
 612                 .proc_handler   = proc_dointvec_minmax,
 613                 .extra1         = &minolduid,
 614                 .extra2         = &maxolduid,
 615         },
 616 #ifdef CONFIG_S390
 617 #ifdef CONFIG_MATHEMU
 618         {
 619                 .procname       = "ieee_emulation_warnings",
 620                 .data           = &sysctl_ieee_emulation_warnings,
 621                 .maxlen         = sizeof(int),
 622                 .mode           = 0644,
 623                 .proc_handler   = proc_dointvec,
 624         },
 625 #endif
 626         {
 627                 .procname       = "userprocess_debug",
 628                 .data           = &sysctl_userprocess_debug,
 629                 .maxlen         = sizeof(int),
 630                 .mode           = 0644,
 631                 .proc_handler   = proc_dointvec,
 632         },
 633 #endif
 634         {
 635                 .procname       = "pid_max",
 636                 .data           = &pid_max,
 637                 .maxlen         = sizeof (int),
 638                 .mode           = 0644,
 639                 .proc_handler   = proc_dointvec_minmax,
 640                 .extra1         = &pid_max_min,
 641                 .extra2         = &pid_max_max,
 642         },
 643         {
 644                 .procname       = "panic_on_oops",
 645                 .data           = &panic_on_oops,
 646                 .maxlen         = sizeof(int),
 647                 .mode           = 0644,
 648                 .proc_handler   = proc_dointvec,
 649         },
 650 #if defined CONFIG_PRINTK
 651         {
 652                 .procname       = "printk",
 653                 .data           = &console_loglevel,
 654                 .maxlen         = 4*sizeof(int),
 655                 .mode           = 0644,
 656                 .proc_handler   = proc_dointvec,
 657         },
 658         {
 659                 .procname       = "printk_ratelimit",
 660                 .data           = &printk_ratelimit_state.interval,
 661                 .maxlen         = sizeof(int),
 662                 .mode           = 0644,
 663                 .proc_handler   = proc_dointvec_jiffies,
 664         },
 665         {
 666                 .procname       = "printk_ratelimit_burst",
 667                 .data           = &printk_ratelimit_state.burst,
 668                 .maxlen         = sizeof(int),
 669                 .mode           = 0644,
 670                 .proc_handler   = proc_dointvec,
 671         },
 672         {
 673                 .procname       = "printk_delay",
 674                 .data           = &printk_delay_msec,
 675                 .maxlen         = sizeof(int),
 676                 .mode           = 0644,
 677                 .proc_handler   = proc_dointvec_minmax,
 678                 .extra1         = &zero,
 679                 .extra2         = &ten_thousand,
 680         },
 681 #endif
 682         {
 683                 .procname       = "ngroups_max",
 684                 .data           = &ngroups_max,
 685                 .maxlen         = sizeof (int),
 686                 .mode           = 0444,
 687                 .proc_handler   = proc_dointvec,
 688         },
 689 #if defined(CONFIG_NMI_WATCHDOG)
 690         {
 691                 .procname       = "nmi_watchdog",
 692                 .data           = &nmi_watchdog_enabled,
 693                 .maxlen         = sizeof (int),
 694                 .mode           = 0644,
 695                 .proc_handler   = proc_nmi_enabled,
 696         },
 697 #endif
 698 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86) && !defined(CONFIG_NMI_WATCHDOG)
 699         {
 700                 .procname       = "unknown_nmi_panic",
 701                 .data           = &unknown_nmi_panic,
 702                 .maxlen         = sizeof (int),
 703                 .mode           = 0644,
 704                 .proc_handler   = proc_dointvec,
 705         },
 706         {
 707                 .procname       = "nmi_watchdog",
 708                 .data           = &nmi_watchdog_enabled,
 709                 .maxlen         = sizeof (int),
 710                 .mode           = 0644,
 711                 .proc_handler   = proc_nmi_enabled,
 712         },
 713 #endif
 714 #if defined(CONFIG_X86)
 715         {
 716                 .procname       = "panic_on_unrecovered_nmi",
 717                 .data           = &panic_on_unrecovered_nmi,
 718                 .maxlen         = sizeof(int),
 719                 .mode           = 0644,
 720                 .proc_handler   = proc_dointvec,
 721         },
 722         {
 723                 .procname       = "panic_on_io_nmi",
 724                 .data           = &panic_on_io_nmi,
 725                 .maxlen         = sizeof(int),
 726                 .mode           = 0644,
 727                 .proc_handler   = proc_dointvec,
 728         },
 729         {
 730                 .procname       = "bootloader_type",
 731                 .data           = &bootloader_type,
 732                 .maxlen         = sizeof (int),
 733                 .mode           = 0444,
 734                 .proc_handler   = proc_dointvec,
 735         },
 736         {
 737                 .procname       = "bootloader_version",
 738                 .data           = &bootloader_version,
 739                 .maxlen         = sizeof (int),
 740                 .mode           = 0444,
 741                 .proc_handler   = proc_dointvec,
 742         },
 743         {
 744                 .procname       = "kstack_depth_to_print",
 745                 .data           = &kstack_depth_to_print,
 746                 .maxlen         = sizeof(int),
 747                 .mode           = 0644,
 748                 .proc_handler   = proc_dointvec,
 749         },
 750         {
 751                 .procname       = "io_delay_type",
 752                 .data           = &io_delay_type,
 753                 .maxlen         = sizeof(int),
 754                 .mode           = 0644,
 755                 .proc_handler   = proc_dointvec,
 756         },
 757 #endif
 758 #if defined(CONFIG_MMU)
 759         {
 760                 .procname       = "randomize_va_space",
 761                 .data           = &randomize_va_space,
 762                 .maxlen         = sizeof(int),
 763                 .mode           = 0644,
 764                 .proc_handler   = proc_dointvec,
 765         },
 766 #endif
 767 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 768         {
 769                 .procname       = "spin_retry",
 770                 .data           = &spin_retry,
 771                 .maxlen         = sizeof (int),
 772                 .mode           = 0644,
 773                 .proc_handler   = proc_dointvec,
 774         },
 775 #endif
 776 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 777         {
 778                 .procname       = "acpi_video_flags",
 779                 .data           = &acpi_realmode_flags,
 780                 .maxlen         = sizeof (unsigned long),
 781                 .mode           = 0644,
 782                 .proc_handler   = proc_doulongvec_minmax,
 783         },
 784 #endif
 785 #ifdef CONFIG_IA64
 786         {
 787                 .procname       = "ignore-unaligned-usertrap",
 788                 .data           = &no_unaligned_warning,
 789                 .maxlen         = sizeof (int),
 790                 .mode           = 0644,
 791                 .proc_handler   = proc_dointvec,
 792         },
 793         {
 794                 .procname       = "unaligned-dump-stack",
 795                 .data           = &unaligned_dump_stack,
 796                 .maxlen         = sizeof (int),
 797                 .mode           = 0644,
 798                 .proc_handler   = proc_dointvec,
 799         },
 800 #endif
 801 #ifdef CONFIG_DETECT_SOFTLOCKUP
 802         {
 803                 .procname       = "softlockup_panic",
 804                 .data           = &softlockup_panic,
 805                 .maxlen         = sizeof(int),
 806                 .mode           = 0644,
 807                 .proc_handler   = proc_dointvec_minmax,
 808                 .extra1         = &zero,
 809                 .extra2         = &one,
 810         },
 811         {
 812                 .procname       = "softlockup_thresh",
 813                 .data           = &softlockup_thresh,
 814                 .maxlen         = sizeof(int),
 815                 .mode           = 0644,
 816                 .proc_handler   = proc_dosoftlockup_thresh,
 817                 .extra1         = &neg_one,
 818                 .extra2         = &sixty,
 819         },
 820 #endif
 821 #ifdef CONFIG_DETECT_HUNG_TASK
 822         {
 823                 .procname       = "hung_task_panic",
 824                 .data           = &sysctl_hung_task_panic,
 825                 .maxlen         = sizeof(int),
 826                 .mode           = 0644,
 827                 .proc_handler   = proc_dointvec_minmax,
 828                 .extra1         = &zero,
 829                 .extra2         = &one,
 830         },
 831         {
 832                 .procname       = "hung_task_check_count",
 833                 .data           = &sysctl_hung_task_check_count,
 834                 .maxlen         = sizeof(unsigned long),
 835                 .mode           = 0644,
 836                 .proc_handler   = proc_doulongvec_minmax,
 837         },
 838         {
 839                 .procname       = "hung_task_timeout_secs",
 840                 .data           = &sysctl_hung_task_timeout_secs,
 841                 .maxlen         = sizeof(unsigned long),
 842                 .mode           = 0644,
 843                 .proc_handler   = proc_dohung_task_timeout_secs,
 844         },
 845         {
 846                 .procname       = "hung_task_warnings",
 847                 .data           = &sysctl_hung_task_warnings,
 848                 .maxlen         = sizeof(unsigned long),
 849                 .mode           = 0644,
 850                 .proc_handler   = proc_doulongvec_minmax,
 851         },
 852 #endif
 853 #ifdef CONFIG_COMPAT
 854         {
 855                 .procname       = "compat-log",
 856                 .data           = &compat_log,
 857                 .maxlen         = sizeof (int),
 858                 .mode           = 0644,
 859                 .proc_handler   = proc_dointvec,
 860         },
 861 #endif
 862 #ifdef CONFIG_RT_MUTEXES
 863         {
 864                 .procname       = "max_lock_depth",
 865                 .data           = &max_lock_depth,
 866                 .maxlen         = sizeof(int),
 867                 .mode           = 0644,
 868                 .proc_handler   = proc_dointvec,
 869         },
 870 #endif
 871         {
 872                 .procname       = "poweroff_cmd",
 873                 .data           = &poweroff_cmd,
 874                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 875                 .mode           = 0644,
 876                 .proc_handler   = proc_dostring,
 877         },
 878 #ifdef CONFIG_KEYS
 879         {
 880                 .procname       = "keys",
 881                 .mode           = 0555,
 882                 .child          = key_sysctls,
 883         },
 884 #endif
 885 #ifdef CONFIG_RCU_TORTURE_TEST
 886         {
 887                 .procname       = "rcutorture_runnable",
 888                 .data           = &rcutorture_runnable,
 889                 .maxlen         = sizeof(int),
 890                 .mode           = 0644,
 891                 .proc_handler   = proc_dointvec,
 892         },
 893 #endif
 894 #ifdef CONFIG_SLOW_WORK
 895         {
 896                 .procname       = "slow-work",
 897                 .mode           = 0555,
 898                 .child          = slow_work_sysctls,
 899         },
 900 #endif
 901 #ifdef CONFIG_PERF_EVENTS
 902         {
 903                 .procname       = "perf_event_paranoid",
 904                 .data           = &sysctl_perf_event_paranoid,
 905                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
 906                 .mode           = 0644,
 907                 .proc_handler   = proc_dointvec,
 908         },
 909         {
 910                 .procname       = "perf_event_mlock_kb",
 911                 .data           = &sysctl_perf_event_mlock,
 912                 .maxlen         = sizeof(sysctl_perf_event_mlock),
 913                 .mode           = 0644,
 914                 .proc_handler   = proc_dointvec,
 915         },
 916         {
 917                 .procname       = "perf_event_max_sample_rate",
 918                 .data           = &sysctl_perf_event_sample_rate,
 919                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
 920                 .mode           = 0644,
 921                 .proc_handler   = proc_dointvec,
 922         },
 923 #endif
 924 #ifdef CONFIG_KMEMCHECK
 925         {
 926                 .procname       = "kmemcheck",
 927                 .data           = &kmemcheck_enabled,
 928                 .maxlen         = sizeof(int),
 929                 .mode           = 0644,
 930                 .proc_handler   = proc_dointvec,
 931         },
 932 #endif
 933 #ifdef CONFIG_BLOCK
 934         {
 935                 .procname       = "blk_iopoll",
 936                 .data           = &blk_iopoll_enabled,
 937                 .maxlen         = sizeof(int),
 938                 .mode           = 0644,
 939                 .proc_handler   = proc_dointvec,
 940         },
 941 #endif
 942 /*
 943  * NOTE: do not add new entries to this table unless you have read
 944  * Documentation/sysctl/ctl_unnumbered.txt
 945  */
 946         { }
 947 };
 948
 949 static struct ctl_table vm_table[] = {
 950         {
 951                 .procname       = "overcommit_memory",
 952                 .data           = &sysctl_overcommit_memory,
 953                 .maxlen         = sizeof(sysctl_overcommit_memory),
 954                 .mode           = 0644,
 955                 .proc_handler   = proc_dointvec,
 956         },
 957         {
 958                 .procname       = "panic_on_oom",
 959                 .data           = &sysctl_panic_on_oom,
 960                 .maxlen         = sizeof(sysctl_panic_on_oom),
 961                 .mode           = 0644,
 962                 .proc_handler   = proc_dointvec,
 963         },
 964         {
 965                 .procname       = "oom_kill_allocating_task",
 966                 .data           = &sysctl_oom_kill_allocating_task,
 967                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
 968                 .mode           = 0644,
 969                 .proc_handler   = proc_dointvec,
 970         },
 971         {
 972                 .procname       = "oom_dump_tasks",
 973                 .data           = &sysctl_oom_dump_tasks,
 974                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
 975                 .mode           = 0644,
 976                 .proc_handler   = proc_dointvec,
 977         },
 978         {
 979                 .procname       = "overcommit_ratio",
 980                 .data           = &sysctl_overcommit_ratio,
 981                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 982                 .mode           = 0644,
 983                 .proc_handler   = proc_dointvec,
 984         },
 985         {
 986                 .procname       = "page-cluster",
 987                 .data           = &page_cluster,
 988                 .maxlen         = sizeof(int),
 989                 .mode           = 0644,
 990                 .proc_handler   = proc_dointvec,
 991         },
 992         {
 993                 .procname       = "dirty_background_ratio",
 994                 .data           = &dirty_background_ratio,
 995                 .maxlen         = sizeof(dirty_background_ratio),
 996                 .mode           = 0644,
 997                 .proc_handler   = dirty_background_ratio_handler,
 998                 .extra1         = &zero,
 999                 .extra2         = &one_hundred,
1000         },
1001         {
1002                 .procname       = "dirty_background_bytes",
1003                 .data           = &dirty_background_bytes,
1004                 .maxlen         = sizeof(dirty_background_bytes),
1005                 .mode           = 0644,
1006                 .proc_handler   = dirty_background_bytes_handler,
1007                 .extra1         = &one_ul,
1008         },
1009         {
1010                 .procname       = "dirty_ratio",
1011                 .data           = &vm_dirty_ratio,
1012                 .maxlen         = sizeof(vm_dirty_ratio),
1013                 .mode           = 0644,
1014                 .proc_handler   = dirty_ratio_handler,
1015                 .extra1         = &zero,
1016                 .extra2         = &one_hundred,
1017         },
1018         {
1019                 .procname       = "dirty_bytes",
1020                 .data           = &vm_dirty_bytes,
1021                 .maxlen         = sizeof(vm_dirty_bytes),
1022                 .mode           = 0644,
1023                 .proc_handler   = dirty_bytes_handler,
1024                 .extra1         = &dirty_bytes_min,
1025         },
1026         {
1027                 .procname       = "dirty_writeback_centisecs",
1028                 .data           = &dirty_writeback_interval,
1029                 .maxlen         = sizeof(dirty_writeback_interval),
1030                 .mode           = 0644,
1031                 .proc_handler   = dirty_writeback_centisecs_handler,
1032         },
1033         {
1034                 .procname       = "dirty_expire_centisecs",
1035                 .data           = &dirty_expire_interval,
1036                 .maxlen         = sizeof(dirty_expire_interval),
1037                 .mode           = 0644,
1038                 .proc_handler   = proc_dointvec,
1039         },
1040         {
1041                 .procname       = "nr_pdflush_threads",
1042                 .data           = &nr_pdflush_threads,
1043                 .maxlen         = sizeof nr_pdflush_threads,
1044                 .mode           = 0444 /* read-only*/,
1045                 .proc_handler   = proc_dointvec,
1046         },
1047         {
1048                 .procname       = "swappiness",
1049                 .data           = &vm_swappiness,
1050                 .maxlen         = sizeof(vm_swappiness),
1051                 .mode           = 0644,
1052                 .proc_handler   = proc_dointvec_minmax,
1053                 .extra1         = &zero,
1054                 .extra2         = &one_hundred,
1055         },
1056 #ifdef CONFIG_HUGETLB_PAGE
1057         {
1058                 .procname       = "nr_hugepages",
1059                 .data           = NULL,
1060                 .maxlen         = sizeof(unsigned long),
1061                 .mode           = 0644,
1062                 .proc_handler   = hugetlb_sysctl_handler,
1063                 .extra1         = (void *)&hugetlb_zero,
1064                 .extra2         = (void *)&hugetlb_infinity,
1065         },
1066 #ifdef CONFIG_NUMA
1067         {
1068                 .procname       = "nr_hugepages_mempolicy",
1069                 .data           = NULL,
1070                 .maxlen         = sizeof(unsigned long),
1071                 .mode           = 0644,
1072                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1073                 .extra1         = (void *)&hugetlb_zero,
1074                 .extra2         = (void *)&hugetlb_infinity,
1075         },
1076 #endif
1077          {
1078                 .procname       = "hugetlb_shm_group",
1079                 .data           = &sysctl_hugetlb_shm_group,
1080                 .maxlen         = sizeof(gid_t),
1081                 .mode           = 0644,
1082                 .proc_handler   = proc_dointvec,
1083          },
1084          {
1085                 .procname       = "hugepages_treat_as_movable",
1086                 .data           = &hugepages_treat_as_movable,
1087                 .maxlen         = sizeof(int),
1088                 .mode           = 0644,
1089                 .proc_handler   = hugetlb_treat_movable_handler,
1090         },
1091         {
1092                 .procname       = "nr_overcommit_hugepages",
1093                 .data           = NULL,
1094                 .maxlen         = sizeof(unsigned long),
1095                 .mode           = 0644,
1096                 .proc_handler   = hugetlb_overcommit_handler,
1097                 .extra1         = (void *)&hugetlb_zero,
1098                 .extra2         = (void *)&hugetlb_infinity,
1099         },
1100 #endif
1101         {
1102                 .procname       = "lowmem_reserve_ratio",
1103                 .data           = &sysctl_lowmem_reserve_ratio,
1104                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1105                 .mode           = 0644,
1106                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1107         },
1108         {
1109                 .procname       = "drop_caches",
1110                 .data           = &sysctl_drop_caches,
1111                 .maxlen         = sizeof(int),
1112                 .mode           = 0644,
1113                 .proc_handler   = drop_caches_sysctl_handler,
1114         },
1115         {
1116                 .procname       = "min_free_kbytes",
1117                 .data           = &min_free_kbytes,
1118                 .maxlen         = sizeof(min_free_kbytes),
1119                 .mode           = 0644,
1120                 .proc_handler   = min_free_kbytes_sysctl_handler,
1121                 .extra1         = &zero,
1122         },
1123         {
1124                 .procname       = "percpu_pagelist_fraction",
1125                 .data           = &percpu_pagelist_fraction,
1126                 .maxlen         = sizeof(percpu_pagelist_fraction),
1127                 .mode           = 0644,
1128                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1129                 .extra1         = &min_percpu_pagelist_fract,
1130         },
1131 #ifdef CONFIG_MMU
1132         {
1133                 .procname       = "max_map_count",
1134                 .data           = &sysctl_max_map_count,
1135                 .maxlen         = sizeof(sysctl_max_map_count),
1136                 .mode           = 0644,
1137                 .proc_handler   = proc_dointvec_minmax,
1138                 .extra1         = &zero,
1139         },
1140 #else
1141         {
1142                 .procname       = "nr_trim_pages",
1143                 .data           = &sysctl_nr_trim_pages,
1144                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1145                 .mode           = 0644,
1146                 .proc_handler   = proc_dointvec_minmax,
1147                 .extra1         = &zero,
1148         },
1149 #endif
1150         {
1151                 .procname       = "laptop_mode",
1152                 .data           = &laptop_mode,
1153                 .maxlen         = sizeof(laptop_mode),
1154                 .mode           = 0644,
1155                 .proc_handler   = proc_dointvec_jiffies,
1156         },
1157         {
1158                 .procname       = "block_dump",
1159                 .data           = &block_dump,
1160                 .maxlen         = sizeof(block_dump),
1161                 .mode           = 0644,
1162                 .proc_handler   = proc_dointvec,
1163                 .extra1         = &zero,
1164         },
1165         {
1166                 .procname       = "vfs_cache_pressure",
1167                 .data           = &sysctl_vfs_cache_pressure,
1168                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1169                 .mode           = 0644,
1170                 .proc_handler   = proc_dointvec,
1171                 .extra1         = &zero,
1172         },
1173 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1174         {
1175                 .procname       = "legacy_va_layout",
1176                 .data           = &sysctl_legacy_va_layout,
1177                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1178                 .mode           = 0644,
1179                 .proc_handler   = proc_dointvec,
1180                 .extra1         = &zero,
1181         },
1182 #endif
1183 #ifdef CONFIG_NUMA
1184         {
1185                 .procname       = "zone_reclaim_mode",
1186                 .data           = &zone_reclaim_mode,
1187                 .maxlen         = sizeof(zone_reclaim_mode),
1188                 .mode           = 0644,
1189                 .proc_handler   = proc_dointvec,
1190                 .extra1         = &zero,
1191         },
1192         {
1193                 .procname       = "min_unmapped_ratio",
1194                 .data           = &sysctl_min_unmapped_ratio,
1195                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1196                 .mode           = 0644,
1197                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1198                 .extra1         = &zero,
1199                 .extra2         = &one_hundred,
1200         },
1201         {
1202                 .procname       = "min_slab_ratio",
1203                 .data           = &sysctl_min_slab_ratio,
1204                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1205                 .mode           = 0644,
1206                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1207                 .extra1         = &zero,
1208                 .extra2         = &one_hundred,
1209         },
1210 #endif
1211 #ifdef CONFIG_SMP
1212         {
1213                 .procname       = "stat_interval",
1214                 .data           = &sysctl_stat_interval,
1215                 .maxlen         = sizeof(sysctl_stat_interval),
1216                 .mode           = 0644,
1217                 .proc_handler   = proc_dointvec_jiffies,
1218         },
1219 #endif
1220 #ifdef CONFIG_MMU
1221         {
1222                 .procname       = "mmap_min_addr",
1223                 .data           = &dac_mmap_min_addr,
1224                 .maxlen         = sizeof(unsigned long),
1225                 .mode           = 0644,
1226                 .proc_handler   = mmap_min_addr_handler,
1227         },
1228 #endif
1229 #ifdef CONFIG_NUMA
1230         {
1231                 .procname       = "numa_zonelist_order",
1232                 .data           = &numa_zonelist_order,
1233                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1234                 .mode           = 0644,
1235                 .proc_handler   = numa_zonelist_order_handler,
1236         },
1237 #endif
1238 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1239    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1240         {
1241                 .procname       = "vdso_enabled",
1242                 .data           = &vdso_enabled,
1243                 .maxlen         = sizeof(vdso_enabled),
1244                 .mode           = 0644,
1245                 .proc_handler   = proc_dointvec,
1246                 .extra1         = &zero,
1247         },
1248 #endif
1249 #ifdef CONFIG_HIGHMEM
1250         {
1251                 .procname       = "highmem_is_dirtyable",
1252                 .data           = &vm_highmem_is_dirtyable,
1253                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1254                 .mode           = 0644,
1255                 .proc_handler   = proc_dointvec_minmax,
1256                 .extra1         = &zero,
1257                 .extra2         = &one,
1258         },
1259 #endif
1260         {
1261                 .procname       = "scan_unevictable_pages",
1262                 .data           = &scan_unevictable_pages,
1263                 .maxlen         = sizeof(scan_unevictable_pages),
1264                 .mode           = 0644,
1265                 .proc_handler   = scan_unevictable_handler,
1266         },
1267 #ifdef CONFIG_MEMORY_FAILURE
1268         {
1269                 .procname       = "memory_failure_early_kill",
1270                 .data           = &sysctl_memory_failure_early_kill,
1271                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1272                 .mode           = 0644,
1273                 .proc_handler   = proc_dointvec_minmax,
1274                 .extra1         = &zero,
1275                 .extra2         = &one,
1276         },
1277         {
1278                 .procname       = "memory_failure_recovery",
1279                 .data           = &sysctl_memory_failure_recovery,
1280                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1281                 .mode           = 0644,
1282                 .proc_handler   = proc_dointvec_minmax,
1283                 .extra1         = &zero,
1284                 .extra2         = &one,
1285         },
1286 #endif
1287
1288 /*
1289  * NOTE: do not add new entries to this table unless you have read
1290  * Documentation/sysctl/ctl_unnumbered.txt
1291  */
1292         { }
1293 };
1294
1295 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1296 static struct ctl_table binfmt_misc_table[] = {
1297         { }
1298 };
1299 #endif
1300
1301 static struct ctl_table fs_table[] = {
1302         {
1303                 .procname       = "inode-nr",
1304                 .data           = &inodes_stat,
1305                 .maxlen         = 2*sizeof(int),
1306                 .mode           = 0444,
1307                 .proc_handler   = proc_dointvec,
1308         },
1309         {
1310                 .procname       = "inode-state",
1311                 .data           = &inodes_stat,
1312                 .maxlen         = 7*sizeof(int),
1313                 .mode           = 0444,
1314                 .proc_handler   = proc_dointvec,
1315         },
1316         {
1317                 .procname       = "file-nr",
1318                 .data           = &files_stat,
1319                 .maxlen         = 3*sizeof(int),
1320                 .mode           = 0444,
1321                 .proc_handler   = proc_nr_files,
1322         },
1323         {
1324                 .procname       = "file-max",
1325                 .data           = &files_stat.max_files,
1326                 .maxlen         = sizeof(int),
1327                 .mode           = 0644,
1328                 .proc_handler   = proc_dointvec,
1329         },
1330         {
1331                 .procname       = "nr_open",
1332                 .data           = &sysctl_nr_open,
1333                 .maxlen         = sizeof(int),
1334                 .mode           = 0644,
1335                 .proc_handler   = proc_dointvec_minmax,
1336                 .extra1         = &sysctl_nr_open_min,
1337                 .extra2         = &sysctl_nr_open_max,
1338         },
1339         {
1340                 .procname       = "dentry-state",
1341                 .data           = &dentry_stat,
1342                 .maxlen         = 6*sizeof(int),
1343                 .mode           = 0444,
1344                 .proc_handler   = proc_dointvec,
1345         },
1346         {
1347                 .procname       = "overflowuid",
1348                 .data           = &fs_overflowuid,
1349                 .maxlen         = sizeof(int),
1350                 .mode           = 0644,
1351                 .proc_handler   = proc_dointvec_minmax,
1352                 .extra1         = &minolduid,
1353                 .extra2         = &maxolduid,
1354         },
1355         {
1356                 .procname       = "overflowgid",
1357                 .data           = &fs_overflowgid,
1358                 .maxlen         = sizeof(int),
1359                 .mode           = 0644,
1360                 .proc_handler   = proc_dointvec_minmax,
1361                 .extra1         = &minolduid,
1362                 .extra2         = &maxolduid,
1363         },
1364 #ifdef CONFIG_FILE_LOCKING
1365         {
1366                 .procname       = "leases-enable",
1367                 .data           = &leases_enable,
1368                 .maxlen         = sizeof(int),
1369                 .mode           = 0644,
1370                 .proc_handler   = proc_dointvec,
1371         },
1372 #endif
1373 #ifdef CONFIG_DNOTIFY
1374         {
1375                 .procname       = "dir-notify-enable",
1376                 .data           = &dir_notify_enable,
1377                 .maxlen         = sizeof(int),
1378                 .mode           = 0644,
1379                 .proc_handler   = proc_dointvec,
1380         },
1381 #endif
1382 #ifdef CONFIG_MMU
1383 #ifdef CONFIG_FILE_LOCKING
1384         {
1385                 .procname       = "lease-break-time",
1386                 .data           = &lease_break_time,
1387                 .maxlen         = sizeof(int),
1388                 .mode           = 0644,
1389                 .proc_handler   = proc_dointvec,
1390         },
1391 #endif
1392 #ifdef CONFIG_AIO
1393         {
1394                 .procname       = "aio-nr",
1395                 .data           = &aio_nr,
1396                 .maxlen         = sizeof(aio_nr),
1397                 .mode           = 0444,
1398                 .proc_handler   = proc_doulongvec_minmax,
1399         },
1400         {
1401                 .procname       = "aio-max-nr",
1402                 .data           = &aio_max_nr,
1403                 .maxlen         = sizeof(aio_max_nr),
1404                 .mode           = 0644,
1405                 .proc_handler   = proc_doulongvec_minmax,
1406         },
1407 #endif /* CONFIG_AIO */
1408 #ifdef CONFIG_INOTIFY_USER
1409         {
1410                 .procname       = "inotify",
1411                 .mode           = 0555,
1412                 .child          = inotify_table,
1413         },
1414 #endif
1415 #ifdef CONFIG_EPOLL
1416         {
1417                 .procname       = "epoll",
1418                 .mode           = 0555,
1419                 .child          = epoll_table,
1420         },
1421 #endif
1422 #endif
1423         {
1424                 .procname       = "suid_dumpable",
1425                 .data           = &suid_dumpable,
1426                 .maxlen         = sizeof(int),
1427                 .mode           = 0644,
1428                 .proc_handler   = proc_dointvec_minmax,
1429                 .extra1         = &zero,
1430                 .extra2         = &two,
1431         },
1432 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1433         {
1434                 .procname       = "binfmt_misc",
1435                 .mode           = 0555,
1436                 .child          = binfmt_misc_table,
1437         },
1438 #endif
1439 /*
1440  * NOTE: do not add new entries to this table unless you have read
1441  * Documentation/sysctl/ctl_unnumbered.txt
1442  */
1443         { }
1444 };
1445
1446 static struct ctl_table debug_table[] = {
1447 #if defined(CONFIG_X86) || defined(CONFIG_PPC) || defined(CONFIG_SPARC)
1448         {
1449                 .procname       = "exception-trace",
1450                 .data           = &show_unhandled_signals,
1451                 .maxlen         = sizeof(int),
1452                 .mode           = 0644,
1453                 .proc_handler   = proc_dointvec
1454         },
1455 #endif
1456 #if defined(CONFIG_OPTPROBES)
1457         {
1458                 .procname       = "kprobes-optimization",
1459                 .data           = &sysctl_kprobes_optimization,
1460                 .maxlen         = sizeof(int),
1461                 .mode           = 0644,
1462                 .proc_handler   = proc_kprobes_optimization_handler,
1463                 .extra1         = &zero,
1464                 .extra2         = &one,
1465         },
1466 #endif
1467         { }
1468 };
1469
1470 static struct ctl_table dev_table[] = {
1471         { }
1472 };
1473
1474 static DEFINE_SPINLOCK(sysctl_lock);
1475
1476 /* called under sysctl_lock */
1477 static int use_table(struct ctl_table_header *p)
1478 {
1479         if (unlikely(p->unregistering))
1480                 return 0;
1481         p->used++;
1482         return 1;
1483 }
1484
1485 /* called under sysctl_lock */
1486 static void unuse_table(struct ctl_table_header *p)
1487 {
1488         if (!--p->used)
1489                 if (unlikely(p->unregistering))
1490                         complete(p->unregistering);
1491 }
1492
1493 /* called under sysctl_lock, will reacquire if has to wait */
1494 static void start_unregistering(struct ctl_table_header *p)
1495 {
1496         /*
1497          * if p->used is 0, nobody will ever touch that entry again;
1498          * we'll eliminate all paths to it before dropping sysctl_lock
1499          */
1500         if (unlikely(p->used)) {
1501                 struct completion wait;
1502                 init_completion(&wait);
1503                 p->unregistering = &wait;
1504                 spin_unlock(&sysctl_lock);
1505                 wait_for_completion(&wait);
1506                 spin_lock(&sysctl_lock);
1507         } else {
1508                 /* anything non-NULL; we'll never dereference it */
1509                 p->unregistering = ERR_PTR(-EINVAL);
1510         }
1511         /*
1512          * do not remove from the list until nobody holds it; walking the
1513          * list in do_sysctl() relies on that.
1514          */
1515         list_del_init(&p->ctl_entry);
1516 }
1517
1518 void sysctl_head_get(struct ctl_table_header *head)
1519 {
1520         spin_lock(&sysctl_lock);
1521         head->count++;
1522         spin_unlock(&sysctl_lock);
1523 }
1524
1525 void sysctl_head_put(struct ctl_table_header *head)
1526 {
1527         spin_lock(&sysctl_lock);
1528         if (!--head->count)
1529                 kfree(head);
1530         spin_unlock(&sysctl_lock);
1531 }
1532
1533 struct ctl_table_header *sysctl_head_grab(struct ctl_table_header *head)
1534 {
1535         if (!head)
1536                 BUG();
1537         spin_lock(&sysctl_lock);
1538         if (!use_table(head))
1539                 head = ERR_PTR(-ENOENT);
1540         spin_unlock(&sysctl_lock);
1541         return head;
1542 }
1543
1544 void sysctl_head_finish(struct ctl_table_header *head)
1545 {
1546         if (!head)
1547                 return;
1548         spin_lock(&sysctl_lock);
1549         unuse_table(head);
1550         spin_unlock(&sysctl_lock);
1551 }
1552
1553 static struct ctl_table_set *
1554 lookup_header_set(struct ctl_table_root *root, struct nsproxy *namespaces)
1555 {
1556         struct ctl_table_set *set = &root->default_set;
1557         if (root->lookup)
1558                 set = root->lookup(root, namespaces);
1559         return set;
1560 }
1561
1562 static struct list_head *
1563 lookup_header_list(struct ctl_table_root *root, struct nsproxy *namespaces)
1564 {
1565         struct ctl_table_set *set = lookup_header_set(root, namespaces);
1566         return &set->list;
1567 }
1568
1569 struct ctl_table_header *__sysctl_head_next(struct nsproxy *namespaces,
1570                                             struct ctl_table_header *prev)
1571 {
1572         struct ctl_table_root *root;
1573         struct list_head *header_list;
1574         struct ctl_table_header *head;
1575         struct list_head *tmp;
1576
1577         spin_lock(&sysctl_lock);
1578         if (prev) {
1579                 head = prev;
1580                 tmp = &prev->ctl_entry;
1581                 unuse_table(prev);
1582                 goto next;
1583         }
1584         tmp = &root_table_header.ctl_entry;
1585         for (;;) {
1586                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1587
1588                 if (!use_table(head))
1589                         goto next;
1590                 spin_unlock(&sysctl_lock);
1591                 return head;
1592         next:
1593                 root = head->root;
1594                 tmp = tmp->next;
1595                 header_list = lookup_header_list(root, namespaces);
1596                 if (tmp != header_list)
1597                         continue;
1598
1599                 do {
1600                         root = list_entry(root->root_list.next,
1601                                         struct ctl_table_root, root_list);
1602                         if (root == &sysctl_table_root)
1603                                 goto out;
1604                         header_list = lookup_header_list(root, namespaces);
1605                 } while (list_empty(header_list));
1606                 tmp = header_list->next;
1607         }
1608 out:
1609         spin_unlock(&sysctl_lock);
1610         return NULL;
1611 }
1612
1613 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1614 {
1615         return __sysctl_head_next(current->nsproxy, prev);
1616 }
1617
1618 void register_sysctl_root(struct ctl_table_root *root)
1619 {
1620         spin_lock(&sysctl_lock);
1621         list_add_tail(&root->root_list, &sysctl_table_root.root_list);
1622         spin_unlock(&sysctl_lock);
1623 }
1624
1625 /*
1626  * sysctl_perm does NOT grant the superuser all rights automatically, because
1627  * some sysctl variables are readonly even to root.
1628  */
1629
1630 static int test_perm(int mode, int op)
1631 {
1632         if (!current_euid())
1633                 mode >>= 6;
1634         else if (in_egroup_p(0))
1635                 mode >>= 3;
1636         if ((op & ~mode & (MAY_READ|MAY_WRITE|MAY_EXEC)) == 0)
1637                 return 0;
1638         return -EACCES;
1639 }
1640
1641 int sysctl_perm(struct ctl_table_root *root, struct ctl_table *table, int op)
1642 {
1643         int error;
1644         int mode;
1645
1646         error = security_sysctl(table, op & (MAY_READ | MAY_WRITE | MAY_EXEC));
1647         if (error)
1648                 return error;
1649
1650         if (root->permissions)
1651                 mode = root->permissions(root, current->nsproxy, table);
1652         else
1653                 mode = table->mode;
1654
1655         return test_perm(mode, op);
1656 }
1657
1658 static void sysctl_set_parent(struct ctl_table *parent, struct ctl_table *table)
1659 {
1660         for (; table->procname; table++) {
1661                 table->parent = parent;
1662                 if (table->child)
1663                         sysctl_set_parent(table, table->child);
1664         }
1665 }
1666
1667 static __init int sysctl_init(void)
1668 {
1669         sysctl_set_parent(NULL, root_table);
1670 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1671         {
1672                 int err;
1673                 err = sysctl_check_table(current->nsproxy, root_table);
1674         }
1675 #endif
1676         return 0;
1677 }
1678
1679 core_initcall(sysctl_init);
1680
1681 static struct ctl_table *is_branch_in(struct ctl_table *branch,
1682                                       struct ctl_table *table)
1683 {
1684         struct ctl_table *p;
1685         const char *s = branch->procname;
1686
1687         /* branch should have named subdirectory as its first element */
1688         if (!s || !branch->child)
1689                 return NULL;
1690
1691         /* ... and nothing else */
1692         if (branch[1].procname)
1693                 return NULL;
1694
1695         /* table should contain subdirectory with the same name */
1696         for (p = table; p->procname; p++) {
1697                 if (!p->child)
1698                         continue;
1699                 if (p->procname && strcmp(p->procname, s) == 0)
1700                         return p;
1701         }
1702         return NULL;
1703 }
1704
1705 /* see if attaching q to p would be an improvement */
1706 static void try_attach(struct ctl_table_header *p, struct ctl_table_header *q)
1707 {
1708         struct ctl_table *to = p->ctl_table, *by = q->ctl_table;
1709         struct ctl_table *next;
1710         int is_better = 0;
1711         int not_in_parent = !p->attached_by;
1712
1713         while ((next = is_branch_in(by, to)) != NULL) {
1714                 if (by == q->attached_by)
1715                         is_better = 1;
1716                 if (to == p->attached_by)
1717                         not_in_parent = 1;
1718                 by = by->child;
1719                 to = next->child;
1720         }
1721
1722         if (is_better && not_in_parent) {
1723                 q->attached_by = by;
1724                 q->attached_to = to;
1725                 q->parent = p;
1726         }
1727 }
1728
1729 /**
1730  * __register_sysctl_paths - register a sysctl hierarchy
1731  * @root: List of sysctl headers to register on
1732  * @namespaces: Data to compute which lists of sysctl entries are visible
1733  * @path: The path to the directory the sysctl table is in.
1734  * @table: the top-level table structure
1735  *
1736  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1737  * array. A completely 0 filled entry terminates the table.
1738  *
1739  * The members of the &struct ctl_table structure are used as follows:
1740  *
1741  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1742  *            enter a sysctl file
1743  *
1744  * data - a pointer to data for use by proc_handler
1745  *
1746  * maxlen - the maximum size in bytes of the data
1747  *
1748  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1749  *
1750  * child - a pointer to the child sysctl table if this entry is a directory, or
1751  *         %NULL.
1752  *
1753  * proc_handler - the text handler routine (described below)
1754  *
1755  * de - for internal use by the sysctl routines
1756  *
1757  * extra1, extra2 - extra pointers usable by the proc handler routines
1758  *
1759  * Leaf nodes in the sysctl tree will be represented by a single file
1760  * under /proc; non-leaf nodes will be represented by directories.
1761  *
1762  * sysctl(2) can automatically manage read and write requests through
1763  * the sysctl table.  The data and maxlen fields of the ctl_table
1764  * struct enable minimal validation of the values being written to be
1765  * performed, and the mode field allows minimal authentication.
1766  *
1767  * There must be a proc_handler routine for any terminal nodes
1768  * mirrored under /proc/sys (non-terminals are handled by a built-in
1769  * directory handler).  Several default handlers are available to
1770  * cover common cases -
1771  *
1772  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1773  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1774  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1775  *
1776  * It is the handler's job to read the input buffer from user memory
1777  * and process it. The handler should return 0 on success.
1778  *
1779  * This routine returns %NULL on a failure to register, and a pointer
1780  * to the table header on success.
1781  */
1782 struct ctl_table_header *__register_sysctl_paths(
1783         struct ctl_table_root *root,
1784         struct nsproxy *namespaces,
1785         const struct ctl_path *path, struct ctl_table *table)
1786 {
1787         struct ctl_table_header *header;
1788         struct ctl_table *new, **prevp;
1789         unsigned int n, npath;
1790         struct ctl_table_set *set;
1791
1792         /* Count the path components */
1793         for (npath = 0; path[npath].procname; ++npath)
1794                 ;
1795
1796         /*
1797          * For each path component, allocate a 2-element ctl_table array.
1798          * The first array element will be filled with the sysctl entry
1799          * for this, the second will be the sentinel (procname == 0).
1800          *
1801          * We allocate everything in one go so that we don't have to
1802          * worry about freeing additional memory in unregister_sysctl_table.
1803          */
1804         header = kzalloc(sizeof(struct ctl_table_header) +
1805                          (2 * npath * sizeof(struct ctl_table)), GFP_KERNEL);
1806         if (!header)
1807                 return NULL;
1808
1809         new = (struct ctl_table *) (header + 1);
1810
1811         /* Now connect the dots */
1812         prevp = &header->ctl_table;
1813         for (n = 0; n < npath; ++n, ++path) {
1814                 /* Copy the procname */
1815                 new->procname = path->procname;
1816                 new->mode     = 0555;
1817
1818                 *prevp = new;
1819                 prevp = &new->child;
1820
1821                 new += 2;
1822         }
1823         *prevp = table;
1824         header->ctl_table_arg = table;
1825
1826         INIT_LIST_HEAD(&header->ctl_entry);
1827         header->used = 0;
1828         header->unregistering = NULL;
1829         header->root = root;
1830         sysctl_set_parent(NULL, header->ctl_table);
1831         header->count = 1;
1832 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1833         if (sysctl_check_table(namespaces, header->ctl_table)) {
1834                 kfree(header);
1835                 return NULL;
1836         }
1837 #endif
1838         spin_lock(&sysctl_lock);
1839         header->set = lookup_header_set(root, namespaces);
1840         header->attached_by = header->ctl_table;
1841         header->attached_to = root_table;
1842         header->parent = &root_table_header;
1843         for (set = header->set; set; set = set->parent) {
1844                 struct ctl_table_header *p;
1845                 list_for_each_entry(p, &set->list, ctl_entry) {
1846                         if (p->unregistering)
1847                                 continue;
1848                         try_attach(p, header);
1849                 }
1850         }
1851         header->parent->count++;
1852         list_add_tail(&header->ctl_entry, &header->set->list);
1853         spin_unlock(&sysctl_lock);
1854
1855         return header;
1856 }
1857
1858 /**
1859  * register_sysctl_table_path - register a sysctl table hierarchy
1860  * @path: The path to the directory the sysctl table is in.
1861  * @table: the top-level table structure
1862  *
1863  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1864  * array. A completely 0 filled entry terminates the table.
1865  *
1866  * See __register_sysctl_paths for more details.
1867  */
1868 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
1869                                                 struct ctl_table *table)
1870 {
1871         return __register_sysctl_paths(&sysctl_table_root, current->nsproxy,
1872                                         path, table);
1873 }
1874
1875 /**
1876  * register_sysctl_table - register a sysctl table hierarchy
1877  * @table: the top-level table structure
1878  *
1879  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1880  * array. A completely 0 filled entry terminates the table.
1881  *
1882  * See register_sysctl_paths for more details.
1883  */
1884 struct ctl_table_header *register_sysctl_table(struct ctl_table *table)
1885 {
1886         static const struct ctl_path null_path[] = { {} };
1887
1888         return register_sysctl_paths(null_path, table);
1889 }
1890
1891 /**
1892  * unregister_sysctl_table - unregister a sysctl table hierarchy
1893  * @header: the header returned from register_sysctl_table
1894  *
1895  * Unregisters the sysctl table and all children. proc entries may not
1896  * actually be removed until they are no longer used by anyone.
1897  */
1898 void unregister_sysctl_table(struct ctl_table_header * header)
1899 {
1900         might_sleep();
1901
1902         if (header == NULL)
1903                 return;
1904
1905         spin_lock(&sysctl_lock);
1906         start_unregistering(header);
1907         if (!--header->parent->count) {
1908                 WARN_ON(1);
1909                 kfree(header->parent);
1910         }
1911         if (!--header->count)
1912                 kfree(header);
1913         spin_unlock(&sysctl_lock);
1914 }
1915
1916 int sysctl_is_seen(struct ctl_table_header *p)
1917 {
1918         struct ctl_table_set *set = p->set;
1919         int res;
1920         spin_lock(&sysctl_lock);
1921         if (p->unregistering)
1922                 res = 0;
1923         else if (!set->is_seen)
1924                 res = 1;
1925         else
1926                 res = set->is_seen(set);
1927         spin_unlock(&sysctl_lock);
1928         return res;
1929 }
1930
1931 void setup_sysctl_set(struct ctl_table_set *p,
1932         struct ctl_table_set *parent,
1933         int (*is_seen)(struct ctl_table_set *))
1934 {
1935         INIT_LIST_HEAD(&p->list);
1936         p->parent = parent ? parent : &sysctl_table_root.default_set;
1937         p->is_seen = is_seen;
1938 }
1939
1940 #else /* !CONFIG_SYSCTL */
1941 struct ctl_table_header *register_sysctl_table(struct ctl_table * table)
1942 {
1943         return NULL;
1944 }
1945
1946 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
1947                                                     struct ctl_table *table)
1948 {
1949         return NULL;
1950 }
1951
1952 void unregister_sysctl_table(struct ctl_table_header * table)
1953 {
1954 }
1955
1956 void setup_sysctl_set(struct ctl_table_set *p,
1957         struct ctl_table_set *parent,
1958         int (*is_seen)(struct ctl_table_set *))
1959 {
1960 }
1961
1962 void sysctl_head_put(struct ctl_table_header *head)
1963 {
1964 }
1965
1966 #endif /* CONFIG_SYSCTL */
1967
1968 /*
1969  * /proc/sys support
1970  */
1971
1972 #ifdef CONFIG_PROC_SYSCTL
1973
1974 static int _proc_do_string(void* data, int maxlen, int write,
1975                            void __user *buffer,
1976                            size_t *lenp, loff_t *ppos)
1977 {
1978         size_t len;
1979         char __user *p;
1980         char c;
1981
1982         if (!data || !maxlen || !*lenp) {
1983                 *lenp = 0;
1984                 return 0;
1985         }
1986
1987         if (write) {
1988                 len = 0;
1989                 p = buffer;
1990                 while (len < *lenp) {
1991                         if (get_user(c, p++))
1992                                 return -EFAULT;
1993                         if (c == 0 || c == '\n')
1994                                 break;
1995                         len++;
1996                 }
1997                 if (len >= maxlen)
1998                         len = maxlen-1;
1999                 if(copy_from_user(data, buffer, len))
2000                         return -EFAULT;
2001                 ((char *) data)[len] = 0;
2002                 *ppos += *lenp;
2003         } else {
2004                 len = strlen(data);
2005                 if (len > maxlen)
2006                         len = maxlen;
2007
2008                 if (*ppos > len) {
2009                         *lenp = 0;
2010                         return 0;
2011                 }
2012
2013                 data += *ppos;
2014                 len  -= *ppos;
2015
2016                 if (len > *lenp)
2017                         len = *lenp;
2018                 if (len)
2019                         if(copy_to_user(buffer, data, len))
2020                                 return -EFAULT;
2021                 if (len < *lenp) {
2022                         if(put_user('\n', ((char __user *) buffer) + len))
2023                                 return -EFAULT;
2024                         len++;
2025                 }
2026                 *lenp = len;
2027                 *ppos += len;
2028         }
2029         return 0;
2030 }
2031
2032 /**
2033  * proc_dostring - read a string sysctl
2034  * @table: the sysctl table
2035  * @write: %TRUE if this is a write to the sysctl file
2036  * @buffer: the user buffer
2037  * @lenp: the size of the user buffer
2038  * @ppos: file position
2039  *
2040  * Reads/writes a string from/to the user buffer. If the kernel
2041  * buffer provided is not large enough to hold the string, the
2042  * string is truncated. The copied string is %NULL-terminated.
2043  * If the string is being read by the user process, it is copied
2044  * and a newline '\n' is added. It is truncated if the buffer is
2045  * not large enough.
2046  *
2047  * Returns 0 on success.
2048  */
2049 int proc_dostring(struct ctl_table *table, int write,
2050                   void __user *buffer, size_t *lenp, loff_t *ppos)
2051 {
2052         return _proc_do_string(table->data, table->maxlen, write,
2053                                buffer, lenp, ppos);
2054 }
2055
2056
2057 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
2058                                  int *valp,
2059                                  int write, void *data)
2060 {
2061         if (write) {
2062                 *valp = *negp ? -*lvalp : *lvalp;
2063         } else {
2064                 int val = *valp;
2065                 if (val < 0) {
2066                         *negp = -1;
2067                         *lvalp = (unsigned long)-val;
2068                 } else {
2069                         *negp = 0;
2070                         *lvalp = (unsigned long)val;
2071                 }
2072         }
2073         return 0;
2074 }
2075
2076 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2077                   int write, void __user *buffer,
2078                   size_t *lenp, loff_t *ppos,
2079                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2080                               int write, void *data),
2081                   void *data)
2082 {
2083 #define TMPBUFLEN 21
2084         int *i, vleft, first = 1, neg;
2085         unsigned long lval;
2086         size_t left, len;
2087
2088         char buf[TMPBUFLEN], *p;
2089         char __user *s = buffer;
2090
2091         if (!tbl_data || !table->maxlen || !*lenp ||
2092             (*ppos && !write)) {
2093                 *lenp = 0;
2094                 return 0;
2095         }
2096
2097         i = (int *) tbl_data;
2098         vleft = table->maxlen / sizeof(*i);
2099         left = *lenp;
2100
2101         if (!conv)
2102                 conv = do_proc_dointvec_conv;
2103
2104         for (; left && vleft--; i++, first=0) {
2105                 if (write) {
2106                         while (left) {
2107                                 char c;
2108                                 if (get_user(c, s))
2109                                         return -EFAULT;
2110                                 if (!isspace(c))
2111                                         break;
2112                                 left--;
2113                                 s++;
2114                         }
2115                         if (!left)
2116                                 break;
2117                         neg = 0;
2118                         len = left;
2119                         if (len > sizeof(buf) - 1)
2120                                 len = sizeof(buf) - 1;
2121                         if (copy_from_user(buf, s, len))
2122                                 return -EFAULT;
2123                         buf[len] = 0;
2124                         p = buf;
2125                         if (*p == '-' && left > 1) {
2126                                 neg = 1;
2127                                 p++;
2128                         }
2129                         if (*p < '0' || *p > '9')
2130                                 break;
2131
2132                         lval = simple_strtoul(p, &p, 0);
2133
2134                         len = p-buf;
2135                         if ((len < left) && *p && !isspace(*p))
2136                                 break;
2137                         s += len;
2138                         left -= len;
2139
2140                         if (conv(&neg, &lval, i, 1, data))
2141                                 break;
2142                 } else {
2143                         p = buf;
2144                         if (!first)
2145                                 *p++ = '\t';
2146
2147                         if (conv(&neg, &lval, i, 0, data))
2148                                 break;
2149
2150                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
2151                         len = strlen(buf);
2152                         if (len > left)
2153                                 len = left;
2154                         if(copy_to_user(s, buf, len))
2155                                 return -EFAULT;
2156                         left -= len;
2157                         s += len;
2158                 }
2159         }
2160
2161         if (!write && !first && left) {
2162                 if(put_user('\n', s))
2163                         return -EFAULT;
2164                 left--, s++;
2165         }
2166         if (write) {
2167                 while (left) {
2168                         char c;
2169                         if (get_user(c, s++))
2170                                 return -EFAULT;
2171                         if (!isspace(c))
2172                                 break;
2173                         left--;
2174                 }
2175         }
2176         if (write && first)
2177                 return -EINVAL;
2178         *lenp -= left;
2179         *ppos += *lenp;
2180         return 0;
2181 #undef TMPBUFLEN
2182 }
2183
2184 static int do_proc_dointvec(struct ctl_table *table, int write,
2185                   void __user *buffer, size_t *lenp, loff_t *ppos,
2186                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2187                               int write, void *data),
2188                   void *data)
2189 {
2190         return __do_proc_dointvec(table->data, table, write,
2191                         buffer, lenp, ppos, conv, data);
2192 }
2193
2194 /**
2195  * proc_dointvec - read a vector of integers
2196  * @table: the sysctl table
2197  * @write: %TRUE if this is a write to the sysctl file
2198  * @buffer: the user buffer
2199  * @lenp: the size of the user buffer
2200  * @ppos: file position
2201  *
2202  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2203  * values from/to the user buffer, treated as an ASCII string.
2204  *
2205  * Returns 0 on success.
2206  */
2207 int proc_dointvec(struct ctl_table *table, int write,
2208                      void __user *buffer, size_t *lenp, loff_t *ppos)
2209 {
2210     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2211                             NULL,NULL);
2212 }
2213
2214 /*
2215  * Taint values can only be increased
2216  * This means we can safely use a temporary.
2217  */
2218 static int proc_taint(struct ctl_table *table, int write,
2219                                void __user *buffer, size_t *lenp, loff_t *ppos)
2220 {
2221         struct ctl_table t;
2222         unsigned long tmptaint = get_taint();
2223         int err;
2224
2225         if (write && !capable(CAP_SYS_ADMIN))
2226                 return -EPERM;
2227
2228         t = *table;
2229         t.data = &tmptaint;
2230         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2231         if (err < 0)
2232                 return err;
2233
2234         if (write) {
2235                 /*
2236                  * Poor man's atomic or. Not worth adding a primitive
2237                  * to everyone's atomic.h for this
2238                  */
2239                 int i;
2240                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2241                         if ((tmptaint >> i) & 1)
2242                                 add_taint(i);
2243                 }
2244         }
2245
2246         return err;
2247 }
2248
2249 struct do_proc_dointvec_minmax_conv_param {
2250         int *min;
2251         int *max;
2252 };
2253
2254 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
2255                                         int *valp,
2256                                         int write, void *data)
2257 {
2258         struct do_proc_dointvec_minmax_conv_param *param = data;
2259         if (write) {
2260                 int val = *negp ? -*lvalp : *lvalp;
2261                 if ((param->min && *param->min > val) ||
2262                     (param->max && *param->max < val))
2263                         return -EINVAL;
2264                 *valp = val;
2265         } else {
2266                 int val = *valp;
2267                 if (val < 0) {
2268                         *negp = -1;
2269                         *lvalp = (unsigned long)-val;
2270                 } else {
2271                         *negp = 0;
2272                         *lvalp = (unsigned long)val;
2273                 }
2274         }
2275         return 0;
2276 }
2277
2278 /**
2279  * proc_dointvec_minmax - read a vector of integers with min/max values
2280  * @table: the sysctl table
2281  * @write: %TRUE if this is a write to the sysctl file
2282  * @buffer: the user buffer
2283  * @lenp: the size of the user buffer
2284  * @ppos: file position
2285  *
2286  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2287  * values from/to the user buffer, treated as an ASCII string.
2288  *
2289  * This routine will ensure the values are within the range specified by
2290  * table->extra1 (min) and table->extra2 (max).
2291  *
2292  * Returns 0 on success.
2293  */
2294 int proc_dointvec_minmax(struct ctl_table *table, int write,
2295                   void __user *buffer, size_t *lenp, loff_t *ppos)
2296 {
2297         struct do_proc_dointvec_minmax_conv_param param = {
2298                 .min = (int *) table->extra1,
2299                 .max = (int *) table->extra2,
2300         };
2301         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2302                                 do_proc_dointvec_minmax_conv, &param);
2303 }
2304
2305 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2306                                      void __user *buffer,
2307                                      size_t *lenp, loff_t *ppos,
2308                                      unsigned long convmul,
2309                                      unsigned long convdiv)
2310 {
2311 #define TMPBUFLEN 21
2312         unsigned long *i, *min, *max, val;
2313         int vleft, first=1, neg;
2314         size_t len, left;
2315         char buf[TMPBUFLEN], *p;
2316         char __user *s = buffer;
2317
2318         if (!data || !table->maxlen || !*lenp ||
2319             (*ppos && !write)) {
2320                 *lenp = 0;
2321                 return 0;
2322         }
2323
2324         i = (unsigned long *) data;
2325         min = (unsigned long *) table->extra1;
2326         max = (unsigned long *) table->extra2;
2327         vleft = table->maxlen / sizeof(unsigned long);
2328         left = *lenp;
2329
2330         for (; left && vleft--; i++, min++, max++, first=0) {
2331                 if (write) {
2332                         while (left) {
2333                                 char c;
2334                                 if (get_user(c, s))
2335                                         return -EFAULT;
2336                                 if (!isspace(c))
2337                                         break;
2338                                 left--;
2339                                 s++;
2340                         }
2341                         if (!left)
2342                                 break;
2343                         neg = 0;
2344                         len = left;
2345                         if (len > TMPBUFLEN-1)
2346                                 len = TMPBUFLEN-1;
2347                         if (copy_from_user(buf, s, len))
2348                                 return -EFAULT;
2349                         buf[len] = 0;
2350                         p = buf;
2351                         if (*p == '-' && left > 1) {
2352                                 neg = 1;
2353                                 p++;
2354                         }
2355                         if (*p < '0' || *p > '9')
2356                                 break;
2357                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
2358                         len = p-buf;
2359                         if ((len < left) && *p && !isspace(*p))
2360                                 break;
2361                         if (neg)
2362                                 val = -val;
2363                         s += len;
2364                         left -= len;
2365
2366                         if(neg)
2367                                 continue;
2368                         if ((min && val < *min) || (max && val > *max))
2369                                 continue;
2370                         *i = val;
2371                 } else {
2372                         p = buf;
2373                         if (!first)
2374                                 *p++ = '\t';
2375                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2376                         len = strlen(buf);
2377                         if (len > left)
2378                                 len = left;
2379                         if(copy_to_user(s, buf, len))
2380                                 return -EFAULT;
2381                         left -= len;
2382                         s += len;
2383                 }
2384         }
2385
2386         if (!write && !first && left) {
2387                 if(put_user('\n', s))
2388                         return -EFAULT;
2389                 left--, s++;
2390         }
2391         if (write) {
2392                 while (left) {
2393                         char c;
2394                         if (get_user(c, s++))
2395                                 return -EFAULT;
2396                         if (!isspace(c))
2397                                 break;
2398                         left--;
2399                 }
2400         }
2401         if (write && first)
2402                 return -EINVAL;
2403         *lenp -= left;
2404         *ppos += *lenp;
2405         return 0;
2406 #undef TMPBUFLEN
2407 }
2408
2409 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2410                                      void __user *buffer,
2411                                      size_t *lenp, loff_t *ppos,
2412                                      unsigned long convmul,
2413                                      unsigned long convdiv)
2414 {
2415         return __do_proc_doulongvec_minmax(table->data, table, write,
2416                         buffer, lenp, ppos, convmul, convdiv);
2417 }
2418
2419 /**
2420  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2421  * @table: the sysctl table
2422  * @write: %TRUE if this is a write to the sysctl file
2423  * @buffer: the user buffer
2424  * @lenp: the size of the user buffer
2425  * @ppos: file position
2426  *
2427  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2428  * values from/to the user buffer, treated as an ASCII string.
2429  *
2430  * This routine will ensure the values are within the range specified by
2431  * table->extra1 (min) and table->extra2 (max).
2432  *
2433  * Returns 0 on success.
2434  */
2435 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2436                            void __user *buffer, size_t *lenp, loff_t *ppos)
2437 {
2438     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2439 }
2440
2441 /**
2442  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2443  * @table: the sysctl table
2444  * @write: %TRUE if this is a write to the sysctl file
2445  * @buffer: the user buffer
2446  * @lenp: the size of the user buffer
2447  * @ppos: file position
2448  *
2449  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2450  * values from/to the user buffer, treated as an ASCII string. The values
2451  * are treated as milliseconds, and converted to jiffies when they are stored.
2452  *
2453  * This routine will ensure the values are within the range specified by
2454  * table->extra1 (min) and table->extra2 (max).
2455  *
2456  * Returns 0 on success.
2457  */
2458 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2459                                       void __user *buffer,
2460                                       size_t *lenp, loff_t *ppos)
2461 {
2462     return do_proc_doulongvec_minmax(table, write, buffer,
2463                                      lenp, ppos, HZ, 1000l);
2464 }
2465
2466
2467 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2468                                          int *valp,
2469                                          int write, void *data)
2470 {
2471         if (write) {
2472                 if (*lvalp > LONG_MAX / HZ)
2473                         return 1;
2474                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2475         } else {
2476                 int val = *valp;
2477                 unsigned long lval;
2478                 if (val < 0) {
2479                         *negp = -1;
2480                         lval = (unsigned long)-val;
2481                 } else {
2482                         *negp = 0;
2483                         lval = (unsigned long)val;
2484                 }
2485                 *lvalp = lval / HZ;
2486         }
2487         return 0;
2488 }
2489
2490 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2491                                                 int *valp,
2492                                                 int write, void *data)
2493 {
2494         if (write) {
2495                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2496                         return 1;
2497                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2498         } else {
2499                 int val = *valp;
2500                 unsigned long lval;
2501                 if (val < 0) {
2502                         *negp = -1;
2503                         lval = (unsigned long)-val;
2504                 } else {
2505                         *negp = 0;
2506                         lval = (unsigned long)val;
2507                 }
2508                 *lvalp = jiffies_to_clock_t(lval);
2509         }
2510         return 0;
2511 }
2512
2513 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2514                                             int *valp,
2515                                             int write, void *data)
2516 {
2517         if (write) {
2518                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2519         } else {
2520                 int val = *valp;
2521                 unsigned long lval;
2522                 if (val < 0) {
2523                         *negp = -1;
2524                         lval = (unsigned long)-val;
2525                 } else {
2526                         *negp = 0;
2527                         lval = (unsigned long)val;
2528                 }
2529                 *lvalp = jiffies_to_msecs(lval);
2530         }
2531         return 0;
2532 }
2533
2534 /**
2535  * proc_dointvec_jiffies - read a vector of integers as seconds
2536  * @table: the sysctl table
2537  * @write: %TRUE if this is a write to the sysctl file
2538  * @buffer: the user buffer
2539  * @lenp: the size of the user buffer
2540  * @ppos: file position
2541  *
2542  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2543  * values from/to the user buffer, treated as an ASCII string.
2544  * The values read are assumed to be in seconds, and are converted into
2545  * jiffies.
2546  *
2547  * Returns 0 on success.
2548  */
2549 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2550                           void __user *buffer, size_t *lenp, loff_t *ppos)
2551 {
2552     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2553                             do_proc_dointvec_jiffies_conv,NULL);
2554 }
2555
2556 /**
2557  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2558  * @table: the sysctl table
2559  * @write: %TRUE if this is a write to the sysctl file
2560  * @buffer: the user buffer
2561  * @lenp: the size of the user buffer
2562  * @ppos: pointer to the file position
2563  *
2564  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2565  * values from/to the user buffer, treated as an ASCII string.
2566  * The values read are assumed to be in 1/USER_HZ seconds, and
2567  * are converted into jiffies.
2568  *
2569  * Returns 0 on success.
2570  */
2571 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2572                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2573 {
2574     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2575                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2576 }
2577
2578 /**
2579  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2580  * @table: the sysctl table
2581  * @write: %TRUE if this is a write to the sysctl file
2582  * @buffer: the user buffer
2583  * @lenp: the size of the user buffer
2584  * @ppos: file position
2585  * @ppos: the current position in the file
2586  *
2587  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2588  * values from/to the user buffer, treated as an ASCII string.
2589  * The values read are assumed to be in 1/1000 seconds, and
2590  * are converted into jiffies.
2591  *
2592  * Returns 0 on success.
2593  */
2594 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2595                              void __user *buffer, size_t *lenp, loff_t *ppos)
2596 {
2597         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2598                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2599 }
2600
2601 static int proc_do_cad_pid(struct ctl_table *table, int write,
2602                            void __user *buffer, size_t *lenp, loff_t *ppos)
2603 {
2604         struct pid *new_pid;
2605         pid_t tmp;
2606         int r;
2607
2608         tmp = pid_vnr(cad_pid);
2609
2610         r = __do_proc_dointvec(&tmp, table, write, buffer,
2611                                lenp, ppos, NULL, NULL);
2612         if (r || !write)
2613                 return r;
2614
2615         new_pid = find_get_pid(tmp);
2616         if (!new_pid)
2617                 return -ESRCH;
2618
2619         put_pid(xchg(&cad_pid, new_pid));
2620         return 0;
2621 }
2622
2623 #else /* CONFIG_PROC_FS */
2624
2625 int proc_dostring(struct ctl_table *table, int write,
2626                   void __user *buffer, size_t *lenp, loff_t *ppos)
2627 {
2628         return -ENOSYS;
2629 }
2630
2631 int proc_dointvec(struct ctl_table *table, int write,
2632                   void __user *buffer, size_t *lenp, loff_t *ppos)
2633 {
2634         return -ENOSYS;
2635 }
2636
2637 int proc_dointvec_minmax(struct ctl_table *table, int write,
2638                     void __user *buffer, size_t *lenp, loff_t *ppos)
2639 {
2640         return -ENOSYS;
2641 }
2642
2643 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2644                     void __user *buffer, size_t *lenp, loff_t *ppos)
2645 {
2646         return -ENOSYS;
2647 }
2648
2649 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2650                     void __user *buffer, size_t *lenp, loff_t *ppos)
2651 {
2652         return -ENOSYS;
2653 }
2654
2655 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2656                              void __user *buffer, size_t *lenp, loff_t *ppos)
2657 {
2658         return -ENOSYS;
2659 }
2660
2661 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2662                     void __user *buffer, size_t *lenp, loff_t *ppos)
2663 {
2664         return -ENOSYS;
2665 }
2666
2667 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2668                                       void __user *buffer,
2669                                       size_t *lenp, loff_t *ppos)
2670 {
2671     return -ENOSYS;
2672 }
2673
2674
2675 #endif /* CONFIG_PROC_FS */
2676
2677 /*
2678  * No sense putting this after each symbol definition, twice,
2679  * exception granted :-)
2680  */
2681 EXPORT_SYMBOL(proc_dointvec);
2682 EXPORT_SYMBOL(proc_dointvec_jiffies);
2683 EXPORT_SYMBOL(proc_dointvec_minmax);
2684 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2685 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2686 EXPORT_SYMBOL(proc_dostring);
2687 EXPORT_SYMBOL(proc_doulongvec_minmax);
2688 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2689 EXPORT_SYMBOL(register_sysctl_table);
2690 EXPORT_SYMBOL(register_sysctl_paths);
2691 EXPORT_SYMBOL(unregister_sysctl_table);