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/aio.h>
  23 #include <linux/mm.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/bitmap.h>
  28 #include <linux/signal.h>
  29 #include <linux/printk.h>
  30 #include <linux/proc_fs.h>
  31 #include <linux/security.h>
  32 #include <linux/ctype.h>
  33 #include <linux/kmemleak.h>
  34 #include <linux/fs.h>
  35 #include <linux/init.h>
  36 #include <linux/kernel.h>
  37 #include <linux/kobject.h>
  38 #include <linux/net.h>
  39 #include <linux/sysrq.h>
  40 #include <linux/highuid.h>
  41 #include <linux/writeback.h>
  42 #include <linux/ratelimit.h>
  43 #include <linux/compaction.h>
  44 #include <linux/hugetlb.h>
  45 #include <linux/initrd.h>
  46 #include <linux/key.h>
  47 #include <linux/times.h>
  48 #include <linux/limits.h>
  49 #include <linux/dcache.h>
  50 #include <linux/dnotify.h>
  51 #include <linux/syscalls.h>
  52 #include <linux/vmstat.h>
  53 #include <linux/nfs_fs.h>
  54 #include <linux/acpi.h>
  55 #include <linux/reboot.h>
  56 #include <linux/ftrace.h>
  57 #include <linux/perf_event.h>
  58 #include <linux/kprobes.h>
  59 #include <linux/pipe_fs_i.h>
  60 #include <linux/oom.h>
  61 #include <linux/kmod.h>
  62 #include <linux/capability.h>
  63 #include <linux/binfmts.h>
  64 #include <linux/sched/sysctl.h>
  65 #include <linux/sched/coredump.h>
  66 #include <linux/kexec.h>
  67 #include <linux/bpf.h>
  68 #include <linux/mount.h>
  69 #include <linux/pipe_fs_i.h>
  70
  71 #include <linux/uaccess.h>
  72 #include <asm/processor.h>
  73
  74 #ifdef CONFIG_X86
  75 #include <asm/nmi.h>
  76 #include <asm/stacktrace.h>
  77 #include <asm/io.h>
  78 #endif
  79 #ifdef CONFIG_SPARC
  80 #include <asm/setup.h>
  81 #endif
  82 #ifdef CONFIG_BSD_PROCESS_ACCT
  83 #include <linux/acct.h>
  84 #endif
  85 #ifdef CONFIG_RT_MUTEXES
  86 #include <linux/rtmutex.h>
  87 #endif
  88 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  89 #include <linux/lockdep.h>
  90 #endif
  91 #ifdef CONFIG_CHR_DEV_SG
  92 #include <scsi/sg.h>
  93 #endif
  94 #ifdef CONFIG_STACKLEAK_RUNTIME_DISABLE
  95 #include <linux/stackleak.h>
  96 #endif
  97 #ifdef CONFIG_LOCKUP_DETECTOR
  98 #include <linux/nmi.h>
  99 #endif
 100
 101 #if defined(CONFIG_SYSCTL)
 102
 103 /* External variables not in a header file. */
 104 extern int suid_dumpable;
 105 #ifdef CONFIG_COREDUMP
 106 extern int core_uses_pid;
 107 extern char core_pattern[];
 108 extern unsigned int core_pipe_limit;
 109 #endif
 110 extern int pid_max;
 111 extern int pid_max_min, pid_max_max;
 112 extern int percpu_pagelist_fraction;
 113 extern int latencytop_enabled;
 114 extern unsigned int sysctl_nr_open_min, sysctl_nr_open_max;
 115 #ifndef CONFIG_MMU
 116 extern int sysctl_nr_trim_pages;
 117 #endif
 118
 119 /* Constants used for minimum and  maximum */
 120 #ifdef CONFIG_LOCKUP_DETECTOR
 121 static int sixty = 60;
 122 #endif
 123
 124 static int __maybe_unused neg_one = -1;
 125
 126 static int zero;
 127 static int __maybe_unused one = 1;
 128 static int __maybe_unused two = 2;
 129 static int __maybe_unused four = 4;
 130 static unsigned long one_ul = 1;
 131 static int one_hundred = 100;
 132 static int one_thousand = 1000;
 133 #ifdef CONFIG_PRINTK
 134 static int ten_thousand = 10000;
 135 #endif
 136 #ifdef CONFIG_PERF_EVENTS
 137 static int six_hundred_forty_kb = 640 * 1024;
 138 #endif
 139
 140 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 141 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 142
 143 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 144 static int maxolduid = 65535;
 145 static int minolduid;
 146
 147 static int ngroups_max = NGROUPS_MAX;
 148 static const int cap_last_cap = CAP_LAST_CAP;
 149
 150 /*
 151  * This is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs
 152  * and hung_task_check_interval_secs
 153  */
 154 #ifdef CONFIG_DETECT_HUNG_TASK
 155 static unsigned long hung_task_timeout_max = (LONG_MAX/HZ);
 156 #endif
 157
 158 #ifdef CONFIG_INOTIFY_USER
 159 #include <linux/inotify.h>
 160 #endif
 161 #ifdef CONFIG_SPARC
 162 #endif
 163
 164 #ifdef __hppa__
 165 extern int pwrsw_enabled;
 166 #endif
 167
 168 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 169 extern int unaligned_enabled;
 170 #endif
 171
 172 #ifdef CONFIG_IA64
 173 extern int unaligned_dump_stack;
 174 #endif
 175
 176 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 177 extern int no_unaligned_warning;
 178 #endif
 179
 180 #ifdef CONFIG_PROC_SYSCTL
 181
 182 /**
 183  * enum sysctl_writes_mode - supported sysctl write modes
 184  *
 185  * @SYSCTL_WRITES_LEGACY: each write syscall must fully contain the sysctl value
 186  *      to be written, and multiple writes on the same sysctl file descriptor
 187  *      will rewrite the sysctl value, regardless of file position. No warning
 188  *      is issued when the initial position is not 0.
 189  * @SYSCTL_WRITES_WARN: same as above but warn when the initial file position is
 190  *      not 0.
 191  * @SYSCTL_WRITES_STRICT: writes to numeric sysctl entries must always be at
 192  *      file position 0 and the value must be fully contained in the buffer
 193  *      sent to the write syscall. If dealing with strings respect the file
 194  *      position, but restrict this to the max length of the buffer, anything
 195  *      passed the max lenght will be ignored. Multiple writes will append
 196  *      to the buffer.
 197  *
 198  * These write modes control how current file position affects the behavior of
 199  * updating sysctl values through the proc interface on each write.
 200  */
 201 enum sysctl_writes_mode {
 202         SYSCTL_WRITES_LEGACY            = -1,
 203         SYSCTL_WRITES_WARN              = 0,
 204         SYSCTL_WRITES_STRICT            = 1,
 205 };
 206
 207 static enum sysctl_writes_mode sysctl_writes_strict = SYSCTL_WRITES_STRICT;
 208
 209 static int proc_do_cad_pid(struct ctl_table *table, int write,
 210                   void __user *buffer, size_t *lenp, loff_t *ppos);
 211 static int proc_taint(struct ctl_table *table, int write,
 212                                void __user *buffer, size_t *lenp, loff_t *ppos);
 213 #endif
 214
 215 #ifdef CONFIG_PRINTK
 216 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 217                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 218 #endif
 219
 220 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 221                 void __user *buffer, size_t *lenp, loff_t *ppos);
 222 #ifdef CONFIG_COREDUMP
 223 static int proc_dostring_coredump(struct ctl_table *table, int write,
 224                 void __user *buffer, size_t *lenp, loff_t *ppos);
 225 #endif
 226 static int proc_dopipe_max_size(struct ctl_table *table, int write,
 227                 void __user *buffer, size_t *lenp, loff_t *ppos);
 228
 229 #ifdef CONFIG_MAGIC_SYSRQ
 230 /* Note: sysrq code uses its own private copy */
 231 static int __sysrq_enabled = CONFIG_MAGIC_SYSRQ_DEFAULT_ENABLE;
 232
 233 static int sysrq_sysctl_handler(struct ctl_table *table, int write,
 234                                 void __user *buffer, size_t *lenp,
 235                                 loff_t *ppos)
 236 {
 237         int error;
 238
 239         error = proc_dointvec(table, write, buffer, lenp, ppos);
 240         if (error)
 241                 return error;
 242
 243         if (write)
 244                 sysrq_toggle_support(__sysrq_enabled);
 245
 246         return 0;
 247 }
 248
 249 #endif
 250
 251 static struct ctl_table kern_table[];
 252 static struct ctl_table vm_table[];
 253 static struct ctl_table fs_table[];
 254 static struct ctl_table debug_table[];
 255 static struct ctl_table dev_table[];
 256 extern struct ctl_table random_table[];
 257 #ifdef CONFIG_EPOLL
 258 extern struct ctl_table epoll_table[];
 259 #endif
 260
 261 #ifdef CONFIG_FW_LOADER_USER_HELPER
 262 extern struct ctl_table firmware_config_table[];
 263 #endif
 264
 265 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 266 int sysctl_legacy_va_layout;
 267 #endif
 268
 269 /* The default sysctl tables: */
 270
 271 static struct ctl_table sysctl_base_table[] = {
 272         {
 273                 .procname       = "kernel",
 274                 .mode           = 0555,
 275                 .child          = kern_table,
 276         },
 277         {
 278                 .procname       = "vm",
 279                 .mode           = 0555,
 280                 .child          = vm_table,
 281         },
 282         {
 283                 .procname       = "fs",
 284                 .mode           = 0555,
 285                 .child          = fs_table,
 286         },
 287         {
 288                 .procname       = "debug",
 289                 .mode           = 0555,
 290                 .child          = debug_table,
 291         },
 292         {
 293                 .procname       = "dev",
 294                 .mode           = 0555,
 295                 .child          = dev_table,
 296         },
 297         { }
 298 };
 299
 300 #ifdef CONFIG_SCHED_DEBUG
 301 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 302 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 303 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 304 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 305 #ifdef CONFIG_SMP
 306 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 307 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 308 #endif /* CONFIG_SMP */
 309 #endif /* CONFIG_SCHED_DEBUG */
 310
 311 #ifdef CONFIG_COMPACTION
 312 static int min_extfrag_threshold;
 313 static int max_extfrag_threshold = 1000;
 314 #endif
 315
 316 static struct ctl_table kern_table[] = {
 317         {
 318                 .procname       = "sched_child_runs_first",
 319                 .data           = &sysctl_sched_child_runs_first,
 320                 .maxlen         = sizeof(unsigned int),
 321                 .mode           = 0644,
 322                 .proc_handler   = proc_dointvec,
 323         },
 324 #ifdef CONFIG_SCHED_DEBUG
 325         {
 326                 .procname       = "sched_min_granularity_ns",
 327                 .data           = &sysctl_sched_min_granularity,
 328                 .maxlen         = sizeof(unsigned int),
 329                 .mode           = 0644,
 330                 .proc_handler   = sched_proc_update_handler,
 331                 .extra1         = &min_sched_granularity_ns,
 332                 .extra2         = &max_sched_granularity_ns,
 333         },
 334         {
 335                 .procname       = "sched_latency_ns",
 336                 .data           = &sysctl_sched_latency,
 337                 .maxlen         = sizeof(unsigned int),
 338                 .mode           = 0644,
 339                 .proc_handler   = sched_proc_update_handler,
 340                 .extra1         = &min_sched_granularity_ns,
 341                 .extra2         = &max_sched_granularity_ns,
 342         },
 343         {
 344                 .procname       = "sched_wakeup_granularity_ns",
 345                 .data           = &sysctl_sched_wakeup_granularity,
 346                 .maxlen         = sizeof(unsigned int),
 347                 .mode           = 0644,
 348                 .proc_handler   = sched_proc_update_handler,
 349                 .extra1         = &min_wakeup_granularity_ns,
 350                 .extra2         = &max_wakeup_granularity_ns,
 351         },
 352 #ifdef CONFIG_SMP
 353         {
 354                 .procname       = "sched_tunable_scaling",
 355                 .data           = &sysctl_sched_tunable_scaling,
 356                 .maxlen         = sizeof(enum sched_tunable_scaling),
 357                 .mode           = 0644,
 358                 .proc_handler   = sched_proc_update_handler,
 359                 .extra1         = &min_sched_tunable_scaling,
 360                 .extra2         = &max_sched_tunable_scaling,
 361         },
 362         {
 363                 .procname       = "sched_migration_cost_ns",
 364                 .data           = &sysctl_sched_migration_cost,
 365                 .maxlen         = sizeof(unsigned int),
 366                 .mode           = 0644,
 367                 .proc_handler   = proc_dointvec,
 368         },
 369         {
 370                 .procname       = "sched_nr_migrate",
 371                 .data           = &sysctl_sched_nr_migrate,
 372                 .maxlen         = sizeof(unsigned int),
 373                 .mode           = 0644,
 374                 .proc_handler   = proc_dointvec,
 375         },
 376 #ifdef CONFIG_SCHEDSTATS
 377         {
 378                 .procname       = "sched_schedstats",
 379                 .data           = NULL,
 380                 .maxlen         = sizeof(unsigned int),
 381                 .mode           = 0644,
 382                 .proc_handler   = sysctl_schedstats,
 383                 .extra1         = &zero,
 384                 .extra2         = &one,
 385         },
 386 #endif /* CONFIG_SCHEDSTATS */
 387 #endif /* CONFIG_SMP */
 388 #ifdef CONFIG_NUMA_BALANCING
 389         {
 390                 .procname       = "numa_balancing_scan_delay_ms",
 391                 .data           = &sysctl_numa_balancing_scan_delay,
 392                 .maxlen         = sizeof(unsigned int),
 393                 .mode           = 0644,
 394                 .proc_handler   = proc_dointvec,
 395         },
 396         {
 397                 .procname       = "numa_balancing_scan_period_min_ms",
 398                 .data           = &sysctl_numa_balancing_scan_period_min,
 399                 .maxlen         = sizeof(unsigned int),
 400                 .mode           = 0644,
 401                 .proc_handler   = proc_dointvec,
 402         },
 403         {
 404                 .procname       = "numa_balancing_scan_period_max_ms",
 405                 .data           = &sysctl_numa_balancing_scan_period_max,
 406                 .maxlen         = sizeof(unsigned int),
 407                 .mode           = 0644,
 408                 .proc_handler   = proc_dointvec,
 409         },
 410         {
 411                 .procname       = "numa_balancing_scan_size_mb",
 412                 .data           = &sysctl_numa_balancing_scan_size,
 413                 .maxlen         = sizeof(unsigned int),
 414                 .mode           = 0644,
 415                 .proc_handler   = proc_dointvec_minmax,
 416                 .extra1         = &one,
 417         },
 418         {
 419                 .procname       = "numa_balancing",
 420                 .data           = NULL, /* filled in by handler */
 421                 .maxlen         = sizeof(unsigned int),
 422                 .mode           = 0644,
 423                 .proc_handler   = sysctl_numa_balancing,
 424                 .extra1         = &zero,
 425                 .extra2         = &one,
 426         },
 427 #endif /* CONFIG_NUMA_BALANCING */
 428 #endif /* CONFIG_SCHED_DEBUG */
 429         {
 430                 .procname       = "sched_rt_period_us",
 431                 .data           = &sysctl_sched_rt_period,
 432                 .maxlen         = sizeof(unsigned int),
 433                 .mode           = 0644,
 434                 .proc_handler   = sched_rt_handler,
 435         },
 436         {
 437                 .procname       = "sched_rt_runtime_us",
 438                 .data           = &sysctl_sched_rt_runtime,
 439                 .maxlen         = sizeof(int),
 440                 .mode           = 0644,
 441                 .proc_handler   = sched_rt_handler,
 442         },
 443         {
 444                 .procname       = "sched_rr_timeslice_ms",
 445                 .data           = &sysctl_sched_rr_timeslice,
 446                 .maxlen         = sizeof(int),
 447                 .mode           = 0644,
 448                 .proc_handler   = sched_rr_handler,
 449         },
 450 #ifdef CONFIG_SCHED_AUTOGROUP
 451         {
 452                 .procname       = "sched_autogroup_enabled",
 453                 .data           = &sysctl_sched_autogroup_enabled,
 454                 .maxlen         = sizeof(unsigned int),
 455                 .mode           = 0644,
 456                 .proc_handler   = proc_dointvec_minmax,
 457                 .extra1         = &zero,
 458                 .extra2         = &one,
 459         },
 460 #endif
 461 #ifdef CONFIG_CFS_BANDWIDTH
 462         {
 463                 .procname       = "sched_cfs_bandwidth_slice_us",
 464                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 465                 .maxlen         = sizeof(unsigned int),
 466                 .mode           = 0644,
 467                 .proc_handler   = proc_dointvec_minmax,
 468                 .extra1         = &one,
 469         },
 470 #endif
 471 #ifdef CONFIG_PROVE_LOCKING
 472         {
 473                 .procname       = "prove_locking",
 474                 .data           = &prove_locking,
 475                 .maxlen         = sizeof(int),
 476                 .mode           = 0644,
 477                 .proc_handler   = proc_dointvec,
 478         },
 479 #endif
 480 #ifdef CONFIG_LOCK_STAT
 481         {
 482                 .procname       = "lock_stat",
 483                 .data           = &lock_stat,
 484                 .maxlen         = sizeof(int),
 485                 .mode           = 0644,
 486                 .proc_handler   = proc_dointvec,
 487         },
 488 #endif
 489         {
 490                 .procname       = "panic",
 491                 .data           = &panic_timeout,
 492                 .maxlen         = sizeof(int),
 493                 .mode           = 0644,
 494                 .proc_handler   = proc_dointvec,
 495         },
 496 #ifdef CONFIG_COREDUMP
 497         {
 498                 .procname       = "core_uses_pid",
 499                 .data           = &core_uses_pid,
 500                 .maxlen         = sizeof(int),
 501                 .mode           = 0644,
 502                 .proc_handler   = proc_dointvec,
 503         },
 504         {
 505                 .procname       = "core_pattern",
 506                 .data           = core_pattern,
 507                 .maxlen         = CORENAME_MAX_SIZE,
 508                 .mode           = 0644,
 509                 .proc_handler   = proc_dostring_coredump,
 510         },
 511         {
 512                 .procname       = "core_pipe_limit",
 513                 .data           = &core_pipe_limit,
 514                 .maxlen         = sizeof(unsigned int),
 515                 .mode           = 0644,
 516                 .proc_handler   = proc_dointvec,
 517         },
 518 #endif
 519 #ifdef CONFIG_PROC_SYSCTL
 520         {
 521                 .procname       = "tainted",
 522                 .maxlen         = sizeof(long),
 523                 .mode           = 0644,
 524                 .proc_handler   = proc_taint,
 525         },
 526         {
 527                 .procname       = "sysctl_writes_strict",
 528                 .data           = &sysctl_writes_strict,
 529                 .maxlen         = sizeof(int),
 530                 .mode           = 0644,
 531                 .proc_handler   = proc_dointvec_minmax,
 532                 .extra1         = &neg_one,
 533                 .extra2         = &one,
 534         },
 535 #endif
 536 #ifdef CONFIG_LATENCYTOP
 537         {
 538                 .procname       = "latencytop",
 539                 .data           = &latencytop_enabled,
 540                 .maxlen         = sizeof(int),
 541                 .mode           = 0644,
 542                 .proc_handler   = sysctl_latencytop,
 543         },
 544 #endif
 545 #ifdef CONFIG_BLK_DEV_INITRD
 546         {
 547                 .procname       = "real-root-dev",
 548                 .data           = &real_root_dev,
 549                 .maxlen         = sizeof(int),
 550                 .mode           = 0644,
 551                 .proc_handler   = proc_dointvec,
 552         },
 553 #endif
 554         {
 555                 .procname       = "print-fatal-signals",
 556                 .data           = &print_fatal_signals,
 557                 .maxlen         = sizeof(int),
 558                 .mode           = 0644,
 559                 .proc_handler   = proc_dointvec,
 560         },
 561 #ifdef CONFIG_SPARC
 562         {
 563                 .procname       = "reboot-cmd",
 564                 .data           = reboot_command,
 565                 .maxlen         = 256,
 566                 .mode           = 0644,
 567                 .proc_handler   = proc_dostring,
 568         },
 569         {
 570                 .procname       = "stop-a",
 571                 .data           = &stop_a_enabled,
 572                 .maxlen         = sizeof (int),
 573                 .mode           = 0644,
 574                 .proc_handler   = proc_dointvec,
 575         },
 576         {
 577                 .procname       = "scons-poweroff",
 578                 .data           = &scons_pwroff,
 579                 .maxlen         = sizeof (int),
 580                 .mode           = 0644,
 581                 .proc_handler   = proc_dointvec,
 582         },
 583 #endif
 584 #ifdef CONFIG_SPARC64
 585         {
 586                 .procname       = "tsb-ratio",
 587                 .data           = &sysctl_tsb_ratio,
 588                 .maxlen         = sizeof (int),
 589                 .mode           = 0644,
 590                 .proc_handler   = proc_dointvec,
 591         },
 592 #endif
 593 #ifdef __hppa__
 594         {
 595                 .procname       = "soft-power",
 596                 .data           = &pwrsw_enabled,
 597                 .maxlen         = sizeof (int),
 598                 .mode           = 0644,
 599                 .proc_handler   = proc_dointvec,
 600         },
 601 #endif
 602 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 603         {
 604                 .procname       = "unaligned-trap",
 605                 .data           = &unaligned_enabled,
 606                 .maxlen         = sizeof (int),
 607                 .mode           = 0644,
 608                 .proc_handler   = proc_dointvec,
 609         },
 610 #endif
 611         {
 612                 .procname       = "ctrl-alt-del",
 613                 .data           = &C_A_D,
 614                 .maxlen         = sizeof(int),
 615                 .mode           = 0644,
 616                 .proc_handler   = proc_dointvec,
 617         },
 618 #ifdef CONFIG_FUNCTION_TRACER
 619         {
 620                 .procname       = "ftrace_enabled",
 621                 .data           = &ftrace_enabled,
 622                 .maxlen         = sizeof(int),
 623                 .mode           = 0644,
 624                 .proc_handler   = ftrace_enable_sysctl,
 625         },
 626 #endif
 627 #ifdef CONFIG_STACK_TRACER
 628         {
 629                 .procname       = "stack_tracer_enabled",
 630                 .data           = &stack_tracer_enabled,
 631                 .maxlen         = sizeof(int),
 632                 .mode           = 0644,
 633                 .proc_handler   = stack_trace_sysctl,
 634         },
 635 #endif
 636 #ifdef CONFIG_TRACING
 637         {
 638                 .procname       = "ftrace_dump_on_oops",
 639                 .data           = &ftrace_dump_on_oops,
 640                 .maxlen         = sizeof(int),
 641                 .mode           = 0644,
 642                 .proc_handler   = proc_dointvec,
 643         },
 644         {
 645                 .procname       = "traceoff_on_warning",
 646                 .data           = &__disable_trace_on_warning,
 647                 .maxlen         = sizeof(__disable_trace_on_warning),
 648                 .mode           = 0644,
 649                 .proc_handler   = proc_dointvec,
 650         },
 651         {
 652                 .procname       = "tracepoint_printk",
 653                 .data           = &tracepoint_printk,
 654                 .maxlen         = sizeof(tracepoint_printk),
 655                 .mode           = 0644,
 656                 .proc_handler   = tracepoint_printk_sysctl,
 657         },
 658 #endif
 659 #ifdef CONFIG_KEXEC_CORE
 660         {
 661                 .procname       = "kexec_load_disabled",
 662                 .data           = &kexec_load_disabled,
 663                 .maxlen         = sizeof(int),
 664                 .mode           = 0644,
 665                 /* only handle a transition from default "0" to "1" */
 666                 .proc_handler   = proc_dointvec_minmax,
 667                 .extra1         = &one,
 668                 .extra2         = &one,
 669         },
 670 #endif
 671 #ifdef CONFIG_MODULES
 672         {
 673                 .procname       = "modprobe",
 674                 .data           = &modprobe_path,
 675                 .maxlen         = KMOD_PATH_LEN,
 676                 .mode           = 0644,
 677                 .proc_handler   = proc_dostring,
 678         },
 679         {
 680                 .procname       = "modules_disabled",
 681                 .data           = &modules_disabled,
 682                 .maxlen         = sizeof(int),
 683                 .mode           = 0644,
 684                 /* only handle a transition from default "0" to "1" */
 685                 .proc_handler   = proc_dointvec_minmax,
 686                 .extra1         = &one,
 687                 .extra2         = &one,
 688         },
 689 #endif
 690 #ifdef CONFIG_UEVENT_HELPER
 691         {
 692                 .procname       = "hotplug",
 693                 .data           = &uevent_helper,
 694                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 695                 .mode           = 0644,
 696                 .proc_handler   = proc_dostring,
 697         },
 698 #endif
 699 #ifdef CONFIG_CHR_DEV_SG
 700         {
 701                 .procname       = "sg-big-buff",
 702                 .data           = &sg_big_buff,
 703                 .maxlen         = sizeof (int),
 704                 .mode           = 0444,
 705                 .proc_handler   = proc_dointvec,
 706         },
 707 #endif
 708 #ifdef CONFIG_BSD_PROCESS_ACCT
 709         {
 710                 .procname       = "acct",
 711                 .data           = &acct_parm,
 712                 .maxlen         = 3*sizeof(int),
 713                 .mode           = 0644,
 714                 .proc_handler   = proc_dointvec,
 715         },
 716 #endif
 717 #ifdef CONFIG_MAGIC_SYSRQ
 718         {
 719                 .procname       = "sysrq",
 720                 .data           = &__sysrq_enabled,
 721                 .maxlen         = sizeof (int),
 722                 .mode           = 0644,
 723                 .proc_handler   = sysrq_sysctl_handler,
 724         },
 725 #endif
 726 #ifdef CONFIG_PROC_SYSCTL
 727         {
 728                 .procname       = "cad_pid",
 729                 .data           = NULL,
 730                 .maxlen         = sizeof (int),
 731                 .mode           = 0600,
 732                 .proc_handler   = proc_do_cad_pid,
 733         },
 734 #endif
 735         {
 736                 .procname       = "threads-max",
 737                 .data           = NULL,
 738                 .maxlen         = sizeof(int),
 739                 .mode           = 0644,
 740                 .proc_handler   = sysctl_max_threads,
 741         },
 742         {
 743                 .procname       = "random",
 744                 .mode           = 0555,
 745                 .child          = random_table,
 746         },
 747         {
 748                 .procname       = "usermodehelper",
 749                 .mode           = 0555,
 750                 .child          = usermodehelper_table,
 751         },
 752 #ifdef CONFIG_FW_LOADER_USER_HELPER
 753         {
 754                 .procname       = "firmware_config",
 755                 .mode           = 0555,
 756                 .child          = firmware_config_table,
 757         },
 758 #endif
 759         {
 760                 .procname       = "overflowuid",
 761                 .data           = &overflowuid,
 762                 .maxlen         = sizeof(int),
 763                 .mode           = 0644,
 764                 .proc_handler   = proc_dointvec_minmax,
 765                 .extra1         = &minolduid,
 766                 .extra2         = &maxolduid,
 767         },
 768         {
 769                 .procname       = "overflowgid",
 770                 .data           = &overflowgid,
 771                 .maxlen         = sizeof(int),
 772                 .mode           = 0644,
 773                 .proc_handler   = proc_dointvec_minmax,
 774                 .extra1         = &minolduid,
 775                 .extra2         = &maxolduid,
 776         },
 777 #ifdef CONFIG_S390
 778 #ifdef CONFIG_MATHEMU
 779         {
 780                 .procname       = "ieee_emulation_warnings",
 781                 .data           = &sysctl_ieee_emulation_warnings,
 782                 .maxlen         = sizeof(int),
 783                 .mode           = 0644,
 784                 .proc_handler   = proc_dointvec,
 785         },
 786 #endif
 787         {
 788                 .procname       = "userprocess_debug",
 789                 .data           = &show_unhandled_signals,
 790                 .maxlen         = sizeof(int),
 791                 .mode           = 0644,
 792                 .proc_handler   = proc_dointvec,
 793         },
 794 #endif
 795         {
 796                 .procname       = "pid_max",
 797                 .data           = &pid_max,
 798                 .maxlen         = sizeof (int),
 799                 .mode           = 0644,
 800                 .proc_handler   = proc_dointvec_minmax,
 801                 .extra1         = &pid_max_min,
 802                 .extra2         = &pid_max_max,
 803         },
 804         {
 805                 .procname       = "panic_on_oops",
 806                 .data           = &panic_on_oops,
 807                 .maxlen         = sizeof(int),
 808                 .mode           = 0644,
 809                 .proc_handler   = proc_dointvec,
 810         },
 811 #if defined CONFIG_PRINTK
 812         {
 813                 .procname       = "printk",
 814                 .data           = &console_loglevel,
 815                 .maxlen         = 4*sizeof(int),
 816                 .mode           = 0644,
 817                 .proc_handler   = proc_dointvec,
 818         },
 819         {
 820                 .procname       = "printk_ratelimit",
 821                 .data           = &printk_ratelimit_state.interval,
 822                 .maxlen         = sizeof(int),
 823                 .mode           = 0644,
 824                 .proc_handler   = proc_dointvec_jiffies,
 825         },
 826         {
 827                 .procname       = "printk_ratelimit_burst",
 828                 .data           = &printk_ratelimit_state.burst,
 829                 .maxlen         = sizeof(int),
 830                 .mode           = 0644,
 831                 .proc_handler   = proc_dointvec,
 832         },
 833         {
 834                 .procname       = "printk_delay",
 835                 .data           = &printk_delay_msec,
 836                 .maxlen         = sizeof(int),
 837                 .mode           = 0644,
 838                 .proc_handler   = proc_dointvec_minmax,
 839                 .extra1         = &zero,
 840                 .extra2         = &ten_thousand,
 841         },
 842         {
 843                 .procname       = "printk_devkmsg",
 844                 .data           = devkmsg_log_str,
 845                 .maxlen         = DEVKMSG_STR_MAX_SIZE,
 846                 .mode           = 0644,
 847                 .proc_handler   = devkmsg_sysctl_set_loglvl,
 848         },
 849         {
 850                 .procname       = "dmesg_restrict",
 851                 .data           = &dmesg_restrict,
 852                 .maxlen         = sizeof(int),
 853                 .mode           = 0644,
 854                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 855                 .extra1         = &zero,
 856                 .extra2         = &one,
 857         },
 858         {
 859                 .procname       = "kptr_restrict",
 860                 .data           = &kptr_restrict,
 861                 .maxlen         = sizeof(int),
 862                 .mode           = 0644,
 863                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 864                 .extra1         = &zero,
 865                 .extra2         = &two,
 866         },
 867 #endif
 868         {
 869                 .procname       = "ngroups_max",
 870                 .data           = &ngroups_max,
 871                 .maxlen         = sizeof (int),
 872                 .mode           = 0444,
 873                 .proc_handler   = proc_dointvec,
 874         },
 875         {
 876                 .procname       = "cap_last_cap",
 877                 .data           = (void *)&cap_last_cap,
 878                 .maxlen         = sizeof(int),
 879                 .mode           = 0444,
 880                 .proc_handler   = proc_dointvec,
 881         },
 882 #if defined(CONFIG_LOCKUP_DETECTOR)
 883         {
 884                 .procname       = "watchdog",
 885                 .data           = &watchdog_user_enabled,
 886                 .maxlen         = sizeof(int),
 887                 .mode           = 0644,
 888                 .proc_handler   = proc_watchdog,
 889                 .extra1         = &zero,
 890                 .extra2         = &one,
 891         },
 892         {
 893                 .procname       = "watchdog_thresh",
 894                 .data           = &watchdog_thresh,
 895                 .maxlen         = sizeof(int),
 896                 .mode           = 0644,
 897                 .proc_handler   = proc_watchdog_thresh,
 898                 .extra1         = &zero,
 899                 .extra2         = &sixty,
 900         },
 901         {
 902                 .procname       = "nmi_watchdog",
 903                 .data           = &nmi_watchdog_user_enabled,
 904                 .maxlen         = sizeof(int),
 905                 .mode           = NMI_WATCHDOG_SYSCTL_PERM,
 906                 .proc_handler   = proc_nmi_watchdog,
 907                 .extra1         = &zero,
 908                 .extra2         = &one,
 909         },
 910         {
 911                 .procname       = "watchdog_cpumask",
 912                 .data           = &watchdog_cpumask_bits,
 913                 .maxlen         = NR_CPUS,
 914                 .mode           = 0644,
 915                 .proc_handler   = proc_watchdog_cpumask,
 916         },
 917 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
 918         {
 919                 .procname       = "soft_watchdog",
 920                 .data           = &soft_watchdog_user_enabled,
 921                 .maxlen         = sizeof(int),
 922                 .mode           = 0644,
 923                 .proc_handler   = proc_soft_watchdog,
 924                 .extra1         = &zero,
 925                 .extra2         = &one,
 926         },
 927         {
 928                 .procname       = "softlockup_panic",
 929                 .data           = &softlockup_panic,
 930                 .maxlen         = sizeof(int),
 931                 .mode           = 0644,
 932                 .proc_handler   = proc_dointvec_minmax,
 933                 .extra1         = &zero,
 934                 .extra2         = &one,
 935         },
 936 #ifdef CONFIG_SMP
 937         {
 938                 .procname       = "softlockup_all_cpu_backtrace",
 939                 .data           = &sysctl_softlockup_all_cpu_backtrace,
 940                 .maxlen         = sizeof(int),
 941                 .mode           = 0644,
 942                 .proc_handler   = proc_dointvec_minmax,
 943                 .extra1         = &zero,
 944                 .extra2         = &one,
 945         },
 946 #endif /* CONFIG_SMP */
 947 #endif
 948 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 949         {
 950                 .procname       = "hardlockup_panic",
 951                 .data           = &hardlockup_panic,
 952                 .maxlen         = sizeof(int),
 953                 .mode           = 0644,
 954                 .proc_handler   = proc_dointvec_minmax,
 955                 .extra1         = &zero,
 956                 .extra2         = &one,
 957         },
 958 #ifdef CONFIG_SMP
 959         {
 960                 .procname       = "hardlockup_all_cpu_backtrace",
 961                 .data           = &sysctl_hardlockup_all_cpu_backtrace,
 962                 .maxlen         = sizeof(int),
 963                 .mode           = 0644,
 964                 .proc_handler   = proc_dointvec_minmax,
 965                 .extra1         = &zero,
 966                 .extra2         = &one,
 967         },
 968 #endif /* CONFIG_SMP */
 969 #endif
 970 #endif
 971
 972 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 973         {
 974                 .procname       = "unknown_nmi_panic",
 975                 .data           = &unknown_nmi_panic,
 976                 .maxlen         = sizeof (int),
 977                 .mode           = 0644,
 978                 .proc_handler   = proc_dointvec,
 979         },
 980 #endif
 981 #if defined(CONFIG_X86)
 982         {
 983                 .procname       = "panic_on_unrecovered_nmi",
 984                 .data           = &panic_on_unrecovered_nmi,
 985                 .maxlen         = sizeof(int),
 986                 .mode           = 0644,
 987                 .proc_handler   = proc_dointvec,
 988         },
 989         {
 990                 .procname       = "panic_on_io_nmi",
 991                 .data           = &panic_on_io_nmi,
 992                 .maxlen         = sizeof(int),
 993                 .mode           = 0644,
 994                 .proc_handler   = proc_dointvec,
 995         },
 996 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 997         {
 998                 .procname       = "panic_on_stackoverflow",
 999                 .data           = &sysctl_panic_on_stackoverflow,
1000                 .maxlen         = sizeof(int),
1001                 .mode           = 0644,
1002                 .proc_handler   = proc_dointvec,
1003         },
1004 #endif
1005         {
1006                 .procname       = "bootloader_type",
1007                 .data           = &bootloader_type,
1008                 .maxlen         = sizeof (int),
1009                 .mode           = 0444,
1010                 .proc_handler   = proc_dointvec,
1011         },
1012         {
1013                 .procname       = "bootloader_version",
1014                 .data           = &bootloader_version,
1015                 .maxlen         = sizeof (int),
1016                 .mode           = 0444,
1017                 .proc_handler   = proc_dointvec,
1018         },
1019         {
1020                 .procname       = "io_delay_type",
1021                 .data           = &io_delay_type,
1022                 .maxlen         = sizeof(int),
1023                 .mode           = 0644,
1024                 .proc_handler   = proc_dointvec,
1025         },
1026 #endif
1027 #if defined(CONFIG_MMU)
1028         {
1029                 .procname       = "randomize_va_space",
1030                 .data           = &randomize_va_space,
1031                 .maxlen         = sizeof(int),
1032                 .mode           = 0644,
1033                 .proc_handler   = proc_dointvec,
1034         },
1035 #endif
1036 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
1037         {
1038                 .procname       = "spin_retry",
1039                 .data           = &spin_retry,
1040                 .maxlen         = sizeof (int),
1041                 .mode           = 0644,
1042                 .proc_handler   = proc_dointvec,
1043         },
1044 #endif
1045 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
1046         {
1047                 .procname       = "acpi_video_flags",
1048                 .data           = &acpi_realmode_flags,
1049                 .maxlen         = sizeof (unsigned long),
1050                 .mode           = 0644,
1051                 .proc_handler   = proc_doulongvec_minmax,
1052         },
1053 #endif
1054 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
1055         {
1056                 .procname       = "ignore-unaligned-usertrap",
1057                 .data           = &no_unaligned_warning,
1058                 .maxlen         = sizeof (int),
1059                 .mode           = 0644,
1060                 .proc_handler   = proc_dointvec,
1061         },
1062 #endif
1063 #ifdef CONFIG_IA64
1064         {
1065                 .procname       = "unaligned-dump-stack",
1066                 .data           = &unaligned_dump_stack,
1067                 .maxlen         = sizeof (int),
1068                 .mode           = 0644,
1069                 .proc_handler   = proc_dointvec,
1070         },
1071 #endif
1072 #ifdef CONFIG_DETECT_HUNG_TASK
1073         {
1074                 .procname       = "hung_task_panic",
1075                 .data           = &sysctl_hung_task_panic,
1076                 .maxlen         = sizeof(int),
1077                 .mode           = 0644,
1078                 .proc_handler   = proc_dointvec_minmax,
1079                 .extra1         = &zero,
1080                 .extra2         = &one,
1081         },
1082         {
1083                 .procname       = "hung_task_check_count",
1084                 .data           = &sysctl_hung_task_check_count,
1085                 .maxlen         = sizeof(int),
1086                 .mode           = 0644,
1087                 .proc_handler   = proc_dointvec_minmax,
1088                 .extra1         = &zero,
1089         },
1090         {
1091                 .procname       = "hung_task_timeout_secs",
1092                 .data           = &sysctl_hung_task_timeout_secs,
1093                 .maxlen         = sizeof(unsigned long),
1094                 .mode           = 0644,
1095                 .proc_handler   = proc_dohung_task_timeout_secs,
1096                 .extra2         = &hung_task_timeout_max,
1097         },
1098         {
1099                 .procname       = "hung_task_check_interval_secs",
1100                 .data           = &sysctl_hung_task_check_interval_secs,
1101                 .maxlen         = sizeof(unsigned long),
1102                 .mode           = 0644,
1103                 .proc_handler   = proc_dohung_task_timeout_secs,
1104                 .extra2         = &hung_task_timeout_max,
1105         },
1106         {
1107                 .procname       = "hung_task_warnings",
1108                 .data           = &sysctl_hung_task_warnings,
1109                 .maxlen         = sizeof(int),
1110                 .mode           = 0644,
1111                 .proc_handler   = proc_dointvec_minmax,
1112                 .extra1         = &neg_one,
1113         },
1114 #endif
1115 #ifdef CONFIG_RT_MUTEXES
1116         {
1117                 .procname       = "max_lock_depth",
1118                 .data           = &max_lock_depth,
1119                 .maxlen         = sizeof(int),
1120                 .mode           = 0644,
1121                 .proc_handler   = proc_dointvec,
1122         },
1123 #endif
1124         {
1125                 .procname       = "poweroff_cmd",
1126                 .data           = &poweroff_cmd,
1127                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1128                 .mode           = 0644,
1129                 .proc_handler   = proc_dostring,
1130         },
1131 #ifdef CONFIG_KEYS
1132         {
1133                 .procname       = "keys",
1134                 .mode           = 0555,
1135                 .child          = key_sysctls,
1136         },
1137 #endif
1138 #ifdef CONFIG_PERF_EVENTS
1139         /*
1140          * User-space scripts rely on the existence of this file
1141          * as a feature check for perf_events being enabled.
1142          *
1143          * So it's an ABI, do not remove!
1144          */
1145         {
1146                 .procname       = "perf_event_paranoid",
1147                 .data           = &sysctl_perf_event_paranoid,
1148                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1149                 .mode           = 0644,
1150                 .proc_handler   = proc_dointvec,
1151         },
1152         {
1153                 .procname       = "perf_event_mlock_kb",
1154                 .data           = &sysctl_perf_event_mlock,
1155                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1156                 .mode           = 0644,
1157                 .proc_handler   = proc_dointvec,
1158         },
1159         {
1160                 .procname       = "perf_event_max_sample_rate",
1161                 .data           = &sysctl_perf_event_sample_rate,
1162                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1163                 .mode           = 0644,
1164                 .proc_handler   = perf_proc_update_handler,
1165                 .extra1         = &one,
1166         },
1167         {
1168                 .procname       = "perf_cpu_time_max_percent",
1169                 .data           = &sysctl_perf_cpu_time_max_percent,
1170                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1171                 .mode           = 0644,
1172                 .proc_handler   = perf_cpu_time_max_percent_handler,
1173                 .extra1         = &zero,
1174                 .extra2         = &one_hundred,
1175         },
1176         {
1177                 .procname       = "perf_event_max_stack",
1178                 .data           = &sysctl_perf_event_max_stack,
1179                 .maxlen         = sizeof(sysctl_perf_event_max_stack),
1180                 .mode           = 0644,
1181                 .proc_handler   = perf_event_max_stack_handler,
1182                 .extra1         = &zero,
1183                 .extra2         = &six_hundred_forty_kb,
1184         },
1185         {
1186                 .procname       = "perf_event_max_contexts_per_stack",
1187                 .data           = &sysctl_perf_event_max_contexts_per_stack,
1188                 .maxlen         = sizeof(sysctl_perf_event_max_contexts_per_stack),
1189                 .mode           = 0644,
1190                 .proc_handler   = perf_event_max_stack_handler,
1191                 .extra1         = &zero,
1192                 .extra2         = &one_thousand,
1193         },
1194 #endif
1195         {
1196                 .procname       = "panic_on_warn",
1197                 .data           = &panic_on_warn,
1198                 .maxlen         = sizeof(int),
1199                 .mode           = 0644,
1200                 .proc_handler   = proc_dointvec_minmax,
1201                 .extra1         = &zero,
1202                 .extra2         = &one,
1203         },
1204 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
1205         {
1206                 .procname       = "timer_migration",
1207                 .data           = &sysctl_timer_migration,
1208                 .maxlen         = sizeof(unsigned int),
1209                 .mode           = 0644,
1210                 .proc_handler   = timer_migration_handler,
1211                 .extra1         = &zero,
1212                 .extra2         = &one,
1213         },
1214 #endif
1215 #ifdef CONFIG_BPF_SYSCALL
1216         {
1217                 .procname       = "unprivileged_bpf_disabled",
1218                 .data           = &sysctl_unprivileged_bpf_disabled,
1219                 .maxlen         = sizeof(sysctl_unprivileged_bpf_disabled),
1220                 .mode           = 0644,
1221                 /* only handle a transition from default "0" to "1" */
1222                 .proc_handler   = proc_dointvec_minmax,
1223                 .extra1         = &one,
1224                 .extra2         = &one,
1225         },
1226 #endif
1227 #if defined(CONFIG_TREE_RCU) || defined(CONFIG_PREEMPT_RCU)
1228         {
1229                 .procname       = "panic_on_rcu_stall",
1230                 .data           = &sysctl_panic_on_rcu_stall,
1231                 .maxlen         = sizeof(sysctl_panic_on_rcu_stall),
1232                 .mode           = 0644,
1233                 .proc_handler   = proc_dointvec_minmax,
1234                 .extra1         = &zero,
1235                 .extra2         = &one,
1236         },
1237 #endif
1238 #ifdef CONFIG_STACKLEAK_RUNTIME_DISABLE
1239         {
1240                 .procname       = "stack_erasing",
1241                 .data           = NULL,
1242                 .maxlen         = sizeof(int),
1243                 .mode           = 0600,
1244                 .proc_handler   = stack_erasing_sysctl,
1245                 .extra1         = &zero,
1246                 .extra2         = &one,
1247         },
1248 #endif
1249         { }
1250 };
1251
1252 static struct ctl_table vm_table[] = {
1253         {
1254                 .procname       = "overcommit_memory",
1255                 .data           = &sysctl_overcommit_memory,
1256                 .maxlen         = sizeof(sysctl_overcommit_memory),
1257                 .mode           = 0644,
1258                 .proc_handler   = proc_dointvec_minmax,
1259                 .extra1         = &zero,
1260                 .extra2         = &two,
1261         },
1262         {
1263                 .procname       = "panic_on_oom",
1264                 .data           = &sysctl_panic_on_oom,
1265                 .maxlen         = sizeof(sysctl_panic_on_oom),
1266                 .mode           = 0644,
1267                 .proc_handler   = proc_dointvec_minmax,
1268                 .extra1         = &zero,
1269                 .extra2         = &two,
1270         },
1271         {
1272                 .procname       = "oom_kill_allocating_task",
1273                 .data           = &sysctl_oom_kill_allocating_task,
1274                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1275                 .mode           = 0644,
1276                 .proc_handler   = proc_dointvec,
1277         },
1278         {
1279                 .procname       = "oom_dump_tasks",
1280                 .data           = &sysctl_oom_dump_tasks,
1281                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1282                 .mode           = 0644,
1283                 .proc_handler   = proc_dointvec,
1284         },
1285         {
1286                 .procname       = "overcommit_ratio",
1287                 .data           = &sysctl_overcommit_ratio,
1288                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1289                 .mode           = 0644,
1290                 .proc_handler   = overcommit_ratio_handler,
1291         },
1292         {
1293                 .procname       = "overcommit_kbytes",
1294                 .data           = &sysctl_overcommit_kbytes,
1295                 .maxlen         = sizeof(sysctl_overcommit_kbytes),
1296                 .mode           = 0644,
1297                 .proc_handler   = overcommit_kbytes_handler,
1298         },
1299         {
1300                 .procname       = "page-cluster",
1301                 .data           = &page_cluster,
1302                 .maxlen         = sizeof(int),
1303                 .mode           = 0644,
1304                 .proc_handler   = proc_dointvec_minmax,
1305                 .extra1         = &zero,
1306         },
1307         {
1308                 .procname       = "dirty_background_ratio",
1309                 .data           = &dirty_background_ratio,
1310                 .maxlen         = sizeof(dirty_background_ratio),
1311                 .mode           = 0644,
1312                 .proc_handler   = dirty_background_ratio_handler,
1313                 .extra1         = &zero,
1314                 .extra2         = &one_hundred,
1315         },
1316         {
1317                 .procname       = "dirty_background_bytes",
1318                 .data           = &dirty_background_bytes,
1319                 .maxlen         = sizeof(dirty_background_bytes),
1320                 .mode           = 0644,
1321                 .proc_handler   = dirty_background_bytes_handler,
1322                 .extra1         = &one_ul,
1323         },
1324         {
1325                 .procname       = "dirty_ratio",
1326                 .data           = &vm_dirty_ratio,
1327                 .maxlen         = sizeof(vm_dirty_ratio),
1328                 .mode           = 0644,
1329                 .proc_handler   = dirty_ratio_handler,
1330                 .extra1         = &zero,
1331                 .extra2         = &one_hundred,
1332         },
1333         {
1334                 .procname       = "dirty_bytes",
1335                 .data           = &vm_dirty_bytes,
1336                 .maxlen         = sizeof(vm_dirty_bytes),
1337                 .mode           = 0644,
1338                 .proc_handler   = dirty_bytes_handler,
1339                 .extra1         = &dirty_bytes_min,
1340         },
1341         {
1342                 .procname       = "dirty_writeback_centisecs",
1343                 .data           = &dirty_writeback_interval,
1344                 .maxlen         = sizeof(dirty_writeback_interval),
1345                 .mode           = 0644,
1346                 .proc_handler   = dirty_writeback_centisecs_handler,
1347         },
1348         {
1349                 .procname       = "dirty_expire_centisecs",
1350                 .data           = &dirty_expire_interval,
1351                 .maxlen         = sizeof(dirty_expire_interval),
1352                 .mode           = 0644,
1353                 .proc_handler   = proc_dointvec_minmax,
1354                 .extra1         = &zero,
1355         },
1356         {
1357                 .procname       = "dirtytime_expire_seconds",
1358                 .data           = &dirtytime_expire_interval,
1359                 .maxlen         = sizeof(dirtytime_expire_interval),
1360                 .mode           = 0644,
1361                 .proc_handler   = dirtytime_interval_handler,
1362                 .extra1         = &zero,
1363         },
1364         {
1365                 .procname       = "swappiness",
1366                 .data           = &vm_swappiness,
1367                 .maxlen         = sizeof(vm_swappiness),
1368                 .mode           = 0644,
1369                 .proc_handler   = proc_dointvec_minmax,
1370                 .extra1         = &zero,
1371                 .extra2         = &one_hundred,
1372         },
1373 #ifdef CONFIG_HUGETLB_PAGE
1374         {
1375                 .procname       = "nr_hugepages",
1376                 .data           = NULL,
1377                 .maxlen         = sizeof(unsigned long),
1378                 .mode           = 0644,
1379                 .proc_handler   = hugetlb_sysctl_handler,
1380         },
1381 #ifdef CONFIG_NUMA
1382         {
1383                 .procname       = "nr_hugepages_mempolicy",
1384                 .data           = NULL,
1385                 .maxlen         = sizeof(unsigned long),
1386                 .mode           = 0644,
1387                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1388         },
1389         {
1390                 .procname               = "numa_stat",
1391                 .data                   = &sysctl_vm_numa_stat,
1392                 .maxlen                 = sizeof(int),
1393                 .mode                   = 0644,
1394                 .proc_handler   = sysctl_vm_numa_stat_handler,
1395                 .extra1                 = &zero,
1396                 .extra2                 = &one,
1397         },
1398 #endif
1399          {
1400                 .procname       = "hugetlb_shm_group",
1401                 .data           = &sysctl_hugetlb_shm_group,
1402                 .maxlen         = sizeof(gid_t),
1403                 .mode           = 0644,
1404                 .proc_handler   = proc_dointvec,
1405          },
1406         {
1407                 .procname       = "nr_overcommit_hugepages",
1408                 .data           = NULL,
1409                 .maxlen         = sizeof(unsigned long),
1410                 .mode           = 0644,
1411                 .proc_handler   = hugetlb_overcommit_handler,
1412         },
1413 #endif
1414         {
1415                 .procname       = "lowmem_reserve_ratio",
1416                 .data           = &sysctl_lowmem_reserve_ratio,
1417                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1418                 .mode           = 0644,
1419                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1420         },
1421         {
1422                 .procname       = "drop_caches",
1423                 .data           = &sysctl_drop_caches,
1424                 .maxlen         = sizeof(int),
1425                 .mode           = 0644,
1426                 .proc_handler   = drop_caches_sysctl_handler,
1427                 .extra1         = &one,
1428                 .extra2         = &four,
1429         },
1430 #ifdef CONFIG_COMPACTION
1431         {
1432                 .procname       = "compact_memory",
1433                 .data           = &sysctl_compact_memory,
1434                 .maxlen         = sizeof(int),
1435                 .mode           = 0200,
1436                 .proc_handler   = sysctl_compaction_handler,
1437         },
1438         {
1439                 .procname       = "extfrag_threshold",
1440                 .data           = &sysctl_extfrag_threshold,
1441                 .maxlen         = sizeof(int),
1442                 .mode           = 0644,
1443                 .proc_handler   = sysctl_extfrag_handler,
1444                 .extra1         = &min_extfrag_threshold,
1445                 .extra2         = &max_extfrag_threshold,
1446         },
1447         {
1448                 .procname       = "compact_unevictable_allowed",
1449                 .data           = &sysctl_compact_unevictable_allowed,
1450                 .maxlen         = sizeof(int),
1451                 .mode           = 0644,
1452                 .proc_handler   = proc_dointvec,
1453                 .extra1         = &zero,
1454                 .extra2         = &one,
1455         },
1456
1457 #endif /* CONFIG_COMPACTION */
1458         {
1459                 .procname       = "min_free_kbytes",
1460                 .data           = &min_free_kbytes,
1461                 .maxlen         = sizeof(min_free_kbytes),
1462                 .mode           = 0644,
1463                 .proc_handler   = min_free_kbytes_sysctl_handler,
1464                 .extra1         = &zero,
1465         },
1466         {
1467                 .procname       = "watermark_scale_factor",
1468                 .data           = &watermark_scale_factor,
1469                 .maxlen         = sizeof(watermark_scale_factor),
1470                 .mode           = 0644,
1471                 .proc_handler   = watermark_scale_factor_sysctl_handler,
1472                 .extra1         = &one,
1473                 .extra2         = &one_thousand,
1474         },
1475         {
1476                 .procname       = "percpu_pagelist_fraction",
1477                 .data           = &percpu_pagelist_fraction,
1478                 .maxlen         = sizeof(percpu_pagelist_fraction),
1479                 .mode           = 0644,
1480                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1481                 .extra1         = &zero,
1482         },
1483 #ifdef CONFIG_MMU
1484         {
1485                 .procname       = "max_map_count",
1486                 .data           = &sysctl_max_map_count,
1487                 .maxlen         = sizeof(sysctl_max_map_count),
1488                 .mode           = 0644,
1489                 .proc_handler   = proc_dointvec_minmax,
1490                 .extra1         = &zero,
1491         },
1492 #else
1493         {
1494                 .procname       = "nr_trim_pages",
1495                 .data           = &sysctl_nr_trim_pages,
1496                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1497                 .mode           = 0644,
1498                 .proc_handler   = proc_dointvec_minmax,
1499                 .extra1         = &zero,
1500         },
1501 #endif
1502         {
1503                 .procname       = "laptop_mode",
1504                 .data           = &laptop_mode,
1505                 .maxlen         = sizeof(laptop_mode),
1506                 .mode           = 0644,
1507                 .proc_handler   = proc_dointvec_jiffies,
1508         },
1509         {
1510                 .procname       = "block_dump",
1511                 .data           = &block_dump,
1512                 .maxlen         = sizeof(block_dump),
1513                 .mode           = 0644,
1514                 .proc_handler   = proc_dointvec,
1515                 .extra1         = &zero,
1516         },
1517         {
1518                 .procname       = "vfs_cache_pressure",
1519                 .data           = &sysctl_vfs_cache_pressure,
1520                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1521                 .mode           = 0644,
1522                 .proc_handler   = proc_dointvec,
1523                 .extra1         = &zero,
1524         },
1525 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1526         {
1527                 .procname       = "legacy_va_layout",
1528                 .data           = &sysctl_legacy_va_layout,
1529                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1530                 .mode           = 0644,
1531                 .proc_handler   = proc_dointvec,
1532                 .extra1         = &zero,
1533         },
1534 #endif
1535 #ifdef CONFIG_NUMA
1536         {
1537                 .procname       = "zone_reclaim_mode",
1538                 .data           = &node_reclaim_mode,
1539                 .maxlen         = sizeof(node_reclaim_mode),
1540                 .mode           = 0644,
1541                 .proc_handler   = proc_dointvec,
1542                 .extra1         = &zero,
1543         },
1544         {
1545                 .procname       = "min_unmapped_ratio",
1546                 .data           = &sysctl_min_unmapped_ratio,
1547                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1548                 .mode           = 0644,
1549                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1550                 .extra1         = &zero,
1551                 .extra2         = &one_hundred,
1552         },
1553         {
1554                 .procname       = "min_slab_ratio",
1555                 .data           = &sysctl_min_slab_ratio,
1556                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1557                 .mode           = 0644,
1558                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1559                 .extra1         = &zero,
1560                 .extra2         = &one_hundred,
1561         },
1562 #endif
1563 #ifdef CONFIG_SMP
1564         {
1565                 .procname       = "stat_interval",
1566                 .data           = &sysctl_stat_interval,
1567                 .maxlen         = sizeof(sysctl_stat_interval),
1568                 .mode           = 0644,
1569                 .proc_handler   = proc_dointvec_jiffies,
1570         },
1571         {
1572                 .procname       = "stat_refresh",
1573                 .data           = NULL,
1574                 .maxlen         = 0,
1575                 .mode           = 0600,
1576                 .proc_handler   = vmstat_refresh,
1577         },
1578 #endif
1579 #ifdef CONFIG_MMU
1580         {
1581                 .procname       = "mmap_min_addr",
1582                 .data           = &dac_mmap_min_addr,
1583                 .maxlen         = sizeof(unsigned long),
1584                 .mode           = 0644,
1585                 .proc_handler   = mmap_min_addr_handler,
1586         },
1587 #endif
1588 #ifdef CONFIG_NUMA
1589         {
1590                 .procname       = "numa_zonelist_order",
1591                 .data           = &numa_zonelist_order,
1592                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1593                 .mode           = 0644,
1594                 .proc_handler   = numa_zonelist_order_handler,
1595         },
1596 #endif
1597 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1598    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1599         {
1600                 .procname       = "vdso_enabled",
1601 #ifdef CONFIG_X86_32
1602                 .data           = &vdso32_enabled,
1603                 .maxlen         = sizeof(vdso32_enabled),
1604 #else
1605                 .data           = &vdso_enabled,
1606                 .maxlen         = sizeof(vdso_enabled),
1607 #endif
1608                 .mode           = 0644,
1609                 .proc_handler   = proc_dointvec,
1610                 .extra1         = &zero,
1611         },
1612 #endif
1613 #ifdef CONFIG_HIGHMEM
1614         {
1615                 .procname       = "highmem_is_dirtyable",
1616                 .data           = &vm_highmem_is_dirtyable,
1617                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1618                 .mode           = 0644,
1619                 .proc_handler   = proc_dointvec_minmax,
1620                 .extra1         = &zero,
1621                 .extra2         = &one,
1622         },
1623 #endif
1624 #ifdef CONFIG_MEMORY_FAILURE
1625         {
1626                 .procname       = "memory_failure_early_kill",
1627                 .data           = &sysctl_memory_failure_early_kill,
1628                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1629                 .mode           = 0644,
1630                 .proc_handler   = proc_dointvec_minmax,
1631                 .extra1         = &zero,
1632                 .extra2         = &one,
1633         },
1634         {
1635                 .procname       = "memory_failure_recovery",
1636                 .data           = &sysctl_memory_failure_recovery,
1637                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1638                 .mode           = 0644,
1639                 .proc_handler   = proc_dointvec_minmax,
1640                 .extra1         = &zero,
1641                 .extra2         = &one,
1642         },
1643 #endif
1644         {
1645                 .procname       = "user_reserve_kbytes",
1646                 .data           = &sysctl_user_reserve_kbytes,
1647                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1648                 .mode           = 0644,
1649                 .proc_handler   = proc_doulongvec_minmax,
1650         },
1651         {
1652                 .procname       = "admin_reserve_kbytes",
1653                 .data           = &sysctl_admin_reserve_kbytes,
1654                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1655                 .mode           = 0644,
1656                 .proc_handler   = proc_doulongvec_minmax,
1657         },
1658 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
1659         {
1660                 .procname       = "mmap_rnd_bits",
1661                 .data           = &mmap_rnd_bits,
1662                 .maxlen         = sizeof(mmap_rnd_bits),
1663                 .mode           = 0600,
1664                 .proc_handler   = proc_dointvec_minmax,
1665                 .extra1         = (void *)&mmap_rnd_bits_min,
1666                 .extra2         = (void *)&mmap_rnd_bits_max,
1667         },
1668 #endif
1669 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
1670         {
1671                 .procname       = "mmap_rnd_compat_bits",
1672                 .data           = &mmap_rnd_compat_bits,
1673                 .maxlen         = sizeof(mmap_rnd_compat_bits),
1674                 .mode           = 0600,
1675                 .proc_handler   = proc_dointvec_minmax,
1676                 .extra1         = (void *)&mmap_rnd_compat_bits_min,
1677                 .extra2         = (void *)&mmap_rnd_compat_bits_max,
1678         },
1679 #endif
1680         { }
1681 };
1682
1683 static struct ctl_table fs_table[] = {
1684         {
1685                 .procname       = "inode-nr",
1686                 .data           = &inodes_stat,
1687                 .maxlen         = 2*sizeof(long),
1688                 .mode           = 0444,
1689                 .proc_handler   = proc_nr_inodes,
1690         },
1691         {
1692                 .procname       = "inode-state",
1693                 .data           = &inodes_stat,
1694                 .maxlen         = 7*sizeof(long),
1695                 .mode           = 0444,
1696                 .proc_handler   = proc_nr_inodes,
1697         },
1698         {
1699                 .procname       = "file-nr",
1700                 .data           = &files_stat,
1701                 .maxlen         = sizeof(files_stat),
1702                 .mode           = 0444,
1703                 .proc_handler   = proc_nr_files,
1704         },
1705         {
1706                 .procname       = "file-max",
1707                 .data           = &files_stat.max_files,
1708                 .maxlen         = sizeof(files_stat.max_files),
1709                 .mode           = 0644,
1710                 .proc_handler   = proc_doulongvec_minmax,
1711         },
1712         {
1713                 .procname       = "nr_open",
1714                 .data           = &sysctl_nr_open,
1715                 .maxlen         = sizeof(unsigned int),
1716                 .mode           = 0644,
1717                 .proc_handler   = proc_dointvec_minmax,
1718                 .extra1         = &sysctl_nr_open_min,
1719                 .extra2         = &sysctl_nr_open_max,
1720         },
1721         {
1722                 .procname       = "dentry-state",
1723                 .data           = &dentry_stat,
1724                 .maxlen         = 6*sizeof(long),
1725                 .mode           = 0444,
1726                 .proc_handler   = proc_nr_dentry,
1727         },
1728         {
1729                 .procname       = "overflowuid",
1730                 .data           = &fs_overflowuid,
1731                 .maxlen         = sizeof(int),
1732                 .mode           = 0644,
1733                 .proc_handler   = proc_dointvec_minmax,
1734                 .extra1         = &minolduid,
1735                 .extra2         = &maxolduid,
1736         },
1737         {
1738                 .procname       = "overflowgid",
1739                 .data           = &fs_overflowgid,
1740                 .maxlen         = sizeof(int),
1741                 .mode           = 0644,
1742                 .proc_handler   = proc_dointvec_minmax,
1743                 .extra1         = &minolduid,
1744                 .extra2         = &maxolduid,
1745         },
1746 #ifdef CONFIG_FILE_LOCKING
1747         {
1748                 .procname       = "leases-enable",
1749                 .data           = &leases_enable,
1750                 .maxlen         = sizeof(int),
1751                 .mode           = 0644,
1752                 .proc_handler   = proc_dointvec,
1753         },
1754 #endif
1755 #ifdef CONFIG_DNOTIFY
1756         {
1757                 .procname       = "dir-notify-enable",
1758                 .data           = &dir_notify_enable,
1759                 .maxlen         = sizeof(int),
1760                 .mode           = 0644,
1761                 .proc_handler   = proc_dointvec,
1762         },
1763 #endif
1764 #ifdef CONFIG_MMU
1765 #ifdef CONFIG_FILE_LOCKING
1766         {
1767                 .procname       = "lease-break-time",
1768                 .data           = &lease_break_time,
1769                 .maxlen         = sizeof(int),
1770                 .mode           = 0644,
1771                 .proc_handler   = proc_dointvec,
1772         },
1773 #endif
1774 #ifdef CONFIG_AIO
1775         {
1776                 .procname       = "aio-nr",
1777                 .data           = &aio_nr,
1778                 .maxlen         = sizeof(aio_nr),
1779                 .mode           = 0444,
1780                 .proc_handler   = proc_doulongvec_minmax,
1781         },
1782         {
1783                 .procname       = "aio-max-nr",
1784                 .data           = &aio_max_nr,
1785                 .maxlen         = sizeof(aio_max_nr),
1786                 .mode           = 0644,
1787                 .proc_handler   = proc_doulongvec_minmax,
1788         },
1789 #endif /* CONFIG_AIO */
1790 #ifdef CONFIG_INOTIFY_USER
1791         {
1792                 .procname       = "inotify",
1793                 .mode           = 0555,
1794                 .child          = inotify_table,
1795         },
1796 #endif
1797 #ifdef CONFIG_EPOLL
1798         {
1799                 .procname       = "epoll",
1800                 .mode           = 0555,
1801                 .child          = epoll_table,
1802         },
1803 #endif
1804 #endif
1805         {
1806                 .procname       = "protected_symlinks",
1807                 .data           = &sysctl_protected_symlinks,
1808                 .maxlen         = sizeof(int),
1809                 .mode           = 0600,
1810                 .proc_handler   = proc_dointvec_minmax,
1811                 .extra1         = &zero,
1812                 .extra2         = &one,
1813         },
1814         {
1815                 .procname       = "protected_hardlinks",
1816                 .data           = &sysctl_protected_hardlinks,
1817                 .maxlen         = sizeof(int),
1818                 .mode           = 0600,
1819                 .proc_handler   = proc_dointvec_minmax,
1820                 .extra1         = &zero,
1821                 .extra2         = &one,
1822         },
1823         {
1824                 .procname       = "protected_fifos",
1825                 .data           = &sysctl_protected_fifos,
1826                 .maxlen         = sizeof(int),
1827                 .mode           = 0600,
1828                 .proc_handler   = proc_dointvec_minmax,
1829                 .extra1         = &zero,
1830                 .extra2         = &two,
1831         },
1832         {
1833                 .procname       = "protected_regular",
1834                 .data           = &sysctl_protected_regular,
1835                 .maxlen         = sizeof(int),
1836                 .mode           = 0600,
1837                 .proc_handler   = proc_dointvec_minmax,
1838                 .extra1         = &zero,
1839                 .extra2         = &two,
1840         },
1841         {
1842                 .procname       = "suid_dumpable",
1843                 .data           = &suid_dumpable,
1844                 .maxlen         = sizeof(int),
1845                 .mode           = 0644,
1846                 .proc_handler   = proc_dointvec_minmax_coredump,
1847                 .extra1         = &zero,
1848                 .extra2         = &two,
1849         },
1850 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1851         {
1852                 .procname       = "binfmt_misc",
1853                 .mode           = 0555,
1854                 .child          = sysctl_mount_point,
1855         },
1856 #endif
1857         {
1858                 .procname       = "pipe-max-size",
1859                 .data           = &pipe_max_size,
1860                 .maxlen         = sizeof(pipe_max_size),
1861                 .mode           = 0644,
1862                 .proc_handler   = proc_dopipe_max_size,
1863         },
1864         {
1865                 .procname       = "pipe-user-pages-hard",
1866                 .data           = &pipe_user_pages_hard,
1867                 .maxlen         = sizeof(pipe_user_pages_hard),
1868                 .mode           = 0644,
1869                 .proc_handler   = proc_doulongvec_minmax,
1870         },
1871         {
1872                 .procname       = "pipe-user-pages-soft",
1873                 .data           = &pipe_user_pages_soft,
1874                 .maxlen         = sizeof(pipe_user_pages_soft),
1875                 .mode           = 0644,
1876                 .proc_handler   = proc_doulongvec_minmax,
1877         },
1878         {
1879                 .procname       = "mount-max",
1880                 .data           = &sysctl_mount_max,
1881                 .maxlen         = sizeof(unsigned int),
1882                 .mode           = 0644,
1883                 .proc_handler   = proc_dointvec_minmax,
1884                 .extra1         = &one,
1885         },
1886         { }
1887 };
1888
1889 static struct ctl_table debug_table[] = {
1890 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1891         {
1892                 .procname       = "exception-trace",
1893                 .data           = &show_unhandled_signals,
1894                 .maxlen         = sizeof(int),
1895                 .mode           = 0644,
1896                 .proc_handler   = proc_dointvec
1897         },
1898 #endif
1899 #if defined(CONFIG_OPTPROBES)
1900         {
1901                 .procname       = "kprobes-optimization",
1902                 .data           = &sysctl_kprobes_optimization,
1903                 .maxlen         = sizeof(int),
1904                 .mode           = 0644,
1905                 .proc_handler   = proc_kprobes_optimization_handler,
1906                 .extra1         = &zero,
1907                 .extra2         = &one,
1908         },
1909 #endif
1910         { }
1911 };
1912
1913 static struct ctl_table dev_table[] = {
1914         { }
1915 };
1916
1917 int __init sysctl_init(void)
1918 {
1919         struct ctl_table_header *hdr;
1920
1921         hdr = register_sysctl_table(sysctl_base_table);
1922         kmemleak_not_leak(hdr);
1923         return 0;
1924 }
1925
1926 #endif /* CONFIG_SYSCTL */
1927
1928 /*
1929  * /proc/sys support
1930  */
1931
1932 #ifdef CONFIG_PROC_SYSCTL
1933
1934 static int _proc_do_string(char *data, int maxlen, int write,
1935                            char __user *buffer,
1936                            size_t *lenp, loff_t *ppos)
1937 {
1938         size_t len;
1939         char __user *p;
1940         char c;
1941
1942         if (!data || !maxlen || !*lenp) {
1943                 *lenp = 0;
1944                 return 0;
1945         }
1946
1947         if (write) {
1948                 if (sysctl_writes_strict == SYSCTL_WRITES_STRICT) {
1949                         /* Only continue writes not past the end of buffer. */
1950                         len = strlen(data);
1951                         if (len > maxlen - 1)
1952                                 len = maxlen - 1;
1953
1954                         if (*ppos > len)
1955                                 return 0;
1956                         len = *ppos;
1957                 } else {
1958                         /* Start writing from beginning of buffer. */
1959                         len = 0;
1960                 }
1961
1962                 *ppos += *lenp;
1963                 p = buffer;
1964                 while ((p - buffer) < *lenp && len < maxlen - 1) {
1965                         if (get_user(c, p++))
1966                                 return -EFAULT;
1967                         if (c == 0 || c == '\n')
1968                                 break;
1969                         data[len++] = c;
1970                 }
1971                 data[len] = 0;
1972         } else {
1973                 len = strlen(data);
1974                 if (len > maxlen)
1975                         len = maxlen;
1976
1977                 if (*ppos > len) {
1978                         *lenp = 0;
1979                         return 0;
1980                 }
1981
1982                 data += *ppos;
1983                 len  -= *ppos;
1984
1985                 if (len > *lenp)
1986                         len = *lenp;
1987                 if (len)
1988                         if (copy_to_user(buffer, data, len))
1989                                 return -EFAULT;
1990                 if (len < *lenp) {
1991                         if (put_user('\n', buffer + len))
1992                                 return -EFAULT;
1993                         len++;
1994                 }
1995                 *lenp = len;
1996                 *ppos += len;
1997         }
1998         return 0;
1999 }
2000
2001 static void warn_sysctl_write(struct ctl_table *table)
2002 {
2003         pr_warn_once("%s wrote to %s when file position was not 0!\n"
2004                 "This will not be supported in the future. To silence this\n"
2005                 "warning, set kernel.sysctl_writes_strict = -1\n",
2006                 current->comm, table->procname);
2007 }
2008
2009 /**
2010  * proc_first_pos_non_zero_ignore - check if first position is allowed
2011  * @ppos: file position
2012  * @table: the sysctl table
2013  *
2014  * Returns true if the first position is non-zero and the sysctl_writes_strict
2015  * mode indicates this is not allowed for numeric input types. String proc
2016  * handlers can ignore the return value.
2017  */
2018 static bool proc_first_pos_non_zero_ignore(loff_t *ppos,
2019                                            struct ctl_table *table)
2020 {
2021         if (!*ppos)
2022                 return false;
2023
2024         switch (sysctl_writes_strict) {
2025         case SYSCTL_WRITES_STRICT:
2026                 return true;
2027         case SYSCTL_WRITES_WARN:
2028                 warn_sysctl_write(table);
2029                 return false;
2030         default:
2031                 return false;
2032         }
2033 }
2034
2035 /**
2036  * proc_dostring - read a string sysctl
2037  * @table: the sysctl table
2038  * @write: %TRUE if this is a write to the sysctl file
2039  * @buffer: the user buffer
2040  * @lenp: the size of the user buffer
2041  * @ppos: file position
2042  *
2043  * Reads/writes a string from/to the user buffer. If the kernel
2044  * buffer provided is not large enough to hold the string, the
2045  * string is truncated. The copied string is %NULL-terminated.
2046  * If the string is being read by the user process, it is copied
2047  * and a newline '\n' is added. It is truncated if the buffer is
2048  * not large enough.
2049  *
2050  * Returns 0 on success.
2051  */
2052 int proc_dostring(struct ctl_table *table, int write,
2053                   void __user *buffer, size_t *lenp, loff_t *ppos)
2054 {
2055         if (write)
2056                 proc_first_pos_non_zero_ignore(ppos, table);
2057
2058         return _proc_do_string((char *)(table->data), table->maxlen, write,
2059                                (char __user *)buffer, lenp, ppos);
2060 }
2061
2062 static size_t proc_skip_spaces(char **buf)
2063 {
2064         size_t ret;
2065         char *tmp = skip_spaces(*buf);
2066         ret = tmp - *buf;
2067         *buf = tmp;
2068         return ret;
2069 }
2070
2071 static void proc_skip_char(char **buf, size_t *size, const char v)
2072 {
2073         while (*size) {
2074                 if (**buf != v)
2075                         break;
2076                 (*size)--;
2077                 (*buf)++;
2078         }
2079 }
2080
2081 #define TMPBUFLEN 22
2082 /**
2083  * proc_get_long - reads an ASCII formatted integer from a user buffer
2084  *
2085  * @buf: a kernel buffer
2086  * @size: size of the kernel buffer
2087  * @val: this is where the number will be stored
2088  * @neg: set to %TRUE if number is negative
2089  * @perm_tr: a vector which contains the allowed trailers
2090  * @perm_tr_len: size of the perm_tr vector
2091  * @tr: pointer to store the trailer character
2092  *
2093  * In case of success %0 is returned and @buf and @size are updated with
2094  * the amount of bytes read. If @tr is non-NULL and a trailing
2095  * character exists (size is non-zero after returning from this
2096  * function), @tr is updated with the trailing character.
2097  */
2098 static int proc_get_long(char **buf, size_t *size,
2099                           unsigned long *val, bool *neg,
2100                           const char *perm_tr, unsigned perm_tr_len, char *tr)
2101 {
2102         int len;
2103         char *p, tmp[TMPBUFLEN];
2104
2105         if (!*size)
2106                 return -EINVAL;
2107
2108         len = *size;
2109         if (len > TMPBUFLEN - 1)
2110                 len = TMPBUFLEN - 1;
2111
2112         memcpy(tmp, *buf, len);
2113
2114         tmp[len] = 0;
2115         p = tmp;
2116         if (*p == '-' && *size > 1) {
2117                 *neg = true;
2118                 p++;
2119         } else
2120                 *neg = false;
2121         if (!isdigit(*p))
2122                 return -EINVAL;
2123
2124         *val = simple_strtoul(p, &p, 0);
2125
2126         len = p - tmp;
2127
2128         /* We don't know if the next char is whitespace thus we may accept
2129          * invalid integers (e.g. 1234...a) or two integers instead of one
2130          * (e.g. 123...1). So lets not allow such large numbers. */
2131         if (len == TMPBUFLEN - 1)
2132                 return -EINVAL;
2133
2134         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
2135                 return -EINVAL;
2136
2137         if (tr && (len < *size))
2138                 *tr = *p;
2139
2140         *buf += len;
2141         *size -= len;
2142
2143         return 0;
2144 }
2145
2146 /**
2147  * proc_put_long - converts an integer to a decimal ASCII formatted string
2148  *
2149  * @buf: the user buffer
2150  * @size: the size of the user buffer
2151  * @val: the integer to be converted
2152  * @neg: sign of the number, %TRUE for negative
2153  *
2154  * In case of success %0 is returned and @buf and @size are updated with
2155  * the amount of bytes written.
2156  */
2157 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
2158                           bool neg)
2159 {
2160         int len;
2161         char tmp[TMPBUFLEN], *p = tmp;
2162
2163         sprintf(p, "%s%lu", neg ? "-" : "", val);
2164         len = strlen(tmp);
2165         if (len > *size)
2166                 len = *size;
2167         if (copy_to_user(*buf, tmp, len))
2168                 return -EFAULT;
2169         *size -= len;
2170         *buf += len;
2171         return 0;
2172 }
2173 #undef TMPBUFLEN
2174
2175 static int proc_put_char(void __user **buf, size_t *size, char c)
2176 {
2177         if (*size) {
2178                 char __user **buffer = (char __user **)buf;
2179                 if (put_user(c, *buffer))
2180                         return -EFAULT;
2181                 (*size)--, (*buffer)++;
2182                 *buf = *buffer;
2183         }
2184         return 0;
2185 }
2186
2187 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
2188                                  int *valp,
2189                                  int write, void *data)
2190 {
2191         if (write) {
2192                 if (*negp) {
2193                         if (*lvalp > (unsigned long) INT_MAX + 1)
2194                                 return -EINVAL;
2195                         *valp = -*lvalp;
2196                 } else {
2197                         if (*lvalp > (unsigned long) INT_MAX)
2198                                 return -EINVAL;
2199                         *valp = *lvalp;
2200                 }
2201         } else {
2202                 int val = *valp;
2203                 if (val < 0) {
2204                         *negp = true;
2205                         *lvalp = -(unsigned long)val;
2206                 } else {
2207                         *negp = false;
2208                         *lvalp = (unsigned long)val;
2209                 }
2210         }
2211         return 0;
2212 }
2213
2214 static int do_proc_douintvec_conv(unsigned long *lvalp,
2215                                   unsigned int *valp,
2216                                   int write, void *data)
2217 {
2218         if (write) {
2219                 if (*lvalp > UINT_MAX)
2220                         return -EINVAL;
2221                 *valp = *lvalp;
2222         } else {
2223                 unsigned int val = *valp;
2224                 *lvalp = (unsigned long)val;
2225         }
2226         return 0;
2227 }
2228
2229 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
2230
2231 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2232                   int write, void __user *buffer,
2233                   size_t *lenp, loff_t *ppos,
2234                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2235                               int write, void *data),
2236                   void *data)
2237 {
2238         int *i, vleft, first = 1, err = 0;
2239         size_t left;
2240         char *kbuf = NULL, *p;
2241
2242         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
2243                 *lenp = 0;
2244                 return 0;
2245         }
2246
2247         i = (int *) tbl_data;
2248         vleft = table->maxlen / sizeof(*i);
2249         left = *lenp;
2250
2251         if (!conv)
2252                 conv = do_proc_dointvec_conv;
2253
2254         if (write) {
2255                 if (proc_first_pos_non_zero_ignore(ppos, table))
2256                         goto out;
2257
2258                 if (left > PAGE_SIZE - 1)
2259                         left = PAGE_SIZE - 1;
2260                 p = kbuf = memdup_user_nul(buffer, left);
2261                 if (IS_ERR(kbuf))
2262                         return PTR_ERR(kbuf);
2263         }
2264
2265         for (; left && vleft--; i++, first=0) {
2266                 unsigned long lval;
2267                 bool neg;
2268
2269                 if (write) {
2270                         left -= proc_skip_spaces(&p);
2271
2272                         if (!left)
2273                                 break;
2274                         err = proc_get_long(&p, &left, &lval, &neg,
2275                                              proc_wspace_sep,
2276                                              sizeof(proc_wspace_sep), NULL);
2277                         if (err)
2278                                 break;
2279                         if (conv(&neg, &lval, i, 1, data)) {
2280                                 err = -EINVAL;
2281                                 break;
2282                         }
2283                 } else {
2284                         if (conv(&neg, &lval, i, 0, data)) {
2285                                 err = -EINVAL;
2286                                 break;
2287                         }
2288                         if (!first)
2289                                 err = proc_put_char(&buffer, &left, '\t');
2290                         if (err)
2291                                 break;
2292                         err = proc_put_long(&buffer, &left, lval, neg);
2293                         if (err)
2294                                 break;
2295                 }
2296         }
2297
2298         if (!write && !first && left && !err)
2299                 err = proc_put_char(&buffer, &left, '\n');
2300         if (write && !err && left)
2301                 left -= proc_skip_spaces(&p);
2302         if (write) {
2303                 kfree(kbuf);
2304                 if (first)
2305                         return err ? : -EINVAL;
2306         }
2307         *lenp -= left;
2308 out:
2309         *ppos += *lenp;
2310         return err;
2311 }
2312
2313 static int do_proc_dointvec(struct ctl_table *table, int write,
2314                   void __user *buffer, size_t *lenp, loff_t *ppos,
2315                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2316                               int write, void *data),
2317                   void *data)
2318 {
2319         return __do_proc_dointvec(table->data, table, write,
2320                         buffer, lenp, ppos, conv, data);
2321 }
2322
2323 static int do_proc_douintvec_w(unsigned int *tbl_data,
2324                                struct ctl_table *table,
2325                                void __user *buffer,
2326                                size_t *lenp, loff_t *ppos,
2327                                int (*conv)(unsigned long *lvalp,
2328                                            unsigned int *valp,
2329                                            int write, void *data),
2330                                void *data)
2331 {
2332         unsigned long lval;
2333         int err = 0;
2334         size_t left;
2335         bool neg;
2336         char *kbuf = NULL, *p;
2337
2338         left = *lenp;
2339
2340         if (proc_first_pos_non_zero_ignore(ppos, table))
2341                 goto bail_early;
2342
2343         if (left > PAGE_SIZE - 1)
2344                 left = PAGE_SIZE - 1;
2345
2346         p = kbuf = memdup_user_nul(buffer, left);
2347         if (IS_ERR(kbuf))
2348                 return -EINVAL;
2349
2350         left -= proc_skip_spaces(&p);
2351         if (!left) {
2352                 err = -EINVAL;
2353                 goto out_free;
2354         }
2355
2356         err = proc_get_long(&p, &left, &lval, &neg,
2357                              proc_wspace_sep,
2358                              sizeof(proc_wspace_sep), NULL);
2359         if (err || neg) {
2360                 err = -EINVAL;
2361                 goto out_free;
2362         }
2363
2364         if (conv(&lval, tbl_data, 1, data)) {
2365                 err = -EINVAL;
2366                 goto out_free;
2367         }
2368
2369         if (!err && left)
2370                 left -= proc_skip_spaces(&p);
2371
2372 out_free:
2373         kfree(kbuf);
2374         if (err)
2375                 return -EINVAL;
2376
2377         return 0;
2378
2379         /* This is in keeping with old __do_proc_dointvec() */
2380 bail_early:
2381         *ppos += *lenp;
2382         return err;
2383 }
2384
2385 static int do_proc_douintvec_r(unsigned int *tbl_data, void __user *buffer,
2386                                size_t *lenp, loff_t *ppos,
2387                                int (*conv)(unsigned long *lvalp,
2388                                            unsigned int *valp,
2389                                            int write, void *data),
2390                                void *data)
2391 {
2392         unsigned long lval;
2393         int err = 0;
2394         size_t left;
2395
2396         left = *lenp;
2397
2398         if (conv(&lval, tbl_data, 0, data)) {
2399                 err = -EINVAL;
2400                 goto out;
2401         }
2402
2403         err = proc_put_long(&buffer, &left, lval, false);
2404         if (err || !left)
2405                 goto out;
2406
2407         err = proc_put_char(&buffer, &left, '\n');
2408
2409 out:
2410         *lenp -= left;
2411         *ppos += *lenp;
2412
2413         return err;
2414 }
2415
2416 static int __do_proc_douintvec(void *tbl_data, struct ctl_table *table,
2417                                int write, void __user *buffer,
2418                                size_t *lenp, loff_t *ppos,
2419                                int (*conv)(unsigned long *lvalp,
2420                                            unsigned int *valp,
2421                                            int write, void *data),
2422                                void *data)
2423 {
2424         unsigned int *i, vleft;
2425
2426         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
2427                 *lenp = 0;
2428                 return 0;
2429         }
2430
2431         i = (unsigned int *) tbl_data;
2432         vleft = table->maxlen / sizeof(*i);
2433
2434         /*
2435          * Arrays are not supported, keep this simple. *Do not* add
2436          * support for them.
2437          */
2438         if (vleft != 1) {
2439                 *lenp = 0;
2440                 return -EINVAL;
2441         }
2442
2443         if (!conv)
2444                 conv = do_proc_douintvec_conv;
2445
2446         if (write)
2447                 return do_proc_douintvec_w(i, table, buffer, lenp, ppos,
2448                                            conv, data);
2449         return do_proc_douintvec_r(i, buffer, lenp, ppos, conv, data);
2450 }
2451
2452 static int do_proc_douintvec(struct ctl_table *table, int write,
2453                              void __user *buffer, size_t *lenp, loff_t *ppos,
2454                              int (*conv)(unsigned long *lvalp,
2455                                          unsigned int *valp,
2456                                          int write, void *data),
2457                              void *data)
2458 {
2459         return __do_proc_douintvec(table->data, table, write,
2460                                    buffer, lenp, ppos, conv, data);
2461 }
2462
2463 /**
2464  * proc_dointvec - read a vector of integers
2465  * @table: the sysctl table
2466  * @write: %TRUE if this is a write to the sysctl file
2467  * @buffer: the user buffer
2468  * @lenp: the size of the user buffer
2469  * @ppos: file position
2470  *
2471  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2472  * values from/to the user buffer, treated as an ASCII string.
2473  *
2474  * Returns 0 on success.
2475  */
2476 int proc_dointvec(struct ctl_table *table, int write,
2477                      void __user *buffer, size_t *lenp, loff_t *ppos)
2478 {
2479         return do_proc_dointvec(table, write, buffer, lenp, ppos, NULL, NULL);
2480 }
2481
2482 /**
2483  * proc_douintvec - read a vector of unsigned integers
2484  * @table: the sysctl table
2485  * @write: %TRUE if this is a write to the sysctl file
2486  * @buffer: the user buffer
2487  * @lenp: the size of the user buffer
2488  * @ppos: file position
2489  *
2490  * Reads/writes up to table->maxlen/sizeof(unsigned int) unsigned integer
2491  * values from/to the user buffer, treated as an ASCII string.
2492  *
2493  * Returns 0 on success.
2494  */
2495 int proc_douintvec(struct ctl_table *table, int write,
2496                      void __user *buffer, size_t *lenp, loff_t *ppos)
2497 {
2498         return do_proc_douintvec(table, write, buffer, lenp, ppos,
2499                                  do_proc_douintvec_conv, NULL);
2500 }
2501
2502 /*
2503  * Taint values can only be increased
2504  * This means we can safely use a temporary.
2505  */
2506 static int proc_taint(struct ctl_table *table, int write,
2507                                void __user *buffer, size_t *lenp, loff_t *ppos)
2508 {
2509         struct ctl_table t;
2510         unsigned long tmptaint = get_taint();
2511         int err;
2512
2513         if (write && !capable(CAP_SYS_ADMIN))
2514                 return -EPERM;
2515
2516         t = *table;
2517         t.data = &tmptaint;
2518         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2519         if (err < 0)
2520                 return err;
2521
2522         if (write) {
2523                 /*
2524                  * Poor man's atomic or. Not worth adding a primitive
2525                  * to everyone's atomic.h for this
2526                  */
2527                 int i;
2528                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2529                         if ((tmptaint >> i) & 1)
2530                                 add_taint(i, LOCKDEP_STILL_OK);
2531                 }
2532         }
2533
2534         return err;
2535 }
2536
2537 #ifdef CONFIG_PRINTK
2538 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2539                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2540 {
2541         if (write && !capable(CAP_SYS_ADMIN))
2542                 return -EPERM;
2543
2544         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2545 }
2546 #endif
2547
2548 /**
2549  * struct do_proc_dointvec_minmax_conv_param - proc_dointvec_minmax() range checking structure
2550  * @min: pointer to minimum allowable value
2551  * @max: pointer to maximum allowable value
2552  *
2553  * The do_proc_dointvec_minmax_conv_param structure provides the
2554  * minimum and maximum values for doing range checking for those sysctl
2555  * parameters that use the proc_dointvec_minmax() handler.
2556  */
2557 struct do_proc_dointvec_minmax_conv_param {
2558         int *min;
2559         int *max;
2560 };
2561
2562 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2563                                         int *valp,
2564                                         int write, void *data)
2565 {
2566         struct do_proc_dointvec_minmax_conv_param *param = data;
2567         if (write) {
2568                 int val = *negp ? -*lvalp : *lvalp;
2569                 if ((param->min && *param->min > val) ||
2570                     (param->max && *param->max < val))
2571                         return -EINVAL;
2572                 *valp = val;
2573         } else {
2574                 int val = *valp;
2575                 if (val < 0) {
2576                         *negp = true;
2577                         *lvalp = -(unsigned long)val;
2578                 } else {
2579                         *negp = false;
2580                         *lvalp = (unsigned long)val;
2581                 }
2582         }
2583         return 0;
2584 }
2585
2586 /**
2587  * proc_dointvec_minmax - read a vector of integers with min/max values
2588  * @table: the sysctl table
2589  * @write: %TRUE if this is a write to the sysctl file
2590  * @buffer: the user buffer
2591  * @lenp: the size of the user buffer
2592  * @ppos: file position
2593  *
2594  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2595  * values from/to the user buffer, treated as an ASCII string.
2596  *
2597  * This routine will ensure the values are within the range specified by
2598  * table->extra1 (min) and table->extra2 (max).
2599  *
2600  * Returns 0 on success or -EINVAL on write when the range check fails.
2601  */
2602 int proc_dointvec_minmax(struct ctl_table *table, int write,
2603                   void __user *buffer, size_t *lenp, loff_t *ppos)
2604 {
2605         struct do_proc_dointvec_minmax_conv_param param = {
2606                 .min = (int *) table->extra1,
2607                 .max = (int *) table->extra2,
2608         };
2609         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2610                                 do_proc_dointvec_minmax_conv, &param);
2611 }
2612
2613 /**
2614  * struct do_proc_douintvec_minmax_conv_param - proc_douintvec_minmax() range checking structure
2615  * @min: pointer to minimum allowable value
2616  * @max: pointer to maximum allowable value
2617  *
2618  * The do_proc_douintvec_minmax_conv_param structure provides the
2619  * minimum and maximum values for doing range checking for those sysctl
2620  * parameters that use the proc_douintvec_minmax() handler.
2621  */
2622 struct do_proc_douintvec_minmax_conv_param {
2623         unsigned int *min;
2624         unsigned int *max;
2625 };
2626
2627 static int do_proc_douintvec_minmax_conv(unsigned long *lvalp,
2628                                          unsigned int *valp,
2629                                          int write, void *data)
2630 {
2631         struct do_proc_douintvec_minmax_conv_param *param = data;
2632
2633         if (write) {
2634                 unsigned int val = *lvalp;
2635
2636                 if (*lvalp > UINT_MAX)
2637                         return -EINVAL;
2638
2639                 if ((param->min && *param->min > val) ||
2640                     (param->max && *param->max < val))
2641                         return -ERANGE;
2642
2643                 *valp = val;
2644         } else {
2645                 unsigned int val = *valp;
2646                 *lvalp = (unsigned long) val;
2647         }
2648
2649         return 0;
2650 }
2651
2652 /**
2653  * proc_douintvec_minmax - read a vector of unsigned ints with min/max values
2654  * @table: the sysctl table
2655  * @write: %TRUE if this is a write to the sysctl file
2656  * @buffer: the user buffer
2657  * @lenp: the size of the user buffer
2658  * @ppos: file position
2659  *
2660  * Reads/writes up to table->maxlen/sizeof(unsigned int) unsigned integer
2661  * values from/to the user buffer, treated as an ASCII string. Negative
2662  * strings are not allowed.
2663  *
2664  * This routine will ensure the values are within the range specified by
2665  * table->extra1 (min) and table->extra2 (max). There is a final sanity
2666  * check for UINT_MAX to avoid having to support wrap around uses from
2667  * userspace.
2668  *
2669  * Returns 0 on success or -ERANGE on write when the range check fails.
2670  */
2671 int proc_douintvec_minmax(struct ctl_table *table, int write,
2672                           void __user *buffer, size_t *lenp, loff_t *ppos)
2673 {
2674         struct do_proc_douintvec_minmax_conv_param param = {
2675                 .min = (unsigned int *) table->extra1,
2676                 .max = (unsigned int *) table->extra2,
2677         };
2678         return do_proc_douintvec(table, write, buffer, lenp, ppos,
2679                                  do_proc_douintvec_minmax_conv, &param);
2680 }
2681
2682 static int do_proc_dopipe_max_size_conv(unsigned long *lvalp,
2683                                         unsigned int *valp,
2684                                         int write, void *data)
2685 {
2686         if (write) {
2687                 unsigned int val;
2688
2689                 val = round_pipe_size(*lvalp);
2690                 if (val == 0)
2691                         return -EINVAL;
2692
2693                 *valp = val;
2694         } else {
2695                 unsigned int val = *valp;
2696                 *lvalp = (unsigned long) val;
2697         }
2698
2699         return 0;
2700 }
2701
2702 static int proc_dopipe_max_size(struct ctl_table *table, int write,
2703                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2704 {
2705         return do_proc_douintvec(table, write, buffer, lenp, ppos,
2706                                  do_proc_dopipe_max_size_conv, NULL);
2707 }
2708
2709 static void validate_coredump_safety(void)
2710 {
2711 #ifdef CONFIG_COREDUMP
2712         if (suid_dumpable == SUID_DUMP_ROOT &&
2713             core_pattern[0] != '/' && core_pattern[0] != '|') {
2714                 printk(KERN_WARNING
2715 "Unsafe core_pattern used with fs.suid_dumpable=2.\n"
2716 "Pipe handler or fully qualified core dump path required.\n"
2717 "Set kernel.core_pattern before fs.suid_dumpable.\n"
2718                 );
2719         }
2720 #endif
2721 }
2722
2723 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2724                 void __user *buffer, size_t *lenp, loff_t *ppos)
2725 {
2726         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2727         if (!error)
2728                 validate_coredump_safety();
2729         return error;
2730 }
2731
2732 #ifdef CONFIG_COREDUMP
2733 static int proc_dostring_coredump(struct ctl_table *table, int write,
2734                   void __user *buffer, size_t *lenp, loff_t *ppos)
2735 {
2736         int error = proc_dostring(table, write, buffer, lenp, ppos);
2737         if (!error)
2738                 validate_coredump_safety();
2739         return error;
2740 }
2741 #endif
2742
2743 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2744                                      void __user *buffer,
2745                                      size_t *lenp, loff_t *ppos,
2746                                      unsigned long convmul,
2747                                      unsigned long convdiv)
2748 {
2749         unsigned long *i, *min, *max;
2750         int vleft, first = 1, err = 0;
2751         size_t left;
2752         char *kbuf = NULL, *p;
2753
2754         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2755                 *lenp = 0;
2756                 return 0;
2757         }
2758
2759         i = (unsigned long *) data;
2760         min = (unsigned long *) table->extra1;
2761         max = (unsigned long *) table->extra2;
2762         vleft = table->maxlen / sizeof(unsigned long);
2763         left = *lenp;
2764
2765         if (write) {
2766                 if (proc_first_pos_non_zero_ignore(ppos, table))
2767                         goto out;
2768
2769                 if (left > PAGE_SIZE - 1)
2770                         left = PAGE_SIZE - 1;
2771                 p = kbuf = memdup_user_nul(buffer, left);
2772                 if (IS_ERR(kbuf))
2773                         return PTR_ERR(kbuf);
2774         }
2775
2776         for (; left && vleft--; i++, first = 0) {
2777                 unsigned long val;
2778
2779                 if (write) {
2780                         bool neg;
2781
2782                         left -= proc_skip_spaces(&p);
2783
2784                         err = proc_get_long(&p, &left, &val, &neg,
2785                                              proc_wspace_sep,
2786                                              sizeof(proc_wspace_sep), NULL);
2787                         if (err)
2788                                 break;
2789                         if (neg)
2790                                 continue;
2791                         val = convmul * val / convdiv;
2792                         if ((min && val < *min) || (max && val > *max))
2793                                 continue;
2794                         *i = val;
2795                 } else {
2796                         val = convdiv * (*i) / convmul;
2797                         if (!first) {
2798                                 err = proc_put_char(&buffer, &left, '\t');
2799                                 if (err)
2800                                         break;
2801                         }
2802                         err = proc_put_long(&buffer, &left, val, false);
2803                         if (err)
2804                                 break;
2805                 }
2806         }
2807
2808         if (!write && !first && left && !err)
2809                 err = proc_put_char(&buffer, &left, '\n');
2810         if (write && !err)
2811                 left -= proc_skip_spaces(&p);
2812         if (write) {
2813                 kfree(kbuf);
2814                 if (first)
2815                         return err ? : -EINVAL;
2816         }
2817         *lenp -= left;
2818 out:
2819         *ppos += *lenp;
2820         return err;
2821 }
2822
2823 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2824                                      void __user *buffer,
2825                                      size_t *lenp, loff_t *ppos,
2826                                      unsigned long convmul,
2827                                      unsigned long convdiv)
2828 {
2829         return __do_proc_doulongvec_minmax(table->data, table, write,
2830                         buffer, lenp, ppos, convmul, convdiv);
2831 }
2832
2833 /**
2834  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2835  * @table: the sysctl table
2836  * @write: %TRUE if this is a write to the sysctl file
2837  * @buffer: the user buffer
2838  * @lenp: the size of the user buffer
2839  * @ppos: file position
2840  *
2841  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2842  * values from/to the user buffer, treated as an ASCII string.
2843  *
2844  * This routine will ensure the values are within the range specified by
2845  * table->extra1 (min) and table->extra2 (max).
2846  *
2847  * Returns 0 on success.
2848  */
2849 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2850                            void __user *buffer, size_t *lenp, loff_t *ppos)
2851 {
2852     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2853 }
2854
2855 /**
2856  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2857  * @table: the sysctl table
2858  * @write: %TRUE if this is a write to the sysctl file
2859  * @buffer: the user buffer
2860  * @lenp: the size of the user buffer
2861  * @ppos: file position
2862  *
2863  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2864  * values from/to the user buffer, treated as an ASCII string. The values
2865  * are treated as milliseconds, and converted to jiffies when they are stored.
2866  *
2867  * This routine will ensure the values are within the range specified by
2868  * table->extra1 (min) and table->extra2 (max).
2869  *
2870  * Returns 0 on success.
2871  */
2872 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2873                                       void __user *buffer,
2874                                       size_t *lenp, loff_t *ppos)
2875 {
2876     return do_proc_doulongvec_minmax(table, write, buffer,
2877                                      lenp, ppos, HZ, 1000l);
2878 }
2879
2880
2881 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2882                                          int *valp,
2883                                          int write, void *data)
2884 {
2885         if (write) {
2886                 if (*lvalp > INT_MAX / HZ)
2887                         return 1;
2888                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2889         } else {
2890                 int val = *valp;
2891                 unsigned long lval;
2892                 if (val < 0) {
2893                         *negp = true;
2894                         lval = -(unsigned long)val;
2895                 } else {
2896                         *negp = false;
2897                         lval = (unsigned long)val;
2898                 }
2899                 *lvalp = lval / HZ;
2900         }
2901         return 0;
2902 }
2903
2904 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2905                                                 int *valp,
2906                                                 int write, void *data)
2907 {
2908         if (write) {
2909                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2910                         return 1;
2911                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2912         } else {
2913                 int val = *valp;
2914                 unsigned long lval;
2915                 if (val < 0) {
2916                         *negp = true;
2917                         lval = -(unsigned long)val;
2918                 } else {
2919                         *negp = false;
2920                         lval = (unsigned long)val;
2921                 }
2922                 *lvalp = jiffies_to_clock_t(lval);
2923         }
2924         return 0;
2925 }
2926
2927 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2928                                             int *valp,
2929                                             int write, void *data)
2930 {
2931         if (write) {
2932                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2933
2934                 if (jif > INT_MAX)
2935                         return 1;
2936                 *valp = (int)jif;
2937         } else {
2938                 int val = *valp;
2939                 unsigned long lval;
2940                 if (val < 0) {
2941                         *negp = true;
2942                         lval = -(unsigned long)val;
2943                 } else {
2944                         *negp = false;
2945                         lval = (unsigned long)val;
2946                 }
2947                 *lvalp = jiffies_to_msecs(lval);
2948         }
2949         return 0;
2950 }
2951
2952 /**
2953  * proc_dointvec_jiffies - read a vector of integers as seconds
2954  * @table: the sysctl table
2955  * @write: %TRUE if this is a write to the sysctl file
2956  * @buffer: the user buffer
2957  * @lenp: the size of the user buffer
2958  * @ppos: file position
2959  *
2960  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2961  * values from/to the user buffer, treated as an ASCII string.
2962  * The values read are assumed to be in seconds, and are converted into
2963  * jiffies.
2964  *
2965  * Returns 0 on success.
2966  */
2967 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2968                           void __user *buffer, size_t *lenp, loff_t *ppos)
2969 {
2970     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2971                             do_proc_dointvec_jiffies_conv,NULL);
2972 }
2973
2974 /**
2975  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2976  * @table: the sysctl table
2977  * @write: %TRUE if this is a write to the sysctl file
2978  * @buffer: the user buffer
2979  * @lenp: the size of the user buffer
2980  * @ppos: pointer to the file position
2981  *
2982  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2983  * values from/to the user buffer, treated as an ASCII string.
2984  * The values read are assumed to be in 1/USER_HZ seconds, and
2985  * are converted into jiffies.
2986  *
2987  * Returns 0 on success.
2988  */
2989 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2990                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2991 {
2992     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2993                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2994 }
2995
2996 /**
2997  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2998  * @table: the sysctl table
2999  * @write: %TRUE if this is a write to the sysctl file
3000  * @buffer: the user buffer
3001  * @lenp: the size of the user buffer
3002  * @ppos: file position
3003  * @ppos: the current position in the file
3004  *
3005  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
3006  * values from/to the user buffer, treated as an ASCII string.
3007  * The values read are assumed to be in 1/1000 seconds, and
3008  * are converted into jiffies.
3009  *
3010  * Returns 0 on success.
3011  */
3012 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
3013                              void __user *buffer, size_t *lenp, loff_t *ppos)
3014 {
3015         return do_proc_dointvec(table, write, buffer, lenp, ppos,
3016                                 do_proc_dointvec_ms_jiffies_conv, NULL);
3017 }
3018
3019 static int proc_do_cad_pid(struct ctl_table *table, int write,
3020                            void __user *buffer, size_t *lenp, loff_t *ppos)
3021 {
3022         struct pid *new_pid;
3023         pid_t tmp;
3024         int r;
3025
3026         tmp = pid_vnr(cad_pid);
3027
3028         r = __do_proc_dointvec(&tmp, table, write, buffer,
3029                                lenp, ppos, NULL, NULL);
3030         if (r || !write)
3031                 return r;
3032
3033         new_pid = find_get_pid(tmp);
3034         if (!new_pid)
3035                 return -ESRCH;
3036
3037         put_pid(xchg(&cad_pid, new_pid));
3038         return 0;
3039 }
3040
3041 /**
3042  * proc_do_large_bitmap - read/write from/to a large bitmap
3043  * @table: the sysctl table
3044  * @write: %TRUE if this is a write to the sysctl file
3045  * @buffer: the user buffer
3046  * @lenp: the size of the user buffer
3047  * @ppos: file position
3048  *
3049  * The bitmap is stored at table->data and the bitmap length (in bits)
3050  * in table->maxlen.
3051  *
3052  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
3053  * large bitmaps may be represented in a compact manner. Writing into
3054  * the file will clear the bitmap then update it with the given input.
3055  *
3056  * Returns 0 on success.
3057  */
3058 int proc_do_large_bitmap(struct ctl_table *table, int write,
3059                          void __user *buffer, size_t *lenp, loff_t *ppos)
3060 {
3061         int err = 0;
3062         bool first = 1;
3063         size_t left = *lenp;
3064         unsigned long bitmap_len = table->maxlen;
3065         unsigned long *bitmap = *(unsigned long **) table->data;
3066         unsigned long *tmp_bitmap = NULL;
3067         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
3068
3069         if (!bitmap || !bitmap_len || !left || (*ppos && !write)) {
3070                 *lenp = 0;
3071                 return 0;
3072         }
3073
3074         if (write) {
3075                 char *kbuf, *p;
3076
3077                 if (left > PAGE_SIZE - 1)
3078                         left = PAGE_SIZE - 1;
3079
3080                 p = kbuf = memdup_user_nul(buffer, left);
3081                 if (IS_ERR(kbuf))
3082                         return PTR_ERR(kbuf);
3083
3084                 tmp_bitmap = kcalloc(BITS_TO_LONGS(bitmap_len),
3085                                      sizeof(unsigned long),
3086                                      GFP_KERNEL);
3087                 if (!tmp_bitmap) {
3088                         kfree(kbuf);
3089                         return -ENOMEM;
3090                 }
3091                 proc_skip_char(&p, &left, '\n');
3092                 while (!err && left) {
3093                         unsigned long val_a, val_b;
3094                         bool neg;
3095
3096                         err = proc_get_long(&p, &left, &val_a, &neg, tr_a,
3097                                              sizeof(tr_a), &c);
3098                         if (err)
3099                                 break;
3100                         if (val_a >= bitmap_len || neg) {
3101                                 err = -EINVAL;
3102                                 break;
3103                         }
3104
3105                         val_b = val_a;
3106                         if (left) {
3107                                 p++;
3108                                 left--;
3109                         }
3110
3111                         if (c == '-') {
3112                                 err = proc_get_long(&p, &left, &val_b,
3113                                                      &neg, tr_b, sizeof(tr_b),
3114                                                      &c);
3115                                 if (err)
3116                                         break;
3117                                 if (val_b >= bitmap_len || neg ||
3118                                     val_a > val_b) {
3119                                         err = -EINVAL;
3120                                         break;
3121                                 }
3122                                 if (left) {
3123                                         p++;
3124                                         left--;
3125                                 }
3126                         }
3127
3128                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
3129                         first = 0;
3130                         proc_skip_char(&p, &left, '\n');
3131                 }
3132                 kfree(kbuf);
3133         } else {
3134                 unsigned long bit_a, bit_b = 0;
3135
3136                 while (left) {
3137                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
3138                         if (bit_a >= bitmap_len)
3139                                 break;
3140                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
3141                                                    bit_a + 1) - 1;
3142
3143                         if (!first) {
3144                                 err = proc_put_char(&buffer, &left, ',');
3145                                 if (err)
3146                                         break;
3147                         }
3148                         err = proc_put_long(&buffer, &left, bit_a, false);
3149                         if (err)
3150                                 break;
3151                         if (bit_a != bit_b) {
3152                                 err = proc_put_char(&buffer, &left, '-');
3153                                 if (err)
3154                                         break;
3155                                 err = proc_put_long(&buffer, &left, bit_b, false);
3156                                 if (err)
3157                                         break;
3158                         }
3159
3160                         first = 0; bit_b++;
3161                 }
3162                 if (!err)
3163                         err = proc_put_char(&buffer, &left, '\n');
3164         }
3165
3166         if (!err) {
3167                 if (write) {
3168                         if (*ppos)
3169                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
3170                         else
3171                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
3172                 }
3173                 *lenp -= left;
3174                 *ppos += *lenp;
3175         }
3176
3177         kfree(tmp_bitmap);
3178         return err;
3179 }
3180
3181 #else /* CONFIG_PROC_SYSCTL */
3182
3183 int proc_dostring(struct ctl_table *table, int write,
3184                   void __user *buffer, size_t *lenp, loff_t *ppos)
3185 {
3186         return -ENOSYS;
3187 }
3188
3189 int proc_dointvec(struct ctl_table *table, int write,
3190                   void __user *buffer, size_t *lenp, loff_t *ppos)
3191 {
3192         return -ENOSYS;
3193 }
3194
3195 int proc_douintvec(struct ctl_table *table, int write,
3196                   void __user *buffer, size_t *lenp, loff_t *ppos)
3197 {
3198         return -ENOSYS;
3199 }
3200
3201 int proc_dointvec_minmax(struct ctl_table *table, int write,
3202                     void __user *buffer, size_t *lenp, loff_t *ppos)
3203 {
3204         return -ENOSYS;
3205 }
3206
3207 int proc_douintvec_minmax(struct ctl_table *table, int write,
3208                           void __user *buffer, size_t *lenp, loff_t *ppos)
3209 {
3210         return -ENOSYS;
3211 }
3212
3213 int proc_dointvec_jiffies(struct ctl_table *table, int write,
3214                     void __user *buffer, size_t *lenp, loff_t *ppos)
3215 {
3216         return -ENOSYS;
3217 }
3218
3219 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
3220                     void __user *buffer, size_t *lenp, loff_t *ppos)
3221 {
3222         return -ENOSYS;
3223 }
3224
3225 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
3226                              void __user *buffer, size_t *lenp, loff_t *ppos)
3227 {
3228         return -ENOSYS;
3229 }
3230
3231 int proc_doulongvec_minmax(struct ctl_table *table, int write,
3232                     void __user *buffer, size_t *lenp, loff_t *ppos)
3233 {
3234         return -ENOSYS;
3235 }
3236
3237 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
3238                                       void __user *buffer,
3239                                       size_t *lenp, loff_t *ppos)
3240 {
3241     return -ENOSYS;
3242 }
3243
3244
3245 #endif /* CONFIG_PROC_SYSCTL */
3246
3247 /*
3248  * No sense putting this after each symbol definition, twice,
3249  * exception granted :-)
3250  */
3251 EXPORT_SYMBOL(proc_dointvec);
3252 EXPORT_SYMBOL(proc_douintvec);
3253 EXPORT_SYMBOL(proc_dointvec_jiffies);
3254 EXPORT_SYMBOL(proc_dointvec_minmax);
3255 EXPORT_SYMBOL_GPL(proc_douintvec_minmax);
3256 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
3257 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
3258 EXPORT_SYMBOL(proc_dostring);
3259 EXPORT_SYMBOL(proc_doulongvec_minmax);
3260 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);