kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/mm.h>
  23 #include <linux/swap.h>
  24 #include <linux/slab.h>
  25 #include <linux/sysctl.h>
  26 #include <linux/bitmap.h>
  27 #include <linux/signal.h>
  28 #include <linux/printk.h>
  29 #include <linux/proc_fs.h>
  30 #include <linux/security.h>
  31 #include <linux/ctype.h>
  32 #include <linux/kmemcheck.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
  66 #include <asm/uaccess.h>
  67 #include <asm/processor.h>
  68
  69 #ifdef CONFIG_X86
  70 #include <asm/nmi.h>
  71 #include <asm/stacktrace.h>
  72 #include <asm/io.h>
  73 #endif
  74 #ifdef CONFIG_SPARC
  75 #include <asm/setup.h>
  76 #endif
  77 #ifdef CONFIG_BSD_PROCESS_ACCT
  78 #include <linux/acct.h>
  79 #endif
  80 #ifdef CONFIG_RT_MUTEXES
  81 #include <linux/rtmutex.h>
  82 #endif
  83 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  84 #include <linux/lockdep.h>
  85 #endif
  86 #ifdef CONFIG_CHR_DEV_SG
  87 #include <scsi/sg.h>
  88 #endif
  89
  90 #ifdef CONFIG_LOCKUP_DETECTOR
  91 #include <linux/nmi.h>
  92 #endif
  93
  94
  95 #if defined(CONFIG_SYSCTL)
  96
  97 /* External variables not in a header file. */
  98 extern int sysctl_overcommit_memory;
  99 extern int sysctl_overcommit_ratio;
 100 extern int max_threads;
 101 extern int suid_dumpable;
 102 #ifdef CONFIG_COREDUMP
 103 extern int core_uses_pid;
 104 extern char core_pattern[];
 105 extern unsigned int core_pipe_limit;
 106 #endif
 107 extern int pid_max;
 108 extern int pid_max_min, pid_max_max;
 109 extern int percpu_pagelist_fraction;
 110 extern int compat_log;
 111 extern int latencytop_enabled;
 112 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 113 #ifndef CONFIG_MMU
 114 extern int sysctl_nr_trim_pages;
 115 #endif
 116 #ifdef CONFIG_BLOCK
 117 extern int blk_iopoll_enabled;
 118 #endif
 119
 120 /* Constants used for minimum and  maximum */
 121 #ifdef CONFIG_LOCKUP_DETECTOR
 122 static int sixty = 60;
 123 #endif
 124
 125 static int neg_one = -1;
 126 static int zero;
 127 static int __maybe_unused one = 1;
 128 static int __maybe_unused two = 2;
 129 static int __maybe_unused three = 3;
 130 static unsigned long one_ul = 1;
 131 static int one_hundred = 100;
 132 #ifdef CONFIG_PRINTK
 133 static int ten_thousand = 10000;
 134 #endif
 135
 136 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 137 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 138
 139 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 140 static int maxolduid = 65535;
 141 static int minolduid;
 142 static int min_percpu_pagelist_fract = 8;
 143
 144 static int ngroups_max = NGROUPS_MAX;
 145 static const int cap_last_cap = CAP_LAST_CAP;
 146
 147 #ifdef CONFIG_INOTIFY_USER
 148 #include <linux/inotify.h>
 149 #endif
 150 #ifdef CONFIG_SPARC
 151 #endif
 152
 153 #ifdef CONFIG_SPARC64
 154 extern int sysctl_tsb_ratio;
 155 #endif
 156
 157 #ifdef __hppa__
 158 extern int pwrsw_enabled;
 159 #endif
 160
 161 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 162 extern int unaligned_enabled;
 163 #endif
 164
 165 #ifdef CONFIG_IA64
 166 extern int unaligned_dump_stack;
 167 #endif
 168
 169 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 170 extern int no_unaligned_warning;
 171 #endif
 172
 173 #ifdef CONFIG_PROC_SYSCTL
 174 static int proc_do_cad_pid(struct ctl_table *table, int write,
 175                   void __user *buffer, size_t *lenp, loff_t *ppos);
 176 static int proc_taint(struct ctl_table *table, int write,
 177                                void __user *buffer, size_t *lenp, loff_t *ppos);
 178 #endif
 179
 180 #ifdef CONFIG_PRINTK
 181 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 182                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 183 #endif
 184
 185 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 186                 void __user *buffer, size_t *lenp, loff_t *ppos);
 187 #ifdef CONFIG_COREDUMP
 188 static int proc_dostring_coredump(struct ctl_table *table, int write,
 189                 void __user *buffer, size_t *lenp, loff_t *ppos);
 190 #endif
 191
 192 #ifdef CONFIG_MAGIC_SYSRQ
 193 /* Note: sysrq code uses it's own private copy */
 194 static int __sysrq_enabled = CONFIG_MAGIC_SYSRQ_DEFAULT_ENABLE;
 195
 196 static int sysrq_sysctl_handler(ctl_table *table, int write,
 197                                 void __user *buffer, size_t *lenp,
 198                                 loff_t *ppos)
 199 {
 200         int error;
 201
 202         error = proc_dointvec(table, write, buffer, lenp, ppos);
 203         if (error)
 204                 return error;
 205
 206         if (write)
 207                 sysrq_toggle_support(__sysrq_enabled);
 208
 209         return 0;
 210 }
 211
 212 #endif
 213
 214 static struct ctl_table kern_table[];
 215 static struct ctl_table vm_table[];
 216 static struct ctl_table fs_table[];
 217 static struct ctl_table debug_table[];
 218 static struct ctl_table dev_table[];
 219 extern struct ctl_table random_table[];
 220 #ifdef CONFIG_EPOLL
 221 extern struct ctl_table epoll_table[];
 222 #endif
 223
 224 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 225 int sysctl_legacy_va_layout;
 226 #endif
 227
 228 /* The default sysctl tables: */
 229
 230 static struct ctl_table sysctl_base_table[] = {
 231         {
 232                 .procname       = "kernel",
 233                 .mode           = 0555,
 234                 .child          = kern_table,
 235         },
 236         {
 237                 .procname       = "vm",
 238                 .mode           = 0555,
 239                 .child          = vm_table,
 240         },
 241         {
 242                 .procname       = "fs",
 243                 .mode           = 0555,
 244                 .child          = fs_table,
 245         },
 246         {
 247                 .procname       = "debug",
 248                 .mode           = 0555,
 249                 .child          = debug_table,
 250         },
 251         {
 252                 .procname       = "dev",
 253                 .mode           = 0555,
 254                 .child          = dev_table,
 255         },
 256         { }
 257 };
 258
 259 #ifdef CONFIG_SCHED_DEBUG
 260 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 261 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 262 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 263 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 264 #ifdef CONFIG_SMP
 265 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 266 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 267 #endif /* CONFIG_SMP */
 268 #endif /* CONFIG_SCHED_DEBUG */
 269
 270 #ifdef CONFIG_COMPACTION
 271 static int min_extfrag_threshold;
 272 static int max_extfrag_threshold = 1000;
 273 #endif
 274
 275 static struct ctl_table kern_table[] = {
 276         {
 277                 .procname       = "sched_child_runs_first",
 278                 .data           = &sysctl_sched_child_runs_first,
 279                 .maxlen         = sizeof(unsigned int),
 280                 .mode           = 0644,
 281                 .proc_handler   = proc_dointvec,
 282         },
 283 #ifdef CONFIG_SCHED_DEBUG
 284         {
 285                 .procname       = "sched_min_granularity_ns",
 286                 .data           = &sysctl_sched_min_granularity,
 287                 .maxlen         = sizeof(unsigned int),
 288                 .mode           = 0644,
 289                 .proc_handler   = sched_proc_update_handler,
 290                 .extra1         = &min_sched_granularity_ns,
 291                 .extra2         = &max_sched_granularity_ns,
 292         },
 293         {
 294                 .procname       = "sched_latency_ns",
 295                 .data           = &sysctl_sched_latency,
 296                 .maxlen         = sizeof(unsigned int),
 297                 .mode           = 0644,
 298                 .proc_handler   = sched_proc_update_handler,
 299                 .extra1         = &min_sched_granularity_ns,
 300                 .extra2         = &max_sched_granularity_ns,
 301         },
 302         {
 303                 .procname       = "sched_wakeup_granularity_ns",
 304                 .data           = &sysctl_sched_wakeup_granularity,
 305                 .maxlen         = sizeof(unsigned int),
 306                 .mode           = 0644,
 307                 .proc_handler   = sched_proc_update_handler,
 308                 .extra1         = &min_wakeup_granularity_ns,
 309                 .extra2         = &max_wakeup_granularity_ns,
 310         },
 311 #ifdef CONFIG_SMP
 312         {
 313                 .procname       = "sched_tunable_scaling",
 314                 .data           = &sysctl_sched_tunable_scaling,
 315                 .maxlen         = sizeof(enum sched_tunable_scaling),
 316                 .mode           = 0644,
 317                 .proc_handler   = sched_proc_update_handler,
 318                 .extra1         = &min_sched_tunable_scaling,
 319                 .extra2         = &max_sched_tunable_scaling,
 320         },
 321         {
 322                 .procname       = "sched_migration_cost_ns",
 323                 .data           = &sysctl_sched_migration_cost,
 324                 .maxlen         = sizeof(unsigned int),
 325                 .mode           = 0644,
 326                 .proc_handler   = proc_dointvec,
 327         },
 328         {
 329                 .procname       = "sched_nr_migrate",
 330                 .data           = &sysctl_sched_nr_migrate,
 331                 .maxlen         = sizeof(unsigned int),
 332                 .mode           = 0644,
 333                 .proc_handler   = proc_dointvec,
 334         },
 335         {
 336                 .procname       = "sched_time_avg_ms",
 337                 .data           = &sysctl_sched_time_avg,
 338                 .maxlen         = sizeof(unsigned int),
 339                 .mode           = 0644,
 340                 .proc_handler   = proc_dointvec,
 341         },
 342         {
 343                 .procname       = "sched_shares_window_ns",
 344                 .data           = &sysctl_sched_shares_window,
 345                 .maxlen         = sizeof(unsigned int),
 346                 .mode           = 0644,
 347                 .proc_handler   = proc_dointvec,
 348         },
 349         {
 350                 .procname       = "timer_migration",
 351                 .data           = &sysctl_timer_migration,
 352                 .maxlen         = sizeof(unsigned int),
 353                 .mode           = 0644,
 354                 .proc_handler   = proc_dointvec_minmax,
 355                 .extra1         = &zero,
 356                 .extra2         = &one,
 357         },
 358 #endif /* CONFIG_SMP */
 359 #ifdef CONFIG_NUMA_BALANCING
 360         {
 361                 .procname       = "numa_balancing_scan_delay_ms",
 362                 .data           = &sysctl_numa_balancing_scan_delay,
 363                 .maxlen         = sizeof(unsigned int),
 364                 .mode           = 0644,
 365                 .proc_handler   = proc_dointvec,
 366         },
 367         {
 368                 .procname       = "numa_balancing_scan_period_min_ms",
 369                 .data           = &sysctl_numa_balancing_scan_period_min,
 370                 .maxlen         = sizeof(unsigned int),
 371                 .mode           = 0644,
 372                 .proc_handler   = proc_dointvec,
 373         },
 374         {
 375                 .procname       = "numa_balancing_scan_period_max_ms",
 376                 .data           = &sysctl_numa_balancing_scan_period_max,
 377                 .maxlen         = sizeof(unsigned int),
 378                 .mode           = 0644,
 379                 .proc_handler   = proc_dointvec,
 380         },
 381         {
 382                 .procname       = "numa_balancing_scan_size_mb",
 383                 .data           = &sysctl_numa_balancing_scan_size,
 384                 .maxlen         = sizeof(unsigned int),
 385                 .mode           = 0644,
 386                 .proc_handler   = proc_dointvec,
 387         },
 388         {
 389                 .procname       = "numa_balancing_migrate_deferred",
 390                 .data           = &sysctl_numa_balancing_migrate_deferred,
 391                 .maxlen         = sizeof(unsigned int),
 392                 .mode           = 0644,
 393                 .proc_handler   = proc_dointvec,
 394         },
 395 #endif /* CONFIG_NUMA_BALANCING */
 396 #endif /* CONFIG_SCHED_DEBUG */
 397         {
 398                 .procname       = "sched_rt_period_us",
 399                 .data           = &sysctl_sched_rt_period,
 400                 .maxlen         = sizeof(unsigned int),
 401                 .mode           = 0644,
 402                 .proc_handler   = sched_rt_handler,
 403         },
 404         {
 405                 .procname       = "sched_rt_runtime_us",
 406                 .data           = &sysctl_sched_rt_runtime,
 407                 .maxlen         = sizeof(int),
 408                 .mode           = 0644,
 409                 .proc_handler   = sched_rt_handler,
 410         },
 411         {
 412                 .procname       = "sched_rr_timeslice_ms",
 413                 .data           = &sched_rr_timeslice,
 414                 .maxlen         = sizeof(int),
 415                 .mode           = 0644,
 416                 .proc_handler   = sched_rr_handler,
 417         },
 418 #ifdef CONFIG_SCHED_AUTOGROUP
 419         {
 420                 .procname       = "sched_autogroup_enabled",
 421                 .data           = &sysctl_sched_autogroup_enabled,
 422                 .maxlen         = sizeof(unsigned int),
 423                 .mode           = 0644,
 424                 .proc_handler   = proc_dointvec_minmax,
 425                 .extra1         = &zero,
 426                 .extra2         = &one,
 427         },
 428 #endif
 429 #ifdef CONFIG_CFS_BANDWIDTH
 430         {
 431                 .procname       = "sched_cfs_bandwidth_slice_us",
 432                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 433                 .maxlen         = sizeof(unsigned int),
 434                 .mode           = 0644,
 435                 .proc_handler   = proc_dointvec_minmax,
 436                 .extra1         = &one,
 437         },
 438 #endif
 439 #ifdef CONFIG_PROVE_LOCKING
 440         {
 441                 .procname       = "prove_locking",
 442                 .data           = &prove_locking,
 443                 .maxlen         = sizeof(int),
 444                 .mode           = 0644,
 445                 .proc_handler   = proc_dointvec,
 446         },
 447 #endif
 448 #ifdef CONFIG_LOCK_STAT
 449         {
 450                 .procname       = "lock_stat",
 451                 .data           = &lock_stat,
 452                 .maxlen         = sizeof(int),
 453                 .mode           = 0644,
 454                 .proc_handler   = proc_dointvec,
 455         },
 456 #endif
 457         {
 458                 .procname       = "panic",
 459                 .data           = &panic_timeout,
 460                 .maxlen         = sizeof(int),
 461                 .mode           = 0644,
 462                 .proc_handler   = proc_dointvec,
 463         },
 464 #ifdef CONFIG_COREDUMP
 465         {
 466                 .procname       = "core_uses_pid",
 467                 .data           = &core_uses_pid,
 468                 .maxlen         = sizeof(int),
 469                 .mode           = 0644,
 470                 .proc_handler   = proc_dointvec,
 471         },
 472         {
 473                 .procname       = "core_pattern",
 474                 .data           = core_pattern,
 475                 .maxlen         = CORENAME_MAX_SIZE,
 476                 .mode           = 0644,
 477                 .proc_handler   = proc_dostring_coredump,
 478         },
 479         {
 480                 .procname       = "core_pipe_limit",
 481                 .data           = &core_pipe_limit,
 482                 .maxlen         = sizeof(unsigned int),
 483                 .mode           = 0644,
 484                 .proc_handler   = proc_dointvec,
 485         },
 486 #endif
 487 #ifdef CONFIG_PROC_SYSCTL
 488         {
 489                 .procname       = "tainted",
 490                 .maxlen         = sizeof(long),
 491                 .mode           = 0644,
 492                 .proc_handler   = proc_taint,
 493         },
 494 #endif
 495 #ifdef CONFIG_LATENCYTOP
 496         {
 497                 .procname       = "latencytop",
 498                 .data           = &latencytop_enabled,
 499                 .maxlen         = sizeof(int),
 500                 .mode           = 0644,
 501                 .proc_handler   = proc_dointvec,
 502         },
 503 #endif
 504 #ifdef CONFIG_BLK_DEV_INITRD
 505         {
 506                 .procname       = "real-root-dev",
 507                 .data           = &real_root_dev,
 508                 .maxlen         = sizeof(int),
 509                 .mode           = 0644,
 510                 .proc_handler   = proc_dointvec,
 511         },
 512 #endif
 513         {
 514                 .procname       = "print-fatal-signals",
 515                 .data           = &print_fatal_signals,
 516                 .maxlen         = sizeof(int),
 517                 .mode           = 0644,
 518                 .proc_handler   = proc_dointvec,
 519         },
 520 #ifdef CONFIG_SPARC
 521         {
 522                 .procname       = "reboot-cmd",
 523                 .data           = reboot_command,
 524                 .maxlen         = 256,
 525                 .mode           = 0644,
 526                 .proc_handler   = proc_dostring,
 527         },
 528         {
 529                 .procname       = "stop-a",
 530                 .data           = &stop_a_enabled,
 531                 .maxlen         = sizeof (int),
 532                 .mode           = 0644,
 533                 .proc_handler   = proc_dointvec,
 534         },
 535         {
 536                 .procname       = "scons-poweroff",
 537                 .data           = &scons_pwroff,
 538                 .maxlen         = sizeof (int),
 539                 .mode           = 0644,
 540                 .proc_handler   = proc_dointvec,
 541         },
 542 #endif
 543 #ifdef CONFIG_SPARC64
 544         {
 545                 .procname       = "tsb-ratio",
 546                 .data           = &sysctl_tsb_ratio,
 547                 .maxlen         = sizeof (int),
 548                 .mode           = 0644,
 549                 .proc_handler   = proc_dointvec,
 550         },
 551 #endif
 552 #ifdef __hppa__
 553         {
 554                 .procname       = "soft-power",
 555                 .data           = &pwrsw_enabled,
 556                 .maxlen         = sizeof (int),
 557                 .mode           = 0644,
 558                 .proc_handler   = proc_dointvec,
 559         },
 560 #endif
 561 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 562         {
 563                 .procname       = "unaligned-trap",
 564                 .data           = &unaligned_enabled,
 565                 .maxlen         = sizeof (int),
 566                 .mode           = 0644,
 567                 .proc_handler   = proc_dointvec,
 568         },
 569 #endif
 570         {
 571                 .procname       = "ctrl-alt-del",
 572                 .data           = &C_A_D,
 573                 .maxlen         = sizeof(int),
 574                 .mode           = 0644,
 575                 .proc_handler   = proc_dointvec,
 576         },
 577 #ifdef CONFIG_FUNCTION_TRACER
 578         {
 579                 .procname       = "ftrace_enabled",
 580                 .data           = &ftrace_enabled,
 581                 .maxlen         = sizeof(int),
 582                 .mode           = 0644,
 583                 .proc_handler   = ftrace_enable_sysctl,
 584         },
 585 #endif
 586 #ifdef CONFIG_STACK_TRACER
 587         {
 588                 .procname       = "stack_tracer_enabled",
 589                 .data           = &stack_tracer_enabled,
 590                 .maxlen         = sizeof(int),
 591                 .mode           = 0644,
 592                 .proc_handler   = stack_trace_sysctl,
 593         },
 594 #endif
 595 #ifdef CONFIG_TRACING
 596         {
 597                 .procname       = "ftrace_dump_on_oops",
 598                 .data           = &ftrace_dump_on_oops,
 599                 .maxlen         = sizeof(int),
 600                 .mode           = 0644,
 601                 .proc_handler   = proc_dointvec,
 602         },
 603         {
 604                 .procname       = "traceoff_on_warning",
 605                 .data           = &__disable_trace_on_warning,
 606                 .maxlen         = sizeof(__disable_trace_on_warning),
 607                 .mode           = 0644,
 608                 .proc_handler   = proc_dointvec,
 609         },
 610 #endif
 611 #ifdef CONFIG_MODULES
 612         {
 613                 .procname       = "modprobe",
 614                 .data           = &modprobe_path,
 615                 .maxlen         = KMOD_PATH_LEN,
 616                 .mode           = 0644,
 617                 .proc_handler   = proc_dostring,
 618         },
 619         {
 620                 .procname       = "modules_disabled",
 621                 .data           = &modules_disabled,
 622                 .maxlen         = sizeof(int),
 623                 .mode           = 0644,
 624                 /* only handle a transition from default "0" to "1" */
 625                 .proc_handler   = proc_dointvec_minmax,
 626                 .extra1         = &one,
 627                 .extra2         = &one,
 628         },
 629 #endif
 630
 631         {
 632                 .procname       = "hotplug",
 633                 .data           = &uevent_helper,
 634                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 635                 .mode           = 0644,
 636                 .proc_handler   = proc_dostring,
 637         },
 638
 639 #ifdef CONFIG_CHR_DEV_SG
 640         {
 641                 .procname       = "sg-big-buff",
 642                 .data           = &sg_big_buff,
 643                 .maxlen         = sizeof (int),
 644                 .mode           = 0444,
 645                 .proc_handler   = proc_dointvec,
 646         },
 647 #endif
 648 #ifdef CONFIG_BSD_PROCESS_ACCT
 649         {
 650                 .procname       = "acct",
 651                 .data           = &acct_parm,
 652                 .maxlen         = 3*sizeof(int),
 653                 .mode           = 0644,
 654                 .proc_handler   = proc_dointvec,
 655         },
 656 #endif
 657 #ifdef CONFIG_MAGIC_SYSRQ
 658         {
 659                 .procname       = "sysrq",
 660                 .data           = &__sysrq_enabled,
 661                 .maxlen         = sizeof (int),
 662                 .mode           = 0644,
 663                 .proc_handler   = sysrq_sysctl_handler,
 664         },
 665 #endif
 666 #ifdef CONFIG_PROC_SYSCTL
 667         {
 668                 .procname       = "cad_pid",
 669                 .data           = NULL,
 670                 .maxlen         = sizeof (int),
 671                 .mode           = 0600,
 672                 .proc_handler   = proc_do_cad_pid,
 673         },
 674 #endif
 675         {
 676                 .procname       = "threads-max",
 677                 .data           = &max_threads,
 678                 .maxlen         = sizeof(int),
 679                 .mode           = 0644,
 680                 .proc_handler   = proc_dointvec,
 681         },
 682         {
 683                 .procname       = "random",
 684                 .mode           = 0555,
 685                 .child          = random_table,
 686         },
 687         {
 688                 .procname       = "usermodehelper",
 689                 .mode           = 0555,
 690                 .child          = usermodehelper_table,
 691         },
 692         {
 693                 .procname       = "overflowuid",
 694                 .data           = &overflowuid,
 695                 .maxlen         = sizeof(int),
 696                 .mode           = 0644,
 697                 .proc_handler   = proc_dointvec_minmax,
 698                 .extra1         = &minolduid,
 699                 .extra2         = &maxolduid,
 700         },
 701         {
 702                 .procname       = "overflowgid",
 703                 .data           = &overflowgid,
 704                 .maxlen         = sizeof(int),
 705                 .mode           = 0644,
 706                 .proc_handler   = proc_dointvec_minmax,
 707                 .extra1         = &minolduid,
 708                 .extra2         = &maxolduid,
 709         },
 710 #ifdef CONFIG_S390
 711 #ifdef CONFIG_MATHEMU
 712         {
 713                 .procname       = "ieee_emulation_warnings",
 714                 .data           = &sysctl_ieee_emulation_warnings,
 715                 .maxlen         = sizeof(int),
 716                 .mode           = 0644,
 717                 .proc_handler   = proc_dointvec,
 718         },
 719 #endif
 720         {
 721                 .procname       = "userprocess_debug",
 722                 .data           = &show_unhandled_signals,
 723                 .maxlen         = sizeof(int),
 724                 .mode           = 0644,
 725                 .proc_handler   = proc_dointvec,
 726         },
 727 #endif
 728         {
 729                 .procname       = "pid_max",
 730                 .data           = &pid_max,
 731                 .maxlen         = sizeof (int),
 732                 .mode           = 0644,
 733                 .proc_handler   = proc_dointvec_minmax,
 734                 .extra1         = &pid_max_min,
 735                 .extra2         = &pid_max_max,
 736         },
 737         {
 738                 .procname       = "panic_on_oops",
 739                 .data           = &panic_on_oops,
 740                 .maxlen         = sizeof(int),
 741                 .mode           = 0644,
 742                 .proc_handler   = proc_dointvec,
 743         },
 744 #if defined CONFIG_PRINTK
 745         {
 746                 .procname       = "printk",
 747                 .data           = &console_loglevel,
 748                 .maxlen         = 4*sizeof(int),
 749                 .mode           = 0644,
 750                 .proc_handler   = proc_dointvec,
 751         },
 752         {
 753                 .procname       = "printk_ratelimit",
 754                 .data           = &printk_ratelimit_state.interval,
 755                 .maxlen         = sizeof(int),
 756                 .mode           = 0644,
 757                 .proc_handler   = proc_dointvec_jiffies,
 758         },
 759         {
 760                 .procname       = "printk_ratelimit_burst",
 761                 .data           = &printk_ratelimit_state.burst,
 762                 .maxlen         = sizeof(int),
 763                 .mode           = 0644,
 764                 .proc_handler   = proc_dointvec,
 765         },
 766         {
 767                 .procname       = "printk_delay",
 768                 .data           = &printk_delay_msec,
 769                 .maxlen         = sizeof(int),
 770                 .mode           = 0644,
 771                 .proc_handler   = proc_dointvec_minmax,
 772                 .extra1         = &zero,
 773                 .extra2         = &ten_thousand,
 774         },
 775         {
 776                 .procname       = "dmesg_restrict",
 777                 .data           = &dmesg_restrict,
 778                 .maxlen         = sizeof(int),
 779                 .mode           = 0644,
 780                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 781                 .extra1         = &zero,
 782                 .extra2         = &one,
 783         },
 784         {
 785                 .procname       = "kptr_restrict",
 786                 .data           = &kptr_restrict,
 787                 .maxlen         = sizeof(int),
 788                 .mode           = 0644,
 789                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 790                 .extra1         = &zero,
 791                 .extra2         = &two,
 792         },
 793 #endif
 794         {
 795                 .procname       = "ngroups_max",
 796                 .data           = &ngroups_max,
 797                 .maxlen         = sizeof (int),
 798                 .mode           = 0444,
 799                 .proc_handler   = proc_dointvec,
 800         },
 801         {
 802                 .procname       = "cap_last_cap",
 803                 .data           = (void *)&cap_last_cap,
 804                 .maxlen         = sizeof(int),
 805                 .mode           = 0444,
 806                 .proc_handler   = proc_dointvec,
 807         },
 808 #if defined(CONFIG_LOCKUP_DETECTOR)
 809         {
 810                 .procname       = "watchdog",
 811                 .data           = &watchdog_user_enabled,
 812                 .maxlen         = sizeof (int),
 813                 .mode           = 0644,
 814                 .proc_handler   = proc_dowatchdog,
 815                 .extra1         = &zero,
 816                 .extra2         = &one,
 817         },
 818         {
 819                 .procname       = "watchdog_thresh",
 820                 .data           = &watchdog_thresh,
 821                 .maxlen         = sizeof(int),
 822                 .mode           = 0644,
 823                 .proc_handler   = proc_dowatchdog,
 824                 .extra1         = &zero,
 825                 .extra2         = &sixty,
 826         },
 827         {
 828                 .procname       = "softlockup_panic",
 829                 .data           = &softlockup_panic,
 830                 .maxlen         = sizeof(int),
 831                 .mode           = 0644,
 832                 .proc_handler   = proc_dointvec_minmax,
 833                 .extra1         = &zero,
 834                 .extra2         = &one,
 835         },
 836         {
 837                 .procname       = "nmi_watchdog",
 838                 .data           = &watchdog_user_enabled,
 839                 .maxlen         = sizeof (int),
 840                 .mode           = 0644,
 841                 .proc_handler   = proc_dowatchdog,
 842                 .extra1         = &zero,
 843                 .extra2         = &one,
 844         },
 845 #endif
 846 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 847         {
 848                 .procname       = "unknown_nmi_panic",
 849                 .data           = &unknown_nmi_panic,
 850                 .maxlen         = sizeof (int),
 851                 .mode           = 0644,
 852                 .proc_handler   = proc_dointvec,
 853         },
 854 #endif
 855 #if defined(CONFIG_X86)
 856         {
 857                 .procname       = "panic_on_unrecovered_nmi",
 858                 .data           = &panic_on_unrecovered_nmi,
 859                 .maxlen         = sizeof(int),
 860                 .mode           = 0644,
 861                 .proc_handler   = proc_dointvec,
 862         },
 863         {
 864                 .procname       = "panic_on_io_nmi",
 865                 .data           = &panic_on_io_nmi,
 866                 .maxlen         = sizeof(int),
 867                 .mode           = 0644,
 868                 .proc_handler   = proc_dointvec,
 869         },
 870 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 871         {
 872                 .procname       = "panic_on_stackoverflow",
 873                 .data           = &sysctl_panic_on_stackoverflow,
 874                 .maxlen         = sizeof(int),
 875                 .mode           = 0644,
 876                 .proc_handler   = proc_dointvec,
 877         },
 878 #endif
 879         {
 880                 .procname       = "bootloader_type",
 881                 .data           = &bootloader_type,
 882                 .maxlen         = sizeof (int),
 883                 .mode           = 0444,
 884                 .proc_handler   = proc_dointvec,
 885         },
 886         {
 887                 .procname       = "bootloader_version",
 888                 .data           = &bootloader_version,
 889                 .maxlen         = sizeof (int),
 890                 .mode           = 0444,
 891                 .proc_handler   = proc_dointvec,
 892         },
 893         {
 894                 .procname       = "kstack_depth_to_print",
 895                 .data           = &kstack_depth_to_print,
 896                 .maxlen         = sizeof(int),
 897                 .mode           = 0644,
 898                 .proc_handler   = proc_dointvec,
 899         },
 900         {
 901                 .procname       = "io_delay_type",
 902                 .data           = &io_delay_type,
 903                 .maxlen         = sizeof(int),
 904                 .mode           = 0644,
 905                 .proc_handler   = proc_dointvec,
 906         },
 907 #endif
 908 #if defined(CONFIG_MMU)
 909         {
 910                 .procname       = "randomize_va_space",
 911                 .data           = &randomize_va_space,
 912                 .maxlen         = sizeof(int),
 913                 .mode           = 0644,
 914                 .proc_handler   = proc_dointvec,
 915         },
 916 #endif
 917 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 918         {
 919                 .procname       = "spin_retry",
 920                 .data           = &spin_retry,
 921                 .maxlen         = sizeof (int),
 922                 .mode           = 0644,
 923                 .proc_handler   = proc_dointvec,
 924         },
 925 #endif
 926 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 927         {
 928                 .procname       = "acpi_video_flags",
 929                 .data           = &acpi_realmode_flags,
 930                 .maxlen         = sizeof (unsigned long),
 931                 .mode           = 0644,
 932                 .proc_handler   = proc_doulongvec_minmax,
 933         },
 934 #endif
 935 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 936         {
 937                 .procname       = "ignore-unaligned-usertrap",
 938                 .data           = &no_unaligned_warning,
 939                 .maxlen         = sizeof (int),
 940                 .mode           = 0644,
 941                 .proc_handler   = proc_dointvec,
 942         },
 943 #endif
 944 #ifdef CONFIG_IA64
 945         {
 946                 .procname       = "unaligned-dump-stack",
 947                 .data           = &unaligned_dump_stack,
 948                 .maxlen         = sizeof (int),
 949                 .mode           = 0644,
 950                 .proc_handler   = proc_dointvec,
 951         },
 952 #endif
 953 #ifdef CONFIG_DETECT_HUNG_TASK
 954         {
 955                 .procname       = "hung_task_panic",
 956                 .data           = &sysctl_hung_task_panic,
 957                 .maxlen         = sizeof(int),
 958                 .mode           = 0644,
 959                 .proc_handler   = proc_dointvec_minmax,
 960                 .extra1         = &zero,
 961                 .extra2         = &one,
 962         },
 963         {
 964                 .procname       = "hung_task_check_count",
 965                 .data           = &sysctl_hung_task_check_count,
 966                 .maxlen         = sizeof(int),
 967                 .mode           = 0644,
 968                 .proc_handler   = proc_dointvec_minmax,
 969                 .extra1         = &zero,
 970         },
 971         {
 972                 .procname       = "hung_task_timeout_secs",
 973                 .data           = &sysctl_hung_task_timeout_secs,
 974                 .maxlen         = sizeof(unsigned long),
 975                 .mode           = 0644,
 976                 .proc_handler   = proc_dohung_task_timeout_secs,
 977         },
 978         {
 979                 .procname       = "hung_task_warnings",
 980                 .data           = &sysctl_hung_task_warnings,
 981                 .maxlen         = sizeof(unsigned long),
 982                 .mode           = 0644,
 983                 .proc_handler   = proc_doulongvec_minmax,
 984         },
 985 #endif
 986 #ifdef CONFIG_COMPAT
 987         {
 988                 .procname       = "compat-log",
 989                 .data           = &compat_log,
 990                 .maxlen         = sizeof (int),
 991                 .mode           = 0644,
 992                 .proc_handler   = proc_dointvec,
 993         },
 994 #endif
 995 #ifdef CONFIG_RT_MUTEXES
 996         {
 997                 .procname       = "max_lock_depth",
 998                 .data           = &max_lock_depth,
 999                 .maxlen         = sizeof(int),
1000                 .mode           = 0644,
1001                 .proc_handler   = proc_dointvec,
1002         },
1003 #endif
1004         {
1005                 .procname       = "poweroff_cmd",
1006                 .data           = &poweroff_cmd,
1007                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1008                 .mode           = 0644,
1009                 .proc_handler   = proc_dostring,
1010         },
1011 #ifdef CONFIG_KEYS
1012         {
1013                 .procname       = "keys",
1014                 .mode           = 0555,
1015                 .child          = key_sysctls,
1016         },
1017 #endif
1018 #ifdef CONFIG_RCU_TORTURE_TEST
1019         {
1020                 .procname       = "rcutorture_runnable",
1021                 .data           = &rcutorture_runnable,
1022                 .maxlen         = sizeof(int),
1023                 .mode           = 0644,
1024                 .proc_handler   = proc_dointvec,
1025         },
1026 #endif
1027 #ifdef CONFIG_PERF_EVENTS
1028         /*
1029          * User-space scripts rely on the existence of this file
1030          * as a feature check for perf_events being enabled.
1031          *
1032          * So it's an ABI, do not remove!
1033          */
1034         {
1035                 .procname       = "perf_event_paranoid",
1036                 .data           = &sysctl_perf_event_paranoid,
1037                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1038                 .mode           = 0644,
1039                 .proc_handler   = proc_dointvec,
1040         },
1041         {
1042                 .procname       = "perf_event_mlock_kb",
1043                 .data           = &sysctl_perf_event_mlock,
1044                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1045                 .mode           = 0644,
1046                 .proc_handler   = proc_dointvec,
1047         },
1048         {
1049                 .procname       = "perf_event_max_sample_rate",
1050                 .data           = &sysctl_perf_event_sample_rate,
1051                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1052                 .mode           = 0644,
1053                 .proc_handler   = perf_proc_update_handler,
1054                 .extra1         = &one,
1055         },
1056         {
1057                 .procname       = "perf_cpu_time_max_percent",
1058                 .data           = &sysctl_perf_cpu_time_max_percent,
1059                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1060                 .mode           = 0644,
1061                 .proc_handler   = perf_cpu_time_max_percent_handler,
1062                 .extra1         = &zero,
1063                 .extra2         = &one_hundred,
1064         },
1065 #endif
1066 #ifdef CONFIG_KMEMCHECK
1067         {
1068                 .procname       = "kmemcheck",
1069                 .data           = &kmemcheck_enabled,
1070                 .maxlen         = sizeof(int),
1071                 .mode           = 0644,
1072                 .proc_handler   = proc_dointvec,
1073         },
1074 #endif
1075 #ifdef CONFIG_BLOCK
1076         {
1077                 .procname       = "blk_iopoll",
1078                 .data           = &blk_iopoll_enabled,
1079                 .maxlen         = sizeof(int),
1080                 .mode           = 0644,
1081                 .proc_handler   = proc_dointvec,
1082         },
1083 #endif
1084         { }
1085 };
1086
1087 static struct ctl_table vm_table[] = {
1088         {
1089                 .procname       = "overcommit_memory",
1090                 .data           = &sysctl_overcommit_memory,
1091                 .maxlen         = sizeof(sysctl_overcommit_memory),
1092                 .mode           = 0644,
1093                 .proc_handler   = proc_dointvec_minmax,
1094                 .extra1         = &zero,
1095                 .extra2         = &two,
1096         },
1097         {
1098                 .procname       = "panic_on_oom",
1099                 .data           = &sysctl_panic_on_oom,
1100                 .maxlen         = sizeof(sysctl_panic_on_oom),
1101                 .mode           = 0644,
1102                 .proc_handler   = proc_dointvec_minmax,
1103                 .extra1         = &zero,
1104                 .extra2         = &two,
1105         },
1106         {
1107                 .procname       = "oom_kill_allocating_task",
1108                 .data           = &sysctl_oom_kill_allocating_task,
1109                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1110                 .mode           = 0644,
1111                 .proc_handler   = proc_dointvec,
1112         },
1113         {
1114                 .procname       = "oom_dump_tasks",
1115                 .data           = &sysctl_oom_dump_tasks,
1116                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1117                 .mode           = 0644,
1118                 .proc_handler   = proc_dointvec,
1119         },
1120         {
1121                 .procname       = "overcommit_ratio",
1122                 .data           = &sysctl_overcommit_ratio,
1123                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1124                 .mode           = 0644,
1125                 .proc_handler   = proc_dointvec,
1126         },
1127         {
1128                 .procname       = "page-cluster",
1129                 .data           = &page_cluster,
1130                 .maxlen         = sizeof(int),
1131                 .mode           = 0644,
1132                 .proc_handler   = proc_dointvec_minmax,
1133                 .extra1         = &zero,
1134         },
1135         {
1136                 .procname       = "dirty_background_ratio",
1137                 .data           = &dirty_background_ratio,
1138                 .maxlen         = sizeof(dirty_background_ratio),
1139                 .mode           = 0644,
1140                 .proc_handler   = dirty_background_ratio_handler,
1141                 .extra1         = &zero,
1142                 .extra2         = &one_hundred,
1143         },
1144         {
1145                 .procname       = "dirty_background_bytes",
1146                 .data           = &dirty_background_bytes,
1147                 .maxlen         = sizeof(dirty_background_bytes),
1148                 .mode           = 0644,
1149                 .proc_handler   = dirty_background_bytes_handler,
1150                 .extra1         = &one_ul,
1151         },
1152         {
1153                 .procname       = "dirty_ratio",
1154                 .data           = &vm_dirty_ratio,
1155                 .maxlen         = sizeof(vm_dirty_ratio),
1156                 .mode           = 0644,
1157                 .proc_handler   = dirty_ratio_handler,
1158                 .extra1         = &zero,
1159                 .extra2         = &one_hundred,
1160         },
1161         {
1162                 .procname       = "dirty_bytes",
1163                 .data           = &vm_dirty_bytes,
1164                 .maxlen         = sizeof(vm_dirty_bytes),
1165                 .mode           = 0644,
1166                 .proc_handler   = dirty_bytes_handler,
1167                 .extra1         = &dirty_bytes_min,
1168         },
1169         {
1170                 .procname       = "dirty_writeback_centisecs",
1171                 .data           = &dirty_writeback_interval,
1172                 .maxlen         = sizeof(dirty_writeback_interval),
1173                 .mode           = 0644,
1174                 .proc_handler   = dirty_writeback_centisecs_handler,
1175         },
1176         {
1177                 .procname       = "dirty_expire_centisecs",
1178                 .data           = &dirty_expire_interval,
1179                 .maxlen         = sizeof(dirty_expire_interval),
1180                 .mode           = 0644,
1181                 .proc_handler   = proc_dointvec_minmax,
1182                 .extra1         = &zero,
1183         },
1184         {
1185                 .procname       = "nr_pdflush_threads",
1186                 .mode           = 0444 /* read-only */,
1187                 .proc_handler   = pdflush_proc_obsolete,
1188         },
1189         {
1190                 .procname       = "swappiness",
1191                 .data           = &vm_swappiness,
1192                 .maxlen         = sizeof(vm_swappiness),
1193                 .mode           = 0644,
1194                 .proc_handler   = proc_dointvec_minmax,
1195                 .extra1         = &zero,
1196                 .extra2         = &one_hundred,
1197         },
1198 #ifdef CONFIG_HUGETLB_PAGE
1199         {
1200                 .procname       = "nr_hugepages",
1201                 .data           = NULL,
1202                 .maxlen         = sizeof(unsigned long),
1203                 .mode           = 0644,
1204                 .proc_handler   = hugetlb_sysctl_handler,
1205                 .extra1         = (void *)&hugetlb_zero,
1206                 .extra2         = (void *)&hugetlb_infinity,
1207         },
1208 #ifdef CONFIG_NUMA
1209         {
1210                 .procname       = "nr_hugepages_mempolicy",
1211                 .data           = NULL,
1212                 .maxlen         = sizeof(unsigned long),
1213                 .mode           = 0644,
1214                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1215                 .extra1         = (void *)&hugetlb_zero,
1216                 .extra2         = (void *)&hugetlb_infinity,
1217         },
1218 #endif
1219          {
1220                 .procname       = "hugetlb_shm_group",
1221                 .data           = &sysctl_hugetlb_shm_group,
1222                 .maxlen         = sizeof(gid_t),
1223                 .mode           = 0644,
1224                 .proc_handler   = proc_dointvec,
1225          },
1226          {
1227                 .procname       = "hugepages_treat_as_movable",
1228                 .data           = &hugepages_treat_as_movable,
1229                 .maxlen         = sizeof(int),
1230                 .mode           = 0644,
1231                 .proc_handler   = proc_dointvec,
1232         },
1233         {
1234                 .procname       = "nr_overcommit_hugepages",
1235                 .data           = NULL,
1236                 .maxlen         = sizeof(unsigned long),
1237                 .mode           = 0644,
1238                 .proc_handler   = hugetlb_overcommit_handler,
1239                 .extra1         = (void *)&hugetlb_zero,
1240                 .extra2         = (void *)&hugetlb_infinity,
1241         },
1242 #endif
1243         {
1244                 .procname       = "lowmem_reserve_ratio",
1245                 .data           = &sysctl_lowmem_reserve_ratio,
1246                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1247                 .mode           = 0644,
1248                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1249         },
1250         {
1251                 .procname       = "drop_caches",
1252                 .data           = &sysctl_drop_caches,
1253                 .maxlen         = sizeof(int),
1254                 .mode           = 0644,
1255                 .proc_handler   = drop_caches_sysctl_handler,
1256                 .extra1         = &one,
1257                 .extra2         = &three,
1258         },
1259 #ifdef CONFIG_COMPACTION
1260         {
1261                 .procname       = "compact_memory",
1262                 .data           = &sysctl_compact_memory,
1263                 .maxlen         = sizeof(int),
1264                 .mode           = 0200,
1265                 .proc_handler   = sysctl_compaction_handler,
1266         },
1267         {
1268                 .procname       = "extfrag_threshold",
1269                 .data           = &sysctl_extfrag_threshold,
1270                 .maxlen         = sizeof(int),
1271                 .mode           = 0644,
1272                 .proc_handler   = sysctl_extfrag_handler,
1273                 .extra1         = &min_extfrag_threshold,
1274                 .extra2         = &max_extfrag_threshold,
1275         },
1276
1277 #endif /* CONFIG_COMPACTION */
1278         {
1279                 .procname       = "min_free_kbytes",
1280                 .data           = &min_free_kbytes,
1281                 .maxlen         = sizeof(min_free_kbytes),
1282                 .mode           = 0644,
1283                 .proc_handler   = min_free_kbytes_sysctl_handler,
1284                 .extra1         = &zero,
1285         },
1286         {
1287                 .procname       = "percpu_pagelist_fraction",
1288                 .data           = &percpu_pagelist_fraction,
1289                 .maxlen         = sizeof(percpu_pagelist_fraction),
1290                 .mode           = 0644,
1291                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1292                 .extra1         = &min_percpu_pagelist_fract,
1293         },
1294 #ifdef CONFIG_MMU
1295         {
1296                 .procname       = "max_map_count",
1297                 .data           = &sysctl_max_map_count,
1298                 .maxlen         = sizeof(sysctl_max_map_count),
1299                 .mode           = 0644,
1300                 .proc_handler   = proc_dointvec_minmax,
1301                 .extra1         = &zero,
1302         },
1303 #else
1304         {
1305                 .procname       = "nr_trim_pages",
1306                 .data           = &sysctl_nr_trim_pages,
1307                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1308                 .mode           = 0644,
1309                 .proc_handler   = proc_dointvec_minmax,
1310                 .extra1         = &zero,
1311         },
1312 #endif
1313         {
1314                 .procname       = "laptop_mode",
1315                 .data           = &laptop_mode,
1316                 .maxlen         = sizeof(laptop_mode),
1317                 .mode           = 0644,
1318                 .proc_handler   = proc_dointvec_jiffies,
1319         },
1320         {
1321                 .procname       = "block_dump",
1322                 .data           = &block_dump,
1323                 .maxlen         = sizeof(block_dump),
1324                 .mode           = 0644,
1325                 .proc_handler   = proc_dointvec,
1326                 .extra1         = &zero,
1327         },
1328         {
1329                 .procname       = "vfs_cache_pressure",
1330                 .data           = &sysctl_vfs_cache_pressure,
1331                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1332                 .mode           = 0644,
1333                 .proc_handler   = proc_dointvec,
1334                 .extra1         = &zero,
1335         },
1336 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1337         {
1338                 .procname       = "legacy_va_layout",
1339                 .data           = &sysctl_legacy_va_layout,
1340                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1341                 .mode           = 0644,
1342                 .proc_handler   = proc_dointvec,
1343                 .extra1         = &zero,
1344         },
1345 #endif
1346 #ifdef CONFIG_NUMA
1347         {
1348                 .procname       = "zone_reclaim_mode",
1349                 .data           = &zone_reclaim_mode,
1350                 .maxlen         = sizeof(zone_reclaim_mode),
1351                 .mode           = 0644,
1352                 .proc_handler   = proc_dointvec,
1353                 .extra1         = &zero,
1354         },
1355         {
1356                 .procname       = "min_unmapped_ratio",
1357                 .data           = &sysctl_min_unmapped_ratio,
1358                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1359                 .mode           = 0644,
1360                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1361                 .extra1         = &zero,
1362                 .extra2         = &one_hundred,
1363         },
1364         {
1365                 .procname       = "min_slab_ratio",
1366                 .data           = &sysctl_min_slab_ratio,
1367                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1368                 .mode           = 0644,
1369                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1370                 .extra1         = &zero,
1371                 .extra2         = &one_hundred,
1372         },
1373 #endif
1374 #ifdef CONFIG_SMP
1375         {
1376                 .procname       = "stat_interval",
1377                 .data           = &sysctl_stat_interval,
1378                 .maxlen         = sizeof(sysctl_stat_interval),
1379                 .mode           = 0644,
1380                 .proc_handler   = proc_dointvec_jiffies,
1381         },
1382 #endif
1383 #ifdef CONFIG_MMU
1384         {
1385                 .procname       = "mmap_min_addr",
1386                 .data           = &dac_mmap_min_addr,
1387                 .maxlen         = sizeof(unsigned long),
1388                 .mode           = 0644,
1389                 .proc_handler   = mmap_min_addr_handler,
1390         },
1391 #endif
1392 #ifdef CONFIG_NUMA
1393         {
1394                 .procname       = "numa_zonelist_order",
1395                 .data           = &numa_zonelist_order,
1396                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1397                 .mode           = 0644,
1398                 .proc_handler   = numa_zonelist_order_handler,
1399         },
1400 #endif
1401 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1402    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1403         {
1404                 .procname       = "vdso_enabled",
1405                 .data           = &vdso_enabled,
1406                 .maxlen         = sizeof(vdso_enabled),
1407                 .mode           = 0644,
1408                 .proc_handler   = proc_dointvec,
1409                 .extra1         = &zero,
1410         },
1411 #endif
1412 #ifdef CONFIG_HIGHMEM
1413         {
1414                 .procname       = "highmem_is_dirtyable",
1415                 .data           = &vm_highmem_is_dirtyable,
1416                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1417                 .mode           = 0644,
1418                 .proc_handler   = proc_dointvec_minmax,
1419                 .extra1         = &zero,
1420                 .extra2         = &one,
1421         },
1422 #endif
1423         {
1424                 .procname       = "scan_unevictable_pages",
1425                 .data           = &scan_unevictable_pages,
1426                 .maxlen         = sizeof(scan_unevictable_pages),
1427                 .mode           = 0644,
1428                 .proc_handler   = scan_unevictable_handler,
1429         },
1430 #ifdef CONFIG_MEMORY_FAILURE
1431         {
1432                 .procname       = "memory_failure_early_kill",
1433                 .data           = &sysctl_memory_failure_early_kill,
1434                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1435                 .mode           = 0644,
1436                 .proc_handler   = proc_dointvec_minmax,
1437                 .extra1         = &zero,
1438                 .extra2         = &one,
1439         },
1440         {
1441                 .procname       = "memory_failure_recovery",
1442                 .data           = &sysctl_memory_failure_recovery,
1443                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1444                 .mode           = 0644,
1445                 .proc_handler   = proc_dointvec_minmax,
1446                 .extra1         = &zero,
1447                 .extra2         = &one,
1448         },
1449 #endif
1450         {
1451                 .procname       = "user_reserve_kbytes",
1452                 .data           = &sysctl_user_reserve_kbytes,
1453                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1454                 .mode           = 0644,
1455                 .proc_handler   = proc_doulongvec_minmax,
1456         },
1457         {
1458                 .procname       = "admin_reserve_kbytes",
1459                 .data           = &sysctl_admin_reserve_kbytes,
1460                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1461                 .mode           = 0644,
1462                 .proc_handler   = proc_doulongvec_minmax,
1463         },
1464         { }
1465 };
1466
1467 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1468 static struct ctl_table binfmt_misc_table[] = {
1469         { }
1470 };
1471 #endif
1472
1473 static struct ctl_table fs_table[] = {
1474         {
1475                 .procname       = "inode-nr",
1476                 .data           = &inodes_stat,
1477                 .maxlen         = 2*sizeof(long),
1478                 .mode           = 0444,
1479                 .proc_handler   = proc_nr_inodes,
1480         },
1481         {
1482                 .procname       = "inode-state",
1483                 .data           = &inodes_stat,
1484                 .maxlen         = 7*sizeof(long),
1485                 .mode           = 0444,
1486                 .proc_handler   = proc_nr_inodes,
1487         },
1488         {
1489                 .procname       = "file-nr",
1490                 .data           = &files_stat,
1491                 .maxlen         = sizeof(files_stat),
1492                 .mode           = 0444,
1493                 .proc_handler   = proc_nr_files,
1494         },
1495         {
1496                 .procname       = "file-max",
1497                 .data           = &files_stat.max_files,
1498                 .maxlen         = sizeof(files_stat.max_files),
1499                 .mode           = 0644,
1500                 .proc_handler   = proc_doulongvec_minmax,
1501         },
1502         {
1503                 .procname       = "nr_open",
1504                 .data           = &sysctl_nr_open,
1505                 .maxlen         = sizeof(int),
1506                 .mode           = 0644,
1507                 .proc_handler   = proc_dointvec_minmax,
1508                 .extra1         = &sysctl_nr_open_min,
1509                 .extra2         = &sysctl_nr_open_max,
1510         },
1511         {
1512                 .procname       = "dentry-state",
1513                 .data           = &dentry_stat,
1514                 .maxlen         = 6*sizeof(long),
1515                 .mode           = 0444,
1516                 .proc_handler   = proc_nr_dentry,
1517         },
1518         {
1519                 .procname       = "overflowuid",
1520                 .data           = &fs_overflowuid,
1521                 .maxlen         = sizeof(int),
1522                 .mode           = 0644,
1523                 .proc_handler   = proc_dointvec_minmax,
1524                 .extra1         = &minolduid,
1525                 .extra2         = &maxolduid,
1526         },
1527         {
1528                 .procname       = "overflowgid",
1529                 .data           = &fs_overflowgid,
1530                 .maxlen         = sizeof(int),
1531                 .mode           = 0644,
1532                 .proc_handler   = proc_dointvec_minmax,
1533                 .extra1         = &minolduid,
1534                 .extra2         = &maxolduid,
1535         },
1536 #ifdef CONFIG_FILE_LOCKING
1537         {
1538                 .procname       = "leases-enable",
1539                 .data           = &leases_enable,
1540                 .maxlen         = sizeof(int),
1541                 .mode           = 0644,
1542                 .proc_handler   = proc_dointvec,
1543         },
1544 #endif
1545 #ifdef CONFIG_DNOTIFY
1546         {
1547                 .procname       = "dir-notify-enable",
1548                 .data           = &dir_notify_enable,
1549                 .maxlen         = sizeof(int),
1550                 .mode           = 0644,
1551                 .proc_handler   = proc_dointvec,
1552         },
1553 #endif
1554 #ifdef CONFIG_MMU
1555 #ifdef CONFIG_FILE_LOCKING
1556         {
1557                 .procname       = "lease-break-time",
1558                 .data           = &lease_break_time,
1559                 .maxlen         = sizeof(int),
1560                 .mode           = 0644,
1561                 .proc_handler   = proc_dointvec,
1562         },
1563 #endif
1564 #ifdef CONFIG_AIO
1565         {
1566                 .procname       = "aio-nr",
1567                 .data           = &aio_nr,
1568                 .maxlen         = sizeof(aio_nr),
1569                 .mode           = 0444,
1570                 .proc_handler   = proc_doulongvec_minmax,
1571         },
1572         {
1573                 .procname       = "aio-max-nr",
1574                 .data           = &aio_max_nr,
1575                 .maxlen         = sizeof(aio_max_nr),
1576                 .mode           = 0644,
1577                 .proc_handler   = proc_doulongvec_minmax,
1578         },
1579 #endif /* CONFIG_AIO */
1580 #ifdef CONFIG_INOTIFY_USER
1581         {
1582                 .procname       = "inotify",
1583                 .mode           = 0555,
1584                 .child          = inotify_table,
1585         },
1586 #endif
1587 #ifdef CONFIG_EPOLL
1588         {
1589                 .procname       = "epoll",
1590                 .mode           = 0555,
1591                 .child          = epoll_table,
1592         },
1593 #endif
1594 #endif
1595         {
1596                 .procname       = "protected_symlinks",
1597                 .data           = &sysctl_protected_symlinks,
1598                 .maxlen         = sizeof(int),
1599                 .mode           = 0600,
1600                 .proc_handler   = proc_dointvec_minmax,
1601                 .extra1         = &zero,
1602                 .extra2         = &one,
1603         },
1604         {
1605                 .procname       = "protected_hardlinks",
1606                 .data           = &sysctl_protected_hardlinks,
1607                 .maxlen         = sizeof(int),
1608                 .mode           = 0600,
1609                 .proc_handler   = proc_dointvec_minmax,
1610                 .extra1         = &zero,
1611                 .extra2         = &one,
1612         },
1613         {
1614                 .procname       = "suid_dumpable",
1615                 .data           = &suid_dumpable,
1616                 .maxlen         = sizeof(int),
1617                 .mode           = 0644,
1618                 .proc_handler   = proc_dointvec_minmax_coredump,
1619                 .extra1         = &zero,
1620                 .extra2         = &two,
1621         },
1622 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1623         {
1624                 .procname       = "binfmt_misc",
1625                 .mode           = 0555,
1626                 .child          = binfmt_misc_table,
1627         },
1628 #endif
1629         {
1630                 .procname       = "pipe-max-size",
1631                 .data           = &pipe_max_size,
1632                 .maxlen         = sizeof(int),
1633                 .mode           = 0644,
1634                 .proc_handler   = &pipe_proc_fn,
1635                 .extra1         = &pipe_min_size,
1636         },
1637         { }
1638 };
1639
1640 static struct ctl_table debug_table[] = {
1641 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1642         {
1643                 .procname       = "exception-trace",
1644                 .data           = &show_unhandled_signals,
1645                 .maxlen         = sizeof(int),
1646                 .mode           = 0644,
1647                 .proc_handler   = proc_dointvec
1648         },
1649 #endif
1650 #if defined(CONFIG_OPTPROBES)
1651         {
1652                 .procname       = "kprobes-optimization",
1653                 .data           = &sysctl_kprobes_optimization,
1654                 .maxlen         = sizeof(int),
1655                 .mode           = 0644,
1656                 .proc_handler   = proc_kprobes_optimization_handler,
1657                 .extra1         = &zero,
1658                 .extra2         = &one,
1659         },
1660 #endif
1661         { }
1662 };
1663
1664 static struct ctl_table dev_table[] = {
1665         { }
1666 };
1667
1668 int __init sysctl_init(void)
1669 {
1670         struct ctl_table_header *hdr;
1671
1672         hdr = register_sysctl_table(sysctl_base_table);
1673         kmemleak_not_leak(hdr);
1674         return 0;
1675 }
1676
1677 #endif /* CONFIG_SYSCTL */
1678
1679 /*
1680  * /proc/sys support
1681  */
1682
1683 #ifdef CONFIG_PROC_SYSCTL
1684
1685 static int _proc_do_string(void* data, int maxlen, int write,
1686                            void __user *buffer,
1687                            size_t *lenp, loff_t *ppos)
1688 {
1689         size_t len;
1690         char __user *p;
1691         char c;
1692
1693         if (!data || !maxlen || !*lenp) {
1694                 *lenp = 0;
1695                 return 0;
1696         }
1697
1698         if (write) {
1699                 len = 0;
1700                 p = buffer;
1701                 while (len < *lenp) {
1702                         if (get_user(c, p++))
1703                                 return -EFAULT;
1704                         if (c == 0 || c == '\n')
1705                                 break;
1706                         len++;
1707                 }
1708                 if (len >= maxlen)
1709                         len = maxlen-1;
1710                 if(copy_from_user(data, buffer, len))
1711                         return -EFAULT;
1712                 ((char *) data)[len] = 0;
1713                 *ppos += *lenp;
1714         } else {
1715                 len = strlen(data);
1716                 if (len > maxlen)
1717                         len = maxlen;
1718
1719                 if (*ppos > len) {
1720                         *lenp = 0;
1721                         return 0;
1722                 }
1723
1724                 data += *ppos;
1725                 len  -= *ppos;
1726
1727                 if (len > *lenp)
1728                         len = *lenp;
1729                 if (len)
1730                         if(copy_to_user(buffer, data, len))
1731                                 return -EFAULT;
1732                 if (len < *lenp) {
1733                         if(put_user('\n', ((char __user *) buffer) + len))
1734                                 return -EFAULT;
1735                         len++;
1736                 }
1737                 *lenp = len;
1738                 *ppos += len;
1739         }
1740         return 0;
1741 }
1742
1743 /**
1744  * proc_dostring - read a string sysctl
1745  * @table: the sysctl table
1746  * @write: %TRUE if this is a write to the sysctl file
1747  * @buffer: the user buffer
1748  * @lenp: the size of the user buffer
1749  * @ppos: file position
1750  *
1751  * Reads/writes a string from/to the user buffer. If the kernel
1752  * buffer provided is not large enough to hold the string, the
1753  * string is truncated. The copied string is %NULL-terminated.
1754  * If the string is being read by the user process, it is copied
1755  * and a newline '\n' is added. It is truncated if the buffer is
1756  * not large enough.
1757  *
1758  * Returns 0 on success.
1759  */
1760 int proc_dostring(struct ctl_table *table, int write,
1761                   void __user *buffer, size_t *lenp, loff_t *ppos)
1762 {
1763         return _proc_do_string(table->data, table->maxlen, write,
1764                                buffer, lenp, ppos);
1765 }
1766
1767 static size_t proc_skip_spaces(char **buf)
1768 {
1769         size_t ret;
1770         char *tmp = skip_spaces(*buf);
1771         ret = tmp - *buf;
1772         *buf = tmp;
1773         return ret;
1774 }
1775
1776 static void proc_skip_char(char **buf, size_t *size, const char v)
1777 {
1778         while (*size) {
1779                 if (**buf != v)
1780                         break;
1781                 (*size)--;
1782                 (*buf)++;
1783         }
1784 }
1785
1786 #define TMPBUFLEN 22
1787 /**
1788  * proc_get_long - reads an ASCII formatted integer from a user buffer
1789  *
1790  * @buf: a kernel buffer
1791  * @size: size of the kernel buffer
1792  * @val: this is where the number will be stored
1793  * @neg: set to %TRUE if number is negative
1794  * @perm_tr: a vector which contains the allowed trailers
1795  * @perm_tr_len: size of the perm_tr vector
1796  * @tr: pointer to store the trailer character
1797  *
1798  * In case of success %0 is returned and @buf and @size are updated with
1799  * the amount of bytes read. If @tr is non-NULL and a trailing
1800  * character exists (size is non-zero after returning from this
1801  * function), @tr is updated with the trailing character.
1802  */
1803 static int proc_get_long(char **buf, size_t *size,
1804                           unsigned long *val, bool *neg,
1805                           const char *perm_tr, unsigned perm_tr_len, char *tr)
1806 {
1807         int len;
1808         char *p, tmp[TMPBUFLEN];
1809
1810         if (!*size)
1811                 return -EINVAL;
1812
1813         len = *size;
1814         if (len > TMPBUFLEN - 1)
1815                 len = TMPBUFLEN - 1;
1816
1817         memcpy(tmp, *buf, len);
1818
1819         tmp[len] = 0;
1820         p = tmp;
1821         if (*p == '-' && *size > 1) {
1822                 *neg = true;
1823                 p++;
1824         } else
1825                 *neg = false;
1826         if (!isdigit(*p))
1827                 return -EINVAL;
1828
1829         *val = simple_strtoul(p, &p, 0);
1830
1831         len = p - tmp;
1832
1833         /* We don't know if the next char is whitespace thus we may accept
1834          * invalid integers (e.g. 1234...a) or two integers instead of one
1835          * (e.g. 123...1). So lets not allow such large numbers. */
1836         if (len == TMPBUFLEN - 1)
1837                 return -EINVAL;
1838
1839         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
1840                 return -EINVAL;
1841
1842         if (tr && (len < *size))
1843                 *tr = *p;
1844
1845         *buf += len;
1846         *size -= len;
1847
1848         return 0;
1849 }
1850
1851 /**
1852  * proc_put_long - converts an integer to a decimal ASCII formatted string
1853  *
1854  * @buf: the user buffer
1855  * @size: the size of the user buffer
1856  * @val: the integer to be converted
1857  * @neg: sign of the number, %TRUE for negative
1858  *
1859  * In case of success %0 is returned and @buf and @size are updated with
1860  * the amount of bytes written.
1861  */
1862 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
1863                           bool neg)
1864 {
1865         int len;
1866         char tmp[TMPBUFLEN], *p = tmp;
1867
1868         sprintf(p, "%s%lu", neg ? "-" : "", val);
1869         len = strlen(tmp);
1870         if (len > *size)
1871                 len = *size;
1872         if (copy_to_user(*buf, tmp, len))
1873                 return -EFAULT;
1874         *size -= len;
1875         *buf += len;
1876         return 0;
1877 }
1878 #undef TMPBUFLEN
1879
1880 static int proc_put_char(void __user **buf, size_t *size, char c)
1881 {
1882         if (*size) {
1883                 char __user **buffer = (char __user **)buf;
1884                 if (put_user(c, *buffer))
1885                         return -EFAULT;
1886                 (*size)--, (*buffer)++;
1887                 *buf = *buffer;
1888         }
1889         return 0;
1890 }
1891
1892 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
1893                                  int *valp,
1894                                  int write, void *data)
1895 {
1896         if (write) {
1897                 *valp = *negp ? -*lvalp : *lvalp;
1898         } else {
1899                 int val = *valp;
1900                 if (val < 0) {
1901                         *negp = true;
1902                         *lvalp = (unsigned long)-val;
1903                 } else {
1904                         *negp = false;
1905                         *lvalp = (unsigned long)val;
1906                 }
1907         }
1908         return 0;
1909 }
1910
1911 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
1912
1913 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
1914                   int write, void __user *buffer,
1915                   size_t *lenp, loff_t *ppos,
1916                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
1917                               int write, void *data),
1918                   void *data)
1919 {
1920         int *i, vleft, first = 1, err = 0;
1921         unsigned long page = 0;
1922         size_t left;
1923         char *kbuf;
1924
1925         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
1926                 *lenp = 0;
1927                 return 0;
1928         }
1929
1930         i = (int *) tbl_data;
1931         vleft = table->maxlen / sizeof(*i);
1932         left = *lenp;
1933
1934         if (!conv)
1935                 conv = do_proc_dointvec_conv;
1936
1937         if (write) {
1938                 if (left > PAGE_SIZE - 1)
1939                         left = PAGE_SIZE - 1;
1940                 page = __get_free_page(GFP_TEMPORARY);
1941                 kbuf = (char *) page;
1942                 if (!kbuf)
1943                         return -ENOMEM;
1944                 if (copy_from_user(kbuf, buffer, left)) {
1945                         err = -EFAULT;
1946                         goto free;
1947                 }
1948                 kbuf[left] = 0;
1949         }
1950
1951         for (; left && vleft--; i++, first=0) {
1952                 unsigned long lval;
1953                 bool neg;
1954
1955                 if (write) {
1956                         left -= proc_skip_spaces(&kbuf);
1957
1958                         if (!left)
1959                                 break;
1960                         err = proc_get_long(&kbuf, &left, &lval, &neg,
1961                                              proc_wspace_sep,
1962                                              sizeof(proc_wspace_sep), NULL);
1963                         if (err)
1964                                 break;
1965                         if (conv(&neg, &lval, i, 1, data)) {
1966                                 err = -EINVAL;
1967                                 break;
1968                         }
1969                 } else {
1970                         if (conv(&neg, &lval, i, 0, data)) {
1971                                 err = -EINVAL;
1972                                 break;
1973                         }
1974                         if (!first)
1975                                 err = proc_put_char(&buffer, &left, '\t');
1976                         if (err)
1977                                 break;
1978                         err = proc_put_long(&buffer, &left, lval, neg);
1979                         if (err)
1980                                 break;
1981                 }
1982         }
1983
1984         if (!write && !first && left && !err)
1985                 err = proc_put_char(&buffer, &left, '\n');
1986         if (write && !err && left)
1987                 left -= proc_skip_spaces(&kbuf);
1988 free:
1989         if (write) {
1990                 free_page(page);
1991                 if (first)
1992                         return err ? : -EINVAL;
1993         }
1994         *lenp -= left;
1995         *ppos += *lenp;
1996         return err;
1997 }
1998
1999 static int do_proc_dointvec(struct ctl_table *table, int write,
2000                   void __user *buffer, size_t *lenp, loff_t *ppos,
2001                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2002                               int write, void *data),
2003                   void *data)
2004 {
2005         return __do_proc_dointvec(table->data, table, write,
2006                         buffer, lenp, ppos, conv, data);
2007 }
2008
2009 /**
2010  * proc_dointvec - read a vector of integers
2011  * @table: the sysctl table
2012  * @write: %TRUE if this is a write to the sysctl file
2013  * @buffer: the user buffer
2014  * @lenp: the size of the user buffer
2015  * @ppos: file position
2016  *
2017  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2018  * values from/to the user buffer, treated as an ASCII string.
2019  *
2020  * Returns 0 on success.
2021  */
2022 int proc_dointvec(struct ctl_table *table, int write,
2023                      void __user *buffer, size_t *lenp, loff_t *ppos)
2024 {
2025     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2026                             NULL,NULL);
2027 }
2028
2029 /*
2030  * Taint values can only be increased
2031  * This means we can safely use a temporary.
2032  */
2033 static int proc_taint(struct ctl_table *table, int write,
2034                                void __user *buffer, size_t *lenp, loff_t *ppos)
2035 {
2036         struct ctl_table t;
2037         unsigned long tmptaint = get_taint();
2038         int err;
2039
2040         if (write && !capable(CAP_SYS_ADMIN))
2041                 return -EPERM;
2042
2043         t = *table;
2044         t.data = &tmptaint;
2045         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2046         if (err < 0)
2047                 return err;
2048
2049         if (write) {
2050                 /*
2051                  * Poor man's atomic or. Not worth adding a primitive
2052                  * to everyone's atomic.h for this
2053                  */
2054                 int i;
2055                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2056                         if ((tmptaint >> i) & 1)
2057                                 add_taint(i, LOCKDEP_STILL_OK);
2058                 }
2059         }
2060
2061         return err;
2062 }
2063
2064 #ifdef CONFIG_PRINTK
2065 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2066                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2067 {
2068         if (write && !capable(CAP_SYS_ADMIN))
2069                 return -EPERM;
2070
2071         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2072 }
2073 #endif
2074
2075 struct do_proc_dointvec_minmax_conv_param {
2076         int *min;
2077         int *max;
2078 };
2079
2080 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2081                                         int *valp,
2082                                         int write, void *data)
2083 {
2084         struct do_proc_dointvec_minmax_conv_param *param = data;
2085         if (write) {
2086                 int val = *negp ? -*lvalp : *lvalp;
2087                 if ((param->min && *param->min > val) ||
2088                     (param->max && *param->max < val))
2089                         return -EINVAL;
2090                 *valp = val;
2091         } else {
2092                 int val = *valp;
2093                 if (val < 0) {
2094                         *negp = true;
2095                         *lvalp = (unsigned long)-val;
2096                 } else {
2097                         *negp = false;
2098                         *lvalp = (unsigned long)val;
2099                 }
2100         }
2101         return 0;
2102 }
2103
2104 /**
2105  * proc_dointvec_minmax - read a vector of integers with min/max values
2106  * @table: the sysctl table
2107  * @write: %TRUE if this is a write to the sysctl file
2108  * @buffer: the user buffer
2109  * @lenp: the size of the user buffer
2110  * @ppos: file position
2111  *
2112  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2113  * values from/to the user buffer, treated as an ASCII string.
2114  *
2115  * This routine will ensure the values are within the range specified by
2116  * table->extra1 (min) and table->extra2 (max).
2117  *
2118  * Returns 0 on success.
2119  */
2120 int proc_dointvec_minmax(struct ctl_table *table, int write,
2121                   void __user *buffer, size_t *lenp, loff_t *ppos)
2122 {
2123         struct do_proc_dointvec_minmax_conv_param param = {
2124                 .min = (int *) table->extra1,
2125                 .max = (int *) table->extra2,
2126         };
2127         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2128                                 do_proc_dointvec_minmax_conv, &param);
2129 }
2130
2131 static void validate_coredump_safety(void)
2132 {
2133 #ifdef CONFIG_COREDUMP
2134         if (suid_dumpable == SUID_DUMP_ROOT &&
2135             core_pattern[0] != '/' && core_pattern[0] != '|') {
2136                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2137                         "suid_dumpable=2. Pipe handler or fully qualified "\
2138                         "core dump path required.\n");
2139         }
2140 #endif
2141 }
2142
2143 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2144                 void __user *buffer, size_t *lenp, loff_t *ppos)
2145 {
2146         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2147         if (!error)
2148                 validate_coredump_safety();
2149         return error;
2150 }
2151
2152 #ifdef CONFIG_COREDUMP
2153 static int proc_dostring_coredump(struct ctl_table *table, int write,
2154                   void __user *buffer, size_t *lenp, loff_t *ppos)
2155 {
2156         int error = proc_dostring(table, write, buffer, lenp, ppos);
2157         if (!error)
2158                 validate_coredump_safety();
2159         return error;
2160 }
2161 #endif
2162
2163 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2164                                      void __user *buffer,
2165                                      size_t *lenp, loff_t *ppos,
2166                                      unsigned long convmul,
2167                                      unsigned long convdiv)
2168 {
2169         unsigned long *i, *min, *max;
2170         int vleft, first = 1, err = 0;
2171         unsigned long page = 0;
2172         size_t left;
2173         char *kbuf;
2174
2175         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2176                 *lenp = 0;
2177                 return 0;
2178         }
2179
2180         i = (unsigned long *) data;
2181         min = (unsigned long *) table->extra1;
2182         max = (unsigned long *) table->extra2;
2183         vleft = table->maxlen / sizeof(unsigned long);
2184         left = *lenp;
2185
2186         if (write) {
2187                 if (left > PAGE_SIZE - 1)
2188                         left = PAGE_SIZE - 1;
2189                 page = __get_free_page(GFP_TEMPORARY);
2190                 kbuf = (char *) page;
2191                 if (!kbuf)
2192                         return -ENOMEM;
2193                 if (copy_from_user(kbuf, buffer, left)) {
2194                         err = -EFAULT;
2195                         goto free;
2196                 }
2197                 kbuf[left] = 0;
2198         }
2199
2200         for (; left && vleft--; i++, first = 0) {
2201                 unsigned long val;
2202
2203                 if (write) {
2204                         bool neg;
2205
2206                         left -= proc_skip_spaces(&kbuf);
2207
2208                         err = proc_get_long(&kbuf, &left, &val, &neg,
2209                                              proc_wspace_sep,
2210                                              sizeof(proc_wspace_sep), NULL);
2211                         if (err)
2212                                 break;
2213                         if (neg)
2214                                 continue;
2215                         if ((min && val < *min) || (max && val > *max))
2216                                 continue;
2217                         *i = val;
2218                 } else {
2219                         val = convdiv * (*i) / convmul;
2220                         if (!first) {
2221                                 err = proc_put_char(&buffer, &left, '\t');
2222                                 if (err)
2223                                         break;
2224                         }
2225                         err = proc_put_long(&buffer, &left, val, false);
2226                         if (err)
2227                                 break;
2228                 }
2229         }
2230
2231         if (!write && !first && left && !err)
2232                 err = proc_put_char(&buffer, &left, '\n');
2233         if (write && !err)
2234                 left -= proc_skip_spaces(&kbuf);
2235 free:
2236         if (write) {
2237                 free_page(page);
2238                 if (first)
2239                         return err ? : -EINVAL;
2240         }
2241         *lenp -= left;
2242         *ppos += *lenp;
2243         return err;
2244 }
2245
2246 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2247                                      void __user *buffer,
2248                                      size_t *lenp, loff_t *ppos,
2249                                      unsigned long convmul,
2250                                      unsigned long convdiv)
2251 {
2252         return __do_proc_doulongvec_minmax(table->data, table, write,
2253                         buffer, lenp, ppos, convmul, convdiv);
2254 }
2255
2256 /**
2257  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2258  * @table: the sysctl table
2259  * @write: %TRUE if this is a write to the sysctl file
2260  * @buffer: the user buffer
2261  * @lenp: the size of the user buffer
2262  * @ppos: file position
2263  *
2264  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2265  * values from/to the user buffer, treated as an ASCII string.
2266  *
2267  * This routine will ensure the values are within the range specified by
2268  * table->extra1 (min) and table->extra2 (max).
2269  *
2270  * Returns 0 on success.
2271  */
2272 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2273                            void __user *buffer, size_t *lenp, loff_t *ppos)
2274 {
2275     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2276 }
2277
2278 /**
2279  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2280  * @table: the sysctl table
2281  * @write: %TRUE if this is a write to the sysctl file
2282  * @buffer: the user buffer
2283  * @lenp: the size of the user buffer
2284  * @ppos: file position
2285  *
2286  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2287  * values from/to the user buffer, treated as an ASCII string. The values
2288  * are treated as milliseconds, and converted to jiffies when they are stored.
2289  *
2290  * This routine will ensure the values are within the range specified by
2291  * table->extra1 (min) and table->extra2 (max).
2292  *
2293  * Returns 0 on success.
2294  */
2295 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2296                                       void __user *buffer,
2297                                       size_t *lenp, loff_t *ppos)
2298 {
2299     return do_proc_doulongvec_minmax(table, write, buffer,
2300                                      lenp, ppos, HZ, 1000l);
2301 }
2302
2303
2304 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2305                                          int *valp,
2306                                          int write, void *data)
2307 {
2308         if (write) {
2309                 if (*lvalp > LONG_MAX / HZ)
2310                         return 1;
2311                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2312         } else {
2313                 int val = *valp;
2314                 unsigned long lval;
2315                 if (val < 0) {
2316                         *negp = true;
2317                         lval = (unsigned long)-val;
2318                 } else {
2319                         *negp = false;
2320                         lval = (unsigned long)val;
2321                 }
2322                 *lvalp = lval / HZ;
2323         }
2324         return 0;
2325 }
2326
2327 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2328                                                 int *valp,
2329                                                 int write, void *data)
2330 {
2331         if (write) {
2332                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2333                         return 1;
2334                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2335         } else {
2336                 int val = *valp;
2337                 unsigned long lval;
2338                 if (val < 0) {
2339                         *negp = true;
2340                         lval = (unsigned long)-val;
2341                 } else {
2342                         *negp = false;
2343                         lval = (unsigned long)val;
2344                 }
2345                 *lvalp = jiffies_to_clock_t(lval);
2346         }
2347         return 0;
2348 }
2349
2350 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2351                                             int *valp,
2352                                             int write, void *data)
2353 {
2354         if (write) {
2355                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2356
2357                 if (jif > INT_MAX)
2358                         return 1;
2359                 *valp = (int)jif;
2360         } else {
2361                 int val = *valp;
2362                 unsigned long lval;
2363                 if (val < 0) {
2364                         *negp = true;
2365                         lval = (unsigned long)-val;
2366                 } else {
2367                         *negp = false;
2368                         lval = (unsigned long)val;
2369                 }
2370                 *lvalp = jiffies_to_msecs(lval);
2371         }
2372         return 0;
2373 }
2374
2375 /**
2376  * proc_dointvec_jiffies - read a vector of integers as seconds
2377  * @table: the sysctl table
2378  * @write: %TRUE if this is a write to the sysctl file
2379  * @buffer: the user buffer
2380  * @lenp: the size of the user buffer
2381  * @ppos: file position
2382  *
2383  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2384  * values from/to the user buffer, treated as an ASCII string.
2385  * The values read are assumed to be in seconds, and are converted into
2386  * jiffies.
2387  *
2388  * Returns 0 on success.
2389  */
2390 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2391                           void __user *buffer, size_t *lenp, loff_t *ppos)
2392 {
2393     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2394                             do_proc_dointvec_jiffies_conv,NULL);
2395 }
2396
2397 /**
2398  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2399  * @table: the sysctl table
2400  * @write: %TRUE if this is a write to the sysctl file
2401  * @buffer: the user buffer
2402  * @lenp: the size of the user buffer
2403  * @ppos: pointer to the file position
2404  *
2405  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2406  * values from/to the user buffer, treated as an ASCII string.
2407  * The values read are assumed to be in 1/USER_HZ seconds, and
2408  * are converted into jiffies.
2409  *
2410  * Returns 0 on success.
2411  */
2412 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2413                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2414 {
2415     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2416                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2417 }
2418
2419 /**
2420  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2421  * @table: the sysctl table
2422  * @write: %TRUE if this is a write to the sysctl file
2423  * @buffer: the user buffer
2424  * @lenp: the size of the user buffer
2425  * @ppos: file position
2426  * @ppos: the current position in the file
2427  *
2428  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2429  * values from/to the user buffer, treated as an ASCII string.
2430  * The values read are assumed to be in 1/1000 seconds, and
2431  * are converted into jiffies.
2432  *
2433  * Returns 0 on success.
2434  */
2435 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2436                              void __user *buffer, size_t *lenp, loff_t *ppos)
2437 {
2438         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2439                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2440 }
2441
2442 static int proc_do_cad_pid(struct ctl_table *table, int write,
2443                            void __user *buffer, size_t *lenp, loff_t *ppos)
2444 {
2445         struct pid *new_pid;
2446         pid_t tmp;
2447         int r;
2448
2449         tmp = pid_vnr(cad_pid);
2450
2451         r = __do_proc_dointvec(&tmp, table, write, buffer,
2452                                lenp, ppos, NULL, NULL);
2453         if (r || !write)
2454                 return r;
2455
2456         new_pid = find_get_pid(tmp);
2457         if (!new_pid)
2458                 return -ESRCH;
2459
2460         put_pid(xchg(&cad_pid, new_pid));
2461         return 0;
2462 }
2463
2464 /**
2465  * proc_do_large_bitmap - read/write from/to a large bitmap
2466  * @table: the sysctl table
2467  * @write: %TRUE if this is a write to the sysctl file
2468  * @buffer: the user buffer
2469  * @lenp: the size of the user buffer
2470  * @ppos: file position
2471  *
2472  * The bitmap is stored at table->data and the bitmap length (in bits)
2473  * in table->maxlen.
2474  *
2475  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2476  * large bitmaps may be represented in a compact manner. Writing into
2477  * the file will clear the bitmap then update it with the given input.
2478  *
2479  * Returns 0 on success.
2480  */
2481 int proc_do_large_bitmap(struct ctl_table *table, int write,
2482                          void __user *buffer, size_t *lenp, loff_t *ppos)
2483 {
2484         int err = 0;
2485         bool first = 1;
2486         size_t left = *lenp;
2487         unsigned long bitmap_len = table->maxlen;
2488         unsigned long *bitmap = (unsigned long *) table->data;
2489         unsigned long *tmp_bitmap = NULL;
2490         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2491
2492         if (!bitmap_len || !left || (*ppos && !write)) {
2493                 *lenp = 0;
2494                 return 0;
2495         }
2496
2497         if (write) {
2498                 unsigned long page = 0;
2499                 char *kbuf;
2500
2501                 if (left > PAGE_SIZE - 1)
2502                         left = PAGE_SIZE - 1;
2503
2504                 page = __get_free_page(GFP_TEMPORARY);
2505                 kbuf = (char *) page;
2506                 if (!kbuf)
2507                         return -ENOMEM;
2508                 if (copy_from_user(kbuf, buffer, left)) {
2509                         free_page(page);
2510                         return -EFAULT;
2511                 }
2512                 kbuf[left] = 0;
2513
2514                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2515                                      GFP_KERNEL);
2516                 if (!tmp_bitmap) {
2517                         free_page(page);
2518                         return -ENOMEM;
2519                 }
2520                 proc_skip_char(&kbuf, &left, '\n');
2521                 while (!err && left) {
2522                         unsigned long val_a, val_b;
2523                         bool neg;
2524
2525                         err = proc_get_long(&kbuf, &left, &val_a, &neg, tr_a,
2526                                              sizeof(tr_a), &c);
2527                         if (err)
2528                                 break;
2529                         if (val_a >= bitmap_len || neg) {
2530                                 err = -EINVAL;
2531                                 break;
2532                         }
2533
2534                         val_b = val_a;
2535                         if (left) {
2536                                 kbuf++;
2537                                 left--;
2538                         }
2539
2540                         if (c == '-') {
2541                                 err = proc_get_long(&kbuf, &left, &val_b,
2542                                                      &neg, tr_b, sizeof(tr_b),
2543                                                      &c);
2544                                 if (err)
2545                                         break;
2546                                 if (val_b >= bitmap_len || neg ||
2547                                     val_a > val_b) {
2548                                         err = -EINVAL;
2549                                         break;
2550                                 }
2551                                 if (left) {
2552                                         kbuf++;
2553                                         left--;
2554                                 }
2555                         }
2556
2557                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2558                         first = 0;
2559                         proc_skip_char(&kbuf, &left, '\n');
2560                 }
2561                 free_page(page);
2562         } else {
2563                 unsigned long bit_a, bit_b = 0;
2564
2565                 while (left) {
2566                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2567                         if (bit_a >= bitmap_len)
2568                                 break;
2569                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2570                                                    bit_a + 1) - 1;
2571
2572                         if (!first) {
2573                                 err = proc_put_char(&buffer, &left, ',');
2574                                 if (err)
2575                                         break;
2576                         }
2577                         err = proc_put_long(&buffer, &left, bit_a, false);
2578                         if (err)
2579                                 break;
2580                         if (bit_a != bit_b) {
2581                                 err = proc_put_char(&buffer, &left, '-');
2582                                 if (err)
2583                                         break;
2584                                 err = proc_put_long(&buffer, &left, bit_b, false);
2585                                 if (err)
2586                                         break;
2587                         }
2588
2589                         first = 0; bit_b++;
2590                 }
2591                 if (!err)
2592                         err = proc_put_char(&buffer, &left, '\n');
2593         }
2594
2595         if (!err) {
2596                 if (write) {
2597                         if (*ppos)
2598                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2599                         else
2600                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2601                 }
2602                 kfree(tmp_bitmap);
2603                 *lenp -= left;
2604                 *ppos += *lenp;
2605                 return 0;
2606         } else {
2607                 kfree(tmp_bitmap);
2608                 return err;
2609         }
2610 }
2611
2612 #else /* CONFIG_PROC_SYSCTL */
2613
2614 int proc_dostring(struct ctl_table *table, int write,
2615                   void __user *buffer, size_t *lenp, loff_t *ppos)
2616 {
2617         return -ENOSYS;
2618 }
2619
2620 int proc_dointvec(struct ctl_table *table, int write,
2621                   void __user *buffer, size_t *lenp, loff_t *ppos)
2622 {
2623         return -ENOSYS;
2624 }
2625
2626 int proc_dointvec_minmax(struct ctl_table *table, int write,
2627                     void __user *buffer, size_t *lenp, loff_t *ppos)
2628 {
2629         return -ENOSYS;
2630 }
2631
2632 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2633                     void __user *buffer, size_t *lenp, loff_t *ppos)
2634 {
2635         return -ENOSYS;
2636 }
2637
2638 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2639                     void __user *buffer, size_t *lenp, loff_t *ppos)
2640 {
2641         return -ENOSYS;
2642 }
2643
2644 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2645                              void __user *buffer, size_t *lenp, loff_t *ppos)
2646 {
2647         return -ENOSYS;
2648 }
2649
2650 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2651                     void __user *buffer, size_t *lenp, loff_t *ppos)
2652 {
2653         return -ENOSYS;
2654 }
2655
2656 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2657                                       void __user *buffer,
2658                                       size_t *lenp, loff_t *ppos)
2659 {
2660     return -ENOSYS;
2661 }
2662
2663
2664 #endif /* CONFIG_PROC_SYSCTL */
2665
2666 /*
2667  * No sense putting this after each symbol definition, twice,
2668  * exception granted :-)
2669  */
2670 EXPORT_SYMBOL(proc_dointvec);
2671 EXPORT_SYMBOL(proc_dointvec_jiffies);
2672 EXPORT_SYMBOL(proc_dointvec_minmax);
2673 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2674 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2675 EXPORT_SYMBOL(proc_dostring);
2676 EXPORT_SYMBOL(proc_doulongvec_minmax);
2677 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);