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/fs.h>
  34 #include <linux/init.h>
  35 #include <linux/kernel.h>
  36 #include <linux/kobject.h>
  37 #include <linux/net.h>
  38 #include <linux/sysrq.h>
  39 #include <linux/highuid.h>
  40 #include <linux/writeback.h>
  41 #include <linux/ratelimit.h>
  42 #include <linux/compaction.h>
  43 #include <linux/hugetlb.h>
  44 #include <linux/initrd.h>
  45 #include <linux/key.h>
  46 #include <linux/times.h>
  47 #include <linux/limits.h>
  48 #include <linux/dcache.h>
  49 #include <linux/dnotify.h>
  50 #include <linux/syscalls.h>
  51 #include <linux/vmstat.h>
  52 #include <linux/nfs_fs.h>
  53 #include <linux/acpi.h>
  54 #include <linux/reboot.h>
  55 #include <linux/ftrace.h>
  56 #include <linux/perf_event.h>
  57 #include <linux/kprobes.h>
  58 #include <linux/pipe_fs_i.h>
  59 #include <linux/oom.h>
  60 #include <linux/kmod.h>
  61 #include <linux/capability.h>
  62 #include <linux/binfmts.h>
  63
  64 #include <asm/uaccess.h>
  65 #include <asm/processor.h>
  66
  67 #ifdef CONFIG_X86
  68 #include <asm/nmi.h>
  69 #include <asm/stacktrace.h>
  70 #include <asm/io.h>
  71 #endif
  72 #ifdef CONFIG_SPARC
  73 #include <asm/setup.h>
  74 #endif
  75 #ifdef CONFIG_BSD_PROCESS_ACCT
  76 #include <linux/acct.h>
  77 #endif
  78 #ifdef CONFIG_RT_MUTEXES
  79 #include <linux/rtmutex.h>
  80 #endif
  81 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  82 #include <linux/lockdep.h>
  83 #endif
  84 #ifdef CONFIG_CHR_DEV_SG
  85 #include <scsi/sg.h>
  86 #endif
  87
  88 #ifdef CONFIG_LOCKUP_DETECTOR
  89 #include <linux/nmi.h>
  90 #endif
  91
  92
  93 #if defined(CONFIG_SYSCTL)
  94
  95 /* External variables not in a header file. */
  96 extern int sysctl_overcommit_memory;
  97 extern int sysctl_overcommit_ratio;
  98 #ifdef CONFIG_MMU
  99 extern int sysctl_mmap_noexec_taint;
 100 #endif
 101 extern int max_threads;
 102 extern int core_uses_pid;
 103 extern int suid_dumpable;
 104 extern char core_pattern[];
 105 extern unsigned int core_pipe_limit;
 106 extern int pid_max;
 107 extern int min_free_kbytes;
 108 extern int pid_max_min, pid_max_max;
 109 extern int sysctl_drop_caches;
 110 extern int percpu_pagelist_fraction;
 111 extern int compat_log;
 112 extern int latencytop_enabled;
 113 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 114 #ifndef CONFIG_MMU
 115 extern int sysctl_nr_trim_pages;
 116 #endif
 117 #ifdef CONFIG_BLOCK
 118 extern int blk_iopoll_enabled;
 119 #endif
 120
 121 /* Constants used for minimum and  maximum */
 122 #ifdef CONFIG_LOCKUP_DETECTOR
 123 static int sixty = 60;
 124 static int neg_one = -1;
 125 #endif
 126
 127 static int zero;
 128 static int __maybe_unused one = 1;
 129 static int __maybe_unused two = 2;
 130 static int __maybe_unused three = 3;
 131 static unsigned long one_ul = 1;
 132 static int one_hundred = 100;
 133 #ifdef CONFIG_PRINTK
 134 static int ten_thousand = 10000;
 135 #endif
 136
 137 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 138 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 139
 140 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 141 static int maxolduid = 65535;
 142 static int minolduid;
 143 static int min_percpu_pagelist_fract = 8;
 144
 145 static int ngroups_max = NGROUPS_MAX;
 146 static const int cap_last_cap = CAP_LAST_CAP;
 147
 148 #ifdef CONFIG_INOTIFY_USER
 149 #include <linux/inotify.h>
 150 #endif
 151 #ifdef CONFIG_SPARC
 152 #endif
 153
 154 #ifdef CONFIG_SPARC64
 155 extern int sysctl_tsb_ratio;
 156 #endif
 157
 158 #ifdef __hppa__
 159 extern int pwrsw_enabled;
 160 extern int unaligned_enabled;
 161 #endif
 162
 163 #ifdef CONFIG_IA64
 164 extern int no_unaligned_warning;
 165 extern int unaligned_dump_stack;
 166 #endif
 167
 168 #ifdef CONFIG_PROC_SYSCTL
 169 static int proc_do_cad_pid(struct ctl_table *table, int write,
 170                   void __user *buffer, size_t *lenp, loff_t *ppos);
 171 static int proc_taint(struct ctl_table *table, int write,
 172                                void __user *buffer, size_t *lenp, loff_t *ppos);
 173 #endif
 174
 175 #ifdef CONFIG_PRINTK
 176 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 177                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 178 #endif
 179
 180 #ifdef CONFIG_MAGIC_SYSRQ
 181 /* Note: sysrq code uses it's own private copy */
 182 static int __sysrq_enabled = SYSRQ_DEFAULT_ENABLE;
 183
 184 static int sysrq_sysctl_handler(ctl_table *table, int write,
 185                                 void __user *buffer, size_t *lenp,
 186                                 loff_t *ppos)
 187 {
 188         int error;
 189
 190         error = proc_dointvec(table, write, buffer, lenp, ppos);
 191         if (error)
 192                 return error;
 193
 194         if (write)
 195                 sysrq_toggle_support(__sysrq_enabled);
 196
 197         return 0;
 198 }
 199
 200 #endif
 201
 202 static struct ctl_table kern_table[];
 203 static struct ctl_table vm_table[];
 204 static struct ctl_table fs_table[];
 205 static struct ctl_table debug_table[];
 206 static struct ctl_table dev_table[];
 207 extern struct ctl_table random_table[];
 208 #ifdef CONFIG_EPOLL
 209 extern struct ctl_table epoll_table[];
 210 #endif
 211
 212 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 213 int sysctl_legacy_va_layout;
 214 #endif
 215
 216 /* The default sysctl tables: */
 217
 218 static struct ctl_table sysctl_base_table[] = {
 219         {
 220                 .procname       = "kernel",
 221                 .mode           = 0555,
 222                 .child          = kern_table,
 223         },
 224         {
 225                 .procname       = "vm",
 226                 .mode           = 0555,
 227                 .child          = vm_table,
 228         },
 229         {
 230                 .procname       = "fs",
 231                 .mode           = 0555,
 232                 .child          = fs_table,
 233         },
 234         {
 235                 .procname       = "debug",
 236                 .mode           = 0555,
 237                 .child          = debug_table,
 238         },
 239         {
 240                 .procname       = "dev",
 241                 .mode           = 0555,
 242                 .child          = dev_table,
 243         },
 244         { }
 245 };
 246
 247 #ifdef CONFIG_SCHED_DEBUG
 248 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 249 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 250 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 251 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 252 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 253 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 254 #endif
 255
 256 #ifdef CONFIG_COMPACTION
 257 static int min_extfrag_threshold;
 258 static int max_extfrag_threshold = 1000;
 259 #endif
 260
 261 static struct ctl_table kern_table[] = {
 262         {
 263                 .procname       = "sched_child_runs_first",
 264                 .data           = &sysctl_sched_child_runs_first,
 265                 .maxlen         = sizeof(unsigned int),
 266                 .mode           = 0644,
 267                 .proc_handler   = proc_dointvec,
 268         },
 269 #ifdef CONFIG_SCHED_DEBUG
 270         {
 271                 .procname       = "sched_min_granularity_ns",
 272                 .data           = &sysctl_sched_min_granularity,
 273                 .maxlen         = sizeof(unsigned int),
 274                 .mode           = 0644,
 275                 .proc_handler   = sched_proc_update_handler,
 276                 .extra1         = &min_sched_granularity_ns,
 277                 .extra2         = &max_sched_granularity_ns,
 278         },
 279         {
 280                 .procname       = "sched_latency_ns",
 281                 .data           = &sysctl_sched_latency,
 282                 .maxlen         = sizeof(unsigned int),
 283                 .mode           = 0644,
 284                 .proc_handler   = sched_proc_update_handler,
 285                 .extra1         = &min_sched_granularity_ns,
 286                 .extra2         = &max_sched_granularity_ns,
 287         },
 288         {
 289                 .procname       = "sched_wakeup_granularity_ns",
 290                 .data           = &sysctl_sched_wakeup_granularity,
 291                 .maxlen         = sizeof(unsigned int),
 292                 .mode           = 0644,
 293                 .proc_handler   = sched_proc_update_handler,
 294                 .extra1         = &min_wakeup_granularity_ns,
 295                 .extra2         = &max_wakeup_granularity_ns,
 296         },
 297         {
 298                 .procname       = "sched_tunable_scaling",
 299                 .data           = &sysctl_sched_tunable_scaling,
 300                 .maxlen         = sizeof(enum sched_tunable_scaling),
 301                 .mode           = 0644,
 302                 .proc_handler   = sched_proc_update_handler,
 303                 .extra1         = &min_sched_tunable_scaling,
 304                 .extra2         = &max_sched_tunable_scaling,
 305         },
 306         {
 307                 .procname       = "sched_migration_cost",
 308                 .data           = &sysctl_sched_migration_cost,
 309                 .maxlen         = sizeof(unsigned int),
 310                 .mode           = 0644,
 311                 .proc_handler   = proc_dointvec,
 312         },
 313         {
 314                 .procname       = "sched_nr_migrate",
 315                 .data           = &sysctl_sched_nr_migrate,
 316                 .maxlen         = sizeof(unsigned int),
 317                 .mode           = 0644,
 318                 .proc_handler   = proc_dointvec,
 319         },
 320         {
 321                 .procname       = "sched_time_avg",
 322                 .data           = &sysctl_sched_time_avg,
 323                 .maxlen         = sizeof(unsigned int),
 324                 .mode           = 0644,
 325                 .proc_handler   = proc_dointvec,
 326         },
 327         {
 328                 .procname       = "sched_shares_window",
 329                 .data           = &sysctl_sched_shares_window,
 330                 .maxlen         = sizeof(unsigned int),
 331                 .mode           = 0644,
 332                 .proc_handler   = proc_dointvec,
 333         },
 334         {
 335                 .procname       = "timer_migration",
 336                 .data           = &sysctl_timer_migration,
 337                 .maxlen         = sizeof(unsigned int),
 338                 .mode           = 0644,
 339                 .proc_handler   = proc_dointvec_minmax,
 340                 .extra1         = &zero,
 341                 .extra2         = &one,
 342         },
 343 #endif
 344         {
 345                 .procname       = "sched_rt_period_us",
 346                 .data           = &sysctl_sched_rt_period,
 347                 .maxlen         = sizeof(unsigned int),
 348                 .mode           = 0644,
 349                 .proc_handler   = sched_rt_handler,
 350         },
 351         {
 352                 .procname       = "sched_rt_runtime_us",
 353                 .data           = &sysctl_sched_rt_runtime,
 354                 .maxlen         = sizeof(int),
 355                 .mode           = 0644,
 356                 .proc_handler   = sched_rt_handler,
 357         },
 358 #ifdef CONFIG_SCHED_AUTOGROUP
 359         {
 360                 .procname       = "sched_autogroup_enabled",
 361                 .data           = &sysctl_sched_autogroup_enabled,
 362                 .maxlen         = sizeof(unsigned int),
 363                 .mode           = 0644,
 364                 .proc_handler   = proc_dointvec_minmax,
 365                 .extra1         = &zero,
 366                 .extra2         = &one,
 367         },
 368 #endif
 369 #ifdef CONFIG_CFS_BANDWIDTH
 370         {
 371                 .procname       = "sched_cfs_bandwidth_slice_us",
 372                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 373                 .maxlen         = sizeof(unsigned int),
 374                 .mode           = 0644,
 375                 .proc_handler   = proc_dointvec_minmax,
 376                 .extra1         = &one,
 377         },
 378 #endif
 379 #ifdef CONFIG_PROVE_LOCKING
 380         {
 381                 .procname       = "prove_locking",
 382                 .data           = &prove_locking,
 383                 .maxlen         = sizeof(int),
 384                 .mode           = 0644,
 385                 .proc_handler   = proc_dointvec,
 386         },
 387 #endif
 388 #ifdef CONFIG_LOCK_STAT
 389         {
 390                 .procname       = "lock_stat",
 391                 .data           = &lock_stat,
 392                 .maxlen         = sizeof(int),
 393                 .mode           = 0644,
 394                 .proc_handler   = proc_dointvec,
 395         },
 396 #endif
 397         {
 398                 .procname       = "panic",
 399                 .data           = &panic_timeout,
 400                 .maxlen         = sizeof(int),
 401                 .mode           = 0644,
 402                 .proc_handler   = proc_dointvec,
 403         },
 404         {
 405                 .procname       = "core_uses_pid",
 406                 .data           = &core_uses_pid,
 407                 .maxlen         = sizeof(int),
 408                 .mode           = 0644,
 409                 .proc_handler   = proc_dointvec,
 410         },
 411         {
 412                 .procname       = "core_pattern",
 413                 .data           = core_pattern,
 414                 .maxlen         = CORENAME_MAX_SIZE,
 415                 .mode           = 0644,
 416                 .proc_handler   = proc_dostring,
 417         },
 418         {
 419                 .procname       = "core_pipe_limit",
 420                 .data           = &core_pipe_limit,
 421                 .maxlen         = sizeof(unsigned int),
 422                 .mode           = 0644,
 423                 .proc_handler   = proc_dointvec,
 424         },
 425 #ifdef CONFIG_PROC_SYSCTL
 426         {
 427                 .procname       = "tainted",
 428                 .maxlen         = sizeof(long),
 429                 .mode           = 0644,
 430                 .proc_handler   = proc_taint,
 431         },
 432 #endif
 433 #ifdef CONFIG_LATENCYTOP
 434         {
 435                 .procname       = "latencytop",
 436                 .data           = &latencytop_enabled,
 437                 .maxlen         = sizeof(int),
 438                 .mode           = 0644,
 439                 .proc_handler   = proc_dointvec,
 440         },
 441 #endif
 442 #ifdef CONFIG_BLK_DEV_INITRD
 443         {
 444                 .procname       = "real-root-dev",
 445                 .data           = &real_root_dev,
 446                 .maxlen         = sizeof(int),
 447                 .mode           = 0644,
 448                 .proc_handler   = proc_dointvec,
 449         },
 450 #endif
 451         {
 452                 .procname       = "print-fatal-signals",
 453                 .data           = &print_fatal_signals,
 454                 .maxlen         = sizeof(int),
 455                 .mode           = 0644,
 456                 .proc_handler   = proc_dointvec,
 457         },
 458 #ifdef CONFIG_SPARC
 459         {
 460                 .procname       = "reboot-cmd",
 461                 .data           = reboot_command,
 462                 .maxlen         = 256,
 463                 .mode           = 0644,
 464                 .proc_handler   = proc_dostring,
 465         },
 466         {
 467                 .procname       = "stop-a",
 468                 .data           = &stop_a_enabled,
 469                 .maxlen         = sizeof (int),
 470                 .mode           = 0644,
 471                 .proc_handler   = proc_dointvec,
 472         },
 473         {
 474                 .procname       = "scons-poweroff",
 475                 .data           = &scons_pwroff,
 476                 .maxlen         = sizeof (int),
 477                 .mode           = 0644,
 478                 .proc_handler   = proc_dointvec,
 479         },
 480 #endif
 481 #ifdef CONFIG_SPARC64
 482         {
 483                 .procname       = "tsb-ratio",
 484                 .data           = &sysctl_tsb_ratio,
 485                 .maxlen         = sizeof (int),
 486                 .mode           = 0644,
 487                 .proc_handler   = proc_dointvec,
 488         },
 489 #endif
 490 #ifdef __hppa__
 491         {
 492                 .procname       = "soft-power",
 493                 .data           = &pwrsw_enabled,
 494                 .maxlen         = sizeof (int),
 495                 .mode           = 0644,
 496                 .proc_handler   = proc_dointvec,
 497         },
 498         {
 499                 .procname       = "unaligned-trap",
 500                 .data           = &unaligned_enabled,
 501                 .maxlen         = sizeof (int),
 502                 .mode           = 0644,
 503                 .proc_handler   = proc_dointvec,
 504         },
 505 #endif
 506         {
 507                 .procname       = "ctrl-alt-del",
 508                 .data           = &C_A_D,
 509                 .maxlen         = sizeof(int),
 510                 .mode           = 0644,
 511                 .proc_handler   = proc_dointvec,
 512         },
 513 #ifdef CONFIG_FUNCTION_TRACER
 514         {
 515                 .procname       = "ftrace_enabled",
 516                 .data           = &ftrace_enabled,
 517                 .maxlen         = sizeof(int),
 518                 .mode           = 0644,
 519                 .proc_handler   = ftrace_enable_sysctl,
 520         },
 521 #endif
 522 #ifdef CONFIG_STACK_TRACER
 523         {
 524                 .procname       = "stack_tracer_enabled",
 525                 .data           = &stack_tracer_enabled,
 526                 .maxlen         = sizeof(int),
 527                 .mode           = 0644,
 528                 .proc_handler   = stack_trace_sysctl,
 529         },
 530 #endif
 531 #ifdef CONFIG_TRACING
 532         {
 533                 .procname       = "ftrace_dump_on_oops",
 534                 .data           = &ftrace_dump_on_oops,
 535                 .maxlen         = sizeof(int),
 536                 .mode           = 0644,
 537                 .proc_handler   = proc_dointvec,
 538         },
 539 #endif
 540 #ifdef CONFIG_MODULES
 541         {
 542                 .procname       = "modprobe",
 543                 .data           = &modprobe_path,
 544                 .maxlen         = KMOD_PATH_LEN,
 545                 .mode           = 0644,
 546                 .proc_handler   = proc_dostring,
 547         },
 548         {
 549                 .procname       = "modules_disabled",
 550                 .data           = &modules_disabled,
 551                 .maxlen         = sizeof(int),
 552                 .mode           = 0644,
 553                 /* only handle a transition from default "0" to "1" */
 554                 .proc_handler   = proc_dointvec_minmax,
 555                 .extra1         = &one,
 556                 .extra2         = &one,
 557         },
 558 #endif
 559 #ifdef CONFIG_HOTPLUG
 560         {
 561                 .procname       = "hotplug",
 562                 .data           = &uevent_helper,
 563                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 564                 .mode           = 0644,
 565                 .proc_handler   = proc_dostring,
 566         },
 567 #endif
 568 #ifdef CONFIG_CHR_DEV_SG
 569         {
 570                 .procname       = "sg-big-buff",
 571                 .data           = &sg_big_buff,
 572                 .maxlen         = sizeof (int),
 573                 .mode           = 0444,
 574                 .proc_handler   = proc_dointvec,
 575         },
 576 #endif
 577 #ifdef CONFIG_BSD_PROCESS_ACCT
 578         {
 579                 .procname       = "acct",
 580                 .data           = &acct_parm,
 581                 .maxlen         = 3*sizeof(int),
 582                 .mode           = 0644,
 583                 .proc_handler   = proc_dointvec,
 584         },
 585 #endif
 586 #ifdef CONFIG_MAGIC_SYSRQ
 587         {
 588                 .procname       = "sysrq",
 589                 .data           = &__sysrq_enabled,
 590                 .maxlen         = sizeof (int),
 591                 .mode           = 0644,
 592                 .proc_handler   = sysrq_sysctl_handler,
 593         },
 594 #endif
 595 #ifdef CONFIG_PROC_SYSCTL
 596         {
 597                 .procname       = "cad_pid",
 598                 .data           = NULL,
 599                 .maxlen         = sizeof (int),
 600                 .mode           = 0600,
 601                 .proc_handler   = proc_do_cad_pid,
 602         },
 603 #endif
 604         {
 605                 .procname       = "threads-max",
 606                 .data           = &max_threads,
 607                 .maxlen         = sizeof(int),
 608                 .mode           = 0644,
 609                 .proc_handler   = proc_dointvec,
 610         },
 611         {
 612                 .procname       = "random",
 613                 .mode           = 0555,
 614                 .child          = random_table,
 615         },
 616         {
 617                 .procname       = "usermodehelper",
 618                 .mode           = 0555,
 619                 .child          = usermodehelper_table,
 620         },
 621         {
 622                 .procname       = "overflowuid",
 623                 .data           = &overflowuid,
 624                 .maxlen         = sizeof(int),
 625                 .mode           = 0644,
 626                 .proc_handler   = proc_dointvec_minmax,
 627                 .extra1         = &minolduid,
 628                 .extra2         = &maxolduid,
 629         },
 630         {
 631                 .procname       = "overflowgid",
 632                 .data           = &overflowgid,
 633                 .maxlen         = sizeof(int),
 634                 .mode           = 0644,
 635                 .proc_handler   = proc_dointvec_minmax,
 636                 .extra1         = &minolduid,
 637                 .extra2         = &maxolduid,
 638         },
 639 #ifdef CONFIG_S390
 640 #ifdef CONFIG_MATHEMU
 641         {
 642                 .procname       = "ieee_emulation_warnings",
 643                 .data           = &sysctl_ieee_emulation_warnings,
 644                 .maxlen         = sizeof(int),
 645                 .mode           = 0644,
 646                 .proc_handler   = proc_dointvec,
 647         },
 648 #endif
 649         {
 650                 .procname       = "userprocess_debug",
 651                 .data           = &show_unhandled_signals,
 652                 .maxlen         = sizeof(int),
 653                 .mode           = 0644,
 654                 .proc_handler   = proc_dointvec,
 655         },
 656 #endif
 657         {
 658                 .procname       = "pid_max",
 659                 .data           = &pid_max,
 660                 .maxlen         = sizeof (int),
 661                 .mode           = 0644,
 662                 .proc_handler   = proc_dointvec_minmax,
 663                 .extra1         = &pid_max_min,
 664                 .extra2         = &pid_max_max,
 665         },
 666         {
 667                 .procname       = "panic_on_oops",
 668                 .data           = &panic_on_oops,
 669                 .maxlen         = sizeof(int),
 670                 .mode           = 0644,
 671                 .proc_handler   = proc_dointvec,
 672         },
 673 #if defined CONFIG_PRINTK
 674         {
 675                 .procname       = "printk",
 676                 .data           = &console_loglevel,
 677                 .maxlen         = 4*sizeof(int),
 678                 .mode           = 0644,
 679                 .proc_handler   = proc_dointvec,
 680         },
 681         {
 682                 .procname       = "printk_ratelimit",
 683                 .data           = &printk_ratelimit_state.interval,
 684                 .maxlen         = sizeof(int),
 685                 .mode           = 0644,
 686                 .proc_handler   = proc_dointvec_jiffies,
 687         },
 688         {
 689                 .procname       = "printk_ratelimit_burst",
 690                 .data           = &printk_ratelimit_state.burst,
 691                 .maxlen         = sizeof(int),
 692                 .mode           = 0644,
 693                 .proc_handler   = proc_dointvec,
 694         },
 695         {
 696                 .procname       = "printk_delay",
 697                 .data           = &printk_delay_msec,
 698                 .maxlen         = sizeof(int),
 699                 .mode           = 0644,
 700                 .proc_handler   = proc_dointvec_minmax,
 701                 .extra1         = &zero,
 702                 .extra2         = &ten_thousand,
 703         },
 704         {
 705                 .procname       = "dmesg_restrict",
 706                 .data           = &dmesg_restrict,
 707                 .maxlen         = sizeof(int),
 708                 .mode           = 0644,
 709                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 710                 .extra1         = &zero,
 711                 .extra2         = &one,
 712         },
 713         {
 714                 .procname       = "kptr_restrict",
 715                 .data           = &kptr_restrict,
 716                 .maxlen         = sizeof(int),
 717                 .mode           = 0644,
 718                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 719                 .extra1         = &zero,
 720                 .extra2         = &two,
 721         },
 722 #endif
 723         {
 724                 .procname       = "ngroups_max",
 725                 .data           = &ngroups_max,
 726                 .maxlen         = sizeof (int),
 727                 .mode           = 0444,
 728                 .proc_handler   = proc_dointvec,
 729         },
 730         {
 731                 .procname       = "cap_last_cap",
 732                 .data           = (void *)&cap_last_cap,
 733                 .maxlen         = sizeof(int),
 734                 .mode           = 0444,
 735                 .proc_handler   = proc_dointvec,
 736         },
 737 #if defined(CONFIG_LOCKUP_DETECTOR)
 738         {
 739                 .procname       = "watchdog",
 740                 .data           = &watchdog_enabled,
 741                 .maxlen         = sizeof (int),
 742                 .mode           = 0644,
 743                 .proc_handler   = proc_dowatchdog,
 744                 .extra1         = &zero,
 745                 .extra2         = &one,
 746         },
 747         {
 748                 .procname       = "watchdog_thresh",
 749                 .data           = &watchdog_thresh,
 750                 .maxlen         = sizeof(int),
 751                 .mode           = 0644,
 752                 .proc_handler   = proc_dowatchdog,
 753                 .extra1         = &neg_one,
 754                 .extra2         = &sixty,
 755         },
 756         {
 757                 .procname       = "softlockup_panic",
 758                 .data           = &softlockup_panic,
 759                 .maxlen         = sizeof(int),
 760                 .mode           = 0644,
 761                 .proc_handler   = proc_dointvec_minmax,
 762                 .extra1         = &zero,
 763                 .extra2         = &one,
 764         },
 765         {
 766                 .procname       = "nmi_watchdog",
 767                 .data           = &watchdog_enabled,
 768                 .maxlen         = sizeof (int),
 769                 .mode           = 0644,
 770                 .proc_handler   = proc_dowatchdog,
 771                 .extra1         = &zero,
 772                 .extra2         = &one,
 773         },
 774 #endif
 775 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 776         {
 777                 .procname       = "unknown_nmi_panic",
 778                 .data           = &unknown_nmi_panic,
 779                 .maxlen         = sizeof (int),
 780                 .mode           = 0644,
 781                 .proc_handler   = proc_dointvec,
 782         },
 783 #endif
 784 #if defined(CONFIG_X86)
 785         {
 786                 .procname       = "panic_on_unrecovered_nmi",
 787                 .data           = &panic_on_unrecovered_nmi,
 788                 .maxlen         = sizeof(int),
 789                 .mode           = 0644,
 790                 .proc_handler   = proc_dointvec,
 791         },
 792         {
 793                 .procname       = "panic_on_io_nmi",
 794                 .data           = &panic_on_io_nmi,
 795                 .maxlen         = sizeof(int),
 796                 .mode           = 0644,
 797                 .proc_handler   = proc_dointvec,
 798         },
 799 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 800         {
 801                 .procname       = "panic_on_stackoverflow",
 802                 .data           = &sysctl_panic_on_stackoverflow,
 803                 .maxlen         = sizeof(int),
 804                 .mode           = 0644,
 805                 .proc_handler   = proc_dointvec,
 806         },
 807 #endif
 808         {
 809                 .procname       = "bootloader_type",
 810                 .data           = &bootloader_type,
 811                 .maxlen         = sizeof (int),
 812                 .mode           = 0444,
 813                 .proc_handler   = proc_dointvec,
 814         },
 815         {
 816                 .procname       = "bootloader_version",
 817                 .data           = &bootloader_version,
 818                 .maxlen         = sizeof (int),
 819                 .mode           = 0444,
 820                 .proc_handler   = proc_dointvec,
 821         },
 822         {
 823                 .procname       = "kstack_depth_to_print",
 824                 .data           = &kstack_depth_to_print,
 825                 .maxlen         = sizeof(int),
 826                 .mode           = 0644,
 827                 .proc_handler   = proc_dointvec,
 828         },
 829         {
 830                 .procname       = "io_delay_type",
 831                 .data           = &io_delay_type,
 832                 .maxlen         = sizeof(int),
 833                 .mode           = 0644,
 834                 .proc_handler   = proc_dointvec,
 835         },
 836 #endif
 837 #if defined(CONFIG_MMU)
 838         {
 839                 .procname       = "randomize_va_space",
 840                 .data           = &randomize_va_space,
 841                 .maxlen         = sizeof(int),
 842                 .mode           = 0644,
 843                 .proc_handler   = proc_dointvec,
 844         },
 845 #endif
 846 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 847         {
 848                 .procname       = "spin_retry",
 849                 .data           = &spin_retry,
 850                 .maxlen         = sizeof (int),
 851                 .mode           = 0644,
 852                 .proc_handler   = proc_dointvec,
 853         },
 854 #endif
 855 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 856         {
 857                 .procname       = "acpi_video_flags",
 858                 .data           = &acpi_realmode_flags,
 859                 .maxlen         = sizeof (unsigned long),
 860                 .mode           = 0644,
 861                 .proc_handler   = proc_doulongvec_minmax,
 862         },
 863 #endif
 864 #ifdef CONFIG_IA64
 865         {
 866                 .procname       = "ignore-unaligned-usertrap",
 867                 .data           = &no_unaligned_warning,
 868                 .maxlen         = sizeof (int),
 869                 .mode           = 0644,
 870                 .proc_handler   = proc_dointvec,
 871         },
 872         {
 873                 .procname       = "unaligned-dump-stack",
 874                 .data           = &unaligned_dump_stack,
 875                 .maxlen         = sizeof (int),
 876                 .mode           = 0644,
 877                 .proc_handler   = proc_dointvec,
 878         },
 879 #endif
 880 #ifdef CONFIG_DETECT_HUNG_TASK
 881         {
 882                 .procname       = "hung_task_panic",
 883                 .data           = &sysctl_hung_task_panic,
 884                 .maxlen         = sizeof(int),
 885                 .mode           = 0644,
 886                 .proc_handler   = proc_dointvec_minmax,
 887                 .extra1         = &zero,
 888                 .extra2         = &one,
 889         },
 890         {
 891                 .procname       = "hung_task_check_count",
 892                 .data           = &sysctl_hung_task_check_count,
 893                 .maxlen         = sizeof(unsigned long),
 894                 .mode           = 0644,
 895                 .proc_handler   = proc_doulongvec_minmax,
 896         },
 897         {
 898                 .procname       = "hung_task_timeout_secs",
 899                 .data           = &sysctl_hung_task_timeout_secs,
 900                 .maxlen         = sizeof(unsigned long),
 901                 .mode           = 0644,
 902                 .proc_handler   = proc_dohung_task_timeout_secs,
 903         },
 904         {
 905                 .procname       = "hung_task_warnings",
 906                 .data           = &sysctl_hung_task_warnings,
 907                 .maxlen         = sizeof(unsigned long),
 908                 .mode           = 0644,
 909                 .proc_handler   = proc_doulongvec_minmax,
 910         },
 911 #endif
 912 #ifdef CONFIG_COMPAT
 913         {
 914                 .procname       = "compat-log",
 915                 .data           = &compat_log,
 916                 .maxlen         = sizeof (int),
 917                 .mode           = 0644,
 918                 .proc_handler   = proc_dointvec,
 919         },
 920 #endif
 921 #ifdef CONFIG_RT_MUTEXES
 922         {
 923                 .procname       = "max_lock_depth",
 924                 .data           = &max_lock_depth,
 925                 .maxlen         = sizeof(int),
 926                 .mode           = 0644,
 927                 .proc_handler   = proc_dointvec,
 928         },
 929 #endif
 930         {
 931                 .procname       = "poweroff_cmd",
 932                 .data           = &poweroff_cmd,
 933                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 934                 .mode           = 0644,
 935                 .proc_handler   = proc_dostring,
 936         },
 937 #ifdef CONFIG_KEYS
 938         {
 939                 .procname       = "keys",
 940                 .mode           = 0555,
 941                 .child          = key_sysctls,
 942         },
 943 #endif
 944 #ifdef CONFIG_RCU_TORTURE_TEST
 945         {
 946                 .procname       = "rcutorture_runnable",
 947                 .data           = &rcutorture_runnable,
 948                 .maxlen         = sizeof(int),
 949                 .mode           = 0644,
 950                 .proc_handler   = proc_dointvec,
 951         },
 952 #endif
 953 #ifdef CONFIG_PERF_EVENTS
 954         /*
 955          * User-space scripts rely on the existence of this file
 956          * as a feature check for perf_events being enabled.
 957          *
 958          * So it's an ABI, do not remove!
 959          */
 960         {
 961                 .procname       = "perf_event_paranoid",
 962                 .data           = &sysctl_perf_event_paranoid,
 963                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
 964                 .mode           = 0644,
 965                 .proc_handler   = proc_dointvec,
 966         },
 967         {
 968                 .procname       = "perf_event_mlock_kb",
 969                 .data           = &sysctl_perf_event_mlock,
 970                 .maxlen         = sizeof(sysctl_perf_event_mlock),
 971                 .mode           = 0644,
 972                 .proc_handler   = proc_dointvec,
 973         },
 974         {
 975                 .procname       = "perf_event_max_sample_rate",
 976                 .data           = &sysctl_perf_event_sample_rate,
 977                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
 978                 .mode           = 0644,
 979                 .proc_handler   = perf_proc_update_handler,
 980         },
 981 #endif
 982 #ifdef CONFIG_KMEMCHECK
 983         {
 984                 .procname       = "kmemcheck",
 985                 .data           = &kmemcheck_enabled,
 986                 .maxlen         = sizeof(int),
 987                 .mode           = 0644,
 988                 .proc_handler   = proc_dointvec,
 989         },
 990 #endif
 991 #ifdef CONFIG_BLOCK
 992         {
 993                 .procname       = "blk_iopoll",
 994                 .data           = &blk_iopoll_enabled,
 995                 .maxlen         = sizeof(int),
 996                 .mode           = 0644,
 997                 .proc_handler   = proc_dointvec,
 998         },
 999 #endif
1000         { }
1001 };
1002
1003 static struct ctl_table vm_table[] = {
1004         {
1005                 .procname       = "overcommit_memory",
1006                 .data           = &sysctl_overcommit_memory,
1007                 .maxlen         = sizeof(sysctl_overcommit_memory),
1008                 .mode           = 0644,
1009                 .proc_handler   = proc_dointvec_minmax,
1010                 .extra1         = &zero,
1011                 .extra2         = &two,
1012         },
1013         {
1014                 .procname       = "panic_on_oom",
1015                 .data           = &sysctl_panic_on_oom,
1016                 .maxlen         = sizeof(sysctl_panic_on_oom),
1017                 .mode           = 0644,
1018                 .proc_handler   = proc_dointvec_minmax,
1019                 .extra1         = &zero,
1020                 .extra2         = &two,
1021         },
1022         {
1023                 .procname       = "oom_kill_allocating_task",
1024                 .data           = &sysctl_oom_kill_allocating_task,
1025                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1026                 .mode           = 0644,
1027                 .proc_handler   = proc_dointvec,
1028         },
1029         {
1030                 .procname       = "oom_dump_tasks",
1031                 .data           = &sysctl_oom_dump_tasks,
1032                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1033                 .mode           = 0644,
1034                 .proc_handler   = proc_dointvec,
1035         },
1036         {
1037                 .procname       = "overcommit_ratio",
1038                 .data           = &sysctl_overcommit_ratio,
1039                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1040                 .mode           = 0644,
1041                 .proc_handler   = proc_dointvec,
1042         },
1043         {
1044                 .procname       = "page-cluster",
1045                 .data           = &page_cluster,
1046                 .maxlen         = sizeof(int),
1047                 .mode           = 0644,
1048                 .proc_handler   = proc_dointvec_minmax,
1049                 .extra1         = &zero,
1050         },
1051         {
1052                 .procname       = "dirty_background_ratio",
1053                 .data           = &dirty_background_ratio,
1054                 .maxlen         = sizeof(dirty_background_ratio),
1055                 .mode           = 0644,
1056                 .proc_handler   = dirty_background_ratio_handler,
1057                 .extra1         = &zero,
1058                 .extra2         = &one_hundred,
1059         },
1060         {
1061                 .procname       = "dirty_background_bytes",
1062                 .data           = &dirty_background_bytes,
1063                 .maxlen         = sizeof(dirty_background_bytes),
1064                 .mode           = 0644,
1065                 .proc_handler   = dirty_background_bytes_handler,
1066                 .extra1         = &one_ul,
1067         },
1068         {
1069                 .procname       = "dirty_ratio",
1070                 .data           = &vm_dirty_ratio,
1071                 .maxlen         = sizeof(vm_dirty_ratio),
1072                 .mode           = 0644,
1073                 .proc_handler   = dirty_ratio_handler,
1074                 .extra1         = &zero,
1075                 .extra2         = &one_hundred,
1076         },
1077         {
1078                 .procname       = "dirty_bytes",
1079                 .data           = &vm_dirty_bytes,
1080                 .maxlen         = sizeof(vm_dirty_bytes),
1081                 .mode           = 0644,
1082                 .proc_handler   = dirty_bytes_handler,
1083                 .extra1         = &dirty_bytes_min,
1084         },
1085         {
1086                 .procname       = "dirty_writeback_centisecs",
1087                 .data           = &dirty_writeback_interval,
1088                 .maxlen         = sizeof(dirty_writeback_interval),
1089                 .mode           = 0644,
1090                 .proc_handler   = dirty_writeback_centisecs_handler,
1091         },
1092         {
1093                 .procname       = "dirty_expire_centisecs",
1094                 .data           = &dirty_expire_interval,
1095                 .maxlen         = sizeof(dirty_expire_interval),
1096                 .mode           = 0644,
1097                 .proc_handler   = proc_dointvec_minmax,
1098                 .extra1         = &zero,
1099         },
1100         {
1101                 .procname       = "nr_pdflush_threads",
1102                 .data           = &nr_pdflush_threads,
1103                 .maxlen         = sizeof nr_pdflush_threads,
1104                 .mode           = 0444 /* read-only*/,
1105                 .proc_handler   = proc_dointvec,
1106         },
1107         {
1108                 .procname       = "swappiness",
1109                 .data           = &vm_swappiness,
1110                 .maxlen         = sizeof(vm_swappiness),
1111                 .mode           = 0644,
1112                 .proc_handler   = proc_dointvec_minmax,
1113                 .extra1         = &zero,
1114                 .extra2         = &one_hundred,
1115         },
1116 #ifdef CONFIG_HUGETLB_PAGE
1117         {
1118                 .procname       = "nr_hugepages",
1119                 .data           = NULL,
1120                 .maxlen         = sizeof(unsigned long),
1121                 .mode           = 0644,
1122                 .proc_handler   = hugetlb_sysctl_handler,
1123                 .extra1         = (void *)&hugetlb_zero,
1124                 .extra2         = (void *)&hugetlb_infinity,
1125         },
1126 #ifdef CONFIG_NUMA
1127         {
1128                 .procname       = "nr_hugepages_mempolicy",
1129                 .data           = NULL,
1130                 .maxlen         = sizeof(unsigned long),
1131                 .mode           = 0644,
1132                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1133                 .extra1         = (void *)&hugetlb_zero,
1134                 .extra2         = (void *)&hugetlb_infinity,
1135         },
1136 #endif
1137          {
1138                 .procname       = "hugetlb_shm_group",
1139                 .data           = &sysctl_hugetlb_shm_group,
1140                 .maxlen         = sizeof(gid_t),
1141                 .mode           = 0644,
1142                 .proc_handler   = proc_dointvec,
1143          },
1144          {
1145                 .procname       = "hugepages_treat_as_movable",
1146                 .data           = &hugepages_treat_as_movable,
1147                 .maxlen         = sizeof(int),
1148                 .mode           = 0644,
1149                 .proc_handler   = hugetlb_treat_movable_handler,
1150         },
1151         {
1152                 .procname       = "nr_overcommit_hugepages",
1153                 .data           = NULL,
1154                 .maxlen         = sizeof(unsigned long),
1155                 .mode           = 0644,
1156                 .proc_handler   = hugetlb_overcommit_handler,
1157                 .extra1         = (void *)&hugetlb_zero,
1158                 .extra2         = (void *)&hugetlb_infinity,
1159         },
1160 #endif
1161         {
1162                 .procname       = "lowmem_reserve_ratio",
1163                 .data           = &sysctl_lowmem_reserve_ratio,
1164                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1165                 .mode           = 0644,
1166                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1167         },
1168         {
1169                 .procname       = "drop_caches",
1170                 .data           = &sysctl_drop_caches,
1171                 .maxlen         = sizeof(int),
1172                 .mode           = 0644,
1173                 .proc_handler   = drop_caches_sysctl_handler,
1174                 .extra1         = &one,
1175                 .extra2         = &three,
1176         },
1177 #ifdef CONFIG_COMPACTION
1178         {
1179                 .procname       = "compact_memory",
1180                 .data           = &sysctl_compact_memory,
1181                 .maxlen         = sizeof(int),
1182                 .mode           = 0200,
1183                 .proc_handler   = sysctl_compaction_handler,
1184         },
1185         {
1186                 .procname       = "extfrag_threshold",
1187                 .data           = &sysctl_extfrag_threshold,
1188                 .maxlen         = sizeof(int),
1189                 .mode           = 0644,
1190                 .proc_handler   = sysctl_extfrag_handler,
1191                 .extra1         = &min_extfrag_threshold,
1192                 .extra2         = &max_extfrag_threshold,
1193         },
1194
1195 #endif /* CONFIG_COMPACTION */
1196         {
1197                 .procname       = "min_free_kbytes",
1198                 .data           = &min_free_kbytes,
1199                 .maxlen         = sizeof(min_free_kbytes),
1200                 .mode           = 0644,
1201                 .proc_handler   = min_free_kbytes_sysctl_handler,
1202                 .extra1         = &zero,
1203         },
1204         {
1205                 .procname       = "percpu_pagelist_fraction",
1206                 .data           = &percpu_pagelist_fraction,
1207                 .maxlen         = sizeof(percpu_pagelist_fraction),
1208                 .mode           = 0644,
1209                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1210                 .extra1         = &min_percpu_pagelist_fract,
1211         },
1212 #ifdef CONFIG_MMU
1213         {
1214                 .procname       = "max_map_count",
1215                 .data           = &sysctl_max_map_count,
1216                 .maxlen         = sizeof(sysctl_max_map_count),
1217                 .mode           = 0644,
1218                 .proc_handler   = proc_dointvec_minmax,
1219                 .extra1         = &zero,
1220         },
1221 #else
1222         {
1223                 .procname       = "nr_trim_pages",
1224                 .data           = &sysctl_nr_trim_pages,
1225                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1226                 .mode           = 0644,
1227                 .proc_handler   = proc_dointvec_minmax,
1228                 .extra1         = &zero,
1229         },
1230 #endif
1231         {
1232                 .procname       = "laptop_mode",
1233                 .data           = &laptop_mode,
1234                 .maxlen         = sizeof(laptop_mode),
1235                 .mode           = 0644,
1236                 .proc_handler   = proc_dointvec_jiffies,
1237         },
1238         {
1239                 .procname       = "block_dump",
1240                 .data           = &block_dump,
1241                 .maxlen         = sizeof(block_dump),
1242                 .mode           = 0644,
1243                 .proc_handler   = proc_dointvec,
1244                 .extra1         = &zero,
1245         },
1246         {
1247                 .procname       = "vfs_cache_pressure",
1248                 .data           = &sysctl_vfs_cache_pressure,
1249                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1250                 .mode           = 0644,
1251                 .proc_handler   = proc_dointvec,
1252                 .extra1         = &zero,
1253         },
1254 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1255         {
1256                 .procname       = "legacy_va_layout",
1257                 .data           = &sysctl_legacy_va_layout,
1258                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1259                 .mode           = 0644,
1260                 .proc_handler   = proc_dointvec,
1261                 .extra1         = &zero,
1262         },
1263 #endif
1264 #ifdef CONFIG_NUMA
1265         {
1266                 .procname       = "zone_reclaim_mode",
1267                 .data           = &zone_reclaim_mode,
1268                 .maxlen         = sizeof(zone_reclaim_mode),
1269                 .mode           = 0644,
1270                 .proc_handler   = proc_dointvec,
1271                 .extra1         = &zero,
1272         },
1273         {
1274                 .procname       = "min_unmapped_ratio",
1275                 .data           = &sysctl_min_unmapped_ratio,
1276                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1277                 .mode           = 0644,
1278                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1279                 .extra1         = &zero,
1280                 .extra2         = &one_hundred,
1281         },
1282         {
1283                 .procname       = "min_slab_ratio",
1284                 .data           = &sysctl_min_slab_ratio,
1285                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1286                 .mode           = 0644,
1287                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1288                 .extra1         = &zero,
1289                 .extra2         = &one_hundred,
1290         },
1291 #endif
1292 #ifdef CONFIG_SMP
1293         {
1294                 .procname       = "stat_interval",
1295                 .data           = &sysctl_stat_interval,
1296                 .maxlen         = sizeof(sysctl_stat_interval),
1297                 .mode           = 0644,
1298                 .proc_handler   = proc_dointvec_jiffies,
1299         },
1300 #endif
1301 #ifdef CONFIG_MMU
1302         {
1303                 .procname       = "mmap_min_addr",
1304                 .data           = &dac_mmap_min_addr,
1305                 .maxlen         = sizeof(unsigned long),
1306                 .mode           = 0644,
1307                 .proc_handler   = mmap_min_addr_handler,
1308         },
1309         {
1310                 .procname       = "mmap_noexec_taint",
1311                 .data           = &sysctl_mmap_noexec_taint,
1312                 .maxlen         = sizeof(unsigned long),
1313                 .mode           = 0644,
1314                 .proc_handler   = proc_dointvec_minmax,
1315                 .extra1         = &zero,
1316                 .extra2         = &one,
1317         },
1318 #endif
1319 #ifdef CONFIG_NUMA
1320         {
1321                 .procname       = "numa_zonelist_order",
1322                 .data           = &numa_zonelist_order,
1323                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1324                 .mode           = 0644,
1325                 .proc_handler   = numa_zonelist_order_handler,
1326         },
1327 #endif
1328 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1329    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1330         {
1331                 .procname       = "vdso_enabled",
1332                 .data           = &vdso_enabled,
1333                 .maxlen         = sizeof(vdso_enabled),
1334                 .mode           = 0644,
1335                 .proc_handler   = proc_dointvec,
1336                 .extra1         = &zero,
1337         },
1338 #endif
1339 #ifdef CONFIG_HIGHMEM
1340         {
1341                 .procname       = "highmem_is_dirtyable",
1342                 .data           = &vm_highmem_is_dirtyable,
1343                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1344                 .mode           = 0644,
1345                 .proc_handler   = proc_dointvec_minmax,
1346                 .extra1         = &zero,
1347                 .extra2         = &one,
1348         },
1349 #endif
1350         {
1351                 .procname       = "scan_unevictable_pages",
1352                 .data           = &scan_unevictable_pages,
1353                 .maxlen         = sizeof(scan_unevictable_pages),
1354                 .mode           = 0644,
1355                 .proc_handler   = scan_unevictable_handler,
1356         },
1357 #ifdef CONFIG_MEMORY_FAILURE
1358         {
1359                 .procname       = "memory_failure_early_kill",
1360                 .data           = &sysctl_memory_failure_early_kill,
1361                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1362                 .mode           = 0644,
1363                 .proc_handler   = proc_dointvec_minmax,
1364                 .extra1         = &zero,
1365                 .extra2         = &one,
1366         },
1367         {
1368                 .procname       = "memory_failure_recovery",
1369                 .data           = &sysctl_memory_failure_recovery,
1370                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1371                 .mode           = 0644,
1372                 .proc_handler   = proc_dointvec_minmax,
1373                 .extra1         = &zero,
1374                 .extra2         = &one,
1375         },
1376 #endif
1377         {
1378                 .procname       = "min_filelist_kbytes",
1379                 .data           = &min_filelist_kbytes,
1380                 .maxlen         = sizeof(min_filelist_kbytes),
1381                 .mode           = 0644,
1382                 .proc_handler   = proc_dointvec,
1383         },
1384         { }
1385 };
1386
1387 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1388 static struct ctl_table binfmt_misc_table[] = {
1389         { }
1390 };
1391 #endif
1392
1393 static struct ctl_table fs_table[] = {
1394         {
1395                 .procname       = "inode-nr",
1396                 .data           = &inodes_stat,
1397                 .maxlen         = 2*sizeof(int),
1398                 .mode           = 0444,
1399                 .proc_handler   = proc_nr_inodes,
1400         },
1401         {
1402                 .procname       = "inode-state",
1403                 .data           = &inodes_stat,
1404                 .maxlen         = 7*sizeof(int),
1405                 .mode           = 0444,
1406                 .proc_handler   = proc_nr_inodes,
1407         },
1408         {
1409                 .procname       = "file-nr",
1410                 .data           = &files_stat,
1411                 .maxlen         = sizeof(files_stat),
1412                 .mode           = 0444,
1413                 .proc_handler   = proc_nr_files,
1414         },
1415         {
1416                 .procname       = "file-max",
1417                 .data           = &files_stat.max_files,
1418                 .maxlen         = sizeof(files_stat.max_files),
1419                 .mode           = 0644,
1420                 .proc_handler   = proc_doulongvec_minmax,
1421         },
1422         {
1423                 .procname       = "nr_open",
1424                 .data           = &sysctl_nr_open,
1425                 .maxlen         = sizeof(int),
1426                 .mode           = 0644,
1427                 .proc_handler   = proc_dointvec_minmax,
1428                 .extra1         = &sysctl_nr_open_min,
1429                 .extra2         = &sysctl_nr_open_max,
1430         },
1431         {
1432                 .procname       = "dentry-state",
1433                 .data           = &dentry_stat,
1434                 .maxlen         = 6*sizeof(int),
1435                 .mode           = 0444,
1436                 .proc_handler   = proc_nr_dentry,
1437         },
1438         {
1439                 .procname       = "overflowuid",
1440                 .data           = &fs_overflowuid,
1441                 .maxlen         = sizeof(int),
1442                 .mode           = 0644,
1443                 .proc_handler   = proc_dointvec_minmax,
1444                 .extra1         = &minolduid,
1445                 .extra2         = &maxolduid,
1446         },
1447         {
1448                 .procname       = "overflowgid",
1449                 .data           = &fs_overflowgid,
1450                 .maxlen         = sizeof(int),
1451                 .mode           = 0644,
1452                 .proc_handler   = proc_dointvec_minmax,
1453                 .extra1         = &minolduid,
1454                 .extra2         = &maxolduid,
1455         },
1456 #ifdef CONFIG_FILE_LOCKING
1457         {
1458                 .procname       = "leases-enable",
1459                 .data           = &leases_enable,
1460                 .maxlen         = sizeof(int),
1461                 .mode           = 0644,
1462                 .proc_handler   = proc_dointvec,
1463         },
1464 #endif
1465 #ifdef CONFIG_DNOTIFY
1466         {
1467                 .procname       = "dir-notify-enable",
1468                 .data           = &dir_notify_enable,
1469                 .maxlen         = sizeof(int),
1470                 .mode           = 0644,
1471                 .proc_handler   = proc_dointvec,
1472         },
1473 #endif
1474 #ifdef CONFIG_MMU
1475 #ifdef CONFIG_FILE_LOCKING
1476         {
1477                 .procname       = "lease-break-time",
1478                 .data           = &lease_break_time,
1479                 .maxlen         = sizeof(int),
1480                 .mode           = 0644,
1481                 .proc_handler   = proc_dointvec,
1482         },
1483 #endif
1484 #ifdef CONFIG_AIO
1485         {
1486                 .procname       = "aio-nr",
1487                 .data           = &aio_nr,
1488                 .maxlen         = sizeof(aio_nr),
1489                 .mode           = 0444,
1490                 .proc_handler   = proc_doulongvec_minmax,
1491         },
1492         {
1493                 .procname       = "aio-max-nr",
1494                 .data           = &aio_max_nr,
1495                 .maxlen         = sizeof(aio_max_nr),
1496                 .mode           = 0644,
1497                 .proc_handler   = proc_doulongvec_minmax,
1498         },
1499 #endif /* CONFIG_AIO */
1500 #ifdef CONFIG_INOTIFY_USER
1501         {
1502                 .procname       = "inotify",
1503                 .mode           = 0555,
1504                 .child          = inotify_table,
1505         },
1506 #endif
1507 #ifdef CONFIG_EPOLL
1508         {
1509                 .procname       = "epoll",
1510                 .mode           = 0555,
1511                 .child          = epoll_table,
1512         },
1513 #endif
1514 #endif
1515         {
1516                 .procname       = "suid_dumpable",
1517                 .data           = &suid_dumpable,
1518                 .maxlen         = sizeof(int),
1519                 .mode           = 0644,
1520                 .proc_handler   = proc_dointvec_minmax,
1521                 .extra1         = &zero,
1522                 .extra2         = &two,
1523         },
1524 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1525         {
1526                 .procname       = "binfmt_misc",
1527                 .mode           = 0555,
1528                 .child          = binfmt_misc_table,
1529         },
1530 #endif
1531         {
1532                 .procname       = "pipe-max-size",
1533                 .data           = &pipe_max_size,
1534                 .maxlen         = sizeof(int),
1535                 .mode           = 0644,
1536                 .proc_handler   = &pipe_proc_fn,
1537                 .extra1         = &pipe_min_size,
1538         },
1539         { }
1540 };
1541
1542 static struct ctl_table debug_table[] = {
1543 #if defined(CONFIG_X86) || defined(CONFIG_PPC) || defined(CONFIG_SPARC) || \
1544     defined(CONFIG_S390) || defined(CONFIG_TILE)
1545         {
1546                 .procname       = "exception-trace",
1547                 .data           = &show_unhandled_signals,
1548                 .maxlen         = sizeof(int),
1549                 .mode           = 0644,
1550                 .proc_handler   = proc_dointvec
1551         },
1552 #endif
1553 #if defined(CONFIG_OPTPROBES)
1554         {
1555                 .procname       = "kprobes-optimization",
1556                 .data           = &sysctl_kprobes_optimization,
1557                 .maxlen         = sizeof(int),
1558                 .mode           = 0644,
1559                 .proc_handler   = proc_kprobes_optimization_handler,
1560                 .extra1         = &zero,
1561                 .extra2         = &one,
1562         },
1563 #endif
1564         { }
1565 };
1566
1567 static struct ctl_table dev_table[] = {
1568         { }
1569 };
1570
1571 int __init sysctl_init(void)
1572 {
1573         register_sysctl_table(sysctl_base_table);
1574         return 0;
1575 }
1576
1577 #endif /* CONFIG_SYSCTL */
1578
1579 /*
1580  * /proc/sys support
1581  */
1582
1583 #ifdef CONFIG_PROC_SYSCTL
1584
1585 static int _proc_do_string(void* data, int maxlen, int write,
1586                            void __user *buffer,
1587                            size_t *lenp, loff_t *ppos)
1588 {
1589         size_t len;
1590         char __user *p;
1591         char c;
1592
1593         if (!data || !maxlen || !*lenp) {
1594                 *lenp = 0;
1595                 return 0;
1596         }
1597
1598         if (write) {
1599                 len = 0;
1600                 p = buffer;
1601                 while (len < *lenp) {
1602                         if (get_user(c, p++))
1603                                 return -EFAULT;
1604                         if (c == 0 || c == '\n')
1605                                 break;
1606                         len++;
1607                 }
1608                 if (len >= maxlen)
1609                         len = maxlen-1;
1610                 if(copy_from_user(data, buffer, len))
1611                         return -EFAULT;
1612                 ((char *) data)[len] = 0;
1613                 *ppos += *lenp;
1614         } else {
1615                 len = strlen(data);
1616                 if (len > maxlen)
1617                         len = maxlen;
1618
1619                 if (*ppos > len) {
1620                         *lenp = 0;
1621                         return 0;
1622                 }
1623
1624                 data += *ppos;
1625                 len  -= *ppos;
1626
1627                 if (len > *lenp)
1628                         len = *lenp;
1629                 if (len)
1630                         if(copy_to_user(buffer, data, len))
1631                                 return -EFAULT;
1632                 if (len < *lenp) {
1633                         if(put_user('\n', ((char __user *) buffer) + len))
1634                                 return -EFAULT;
1635                         len++;
1636                 }
1637                 *lenp = len;
1638                 *ppos += len;
1639         }
1640         return 0;
1641 }
1642
1643 /**
1644  * proc_dostring - read a string sysctl
1645  * @table: the sysctl table
1646  * @write: %TRUE if this is a write to the sysctl file
1647  * @buffer: the user buffer
1648  * @lenp: the size of the user buffer
1649  * @ppos: file position
1650  *
1651  * Reads/writes a string from/to the user buffer. If the kernel
1652  * buffer provided is not large enough to hold the string, the
1653  * string is truncated. The copied string is %NULL-terminated.
1654  * If the string is being read by the user process, it is copied
1655  * and a newline '\n' is added. It is truncated if the buffer is
1656  * not large enough.
1657  *
1658  * Returns 0 on success.
1659  */
1660 int proc_dostring(struct ctl_table *table, int write,
1661                   void __user *buffer, size_t *lenp, loff_t *ppos)
1662 {
1663         return _proc_do_string(table->data, table->maxlen, write,
1664                                buffer, lenp, ppos);
1665 }
1666
1667 static size_t proc_skip_spaces(char **buf)
1668 {
1669         size_t ret;
1670         char *tmp = skip_spaces(*buf);
1671         ret = tmp - *buf;
1672         *buf = tmp;
1673         return ret;
1674 }
1675
1676 static void proc_skip_char(char **buf, size_t *size, const char v)
1677 {
1678         while (*size) {
1679                 if (**buf != v)
1680                         break;
1681                 (*size)--;
1682                 (*buf)++;
1683         }
1684 }
1685
1686 #define TMPBUFLEN 22
1687 /**
1688  * proc_get_long - reads an ASCII formatted integer from a user buffer
1689  *
1690  * @buf: a kernel buffer
1691  * @size: size of the kernel buffer
1692  * @val: this is where the number will be stored
1693  * @neg: set to %TRUE if number is negative
1694  * @perm_tr: a vector which contains the allowed trailers
1695  * @perm_tr_len: size of the perm_tr vector
1696  * @tr: pointer to store the trailer character
1697  *
1698  * In case of success %0 is returned and @buf and @size are updated with
1699  * the amount of bytes read. If @tr is non-NULL and a trailing
1700  * character exists (size is non-zero after returning from this
1701  * function), @tr is updated with the trailing character.
1702  */
1703 static int proc_get_long(char **buf, size_t *size,
1704                           unsigned long *val, bool *neg,
1705                           const char *perm_tr, unsigned perm_tr_len, char *tr)
1706 {
1707         int len;
1708         char *p, tmp[TMPBUFLEN];
1709
1710         if (!*size)
1711                 return -EINVAL;
1712
1713         len = *size;
1714         if (len > TMPBUFLEN - 1)
1715                 len = TMPBUFLEN - 1;
1716
1717         memcpy(tmp, *buf, len);
1718
1719         tmp[len] = 0;
1720         p = tmp;
1721         if (*p == '-' && *size > 1) {
1722                 *neg = true;
1723                 p++;
1724         } else
1725                 *neg = false;
1726         if (!isdigit(*p))
1727                 return -EINVAL;
1728
1729         *val = simple_strtoul(p, &p, 0);
1730
1731         len = p - tmp;
1732
1733         /* We don't know if the next char is whitespace thus we may accept
1734          * invalid integers (e.g. 1234...a) or two integers instead of one
1735          * (e.g. 123...1). So lets not allow such large numbers. */
1736         if (len == TMPBUFLEN - 1)
1737                 return -EINVAL;
1738
1739         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
1740                 return -EINVAL;
1741
1742         if (tr && (len < *size))
1743                 *tr = *p;
1744
1745         *buf += len;
1746         *size -= len;
1747
1748         return 0;
1749 }
1750
1751 /**
1752  * proc_put_long - converts an integer to a decimal ASCII formatted string
1753  *
1754  * @buf: the user buffer
1755  * @size: the size of the user buffer
1756  * @val: the integer to be converted
1757  * @neg: sign of the number, %TRUE for negative
1758  *
1759  * In case of success %0 is returned and @buf and @size are updated with
1760  * the amount of bytes written.
1761  */
1762 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
1763                           bool neg)
1764 {
1765         int len;
1766         char tmp[TMPBUFLEN], *p = tmp;
1767
1768         sprintf(p, "%s%lu", neg ? "-" : "", val);
1769         len = strlen(tmp);
1770         if (len > *size)
1771                 len = *size;
1772         if (copy_to_user(*buf, tmp, len))
1773                 return -EFAULT;
1774         *size -= len;
1775         *buf += len;
1776         return 0;
1777 }
1778 #undef TMPBUFLEN
1779
1780 static int proc_put_char(void __user **buf, size_t *size, char c)
1781 {
1782         if (*size) {
1783                 char __user **buffer = (char __user **)buf;
1784                 if (put_user(c, *buffer))
1785                         return -EFAULT;
1786                 (*size)--, (*buffer)++;
1787                 *buf = *buffer;
1788         }
1789         return 0;
1790 }
1791
1792 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
1793                                  int *valp,
1794                                  int write, void *data)
1795 {
1796         if (write) {
1797                 *valp = *negp ? -*lvalp : *lvalp;
1798         } else {
1799                 int val = *valp;
1800                 if (val < 0) {
1801                         *negp = true;
1802                         *lvalp = (unsigned long)-val;
1803                 } else {
1804                         *negp = false;
1805                         *lvalp = (unsigned long)val;
1806                 }
1807         }
1808         return 0;
1809 }
1810
1811 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
1812
1813 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
1814                   int write, void __user *buffer,
1815                   size_t *lenp, loff_t *ppos,
1816                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
1817                               int write, void *data),
1818                   void *data)
1819 {
1820         int *i, vleft, first = 1, err = 0;
1821         unsigned long page = 0;
1822         size_t left;
1823         char *kbuf;
1824
1825         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
1826                 *lenp = 0;
1827                 return 0;
1828         }
1829
1830         i = (int *) tbl_data;
1831         vleft = table->maxlen / sizeof(*i);
1832         left = *lenp;
1833
1834         if (!conv)
1835                 conv = do_proc_dointvec_conv;
1836
1837         if (write) {
1838                 if (left > PAGE_SIZE - 1)
1839                         left = PAGE_SIZE - 1;
1840                 page = __get_free_page(GFP_TEMPORARY);
1841                 kbuf = (char *) page;
1842                 if (!kbuf)
1843                         return -ENOMEM;
1844                 if (copy_from_user(kbuf, buffer, left)) {
1845                         err = -EFAULT;
1846                         goto free;
1847                 }
1848                 kbuf[left] = 0;
1849         }
1850
1851         for (; left && vleft--; i++, first=0) {
1852                 unsigned long lval;
1853                 bool neg;
1854
1855                 if (write) {
1856                         left -= proc_skip_spaces(&kbuf);
1857
1858                         if (!left)
1859                                 break;
1860                         err = proc_get_long(&kbuf, &left, &lval, &neg,
1861                                              proc_wspace_sep,
1862                                              sizeof(proc_wspace_sep), NULL);
1863                         if (err)
1864                                 break;
1865                         if (conv(&neg, &lval, i, 1, data)) {
1866                                 err = -EINVAL;
1867                                 break;
1868                         }
1869                 } else {
1870                         if (conv(&neg, &lval, i, 0, data)) {
1871                                 err = -EINVAL;
1872                                 break;
1873                         }
1874                         if (!first)
1875                                 err = proc_put_char(&buffer, &left, '\t');
1876                         if (err)
1877                                 break;
1878                         err = proc_put_long(&buffer, &left, lval, neg);
1879                         if (err)
1880                                 break;
1881                 }
1882         }
1883
1884         if (!write && !first && left && !err)
1885                 err = proc_put_char(&buffer, &left, '\n');
1886         if (write && !err && left)
1887                 left -= proc_skip_spaces(&kbuf);
1888 free:
1889         if (write) {
1890                 free_page(page);
1891                 if (first)
1892                         return err ? : -EINVAL;
1893         }
1894         *lenp -= left;
1895         *ppos += *lenp;
1896         return err;
1897 }
1898
1899 static int do_proc_dointvec(struct ctl_table *table, int write,
1900                   void __user *buffer, size_t *lenp, loff_t *ppos,
1901                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
1902                               int write, void *data),
1903                   void *data)
1904 {
1905         return __do_proc_dointvec(table->data, table, write,
1906                         buffer, lenp, ppos, conv, data);
1907 }
1908
1909 /**
1910  * proc_dointvec - read a vector of integers
1911  * @table: the sysctl table
1912  * @write: %TRUE if this is a write to the sysctl file
1913  * @buffer: the user buffer
1914  * @lenp: the size of the user buffer
1915  * @ppos: file position
1916  *
1917  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1918  * values from/to the user buffer, treated as an ASCII string.
1919  *
1920  * Returns 0 on success.
1921  */
1922 int proc_dointvec(struct ctl_table *table, int write,
1923                      void __user *buffer, size_t *lenp, loff_t *ppos)
1924 {
1925     return do_proc_dointvec(table,write,buffer,lenp,ppos,
1926                             NULL,NULL);
1927 }
1928
1929 /*
1930  * Taint values can only be increased
1931  * This means we can safely use a temporary.
1932  */
1933 static int proc_taint(struct ctl_table *table, int write,
1934                                void __user *buffer, size_t *lenp, loff_t *ppos)
1935 {
1936         struct ctl_table t;
1937         unsigned long tmptaint = get_taint();
1938         int err;
1939
1940         if (write && !capable(CAP_SYS_ADMIN))
1941                 return -EPERM;
1942
1943         t = *table;
1944         t.data = &tmptaint;
1945         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
1946         if (err < 0)
1947                 return err;
1948
1949         if (write) {
1950                 /*
1951                  * Poor man's atomic or. Not worth adding a primitive
1952                  * to everyone's atomic.h for this
1953                  */
1954                 int i;
1955                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
1956                         if ((tmptaint >> i) & 1)
1957                                 add_taint(i);
1958                 }
1959         }
1960
1961         return err;
1962 }
1963
1964 #ifdef CONFIG_PRINTK
1965 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
1966                                 void __user *buffer, size_t *lenp, loff_t *ppos)
1967 {
1968         if (write && !capable(CAP_SYS_ADMIN))
1969                 return -EPERM;
1970
1971         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1972 }
1973 #endif
1974
1975 struct do_proc_dointvec_minmax_conv_param {
1976         int *min;
1977         int *max;
1978 };
1979
1980 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
1981                                         int *valp,
1982                                         int write, void *data)
1983 {
1984         struct do_proc_dointvec_minmax_conv_param *param = data;
1985         if (write) {
1986                 int val = *negp ? -*lvalp : *lvalp;
1987                 if ((param->min && *param->min > val) ||
1988                     (param->max && *param->max < val))
1989                         return -EINVAL;
1990                 *valp = val;
1991         } else {
1992                 int val = *valp;
1993                 if (val < 0) {
1994                         *negp = true;
1995                         *lvalp = (unsigned long)-val;
1996                 } else {
1997                         *negp = false;
1998                         *lvalp = (unsigned long)val;
1999                 }
2000         }
2001         return 0;
2002 }
2003
2004 /**
2005  * proc_dointvec_minmax - read a vector of integers with min/max values
2006  * @table: the sysctl table
2007  * @write: %TRUE if this is a write to the sysctl file
2008  * @buffer: the user buffer
2009  * @lenp: the size of the user buffer
2010  * @ppos: file position
2011  *
2012  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2013  * values from/to the user buffer, treated as an ASCII string.
2014  *
2015  * This routine will ensure the values are within the range specified by
2016  * table->extra1 (min) and table->extra2 (max).
2017  *
2018  * Returns 0 on success.
2019  */
2020 int proc_dointvec_minmax(struct ctl_table *table, int write,
2021                   void __user *buffer, size_t *lenp, loff_t *ppos)
2022 {
2023         struct do_proc_dointvec_minmax_conv_param param = {
2024                 .min = (int *) table->extra1,
2025                 .max = (int *) table->extra2,
2026         };
2027         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2028                                 do_proc_dointvec_minmax_conv, &param);
2029 }
2030
2031 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2032                                      void __user *buffer,
2033                                      size_t *lenp, loff_t *ppos,
2034                                      unsigned long convmul,
2035                                      unsigned long convdiv)
2036 {
2037         unsigned long *i, *min, *max;
2038         int vleft, first = 1, err = 0;
2039         unsigned long page = 0;
2040         size_t left;
2041         char *kbuf;
2042
2043         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2044                 *lenp = 0;
2045                 return 0;
2046         }
2047
2048         i = (unsigned long *) data;
2049         min = (unsigned long *) table->extra1;
2050         max = (unsigned long *) table->extra2;
2051         vleft = table->maxlen / sizeof(unsigned long);
2052         left = *lenp;
2053
2054         if (write) {
2055                 if (left > PAGE_SIZE - 1)
2056                         left = PAGE_SIZE - 1;
2057                 page = __get_free_page(GFP_TEMPORARY);
2058                 kbuf = (char *) page;
2059                 if (!kbuf)
2060                         return -ENOMEM;
2061                 if (copy_from_user(kbuf, buffer, left)) {
2062                         err = -EFAULT;
2063                         goto free;
2064                 }
2065                 kbuf[left] = 0;
2066         }
2067
2068         for (; left && vleft--; i++, first = 0) {
2069                 unsigned long val;
2070
2071                 if (write) {
2072                         bool neg;
2073
2074                         left -= proc_skip_spaces(&kbuf);
2075
2076                         err = proc_get_long(&kbuf, &left, &val, &neg,
2077                                              proc_wspace_sep,
2078                                              sizeof(proc_wspace_sep), NULL);
2079                         if (err)
2080                                 break;
2081                         if (neg)
2082                                 continue;
2083                         if ((min && val < *min) || (max && val > *max))
2084                                 continue;
2085                         *i = val;
2086                 } else {
2087                         val = convdiv * (*i) / convmul;
2088                         if (!first)
2089                                 err = proc_put_char(&buffer, &left, '\t');
2090                         err = proc_put_long(&buffer, &left, val, false);
2091                         if (err)
2092                                 break;
2093                 }
2094         }
2095
2096         if (!write && !first && left && !err)
2097                 err = proc_put_char(&buffer, &left, '\n');
2098         if (write && !err)
2099                 left -= proc_skip_spaces(&kbuf);
2100 free:
2101         if (write) {
2102                 free_page(page);
2103                 if (first)
2104                         return err ? : -EINVAL;
2105         }
2106         *lenp -= left;
2107         *ppos += *lenp;
2108         return err;
2109 }
2110
2111 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2112                                      void __user *buffer,
2113                                      size_t *lenp, loff_t *ppos,
2114                                      unsigned long convmul,
2115                                      unsigned long convdiv)
2116 {
2117         return __do_proc_doulongvec_minmax(table->data, table, write,
2118                         buffer, lenp, ppos, convmul, convdiv);
2119 }
2120
2121 /**
2122  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2123  * @table: the sysctl table
2124  * @write: %TRUE if this is a write to the sysctl file
2125  * @buffer: the user buffer
2126  * @lenp: the size of the user buffer
2127  * @ppos: file position
2128  *
2129  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2130  * values from/to the user buffer, treated as an ASCII string.
2131  *
2132  * This routine will ensure the values are within the range specified by
2133  * table->extra1 (min) and table->extra2 (max).
2134  *
2135  * Returns 0 on success.
2136  */
2137 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2138                            void __user *buffer, size_t *lenp, loff_t *ppos)
2139 {
2140     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2141 }
2142
2143 /**
2144  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2145  * @table: the sysctl table
2146  * @write: %TRUE if this is a write to the sysctl file
2147  * @buffer: the user buffer
2148  * @lenp: the size of the user buffer
2149  * @ppos: file position
2150  *
2151  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2152  * values from/to the user buffer, treated as an ASCII string. The values
2153  * are treated as milliseconds, and converted to jiffies when they are stored.
2154  *
2155  * This routine will ensure the values are within the range specified by
2156  * table->extra1 (min) and table->extra2 (max).
2157  *
2158  * Returns 0 on success.
2159  */
2160 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2161                                       void __user *buffer,
2162                                       size_t *lenp, loff_t *ppos)
2163 {
2164     return do_proc_doulongvec_minmax(table, write, buffer,
2165                                      lenp, ppos, HZ, 1000l);
2166 }
2167
2168
2169 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2170                                          int *valp,
2171                                          int write, void *data)
2172 {
2173         if (write) {
2174                 if (*lvalp > LONG_MAX / HZ)
2175                         return 1;
2176                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2177         } else {
2178                 int val = *valp;
2179                 unsigned long lval;
2180                 if (val < 0) {
2181                         *negp = true;
2182                         lval = (unsigned long)-val;
2183                 } else {
2184                         *negp = false;
2185                         lval = (unsigned long)val;
2186                 }
2187                 *lvalp = lval / HZ;
2188         }
2189         return 0;
2190 }
2191
2192 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2193                                                 int *valp,
2194                                                 int write, void *data)
2195 {
2196         if (write) {
2197                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2198                         return 1;
2199                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2200         } else {
2201                 int val = *valp;
2202                 unsigned long lval;
2203                 if (val < 0) {
2204                         *negp = true;
2205                         lval = (unsigned long)-val;
2206                 } else {
2207                         *negp = false;
2208                         lval = (unsigned long)val;
2209                 }
2210                 *lvalp = jiffies_to_clock_t(lval);
2211         }
2212         return 0;
2213 }
2214
2215 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2216                                             int *valp,
2217                                             int write, void *data)
2218 {
2219         if (write) {
2220                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2221         } else {
2222                 int val = *valp;
2223                 unsigned long lval;
2224                 if (val < 0) {
2225                         *negp = true;
2226                         lval = (unsigned long)-val;
2227                 } else {
2228                         *negp = false;
2229                         lval = (unsigned long)val;
2230                 }
2231                 *lvalp = jiffies_to_msecs(lval);
2232         }
2233         return 0;
2234 }
2235
2236 /**
2237  * proc_dointvec_jiffies - read a vector of integers as seconds
2238  * @table: the sysctl table
2239  * @write: %TRUE if this is a write to the sysctl file
2240  * @buffer: the user buffer
2241  * @lenp: the size of the user buffer
2242  * @ppos: file position
2243  *
2244  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2245  * values from/to the user buffer, treated as an ASCII string.
2246  * The values read are assumed to be in seconds, and are converted into
2247  * jiffies.
2248  *
2249  * Returns 0 on success.
2250  */
2251 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2252                           void __user *buffer, size_t *lenp, loff_t *ppos)
2253 {
2254     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2255                             do_proc_dointvec_jiffies_conv,NULL);
2256 }
2257
2258 /**
2259  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2260  * @table: the sysctl table
2261  * @write: %TRUE if this is a write to the sysctl file
2262  * @buffer: the user buffer
2263  * @lenp: the size of the user buffer
2264  * @ppos: pointer to the file position
2265  *
2266  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2267  * values from/to the user buffer, treated as an ASCII string.
2268  * The values read are assumed to be in 1/USER_HZ seconds, and
2269  * are converted into jiffies.
2270  *
2271  * Returns 0 on success.
2272  */
2273 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2274                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2275 {
2276     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2277                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2278 }
2279
2280 /**
2281  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2282  * @table: the sysctl table
2283  * @write: %TRUE if this is a write to the sysctl file
2284  * @buffer: the user buffer
2285  * @lenp: the size of the user buffer
2286  * @ppos: file position
2287  * @ppos: the current position in the file
2288  *
2289  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2290  * values from/to the user buffer, treated as an ASCII string.
2291  * The values read are assumed to be in 1/1000 seconds, and
2292  * are converted into jiffies.
2293  *
2294  * Returns 0 on success.
2295  */
2296 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2297                              void __user *buffer, size_t *lenp, loff_t *ppos)
2298 {
2299         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2300                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2301 }
2302
2303 static int proc_do_cad_pid(struct ctl_table *table, int write,
2304                            void __user *buffer, size_t *lenp, loff_t *ppos)
2305 {
2306         struct pid *new_pid;
2307         pid_t tmp;
2308         int r;
2309
2310         tmp = pid_vnr(cad_pid);
2311
2312         r = __do_proc_dointvec(&tmp, table, write, buffer,
2313                                lenp, ppos, NULL, NULL);
2314         if (r || !write)
2315                 return r;
2316
2317         new_pid = find_get_pid(tmp);
2318         if (!new_pid)
2319                 return -ESRCH;
2320
2321         put_pid(xchg(&cad_pid, new_pid));
2322         return 0;
2323 }
2324
2325 /**
2326  * proc_do_large_bitmap - read/write from/to a large bitmap
2327  * @table: the sysctl table
2328  * @write: %TRUE if this is a write to the sysctl file
2329  * @buffer: the user buffer
2330  * @lenp: the size of the user buffer
2331  * @ppos: file position
2332  *
2333  * The bitmap is stored at table->data and the bitmap length (in bits)
2334  * in table->maxlen.
2335  *
2336  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2337  * large bitmaps may be represented in a compact manner. Writing into
2338  * the file will clear the bitmap then update it with the given input.
2339  *
2340  * Returns 0 on success.
2341  */
2342 int proc_do_large_bitmap(struct ctl_table *table, int write,
2343                          void __user *buffer, size_t *lenp, loff_t *ppos)
2344 {
2345         int err = 0;
2346         bool first = 1;
2347         size_t left = *lenp;
2348         unsigned long bitmap_len = table->maxlen;
2349         unsigned long *bitmap = (unsigned long *) table->data;
2350         unsigned long *tmp_bitmap = NULL;
2351         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2352
2353         if (!bitmap_len || !left || (*ppos && !write)) {
2354                 *lenp = 0;
2355                 return 0;
2356         }
2357
2358         if (write) {
2359                 unsigned long page = 0;
2360                 char *kbuf;
2361
2362                 if (left > PAGE_SIZE - 1)
2363                         left = PAGE_SIZE - 1;
2364
2365                 page = __get_free_page(GFP_TEMPORARY);
2366                 kbuf = (char *) page;
2367                 if (!kbuf)
2368                         return -ENOMEM;
2369                 if (copy_from_user(kbuf, buffer, left)) {
2370                         free_page(page);
2371                         return -EFAULT;
2372                 }
2373                 kbuf[left] = 0;
2374
2375                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2376                                      GFP_KERNEL);
2377                 if (!tmp_bitmap) {
2378                         free_page(page);
2379                         return -ENOMEM;
2380                 }
2381                 proc_skip_char(&kbuf, &left, '\n');
2382                 while (!err && left) {
2383                         unsigned long val_a, val_b;
2384                         bool neg;
2385
2386                         err = proc_get_long(&kbuf, &left, &val_a, &neg, tr_a,
2387                                              sizeof(tr_a), &c);
2388                         if (err)
2389                                 break;
2390                         if (val_a >= bitmap_len || neg) {
2391                                 err = -EINVAL;
2392                                 break;
2393                         }
2394
2395                         val_b = val_a;
2396                         if (left) {
2397                                 kbuf++;
2398                                 left--;
2399                         }
2400
2401                         if (c == '-') {
2402                                 err = proc_get_long(&kbuf, &left, &val_b,
2403                                                      &neg, tr_b, sizeof(tr_b),
2404                                                      &c);
2405                                 if (err)
2406                                         break;
2407                                 if (val_b >= bitmap_len || neg ||
2408                                     val_a > val_b) {
2409                                         err = -EINVAL;
2410                                         break;
2411                                 }
2412                                 if (left) {
2413                                         kbuf++;
2414                                         left--;
2415                                 }
2416                         }
2417
2418                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2419                         first = 0;
2420                         proc_skip_char(&kbuf, &left, '\n');
2421                 }
2422                 free_page(page);
2423         } else {
2424                 unsigned long bit_a, bit_b = 0;
2425
2426                 while (left) {
2427                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2428                         if (bit_a >= bitmap_len)
2429                                 break;
2430                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2431                                                    bit_a + 1) - 1;
2432
2433                         if (!first) {
2434                                 err = proc_put_char(&buffer, &left, ',');
2435                                 if (err)
2436                                         break;
2437                         }
2438                         err = proc_put_long(&buffer, &left, bit_a, false);
2439                         if (err)
2440                                 break;
2441                         if (bit_a != bit_b) {
2442                                 err = proc_put_char(&buffer, &left, '-');
2443                                 if (err)
2444                                         break;
2445                                 err = proc_put_long(&buffer, &left, bit_b, false);
2446                                 if (err)
2447                                         break;
2448                         }
2449
2450                         first = 0; bit_b++;
2451                 }
2452                 if (!err)
2453                         err = proc_put_char(&buffer, &left, '\n');
2454         }
2455
2456         if (!err) {
2457                 if (write) {
2458                         if (*ppos)
2459                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2460                         else
2461                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2462                 }
2463                 kfree(tmp_bitmap);
2464                 *lenp -= left;
2465                 *ppos += *lenp;
2466                 return 0;
2467         } else {
2468                 kfree(tmp_bitmap);
2469                 return err;
2470         }
2471 }
2472
2473 #else /* CONFIG_PROC_SYSCTL */
2474
2475 int proc_dostring(struct ctl_table *table, int write,
2476                   void __user *buffer, size_t *lenp, loff_t *ppos)
2477 {
2478         return -ENOSYS;
2479 }
2480
2481 int proc_dointvec(struct ctl_table *table, int write,
2482                   void __user *buffer, size_t *lenp, loff_t *ppos)
2483 {
2484         return -ENOSYS;
2485 }
2486
2487 int proc_dointvec_minmax(struct ctl_table *table, int write,
2488                     void __user *buffer, size_t *lenp, loff_t *ppos)
2489 {
2490         return -ENOSYS;
2491 }
2492
2493 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2494                     void __user *buffer, size_t *lenp, loff_t *ppos)
2495 {
2496         return -ENOSYS;
2497 }
2498
2499 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2500                     void __user *buffer, size_t *lenp, loff_t *ppos)
2501 {
2502         return -ENOSYS;
2503 }
2504
2505 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2506                              void __user *buffer, size_t *lenp, loff_t *ppos)
2507 {
2508         return -ENOSYS;
2509 }
2510
2511 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2512                     void __user *buffer, size_t *lenp, loff_t *ppos)
2513 {
2514         return -ENOSYS;
2515 }
2516
2517 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2518                                       void __user *buffer,
2519                                       size_t *lenp, loff_t *ppos)
2520 {
2521     return -ENOSYS;
2522 }
2523
2524
2525 #endif /* CONFIG_PROC_SYSCTL */
2526
2527 /*
2528  * No sense putting this after each symbol definition, twice,
2529  * exception granted :-)
2530  */
2531 EXPORT_SYMBOL(proc_dointvec);
2532 EXPORT_SYMBOL(proc_dointvec_jiffies);
2533 EXPORT_SYMBOL(proc_dointvec_minmax);
2534 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2535 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2536 EXPORT_SYMBOL(proc_dostring);
2537 EXPORT_SYMBOL(proc_doulongvec_minmax);
2538 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);