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