kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/mm.h>
  23 #include <linux/swap.h>
  24 #include <linux/slab.h>
  25 #include <linux/sysctl.h>
  26 #include <linux/bitmap.h>
  27 #include <linux/signal.h>
  28 #include <linux/printk.h>
  29 #include <linux/proc_fs.h>
  30 #include <linux/security.h>
  31 #include <linux/ctype.h>
  32 #include <linux/kmemcheck.h>
  33 #include <linux/kmemleak.h>
  34 #include <linux/fs.h>
  35 #include <linux/init.h>
  36 #include <linux/kernel.h>
  37 #include <linux/kobject.h>
  38 #include <linux/net.h>
  39 #include <linux/sysrq.h>
  40 #include <linux/highuid.h>
  41 #include <linux/writeback.h>
  42 #include <linux/ratelimit.h>
  43 #include <linux/compaction.h>
  44 #include <linux/hugetlb.h>
  45 #include <linux/initrd.h>
  46 #include <linux/key.h>
  47 #include <linux/times.h>
  48 #include <linux/limits.h>
  49 #include <linux/dcache.h>
  50 #include <linux/dnotify.h>
  51 #include <linux/syscalls.h>
  52 #include <linux/vmstat.h>
  53 #include <linux/nfs_fs.h>
  54 #include <linux/acpi.h>
  55 #include <linux/reboot.h>
  56 #include <linux/ftrace.h>
  57 #include <linux/perf_event.h>
  58 #include <linux/kprobes.h>
  59 #include <linux/pipe_fs_i.h>
  60 #include <linux/oom.h>
  61 #include <linux/kmod.h>
  62 #include <linux/capability.h>
  63 #include <linux/binfmts.h>
  64 #include <linux/sched/sysctl.h>
  65
  66 #include <asm/uaccess.h>
  67 #include <asm/processor.h>
  68
  69 #ifdef CONFIG_X86
  70 #include <asm/nmi.h>
  71 #include <asm/stacktrace.h>
  72 #include <asm/io.h>
  73 #endif
  74 #ifdef CONFIG_SPARC
  75 #include <asm/setup.h>
  76 #endif
  77 #ifdef CONFIG_BSD_PROCESS_ACCT
  78 #include <linux/acct.h>
  79 #endif
  80 #ifdef CONFIG_RT_MUTEXES
  81 #include <linux/rtmutex.h>
  82 #endif
  83 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  84 #include <linux/lockdep.h>
  85 #endif
  86 #ifdef CONFIG_CHR_DEV_SG
  87 #include <scsi/sg.h>
  88 #endif
  89
  90 #ifdef CONFIG_LOCKUP_DETECTOR
  91 #include <linux/nmi.h>
  92 #endif
  93
  94
  95 #if defined(CONFIG_SYSCTL)
  96
  97 /* External variables not in a header file. */
  98 extern int sysctl_overcommit_memory;
  99 extern int sysctl_overcommit_ratio;
 100 extern int max_threads;
 101 extern int suid_dumpable;
 102 #ifdef CONFIG_COREDUMP
 103 extern int core_uses_pid;
 104 extern char core_pattern[];
 105 extern unsigned int core_pipe_limit;
 106 #endif
 107 extern int pid_max;
 108 extern int pid_max_min, pid_max_max;
 109 extern int percpu_pagelist_fraction;
 110 extern int compat_log;
 111 extern int latencytop_enabled;
 112 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 113 #ifndef CONFIG_MMU
 114 extern int sysctl_nr_trim_pages;
 115 #endif
 116 #ifdef CONFIG_BLOCK
 117 extern int blk_iopoll_enabled;
 118 #endif
 119
 120 /* Constants used for minimum and  maximum */
 121 #ifdef CONFIG_LOCKUP_DETECTOR
 122 static int sixty = 60;
 123 static int neg_one = -1;
 124 #endif
 125
 126 static int zero;
 127 static int __maybe_unused one = 1;
 128 static int __maybe_unused two = 2;
 129 static int __maybe_unused three = 3;
 130 static unsigned long one_ul = 1;
 131 static int one_hundred = 100;
 132 #ifdef CONFIG_PRINTK
 133 static int ten_thousand = 10000;
 134 #endif
 135
 136 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 137 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 138
 139 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 140 static int maxolduid = 65535;
 141 static int minolduid;
 142 static int min_percpu_pagelist_fract = 8;
 143
 144 static int ngroups_max = NGROUPS_MAX;
 145 static const int cap_last_cap = CAP_LAST_CAP;
 146
 147 #ifdef CONFIG_INOTIFY_USER
 148 #include <linux/inotify.h>
 149 #endif
 150 #ifdef CONFIG_SPARC
 151 #endif
 152
 153 #ifdef CONFIG_SPARC64
 154 extern int sysctl_tsb_ratio;
 155 #endif
 156
 157 #ifdef __hppa__
 158 extern int pwrsw_enabled;
 159 #endif
 160
 161 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 162 extern int unaligned_enabled;
 163 #endif
 164
 165 #ifdef CONFIG_IA64
 166 extern int unaligned_dump_stack;
 167 #endif
 168
 169 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 170 extern int no_unaligned_warning;
 171 #endif
 172
 173 #ifdef CONFIG_PROC_SYSCTL
 174 static int proc_do_cad_pid(struct ctl_table *table, int write,
 175                   void __user *buffer, size_t *lenp, loff_t *ppos);
 176 static int proc_taint(struct ctl_table *table, int write,
 177                                void __user *buffer, size_t *lenp, loff_t *ppos);
 178 #endif
 179
 180 #ifdef CONFIG_PRINTK
 181 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 182                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 183 #endif
 184
 185 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 186                 void __user *buffer, size_t *lenp, loff_t *ppos);
 187 #ifdef CONFIG_COREDUMP
 188 static int proc_dostring_coredump(struct ctl_table *table, int write,
 189                 void __user *buffer, size_t *lenp, loff_t *ppos);
 190 #endif
 191
 192 #ifdef CONFIG_MAGIC_SYSRQ
 193 /* Note: sysrq code uses it's own private copy */
 194 static int __sysrq_enabled = SYSRQ_DEFAULT_ENABLE;
 195
 196 static int sysrq_sysctl_handler(ctl_table *table, int write,
 197                                 void __user *buffer, size_t *lenp,
 198                                 loff_t *ppos)
 199 {
 200         int error;
 201
 202         error = proc_dointvec(table, write, buffer, lenp, ppos);
 203         if (error)
 204                 return error;
 205
 206         if (write)
 207                 sysrq_toggle_support(__sysrq_enabled);
 208
 209         return 0;
 210 }
 211
 212 #endif
 213
 214 static struct ctl_table kern_table[];
 215 static struct ctl_table vm_table[];
 216 static struct ctl_table fs_table[];
 217 static struct ctl_table debug_table[];
 218 static struct ctl_table dev_table[];
 219 extern struct ctl_table random_table[];
 220 #ifdef CONFIG_EPOLL
 221 extern struct ctl_table epoll_table[];
 222 #endif
 223
 224 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 225 int sysctl_legacy_va_layout;
 226 #endif
 227
 228 /* The default sysctl tables: */
 229
 230 static struct ctl_table sysctl_base_table[] = {
 231         {
 232                 .procname       = "kernel",
 233                 .mode           = 0555,
 234                 .child          = kern_table,
 235         },
 236         {
 237                 .procname       = "vm",
 238                 .mode           = 0555,
 239                 .child          = vm_table,
 240         },
 241         {
 242                 .procname       = "fs",
 243                 .mode           = 0555,
 244                 .child          = fs_table,
 245         },
 246         {
 247                 .procname       = "debug",
 248                 .mode           = 0555,
 249                 .child          = debug_table,
 250         },
 251         {
 252                 .procname       = "dev",
 253                 .mode           = 0555,
 254                 .child          = dev_table,
 255         },
 256         { }
 257 };
 258
 259 #ifdef CONFIG_SCHED_DEBUG
 260 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 261 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 262 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 263 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 264 #ifdef CONFIG_SMP
 265 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 266 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 267 #endif /* CONFIG_SMP */
 268 #endif /* CONFIG_SCHED_DEBUG */
 269
 270 #ifdef CONFIG_COMPACTION
 271 static int min_extfrag_threshold;
 272 static int max_extfrag_threshold = 1000;
 273 #endif
 274
 275 static struct ctl_table kern_table[] = {
 276         {
 277                 .procname       = "sched_child_runs_first",
 278                 .data           = &sysctl_sched_child_runs_first,
 279                 .maxlen         = sizeof(unsigned int),
 280                 .mode           = 0644,
 281                 .proc_handler   = proc_dointvec,
 282         },
 283 #ifdef CONFIG_SCHED_DEBUG
 284         {
 285                 .procname       = "sched_min_granularity_ns",
 286                 .data           = &sysctl_sched_min_granularity,
 287                 .maxlen         = sizeof(unsigned int),
 288                 .mode           = 0644,
 289                 .proc_handler   = sched_proc_update_handler,
 290                 .extra1         = &min_sched_granularity_ns,
 291                 .extra2         = &max_sched_granularity_ns,
 292         },
 293         {
 294                 .procname       = "sched_latency_ns",
 295                 .data           = &sysctl_sched_latency,
 296                 .maxlen         = sizeof(unsigned int),
 297                 .mode           = 0644,
 298                 .proc_handler   = sched_proc_update_handler,
 299                 .extra1         = &min_sched_granularity_ns,
 300                 .extra2         = &max_sched_granularity_ns,
 301         },
 302         {
 303                 .procname       = "sched_wakeup_granularity_ns",
 304                 .data           = &sysctl_sched_wakeup_granularity,
 305                 .maxlen         = sizeof(unsigned int),
 306                 .mode           = 0644,
 307                 .proc_handler   = sched_proc_update_handler,
 308                 .extra1         = &min_wakeup_granularity_ns,
 309                 .extra2         = &max_wakeup_granularity_ns,
 310         },
 311 #ifdef CONFIG_SMP
 312         {
 313                 .procname       = "sched_tunable_scaling",
 314                 .data           = &sysctl_sched_tunable_scaling,
 315                 .maxlen         = sizeof(enum sched_tunable_scaling),
 316                 .mode           = 0644,
 317                 .proc_handler   = sched_proc_update_handler,
 318                 .extra1         = &min_sched_tunable_scaling,
 319                 .extra2         = &max_sched_tunable_scaling,
 320         },
 321         {
 322                 .procname       = "sched_migration_cost_ns",
 323                 .data           = &sysctl_sched_migration_cost,
 324                 .maxlen         = sizeof(unsigned int),
 325                 .mode           = 0644,
 326                 .proc_handler   = proc_dointvec,
 327         },
 328         {
 329                 .procname       = "sched_nr_migrate",
 330                 .data           = &sysctl_sched_nr_migrate,
 331                 .maxlen         = sizeof(unsigned int),
 332                 .mode           = 0644,
 333                 .proc_handler   = proc_dointvec,
 334         },
 335         {
 336                 .procname       = "sched_time_avg_ms",
 337                 .data           = &sysctl_sched_time_avg,
 338                 .maxlen         = sizeof(unsigned int),
 339                 .mode           = 0644,
 340                 .proc_handler   = proc_dointvec,
 341         },
 342         {
 343                 .procname       = "sched_shares_window_ns",
 344                 .data           = &sysctl_sched_shares_window,
 345                 .maxlen         = sizeof(unsigned int),
 346                 .mode           = 0644,
 347                 .proc_handler   = proc_dointvec,
 348         },
 349         {
 350                 .procname       = "timer_migration",
 351                 .data           = &sysctl_timer_migration,
 352                 .maxlen         = sizeof(unsigned int),
 353                 .mode           = 0644,
 354                 .proc_handler   = proc_dointvec_minmax,
 355                 .extra1         = &zero,
 356                 .extra2         = &one,
 357         },
 358 #endif /* CONFIG_SMP */
 359 #ifdef CONFIG_NUMA_BALANCING
 360         {
 361                 .procname       = "numa_balancing_scan_delay_ms",
 362                 .data           = &sysctl_numa_balancing_scan_delay,
 363                 .maxlen         = sizeof(unsigned int),
 364                 .mode           = 0644,
 365                 .proc_handler   = proc_dointvec,
 366         },
 367         {
 368                 .procname       = "numa_balancing_scan_period_min_ms",
 369                 .data           = &sysctl_numa_balancing_scan_period_min,
 370                 .maxlen         = sizeof(unsigned int),
 371                 .mode           = 0644,
 372                 .proc_handler   = proc_dointvec,
 373         },
 374         {
 375                 .procname       = "numa_balancing_scan_period_reset",
 376                 .data           = &sysctl_numa_balancing_scan_period_reset,
 377                 .maxlen         = sizeof(unsigned int),
 378                 .mode           = 0644,
 379                 .proc_handler   = proc_dointvec,
 380         },
 381         {
 382                 .procname       = "numa_balancing_scan_period_max_ms",
 383                 .data           = &sysctl_numa_balancing_scan_period_max,
 384                 .maxlen         = sizeof(unsigned int),
 385                 .mode           = 0644,
 386                 .proc_handler   = proc_dointvec,
 387         },
 388         {
 389                 .procname       = "numa_balancing_scan_size_mb",
 390                 .data           = &sysctl_numa_balancing_scan_size,
 391                 .maxlen         = sizeof(unsigned int),
 392                 .mode           = 0644,
 393                 .proc_handler   = proc_dointvec,
 394         },
 395 #endif /* CONFIG_NUMA_BALANCING */
 396 #endif /* CONFIG_SCHED_DEBUG */
 397         {
 398                 .procname       = "sched_rt_period_us",
 399                 .data           = &sysctl_sched_rt_period,
 400                 .maxlen         = sizeof(unsigned int),
 401                 .mode           = 0644,
 402                 .proc_handler   = sched_rt_handler,
 403         },
 404         {
 405                 .procname       = "sched_rt_runtime_us",
 406                 .data           = &sysctl_sched_rt_runtime,
 407                 .maxlen         = sizeof(int),
 408                 .mode           = 0644,
 409                 .proc_handler   = sched_rt_handler,
 410         },
 411         {
 412                 .procname       = "sched_rr_timeslice_ms",
 413                 .data           = &sched_rr_timeslice,
 414                 .maxlen         = sizeof(int),
 415                 .mode           = 0644,
 416                 .proc_handler   = sched_rr_handler,
 417         },
 418 #ifdef CONFIG_SCHED_AUTOGROUP
 419         {
 420                 .procname       = "sched_autogroup_enabled",
 421                 .data           = &sysctl_sched_autogroup_enabled,
 422                 .maxlen         = sizeof(unsigned int),
 423                 .mode           = 0644,
 424                 .proc_handler   = proc_dointvec_minmax,
 425                 .extra1         = &zero,
 426                 .extra2         = &one,
 427         },
 428 #endif
 429 #ifdef CONFIG_CFS_BANDWIDTH
 430         {
 431                 .procname       = "sched_cfs_bandwidth_slice_us",
 432                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 433                 .maxlen         = sizeof(unsigned int),
 434                 .mode           = 0644,
 435                 .proc_handler   = proc_dointvec_minmax,
 436                 .extra1         = &one,
 437         },
 438 #endif
 439 #ifdef CONFIG_PROVE_LOCKING
 440         {
 441                 .procname       = "prove_locking",
 442                 .data           = &prove_locking,
 443                 .maxlen         = sizeof(int),
 444                 .mode           = 0644,
 445                 .proc_handler   = proc_dointvec,
 446         },
 447 #endif
 448 #ifdef CONFIG_LOCK_STAT
 449         {
 450                 .procname       = "lock_stat",
 451                 .data           = &lock_stat,
 452                 .maxlen         = sizeof(int),
 453                 .mode           = 0644,
 454                 .proc_handler   = proc_dointvec,
 455         },
 456 #endif
 457         {
 458                 .procname       = "panic",
 459                 .data           = &panic_timeout,
 460                 .maxlen         = sizeof(int),
 461                 .mode           = 0644,
 462                 .proc_handler   = proc_dointvec,
 463         },
 464 #ifdef CONFIG_COREDUMP
 465         {
 466                 .procname       = "core_uses_pid",
 467                 .data           = &core_uses_pid,
 468                 .maxlen         = sizeof(int),
 469                 .mode           = 0644,
 470                 .proc_handler   = proc_dointvec,
 471         },
 472         {
 473                 .procname       = "core_pattern",
 474                 .data           = core_pattern,
 475                 .maxlen         = CORENAME_MAX_SIZE,
 476                 .mode           = 0644,
 477                 .proc_handler   = proc_dostring_coredump,
 478         },
 479         {
 480                 .procname       = "core_pipe_limit",
 481                 .data           = &core_pipe_limit,
 482                 .maxlen         = sizeof(unsigned int),
 483                 .mode           = 0644,
 484                 .proc_handler   = proc_dointvec,
 485         },
 486 #endif
 487 #ifdef CONFIG_PROC_SYSCTL
 488         {
 489                 .procname       = "tainted",
 490                 .maxlen         = sizeof(long),
 491                 .mode           = 0644,
 492                 .proc_handler   = proc_taint,
 493         },
 494 #endif
 495 #ifdef CONFIG_LATENCYTOP
 496         {
 497                 .procname       = "latencytop",
 498                 .data           = &latencytop_enabled,
 499                 .maxlen         = sizeof(int),
 500                 .mode           = 0644,
 501                 .proc_handler   = proc_dointvec,
 502         },
 503 #endif
 504 #ifdef CONFIG_BLK_DEV_INITRD
 505         {
 506                 .procname       = "real-root-dev",
 507                 .data           = &real_root_dev,
 508                 .maxlen         = sizeof(int),
 509                 .mode           = 0644,
 510                 .proc_handler   = proc_dointvec,
 511         },
 512 #endif
 513         {
 514                 .procname       = "print-fatal-signals",
 515                 .data           = &print_fatal_signals,
 516                 .maxlen         = sizeof(int),
 517                 .mode           = 0644,
 518                 .proc_handler   = proc_dointvec,
 519         },
 520 #ifdef CONFIG_SPARC
 521         {
 522                 .procname       = "reboot-cmd",
 523                 .data           = reboot_command,
 524                 .maxlen         = 256,
 525                 .mode           = 0644,
 526                 .proc_handler   = proc_dostring,
 527         },
 528         {
 529                 .procname       = "stop-a",
 530                 .data           = &stop_a_enabled,
 531                 .maxlen         = sizeof (int),
 532                 .mode           = 0644,
 533                 .proc_handler   = proc_dointvec,
 534         },
 535         {
 536                 .procname       = "scons-poweroff",
 537                 .data           = &scons_pwroff,
 538                 .maxlen         = sizeof (int),
 539                 .mode           = 0644,
 540                 .proc_handler   = proc_dointvec,
 541         },
 542 #endif
 543 #ifdef CONFIG_SPARC64
 544         {
 545                 .procname       = "tsb-ratio",
 546                 .data           = &sysctl_tsb_ratio,
 547                 .maxlen         = sizeof (int),
 548                 .mode           = 0644,
 549                 .proc_handler   = proc_dointvec,
 550         },
 551 #endif
 552 #ifdef __hppa__
 553         {
 554                 .procname       = "soft-power",
 555                 .data           = &pwrsw_enabled,
 556                 .maxlen         = sizeof (int),
 557                 .mode           = 0644,
 558                 .proc_handler   = proc_dointvec,
 559         },
 560 #endif
 561 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 562         {
 563                 .procname       = "unaligned-trap",
 564                 .data           = &unaligned_enabled,
 565                 .maxlen         = sizeof (int),
 566                 .mode           = 0644,
 567                 .proc_handler   = proc_dointvec,
 568         },
 569 #endif
 570         {
 571                 .procname       = "ctrl-alt-del",
 572                 .data           = &C_A_D,
 573                 .maxlen         = sizeof(int),
 574                 .mode           = 0644,
 575                 .proc_handler   = proc_dointvec,
 576         },
 577 #ifdef CONFIG_FUNCTION_TRACER
 578         {
 579                 .procname       = "ftrace_enabled",
 580                 .data           = &ftrace_enabled,
 581                 .maxlen         = sizeof(int),
 582                 .mode           = 0644,
 583                 .proc_handler   = ftrace_enable_sysctl,
 584         },
 585 #endif
 586 #ifdef CONFIG_STACK_TRACER
 587         {
 588                 .procname       = "stack_tracer_enabled",
 589                 .data           = &stack_tracer_enabled,
 590                 .maxlen         = sizeof(int),
 591                 .mode           = 0644,
 592                 .proc_handler   = stack_trace_sysctl,
 593         },
 594 #endif
 595 #ifdef CONFIG_TRACING
 596         {
 597                 .procname       = "ftrace_dump_on_oops",
 598                 .data           = &ftrace_dump_on_oops,
 599                 .maxlen         = sizeof(int),
 600                 .mode           = 0644,
 601                 .proc_handler   = proc_dointvec,
 602         },
 603         {
 604                 .procname       = "traceoff_on_warning",
 605                 .data           = &__disable_trace_on_warning,
 606                 .maxlen         = sizeof(__disable_trace_on_warning),
 607                 .mode           = 0644,
 608                 .proc_handler   = proc_dointvec,
 609         },
 610 #endif
 611 #ifdef CONFIG_MODULES
 612         {
 613                 .procname       = "modprobe",
 614                 .data           = &modprobe_path,
 615                 .maxlen         = KMOD_PATH_LEN,
 616                 .mode           = 0644,
 617                 .proc_handler   = proc_dostring,
 618         },
 619         {
 620                 .procname       = "modules_disabled",
 621                 .data           = &modules_disabled,
 622                 .maxlen         = sizeof(int),
 623                 .mode           = 0644,
 624                 /* only handle a transition from default "0" to "1" */
 625                 .proc_handler   = proc_dointvec_minmax,
 626                 .extra1         = &one,
 627                 .extra2         = &one,
 628         },
 629 #endif
 630
 631         {
 632                 .procname       = "hotplug",
 633                 .data           = &uevent_helper,
 634                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 635                 .mode           = 0644,
 636                 .proc_handler   = proc_dostring,
 637         },
 638
 639 #ifdef CONFIG_CHR_DEV_SG
 640         {
 641                 .procname       = "sg-big-buff",
 642                 .data           = &sg_big_buff,
 643                 .maxlen         = sizeof (int),
 644                 .mode           = 0444,
 645                 .proc_handler   = proc_dointvec,
 646         },
 647 #endif
 648 #ifdef CONFIG_BSD_PROCESS_ACCT
 649         {
 650                 .procname       = "acct",
 651                 .data           = &acct_parm,
 652                 .maxlen         = 3*sizeof(int),
 653                 .mode           = 0644,
 654                 .proc_handler   = proc_dointvec,
 655         },
 656 #endif
 657 #ifdef CONFIG_MAGIC_SYSRQ
 658         {
 659                 .procname       = "sysrq",
 660                 .data           = &__sysrq_enabled,
 661                 .maxlen         = sizeof (int),
 662                 .mode           = 0644,
 663                 .proc_handler   = sysrq_sysctl_handler,
 664         },
 665 #endif
 666 #ifdef CONFIG_PROC_SYSCTL
 667         {
 668                 .procname       = "cad_pid",
 669                 .data           = NULL,
 670                 .maxlen         = sizeof (int),
 671                 .mode           = 0600,
 672                 .proc_handler   = proc_do_cad_pid,
 673         },
 674 #endif
 675         {
 676                 .procname       = "threads-max",
 677                 .data           = &max_threads,
 678                 .maxlen         = sizeof(int),
 679                 .mode           = 0644,
 680                 .proc_handler   = proc_dointvec,
 681         },
 682         {
 683                 .procname       = "random",
 684                 .mode           = 0555,
 685                 .child          = random_table,
 686         },
 687         {
 688                 .procname       = "usermodehelper",
 689                 .mode           = 0555,
 690                 .child          = usermodehelper_table,
 691         },
 692         {
 693                 .procname       = "overflowuid",
 694                 .data           = &overflowuid,
 695                 .maxlen         = sizeof(int),
 696                 .mode           = 0644,
 697                 .proc_handler   = proc_dointvec_minmax,
 698                 .extra1         = &minolduid,
 699                 .extra2         = &maxolduid,
 700         },
 701         {
 702                 .procname       = "overflowgid",
 703                 .data           = &overflowgid,
 704                 .maxlen         = sizeof(int),
 705                 .mode           = 0644,
 706                 .proc_handler   = proc_dointvec_minmax,
 707                 .extra1         = &minolduid,
 708                 .extra2         = &maxolduid,
 709         },
 710 #ifdef CONFIG_S390
 711 #ifdef CONFIG_MATHEMU
 712         {
 713                 .procname       = "ieee_emulation_warnings",
 714                 .data           = &sysctl_ieee_emulation_warnings,
 715                 .maxlen         = sizeof(int),
 716                 .mode           = 0644,
 717                 .proc_handler   = proc_dointvec,
 718         },
 719 #endif
 720         {
 721                 .procname       = "userprocess_debug",
 722                 .data           = &show_unhandled_signals,
 723                 .maxlen         = sizeof(int),
 724                 .mode           = 0644,
 725                 .proc_handler   = proc_dointvec,
 726         },
 727 #endif
 728         {
 729                 .procname       = "pid_max",
 730                 .data           = &pid_max,
 731                 .maxlen         = sizeof (int),
 732                 .mode           = 0644,
 733                 .proc_handler   = proc_dointvec_minmax,
 734                 .extra1         = &pid_max_min,
 735                 .extra2         = &pid_max_max,
 736         },
 737         {
 738                 .procname       = "panic_on_oops",
 739                 .data           = &panic_on_oops,
 740                 .maxlen         = sizeof(int),
 741                 .mode           = 0644,
 742                 .proc_handler   = proc_dointvec,
 743         },
 744 #if defined CONFIG_PRINTK
 745         {
 746                 .procname       = "printk",
 747                 .data           = &console_loglevel,
 748                 .maxlen         = 4*sizeof(int),
 749                 .mode           = 0644,
 750                 .proc_handler   = proc_dointvec,
 751         },
 752         {
 753                 .procname       = "printk_ratelimit",
 754                 .data           = &printk_ratelimit_state.interval,
 755                 .maxlen         = sizeof(int),
 756                 .mode           = 0644,
 757                 .proc_handler   = proc_dointvec_jiffies,
 758         },
 759         {
 760                 .procname       = "printk_ratelimit_burst",
 761                 .data           = &printk_ratelimit_state.burst,
 762                 .maxlen         = sizeof(int),
 763                 .mode           = 0644,
 764                 .proc_handler   = proc_dointvec,
 765         },
 766         {
 767                 .procname       = "printk_delay",
 768                 .data           = &printk_delay_msec,
 769                 .maxlen         = sizeof(int),
 770                 .mode           = 0644,
 771                 .proc_handler   = proc_dointvec_minmax,
 772                 .extra1         = &zero,
 773                 .extra2         = &ten_thousand,
 774         },
 775         {
 776                 .procname       = "dmesg_restrict",
 777                 .data           = &dmesg_restrict,
 778                 .maxlen         = sizeof(int),
 779                 .mode           = 0644,
 780                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 781                 .extra1         = &zero,
 782                 .extra2         = &one,
 783         },
 784         {
 785                 .procname       = "kptr_restrict",
 786                 .data           = &kptr_restrict,
 787                 .maxlen         = sizeof(int),
 788                 .mode           = 0644,
 789                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 790                 .extra1         = &zero,
 791                 .extra2         = &two,
 792         },
 793 #endif
 794         {
 795                 .procname       = "ngroups_max",
 796                 .data           = &ngroups_max,
 797                 .maxlen         = sizeof (int),
 798                 .mode           = 0444,
 799                 .proc_handler   = proc_dointvec,
 800         },
 801         {
 802                 .procname       = "cap_last_cap",
 803                 .data           = (void *)&cap_last_cap,
 804                 .maxlen         = sizeof(int),
 805                 .mode           = 0444,
 806                 .proc_handler   = proc_dointvec,
 807         },
 808 #if defined(CONFIG_LOCKUP_DETECTOR)
 809         {
 810                 .procname       = "watchdog",
 811                 .data           = &watchdog_enabled,
 812                 .maxlen         = sizeof (int),
 813                 .mode           = 0644,
 814                 .proc_handler   = proc_dowatchdog,
 815                 .extra1         = &zero,
 816                 .extra2         = &one,
 817         },
 818         {
 819                 .procname       = "watchdog_thresh",
 820                 .data           = &watchdog_thresh,
 821                 .maxlen         = sizeof(int),
 822                 .mode           = 0644,
 823                 .proc_handler   = proc_dowatchdog,
 824                 .extra1         = &neg_one,
 825                 .extra2         = &sixty,
 826         },
 827         {
 828                 .procname       = "softlockup_panic",
 829                 .data           = &softlockup_panic,
 830                 .maxlen         = sizeof(int),
 831                 .mode           = 0644,
 832                 .proc_handler   = proc_dointvec_minmax,
 833                 .extra1         = &zero,
 834                 .extra2         = &one,
 835         },
 836         {
 837                 .procname       = "nmi_watchdog",
 838                 .data           = &watchdog_enabled,
 839                 .maxlen         = sizeof (int),
 840                 .mode           = 0644,
 841                 .proc_handler   = proc_dowatchdog,
 842                 .extra1         = &zero,
 843                 .extra2         = &one,
 844         },
 845 #endif
 846 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 847         {
 848                 .procname       = "unknown_nmi_panic",
 849                 .data           = &unknown_nmi_panic,
 850                 .maxlen         = sizeof (int),
 851                 .mode           = 0644,
 852                 .proc_handler   = proc_dointvec,
 853         },
 854 #endif
 855 #if defined(CONFIG_X86)
 856         {
 857                 .procname       = "panic_on_unrecovered_nmi",
 858                 .data           = &panic_on_unrecovered_nmi,
 859                 .maxlen         = sizeof(int),
 860                 .mode           = 0644,
 861                 .proc_handler   = proc_dointvec,
 862         },
 863         {
 864                 .procname       = "panic_on_io_nmi",
 865                 .data           = &panic_on_io_nmi,
 866                 .maxlen         = sizeof(int),
 867                 .mode           = 0644,
 868                 .proc_handler   = proc_dointvec,
 869         },
 870 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 871         {
 872                 .procname       = "panic_on_stackoverflow",
 873                 .data           = &sysctl_panic_on_stackoverflow,
 874                 .maxlen         = sizeof(int),
 875                 .mode           = 0644,
 876                 .proc_handler   = proc_dointvec,
 877         },
 878 #endif
 879         {
 880                 .procname       = "bootloader_type",
 881                 .data           = &bootloader_type,
 882                 .maxlen         = sizeof (int),
 883                 .mode           = 0444,
 884                 .proc_handler   = proc_dointvec,
 885         },
 886         {
 887                 .procname       = "bootloader_version",
 888                 .data           = &bootloader_version,
 889                 .maxlen         = sizeof (int),
 890                 .mode           = 0444,
 891                 .proc_handler   = proc_dointvec,
 892         },
 893         {
 894                 .procname       = "kstack_depth_to_print",
 895                 .data           = &kstack_depth_to_print,
 896                 .maxlen         = sizeof(int),
 897                 .mode           = 0644,
 898                 .proc_handler   = proc_dointvec,
 899         },
 900         {
 901                 .procname       = "io_delay_type",
 902                 .data           = &io_delay_type,
 903                 .maxlen         = sizeof(int),
 904                 .mode           = 0644,
 905                 .proc_handler   = proc_dointvec,
 906         },
 907 #endif
 908 #if defined(CONFIG_MMU)
 909         {
 910                 .procname       = "randomize_va_space",
 911                 .data           = &randomize_va_space,
 912                 .maxlen         = sizeof(int),
 913                 .mode           = 0644,
 914                 .proc_handler   = proc_dointvec,
 915         },
 916 #endif
 917 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 918         {
 919                 .procname       = "spin_retry",
 920                 .data           = &spin_retry,
 921                 .maxlen         = sizeof (int),
 922                 .mode           = 0644,
 923                 .proc_handler   = proc_dointvec,
 924         },
 925 #endif
 926 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 927         {
 928                 .procname       = "acpi_video_flags",
 929                 .data           = &acpi_realmode_flags,
 930                 .maxlen         = sizeof (unsigned long),
 931                 .mode           = 0644,
 932                 .proc_handler   = proc_doulongvec_minmax,
 933         },
 934 #endif
 935 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 936         {
 937                 .procname       = "ignore-unaligned-usertrap",
 938                 .data           = &no_unaligned_warning,
 939                 .maxlen         = sizeof (int),
 940                 .mode           = 0644,
 941                 .proc_handler   = proc_dointvec,
 942         },
 943 #endif
 944 #ifdef CONFIG_IA64
 945         {
 946                 .procname       = "unaligned-dump-stack",
 947                 .data           = &unaligned_dump_stack,
 948                 .maxlen         = sizeof (int),
 949                 .mode           = 0644,
 950                 .proc_handler   = proc_dointvec,
 951         },
 952 #endif
 953 #ifdef CONFIG_DETECT_HUNG_TASK
 954         {
 955                 .procname       = "hung_task_panic",
 956                 .data           = &sysctl_hung_task_panic,
 957                 .maxlen         = sizeof(int),
 958                 .mode           = 0644,
 959                 .proc_handler   = proc_dointvec_minmax,
 960                 .extra1         = &zero,
 961                 .extra2         = &one,
 962         },
 963         {
 964                 .procname       = "hung_task_check_count",
 965                 .data           = &sysctl_hung_task_check_count,
 966                 .maxlen         = sizeof(unsigned long),
 967                 .mode           = 0644,
 968                 .proc_handler   = proc_doulongvec_minmax,
 969         },
 970         {
 971                 .procname       = "hung_task_timeout_secs",
 972                 .data           = &sysctl_hung_task_timeout_secs,
 973                 .maxlen         = sizeof(unsigned long),
 974                 .mode           = 0644,
 975                 .proc_handler   = proc_dohung_task_timeout_secs,
 976         },
 977         {
 978                 .procname       = "hung_task_warnings",
 979                 .data           = &sysctl_hung_task_warnings,
 980                 .maxlen         = sizeof(unsigned long),
 981                 .mode           = 0644,
 982                 .proc_handler   = proc_doulongvec_minmax,
 983         },
 984 #endif
 985 #ifdef CONFIG_COMPAT
 986         {
 987                 .procname       = "compat-log",
 988                 .data           = &compat_log,
 989                 .maxlen         = sizeof (int),
 990                 .mode           = 0644,
 991                 .proc_handler   = proc_dointvec,
 992         },
 993 #endif
 994 #ifdef CONFIG_RT_MUTEXES
 995         {
 996                 .procname       = "max_lock_depth",
 997                 .data           = &max_lock_depth,
 998                 .maxlen         = sizeof(int),
 999                 .mode           = 0644,
1000                 .proc_handler   = proc_dointvec,
1001         },
1002 #endif
1003         {
1004                 .procname       = "poweroff_cmd",
1005                 .data           = &poweroff_cmd,
1006                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1007                 .mode           = 0644,
1008                 .proc_handler   = proc_dostring,
1009         },
1010 #ifdef CONFIG_KEYS
1011         {
1012                 .procname       = "keys",
1013                 .mode           = 0555,
1014                 .child          = key_sysctls,
1015         },
1016 #endif
1017 #ifdef CONFIG_RCU_TORTURE_TEST
1018         {
1019                 .procname       = "rcutorture_runnable",
1020                 .data           = &rcutorture_runnable,
1021                 .maxlen         = sizeof(int),
1022                 .mode           = 0644,
1023                 .proc_handler   = proc_dointvec,
1024         },
1025 #endif
1026 #ifdef CONFIG_PERF_EVENTS
1027         /*
1028          * User-space scripts rely on the existence of this file
1029          * as a feature check for perf_events being enabled.
1030          *
1031          * So it's an ABI, do not remove!
1032          */
1033         {
1034                 .procname       = "perf_event_paranoid",
1035                 .data           = &sysctl_perf_event_paranoid,
1036                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1037                 .mode           = 0644,
1038                 .proc_handler   = proc_dointvec,
1039         },
1040         {
1041                 .procname       = "perf_event_mlock_kb",
1042                 .data           = &sysctl_perf_event_mlock,
1043                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1044                 .mode           = 0644,
1045                 .proc_handler   = proc_dointvec,
1046         },
1047         {
1048                 .procname       = "perf_event_max_sample_rate",
1049                 .data           = &sysctl_perf_event_sample_rate,
1050                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1051                 .mode           = 0644,
1052                 .proc_handler   = perf_proc_update_handler,
1053         },
1054 #endif
1055 #ifdef CONFIG_KMEMCHECK
1056         {
1057                 .procname       = "kmemcheck",
1058                 .data           = &kmemcheck_enabled,
1059                 .maxlen         = sizeof(int),
1060                 .mode           = 0644,
1061                 .proc_handler   = proc_dointvec,
1062         },
1063 #endif
1064 #ifdef CONFIG_BLOCK
1065         {
1066                 .procname       = "blk_iopoll",
1067                 .data           = &blk_iopoll_enabled,
1068                 .maxlen         = sizeof(int),
1069                 .mode           = 0644,
1070                 .proc_handler   = proc_dointvec,
1071         },
1072 #endif
1073         { }
1074 };
1075
1076 static struct ctl_table vm_table[] = {
1077         {
1078                 .procname       = "overcommit_memory",
1079                 .data           = &sysctl_overcommit_memory,
1080                 .maxlen         = sizeof(sysctl_overcommit_memory),
1081                 .mode           = 0644,
1082                 .proc_handler   = proc_dointvec_minmax,
1083                 .extra1         = &zero,
1084                 .extra2         = &two,
1085         },
1086         {
1087                 .procname       = "panic_on_oom",
1088                 .data           = &sysctl_panic_on_oom,
1089                 .maxlen         = sizeof(sysctl_panic_on_oom),
1090                 .mode           = 0644,
1091                 .proc_handler   = proc_dointvec_minmax,
1092                 .extra1         = &zero,
1093                 .extra2         = &two,
1094         },
1095         {
1096                 .procname       = "oom_kill_allocating_task",
1097                 .data           = &sysctl_oom_kill_allocating_task,
1098                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1099                 .mode           = 0644,
1100                 .proc_handler   = proc_dointvec,
1101         },
1102         {
1103                 .procname       = "oom_dump_tasks",
1104                 .data           = &sysctl_oom_dump_tasks,
1105                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1106                 .mode           = 0644,
1107                 .proc_handler   = proc_dointvec,
1108         },
1109         {
1110                 .procname       = "overcommit_ratio",
1111                 .data           = &sysctl_overcommit_ratio,
1112                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1113                 .mode           = 0644,
1114                 .proc_handler   = proc_dointvec,
1115         },
1116         {
1117                 .procname       = "page-cluster",
1118                 .data           = &page_cluster,
1119                 .maxlen         = sizeof(int),
1120                 .mode           = 0644,
1121                 .proc_handler   = proc_dointvec_minmax,
1122                 .extra1         = &zero,
1123         },
1124         {
1125                 .procname       = "dirty_background_ratio",
1126                 .data           = &dirty_background_ratio,
1127                 .maxlen         = sizeof(dirty_background_ratio),
1128                 .mode           = 0644,
1129                 .proc_handler   = dirty_background_ratio_handler,
1130                 .extra1         = &zero,
1131                 .extra2         = &one_hundred,
1132         },
1133         {
1134                 .procname       = "dirty_background_bytes",
1135                 .data           = &dirty_background_bytes,
1136                 .maxlen         = sizeof(dirty_background_bytes),
1137                 .mode           = 0644,
1138                 .proc_handler   = dirty_background_bytes_handler,
1139                 .extra1         = &one_ul,
1140         },
1141         {
1142                 .procname       = "dirty_ratio",
1143                 .data           = &vm_dirty_ratio,
1144                 .maxlen         = sizeof(vm_dirty_ratio),
1145                 .mode           = 0644,
1146                 .proc_handler   = dirty_ratio_handler,
1147                 .extra1         = &zero,
1148                 .extra2         = &one_hundred,
1149         },
1150         {
1151                 .procname       = "dirty_bytes",
1152                 .data           = &vm_dirty_bytes,
1153                 .maxlen         = sizeof(vm_dirty_bytes),
1154                 .mode           = 0644,
1155                 .proc_handler   = dirty_bytes_handler,
1156                 .extra1         = &dirty_bytes_min,
1157         },
1158         {
1159                 .procname       = "dirty_writeback_centisecs",
1160                 .data           = &dirty_writeback_interval,
1161                 .maxlen         = sizeof(dirty_writeback_interval),
1162                 .mode           = 0644,
1163                 .proc_handler   = dirty_writeback_centisecs_handler,
1164         },
1165         {
1166                 .procname       = "dirty_expire_centisecs",
1167                 .data           = &dirty_expire_interval,
1168                 .maxlen         = sizeof(dirty_expire_interval),
1169                 .mode           = 0644,
1170                 .proc_handler   = proc_dointvec_minmax,
1171                 .extra1         = &zero,
1172         },
1173         {
1174                 .procname       = "nr_pdflush_threads",
1175                 .mode           = 0444 /* read-only */,
1176                 .proc_handler   = pdflush_proc_obsolete,
1177         },
1178         {
1179                 .procname       = "swappiness",
1180                 .data           = &vm_swappiness,
1181                 .maxlen         = sizeof(vm_swappiness),
1182                 .mode           = 0644,
1183                 .proc_handler   = proc_dointvec_minmax,
1184                 .extra1         = &zero,
1185                 .extra2         = &one_hundred,
1186         },
1187 #ifdef CONFIG_HUGETLB_PAGE
1188         {
1189                 .procname       = "nr_hugepages",
1190                 .data           = NULL,
1191                 .maxlen         = sizeof(unsigned long),
1192                 .mode           = 0644,
1193                 .proc_handler   = hugetlb_sysctl_handler,
1194                 .extra1         = (void *)&hugetlb_zero,
1195                 .extra2         = (void *)&hugetlb_infinity,
1196         },
1197 #ifdef CONFIG_NUMA
1198         {
1199                 .procname       = "nr_hugepages_mempolicy",
1200                 .data           = NULL,
1201                 .maxlen         = sizeof(unsigned long),
1202                 .mode           = 0644,
1203                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1204                 .extra1         = (void *)&hugetlb_zero,
1205                 .extra2         = (void *)&hugetlb_infinity,
1206         },
1207 #endif
1208          {
1209                 .procname       = "hugetlb_shm_group",
1210                 .data           = &sysctl_hugetlb_shm_group,
1211                 .maxlen         = sizeof(gid_t),
1212                 .mode           = 0644,
1213                 .proc_handler   = proc_dointvec,
1214          },
1215          {
1216                 .procname       = "hugepages_treat_as_movable",
1217                 .data           = &hugepages_treat_as_movable,
1218                 .maxlen         = sizeof(int),
1219                 .mode           = 0644,
1220                 .proc_handler   = hugetlb_treat_movable_handler,
1221         },
1222         {
1223                 .procname       = "nr_overcommit_hugepages",
1224                 .data           = NULL,
1225                 .maxlen         = sizeof(unsigned long),
1226                 .mode           = 0644,
1227                 .proc_handler   = hugetlb_overcommit_handler,
1228                 .extra1         = (void *)&hugetlb_zero,
1229                 .extra2         = (void *)&hugetlb_infinity,
1230         },
1231 #endif
1232         {
1233                 .procname       = "lowmem_reserve_ratio",
1234                 .data           = &sysctl_lowmem_reserve_ratio,
1235                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1236                 .mode           = 0644,
1237                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1238         },
1239         {
1240                 .procname       = "drop_caches",
1241                 .data           = &sysctl_drop_caches,
1242                 .maxlen         = sizeof(int),
1243                 .mode           = 0644,
1244                 .proc_handler   = drop_caches_sysctl_handler,
1245                 .extra1         = &one,
1246                 .extra2         = &three,
1247         },
1248 #ifdef CONFIG_COMPACTION
1249         {
1250                 .procname       = "compact_memory",
1251                 .data           = &sysctl_compact_memory,
1252                 .maxlen         = sizeof(int),
1253                 .mode           = 0200,
1254                 .proc_handler   = sysctl_compaction_handler,
1255         },
1256         {
1257                 .procname       = "extfrag_threshold",
1258                 .data           = &sysctl_extfrag_threshold,
1259                 .maxlen         = sizeof(int),
1260                 .mode           = 0644,
1261                 .proc_handler   = sysctl_extfrag_handler,
1262                 .extra1         = &min_extfrag_threshold,
1263                 .extra2         = &max_extfrag_threshold,
1264         },
1265
1266 #endif /* CONFIG_COMPACTION */
1267         {
1268                 .procname       = "min_free_kbytes",
1269                 .data           = &min_free_kbytes,
1270                 .maxlen         = sizeof(min_free_kbytes),
1271                 .mode           = 0644,
1272                 .proc_handler   = min_free_kbytes_sysctl_handler,
1273                 .extra1         = &zero,
1274         },
1275         {
1276                 .procname       = "percpu_pagelist_fraction",
1277                 .data           = &percpu_pagelist_fraction,
1278                 .maxlen         = sizeof(percpu_pagelist_fraction),
1279                 .mode           = 0644,
1280                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1281                 .extra1         = &min_percpu_pagelist_fract,
1282         },
1283 #ifdef CONFIG_MMU
1284         {
1285                 .procname       = "max_map_count",
1286                 .data           = &sysctl_max_map_count,
1287                 .maxlen         = sizeof(sysctl_max_map_count),
1288                 .mode           = 0644,
1289                 .proc_handler   = proc_dointvec_minmax,
1290                 .extra1         = &zero,
1291         },
1292 #else
1293         {
1294                 .procname       = "nr_trim_pages",
1295                 .data           = &sysctl_nr_trim_pages,
1296                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1297                 .mode           = 0644,
1298                 .proc_handler   = proc_dointvec_minmax,
1299                 .extra1         = &zero,
1300         },
1301 #endif
1302         {
1303                 .procname       = "laptop_mode",
1304                 .data           = &laptop_mode,
1305                 .maxlen         = sizeof(laptop_mode),
1306                 .mode           = 0644,
1307                 .proc_handler   = proc_dointvec_jiffies,
1308         },
1309         {
1310                 .procname       = "block_dump",
1311                 .data           = &block_dump,
1312                 .maxlen         = sizeof(block_dump),
1313                 .mode           = 0644,
1314                 .proc_handler   = proc_dointvec,
1315                 .extra1         = &zero,
1316         },
1317         {
1318                 .procname       = "vfs_cache_pressure",
1319                 .data           = &sysctl_vfs_cache_pressure,
1320                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1321                 .mode           = 0644,
1322                 .proc_handler   = proc_dointvec,
1323                 .extra1         = &zero,
1324         },
1325 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1326         {
1327                 .procname       = "legacy_va_layout",
1328                 .data           = &sysctl_legacy_va_layout,
1329                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1330                 .mode           = 0644,
1331                 .proc_handler   = proc_dointvec,
1332                 .extra1         = &zero,
1333         },
1334 #endif
1335 #ifdef CONFIG_NUMA
1336         {
1337                 .procname       = "zone_reclaim_mode",
1338                 .data           = &zone_reclaim_mode,
1339                 .maxlen         = sizeof(zone_reclaim_mode),
1340                 .mode           = 0644,
1341                 .proc_handler   = proc_dointvec,
1342                 .extra1         = &zero,
1343         },
1344         {
1345                 .procname       = "min_unmapped_ratio",
1346                 .data           = &sysctl_min_unmapped_ratio,
1347                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1348                 .mode           = 0644,
1349                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1350                 .extra1         = &zero,
1351                 .extra2         = &one_hundred,
1352         },
1353         {
1354                 .procname       = "min_slab_ratio",
1355                 .data           = &sysctl_min_slab_ratio,
1356                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1357                 .mode           = 0644,
1358                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1359                 .extra1         = &zero,
1360                 .extra2         = &one_hundred,
1361         },
1362 #endif
1363 #ifdef CONFIG_SMP
1364         {
1365                 .procname       = "stat_interval",
1366                 .data           = &sysctl_stat_interval,
1367                 .maxlen         = sizeof(sysctl_stat_interval),
1368                 .mode           = 0644,
1369                 .proc_handler   = proc_dointvec_jiffies,
1370         },
1371 #endif
1372 #ifdef CONFIG_MMU
1373         {
1374                 .procname       = "mmap_min_addr",
1375                 .data           = &dac_mmap_min_addr,
1376                 .maxlen         = sizeof(unsigned long),
1377                 .mode           = 0644,
1378                 .proc_handler   = mmap_min_addr_handler,
1379         },
1380 #endif
1381 #ifdef CONFIG_NUMA
1382         {
1383                 .procname       = "numa_zonelist_order",
1384                 .data           = &numa_zonelist_order,
1385                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1386                 .mode           = 0644,
1387                 .proc_handler   = numa_zonelist_order_handler,
1388         },
1389 #endif
1390 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1391    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1392         {
1393                 .procname       = "vdso_enabled",
1394                 .data           = &vdso_enabled,
1395                 .maxlen         = sizeof(vdso_enabled),
1396                 .mode           = 0644,
1397                 .proc_handler   = proc_dointvec,
1398                 .extra1         = &zero,
1399         },
1400 #endif
1401 #ifdef CONFIG_HIGHMEM
1402         {
1403                 .procname       = "highmem_is_dirtyable",
1404                 .data           = &vm_highmem_is_dirtyable,
1405                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1406                 .mode           = 0644,
1407                 .proc_handler   = proc_dointvec_minmax,
1408                 .extra1         = &zero,
1409                 .extra2         = &one,
1410         },
1411 #endif
1412         {
1413                 .procname       = "scan_unevictable_pages",
1414                 .data           = &scan_unevictable_pages,
1415                 .maxlen         = sizeof(scan_unevictable_pages),
1416                 .mode           = 0644,
1417                 .proc_handler   = scan_unevictable_handler,
1418         },
1419 #ifdef CONFIG_MEMORY_FAILURE
1420         {
1421                 .procname       = "memory_failure_early_kill",
1422                 .data           = &sysctl_memory_failure_early_kill,
1423                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1424                 .mode           = 0644,
1425                 .proc_handler   = proc_dointvec_minmax,
1426                 .extra1         = &zero,
1427                 .extra2         = &one,
1428         },
1429         {
1430                 .procname       = "memory_failure_recovery",
1431                 .data           = &sysctl_memory_failure_recovery,
1432                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1433                 .mode           = 0644,
1434                 .proc_handler   = proc_dointvec_minmax,
1435                 .extra1         = &zero,
1436                 .extra2         = &one,
1437         },
1438 #endif
1439         {
1440                 .procname       = "user_reserve_kbytes",
1441                 .data           = &sysctl_user_reserve_kbytes,
1442                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1443                 .mode           = 0644,
1444                 .proc_handler   = proc_doulongvec_minmax,
1445         },
1446         {
1447                 .procname       = "admin_reserve_kbytes",
1448                 .data           = &sysctl_admin_reserve_kbytes,
1449                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1450                 .mode           = 0644,
1451                 .proc_handler   = proc_doulongvec_minmax,
1452         },
1453         { }
1454 };
1455
1456 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1457 static struct ctl_table binfmt_misc_table[] = {
1458         { }
1459 };
1460 #endif
1461
1462 static struct ctl_table fs_table[] = {
1463         {
1464                 .procname       = "inode-nr",
1465                 .data           = &inodes_stat,
1466                 .maxlen         = 2*sizeof(int),
1467                 .mode           = 0444,
1468                 .proc_handler   = proc_nr_inodes,
1469         },
1470         {
1471                 .procname       = "inode-state",
1472                 .data           = &inodes_stat,
1473                 .maxlen         = 7*sizeof(int),
1474                 .mode           = 0444,
1475                 .proc_handler   = proc_nr_inodes,
1476         },
1477         {
1478                 .procname       = "file-nr",
1479                 .data           = &files_stat,
1480                 .maxlen         = sizeof(files_stat),
1481                 .mode           = 0444,
1482                 .proc_handler   = proc_nr_files,
1483         },
1484         {
1485                 .procname       = "file-max",
1486                 .data           = &files_stat.max_files,
1487                 .maxlen         = sizeof(files_stat.max_files),
1488                 .mode           = 0644,
1489                 .proc_handler   = proc_doulongvec_minmax,
1490         },
1491         {
1492                 .procname       = "nr_open",
1493                 .data           = &sysctl_nr_open,
1494                 .maxlen         = sizeof(int),
1495                 .mode           = 0644,
1496                 .proc_handler   = proc_dointvec_minmax,
1497                 .extra1         = &sysctl_nr_open_min,
1498                 .extra2         = &sysctl_nr_open_max,
1499         },
1500         {
1501                 .procname       = "dentry-state",
1502                 .data           = &dentry_stat,
1503                 .maxlen         = 6*sizeof(int),
1504                 .mode           = 0444,
1505                 .proc_handler   = proc_nr_dentry,
1506         },
1507         {
1508                 .procname       = "overflowuid",
1509                 .data           = &fs_overflowuid,
1510                 .maxlen         = sizeof(int),
1511                 .mode           = 0644,
1512                 .proc_handler   = proc_dointvec_minmax,
1513                 .extra1         = &minolduid,
1514                 .extra2         = &maxolduid,
1515         },
1516         {
1517                 .procname       = "overflowgid",
1518                 .data           = &fs_overflowgid,
1519                 .maxlen         = sizeof(int),
1520                 .mode           = 0644,
1521                 .proc_handler   = proc_dointvec_minmax,
1522                 .extra1         = &minolduid,
1523                 .extra2         = &maxolduid,
1524         },
1525 #ifdef CONFIG_FILE_LOCKING
1526         {
1527                 .procname       = "leases-enable",
1528                 .data           = &leases_enable,
1529                 .maxlen         = sizeof(int),
1530                 .mode           = 0644,
1531                 .proc_handler   = proc_dointvec,
1532         },
1533 #endif
1534 #ifdef CONFIG_DNOTIFY
1535         {
1536                 .procname       = "dir-notify-enable",
1537                 .data           = &dir_notify_enable,
1538                 .maxlen         = sizeof(int),
1539                 .mode           = 0644,
1540                 .proc_handler   = proc_dointvec,
1541         },
1542 #endif
1543 #ifdef CONFIG_MMU
1544 #ifdef CONFIG_FILE_LOCKING
1545         {
1546                 .procname       = "lease-break-time",
1547                 .data           = &lease_break_time,
1548                 .maxlen         = sizeof(int),
1549                 .mode           = 0644,
1550                 .proc_handler   = proc_dointvec,
1551         },
1552 #endif
1553 #ifdef CONFIG_AIO
1554         {
1555                 .procname       = "aio-nr",
1556                 .data           = &aio_nr,
1557                 .maxlen         = sizeof(aio_nr),
1558                 .mode           = 0444,
1559                 .proc_handler   = proc_doulongvec_minmax,
1560         },
1561         {
1562                 .procname       = "aio-max-nr",
1563                 .data           = &aio_max_nr,
1564                 .maxlen         = sizeof(aio_max_nr),
1565                 .mode           = 0644,
1566                 .proc_handler   = proc_doulongvec_minmax,
1567         },
1568 #endif /* CONFIG_AIO */
1569 #ifdef CONFIG_INOTIFY_USER
1570         {
1571                 .procname       = "inotify",
1572                 .mode           = 0555,
1573                 .child          = inotify_table,
1574         },
1575 #endif
1576 #ifdef CONFIG_EPOLL
1577         {
1578                 .procname       = "epoll",
1579                 .mode           = 0555,
1580                 .child          = epoll_table,
1581         },
1582 #endif
1583 #endif
1584         {
1585                 .procname       = "protected_symlinks",
1586                 .data           = &sysctl_protected_symlinks,
1587                 .maxlen         = sizeof(int),
1588                 .mode           = 0600,
1589                 .proc_handler   = proc_dointvec_minmax,
1590                 .extra1         = &zero,
1591                 .extra2         = &one,
1592         },
1593         {
1594                 .procname       = "protected_hardlinks",
1595                 .data           = &sysctl_protected_hardlinks,
1596                 .maxlen         = sizeof(int),
1597                 .mode           = 0600,
1598                 .proc_handler   = proc_dointvec_minmax,
1599                 .extra1         = &zero,
1600                 .extra2         = &one,
1601         },
1602         {
1603                 .procname       = "suid_dumpable",
1604                 .data           = &suid_dumpable,
1605                 .maxlen         = sizeof(int),
1606                 .mode           = 0644,
1607                 .proc_handler   = proc_dointvec_minmax_coredump,
1608                 .extra1         = &zero,
1609                 .extra2         = &two,
1610         },
1611 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1612         {
1613                 .procname       = "binfmt_misc",
1614                 .mode           = 0555,
1615                 .child          = binfmt_misc_table,
1616         },
1617 #endif
1618         {
1619                 .procname       = "pipe-max-size",
1620                 .data           = &pipe_max_size,
1621                 .maxlen         = sizeof(int),
1622                 .mode           = 0644,
1623                 .proc_handler   = &pipe_proc_fn,
1624                 .extra1         = &pipe_min_size,
1625         },
1626         { }
1627 };
1628
1629 static struct ctl_table debug_table[] = {
1630 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1631         {
1632                 .procname       = "exception-trace",
1633                 .data           = &show_unhandled_signals,
1634                 .maxlen         = sizeof(int),
1635                 .mode           = 0644,
1636                 .proc_handler   = proc_dointvec
1637         },
1638 #endif
1639 #if defined(CONFIG_OPTPROBES)
1640         {
1641                 .procname       = "kprobes-optimization",
1642                 .data           = &sysctl_kprobes_optimization,
1643                 .maxlen         = sizeof(int),
1644                 .mode           = 0644,
1645                 .proc_handler   = proc_kprobes_optimization_handler,
1646                 .extra1         = &zero,
1647                 .extra2         = &one,
1648         },
1649 #endif
1650         { }
1651 };
1652
1653 static struct ctl_table dev_table[] = {
1654         { }
1655 };
1656
1657 int __init sysctl_init(void)
1658 {
1659         struct ctl_table_header *hdr;
1660
1661         hdr = register_sysctl_table(sysctl_base_table);
1662         kmemleak_not_leak(hdr);
1663         return 0;
1664 }
1665
1666 #endif /* CONFIG_SYSCTL */
1667
1668 /*
1669  * /proc/sys support
1670  */
1671
1672 #ifdef CONFIG_PROC_SYSCTL
1673
1674 static int _proc_do_string(void* data, int maxlen, int write,
1675                            void __user *buffer,
1676                            size_t *lenp, loff_t *ppos)
1677 {
1678         size_t len;
1679         char __user *p;
1680         char c;
1681
1682         if (!data || !maxlen || !*lenp) {
1683                 *lenp = 0;
1684                 return 0;
1685         }
1686
1687         if (write) {
1688                 len = 0;
1689                 p = buffer;
1690                 while (len < *lenp) {
1691                         if (get_user(c, p++))
1692                                 return -EFAULT;
1693                         if (c == 0 || c == '\n')
1694                                 break;
1695                         len++;
1696                 }
1697                 if (len >= maxlen)
1698                         len = maxlen-1;
1699                 if(copy_from_user(data, buffer, len))
1700                         return -EFAULT;
1701                 ((char *) data)[len] = 0;
1702                 *ppos += *lenp;
1703         } else {
1704                 len = strlen(data);
1705                 if (len > maxlen)
1706                         len = maxlen;
1707
1708                 if (*ppos > len) {
1709                         *lenp = 0;
1710                         return 0;
1711                 }
1712
1713                 data += *ppos;
1714                 len  -= *ppos;
1715
1716                 if (len > *lenp)
1717                         len = *lenp;
1718                 if (len)
1719                         if(copy_to_user(buffer, data, len))
1720                                 return -EFAULT;
1721                 if (len < *lenp) {
1722                         if(put_user('\n', ((char __user *) buffer) + len))
1723                                 return -EFAULT;
1724                         len++;
1725                 }
1726                 *lenp = len;
1727                 *ppos += len;
1728         }
1729         return 0;
1730 }
1731
1732 /**
1733  * proc_dostring - read a string sysctl
1734  * @table: the sysctl table
1735  * @write: %TRUE if this is a write to the sysctl file
1736  * @buffer: the user buffer
1737  * @lenp: the size of the user buffer
1738  * @ppos: file position
1739  *
1740  * Reads/writes a string from/to the user buffer. If the kernel
1741  * buffer provided is not large enough to hold the string, the
1742  * string is truncated. The copied string is %NULL-terminated.
1743  * If the string is being read by the user process, it is copied
1744  * and a newline '\n' is added. It is truncated if the buffer is
1745  * not large enough.
1746  *
1747  * Returns 0 on success.
1748  */
1749 int proc_dostring(struct ctl_table *table, int write,
1750                   void __user *buffer, size_t *lenp, loff_t *ppos)
1751 {
1752         return _proc_do_string(table->data, table->maxlen, write,
1753                                buffer, lenp, ppos);
1754 }
1755
1756 static size_t proc_skip_spaces(char **buf)
1757 {
1758         size_t ret;
1759         char *tmp = skip_spaces(*buf);
1760         ret = tmp - *buf;
1761         *buf = tmp;
1762         return ret;
1763 }
1764
1765 static void proc_skip_char(char **buf, size_t *size, const char v)
1766 {
1767         while (*size) {
1768                 if (**buf != v)
1769                         break;
1770                 (*size)--;
1771                 (*buf)++;
1772         }
1773 }
1774
1775 #define TMPBUFLEN 22
1776 /**
1777  * proc_get_long - reads an ASCII formatted integer from a user buffer
1778  *
1779  * @buf: a kernel buffer
1780  * @size: size of the kernel buffer
1781  * @val: this is where the number will be stored
1782  * @neg: set to %TRUE if number is negative
1783  * @perm_tr: a vector which contains the allowed trailers
1784  * @perm_tr_len: size of the perm_tr vector
1785  * @tr: pointer to store the trailer character
1786  *
1787  * In case of success %0 is returned and @buf and @size are updated with
1788  * the amount of bytes read. If @tr is non-NULL and a trailing
1789  * character exists (size is non-zero after returning from this
1790  * function), @tr is updated with the trailing character.
1791  */
1792 static int proc_get_long(char **buf, size_t *size,
1793                           unsigned long *val, bool *neg,
1794                           const char *perm_tr, unsigned perm_tr_len, char *tr)
1795 {
1796         int len;
1797         char *p, tmp[TMPBUFLEN];
1798
1799         if (!*size)
1800                 return -EINVAL;
1801
1802         len = *size;
1803         if (len > TMPBUFLEN - 1)
1804                 len = TMPBUFLEN - 1;
1805
1806         memcpy(tmp, *buf, len);
1807
1808         tmp[len] = 0;
1809         p = tmp;
1810         if (*p == '-' && *size > 1) {
1811                 *neg = true;
1812                 p++;
1813         } else
1814                 *neg = false;
1815         if (!isdigit(*p))
1816                 return -EINVAL;
1817
1818         *val = simple_strtoul(p, &p, 0);
1819
1820         len = p - tmp;
1821
1822         /* We don't know if the next char is whitespace thus we may accept
1823          * invalid integers (e.g. 1234...a) or two integers instead of one
1824          * (e.g. 123...1). So lets not allow such large numbers. */
1825         if (len == TMPBUFLEN - 1)
1826                 return -EINVAL;
1827
1828         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
1829                 return -EINVAL;
1830
1831         if (tr && (len < *size))
1832                 *tr = *p;
1833
1834         *buf += len;
1835         *size -= len;
1836
1837         return 0;
1838 }
1839
1840 /**
1841  * proc_put_long - converts an integer to a decimal ASCII formatted string
1842  *
1843  * @buf: the user buffer
1844  * @size: the size of the user buffer
1845  * @val: the integer to be converted
1846  * @neg: sign of the number, %TRUE for negative
1847  *
1848  * In case of success %0 is returned and @buf and @size are updated with
1849  * the amount of bytes written.
1850  */
1851 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
1852                           bool neg)
1853 {
1854         int len;
1855         char tmp[TMPBUFLEN], *p = tmp;
1856
1857         sprintf(p, "%s%lu", neg ? "-" : "", val);
1858         len = strlen(tmp);
1859         if (len > *size)
1860                 len = *size;
1861         if (copy_to_user(*buf, tmp, len))
1862                 return -EFAULT;
1863         *size -= len;
1864         *buf += len;
1865         return 0;
1866 }
1867 #undef TMPBUFLEN
1868
1869 static int proc_put_char(void __user **buf, size_t *size, char c)
1870 {
1871         if (*size) {
1872                 char __user **buffer = (char __user **)buf;
1873                 if (put_user(c, *buffer))
1874                         return -EFAULT;
1875                 (*size)--, (*buffer)++;
1876                 *buf = *buffer;
1877         }
1878         return 0;
1879 }
1880
1881 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
1882                                  int *valp,
1883                                  int write, void *data)
1884 {
1885         if (write) {
1886                 *valp = *negp ? -*lvalp : *lvalp;
1887         } else {
1888                 int val = *valp;
1889                 if (val < 0) {
1890                         *negp = true;
1891                         *lvalp = (unsigned long)-val;
1892                 } else {
1893                         *negp = false;
1894                         *lvalp = (unsigned long)val;
1895                 }
1896         }
1897         return 0;
1898 }
1899
1900 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
1901
1902 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
1903                   int write, void __user *buffer,
1904                   size_t *lenp, loff_t *ppos,
1905                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
1906                               int write, void *data),
1907                   void *data)
1908 {
1909         int *i, vleft, first = 1, err = 0;
1910         unsigned long page = 0;
1911         size_t left;
1912         char *kbuf;
1913
1914         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
1915                 *lenp = 0;
1916                 return 0;
1917         }
1918
1919         i = (int *) tbl_data;
1920         vleft = table->maxlen / sizeof(*i);
1921         left = *lenp;
1922
1923         if (!conv)
1924                 conv = do_proc_dointvec_conv;
1925
1926         if (write) {
1927                 if (left > PAGE_SIZE - 1)
1928                         left = PAGE_SIZE - 1;
1929                 page = __get_free_page(GFP_TEMPORARY);
1930                 kbuf = (char *) page;
1931                 if (!kbuf)
1932                         return -ENOMEM;
1933                 if (copy_from_user(kbuf, buffer, left)) {
1934                         err = -EFAULT;
1935                         goto free;
1936                 }
1937                 kbuf[left] = 0;
1938         }
1939
1940         for (; left && vleft--; i++, first=0) {
1941                 unsigned long lval;
1942                 bool neg;
1943
1944                 if (write) {
1945                         left -= proc_skip_spaces(&kbuf);
1946
1947                         if (!left)
1948                                 break;
1949                         err = proc_get_long(&kbuf, &left, &lval, &neg,
1950                                              proc_wspace_sep,
1951                                              sizeof(proc_wspace_sep), NULL);
1952                         if (err)
1953                                 break;
1954                         if (conv(&neg, &lval, i, 1, data)) {
1955                                 err = -EINVAL;
1956                                 break;
1957                         }
1958                 } else {
1959                         if (conv(&neg, &lval, i, 0, data)) {
1960                                 err = -EINVAL;
1961                                 break;
1962                         }
1963                         if (!first)
1964                                 err = proc_put_char(&buffer, &left, '\t');
1965                         if (err)
1966                                 break;
1967                         err = proc_put_long(&buffer, &left, lval, neg);
1968                         if (err)
1969                                 break;
1970                 }
1971         }
1972
1973         if (!write && !first && left && !err)
1974                 err = proc_put_char(&buffer, &left, '\n');
1975         if (write && !err && left)
1976                 left -= proc_skip_spaces(&kbuf);
1977 free:
1978         if (write) {
1979                 free_page(page);
1980                 if (first)
1981                         return err ? : -EINVAL;
1982         }
1983         *lenp -= left;
1984         *ppos += *lenp;
1985         return err;
1986 }
1987
1988 static int do_proc_dointvec(struct ctl_table *table, int write,
1989                   void __user *buffer, size_t *lenp, loff_t *ppos,
1990                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
1991                               int write, void *data),
1992                   void *data)
1993 {
1994         return __do_proc_dointvec(table->data, table, write,
1995                         buffer, lenp, ppos, conv, data);
1996 }
1997
1998 /**
1999  * proc_dointvec - read a vector of integers
2000  * @table: the sysctl table
2001  * @write: %TRUE if this is a write to the sysctl file
2002  * @buffer: the user buffer
2003  * @lenp: the size of the user buffer
2004  * @ppos: file position
2005  *
2006  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2007  * values from/to the user buffer, treated as an ASCII string.
2008  *
2009  * Returns 0 on success.
2010  */
2011 int proc_dointvec(struct ctl_table *table, int write,
2012                      void __user *buffer, size_t *lenp, loff_t *ppos)
2013 {
2014     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2015                             NULL,NULL);
2016 }
2017
2018 /*
2019  * Taint values can only be increased
2020  * This means we can safely use a temporary.
2021  */
2022 static int proc_taint(struct ctl_table *table, int write,
2023                                void __user *buffer, size_t *lenp, loff_t *ppos)
2024 {
2025         struct ctl_table t;
2026         unsigned long tmptaint = get_taint();
2027         int err;
2028
2029         if (write && !capable(CAP_SYS_ADMIN))
2030                 return -EPERM;
2031
2032         t = *table;
2033         t.data = &tmptaint;
2034         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2035         if (err < 0)
2036                 return err;
2037
2038         if (write) {
2039                 /*
2040                  * Poor man's atomic or. Not worth adding a primitive
2041                  * to everyone's atomic.h for this
2042                  */
2043                 int i;
2044                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2045                         if ((tmptaint >> i) & 1)
2046                                 add_taint(i, LOCKDEP_STILL_OK);
2047                 }
2048         }
2049
2050         return err;
2051 }
2052
2053 #ifdef CONFIG_PRINTK
2054 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2055                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2056 {
2057         if (write && !capable(CAP_SYS_ADMIN))
2058                 return -EPERM;
2059
2060         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2061 }
2062 #endif
2063
2064 struct do_proc_dointvec_minmax_conv_param {
2065         int *min;
2066         int *max;
2067 };
2068
2069 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2070                                         int *valp,
2071                                         int write, void *data)
2072 {
2073         struct do_proc_dointvec_minmax_conv_param *param = data;
2074         if (write) {
2075                 int val = *negp ? -*lvalp : *lvalp;
2076                 if ((param->min && *param->min > val) ||
2077                     (param->max && *param->max < val))
2078                         return -EINVAL;
2079                 *valp = val;
2080         } else {
2081                 int val = *valp;
2082                 if (val < 0) {
2083                         *negp = true;
2084                         *lvalp = (unsigned long)-val;
2085                 } else {
2086                         *negp = false;
2087                         *lvalp = (unsigned long)val;
2088                 }
2089         }
2090         return 0;
2091 }
2092
2093 /**
2094  * proc_dointvec_minmax - read a vector of integers with min/max values
2095  * @table: the sysctl table
2096  * @write: %TRUE if this is a write to the sysctl file
2097  * @buffer: the user buffer
2098  * @lenp: the size of the user buffer
2099  * @ppos: file position
2100  *
2101  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2102  * values from/to the user buffer, treated as an ASCII string.
2103  *
2104  * This routine will ensure the values are within the range specified by
2105  * table->extra1 (min) and table->extra2 (max).
2106  *
2107  * Returns 0 on success.
2108  */
2109 int proc_dointvec_minmax(struct ctl_table *table, int write,
2110                   void __user *buffer, size_t *lenp, loff_t *ppos)
2111 {
2112         struct do_proc_dointvec_minmax_conv_param param = {
2113                 .min = (int *) table->extra1,
2114                 .max = (int *) table->extra2,
2115         };
2116         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2117                                 do_proc_dointvec_minmax_conv, &param);
2118 }
2119
2120 static void validate_coredump_safety(void)
2121 {
2122 #ifdef CONFIG_COREDUMP
2123         if (suid_dumpable == SUID_DUMP_ROOT &&
2124             core_pattern[0] != '/' && core_pattern[0] != '|') {
2125                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2126                         "suid_dumpable=2. Pipe handler or fully qualified "\
2127                         "core dump path required.\n");
2128         }
2129 #endif
2130 }
2131
2132 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2133                 void __user *buffer, size_t *lenp, loff_t *ppos)
2134 {
2135         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2136         if (!error)
2137                 validate_coredump_safety();
2138         return error;
2139 }
2140
2141 #ifdef CONFIG_COREDUMP
2142 static int proc_dostring_coredump(struct ctl_table *table, int write,
2143                   void __user *buffer, size_t *lenp, loff_t *ppos)
2144 {
2145         int error = proc_dostring(table, write, buffer, lenp, ppos);
2146         if (!error)
2147                 validate_coredump_safety();
2148         return error;
2149 }
2150 #endif
2151
2152 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2153                                      void __user *buffer,
2154                                      size_t *lenp, loff_t *ppos,
2155                                      unsigned long convmul,
2156                                      unsigned long convdiv)
2157 {
2158         unsigned long *i, *min, *max;
2159         int vleft, first = 1, err = 0;
2160         unsigned long page = 0;
2161         size_t left;
2162         char *kbuf;
2163
2164         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2165                 *lenp = 0;
2166                 return 0;
2167         }
2168
2169         i = (unsigned long *) data;
2170         min = (unsigned long *) table->extra1;
2171         max = (unsigned long *) table->extra2;
2172         vleft = table->maxlen / sizeof(unsigned long);
2173         left = *lenp;
2174
2175         if (write) {
2176                 if (left > PAGE_SIZE - 1)
2177                         left = PAGE_SIZE - 1;
2178                 page = __get_free_page(GFP_TEMPORARY);
2179                 kbuf = (char *) page;
2180                 if (!kbuf)
2181                         return -ENOMEM;
2182                 if (copy_from_user(kbuf, buffer, left)) {
2183                         err = -EFAULT;
2184                         goto free;
2185                 }
2186                 kbuf[left] = 0;
2187         }
2188
2189         for (; left && vleft--; i++, first = 0) {
2190                 unsigned long val;
2191
2192                 if (write) {
2193                         bool neg;
2194
2195                         left -= proc_skip_spaces(&kbuf);
2196
2197                         err = proc_get_long(&kbuf, &left, &val, &neg,
2198                                              proc_wspace_sep,
2199                                              sizeof(proc_wspace_sep), NULL);
2200                         if (err)
2201                                 break;
2202                         if (neg)
2203                                 continue;
2204                         if ((min && val < *min) || (max && val > *max))
2205                                 continue;
2206                         *i = val;
2207                 } else {
2208                         val = convdiv * (*i) / convmul;
2209                         if (!first)
2210                                 err = proc_put_char(&buffer, &left, '\t');
2211                         err = proc_put_long(&buffer, &left, val, false);
2212                         if (err)
2213                                 break;
2214                 }
2215         }
2216
2217         if (!write && !first && left && !err)
2218                 err = proc_put_char(&buffer, &left, '\n');
2219         if (write && !err)
2220                 left -= proc_skip_spaces(&kbuf);
2221 free:
2222         if (write) {
2223                 free_page(page);
2224                 if (first)
2225                         return err ? : -EINVAL;
2226         }
2227         *lenp -= left;
2228         *ppos += *lenp;
2229         return err;
2230 }
2231
2232 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2233                                      void __user *buffer,
2234                                      size_t *lenp, loff_t *ppos,
2235                                      unsigned long convmul,
2236                                      unsigned long convdiv)
2237 {
2238         return __do_proc_doulongvec_minmax(table->data, table, write,
2239                         buffer, lenp, ppos, convmul, convdiv);
2240 }
2241
2242 /**
2243  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2244  * @table: the sysctl table
2245  * @write: %TRUE if this is a write to the sysctl file
2246  * @buffer: the user buffer
2247  * @lenp: the size of the user buffer
2248  * @ppos: file position
2249  *
2250  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2251  * values from/to the user buffer, treated as an ASCII string.
2252  *
2253  * This routine will ensure the values are within the range specified by
2254  * table->extra1 (min) and table->extra2 (max).
2255  *
2256  * Returns 0 on success.
2257  */
2258 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2259                            void __user *buffer, size_t *lenp, loff_t *ppos)
2260 {
2261     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2262 }
2263
2264 /**
2265  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2266  * @table: the sysctl table
2267  * @write: %TRUE if this is a write to the sysctl file
2268  * @buffer: the user buffer
2269  * @lenp: the size of the user buffer
2270  * @ppos: file position
2271  *
2272  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2273  * values from/to the user buffer, treated as an ASCII string. The values
2274  * are treated as milliseconds, and converted to jiffies when they are stored.
2275  *
2276  * This routine will ensure the values are within the range specified by
2277  * table->extra1 (min) and table->extra2 (max).
2278  *
2279  * Returns 0 on success.
2280  */
2281 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2282                                       void __user *buffer,
2283                                       size_t *lenp, loff_t *ppos)
2284 {
2285     return do_proc_doulongvec_minmax(table, write, buffer,
2286                                      lenp, ppos, HZ, 1000l);
2287 }
2288
2289
2290 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2291                                          int *valp,
2292                                          int write, void *data)
2293 {
2294         if (write) {
2295                 if (*lvalp > LONG_MAX / HZ)
2296                         return 1;
2297                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2298         } else {
2299                 int val = *valp;
2300                 unsigned long lval;
2301                 if (val < 0) {
2302                         *negp = true;
2303                         lval = (unsigned long)-val;
2304                 } else {
2305                         *negp = false;
2306                         lval = (unsigned long)val;
2307                 }
2308                 *lvalp = lval / HZ;
2309         }
2310         return 0;
2311 }
2312
2313 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2314                                                 int *valp,
2315                                                 int write, void *data)
2316 {
2317         if (write) {
2318                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2319                         return 1;
2320                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2321         } else {
2322                 int val = *valp;
2323                 unsigned long lval;
2324                 if (val < 0) {
2325                         *negp = true;
2326                         lval = (unsigned long)-val;
2327                 } else {
2328                         *negp = false;
2329                         lval = (unsigned long)val;
2330                 }
2331                 *lvalp = jiffies_to_clock_t(lval);
2332         }
2333         return 0;
2334 }
2335
2336 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2337                                             int *valp,
2338                                             int write, void *data)
2339 {
2340         if (write) {
2341                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2342         } else {
2343                 int val = *valp;
2344                 unsigned long lval;
2345                 if (val < 0) {
2346                         *negp = true;
2347                         lval = (unsigned long)-val;
2348                 } else {
2349                         *negp = false;
2350                         lval = (unsigned long)val;
2351                 }
2352                 *lvalp = jiffies_to_msecs(lval);
2353         }
2354         return 0;
2355 }
2356
2357 /**
2358  * proc_dointvec_jiffies - read a vector of integers as seconds
2359  * @table: the sysctl table
2360  * @write: %TRUE if this is a write to the sysctl file
2361  * @buffer: the user buffer
2362  * @lenp: the size of the user buffer
2363  * @ppos: file position
2364  *
2365  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2366  * values from/to the user buffer, treated as an ASCII string.
2367  * The values read are assumed to be in seconds, and are converted into
2368  * jiffies.
2369  *
2370  * Returns 0 on success.
2371  */
2372 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2373                           void __user *buffer, size_t *lenp, loff_t *ppos)
2374 {
2375     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2376                             do_proc_dointvec_jiffies_conv,NULL);
2377 }
2378
2379 /**
2380  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2381  * @table: the sysctl table
2382  * @write: %TRUE if this is a write to the sysctl file
2383  * @buffer: the user buffer
2384  * @lenp: the size of the user buffer
2385  * @ppos: pointer to the file position
2386  *
2387  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2388  * values from/to the user buffer, treated as an ASCII string.
2389  * The values read are assumed to be in 1/USER_HZ seconds, and
2390  * are converted into jiffies.
2391  *
2392  * Returns 0 on success.
2393  */
2394 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2395                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2396 {
2397     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2398                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2399 }
2400
2401 /**
2402  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2403  * @table: the sysctl table
2404  * @write: %TRUE if this is a write to the sysctl file
2405  * @buffer: the user buffer
2406  * @lenp: the size of the user buffer
2407  * @ppos: file position
2408  * @ppos: the current position in the file
2409  *
2410  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2411  * values from/to the user buffer, treated as an ASCII string.
2412  * The values read are assumed to be in 1/1000 seconds, and
2413  * are converted into jiffies.
2414  *
2415  * Returns 0 on success.
2416  */
2417 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2418                              void __user *buffer, size_t *lenp, loff_t *ppos)
2419 {
2420         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2421                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2422 }
2423
2424 static int proc_do_cad_pid(struct ctl_table *table, int write,
2425                            void __user *buffer, size_t *lenp, loff_t *ppos)
2426 {
2427         struct pid *new_pid;
2428         pid_t tmp;
2429         int r;
2430
2431         tmp = pid_vnr(cad_pid);
2432
2433         r = __do_proc_dointvec(&tmp, table, write, buffer,
2434                                lenp, ppos, NULL, NULL);
2435         if (r || !write)
2436                 return r;
2437
2438         new_pid = find_get_pid(tmp);
2439         if (!new_pid)
2440                 return -ESRCH;
2441
2442         put_pid(xchg(&cad_pid, new_pid));
2443         return 0;
2444 }
2445
2446 /**
2447  * proc_do_large_bitmap - read/write from/to a large bitmap
2448  * @table: the sysctl table
2449  * @write: %TRUE if this is a write to the sysctl file
2450  * @buffer: the user buffer
2451  * @lenp: the size of the user buffer
2452  * @ppos: file position
2453  *
2454  * The bitmap is stored at table->data and the bitmap length (in bits)
2455  * in table->maxlen.
2456  *
2457  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2458  * large bitmaps may be represented in a compact manner. Writing into
2459  * the file will clear the bitmap then update it with the given input.
2460  *
2461  * Returns 0 on success.
2462  */
2463 int proc_do_large_bitmap(struct ctl_table *table, int write,
2464                          void __user *buffer, size_t *lenp, loff_t *ppos)
2465 {
2466         int err = 0;
2467         bool first = 1;
2468         size_t left = *lenp;
2469         unsigned long bitmap_len = table->maxlen;
2470         unsigned long *bitmap = (unsigned long *) table->data;
2471         unsigned long *tmp_bitmap = NULL;
2472         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2473
2474         if (!bitmap_len || !left || (*ppos && !write)) {
2475                 *lenp = 0;
2476                 return 0;
2477         }
2478
2479         if (write) {
2480                 unsigned long page = 0;
2481                 char *kbuf;
2482
2483                 if (left > PAGE_SIZE - 1)
2484                         left = PAGE_SIZE - 1;
2485
2486                 page = __get_free_page(GFP_TEMPORARY);
2487                 kbuf = (char *) page;
2488                 if (!kbuf)
2489                         return -ENOMEM;
2490                 if (copy_from_user(kbuf, buffer, left)) {
2491                         free_page(page);
2492                         return -EFAULT;
2493                 }
2494                 kbuf[left] = 0;
2495
2496                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2497                                      GFP_KERNEL);
2498                 if (!tmp_bitmap) {
2499                         free_page(page);
2500                         return -ENOMEM;
2501                 }
2502                 proc_skip_char(&kbuf, &left, '\n');
2503                 while (!err && left) {
2504                         unsigned long val_a, val_b;
2505                         bool neg;
2506
2507                         err = proc_get_long(&kbuf, &left, &val_a, &neg, tr_a,
2508                                              sizeof(tr_a), &c);
2509                         if (err)
2510                                 break;
2511                         if (val_a >= bitmap_len || neg) {
2512                                 err = -EINVAL;
2513                                 break;
2514                         }
2515
2516                         val_b = val_a;
2517                         if (left) {
2518                                 kbuf++;
2519                                 left--;
2520                         }
2521
2522                         if (c == '-') {
2523                                 err = proc_get_long(&kbuf, &left, &val_b,
2524                                                      &neg, tr_b, sizeof(tr_b),
2525                                                      &c);
2526                                 if (err)
2527                                         break;
2528                                 if (val_b >= bitmap_len || neg ||
2529                                     val_a > val_b) {
2530                                         err = -EINVAL;
2531                                         break;
2532                                 }
2533                                 if (left) {
2534                                         kbuf++;
2535                                         left--;
2536                                 }
2537                         }
2538
2539                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2540                         first = 0;
2541                         proc_skip_char(&kbuf, &left, '\n');
2542                 }
2543                 free_page(page);
2544         } else {
2545                 unsigned long bit_a, bit_b = 0;
2546
2547                 while (left) {
2548                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2549                         if (bit_a >= bitmap_len)
2550                                 break;
2551                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2552                                                    bit_a + 1) - 1;
2553
2554                         if (!first) {
2555                                 err = proc_put_char(&buffer, &left, ',');
2556                                 if (err)
2557                                         break;
2558                         }
2559                         err = proc_put_long(&buffer, &left, bit_a, false);
2560                         if (err)
2561                                 break;
2562                         if (bit_a != bit_b) {
2563                                 err = proc_put_char(&buffer, &left, '-');
2564                                 if (err)
2565                                         break;
2566                                 err = proc_put_long(&buffer, &left, bit_b, false);
2567                                 if (err)
2568                                         break;
2569                         }
2570
2571                         first = 0; bit_b++;
2572                 }
2573                 if (!err)
2574                         err = proc_put_char(&buffer, &left, '\n');
2575         }
2576
2577         if (!err) {
2578                 if (write) {
2579                         if (*ppos)
2580                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2581                         else
2582                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2583                 }
2584                 kfree(tmp_bitmap);
2585                 *lenp -= left;
2586                 *ppos += *lenp;
2587                 return 0;
2588         } else {
2589                 kfree(tmp_bitmap);
2590                 return err;
2591         }
2592 }
2593
2594 #else /* CONFIG_PROC_SYSCTL */
2595
2596 int proc_dostring(struct ctl_table *table, int write,
2597                   void __user *buffer, size_t *lenp, loff_t *ppos)
2598 {
2599         return -ENOSYS;
2600 }
2601
2602 int proc_dointvec(struct ctl_table *table, int write,
2603                   void __user *buffer, size_t *lenp, loff_t *ppos)
2604 {
2605         return -ENOSYS;
2606 }
2607
2608 int proc_dointvec_minmax(struct ctl_table *table, int write,
2609                     void __user *buffer, size_t *lenp, loff_t *ppos)
2610 {
2611         return -ENOSYS;
2612 }
2613
2614 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2615                     void __user *buffer, size_t *lenp, loff_t *ppos)
2616 {
2617         return -ENOSYS;
2618 }
2619
2620 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2621                     void __user *buffer, size_t *lenp, loff_t *ppos)
2622 {
2623         return -ENOSYS;
2624 }
2625
2626 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2627                              void __user *buffer, size_t *lenp, loff_t *ppos)
2628 {
2629         return -ENOSYS;
2630 }
2631
2632 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2633                     void __user *buffer, size_t *lenp, loff_t *ppos)
2634 {
2635         return -ENOSYS;
2636 }
2637
2638 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2639                                       void __user *buffer,
2640                                       size_t *lenp, loff_t *ppos)
2641 {
2642     return -ENOSYS;
2643 }
2644
2645
2646 #endif /* CONFIG_PROC_SYSCTL */
2647
2648 /*
2649  * No sense putting this after each symbol definition, twice,
2650  * exception granted :-)
2651  */
2652 EXPORT_SYMBOL(proc_dointvec);
2653 EXPORT_SYMBOL(proc_dointvec_jiffies);
2654 EXPORT_SYMBOL(proc_dointvec_minmax);
2655 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2656 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2657 EXPORT_SYMBOL(proc_dostring);
2658 EXPORT_SYMBOL(proc_doulongvec_minmax);
2659 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);